newt-core 0.7.1

//! Built-in tool definitions and the tool executor for the agentic loop.
//! Moved verbatim from `newt-tui` in Step 9.7 — the Caveats enforcement,
//! shrink guard, build-check feedback, and agent-bridle routing are unchanged.

use super::crew_tool::CrewRunner;
use super::display::{print_denied, print_tool_call, print_tool_output};
use super::git_tool::GitTool;
use super::mcp::McpTools;
use super::memory_fetch::{execute_memory_fetch, memory_fetch_tool_definition, MemorySource};
use super::note_sink::{execute_save_note, save_note_tool_definition, NoteSink};
use super::permissions::{DenialKind, PermissionDecision, PermissionGate, PermissionRequest};
use super::recall::{execute_recall, recall_tool_definition, RecallSource};
use super::spill::{self, SpillStore};
use crate::caveats::CaveatsExt as _;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::LazyLock;

/// #719: default line window for `read_file`'s **model-facing** payload. The
/// on-screen display is capped separately; this bounds what enters the model's
/// context, so one read of a 15k-line file (e.g. `newt-tui/src/lib.rs`) can no
/// longer saturate a small local model's window and abandon the task.
const DEFAULT_READ_LIMIT: usize = 2_000;

/// #726: default token budget for any tool's **model-facing** payload, mirroring
/// Codex's `exec_command.max_output_tokens` (default 10k). One shared budget
/// caps BOTH `read_file` (via [`paginate_read`]'s char backstop) and
/// `run_command` (via [`cap_model_output`] around the shell envelope), so a
/// verbose command can no longer flood the window — the same failure mode #719
/// closed for `read_file`. Overridable by `[tools] max_output_tokens` in config;
/// see [`set_max_output_tokens`].
const DEFAULT_MAX_OUTPUT_TOKENS: usize = 10_000;
const DEFAULT_OUTPUT_HEAD_TOKENS: usize = 1_500;

/// Process-wide model-facing output budget, in tokens. Defaults to
/// [`DEFAULT_MAX_OUTPUT_TOKENS`]; the resolved `[tools] max_output_tokens`
/// config value is pushed here once at the config-resolution entry
/// (`Config::resolve`) so the tool loop never re-reads config from disk. This is
/// the v1 (three-Cs "working code first") seam: a const default with the config
/// override wired at the entry, rather than threading a new `usize` through
/// `ChatCtx` + `execute_tool` + every call site (≈60, mostly tests). Follow-up:
/// thread it per-session like `tool_output_lines` once warranted.
static MAX_OUTPUT_TOKENS: AtomicUsize = AtomicUsize::new(DEFAULT_MAX_OUTPUT_TOKENS);
static OUTPUT_HEAD_TOKENS: AtomicUsize = AtomicUsize::new(DEFAULT_OUTPUT_HEAD_TOKENS);

/// Set the process-wide model-facing output budget (tokens). Called once from
/// `Config::resolve` with the resolved `[tools] max_output_tokens`. `0` means
/// "no cap" — see [`cap_model_output`] / [`paginate_read`].
pub fn set_max_output_tokens(max_tokens: usize) {
    MAX_OUTPUT_TOKENS.store(max_tokens, Ordering::Relaxed);
}

/// Set the head allocation for oversized `run_command` output. The tail gets
/// the remaining budget. `0` means pure-tail; values greater than the max output
/// budget are clamped by [`cap_model_output`].
pub fn set_output_head_tokens(head_tokens: usize) {
    OUTPUT_HEAD_TOKENS.store(head_tokens, Ordering::Relaxed);
}

/// The active model-facing output budget (tokens). [`DEFAULT_MAX_OUTPUT_TOKENS`]
/// until [`set_max_output_tokens`] overrides it.
fn max_output_tokens() -> usize {
    MAX_OUTPUT_TOKENS.load(Ordering::Relaxed)
}

fn output_head_tokens() -> usize {
    OUTPUT_HEAD_TOKENS.load(Ordering::Relaxed)
}

/// #726/#945: cap a tool's **model-facing** output to `max_tokens`' worth of
/// chars, estimated with the default chars/token heuristic
/// ([`crate::tokens::TokenEstimation`], 4 chars/token — the same constant the
/// context estimator uses). Oversized output is rendered as head+tail rather
/// than head-only so command summaries and failures at the end survive. A small
/// output (or `max_tokens == 0`, meaning no cap) passes through verbatim. Pure
/// (no fs / no global) — unit-tested directly.
fn cap_model_output(text: &str, max_tokens: usize) -> String {
    cap_model_output_with_handle(text, max_tokens, output_head_tokens(), None)
}

fn cap_model_output_with_handle(
    text: &str,
    max_tokens: usize,
    head_tokens: usize,
    spill_id: Option<&str>,
) -> String {
    if max_tokens == 0 {
        return text.to_string();
    }
    let est = crate::tokens::TokenEstimation::default();
    if est.tokens_for_chars(text.len()) <= max_tokens {
        return text.to_string();
    }
    let max_chars = est.chars_for_tokens(max_tokens);
    let head_tokens = head_tokens.min(max_tokens);
    let head_chars = est.chars_for_tokens(head_tokens).min(max_chars);
    let tail_chars = max_chars.saturating_sub(head_chars);
    let total_chars = text.chars().count();
    let shown_chars = head_chars.saturating_add(tail_chars).min(total_chars);
    let elided = total_chars.saturating_sub(shown_chars);
    let head = take_chars(text, head_chars);
    let tail = take_tail_chars(text, tail_chars);
    let marker = match spill_id {
        Some(id) => format!(
            "[… {elided} chars elided (head+tail shown). Full output: \
             memory_fetch(\"spill:{id}\"); search it with \
             memory_fetch(\"spill:{id}\", grep=\"<pattern>\") …]"
        ),
        None => format!(
            "[… {elided} chars elided (head+tail shown; ~{max_tokens} token budget). \
             Narrow the command or use a more specific grep/filter if needed …]"
        ),
    };
    format!("{head}\n\n{marker}\n\n{tail}")
}

fn take_chars(text: &str, max_chars: usize) -> String {
    text.chars().take(max_chars).collect()
}

fn take_tail_chars(text: &str, max_chars: usize) -> String {
    let mut chars: Vec<char> = text.chars().rev().take(max_chars).collect();
    chars.reverse();
    chars.into_iter().collect()
}

/// Window + cap a file's contents for `read_file`'s model-facing payload (#719,
/// #726). Returns lines `[offset, offset+limit)` (1-based `offset`, default 1;
/// `limit` default [`DEFAULT_READ_LIMIT`]), with the char backstop derived from
/// the shared token budget (`max_output_tokens` × chars/token — #726, replacing
/// #719's hardcoded 100k so both tools share one budget). A footer points at the
/// next window so the model paginates instead of drowning. A whole-file read
/// that fits both caps is returned verbatim (exact bytes). `max_output_tokens ==
/// 0` disables the char backstop (only the line window applies). Pure (no fs) —
/// unit-tested directly.
fn paginate_read(
    contents: &str,
    offset: Option<usize>,
    limit: Option<usize>,
    max_output_tokens: usize,
) -> String {
    let max_chars = if max_output_tokens == 0 {
        usize::MAX
    } else {
        crate::tokens::TokenEstimation::default().chars_for_tokens(max_output_tokens)
    };
    let total = contents.lines().count();
    let start = offset.filter(|&o| o > 0).unwrap_or(1); // 1-based
    let limit = limit.filter(|&l| l > 0).unwrap_or(DEFAULT_READ_LIMIT);
    // Common case: a whole-file read that fits both caps → return verbatim.
    if start == 1 && limit >= total && contents.len() <= max_chars {
        return contents.to_string();
    }
    let start0 = start - 1;
    if start0 >= total {
        return format!("(offset {start} is past end of file — {total} lines total)");
    }
    let window: Vec<&str> = contents.lines().skip(start0).take(limit).collect();
    let end = start0 + window.len(); // 1-based last line shown == end
    let mut body = window.join("\n");
    let char_capped = body.len() > max_chars;
    if char_capped {
        let mut cut = max_chars;
        while cut > 0 && !body.is_char_boundary(cut) {
            cut -= 1;
        }
        body.truncate(cut);
    }
    let footer = if char_capped {
        Some(format!(
            "payload truncated to {max_chars} chars (~{max_output_tokens} tokens) from line \
             {start}; call read_file with a higher offset (and/or smaller limit) to continue"
        ))
    } else if end < total {
        Some(format!(
            "showing lines {start}-{end} of {total}; \
             call read_file with offset={} to continue",
            end + 1
        ))
    } else {
        None
    };
    match footer {
        Some(f) => format!("{body}\n\n[{f}]"),
        None => body,
    }
}

pub fn tool_definitions() -> serde_json::Value {
    serde_json::json!([
        {
            "type": "function",
            "function": {
                "name": "run_command",
                "description": "Run a shell command in the workspace directory and return its output",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "command": { "type": "string", "description": "The shell command to run" }
                    },
                    "required": ["command"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "read_file",
                "description": "Read a file in the workspace. Returns up to `limit` lines \
                                (default 2000) starting at 1-based `offset` (default 1). Large \
                                files come back with a footer pointing at the next window — read \
                                them in pages with offset/limit rather than all at once, or the \
                                full file can saturate the context window.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "path": { "type": "string", "description": "File path relative to workspace root" },
                        "offset": { "type": "integer", "description": "1-based line number to start at (default 1)" },
                        "limit": { "type": "integer", "description": "Maximum number of lines to return (default 2000)" }
                    },
                    "required": ["path"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "write_file",
                "description": "Write or overwrite a file in the workspace. \
                                WARNING: use edit_file instead when modifying an existing file — \
                                write_file replaces the entire contents and will fail if the new \
                                content is significantly shorter than the original (shrink guard). \
                                Only use write_file for new files or full rewrites you have \
                                explicitly generated in their entirety.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "path": { "type": "string", "description": "File path relative to workspace root" },
                        "content": { "type": "string", "description": "The complete new file contents" }
                    },
                    "required": ["path", "content"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "edit_file",
                "description": "Make a targeted edit to an existing file by replacing one exact \
                                string with another. Safer than write_file for modifying existing \
                                files — you only generate the change, not the whole file. \
                                Fails with a clear error if old_string is not found or matches \
                                multiple times (add more surrounding context to make it unique).",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "path": { "type": "string", "description": "File path relative to workspace root" },
                        "old_string": { "type": "string", "description": "Exact string to find and replace (must match exactly once)" },
                        "new_string": { "type": "string", "description": "Replacement string" }
                    },
                    "required": ["path", "old_string", "new_string"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "list_dir",
                "description": "List files in a directory",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "path": { "type": "string", "description": "Directory path relative to workspace root (use '.' for root)" }
                    },
                    "required": ["path"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "find",
                "description": "Find files and directories by name under the workspace, recursively, WITHOUT a shell (use this instead of the `find` shell command). Returns matching paths relative to the workspace root, one per line, already sorted — no need to pipe to `sort`. Respects .gitignore and skips noise (.git, target, node_modules) by default.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "path": { "type": "string", "description": "Directory to search under, relative to workspace root. Default '.' (the whole workspace)." },
                        "name": { "type": "string", "description": "Glob matched against each entry's basename, e.g. '*.py' or 'pyo3_module.rs'. '*' matches any run, '?' any single char. Omit to match everything." },
                        "type": { "type": "string", "enum": ["f", "d", "any"], "description": "Restrict to files ('f'), directories ('d'), or both ('any', the default)." },
                        "max_depth": { "type": "integer", "description": "Maximum directory depth below `path` (1 = immediate children only). Omit for unlimited." },
                        "max_results": { "type": "integer", "description": "Cap on the number of matches returned. Default 1000; output notes when truncated." },
                        "respect_gitignore": { "type": "boolean", "description": "When true (default) skip .gitignored paths plus .git/target/node_modules/hidden dirs. Set false to search everything." },
                        "case_sensitive": { "type": "boolean", "description": "Case-sensitive basename match. Default true." }
                    },
                    "required": []
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "use_skill",
                "description": "Load a skill's full procedural instructions on demand. The system prompt lists the available skills (name + description); call this with a skill's name to get its complete SKILL.md body plus the paths of any bundled files (scripts/templates) you can read or run.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "name": { "type": "string", "description": "The skill name as shown in the 'Available skills' index" }
                    },
                    "required": ["name"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "web_fetch",
                "description": "Fetch an http(s) URL and return its main content as clean markdown. Use this to read documentation, issues, or pages the task references. Reachable hosts are gated by the session's network capability; the returned text is untrusted page content.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "url": { "type": "string", "description": "The http(s) URL to fetch" },
                        "max_bytes": { "type": "integer", "description": "Optional cap on bytes downloaded (default 5 MiB, max 25 MiB)" }
                    },
                    "required": ["url"]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "request_permissions",
                "description": "Ask the operator to GRANT a capability you were denied — the \
                                capability-grant path (#721). Call this AFTER a `capability denied` \
                                result to request authority you don't currently have. If an operator \
                                is present they may allow it; then retry the original operation. In a \
                                headless session (no operator) you'll be told the capability must be \
                                configured by the owner — change approach. This requests AUTHORITY \
                                (it mints a capability grant); it is NOT a way to ask the user a \
                                free-text question.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "capability": { "type": "string", "enum": ["exec", "fs_read", "fs_write", "net"], "description": "Which capability axis to request" },
                        "target": { "type": "string", "description": "What to grant: a command name (exec), a path (fs_read/fs_write), or a host (net)" },
                        "reason": { "type": "string", "description": "Why you need it — shown to the operator deciding" }
                    },
                    "required": ["capability", "target", "reason"]
                }
            }
        }
    ])
}

/// #728: always-advertised `request_user_input` tool definition — the GENERIC
/// ask-the-human primitive. Pushed unconditionally in [`merged_tool_definitions`]
/// like `resume_context` / `tool_search` / `get_context_remaining`: a model must
/// always be able to ask the human a question, and it degrades honestly headless
/// (the executor answers "no human available this session" when no interactive
/// gate is present). Question-only for v1 — the multiple-choice `options` hint is
/// deferred so no advertised arg is left unused.
fn request_user_input_tool_definition() -> serde_json::Value {
    serde_json::json!({
        "type": "function",
        "function": {
            "name": "request_user_input",
            "description": "Ask the human operator a free-text question and get \
                            their answer. Use this to resolve genuine ambiguity \
                            instead of guessing or narrating (e.g. 'which database \
                            should I target?', 'is this the file you meant?'). This \
                            asks for INFORMATION, not authority — to request a \
                            capability you were denied, use request_permissions \
                            instead. In a headless session (no operator) you'll be \
                            told no human is available — then proceed with your best \
                            judgment and state your assumption explicitly.",
            "parameters": {
                "type": "object",
                "properties": {
                    "question": { "type": "string", "description": "The free-text question to ask the human" }
                },
                "required": ["question"]
            }
        }
    })
}

/// The `lifecycle` tool (#891) — the model-facing surface over the data-driven
/// lifecycle system (#880). Instead of guessing a raw shell command, the model
/// names a phase and newt runs THIS repo's resolved command for it
/// ([`crate::tooling::resolved_phase_commands`]): the `.newt/config.toml`
/// `[lifecycle]` override, else the matching tooling packs (Rust / Python /
/// PyO3 / Go / drop-in / custom). Advertised ALWAYS — like `resume_context`,
/// it degrades honestly (a phase with no configured command returns a clear
/// "no command configured", not an error), so it needs no presence gate. The
/// phase-name enum is built from [`crate::tooling::Phase::ALL`] so the schema
/// can never drift from the vocabulary.
pub fn lifecycle_tool_definition() -> serde_json::Value {
    let phases: Vec<&str> = crate::tooling::Phase::ALL
        .iter()
        .map(|p| p.as_str())
        .collect();
    serde_json::json!({
        "type": "function",
        "function": {
            "name": "lifecycle",
            "description": "Run a named project lifecycle phase using THIS repo's \
                configured command instead of guessing a raw shell command. \
                Phases: setup (resolve deps / prepare a checkout), format \
                (auto-format the tree), lint (static analysis), test (run the \
                tests), check (the full gate a change must pass), clean (remove \
                build artifacts). The command is resolved from \
                .newt/config.toml [lifecycle] and the repo's tooling packs \
                (Rust / Python / PyO3 / Go / custom), so `check` runs the RIGHT \
                gate for this project. Prefer this over run_command for build / \
                test / format / lint / check work so the project's own \
                conventions are honored uniformly across build systems. Use \
                action=list to see the resolved command without running it.",
            "parameters": {
                "type": "object",
                "properties": {
                    "phase": {
                        "type": "string",
                        "enum": phases,
                        "description": "The lifecycle phase to run."
                    },
                    "action": {
                        "type": "string",
                        "enum": ["run", "list"],
                        "description": "run (default) executes the phase's resolved command; \
                                        list returns the command without running it."
                    }
                },
                "required": ["phase"]
            }
        }
    })
}

/// The built-in tool definitions plus every connected MCP server's tools
/// (namespaced `server__tool`). This is what the agent loop advertises to the
/// model so it can call remote MCP tools alongside the built-ins.
///
/// `with_save_note` adds the `save_note` definition (Step 19.3) — true only
/// when the caller supplied a `NoteSink`, so headless/eval sessions (which
/// pass `note_sink: None`) never advertise a tool that can't be executed.
/// `with_recall` gates the `recall` definition (Step 17.5) the same way on
/// a supplied `RecallSource`. `with_memory_fetch` gates the `memory_fetch`
/// definition (progressive-disclosure memory, Workstream A MVP, #319) the same
/// way on a supplied `MemorySource` — `None` ⇒ the tool is never advertised,
/// so eval / headless / ACP sessions are unaffected bit-for-bit.
#[allow(clippy::too_many_arguments)] // presence-gated tool advertisers (one bool each)
pub(crate) fn merged_tool_definitions(
    mcp: &dyn McpTools,
    with_save_note: bool,
    with_recall: bool,
    with_memory_fetch: bool,
    with_git: bool,
    with_team: bool,
    with_scratchpad: bool,
    with_code_search: bool,
    with_experiential: bool,
    with_scheduled: bool,
) -> serde_json::Value {
    let mut defs = match tool_definitions() {
        serde_json::Value::Array(a) => a,
        other => vec![other],
    };
    // #894: everything after the base array is [`EXTENDED_TOOL_REGISTRY`],
    // advertised in declaration order when its gate is satisfied. The always-on
    // tools (resume_context #714 / tool_search #725 / get_context_remaining #727
    // / request_user_input #728 / lifecycle #891) carry `Gate::Always` and ride
    // every session, degrading honestly when their backing source is `None`; the
    // rest are gated on the matching injected `with_*` capability (git PR4,
    // compose_roster+crew #479, scratchpad #583, code_search #582, experiential
    // #585, scheduled #586/#715/#716). Adding a tool is one `ToolSpec`, which
    // updates this listing AND `ALL_TOOL_NAMES` together — no more drift.
    for spec in EXTENDED_TOOL_REGISTRY {
        if gate_satisfied(
            spec.gate,
            with_save_note,
            with_recall,
            with_memory_fetch,
            with_git,
            with_team,
            with_scratchpad,
            with_code_search,
            with_experiential,
            with_scheduled,
        ) {
            defs.push((spec.definition)());
        }
    }
    defs.extend(mcp.tool_defs());
    serde_json::Value::Array(defs)
}

/// Direct tool names the model must call as tool invocations, never as shell
/// commands passed to `run_command`.
const DIRECT_TOOL_NAMES: &[&str] = &[
    "list_dir",
    "read_file",
    "write_file",
    "edit_file",
    "use_skill",
    "web_fetch",
    // #496: `find …` typed at run_command redirects to the embedded `find`
    // tool — which works even when the shell is unavailable in this build.
    "find",
    // PR4: `git …` typed at run_command redirects to the embedded `git` tool —
    // but ONLY its LOCAL ops; the network ops fall through (see
    // [`GIT_NETWORK_SUBCOMMANDS`] / [`run_command_redirect`], #898).
    "git",
];

/// #898: git subcommands that reach the network. The embedded `git` tool
/// (newt-git) is LOCAL-ONLY — `clone`/`fetch`/`push` are deferred — so if
/// run_command bounced *every* `git …` back to that pushless tool, a model
/// could never push a branch (and then never see the "Create a pull request …
/// by visiting: <URL>" line git prints, and never open a PR — issue #898).
/// These ops are therefore allowed to fall through to the confined shell, where
/// the `net` caveat still gates whether the remote is reachable. Local ops
/// (`status`/`log`/`diff`/`add`/`commit`/…) keep redirecting to the embedded
/// tool, which does them better and works even when the shell is unavailable.
const GIT_NETWORK_SUBCOMMANDS: &[&str] = &["push", "fetch", "pull", "clone"];

/// Decide whether a `run_command` invocation is really a misdirected call to a
/// direct tool (`list_dir`/`read_file`/…/`git`), and if so which one — so the
/// executor can bounce it with a correction and [`is_hallucination`] can count
/// it. Returns `None` when the command should run in the shell as-is.
///
/// `git` is special (#898): only its LOCAL ops redirect to the embedded git
/// tool; its network ops ([`GIT_NETWORK_SUBCOMMANDS`]) fall through so the model
/// can actually push a branch and read the PR-creation URL git prints.
fn run_command_redirect(command: &str) -> Option<&'static str> {
    let mut tokens = command.split_ascii_whitespace();
    let first = tokens.next().unwrap_or("");
    if first == "git" {
        let sub = tokens.next().unwrap_or("");
        if GIT_NETWORK_SUBCOMMANDS.contains(&sub) {
            return None;
        }
        return Some("git");
    }
    DIRECT_TOOL_NAMES.iter().copied().find(|&t| t == first)
}

/// #894: the built-in tool registry — ONE self-describing entry per non-base
/// tool (name + JSON schema builder + presence gate), replacing the parallel
/// hand-kept lists that used to drift (the `lifecycle` tool, #891, was
/// advertised + dispatched yet missing from the old `ALL_TOOL_NAMES`, so every
/// legitimate `lifecycle` call was miscounted as a hallucination). Adding a
/// tool here now updates BOTH the advertised set ([`merged_tool_definitions`])
/// AND the real-name set ([`ALL_TOOL_NAMES`]) atomically — you cannot half-wire
/// one without the other. The base tools stay inlined in [`tool_definitions`]
/// (their names mirrored in [`BASE_TOOL_NAMES`], guarded by a test); the
/// executor dispatch match is intentionally left as a separate pass.
///
/// Presence condition for a registered tool. `Always` tools ride every session
/// (they degrade honestly when their backing source is absent); the rest are
/// advertised only when the matching `with_*` capability is injected.
#[derive(Clone, Copy, PartialEq, Eq)]
enum Gate {
    Always,
    SaveNote,
    Recall,
    MemoryFetch,
    Git,
    Team,
    Scratchpad,
    CodeSearch,
    Experiential,
    Scheduled,
}

/// One built-in (non-base) tool, declared in exactly one place.
struct ToolSpec {
    /// The tool name the model calls — must equal `(definition)()`'s function name.
    name: &'static str,
    /// Builds the tool's JSON schema (the same `*_tool_definition()` fns used
    /// before; the registry just references them).
    definition: fn() -> serde_json::Value,
    /// When this tool is advertised + treated as a real (non-hallucinated) name.
    gate: Gate,
}

/// The registered non-base tools, in advertised order. This IS the order
/// [`merged_tool_definitions`] pushes them (after the base array), preserved
/// byte-for-byte from the previous hand-written push ladder.
const EXTENDED_TOOL_REGISTRY: &[ToolSpec] = &[
    // Always-on (degrade gracefully when their source is None) —
    // #714 / #725 / #727 / #728 / #891.
    ToolSpec {
        name: "resume_context",
        definition: super::resume::resume_context_tool_definition,
        gate: Gate::Always,
    },
    ToolSpec {
        name: "tool_search",
        definition: super::tool_search::tool_search_tool_definition,
        gate: Gate::Always,
    },
    ToolSpec {
        name: "get_context_remaining",
        definition: super::budget::get_context_remaining_tool_definition,
        gate: Gate::Always,
    },
    ToolSpec {
        name: "request_user_input",
        definition: request_user_input_tool_definition,
        gate: Gate::Always,
    },
    ToolSpec {
        name: "lifecycle",
        definition: lifecycle_tool_definition,
        gate: Gate::Always,
    },
    // Presence-gated on an injected capability (one `with_*` flag each).
    ToolSpec {
        name: "save_note",
        definition: save_note_tool_definition,
        gate: Gate::SaveNote,
    },
    ToolSpec {
        name: "recall",
        definition: recall_tool_definition,
        gate: Gate::Recall,
    },
    ToolSpec {
        name: "memory_fetch",
        definition: memory_fetch_tool_definition,
        gate: Gate::MemoryFetch,
    },
    ToolSpec {
        name: "git",
        definition: super::git_tool::git_tool_definition,
        gate: Gate::Git,
    },
    ToolSpec {
        name: "compose_roster",
        definition: super::crew_tool::compose_roster_tool_definition,
        gate: Gate::Team,
    },
    ToolSpec {
        name: "crew",
        definition: super::crew_tool::crew_tool_definition,
        gate: Gate::Team,
    },
    ToolSpec {
        name: "state_set",
        definition: super::scratchpad::state_set_tool_definition,
        gate: Gate::Scratchpad,
    },
    ToolSpec {
        name: "state_get",
        definition: super::scratchpad::state_get_tool_definition,
        gate: Gate::Scratchpad,
    },
    ToolSpec {
        name: "state_clear",
        definition: super::scratchpad::state_clear_tool_definition,
        gate: Gate::Scratchpad,
    },
    ToolSpec {
        name: "code_search",
        definition: super::semantic::code_search_tool_definition,
        gate: Gate::CodeSearch,
    },
    ToolSpec {
        name: "experience_record",
        definition: super::experiential::experience_record_tool_definition,
        gate: Gate::Experiential,
    },
    ToolSpec {
        name: "experience_recall",
        definition: super::experiential::experience_recall_tool_definition,
        gate: Gate::Experiential,
    },
    ToolSpec {
        name: "update_plan",
        definition: super::scheduled::update_plan_tool_definition,
        gate: Gate::Scheduled,
    },
    ToolSpec {
        name: "plan_get",
        definition: super::scheduled::plan_get_tool_definition,
        gate: Gate::Scheduled,
    },
];

/// The base tools inlined in [`tool_definitions`], by name. The base array is
/// the one place these tools are declared; this mirror exists only so
/// [`ALL_TOOL_NAMES`] can be assembled without re-parsing JSON, and it is kept
/// in lockstep by `base_tool_names_match_tool_definitions`. The EXTENDED tools'
/// names are NOT duplicated here — they live on their [`ToolSpec`].
const BASE_TOOL_NAMES: &[&str] = &[
    "run_command",
    "read_file",
    "write_file",
    "edit_file",
    "list_dir",
    "find",
    "use_skill",
    "web_fetch",
    "request_permissions",
];

/// Every tool newt can dispatch by name — the base tools plus every registered
/// one — DERIVED from [`BASE_TOOL_NAMES`] + [`EXTENDED_TOOL_REGISTRY`] so it can
/// never drift from the advertised set. Single source of truth for
/// [`is_hallucination`] (which names are real) and [`nearest_tool_name`]
/// (suggestion candidates). MCP `server__tool` names contain `__` and are
/// matched separately.
static ALL_TOOL_NAMES: LazyLock<Vec<&'static str>> = LazyLock::new(|| {
    BASE_TOOL_NAMES
        .iter()
        .copied()
        .chain(EXTENDED_TOOL_REGISTRY.iter().map(|s| s.name))
        .collect()
});

/// Whether `tool_name` is a built-in newt tool name. Dynamic MCP tool names are
/// intentionally excluded; callers that need MCP should check their live MCP
/// registry separately.
pub(crate) fn known_builtin_tool_name(tool_name: &str) -> bool {
    ALL_TOOL_NAMES.contains(&tool_name)
}

/// Whether a [`Gate`] is satisfied given this session's injected capabilities.
/// Extracted so [`merged_tool_definitions`] reads as one loop over the registry.
#[allow(clippy::too_many_arguments)] // mirrors merged_tool_definitions' with_* flags
fn gate_satisfied(
    gate: Gate,
    with_save_note: bool,
    with_recall: bool,
    with_memory_fetch: bool,
    with_git: bool,
    with_team: bool,
    with_scratchpad: bool,
    with_code_search: bool,
    with_experiential: bool,
    with_scheduled: bool,
) -> bool {
    match gate {
        Gate::Always => true,
        Gate::SaveNote => with_save_note,
        Gate::Recall => with_recall,
        Gate::MemoryFetch => with_memory_fetch,
        Gate::Git => with_git,
        Gate::Team => with_team,
        Gate::Scratchpad => with_scratchpad,
        Gate::CodeSearch => with_code_search,
        Gate::Experiential => with_experiential,
        Gate::Scheduled => with_scheduled,
    }
}

/// Returns `true` if a tool call looks like a hallucination:
/// - `run_command` called with a tool name as the shell command, or
/// - An unknown tool name (excluding MCP-namespaced `server__tool` names).
pub(crate) fn is_hallucination(tool_name: &str, args: &serde_json::Value) -> bool {
    if tool_name == "run_command" {
        let cmd = args["command"].as_str().unwrap_or("");
        // A misdirected direct-tool call is a hallucination; a real shell
        // command (including the git NETWORK ops #898 lets through) is not.
        return run_command_redirect(cmd).is_some();
    }
    // MCP tools are namespaced with `__` — never treat them as hallucinations.
    if tool_name.contains("__") {
        return false;
    }
    // #894: derived from BASE_TOOL_NAMES + EXTENDED_TOOL_REGISTRY, so it can
    // never drift from the advertised set.
    !ALL_TOOL_NAMES.contains(&tool_name)
}

/// Outcome of resolving a foreign / hallucinated tool name against newt's real
/// tools (Step 27.1). Weak local models routinely emit tool names learned from
/// other agent harnesses (`str_replace_editor`, `execute`, `bash`); without a
/// resolution layer each such call is a flat "unknown tool" that burns a whole
/// tool-call round, so the model narrates instead of acting (the #215 advisory
/// drift). Resolving them turns a wasted round into a self-correcting one.
pub(crate) enum AliasOutcome {
    /// A foreign name whose argument shape matches a real newt tool: rewrite the
    /// call to this canonical name and dispatch it transparently.
    Rewrite(&'static str),
    /// A foreign name whose arguments do NOT match the real tool: return this
    /// correction (naming the right tool + its signature) so the model retries.
    Correct(String),
}

/// Map a non-newt tool name to a real tool, when we recognize it. Real tool
/// names and MCP `server__tool` names return `None` and dispatch unchanged.
pub(crate) fn resolve_tool_alias(name: &str) -> Option<AliasOutcome> {
    match name {
        // Shell aliases — same single `command` arg shape as run_command, so we
        // rewrite and dispatch. (If the shell is unavailable this build, the
        // run_command arm reports that — Step 27.3 presence-gates it.)
        "execute" | "exec" | "bash" | "shell" | "sh" | "zsh" | "terminal" | "run_shell_command"
        | "shell_command" | "system" => Some(AliasOutcome::Rewrite("run_command")),
        // #891 lifecycle aliases — same `{phase, action}` arg shape as
        // `lifecycle`, so we rewrite and dispatch. Common phrasings a model
        // reaches for when it wants to run a named phase.
        "run_phase" | "run_lifecycle" | "lifecycle_run" => Some(AliasOutcome::Rewrite("lifecycle")),
        // Edit aliases — different arg shape; point at edit_file's signature.
        "str_replace_editor" | "str_replace" | "str-replace-editor" | "apply_patch" | "edit"
        | "editor" | "replace_in_file" | "search_replace" => Some(AliasOutcome::Correct(format!(
            "'{name}' is not a newt tool. To change an existing file, call \
                 edit_file with {{\"path\", \"old_string\", \"new_string\"}} \
                 (replaces one exact occurrence). For a new file or a full \
                 rewrite, call write_file with {{\"path\", \"content\"}}."
        ))),
        // Create-file aliases — point at write_file.
        "create_file" | "new_file" | "createfile" | "add_file" | "touch" => {
            Some(AliasOutcome::Correct(format!(
                "'{name}' is not a newt tool. To create or overwrite a file, call \
                 write_file with {{\"path\", \"content\"}}. To change part of an \
                 existing file, call edit_file with \
                 {{\"path\", \"old_string\", \"new_string\"}}."
            )))
        }
        // #721 mkdir coach — newt has no directory-creation tool, and the model
        // does not need one: write_file creates parent directories automatically
        // (create_dir_all), and an empty file for empty content. A model reaching
        // for `mkdir` (the issue's live `mkdir -p …/src` dead-end) is coached to
        // the tool that already covers it, turning an exec denial into a
        // self-correcting tool call. `touch` is intentionally NOT here — it is
        // already a create-file alias above (→ write_file), and a second arm
        // would be a duplicate match.
        "mkdir" | "make_dir" | "makedirs" | "mkdirs" | "create_dir" | "create_directory" => {
            Some(AliasOutcome::Correct(
                "newt has no mkdir/touch tool — call write_file; it creates parent \
                 directories automatically (create_dir_all). For an empty file, call \
                 write_file with empty content."
                    .to_string(),
            ))
        }
        // Read / list aliases — point at read_file / list_dir.
        "cat" | "open_file" | "view_file" | "view" | "open" => {
            Some(AliasOutcome::Correct(format!(
                "'{name}' is not a newt tool. To read a file, call read_file with \
                 {{\"path\"}}. To list a directory, call list_dir with {{\"path\"}}."
            )))
        }
        // #716 + #715 PR2 PLAN — start/revise a plan. The arg shape is free prose,
        // not the ordered `{"plan":[{"step","status"}]}` array update_plan wants, so
        // Correct (coach), never a silent Rewrite. When `scheduled` is off the
        // dispatch arm for update_plan returns "scheduled planning is off" — the
        // model still gets a coherent answer rather than a dead end. `update_plan`
        // itself is the REAL tool now (falls through to `None`), so it is never an
        // alias of itself; `set_plan`/`plan_advance` no longer exist.
        "enter_plan" | "enter_plan_mode" | "plan_mode" | "start_plan" | "begin_plan"
        | "make_plan" | "create_plan" | "plan" | "planning" | "todo" | "todos" | "todo_write" => {
            Some(AliasOutcome::Correct(format!(
                "'{name}' is not a newt tool. To start or revise your plan, call update_plan with \
                 {{\"plan\":[{{\"step\",\"status\"}}]}} — send the full ordered list each time, \
                 each step's status one of pending/in_progress/completed (exactly one \
                 in_progress)."
            )))
        }
        // #715 PR2 ADVANCE-ish verbs — there is no longer a separate "advance" tool;
        // progress is recorded by re-sending the whole plan with the finished step
        // marked completed. Coach the model back to update_plan.
        "next_step" | "complete_step" | "finish_step" | "mark_done" | "step_done" => {
            Some(AliasOutcome::Correct(format!(
                "'{name}' is not a newt tool. To advance your plan, call update_plan with the \
                 full plan and mark the finished step \"completed\" (and the next one \
                 \"in_progress\")."
            )))
        }
        // #716 PLAN-READ — read the current plan. plan_get takes no args, so the
        // foreign call's (empty) arg shape matches: safe to silently Rewrite.
        // `what_was_i_doing` stays here (→ plan_get) — it asks specifically for
        // the plan; the broader "where were we" reaches go to resume_context.
        "get_plan" | "show_plan" | "read_plan" | "current_plan" | "what_was_i_doing" => {
            Some(AliasOutcome::Rewrite("plan_get"))
        }
        // #714 RESUME — the instinctive "where did we leave off" reach. All take
        // no args (resume_context is a self-read), so the (empty) arg shape
        // matches: safe to silently Rewrite. Meets the dead-end reach the issue
        // observed (the model retrying recall) by landing it on the affordance
        // built for exactly this case.
        "resume" | "where_were_we" | "where_did_we_leave_off" | "catch_me_up" | "recap" => {
            Some(AliasOutcome::Rewrite("resume_context"))
        }
        // #716 CREW / DELEGATE — crew/team is the human-only `/team` toggle a
        // model cannot self-enable, and the targets may be unadvertised, so this
        // can only ever Correct (never silently Rewrite) and the message must NOT
        // imply the model can invoke crew itself.
        "delegate" | "spawn_agent" | "subagent" | "sub_agent" | "crew_dispatch" | "run_crew"
        | "dispatch_crew" | "fork_agent" | "assign" | "team" => {
            Some(AliasOutcome::Correct(format!(
                "'{name}' is not a newt tool. Crew/team delegation is only available once the \
                 human enables /team this session — you cannot turn it on yourself. When /team \
                 is on, compose_roster ({{\"mode\"}}) proposes a roster and crew ({{\"task\"}}) \
                 dispatches it."
            )))
        }
        // #725 TOOL-DISCOVERY — the instinctive "which tool does X?" reach. All
        // mean exactly tool_search (a `query` arg, or none — execute_tool_search
        // lists everything on an empty query), so silently Rewrite. `tool_search`
        // itself is the REAL tool and falls through to `None` below — it is never
        // an alias of itself.
        "find_tool" | "search_tools" | "list_tools" | "which_tool" | "available_tools"
        | "what_tools" | "tools" => Some(AliasOutcome::Rewrite("tool_search")),
        // #716 WORKFLOW — no workflow/pipeline primitive exists; redirect to the
        // plan tools (and crew/team, which needs /team).
        "workflow" | "run_workflow" | "start_workflow" | "pipeline" => Some(AliasOutcome::Correct(
            "newt has no workflow tool; sequence the work with update_plan (the full ordered \
                 plan with statuses), or delegate subtasks via crew/team (needs /team)."
                .to_string(),
        )),
        // #727 BUDGET — "how much context do I have left" reaches. get_context_remaining
        // takes no args (it is a self-read), so the foreign call's (empty) arg shape
        // matches: safe to silently Rewrite. The real name `get_context_remaining` is
        // NOT here — it falls through to `None` and dispatches unchanged (the loop
        // intercepts it), so it is never an alias of itself.
        "context_remaining" | "tokens_left" | "remaining_tokens" | "budget"
        | "how_much_context" | "context_budget" | "token_budget" => {
            Some(AliasOutcome::Rewrite("get_context_remaining"))
        }
        // #728 ASK-THE-HUMAN — the instinctive "ask the user a question" reach.
        // All mean exactly request_user_input (a `question` arg), so silently
        // Rewrite: the executor reads `question` and answers via the gate (or the
        // headless message). The real name `request_user_input` is NOT here — it
        // falls through to `None` and dispatches unchanged, so it is never an
        // alias of itself.
        "ask_user" | "ask_human" | "prompt_user" | "get_user_input" | "ask_question"
        | "clarify" | "ask" => Some(AliasOutcome::Rewrite("request_user_input")),
        _ => None,
    }
}

/// #727: true when `name` is `get_context_remaining` or one of its rewrite
/// aliases. The agentic loop computes the per-turn budget at the dispatch site
/// (where `num_ctx` and the conversation estimate are in scope) and renders it
/// there, bypassing [`execute_tool`] — so the loop must recognize both the
/// canonical name and the aliases that resolve to it.
pub(crate) fn is_context_remaining_call(name: &str) -> bool {
    name == "get_context_remaining"
        || matches!(
            resolve_tool_alias(name),
            Some(AliasOutcome::Rewrite("get_context_remaining"))
        )
}

/// #717: classify a single tool/capability reach for the alias-seam telemetry.
///
/// Pure: given the name the model called, its args, the tool result string, and
/// whether the result read as success, decide whether this reach is phantom and
/// how it resolved. Returns `None` for an ordinary real call (nothing to mine).
/// See [`crate::PhantomReach`] / [`crate::PhantomResolution`].
///
/// `ok` is part of the signature for symmetry with the recording site (it keys
/// the sibling `ToolEvent`); v1's miss-patterns classify on name + result alone.
pub(crate) fn classify_phantom_reach(
    name: &str,
    args: &serde_json::Value,
    result: &str,
    ok: bool,
) -> Option<crate::PhantomResolution> {
    let _ = ok;
    // 1. A recognized foreign/alias name: rewrite to a real tool, or a
    //    correction naming the right one — the canonical alias-seam signal.
    match resolve_tool_alias(name) {
        Some(AliasOutcome::Rewrite(canonical)) => {
            return Some(crate::PhantomResolution::Rewrite(canonical.to_string()))
        }
        Some(AliasOutcome::Correct(msg)) => return Some(crate::PhantomResolution::Correct(msg)),
        None => {}
    }
    // 2. An unknown name with no alias is a true phantom tool (hallucination).
    if is_hallucination(name, args) {
        return Some(crate::PhantomResolution::Unknown);
    }
    // 3. A real tool that returned empty-by-design — a high-signal "miss" that
    //    currently logs ok=true. These are the mineable real-tool reaches; the
    //    loop emits one per call, so a 3x identical-recall loop yields 3 records.
    let r = result.trim_start();
    if name == "state_get" && r.starts_with("no such key") {
        return Some(crate::PhantomResolution::RealToolMiss(
            "state_get on an unset key".into(),
        ));
    }
    if name == "recall" && r.starts_with("no matches in past conversations") {
        return Some(crate::PhantomResolution::RealToolMiss(
            "recall returned no matches".into(),
        ));
    }
    None
}

/// #479 (G4): classify a `crew`/`compose_roster` reach made while the crew/team
/// surface is gated OFF (`advertise_team == false`, the default — the runner is
/// only built when the operator sets `NEWT_TEAM`).
///
/// Pure: given the name the model called and whether the team surface is
/// advertised this session, decide whether this is a gated-off delegation reach
/// worth mining. Returns `None` for everything else (non-crew names, or crew
/// names when the surface is ON — those dispatch normally).
///
/// This is a SEPARATE seam from [`classify_phantom_reach`] on purpose: `crew`
/// and `compose_roster` stay real names in [`ALL_TOOL_NAMES`] (so the ON path is
/// a normal dispatch and [`is_hallucination`] is unchanged), which means
/// `classify_phantom_reach` never flags them. The gated-off detection needs the
/// one fact that function does not have — `advertise_team` — which is known in
/// the agent loop, so the loop composes the two seams there.
pub(crate) fn classify_gated_off_reach(
    name: &str,
    advertise_team: bool,
) -> Option<crate::PhantomResolution> {
    if !advertise_team && (name == "crew" || name == "compose_roster") {
        return Some(crate::PhantomResolution::GatedOff(
            "crew/team surface off (NEWT_TEAM)".into(),
        ));
    }
    None
}

/// Classic Levenshtein edit distance (pure two-row DP). Inputs are tool names
/// (short), so the simple version is plenty — for fuzzy suggestions only.
fn levenshtein(a: &str, b: &str) -> usize {
    let a: Vec<char> = a.chars().collect();
    let b: Vec<char> = b.chars().collect();
    let mut prev: Vec<usize> = (0..=b.len()).collect();
    let mut cur = vec![0usize; b.len() + 1];
    for (i, ca) in a.iter().enumerate() {
        cur[0] = i + 1;
        for (j, cb) in b.iter().enumerate() {
            let cost = usize::from(ca != cb);
            cur[j + 1] = (prev[j + 1] + 1).min(cur[j] + 1).min(prev[j] + cost);
        }
        std::mem::swap(&mut prev, &mut cur);
    }
    prev[b.len()]
}

/// The closest real tool name to `name`, if one is near enough to be a likely
/// typo/variant (distance ≤ ⌈len/3⌉, min 1). Returns `None` when nothing is
/// close, so we never suggest a wildly-unrelated tool.
fn nearest_tool_name(name: &str) -> Option<&'static str> {
    let threshold = (name.chars().count() / 3).max(1);
    ALL_TOOL_NAMES
        .iter()
        .map(|&t| (levenshtein(name, t), t))
        .filter(|(d, _)| *d <= threshold)
        .min_by_key(|(d, _)| *d)
        .map(|(_, t)| t)
}

/// Corrective message for a genuinely-unknown tool name: name the real base
/// tools and, when one is close, suggest it — so a weak model that missed the
/// catalog gets a path back instead of a dead end (Step 27.1). Kept
/// `unknown tool: {name}`-prefixed so existing `starts_with` checks hold.
fn unknown_tool_message(name: &str) -> String {
    const BASE: &str =
        "run_command, read_file, write_file, edit_file, list_dir, find, use_skill, web_fetch";
    match nearest_tool_name(name) {
        Some(sugg) => format!(
            "unknown tool: {name}. Did you mean '{sugg}'? Available tools include: \
             {BASE} (plus git and any memory/plan tools enabled this session)."
        ),
        None => format!(
            "unknown tool: {name}. Available tools include: {BASE} (plus git and \
             any memory/plan tools enabled this session)."
        ),
    }
}

/// Build a shell prefix that exports venv/exec-path vars into the agent-bridle
/// confined shell.
///
/// Agent-bridle's confined shell does not inherit the host environment
/// (`do_not_inherit_env(true)`), so we inject `VIRTUAL_ENV` and prepend
/// venv/extra `bin/` dirs to `PATH` by prefixing every `run_command` cmd.
/// `NEWT_VENV` (set from `--venv` or auto-detected from `$VIRTUAL_ENV` by the
/// CLI) takes precedence; falls back to `$VIRTUAL_ENV` if the TUI was invoked
/// directly without going through the CLI's `dispatch`.
pub fn venv_cmd_prefix() -> Option<String> {
    let venv = std::env::var("NEWT_VENV")
        .or_else(|_| std::env::var("VIRTUAL_ENV"))
        .ok();
    let exec_paths = std::env::var("NEWT_EXEC_PATHS").ok();

    if venv.is_none() && exec_paths.is_none() {
        return None;
    }

    // sh single-quoting: wrap in '', escape any ' as '\''
    let q = |s: &str| format!("'{}'", s.replace('\'', r"'\''"));

    // Build a list of dirs to prepend to PATH (venv/bin first, then exec-paths).
    let mut path_dirs: Vec<String> = Vec::new();
    let mut prefix = String::new();

    if let Some(ref venv) = venv {
        let venv_bin = format!("{venv}/bin");
        prefix.push_str(&format!("export VIRTUAL_ENV={}; ", q(venv)));
        path_dirs.push(venv_bin);
    }
    if let Some(ref paths) = exec_paths {
        for dir in paths.split(':') {
            if !dir.is_empty() {
                path_dirs.push(dir.to_string());
            }
        }
    }

    if !path_dirs.is_empty() {
        let quoted: Vec<String> = path_dirs.iter().map(|d| q(d)).collect();
        prefix.push_str(&format!("export PATH={}:\"$PATH\"; ", quoted.join(":")));
    }

    if prefix.is_empty() {
        None
    } else {
        Some(prefix)
    }
}

/// Build the venv/exec-path environment as a `{KEY:VALUE}` map for the confined
/// shell's structured `env` seam (agent-bridle, newt #783).
///
/// Same inputs as [`venv_cmd_prefix`] — `NEWT_VENV` (preferred) or
/// `$VIRTUAL_ENV`, plus `NEWT_EXEC_PATHS` — but delivered as host-supplied env
/// vars set directly on the spawned child instead of `export …;` text prepended
/// to the command. The `export` form is the #783 root cause: `export` is a
/// shell builtin, not a program, so the confined safe-subset engine refuses it
/// on a compound command (`a; b | c`). Passing the vars through the env seam
/// sidesteps that entirely and never touches the command text.
///
/// `PATH` is the venv `bin` (then any `NEWT_EXEC_PATHS` dirs) *prepended* to the
/// inherited host `PATH`: the env seam sets the value additively over the
/// child's ambient environment, so we read the host `PATH` here and build the
/// full string rather than relying on a `$PATH` expansion inside the value.
/// Returns an empty map when neither input is set (no env key is sent).
fn venv_env_map() -> std::collections::BTreeMap<String, String> {
    let mut map = std::collections::BTreeMap::new();
    let venv = std::env::var("NEWT_VENV")
        .or_else(|_| std::env::var("VIRTUAL_ENV"))
        .ok();
    let exec_paths = std::env::var("NEWT_EXEC_PATHS").ok();

    if venv.is_none() && exec_paths.is_none() {
        return map;
    }

    // Dirs to prepend to PATH (venv/bin first, then any exec-paths), mirroring
    // venv_cmd_prefix's ordering.
    let mut path_dirs: Vec<String> = Vec::new();
    if let Some(ref venv) = venv {
        map.insert("VIRTUAL_ENV".to_string(), venv.clone());
        path_dirs.push(format!("{venv}/bin"));
    }
    if let Some(ref paths) = exec_paths {
        for dir in paths.split(':') {
            if !dir.is_empty() {
                path_dirs.push(dir.to_string());
            }
        }
    }

    if !path_dirs.is_empty() {
        let prepend = path_dirs.join(":");
        let path = match std::env::var("PATH") {
            Ok(inherited) if !inherited.is_empty() => format!("{prepend}:{inherited}"),
            _ => prepend,
        };
        map.insert("PATH".to_string(), path);
    }

    map
}

/// Build the dispatch args for agent-bridle's confined `shell` tool (#783): the
/// RAW user command (free-form `cmd` mode) plus the venv carried through the
/// structured `env` seam ([`venv_env_map`]). Deliberately NO `export …;` prefix
/// on `cmd` — that is what the confined safe-subset engine refuses on a
/// compound command (the #783 root cause); the env seam sets `VIRTUAL_ENV` /
/// `PATH` on the spawned child instead. The host-bypass (`--yolo`) path keeps
/// the prefix form because it runs on a real `/bin/sh` where `export` works.
fn confined_dispatch_args(cmd: &str, workspace: &str) -> serde_json::Value {
    serde_json::json!({
        "cmd": cmd,
        "cwd": workspace,
        "env": venv_env_map(),
    })
}

/// The shell engine selected for this process (ADR 0005 D2 seam). Resolved once
/// at startup by the CLI ([`crate::resolve_shell_engine`] over the `[shell]
/// engine` config, the `--shell-engine` flag, and the `--full-access`
/// auto-upgrade) and published via `NEWT_SHELL_ENGINE`, so the deep
/// `run_command` dispatch reads it without threading it through every signature
/// — the same env-published pattern as [`ocap_disabled`] /
/// [`full_access_requested`]. Absent or unparseable falls back to the
/// `safe-subset` default.
fn shell_engine() -> crate::ShellEngine {
    if let Some(engine) = std::env::var("NEWT_SHELL_ENGINE")
        .ok()
        .and_then(|s| s.parse::<crate::ShellEngine>().ok())
    {
        return engine;
    }
    // No engine was published (e.g. a non-CLI entry point that set
    // NEWT_FULL_ACCESS directly). Honor the same auto-upgrade the CLI applies so
    // `NEWT_FULL_ACCESS=1` alone still gets the full-grammar engine (`host` on
    // unix, `brush` on Windows).
    if full_access_requested() {
        crate::full_access_default_engine()
    } else {
        crate::ShellEngine::default()
    }
}

/// agent-bridle's tool registry with the `"shell"` tool bound to the selected
/// engine (the ADR 0005 D2 seam: `safe-subset` / `host` / `brush` all honor the
/// same `Tool` contract under the `"shell"` name). `web_fetch` is added
/// unchanged. Mirrors `agent_bridle::registry()` but swaps the shell engine so
/// `[shell] engine = "host"` (or `--full-access`) routes `run_command` to the
/// full-grammar, kernel-jailed sandbox-host engine instead of the safe subset.
fn bridle_registry() -> agent_bridle::Registry {
    use std::sync::Arc;
    let shell: Arc<dyn agent_bridle::Tool> = match shell_engine() {
        crate::ShellEngine::SafeSubset => Arc::new(agent_bridle::ShellTool::new()),
        crate::ShellEngine::Host => Arc::new(agent_bridle::HostShellTool::new()),
        crate::ShellEngine::Brush => {
            // The carried brush engine (agent-bridle 0.7): in-process bash + the
            // L2 CommandInterceptor. The cross-platform engine — and on Windows
            // the ONLY full-grammar option, since `host` needs `/bin/sh`.
            #[cfg(windows)]
            {
                use std::sync::Once;
                static WARN: Once = Once::new();
                WARN.call_once(|| {
                    tracing::warn!(
                        "using the 'brush' shell engine on Windows: run_command runs a \
                         bash-in-Rust shell for internal-tooling compatibility. Native \
                         PowerShell/cmd code paths are a FUTURE release — not written yet \
                         (we are opinionated Linux developers who occasionally use a \
                         MacBook). Bash-isms work; Windows-native shell semantics do not."
                    );
                });
            }
            Arc::new(agent_bridle::BrushShellTool::new())
        }
    };
    agent_bridle::Registry::builder()
        .tool(shell)
        .tool(Arc::new(agent_bridle::WebFetchTool::new()))
        .build()
}

// ---------------------------------------------------------------------------
// INTERIM (#297): the --disable-ocap / --yolo exec escape hatch
// ---------------------------------------------------------------------------

/// INTERIM (#297): is the ocap exec bypass asserted for this invocation?
///
/// True only when `NEWT_DISABLE_OCAP=1` — set by the CLI's `--disable-ocap`
/// flag (alias `--yolo`) or exported directly for harness/pod use. The value
/// must be exactly `"1"`: a security bypass reads fail-closed, so anything
/// else (including `true`) leaves confinement on. Deliberately env-only —
/// there is NO config-file key, so the bypass can never silently persist; it
/// must be asserted per invocation.
///
/// Scope: `run_command` only. On stub-shell builds (the only crates.io-
/// publishable configuration) agent-bridle's `shell` tool fails closed on
/// every command, which makes agentic coding impossible without the brush
/// `CommandInterceptor` patch underneath. `web_fetch` is NOT bypassed: the
/// stub-shell branch stubs only the shell tool — `agent-bridle-tool-web`
/// ships the real leash-enforcing implementation (verified at agent-bridle
/// rev `2129c91`), so it does not fail closed. The fs tools keep the
/// newt-native workspace fence untouched: yolo is unconfined exec, fenced fs
/// — never a global authority-off switch.
///
/// Remove (or demote to a debug flag) when brush upstreams the
/// `CommandInterceptor` hook (reubeno/brush#1184) and agent-bridle's real
/// confined shell becomes the default everywhere — see agent-bridle#20 and
/// the `[patch.crates-io]` note in the workspace Cargo.toml.
pub fn ocap_disabled() -> bool {
    std::env::var("NEWT_DISABLE_OCAP").is_ok_and(|v| v == "1")
}

/// Is the per-invocation `full_access` preset override asserted?
///
/// True only when `NEWT_FULL_ACCESS=1` — set by the CLI's `--full-access`
/// flag. The session policy is then built from the `full_access` preset
/// (`Caveats::top()`) regardless of the configured `[tui.permissions]`
/// preset, exactly as if the config said `preset = "full_access"` for this
/// one run. Like [`ocap_disabled`], the value must be exactly `"1"` — a
/// widening switch reads fail-closed — and it is deliberately env-only, so
/// the override can never silently persist.
///
/// This is a DISTINCT switch from [`ocap_disabled`] (`--yolo`): the two
/// compose but never alias. `--full-access` widens the session *authority*
/// (fs fence, net leash, exec allowlist → unrestricted, which also empties
/// the #774 exec floor); `--yolo` changes the exec *mechanism* (host shell
/// instead of the confined shell) and still honors whatever floor is in
/// force. `--yolo --full-access` together yield an unrestricted host shell.
pub fn full_access_requested() -> bool {
    std::env::var("NEWT_FULL_ACCESS").is_ok_and(|v| v == "1")
}

/// facade P4 (#780): is the convenience **routing** turned OFF for this call?
///
/// True only when `NEWT_NO_ROUTE=1` — set by the CLI's `--no-route` flag. It
/// disables the L2 *convenience routing* ([`super::routing`]): a model's
/// `run_command("cat X")` runs the normal exec path as-is instead of being
/// rewritten to the governed `read_file` built-in.
///
/// **This is a DISTINCT switch from [`ocap_disabled`] (§7-F5).**
/// `--no-route` / `NEWT_NO_ROUTE` turns off L2 convenience only; it NEVER
/// disables the L3 boundary — the confined shell still gates exec and the fs
/// fence still governs reads. `--disable-ocap` / `--yolo` / `NEWT_DISABLE_OCAP`
/// (L3-OFF, a full host unconfine) is a completely separate mechanism: the two
/// names never alias, and turning routing off can never imply unconfined exec.
/// Reads fail-closed — only the exact value `1` turns routing off; deliberately
/// env-only, no config key, so it cannot silently persist.
pub fn routing_disabled() -> bool {
    std::env::var("NEWT_NO_ROUTE").is_ok_and(|v| v == "1")
}

/// #307: does the named-permission-preset exec FLOOR permit running `cmd` on
/// the UNCONFINED host shell?
///
/// `None` ⇒ no preset is active; the floor imposes nothing, so the answer is
/// `true` (the `--disable-ocap` bypass behaves exactly as it did pre-#307).
///
/// `Some(scope)` ⇒ the bypass may proceed ONLY for a single, simple command
/// whose program (leading token) the scope authorizes. This is deliberately
/// conservative on TWO counts, because the host shell runs `cmd` verbatim with
/// no per-spawn interceptor:
///
/// 1. A **compound** command (containing a shell metacharacter that could chain
///    another program — `&&`, `||`, `;`, `|`, `` ` ``, `$(`, newline, `&`, `>`,
///    `<`) is NOT allowed to bypass. `echo ok && rm -rf /` would otherwise
///    smuggle `rm` past an `echo` grant. It falls through to the confined
///    shell, which gates every spawn.
/// 2. Only the leading token is matched, so a bare allow-listed program runs;
///    anything else is denied.
///
/// The denied command isn't refused outright — it falls to the confined-shell
/// path, which enforces the (already preset-clamped) `caveats`. So a restricted
/// triage/on-call mode keeps its ceiling even under `--yolo`.
fn exec_floor_permits(floor: Option<&crate::caveats::Scope<String>>, cmd: &str) -> bool {
    use crate::caveats::ScopeExt as _;
    let Some(scope) = floor else {
        return true; // no preset ⇒ bypass unchanged (bit-for-bit)
    };
    // Conservative: any shell control/redirection metacharacter that could
    // introduce a second program defeats leading-token matching, so refuse the
    // bypass and let the confined shell gate each spawn.
    const SHELL_META: &[char] = &['&', '|', ';', '`', '$', '\n', '>', '<', '(', ')'];
    if cmd.contains(SHELL_META) {
        return false;
    }
    match cmd.split_ascii_whitespace().next() {
        // An empty command runs nothing; let it through to the normal path.
        None => true,
        Some(prog) => scope.permits(&prog.to_string()),
    }
}

/// INTERIM (#297): run `cmd` on the PLAIN host shell — no leash, no
/// interceptor, no sandbox — and wrap the outcome in an envelope structurally
/// identical to the confined shell's (`{ exit_code, stdout, stderr,
/// sandbox_kind }`, with `denied` / `denials` omitted exactly as the bridle
/// envelope omits them when nothing was denied). [`envelope_denied`] and
/// [`shell_envelope_output`] — and therefore the loop's truncation / denial /
/// exit-code handling — apply to it unchanged.
///
/// A spawn failure surfaces as `Err`, which the caller formats as the same
/// `error: …` string a bridle dispatch failure produces.
/// Run `cmd` through the SAME confined-shell path the `run_command` tool uses —
/// the venv env seam, the `--disable-ocap` host bypass under the #307 exec
/// floor, the agent-bridle confined shell, and the #263 permission-gate re-ask —
/// and render the envelope. Shared by the `run_command` and `lifecycle` (#891)
/// arms so both honor **identical** exec caveats; the caller prints the
/// tool-call line first.
#[allow(clippy::too_many_arguments)]
async fn exec_confined_command(
    cmd: &str,
    workspace: &str,
    color: bool,
    tool_output_lines: usize,
    caveats: &crate::caveats::Caveats,
    exec_floor: Option<&crate::caveats::Scope<String>>,
    permission_gate: Option<&mut dyn PermissionGate>,
    tool_offload: bool,
    spill_store: Option<&dyn SpillStore>,
) -> String {
    // Venv injection (#783): the confined shell carries the venv via
    // agent-bridle's structured `env` seam (see `confined_dispatch_args` /
    // `venv_env_map`), NOT by prepending `export …;` to the command — an
    // `export` builtin is not a program, so the safe-subset engine refuses it on
    // a compound command (the #783 root cause). `cmd_with_venv` (the
    // `export …;`-prefixed form) is built ONLY for the host-bypass path below:
    // that runs on a real `/bin/sh` where `export` is a genuine builtin.
    let cmd_with_venv = match venv_cmd_prefix() {
        Some(prefix) => format!("{prefix}{cmd}"),
        None => cmd.to_string(),
    };

    // INTERIM (#297): --disable-ocap / --yolo / NEWT_DISABLE_OCAP=1 — run the
    // command UNCONFINED on the host shell instead of the bridle's confined
    // shell. Nothing is denied here, so the #263 permission gate below is never
    // consulted. #307 FLOOR: a named-permission-preset clamp WINS over the
    // bypass — the unconfined host path is taken ONLY if the floor permits this
    // command's leading token; else it falls through to the confined shell,
    // which enforces the already-clamped `caveats`. `None` keeps the bypass
    // bit-for-bit.
    if ocap_disabled() && exec_floor_permits(exec_floor, cmd) {
        return match host_shell_dispatch(&cmd_with_venv, workspace).await {
            Ok(envelope) => shell_envelope_output(
                &envelope,
                tool_output_lines,
                color,
                tool_offload,
                spill_store,
            ),
            Err(e) => format!("error: {e}"),
        };
    }

    // #783: RAW cmd + venv via the env seam — never the `export …;` prefix,
    // which the confined safe-subset engine refuses.
    let dispatch_args = confined_dispatch_args(cmd, workspace);
    match bridle_registry()
        .dispatch("shell", dispatch_args.clone(), caveats)
        .await
    {
        // The confined shell ran. Its envelope carries
        // `{ exit_code, stdout, stderr, timed_out, ... }` plus — when the leash
        // refused a capability — the STRUCTURED denial fields
        // `{ denied: true, denials: [{ kind, target, reason }] }`. In free-form
        // mode an out-of-scope command is denied *inside* the shell by the brush
        // interceptor (the command genuinely does not run); we lift that to the
        // capability-denied UX by reading the structured `denied` field — NEVER
        // a stderr grep.
        Ok(envelope) if envelope_denied(&envelope) => {
            // #263: an interactive gate may turn this denial into a human grant.
            // ONE consult + ONE re-execution per call: a second denial (a
            // different target reached on the re-run) surfaces as the standard
            // envelope — the model can retry, which prompts afresh.
            if let Some(gate) = permission_gate {
                // #905: promptable exec denials OR net-host denials (agent-bridle
                // #196). On Allow, the re-mint widens the matching axis (net adds
                // the host to the allow-list), so the proxy admits it on re-run.
                if let Some(requests) =
                    exec_denial_requests(&envelope).or_else(|| net_denial_requests(&envelope))
                {
                    if let PermissionDecision::Allow(widened) = gate.ask(&requests) {
                        return match bridle_registry()
                            .dispatch("shell", dispatch_args, &widened)
                            .await
                        {
                            Ok(env2) if envelope_denied(&env2) => {
                                denied_run_command_result(&env2, color)
                            }
                            Ok(env2) => shell_envelope_output(
                                &env2,
                                tool_output_lines,
                                color,
                                tool_offload,
                                spill_store,
                            ),
                            Err(e) => format!("error: {e}"),
                        };
                    }
                }
            }
            denied_run_command_result(&envelope, color)
        }
        Ok(envelope) => shell_envelope_output(
            &envelope,
            tool_output_lines,
            color,
            tool_offload,
            spill_store,
        ),
        // An argv-mode leash denial, or an error from inside the tool — surface
        // the reason; the dispatch error Display is safe to show.
        Err(e) => format!("error: {e}"),
    }
}

async fn host_shell_dispatch(cmd: &str, cwd: &str) -> std::io::Result<serde_json::Value> {
    let output = host_shell_output(cmd, cwd).await?;
    Ok(serde_json::json!({
        "exit_code": output.status.code().unwrap_or(-1),
        "stdout": decode_shell_stream(&output.stdout),
        "stderr": decode_shell_stream(&output.stderr),
        // Honest provenance, same field the bridle envelope always carries:
        // nothing sandboxed this run.
        "sandbox_kind": "none",
    }))
}

fn decode_shell_stream(bytes: &[u8]) -> String {
    match std::str::from_utf8(bytes) {
        Ok(text) => text.to_string(),
        Err(_) => repair_bsd_cat_v_utf8(bytes)
            .unwrap_or_else(|| String::from_utf8_lossy(bytes).into_owned()),
    }
}

/// macOS/BSD `cat -v` is not Unicode-aware: for a UTF-8 glyph such as `─`
/// (`e2 94 80`) it emits the lead byte raw (`e2`) and renders only the
/// continuation bytes as ASCII meta-control notation (`M-^TM-^@`). That byte
/// stream is invalid UTF-8, so a plain lossy decode becomes `�M-^TM-^@`.
///
/// Repair only that precise shape. This keeps ordinary valid UTF-8 untouched and
/// leaves unrelated binary output on the existing lossy fallback.
fn repair_bsd_cat_v_utf8(bytes: &[u8]) -> Option<String> {
    let mut repaired = Vec::with_capacity(bytes.len());
    let mut changed = false;
    let mut i = 0;
    while i < bytes.len() {
        let lead = bytes[i];
        let Some(cont_count) = utf8_continuation_count(lead) else {
            repaired.push(lead);
            i += 1;
            continue;
        };

        let mut seq = Vec::with_capacity(cont_count + 1);
        seq.push(lead);
        let mut j = i + 1;
        let mut ok = true;
        for _ in 0..cont_count {
            match parse_cat_v_meta_byte(bytes, j) {
                Some((cont, next)) if (0x80..=0xbf).contains(&cont) => {
                    seq.push(cont);
                    j = next;
                }
                _ => {
                    ok = false;
                    break;
                }
            }
        }

        if ok && std::str::from_utf8(&seq).is_ok() {
            repaired.extend_from_slice(&seq);
            changed = true;
            i = j;
        } else {
            repaired.push(lead);
            i += 1;
        }
    }

    changed.then(|| String::from_utf8(repaired).ok()).flatten()
}

fn utf8_continuation_count(lead: u8) -> Option<usize> {
    match lead {
        0xc2..=0xdf => Some(1),
        0xe0..=0xef => Some(2),
        0xf0..=0xf4 => Some(3),
        _ => None,
    }
}

fn parse_cat_v_meta_byte(bytes: &[u8], start: usize) -> Option<(u8, usize)> {
    if start + 2 > bytes.len() || &bytes[start..start + 2] != b"M-" {
        return None;
    }
    let pos = start + 2;
    match bytes.get(pos).copied()? {
        b'^' => {
            let c = bytes.get(pos + 1).copied()?;
            let low = if c == b'?' {
                0x7f
            } else if (b'@'..=b'_').contains(&c) {
                c - b'@'
            } else {
                return None;
            };
            Some((low | 0x80, pos + 2))
        }
        c if (0x20..=0x7e).contains(&c) => Some((c | 0x80, pos + 1)),
        _ => None,
    }
}

/// INTERIM (#297) host shell selection: `bash -c` with an `sh -c` fallback
/// when bash is absent — the same sh-compatible free-form mode the confined
/// shell ran, so [`venv_cmd_prefix`]'s `export …;` prefix works unchanged.
#[cfg(not(windows))]
async fn host_shell_output(cmd: &str, cwd: &str) -> std::io::Result<std::process::Output> {
    fn shell(program: &str, cmd: &str, cwd: &str) -> tokio::process::Command {
        let mut c = tokio::process::Command::new(program);
        c.arg("-c").arg(cmd).current_dir(cwd);
        c
    }
    match shell("bash", cmd, cwd).output().await {
        Err(e) if e.kind() == std::io::ErrorKind::NotFound => shell("sh", cmd, cwd).output().await,
        other => other,
    }
}

/// INTERIM (#297) host shell selection on Windows: `cmd /C`, the same shape
/// as [`build_check_shell`].
#[cfg(windows)]
async fn host_shell_output(cmd: &str, cwd: &str) -> std::io::Result<std::process::Output> {
    tokio::process::Command::new("cmd")
        .args(["/C", cmd])
        .current_dir(cwd)
        .output()
        .await
}

/// Lexically normalise a path *string* — collapse `.` and `..` components
/// without touching the filesystem — so containment is decided on the location
/// the caller actually named, not on a raw byte prefix. Does NOT resolve
/// symlinks (that needs `canonicalize`, which requires the path to exist and is
/// the still-open `fs-canonical-containment` deviation): a symlink *inside* the
/// workspace can still point out. What this DOES close are the string-only
/// escapes — `..` traversal and sibling-prefix collisions.
fn lexically_normalize(path: &str) -> std::path::PathBuf {
    use std::path::{Component, PathBuf};
    let mut out = PathBuf::new();
    for comp in std::path::Path::new(path).components() {
        match comp {
            Component::CurDir => {}
            Component::ParentDir => {
                // Pop a real segment; never climb above a root/prefix.
                if !out.pop() {
                    out.push(comp.as_os_str());
                }
            }
            other => out.push(other.as_os_str()),
        }
    }
    out
}

/// Returns true if `full_path` is permitted by `scope`.
///
/// The `Caveats` lattice stores workspace-root strings (not individual file
/// paths) with exact-set semantics; this layer adds containment so that "the
/// workspace root is permitted" means "any path *under* it is permitted". Both
/// the candidate and each root are lexically normalised (collapsing `..`) and
/// then compared by whole path components via [`std::path::Path::starts_with`],
/// so `..` traversal (`/ws/../etc/passwd`) and sibling-prefix collisions
/// (`/ws-secret` vs root `/ws`) no longer escape the fence — unlike the raw
/// string prefix match this replaced. Symlink containment is still open
/// (`fs-canonical-containment`); creating one needs exec, which is gated separately.
pub(crate) fn tui_permits_path(scope: &crate::caveats::Scope<String>, full_path: &str) -> bool {
    match scope {
        crate::caveats::Scope::All => true,
        crate::caveats::Scope::Only(set) if set.is_empty() => false,
        crate::caveats::Scope::Only(set) => {
            let candidate = lexically_normalize(full_path);
            set.iter()
                .any(|root| candidate.starts_with(lexically_normalize(root)))
        }
    }
}

/// Run the configured build-check command in `workspace` and return a compact
/// result string appended to the tool output so the model sees it immediately.
pub(crate) fn run_build_check(cmd: &str, workspace: &str) -> String {
    let result = build_check_shell(cmd).current_dir(workspace).output();
    match result {
        Ok(out) if out.status.success() => "  ✓ build check passed".to_string(),
        Ok(out) => {
            let stderr = String::from_utf8_lossy(&out.stderr);
            let stdout = String::from_utf8_lossy(&out.stdout);
            let combined = format!("{stdout}{stderr}");
            let excerpt: String = combined.lines().take(8).collect::<Vec<_>>().join("\n");
            format!("  ✗ build check failed:\n{excerpt}")
        }
        Err(e) => format!("  ⚠ build check could not run: {e}"),
    }
}

#[cfg(windows)]
fn build_check_shell(cmd: &str) -> std::process::Command {
    let mut shell = std::process::Command::new("cmd");
    shell.args(["/C", cmd]);
    shell
}

#[cfg(not(windows))]
fn build_check_shell(cmd: &str) -> std::process::Command {
    let mut shell = std::process::Command::new("sh");
    shell.args(["-c", cmd]);
    shell
}

#[cfg(all(test, windows))]
fn passing_build_check_cmd() -> &'static str {
    "exit /B 0"
}

#[cfg(all(test, not(windows)))]
fn passing_build_check_cmd() -> &'static str {
    "true"
}

#[cfg(all(test, windows))]
fn failing_build_check_cmd(message: &str) -> String {
    format!("echo {message} 1>&2 & exit /B 1")
}

#[cfg(all(test, not(windows)))]
fn failing_build_check_cmd(message: &str) -> String {
    format!("echo {message} >&2; exit 1")
}

/// Whether a confined-shell envelope carries the STRUCTURED `denied: true`
/// flag — the leash's machine-readable signal that the brush interceptor
/// refused an exec / open inside the free-form command. Reads the structured
/// field agent-bridle emits; it does NOT parse stdout/stderr (the old stderr
/// string-match was fragile — a command that merely *printed* a denial-like
/// phrase could be misread, and any wording drift would silently break
/// detection).
fn envelope_denied(envelope: &serde_json::Value) -> bool {
    envelope
        .get("denied")
        .and_then(serde_json::Value::as_bool)
        .unwrap_or(false)
}

/// Build a human-readable denial message from the envelope's structured
/// `denials: [{ kind, target, reason }]` list, joining each entry's `reason`.
/// Falls back to a generic message when the list is missing or empty.
fn envelope_denial_reason(envelope: &serde_json::Value) -> String {
    let reasons: Vec<String> = envelope
        .get("denials")
        .and_then(serde_json::Value::as_array)
        .map(|arr| {
            arr.iter()
                .filter_map(|d| d.get("reason").and_then(serde_json::Value::as_str))
                .map(str::to_string)
                .collect()
        })
        .unwrap_or_default();
    if reasons.is_empty() {
        "denied: the capability leash refused an operation".to_string()
    } else {
        reasons.join("; ")
    }
}

/// The allowlist NAME for an exec target — the trailing path component (the
/// program's basename), so `/usr/bin/env` and `C:\tools\env.exe` resolve to the
/// command a grant would actually allowlist. Used when lifting a denied exec
/// into a #263 [`PermissionRequest`].
fn exec_allowlist_name(target: &str) -> &str {
    target
        .rsplit(['/', '\\'])
        .find(|part| !part.is_empty())
        .unwrap_or(target)
}

/// #721: the model-actionable recovery appended to every capability denial the
/// MODEL sees. A denial used to be a DEAD-END: it told the *human* to edit
/// `[tui.permissions]`, which the model cannot do mid-turn, so the loop stalled
/// (the issue's `mkdir` reproduction). This sentence tells the *model* what IT
/// can do — ask the operator to grant the capability via the
/// `request_permissions` tool, or change approach. The gate is unchanged: the
/// call is STILL denied; only the coaching is added. In a headless flow with no
/// operator, `request_permissions` answers "no operator available", which is
/// itself a recoverable signal (switch strategy) rather than a config edit.
const DENIAL_RECOVERY_HINT: &str =
    "This is outside your granted authority — call request_permissions with the \
     capability, a target, and a reason to ask the operator to grant it, or take \
     a different approach that stays within your current authority.";

/// #479 (G4): the model-facing recovery coach when `crew`/`compose_roster` is
/// reached while the crew/team surface is OFF — the DEFAULT, since the runner is
/// only built when the operator sets `NEWT_TEAM`. Replaces the flat
/// `unknown tool: … (no crew surface …)` dead-end (which left the model nowhere
/// to go) with a model-actionable message in the #721 [`DENIAL_RECOVERY_HINT`]
/// style: it names the operator gesture that enables the surface (`NEWT_TEAM`)
/// AND a real solo alternative (the always-available file/exec tools), so the
/// reach is recoverable instead of a wall. The OCAP presence-gate is unchanged —
/// crew stays `NEWT_TEAM`-gated; only the coaching is added.
const CREW_OFF_RECOVERY_HINT: &str =
    "the crew/team surface is not enabled this session (the operator launches it \
     with NEWT_TEAM). Accomplish this yourself with the available tools \
     (read_file/write_file/edit_file/run_command/...), or ask the operator to \
     enable a crew.";

/// The model-facing result for a `crew`/`compose_roster` dispatch when no
/// `CrewRunner` was injected. One factored message + regression point carrying
/// [`CREW_OFF_RECOVERY_HINT`], so the recoverable wording can never drift.
fn crew_off_recovery_result(name: &str) -> String {
    format!("'{name}' is unavailable: {CREW_OFF_RECOVERY_HINT}")
}

/// #721: the model-facing capability-denial message for an fs tool — the base
/// "{kind} does not permit '{path}'" line plus the recoverable, model-actionable
/// [`DENIAL_RECOVERY_HINT`]. One factored message + regression point shared by
/// every fs denial (read_file / write_file / edit_file / list_dir / find), so
/// the recoverable wording can never drift between them.
fn denied_fs_result(kind: &str, path: &str) -> String {
    format!("capability denied: {kind} does not permit '{path}'. {DENIAL_RECOVERY_HINT}")
}

/// The standard `run_command` capability-denial result, composed EXACTLY ONCE:
/// a single `capability denied: <bare reason>. <recovery hint>` for the model.
///
/// #775 (§2.5): the model-facing message is ONE clean level. Two earlier defects
/// are removed:
///
/// 1. The bare denial `reason` (a full sentence from the leash, e.g.
///    `exec of "export" is not within the granted authority`) is NO LONGER
///    stuffed into [`print_denied`]'s bare `'{target}'` slot. Doing so produced
///    the garbled `capability denied: exec does not permit '<whole reason
///    sentence> - add it via …>'` — a denial sentence nested inside another.
///    [`print_denied`] now receives only the BARE command target, matching its
///    `{axis} does not permit '{target}'` contract (the same shape the fs path
///    uses via [`denied_fs_result`]).
/// 2. The stale `extra_exec` config hint is gone from the model-facing message.
///    #721 superseded "edit your `[tui.permissions]` config" with the
///    model-actionable [`DENIAL_RECOVERY_HINT`] (`request_permissions`), so the
///    model now sees the bare reason once plus that hint — never a config edit
///    it cannot perform mid-turn.
///
/// The #263 prompt path still falls back here on deny (and on a second denial
/// after a re-execution).
fn denied_run_command_result(envelope: &serde_json::Value, color: bool) -> String {
    // Human transcript NOTICE: the bare denied command(s), never the reason
    // sentence (stuffing a sentence into print_denied's `'{target}'` slot was
    // the garble — see #775 above).
    print_denied(
        denial_axis_label(envelope),
        &exec_denial_target_label(envelope),
        color,
    );
    // Model-facing message: composed exactly once.
    format!(
        "capability denied: {}. {DENIAL_RECOVERY_HINT}",
        envelope_denial_reason(envelope)
    )
}

/// The bare target for the human exec-denial NOTICE: the denied command name(s)
/// the leash refused, NEVER the reason sentence. Joins multiple targets with
/// `, `; falls back to a generic label so the notice always prints one clean
/// `{axis} does not permit '{target}'` line. (#775 — restores
/// [`print_denied`]'s bare-`'{target}'` contract.)
fn exec_denial_target_label(envelope: &serde_json::Value) -> String {
    let targets: Vec<&str> = envelope
        .get("denials")
        .and_then(serde_json::Value::as_array)
        .map(|arr| {
            arr.iter()
                .filter_map(|d| d.get("target").and_then(serde_json::Value::as_str))
                .filter(|t| !t.is_empty())
                .collect()
        })
        .unwrap_or_default();
    if targets.is_empty() {
        "a command".to_string()
    } else {
        targets.join(", ")
    }
}

/// The standard `run_command` success path: print + return stdout/stderr,
/// or `(exit N)` when the command produced no output. Factored verbatim so
/// the #263 re-execution path shares one formatter with the first dispatch.
fn shell_envelope_output(
    envelope: &serde_json::Value,
    tool_output_lines: usize,
    color: bool,
    tool_offload: bool,
    spill_store: Option<&dyn SpillStore>,
) -> String {
    let stdout = envelope
        .get("stdout")
        .and_then(serde_json::Value::as_str)
        .unwrap_or("");
    let stderr = envelope
        .get("stderr")
        .and_then(serde_json::Value::as_str)
        .unwrap_or("");
    let out = format!("{stdout}{stderr}");
    // DISPLAY path: on-screen tool output is capped by LINES (unchanged).
    print_tool_output(&out, tool_output_lines, color);
    if out.trim().is_empty() {
        let code = envelope
            .get("exit_code")
            .and_then(serde_json::Value::as_i64)
            .unwrap_or(-1);
        format!("(exit {code})")
    } else {
        // #726/#945: the MODEL-facing payload is capped by the shared TOKEN
        // budget using head+tail. When tool_offload is on, spill the FULL
        // redacted output before capping so the true tail and elided middle stay
        // recoverable via memory_fetch("spill:<id>") and grep.
        let max_tokens = max_output_tokens();
        let est = crate::tokens::TokenEstimation::default();
        let over_model_budget = max_tokens != 0 && est.tokens_for_chars(out.len()) > max_tokens;
        let over_spill_budget = out.chars().count() > spill::TOOL_RESULT_SPILL_CAP;
        let should_spill =
            max_tokens != 0 && tool_offload && (over_model_budget || over_spill_budget);
        let capped = if should_spill {
            match spill_store {
                Some(store) => {
                    let (id, redacted) = spill::store_redacted_full(&out, store);
                    let teaser_tokens =
                        est.tokens_for_chars(spill::TOOL_RESULT_SPILL_CAP.saturating_sub(512));
                    cap_model_output_with_handle(
                        &redacted,
                        max_tokens.min(teaser_tokens),
                        output_head_tokens(),
                        Some(&id),
                    )
                }
                None => cap_model_output(&out, max_tokens),
            }
        } else {
            cap_model_output(&out, max_tokens)
        };
        // #898: if this command's output carries a forge "open a pull/merge
        // request" URL (git prints it on push of a new branch), append an
        // explicit next-step hint so the model opens the PR instead of stalling.
        // Detected from the UNcapped output so a long push log can't truncate the
        // URL away, and appended AFTER the cap so the hint always survives.
        match pr_creation_url(&out) {
            Some(url) => format!("{capped}{}", pr_next_step_hint(url)),
            None => capped,
        }
    }
}

/// #898: the forge "open a pull/merge request" URL that git prints on `push` of
/// a new branch — GitHub `…/pull/new/<branch>` (or a `…/compare/…` link) and
/// GitLab `…/merge_requests/new…`. Returned so [`shell_envelope_output`] can
/// append a next-step hint: models routinely push and then stall instead of
/// opening the PR (issue #898). Scans whitespace-split tokens because git emits
/// the URL on its own `remote:`-prefixed line.
fn pr_creation_url(output: &str) -> Option<&str> {
    output.split_whitespace().find(|tok| {
        tok.starts_with("https://")
            && (tok.contains("/pull/new/")
                || tok.contains("/merge_requests/new")
                || tok.contains("/compare/"))
    })
}

/// The next-step hint appended after a push whose output carries a PR-creation
/// URL (#898). Names the concrete `gh` command AND the tool boundary — the
/// embedded `git` tool cannot push or open PRs — so the model proceeds through
/// run_command + `gh` instead of looping back to the pushless git tool.
fn pr_next_step_hint(url: &str) -> String {
    format!(
        "\n\n[newt] A branch was pushed. To open a pull request now, call \
         run_command with `gh pr create --fill` (the `gh` CLI is available; the \
         `git` tool cannot push or open PRs). Or open this URL: {url}"
    )
}

/// Lift a confined-shell denial envelope into promptable #263 requests.
///
/// Returns `Some` only when EVERY structured denial entry is an `exec` kind
/// with a non-empty target — the case the human can meaningfully grant (the
/// allowlist name, same basename rule as the config hint). Any other kind
/// (e.g. an `open` refused inside the shell) keeps the standard denial:
/// guessing which fs axis an opaque `open` maps to would over-grant.
fn exec_denial_requests(envelope: &serde_json::Value) -> Option<Vec<PermissionRequest>> {
    let denials = envelope.get("denials")?.as_array()?;
    if denials.is_empty() {
        return None;
    }
    let mut requests = Vec::with_capacity(denials.len());
    for d in denials {
        if d.get("kind")?.as_str()? != "exec" {
            return None;
        }
        let target = d.get("target")?.as_str().filter(|t| !t.is_empty())?;
        requests.push(PermissionRequest {
            tool: "run_command".to_string(),
            kind: DenialKind::Exec,
            target: exec_allowlist_name(target).to_string(),
            reason: d
                .get("reason")
                .and_then(serde_json::Value::as_str)
                .unwrap_or_default()
                .to_string(),
        });
    }
    Some(requests)
}

/// #905: lift a confined-shell NET denial envelope into promptable #263 requests
/// — the `net`-axis sibling of [`exec_denial_requests`]. agent-bridle #196
/// surfaces a refused CONNECT host as `Denial { kind: "net", target: <host> }`
/// (with `denied: true`), so when the operator's `net` allow-list refuses a host
/// the shell reached (e.g. `git push` to `github.com`), this turns each into a
/// `PermissionRequest { kind: Net, target: host }` the gate can prompt per-host.
///
/// Returns `Some` only when EVERY denial is a `net` kind with a non-empty host
/// target — the case a grant is meaningful (add the host to the net allow-list).
/// A mixed or non-net batch returns `None` and keeps the standard denial.
fn net_denial_requests(envelope: &serde_json::Value) -> Option<Vec<PermissionRequest>> {
    let denials = envelope.get("denials")?.as_array()?;
    if denials.is_empty() {
        return None;
    }
    let mut requests = Vec::with_capacity(denials.len());
    for d in denials {
        if d.get("kind")?.as_str()? != "net" {
            return None;
        }
        let host = d.get("target")?.as_str().filter(|t| !t.is_empty())?;
        requests.push(PermissionRequest {
            tool: "run_command".to_string(),
            kind: DenialKind::Net,
            target: host.to_string(),
            reason: d
                .get("reason")
                .and_then(serde_json::Value::as_str)
                .unwrap_or_default()
                .to_string(),
        });
    }
    Some(requests)
}

/// #905: the axis label for the human denial NOTICE — `net` when EVERY denial is
/// a net (host) refusal, else `exec` (exec / mixed / empty default). Keeps a net
/// denial from being mislabeled `exec does not permit '<host>'`.
fn denial_axis_label(envelope: &serde_json::Value) -> &'static str {
    let all_net = envelope
        .get("denials")
        .and_then(serde_json::Value::as_array)
        .filter(|arr| !arr.is_empty())
        .is_some_and(|arr| {
            arr.iter()
                .all(|d| d.get("kind").and_then(serde_json::Value::as_str) == Some("net"))
        });
    if all_net {
        "net"
    } else {
        "exec"
    }
}

/// Consult the #263 gate for one denied fs path. Returns `true` only when
/// the human allowed it AND the re-minted caveats actually permit the path —
/// the widened authority is re-checked, never assumed.
fn fs_gate_allows(
    gate: &mut dyn PermissionGate,
    tool: &str,
    kind: DenialKind,
    full_path: &str,
    axis: impl Fn(&crate::caveats::Caveats) -> &crate::caveats::Scope<String>,
) -> bool {
    let request = PermissionRequest {
        tool: tool.to_string(),
        kind,
        target: full_path.to_string(),
        reason: format!("{} does not permit '{full_path}'", kind.as_str()),
    };
    match gate.ask(std::slice::from_ref(&request)) {
        PermissionDecision::Allow(widened) => tui_permits_path(axis(&widened), full_path),
        PermissionDecision::Deny => false,
    }
}

/// #721: map the model-supplied `capability` string for `request_permissions`
/// onto a [`DenialKind`] axis. A small synonym set absorbs the names a weak
/// local model tends to emit; an unrecognized value returns `None` so the tool
/// coaches instead of guessing an axis (guessing would request the WRONG
/// authority). Pure — unit-tested directly.
fn parse_capability(s: &str) -> Option<DenialKind> {
    match s.trim().to_ascii_lowercase().as_str() {
        "exec" | "run" | "run_command" | "command" | "shell" => Some(DenialKind::Exec),
        "fs_read" | "fs-read" | "read" | "read_file" => Some(DenialKind::FsRead),
        "fs_write" | "fs-write" | "write" | "write_file" => Some(DenialKind::FsWrite),
        "net" | "network" | "web" | "web_fetch" => Some(DenialKind::Net),
        _ => None,
    }
}

/// #721: the model-facing `request_permissions` tool — the capability-GRANT
/// path. It builds a [`PermissionRequest`] from `{capability, target, reason}`
/// and consults the SAME #263 [`PermissionGate`] a denial would: `Allow` reports
/// granted (and the gate has remembered any session grant, so the model's retry
/// of the original op rides the existing #263 re-exec machinery), `Deny` reports
/// declined, and **no gate** (headless / eval / ACP) reports that no operator is
/// available to grant — a recoverable signal (switch strategy), never a hang.
///
/// Reconciliation with #728: `request_permissions` (capability GRANT via
/// `gate.ask` / the #263 flow) and `request_user_input` (generic free-text Q&A
/// via `gate.ask_question`) are DISTINCT tools that share the ONE human-interface
/// gate ([`PermissionGate`]). They realize "both surface to the human" without
/// being merged: this one widens authority through the ocap gate, the other only
/// gathers text. `request_permissions` is deliberately NOT routed through
/// `request_user_input` — it mints caveats, which a free-text answer cannot.
fn execute_request_permissions(
    args: &serde_json::Value,
    gate: Option<&mut dyn PermissionGate>,
    color: bool,
    tool_output_lines: usize,
) -> String {
    let capability = args["capability"].as_str().unwrap_or("").trim();
    let target = args["target"].as_str().unwrap_or("").trim();
    let reason = args["reason"].as_str().unwrap_or("").trim();
    print_tool_call("request_permissions", capability, color);

    let Some(kind) = parse_capability(capability) else {
        let out = format!(
            "request_permissions: unknown capability '{capability}'. Use one of: \
             exec, fs_read, fs_write, net."
        );
        print_tool_output(&out, tool_output_lines, color);
        return out;
    };
    if target.is_empty() {
        let out = "request_permissions: 'target' is required — the command name (exec), \
                   the path (fs_read/fs_write), or the host (net)."
            .to_string();
        print_tool_output(&out, tool_output_lines, color);
        return out;
    }

    let request = PermissionRequest {
        tool: "request_permissions".to_string(),
        kind,
        target: target.to_string(),
        reason: if reason.is_empty() {
            format!("model requested {capability} for '{target}'")
        } else {
            reason.to_string()
        },
    };

    let out = match gate {
        // The gate consults the operator and (for a session grant) remembers it,
        // exactly as a denial-driven prompt does. We do not re-execute anything
        // here — the model retries its original tool call, which rides the #263
        // re-exec path under the now-granted caveats.
        Some(g) => match g.ask(std::slice::from_ref(&request)) {
            PermissionDecision::Allow(_widened) => format!(
                "granted: the operator allowed {capability} for '{target}'. \
                 Retry the original operation now."
            ),
            PermissionDecision::Deny => format!(
                "denied: the operator declined {capability} for '{target}'. \
                 Do not retry it — take a different approach."
            ),
        },
        // Headless / eval / ACP: no interactive gate exists to grant authority.
        // This is recoverable (change strategy), not a config edit the model
        // can perform mid-turn.
        None => format!(
            "no operator available to grant {capability} for '{target}' — this session \
             has no interactive permission gate (headless / eval / piped). The capability \
             must be configured by the owner (e.g. [tui.permissions] in newt config); \
             take a different approach for now."
        ),
    };
    print_tool_output(&out, tool_output_lines, color);
    out
}

/// #728: returned by `request_user_input` when there is no human to ask — either
/// no interactive gate this session (headless / eval / ACP / piped) or the gate
/// has no operator available (`ask_question` returned `None`). A recoverable
/// signal the model can act on, NEVER a hang.
const HEADLESS_NO_HUMAN: &str = "no human available this session (running headless) \
    — proceed with your best judgment or state your assumption explicitly.";

/// #728: the model-facing `request_user_input` tool — the GENERIC ask-the-human
/// path. It surfaces a free-text `question` to the operator through the SAME
/// human-interface gate a permission prompt uses ([`PermissionGate::ask_question`])
/// and returns the typed answer. With an operator present the answer is returned
/// verbatim; with NO gate (headless / eval / ACP / piped) — or when the gate has
/// no human to consult (`ask_question` returns `None`) — it returns the
/// [`HEADLESS_NO_HUMAN`] message and NEVER blocks.
///
/// Reconciliation with #721: this is the free-text Q&A path; `request_permissions`
/// is the capability-GRANT path (it mints caveats via the gate). Both surface to
/// the human through the one gate but are DISTINCT tools — one gathers text
/// (`ask_question`), the other widens authority (`ask`) — and are not merged.
fn execute_request_user_input(
    args: &serde_json::Value,
    gate: Option<&mut dyn PermissionGate>,
    color: bool,
    tool_output_lines: usize,
) -> String {
    let question = args["question"].as_str().unwrap_or("").trim();
    print_tool_call("request_user_input", question, color);

    if question.is_empty() {
        let out = "request_user_input: 'question' is required — the free-text \
                   question to ask the human."
            .to_string();
        print_tool_output(&out, tool_output_lines, color);
        return out;
    }

    // `Some(answer)` from the gate → return it verbatim; a `None` gate (headless)
    // OR an `ask_question` that returns `None` (no human to consult) → the
    // recoverable headless message. Either way we never block without an answer.
    let out = match gate.and_then(|g| g.ask_question(question)) {
        Some(answer) => answer,
        None => HEADLESS_NO_HUMAN.to_string(),
    };
    print_tool_output(&out, tool_output_lines, color);
    out
}

/// Best-effort host extraction for the #263 net pre-check. This only gates
/// whether to PROMPT — reachability enforcement stays with the bridle's
/// leash (host allowlist + SSRF screen). `None` (unparseable / non-http URL)
/// skips the pre-check entirely, leaving today's dispatch path untouched.
pub(crate) fn host_of_url(url: &str) -> Option<String> {
    let rest = url
        .strip_prefix("https://")
        .or_else(|| url.strip_prefix("http://"))?;
    let authority = rest.split(['/', '?', '#']).next()?;
    let host_port = authority.rsplit('@').next()?;
    // IPv6 literal `[::1]:8080` — the host is the bracketed part.
    let host = if let Some(stripped) = host_port.strip_prefix('[') {
        stripped.split(']').next()?
    } else {
        host_port.split(':').next()?
    };
    if host.is_empty() {
        None
    } else {
        Some(host.to_ascii_lowercase())
    }
}

/// File-type restriction for the embedded `find` tool (#496).
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
enum FindType {
    Files,
    Dirs,
    Any,
}

/// Parsed, validated options for one `find` invocation.
struct FindOpts<'a> {
    /// Glob matched against each basename; `None` matches everything.
    name: Option<&'a str>,
    type_filter: FindType,
    /// Max depth below the search root (1 = immediate children); `None` =
    /// unlimited.
    max_depth: Option<usize>,
    /// Hard cap on returned matches.
    max_results: usize,
    /// Honour .gitignore + skip .git/target/node_modules/hidden dirs.
    respect_gitignore: bool,
    case_sensitive: bool,
}

/// One-line summary of a `find` invocation for the tool trace (#529): the path
/// plus only the *non-default* filters, so two searches with different filters
/// don't both render as a bare `find: .`. Defaults (any type, unlimited depth,
/// the 1000-match cap, gitignore-respecting, case-sensitive) are omitted.
fn find_detail(path: &str, opts: &FindOpts) -> String {
    let mut parts: Vec<String> = Vec::new();
    if let Some(name) = opts.name {
        parts.push(format!("name={name}"));
    }
    match opts.type_filter {
        FindType::Files => parts.push("type=f".to_string()),
        FindType::Dirs => parts.push("type=d".to_string()),
        FindType::Any => {}
    }
    if let Some(d) = opts.max_depth {
        parts.push(format!("depth={d}"));
    }
    // Mirrors the parse default in the `find` arm.
    if opts.max_results != 1000 {
        parts.push(format!("max={}", opts.max_results));
    }
    if !opts.respect_gitignore {
        parts.push("no-gitignore".to_string());
    }
    if !opts.case_sensitive {
        parts.push("icase".to_string());
    }
    if parts.is_empty() {
        path.to_string()
    } else {
        format!("{path} ({})", parts.join(", "))
    }
}

/// Translate a shell-style basename glob (`*`, `?`) into an anchored regex.
/// Every other character is matched literally (regex metacharacters escaped),
/// so `pyo3_module.rs` matches only that exact basename, not `pyo3Xmodulexrs`.
fn glob_to_regex(glob: &str, case_sensitive: bool) -> Result<regex::Regex, String> {
    let mut re = String::with_capacity(glob.len() + 8);
    if !case_sensitive {
        re.push_str("(?i)");
    }
    re.push('^');
    for ch in glob.chars() {
        match ch {
            '*' => re.push_str(".*"),
            '?' => re.push('.'),
            // Escape every regex metacharacter so the rest is literal.
            '.' | '+' | '(' | ')' | '|' | '[' | ']' | '{' | '}' | '^' | '$' | '\\' => {
                re.push('\\');
                re.push(ch);
            }
            other => re.push(other),
        }
    }
    re.push('$');
    regex::Regex::new(&re).map_err(|e| format!("invalid name pattern: {e}"))
}

/// Recursively walk `root` and collect matches as workspace-relative,
/// `/`-normalised, sorted paths. Pure-`ignore`-crate traversal (no shell, no
/// subprocess) — the whole point of #496. Never follows symlinked directories
/// (avoids cycles and workspace escapes). Returns `(matches, truncated)` where
/// `truncated` is true if `max_results` was reached and more existed.
fn find_walk(
    root: &std::path::Path,
    workspace_root: &std::path::Path,
    opts: &FindOpts<'_>,
) -> Result<(Vec<String>, bool), String> {
    let pattern = match opts.name {
        Some(g) if !g.is_empty() => Some(glob_to_regex(g, opts.case_sensitive)?),
        _ => None,
    };

    let mut builder = ignore::WalkBuilder::new(root);
    builder
        .hidden(opts.respect_gitignore)
        .ignore(opts.respect_gitignore)
        .git_ignore(opts.respect_gitignore)
        .git_global(opts.respect_gitignore)
        .git_exclude(opts.respect_gitignore)
        .parents(opts.respect_gitignore)
        // Honour .gitignore even outside a git repo (the agent's cwd may not be
        // a checkout); without this `ignore` silently ignores gitignore files.
        .require_git(false)
        .follow_links(false);
    if let Some(d) = opts.max_depth {
        builder.max_depth(Some(d));
    }
    // The `ignore` walker prunes via .gitignore/hidden but has no built-in
    // skip for build/dep dirs. Prune them explicitly (and cheaply, before
    // descent) so a default `find` doesn't drown in target/ or node_modules/.
    // `.git` is already covered by `.hidden(true)`. Skipped only when respecting
    // ignores — `respect_gitignore=false` means "search everything".
    if opts.respect_gitignore {
        let mut ob = ignore::overrides::OverrideBuilder::new(root);
        // In override globs a leading `!` excludes; with no whitelist globs
        // present, everything else stays included.
        if ob.add("!target/").is_ok() && ob.add("!node_modules/").is_ok() {
            if let Ok(ov) = ob.build() {
                builder.overrides(ov);
            }
        }
    }

    let mut out: Vec<String> = Vec::new();
    let mut truncated = false;
    for result in builder.build() {
        let entry = match result {
            Ok(e) => e,
            // Skip individual unreadable entries rather than failing the walk.
            Err(_) => continue,
        };
        // depth 0 is the search root itself — never a match.
        if entry.depth() == 0 {
            continue;
        }
        let is_dir = entry.file_type().map(|t| t.is_dir()).unwrap_or(false);
        match opts.type_filter {
            FindType::Files if is_dir => continue,
            FindType::Dirs if !is_dir => continue,
            _ => {}
        }
        if let Some(re) = &pattern {
            let base = entry.file_name().to_string_lossy();
            if !re.is_match(&base) {
                continue;
            }
        }
        if out.len() >= opts.max_results {
            truncated = true;
            break;
        }
        let rel = entry
            .path()
            .strip_prefix(workspace_root)
            .unwrap_or_else(|_| entry.path());
        out.push(rel.to_string_lossy().replace('\\', "/"));
    }
    out.sort();
    out.dedup();
    Ok((out, truncated))
}

/// Execute a single tool call and return the result string sent back to the model.
///
/// `run_command` is routed through agent-bridle's Caveats-confined, brush-backed
/// `shell` tool: the WHOLE command runs inside the leash (`echo ok && rm -rf /`
/// no longer slips `rm` past an `echo` grant — every external spawn passes the
/// interceptor's `before_exec` / `before_open` gate). The fs tools
/// (`read_file` / `write_file` / `list_dir`) keep enforcing the same `caveats`
/// via `permits_*` — rerouting them is out of scope.
///
/// `note_sink` backs the `save_note` tool (Step 19.3), `recall_source` the
/// `recall` tool (Step 17.5), and `memory_source` the `memory_fetch` tool
/// (progressive-disclosure memory, #319). `None` ⇒ the tool was never
/// advertised, so a call here is treated like any unknown tool.
///
/// `permission_gate` is the #263 prompted-grant seam: when present, a
/// capability denial consults the human (allow once / session allow / deny)
/// before failing; an allow re-executes the denied call under the gate's
/// freshly minted caveats. `None` (the default, and every headless caller)
/// keeps every denial exactly as it was — bit-for-bit. #721's
/// `request_permissions` tool also rides this gate: it lets the MODEL proactively
/// request a grant (vs. only reacting to a denial), and reports "no operator
/// available" when the gate is `None`.
///
/// INTERIM (#297): when [`ocap_disabled`] is asserted (`--disable-ocap` /
/// `--yolo` / `NEWT_DISABLE_OCAP=1`), `run_command` skips the confined shell
/// and runs on the plain host shell with the same venv/PATH prefix and an
/// envelope of the same shape — nothing is denied, so the #263 gate is never
/// consulted for exec. Every other tool (fs fence, `web_fetch` leash) is
/// unaffected. Removed when brush upstreams `CommandInterceptor`
/// (agent-bridle#20).
///
/// `exec_floor` (issue #307) is the **named-permission-preset clamp** acting as
/// a hard authority FLOOR over exec. `None` (every existing caller, and the
/// no-preset case) leaves the `--disable-ocap` bypass exactly as it was —
/// bit-for-bit. `Some(scope)` makes the bypass conditional: an out-of-floor
/// command does NOT take the unconfined host path, it falls through to the
/// confined shell, which enforces the already-clamped `caveats` and denies it.
/// This is what makes a deliberately-restricted on-call/triage mode win over a
/// `--yolo` flag — the preset clamp is consulted as a ceiling the bypass
/// cannot cross.
#[allow(clippy::too_many_arguments)]
pub async fn execute_tool(
    name: &str,
    args: &serde_json::Value,
    workspace: &str,
    color: bool,
    tool_output_lines: usize,
    caveats: &crate::caveats::Caveats,
    mcp: &mut dyn McpTools,
    build_check_cmd: Option<&str>,
    note_sink: Option<&mut dyn NoteSink>,
    recall_source: Option<&dyn RecallSource>,
    memory_source: Option<&dyn MemorySource>,
    permission_gate: Option<&mut dyn PermissionGate>,
    exec_floor: Option<&crate::caveats::Scope<String>>,
    git_tool: Option<&dyn GitTool>,
    crew_runner: Option<&dyn CrewRunner>,
    scratchpad_store: Option<&dyn super::scratchpad::ScratchpadStore>,
    code_search: Option<super::semantic::CodeSearch<'_>>,
    experience_store: Option<&dyn super::experiential::ExperienceStore>,
    step_ledger: Option<&dyn super::scheduled::StepLedger>,
) -> String {
    execute_tool_with_offload(
        name,
        args,
        workspace,
        color,
        tool_output_lines,
        caveats,
        mcp,
        build_check_cmd,
        note_sink,
        recall_source,
        memory_source,
        permission_gate,
        exec_floor,
        git_tool,
        crew_runner,
        scratchpad_store,
        code_search,
        experience_store,
        step_ledger,
        false,
        None,
    )
    .await
}

#[allow(clippy::too_many_arguments)]
pub async fn execute_tool_with_offload(
    name: &str,
    args: &serde_json::Value,
    workspace: &str,
    color: bool,
    tool_output_lines: usize,
    caveats: &crate::caveats::Caveats,
    mcp: &mut dyn McpTools,
    build_check_cmd: Option<&str>,
    note_sink: Option<&mut dyn NoteSink>,
    recall_source: Option<&dyn RecallSource>,
    memory_source: Option<&dyn MemorySource>,
    permission_gate: Option<&mut dyn PermissionGate>,
    exec_floor: Option<&crate::caveats::Scope<String>>,
    git_tool: Option<&dyn GitTool>,
    crew_runner: Option<&dyn CrewRunner>,
    scratchpad_store: Option<&dyn super::scratchpad::ScratchpadStore>,
    code_search: Option<super::semantic::CodeSearch<'_>>,
    experience_store: Option<&dyn super::experiential::ExperienceStore>,
    step_ledger: Option<&dyn super::scheduled::StepLedger>,
    tool_offload: bool,
    spill_store: Option<&dyn SpillStore>,
) -> String {
    // Remote MCP tools (namespaced `server__tool`) route to their server before
    // the built-in match. They carry no Caveats leash in this build.
    if mcp.handles(name) {
        print_tool_call(name, &args.to_string(), color);
        let out = mcp.call(name, args).await;
        print_tool_output(&out, tool_output_lines, color);
        return out;
    }

    // Step 27.1: resolve foreign / hallucinated tool names (str_replace_editor,
    // execute, bash, …) BEFORE the dispatch match. Compatible-arg aliases
    // rewrite to the canonical name and dispatch transparently; the rest return
    // a correction that names the right tool. Real names (and MCP `server__tool`
    // names, handled above) fall through unchanged.
    let name = match resolve_tool_alias(name) {
        Some(AliasOutcome::Rewrite(canonical)) => canonical,
        Some(AliasOutcome::Correct(msg)) => return msg,
        None => name,
    };

    // facade P4 (#780): hidden tool-call routing. After alias normalization, a
    // `run_command` (or a shell alias rewritten to one) whose command is a
    // read-only reach (`cat`/`ls`/`find` + read-only `git`) is SILENTLY
    // rewritten to the governed built-in, so the model's instinctive shell
    // calls go through the SAME fs / git caveat checks they would by calling
    // the built-in directly — routing is NOT a bypass (§4.4). The route/gate
    // split is pure DATA ([`super::routing::RouteTable`]). State-modifying git
    // and everything else stay on the exec path (`RouteDecision::Exec`).
    //
    // `--no-route` / `NEWT_NO_ROUTE` ([`routing_disabled`]) turns this L2
    // convenience OFF — the command runs the normal exec path as-is — while the
    // L3 boundary (the confined shell below, the fs fence) STAYS. It is a switch
    // DISTINCT from `--disable-ocap` (§7-F5): the routing escape never disables
    // confinement.
    let routed: Option<(&'static str, serde_json::Value)> =
        if name == "run_command" && !routing_disabled() {
            let command = args.get("command").and_then(|v| v.as_str()).unwrap_or("");
            let decision = super::routing::RouteTable::builtin().classify(command);
            // §4.4: log every silent rewrite (the original command + the
            // governed built-in it routed to). `None` ⇒ nothing was rewritten.
            if let Some(line) = super::routing::audit_line(command, &decision) {
                tracing::debug!(target: "newt::routing", "{line}");
            }
            match decision {
                super::routing::RouteDecision::Route { tool, args } => Some((tool, args)),
                super::routing::RouteDecision::Exec => None,
            }
        } else {
            None
        };
    let (name, args): (&str, &serde_json::Value) = match &routed {
        Some((tool, routed_args)) => (*tool, routed_args),
        None => (name, args),
    };

    match name {
        // Model-curated memory (Step 19.3): routes add / replace / remove
        // through the caller's NoteSink — the same MemoryManager → NoteStore
        // path as `/remember`, so the 19.1 char-cap curator error and the
        // 19.2 write-time security scan apply identically.
        "save_note" => match note_sink {
            Some(sink) => execute_save_note(args, sink, color, tool_output_lines),
            // Without a sink the tool was never advertised — a call here is
            // a model hallucination; answer like any unknown tool.
            None => "unknown tool: save_note (no note store in this session)".to_string(),
        },

        // Cross-session recall (Step 17.5): searches PAST conversations via
        // the caller's RecallSource — workspace-fenced by the store, current
        // conversation excluded by the source implementation.
        "recall" => match recall_source {
            Some(source) => execute_recall(args, source, color, tool_output_lines),
            // Without a source the tool was never advertised — same
            // unknown-tool answer as the sink-less save_note path.
            None => "unknown tool: recall (no conversation store in this session)".to_string(),
        },

        // Progressive-disclosure memory (Workstream A MVP, #319): pulls the
        // verbatim body of one ADDRESSED item (`note:<id>` / `turn:<conv>#<seq>`)
        // via the caller's MemorySource — workspace-fenced by the underlying
        // NoteStore / ConversationStore. Same presence-gating as `recall`.
        "memory_fetch" => match memory_source {
            Some(source) => execute_memory_fetch(args, source, color, tool_output_lines),
            // Without a source the tool was never advertised — same
            // unknown-tool answer as the source-less recall path.
            None => "unknown tool: memory_fetch (no memory source in this session)".to_string(),
        },

        // Step 26.4 (#583): scratchpad state tools — presence-gated on the
        // injected store (advertised only when the `scratchpad` feature is on).
        "state_set" => match scratchpad_store {
            Some(s) => super::scratchpad::execute_state_set(args, s, color, tool_output_lines),
            None => "unknown tool: state_set (no scratchpad in this session)".to_string(),
        },
        "state_get" => match scratchpad_store {
            Some(s) => super::scratchpad::execute_state_get(args, s, color, tool_output_lines),
            None => "unknown tool: state_get (no scratchpad in this session)".to_string(),
        },
        "state_clear" => match scratchpad_store {
            Some(s) => super::scratchpad::execute_state_clear(s, color, tool_output_lines),
            None => "unknown tool: state_clear (no scratchpad in this session)".to_string(),
        },

        // Step 26.5.5 (#582): semantic code search — presence-gated on the
        // injected searcher (advertised only when the `semantic` feature is on).
        "code_search" => match code_search {
            Some(search) => {
                super::semantic::execute_code_search(args, search, color, tool_output_lines).await
            }
            None => {
                "unknown tool: code_search (semantic retrieval is off this session)".to_string()
            }
        },

        // Step 26.6a (#585): experiential record/recall — presence-gated on the
        // store (advertised only when the `experiential` feature is on).
        "experience_record" => match experience_store {
            Some(s) => {
                super::experiential::execute_experience_record(args, s, color, tool_output_lines)
            }
            None => "unknown tool: experience_record (experiential memory is off)".to_string(),
        },
        "experience_recall" => match experience_store {
            Some(s) => super::experiential::execute_experience_recall(
                args,
                s,
                super::experiential::EXPERIENCE_TOP_K,
                color,
                tool_output_lines,
            ),
            None => "unknown tool: experience_recall (experiential memory is off)".to_string(),
        },

        // Step 26.6b (#586) / #715 PR2: scheduled update_plan — the single plan
        // WRITE tool, presence-gated on the ledger (advertised only when the
        // `scheduled` feature is on). Replaces plan_set + plan_advance.
        "update_plan" => match step_ledger {
            Some(l) => super::scheduled::execute_update_plan(args, l, color, tool_output_lines),
            None => "unknown tool: update_plan (scheduled planning is off)".to_string(),
        },
        // #716: read-only plan view (the alias target for "what was I doing?"
        // probes) — same presence gate as update_plan.
        "plan_get" => match step_ledger {
            Some(l) => super::scheduled::execute_plan_get(l, color, tool_output_lines),
            None => "unknown tool: plan_get (scheduled planning is off)".to_string(),
        },

        // #714: self-scoped resume recovery — reads THIS conversation's recent
        // turns (via the RecallSource's this_conversation_recent, the opposite
        // of recall's filter), the <plan>, and the <state>. Advertised ALWAYS,
        // so it reuses the already-present recall_source / step_ledger /
        // scratchpad_store params and degrades gracefully when they are None.
        "resume_context" => super::resume::execute_resume_context(
            recall_source,
            step_ledger,
            scratchpad_store,
            color,
            tool_output_lines,
        ),

        // #721: the capability-GRANT request — rides the SAME #263 gate a denial
        // would consult. Advertised always; `permission_gate` is `None` for
        // headless / eval / ACP, where it answers "no operator available" rather
        // than blocking. Consumes the gate (mutually exclusive with the
        // run_command / fs arms that also use it — only one arm runs per call).
        "request_permissions" => {
            execute_request_permissions(args, permission_gate, color, tool_output_lines)
        }

        // #728: the GENERIC ask-the-human tool — surfaces a free-text question to
        // the operator via the SAME #263 human-interface gate (`ask_question`)
        // and returns the answer. Advertised always; `permission_gate` is `None`
        // for headless / eval / ACP, where it answers "no human available this
        // session" rather than blocking. Consumes the gate (mutually exclusive
        // with the run_command / fs / request_permissions arms that also use it —
        // only one arm runs per call).
        "request_user_input" => {
            execute_request_user_input(args, permission_gate, color, tool_output_lines)
        }

        // #725: tool discovery — search THIS session's advertised catalog by
        // intent so a model that half-remembers a capability finds the real tool
        // name instead of fabricating one. Advertised always; the catalog is
        // rebuilt here from the live presence sources (the `with_*` flags derive
        // from which optional capabilities this call was handed), so the search
        // reflects exactly what was advertised — built-ins, presence-gated tools,
        // AND the connected MCP `server__tool` entries. The matcher
        // ([`super::tool_search::execute_tool_search`]) is pure; the print + the
        // catalog build live here.
        "tool_search" => {
            let query = args.get("query").and_then(|v| v.as_str()).unwrap_or("");
            print_tool_call("tool_search", query, color);
            let catalog = merged_tool_definitions(
                &*mcp,
                note_sink.is_some(),
                recall_source.is_some(),
                memory_source.is_some(),
                git_tool.is_some(),
                crew_runner.is_some(),
                scratchpad_store.is_some(),
                code_search.is_some(),
                experience_store.is_some(),
                step_ledger.is_some(),
            );
            let out = super::tool_search::execute_tool_search(query, &catalog);
            print_tool_output(&out, tool_output_lines, color);
            out
        }

        // Embedded git (PR4, #461): dispatch through the injected GitTool
        // (newt-git's LocalGitTool). `GitCaveats::from_session` projects the
        // session's authority onto the git surface (fail-closed: a read-only
        // session can read but not commit). Same presence-gating as `recall` —
        // without an injected impl the tool was never advertised.
        "git" => match git_tool {
            Some(tool) => {
                let gc = crate::git_caveats::GitCaveats::from_session(caveats);
                let op = args.get("op").and_then(|v| v.as_str()).unwrap_or("");
                print_tool_call("git", op, color);
                let out = match tool.dispatch(op, args, &gc) {
                    Ok(rendered) => rendered,
                    // Denials + engine errors surface verbatim so the model
                    // sees WHY (e.g. "denied: commit" on a read-only session).
                    Err(e) => format!("error: {e}"),
                };
                print_tool_output(&out, tool_output_lines, color);
                out
            }
            None => "unknown tool: git (no git surface in this session)".to_string(),
        },

        // Agent-callable orchestration (#479): compose_roster proposes a crew
        // roster from the live environment; crew dispatches a crew/team on a task
        // and returns the diff + verify status for the overseer to review. Both
        // route through the injected CrewRunner, which runs spawned crews under
        // `meet`-attenuated caveats. Same presence-gating as `git` (the `/team`
        // toggle) — without an injected impl the tools were never advertised.
        "compose_roster" | "crew" => match crew_runner {
            Some(runner) => {
                print_tool_call(name, &args.to_string(), color);
                let out = match runner.dispatch(name, args, caveats).await {
                    Ok(rendered) => rendered,
                    Err(e) => format!("error: {e}"),
                };
                print_tool_output(&out, tool_output_lines, color);
                out
            }
            // #479 (G4): replace the flat dead-end with a recoverable coach —
            // name the operator gesture (NEWT_TEAM) + a real solo alternative.
            None => crew_off_recovery_result(name),
        },

        "run_command" => {
            let cmd = args["command"].as_str().unwrap_or("");

            // Corrective guard: the model tried to call a tool as a shell binary.
            // Return a correction so the model can retry with the right tool call.
            // #898: git NETWORK ops (push/fetch/pull/clone) are NOT bounced — the
            // embedded git tool can't do them, so they fall through to the shell
            // (net-gated), letting the model push a branch and open a PR.
            if let Some(tool) = run_command_redirect(cmd) {
                return format!(
                    "error: '{tool}' is a tool, not a shell command. \
                     Call it as a separate tool invocation — \
                     do not pass '{tool}' as a command argument to run_command."
                );
            }

            print_tool_call("run_command", cmd, color);

            // Route the WHOLE command through agent-bridle's confined shell
            // (free-form `cmd` mode) under the SAME Caveats the TUI resolved from
            // `[tui].permissions`. The confined-exec core is shared with the
            // `lifecycle` arm (#891) so both honor identical exec caveats.
            exec_confined_command(
                cmd,
                workspace,
                color,
                tool_output_lines,
                caveats,
                exec_floor,
                permission_gate,
                tool_offload,
                spill_store,
            )
            .await
        }

        // #891: the model-facing lifecycle surface over the #880 system. Resolve
        // THIS repo's command for the named phase (`.newt/config.toml
        // [lifecycle]` → matching tooling packs) and run it through the SAME
        // confined exec path as run_command. `action=list` returns the resolved
        // command WITHOUT running it — a pure discovery read.
        "lifecycle" => {
            let phase_key = args.get("phase").and_then(|v| v.as_str()).unwrap_or("");
            let Some(phase) = crate::tooling::Phase::from_key(phase_key) else {
                let valid = crate::tooling::Phase::ALL
                    .iter()
                    .map(|p| p.as_str())
                    .collect::<Vec<_>>()
                    .join(", ");
                return format!(
                    "error: unknown lifecycle phase '{phase_key}'. Valid phases: {valid}."
                );
            };
            let action = args.get("action").and_then(|v| v.as_str()).unwrap_or("run");
            // Resolve from the `[lifecycle]` override → matching tooling packs.
            // Several toolchains may each contribute a command; join with `&&`
            // so the confined shell runs them in sequence and short-circuits on
            // the first failure (the safe-subset engine supports `&&`).
            let cmds =
                crate::tooling::resolved_phase_commands(std::path::Path::new(workspace), phase);
            if cmds.is_empty() {
                return format!(
                    "no command configured for lifecycle phase '{}'. Set it in \
                     .newt/config.toml [lifecycle] or a tooling pack; `lifecycle` \
                     only runs commands the project declares.",
                    phase.as_str()
                );
            }
            let joined = cmds.join(" && ");
            match action {
                "list" => format!("lifecycle {} → {joined}", phase.as_str()),
                "run" => {
                    print_tool_call(
                        "lifecycle",
                        &format!("{} → {joined}", phase.as_str()),
                        color,
                    );
                    exec_confined_command(
                        &joined,
                        workspace,
                        color,
                        tool_output_lines,
                        caveats,
                        exec_floor,
                        permission_gate,
                        tool_offload,
                        spill_store,
                    )
                    .await
                }
                other => format!(
                    "error: unknown lifecycle action '{other}'. Use 'run' (default) or 'list'."
                ),
            }
        }

        "read_file" => {
            let path = args["path"].as_str().unwrap_or("");
            let full = std::path::Path::new(workspace).join(path);
            let full_str = full.to_string_lossy();
            if !tui_permits_path(&caveats.fs_read, &full_str) {
                // #263: the gate may grant the read; deny (or no gate) keeps
                // the standard denial text bit-for-bit.
                let allowed = permission_gate.is_some_and(|gate| {
                    fs_gate_allows(gate, "read_file", DenialKind::FsRead, &full_str, |c| {
                        &c.fs_read
                    })
                });
                if !allowed {
                    let msg = denied_fs_result("fs_read", path);
                    print_denied("fs_read", path, color);
                    return msg;
                }
            }
            print_tool_call("read_file", path, color);
            match std::fs::read_to_string(&full) {
                Ok(contents) => {
                    // #719: window + cap the MODEL-facing payload (the on-screen
                    // display is capped separately) so one read of a large file
                    // can't saturate the context window and abandon the task.
                    let offset = args["offset"].as_u64().map(|n| n as usize);
                    let limit = args["limit"].as_u64().map(|n| n as usize);
                    // #726: char backstop now derives from the shared token
                    // budget so read_file and run_command share one cap.
                    let out = paginate_read(&contents, offset, limit, max_output_tokens());
                    print_tool_output(&out, tool_output_lines, color);
                    out
                }
                Err(e) => format!("error reading {path}: {e}"),
            }
        }

        "write_file" => {
            let path = args["path"].as_str().unwrap_or("");
            let content = args["content"].as_str().unwrap_or("");
            let full = std::path::Path::new(workspace).join(path);
            let full_str = full.to_string_lossy();
            if !tui_permits_path(&caveats.fs_write, &full_str) {
                // #263: the gate may grant the write (the human's choice at
                // the prompt is the consent — the y/N confirm below stays
                // governed by the original scope shape, which a denial here
                // proves is `Only`, i.e. no second confirm).
                let allowed = permission_gate.is_some_and(|gate| {
                    fs_gate_allows(gate, "write_file", DenialKind::FsWrite, &full_str, |c| {
                        &c.fs_write
                    })
                });
                if !allowed {
                    let msg = denied_fs_result("fs_write", path);
                    print_denied("fs_write", path, color);
                    return msg;
                }
            }

            // Shrink guard: refuse if the proposed write removes > 30% of
            // lines AND > 30 lines absolute. This catches the failure mode
            // where a model replaces an entire large file with a small
            // fragment (observed in the wild: 4,247 → 107 lines).
            if let Ok(existing) = std::fs::read_to_string(&full) {
                let orig_lines = existing.lines().count();
                let new_lines = content.lines().count();
                let removed = orig_lines.saturating_sub(new_lines);
                if removed > 30 && new_lines < orig_lines * 7 / 10 {
                    let pct = removed * 100 / orig_lines.max(1);
                    let msg = format!(
                        "error: write_file would shrink {path} from {orig_lines} → {new_lines} lines \
                         (-{pct}%). This is likely unintentional. Use edit_file to make targeted \
                         changes, or ensure your content includes the full file."
                    );
                    print_denied("shrink-guard", path, color);
                    return msg;
                }
            }

            print_tool_call(
                "write_file",
                &format!("{path} ({} bytes)", content.len()),
                color,
            );

            // Show first 20 lines as preview.
            let preview: String = content.lines().take(20).collect::<Vec<_>>().join("\n");
            let has_more = content.lines().count() > 20;
            print_tool_output(
                &format!("{preview}{}", if has_more { "\n…" } else { "" }),
                tool_output_lines,
                color,
            );

            // Auto-write when the caveat explicitly scopes fs_write (the
            // preset itself is the user's consent).  Ask y/N only under
            // full_access / custom where fs_write == Scope::All.
            let needs_confirm = matches!(caveats.fs_write, crate::caveats::Scope::All);

            let confirmed = if needs_confirm {
                print!("Write this file? [y/N] ");
                use std::io::Write as _;
                std::io::stdout().flush().ok();
                let mut answer = String::new();
                std::io::stdin().read_line(&mut answer).is_ok()
                    && answer.trim().eq_ignore_ascii_case("y")
            } else {
                true
            };

            if confirmed {
                let full = std::path::Path::new(workspace).join(path);
                if let Some(parent) = full.parent() {
                    let _ = std::fs::create_dir_all(parent);
                }
                match std::fs::write(&full, content) {
                    Ok(_) => {
                        let line_count = content.lines().count();
                        println!("✓ wrote {path} ({line_count} lines)");
                        let check = build_check_cmd
                            .map(|cmd| run_build_check(cmd, workspace))
                            .unwrap_or_default();
                        format!("wrote {path} ({line_count} lines){check}")
                    }
                    Err(e) => format!("error writing {path}: {e}"),
                }
            } else {
                println!("skipped");
                format!("user declined to write {path}")
            }
        }

        "edit_file" => {
            let path = args["path"].as_str().unwrap_or("");
            let old_string = args["old_string"].as_str().unwrap_or("");
            let new_string = args["new_string"].as_str().unwrap_or("");
            let full = std::path::Path::new(workspace).join(path);
            let full_str = full.to_string_lossy();
            if !tui_permits_path(&caveats.fs_write, &full_str) {
                // #263: same prompted-grant path as write_file.
                let allowed = permission_gate.is_some_and(|gate| {
                    fs_gate_allows(gate, "edit_file", DenialKind::FsWrite, &full_str, |c| {
                        &c.fs_write
                    })
                });
                if !allowed {
                    let msg = denied_fs_result("fs_write", path);
                    print_denied("fs_write", path, color);
                    return msg;
                }
            }
            if old_string.is_empty() {
                return "error: old_string must not be empty — use write_file to create new files"
                    .to_string();
            }
            let existing = match std::fs::read_to_string(&full) {
                Ok(s) => s,
                Err(e) => return format!("error reading {path}: {e}"),
            };
            let count = existing.matches(old_string).count();
            if count == 0 {
                // Show the file's actual head so the model can copy the exact
                // text and self-correct on the next call — instead of guessing
                // old_string blind and looping (the failure mode that left a
                // model unable to add a header comment). The content is already
                // in hand from the read above; no extra round needed.
                const HEAD: usize = 40;
                let total = existing.lines().count();
                let head: String = existing
                    .lines()
                    .take(HEAD)
                    .map(|l| format!("  {l}"))
                    .collect::<Vec<_>>()
                    .join("\n");
                let more = if total > HEAD {
                    format!("\n  … ({} more line(s))", total - HEAD)
                } else {
                    String::new()
                };
                return format!(
                    "error: old_string not found in {path} — do not guess again. Copy the \
                     EXACT text (including leading whitespace) from the contents below, then \
                     retry. To add a header/first line, set old_string to the shown first \
                     line and put your header + that line in new_string; to create a new \
                     file use write_file.\n--- {path} (first {shown} of {total} line(s)) ---\n{head}{more}",
                    shown = total.min(HEAD),
                );
            }
            if count > 1 {
                return format!(
                    "error: old_string matches {count} locations in {path}. \
                     Add more surrounding context to make it unique."
                );
            }
            let updated = existing.replacen(old_string, new_string, 1);
            let old_lines = existing.lines().count();
            let new_lines = updated.lines().count();
            let delta = new_lines as i64 - old_lines as i64;
            let delta_str = if delta >= 0 {
                format!("+{delta}")
            } else {
                format!("{delta}")
            };
            print_tool_call("edit_file", &format!("{path} ({delta_str} lines)"), color);
            match std::fs::write(&full, &updated) {
                Ok(_) => {
                    println!("✓ edited {path} ({delta_str} lines, now {new_lines} total)");
                    let check = build_check_cmd
                        .map(|cmd| run_build_check(cmd, workspace))
                        .unwrap_or_default();
                    format!("edited {path} ({delta_str} lines, now {new_lines} total){check}")
                }
                Err(e) => format!("error writing {path}: {e}"),
            }
        }

        "list_dir" => {
            let path = args["path"].as_str().unwrap_or(".");
            let full = std::path::Path::new(workspace).join(path);
            let full_str = full.to_string_lossy();
            if !tui_permits_path(&caveats.fs_read, &full_str) {
                // #263: same prompted-grant path as read_file.
                let allowed = permission_gate.is_some_and(|gate| {
                    fs_gate_allows(gate, "list_dir", DenialKind::FsRead, &full_str, |c| {
                        &c.fs_read
                    })
                });
                if !allowed {
                    let msg = denied_fs_result("fs_read", path);
                    print_denied("fs_read", path, color);
                    return msg;
                }
            }
            print_tool_call("list_dir", path, color);
            match std::fs::read_dir(&full) {
                Ok(entries) => {
                    let mut names: Vec<String> = entries
                        .flatten()
                        .map(|e| e.file_name().to_string_lossy().into_owned())
                        .collect();
                    names.sort();
                    let listing = names.join("\n");
                    print_tool_output(&listing, tool_output_lines, color);
                    listing
                }
                Err(e) => format!("error: {e}"),
            }
        }

        // #496: embedded, shell-free file search. The reported breakage was an
        // agent that needed `find` but the build's shell tool was unavailable;
        // this arm walks the workspace with the `ignore` crate (no subprocess),
        // gated by the same fs_read caveat as list_dir/read_file.
        "find" => {
            let path = args["path"].as_str().unwrap_or(".");
            let full = std::path::Path::new(workspace).join(path);
            let full_str = full.to_string_lossy();
            if !tui_permits_path(&caveats.fs_read, &full_str) {
                let allowed = permission_gate.is_some_and(|gate| {
                    fs_gate_allows(gate, "find", DenialKind::FsRead, &full_str, |c| &c.fs_read)
                });
                if !allowed {
                    let msg = denied_fs_result("fs_read", path);
                    print_denied("fs_read", path, color);
                    return msg;
                }
            }
            let opts = FindOpts {
                name: args["name"].as_str(),
                type_filter: match args["type"].as_str() {
                    Some("f") => FindType::Files,
                    Some("d") => FindType::Dirs,
                    _ => FindType::Any,
                },
                max_depth: args["max_depth"].as_u64().map(|d| d as usize),
                max_results: args["max_results"]
                    .as_u64()
                    .map(|m| m as usize)
                    .unwrap_or(1000),
                respect_gitignore: args["respect_gitignore"].as_bool().unwrap_or(true),
                case_sensitive: args["case_sensitive"].as_bool().unwrap_or(true),
            };
            // #529: echo the active filters, not a bare `find: .` — two very
            // different searches must not render identically in the trace.
            print_tool_call("find", &find_detail(path, &opts), color);
            if !full.exists() {
                return format!("error: no such path '{path}'");
            }
            // Defence-in-depth for a *recursive* read: refuse a root that
            // canonicalises outside the workspace (e.g. via `..`). `find` never
            // follows symlinks, so descent can't escape either.
            if let (Ok(ws_canon), Ok(root_canon)) = (
                std::path::Path::new(workspace).canonicalize(),
                full.canonicalize(),
            ) {
                if !root_canon.starts_with(&ws_canon) {
                    let msg = denied_fs_result("fs_read", path);
                    print_denied("fs_read", path, color);
                    return msg;
                }
            }
            match find_walk(&full, std::path::Path::new(workspace), &opts) {
                Ok((hits, truncated)) => {
                    let mut listing = if hits.is_empty() {
                        "no matches".to_string()
                    } else {
                        hits.join("\n")
                    };
                    if truncated {
                        listing
                            .push_str(&format!("\n… (truncated at {} matches)", opts.max_results));
                    }
                    print_tool_output(&listing, tool_output_lines, color);
                    listing
                }
                Err(e) => format!("error: {e}"),
            }
        }

        "use_skill" => {
            let skill_name = args["name"].as_str().unwrap_or("");
            print_tool_call("use_skill", skill_name, color);
            // Reads from the configured skill search path. This is a read of
            // trusted operator config (procedural knowledge), not an exec of
            // arbitrary code, so it is NOT leash-gated — any SCRIPTS the skill
            // bundles still run through `run_command`'s confined shell and are
            // governed by the session caveats. The same first-directory-wins
            // precedence as the index means we load the copy the model was
            // actually shown.
            let dirs = crate::Config::resolve()
                .map(|c| c.skill_search_dirs())
                .unwrap_or_default();
            match newt_skills::load_body_from(&dirs, skill_name) {
                Ok(body) => {
                    print_tool_output(&body, tool_output_lines, color);
                    body
                }
                Err(e) => format!("error: {e}"),
            }
        }

        "web_fetch" => {
            let url = args["url"].as_str().unwrap_or("");
            print_tool_call("web_fetch", url, color);

            // Route through agent-bridle's `web_fetch` tool under the SAME
            // Caveats. The `net` axis gates which hosts are reachable (host
            // allowlist + SSRF screen); an out-of-scope host is denied by the
            // leash, surfaced via the dispatch error. The tool returns extracted
            // markdown (`{ url, final_url, status, title, markdown }`) — the body
            // is untrusted page content, not a command result.
            let mut fetch_args = serde_json::json!({ "url": url });
            if let Some(max_bytes) = args.get("max_bytes").and_then(serde_json::Value::as_u64) {
                fetch_args["max_bytes"] = serde_json::json!(max_bytes);
            }
            // #263: with a gate present, pre-check the host against the `net`
            // axis so an out-of-allowlist host becomes a prompt instead of a
            // leash error. Allow ⇒ dispatch under the gate's minted caveats;
            // deny (or no gate, or an unparseable URL) ⇒ dispatch under the
            // ORIGINAL caveats — the leash produces today's denial verbatim.
            let widened_for_net = match (permission_gate, host_of_url(url)) {
                (Some(gate), Some(host)) if !caveats.permits_net(&host) => {
                    let request = PermissionRequest {
                        tool: "web_fetch".to_string(),
                        kind: DenialKind::Net,
                        target: host.clone(),
                        reason: format!("net does not permit '{host}'"),
                    };
                    match gate.ask(std::slice::from_ref(&request)) {
                        PermissionDecision::Allow(widened) => Some(widened),
                        PermissionDecision::Deny => None,
                    }
                }
                _ => None,
            };
            let effective_caveats = widened_for_net.as_ref().unwrap_or(caveats);
            match agent_bridle::registry()
                .dispatch("web_fetch", fetch_args, effective_caveats)
                .await
            {
                Ok(result) => {
                    let markdown = result
                        .get("markdown")
                        .and_then(serde_json::Value::as_str)
                        .unwrap_or("");
                    let title = result
                        .get("title")
                        .and_then(serde_json::Value::as_str)
                        .unwrap_or("");
                    let final_url = result
                        .get("final_url")
                        .and_then(serde_json::Value::as_str)
                        .unwrap_or(url);
                    let out = if title.is_empty() {
                        format!("{final_url}\n\n{markdown}")
                    } else {
                        format!("# {title}\n{final_url}\n\n{markdown}")
                    };
                    print_tool_output(&out, tool_output_lines, color);
                    out
                }
                // A `net`-axis leash denial, or a fetch error (SSRF screen,
                // timeout, non-2xx) — surface the reason; Display is safe.
                Err(e) => format!("error: {e}"),
            }
        }

        other => unknown_tool_message(other),
    }
}

/// Classify an [`execute_tool`] result string as success or failure for the
/// turn's recorded tool events (Step 17.6, #246). Best-effort by necessity —
/// tool results are plain strings fed back to the model — so this mirrors
/// the failure prefixes this module (and `McpTools::call`) actually emit:
/// `error:`, `capability denied:`, and `unknown tool`. A successful
/// `run_command` whose *output* happens to start with one of these is
/// misclassified; the recorded event is an outcome claim, not a gate.
pub(crate) fn tool_result_ok(result: &str) -> bool {
    let r = result.trim_start();
    !(r.starts_with("error:")
        || r.starts_with("capability denied:")
        || r.starts_with("unknown tool"))
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::agentic::NoMcp;

    // ---- #717: classify_phantom_reach (pure, no fs) ----

    #[test]
    fn classify_phantom_rewrite_alias() {
        // A shell alias resolves to the canonical run_command rewrite.
        let got = classify_phantom_reach("bash", &serde_json::json!({"command": "ls"}), "ok", true);
        assert_eq!(
            got,
            Some(crate::PhantomResolution::Rewrite("run_command".into()))
        );
    }

    #[test]
    fn classify_phantom_correct_alias() {
        // An edit alias with the wrong arg shape returns Correct guidance.
        let got = classify_phantom_reach(
            "str_replace_editor",
            &serde_json::json!({}),
            "ignored",
            false,
        );
        match got {
            Some(crate::PhantomResolution::Correct(msg)) => {
                assert!(msg.contains("edit_file"), "guidance names the tool: {msg}");
            }
            other => panic!("expected Correct, got {other:?}"),
        }
    }

    #[test]
    fn classify_phantom_unknown_name() {
        // A foreign name with no alias is a true phantom tool. (Note: #716 turned
        // the plan/crew/workflow notions into recognized aliases, so this uses a
        // name no family claims.)
        let got = classify_phantom_reach(
            "summon_kraken",
            &serde_json::json!({}),
            "unknown tool: summon_kraken",
            false,
        );
        assert_eq!(got, Some(crate::PhantomResolution::Unknown));
    }

    #[test]
    fn classify_phantom_plan_alias_is_correct() {
        // #716 + #717: a foreign plan notion now resolves through the alias seam,
        // so the telemetry classifier records it as a Correct (coach) reach — the
        // new arms get phantom-reach telemetry for free.
        let got = classify_phantom_reach("make_plan", &serde_json::json!({}), "ignored", false);
        match got {
            Some(crate::PhantomResolution::Correct(msg)) => {
                assert!(
                    msg.contains("update_plan"),
                    "guidance names the tool: {msg}"
                );
            }
            other => panic!("expected Correct, got {other:?}"),
        }
    }

    #[test]
    fn classify_phantom_state_get_miss() {
        // state_get on an unset key is an empty-by-design real-tool miss.
        let got = classify_phantom_reach(
            "state_get",
            &serde_json::json!({"key": "nope"}),
            "no such key: nope",
            true,
        );
        assert_eq!(
            got,
            Some(crate::PhantomResolution::RealToolMiss(
                "state_get on an unset key".into()
            ))
        );
    }

    #[test]
    fn classify_phantom_recall_miss() {
        // recall with no hits is an empty-by-design real-tool miss.
        let got = classify_phantom_reach(
            "recall",
            &serde_json::json!({"query": "zzz"}),
            "no matches in past conversations for \"zzz\" — try different keywords",
            true,
        );
        assert_eq!(
            got,
            Some(crate::PhantomResolution::RealToolMiss(
                "recall returned no matches".into()
            ))
        );
    }

    #[test]
    fn classify_phantom_resume_reach_is_a_rewrite() {
        // #714 + #717: a "where were we" reach resolves through the alias seam to
        // a Rewrite, so the telemetry already captures it (no new wiring needed).
        let got = classify_phantom_reach("where_were_we", &serde_json::json!({}), "ignored", false);
        assert_eq!(
            got,
            Some(crate::PhantomResolution::Rewrite("resume_context".into()))
        );
    }

    #[test]
    fn classify_phantom_real_success_is_none() {
        // An ordinary successful real tool call is not phantom telemetry.
        let got = classify_phantom_reach(
            "read_file",
            &serde_json::json!({"path": "src/lib.rs"}),
            "line 1\nline 2\n",
            true,
        );
        assert_eq!(got, None);
    }

    // ---- #725: tool_search discovery (alias + name registry) ----

    #[test]
    fn tool_search_is_a_real_tool_name() {
        // It must be in the canonical registry so a model calling it is never
        // treated as a hallucination.
        assert!(ALL_TOOL_NAMES.contains(&"tool_search"));
    }

    #[test]
    fn discovery_verbs_alias_to_tool_search() {
        // The instinctive "which tool does X?" reaches silently Rewrite to the
        // real tool_search.
        for verb in [
            "find_tool",
            "search_tools",
            "list_tools",
            "which_tool",
            "available_tools",
            "what_tools",
            "tools",
        ] {
            match resolve_tool_alias(verb) {
                Some(AliasOutcome::Rewrite(c)) => assert_eq!(c, "tool_search", "verb: {verb}"),
                other => panic!(
                    "expected Rewrite(tool_search) for {verb}, got something else: {}",
                    other.is_some()
                ),
            }
        }
    }

    #[test]
    fn tool_search_is_not_an_alias_of_itself() {
        // The real name must fall through unchanged (no recursive rewrite).
        assert!(resolve_tool_alias("tool_search").is_none());
    }

    #[test]
    fn classify_phantom_discovery_reach_is_a_rewrite() {
        // #725 + #717: a discovery reach resolves through the alias seam to a
        // Rewrite, so the phantom telemetry captures it for free.
        let got = classify_phantom_reach("find_tool", &serde_json::json!({}), "ignored", false);
        assert_eq!(
            got,
            Some(crate::PhantomResolution::Rewrite("tool_search".into()))
        );
    }

    #[test]
    fn classify_phantom_tool_search_real_call_is_none() {
        // A real tool_search call is not phantom telemetry.
        let got = classify_phantom_reach(
            "tool_search",
            &serde_json::json!({"query": "read"}),
            "Tools matching \"read\":\n- read_file — Read a file",
            true,
        );
        assert_eq!(got, None);
    }

    #[test]
    fn tool_search_is_not_a_hallucination() {
        assert!(!is_hallucination(
            "tool_search",
            &serde_json::json!({"query": "x"})
        ));
    }

    // ---- #719: read_file payload window/cap/pagination (pure, no fs) ----

    #[test]
    fn paginate_read_caps_a_large_file_to_the_default_window() {
        // A 15k-line file must NOT flood the model: default window is 2000 lines
        // with a footer to continue (regression for the 12.5k→168k saturation).
        let body: String = (1..=15_057)
            .map(|n| format!("line {n}"))
            .collect::<Vec<_>>()
            .join("\n");
        let out = paginate_read(&body, None, None, DEFAULT_MAX_OUTPUT_TOKENS);
        let lines: Vec<&str> = out.lines().collect();
        assert_eq!(lines[0], "line 1");
        assert_eq!(lines[1999], "line 2000");
        assert!(
            !out.contains("line 2001"),
            "window stops at 2000: {:?}",
            &out[..40]
        );
        assert!(out.contains("of 15057"), "footer names the total");
        assert!(
            out.contains("offset=2001"),
            "footer points at the next window"
        );
    }

    #[test]
    fn paginate_read_offset_and_limit_return_just_that_window() {
        let body: String = (1..=100)
            .map(|n| format!("L{n}"))
            .collect::<Vec<_>>()
            .join("\n");
        let out = paginate_read(&body, Some(10), Some(5), DEFAULT_MAX_OUTPUT_TOKENS);
        assert!(out.starts_with("L10\nL11\nL12\nL13\nL14"), "{out:?}");
        assert!(out.contains("offset=15"), "continues at line 15: {out:?}");
    }

    #[test]
    fn paginate_read_small_file_is_returned_verbatim_without_a_footer() {
        // Whole-file read that fits both caps → exact bytes, no footer.
        assert_eq!(
            paginate_read("a\nb\nc\n", None, None, DEFAULT_MAX_OUTPUT_TOKENS),
            "a\nb\nc\n"
        );
    }

    #[test]
    fn paginate_read_char_backstop_tracks_the_token_budget() {
        // #726: the char backstop is now token-derived (budget × chars/token),
        // NOT a hardcoded 100k. One enormous line: the line window can't help;
        // the token-derived char backstop must. With a 1000-token budget the
        // backstop is ~4000 chars, so a 50k-char line is truncated near there.
        let budget = 1_000;
        let max_chars = crate::tokens::TokenEstimation::default().chars_for_tokens(budget);
        let body = "x".repeat(50_000);
        let out = paginate_read(&body, None, None, budget);
        assert!(
            out.len() < max_chars + 300,
            "char-capped to the token budget (~{max_chars} chars): {} bytes",
            out.len()
        );
        assert!(out.contains("truncated"), "marks the truncation");
        assert!(
            out.contains("~1000 tokens"),
            "footer names the token budget: {out:?}"
        );

        // A LARGER budget keeps more of the same line — the backstop tracks the
        // budget rather than a fixed constant.
        let wide = paginate_read(&body, None, None, 4_000);
        assert!(
            wide.len() > out.len(),
            "a wider token budget keeps more chars: {} vs {}",
            wide.len(),
            out.len()
        );
    }

    #[test]
    fn paginate_read_zero_budget_disables_the_char_backstop() {
        // #726: max_output_tokens == 0 means "no cap" — only the line window
        // applies, so a single huge line comes back verbatim.
        let body = "y".repeat(500_000);
        let out = paginate_read(&body, None, None, 0);
        assert_eq!(out, body, "zero budget = no char backstop");
    }

    #[test]
    fn paginate_read_offset_past_end_is_a_clear_message() {
        let out = paginate_read("a\nb", Some(99), None, DEFAULT_MAX_OUTPUT_TOKENS);
        assert!(out.contains("past end"), "{out:?}");
    }

    // ---- #726: shared token-based model-facing output cap ----

    #[test]
    fn cap_model_output_passes_small_output_through_unchanged() {
        // Well under budget → exact bytes, no marker.
        let small = "hello\nworld\n";
        assert_eq!(cap_model_output(small, DEFAULT_MAX_OUTPUT_TOKENS), small);
    }

    #[test]
    fn cap_model_output_truncates_over_budget_as_head_tail() {
        let big = format!("HEAD_MARKER\n{}\nTAIL_MARKER", "middle\n".repeat(20_000));
        let out = cap_model_output_with_handle(&big, 1_000, 100, None);
        assert!(out.len() < big.len(), "must shrink: {} bytes", out.len());
        assert!(out.contains("HEAD_MARKER"), "head dropped: {out:?}");
        assert!(out.contains("TAIL_MARKER"), "tail dropped: {out:?}");
        assert!(out.contains("head+tail shown"), "marker present: {out:?}");
        assert!(
            !out.contains(&"middle\n".repeat(1_000)),
            "middle should be elided"
        );
    }

    #[test]
    fn cap_model_output_truncates_at_a_char_boundary() {
        // A multi-byte char straddling the cut must not be split — the body must
        // stay valid UTF-8 (no panic, no replacement char).
        let budget = 10; // ~40 chars
        let body = "é".repeat(1_000); // 2 bytes each
        let out = cap_model_output(&body, budget);
        assert!(out.is_char_boundary(out.len()), "valid boundary");
        assert!(
            out.chars()
                .all(|c| c == 'é' || !c.is_control() || c == '\n'),
            "no split char: {out:?}"
        );
    }

    #[test]
    fn cap_model_output_zero_budget_is_no_cap() {
        let body = "z".repeat(500_000);
        assert_eq!(cap_model_output(&body, 0), body);
    }

    #[test]
    fn token_to_char_math_uses_the_default_four_chars_per_token() {
        // The budget→char conversion is the default 4 chars/token (the shared
        // estimator constant), so a 10k-token budget is a 40k-char backstop.
        let est = crate::tokens::TokenEstimation::default();
        assert_eq!(est.chars_for_tokens(DEFAULT_MAX_OUTPUT_TOKENS), 40_000);
    }

    #[test]
    fn find_detail_bare_path_has_no_filters() {
        let opts = FindOpts {
            name: None,
            type_filter: FindType::Any,
            max_depth: None,
            max_results: 1000,
            respect_gitignore: true,
            case_sensitive: true,
        };
        assert_eq!(find_detail(".", &opts), ".");
    }

    #[test]
    fn find_detail_shows_only_non_default_filters() {
        let opts = FindOpts {
            name: Some("*.rs"),
            type_filter: FindType::Files,
            max_depth: Some(2),
            max_results: 50,
            respect_gitignore: false,
            case_sensitive: false,
        };
        assert_eq!(
            find_detail("src", &opts),
            "src (name=*.rs, type=f, depth=2, max=50, no-gitignore, icase)"
        );
    }

    #[test]
    fn find_detail_omits_each_default_independently() {
        let opts = FindOpts {
            name: None,
            type_filter: FindType::Dirs,
            max_depth: None,
            max_results: 1000,
            respect_gitignore: true,
            case_sensitive: true,
        };
        assert_eq!(find_detail(".", &opts), ". (type=d)");
    }

    #[test]
    fn use_skill_tool_is_advertised_in_definitions() {
        let defs = tool_definitions();
        let names: Vec<&str> = defs
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|d| d["function"]["name"].as_str())
            .collect();
        assert!(names.contains(&"use_skill"), "got: {names:?}");
    }

    #[test]
    fn merged_tool_definitions_with_empty_mcp_is_builtin_set() {
        let merged = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        let names: Vec<&str> = merged
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|d| d["function"]["name"].as_str())
            .collect();
        assert_eq!(
            names,
            vec![
                "run_command",
                "read_file",
                "write_file",
                "edit_file",
                "list_dir",
                "find",
                "use_skill",
                "web_fetch",
                // #721: advertised ALWAYS (core capability-grant request, no
                // presence gate) — part of the base tool_definitions() set.
                "request_permissions",
                // #714: advertised ALWAYS (no presence gate), so it joins the
                // base set even with every `with_*` flag off.
                "resume_context",
                // #725: advertised ALWAYS (a discovery tool must always be
                // present), so it too joins the base set with every flag off.
                "tool_search",
                // #727: advertised ALWAYS (read-only budget self-read, no
                // presence gate), pushed right after resume_context.
                "get_context_remaining",
                // #728: advertised ALWAYS (a model must always be able to ask the
                // human; degrades honestly headless), pushed last.
                "request_user_input",
                // #891: advertised ALWAYS (the model-facing lifecycle surface;
                // degrades honestly with "no command configured"), pushed after
                // request_user_input.
                "lifecycle",
            ]
        );
    }

    /// #894: each registry entry's schema-builder produces the SAME name the
    /// entry declares — catches a copy-paste where the `ToolSpec.name` and the
    /// `*_tool_definition()` disagree.
    #[test]
    fn registry_specs_match_their_definition_names() {
        for spec in EXTENDED_TOOL_REGISTRY {
            let def = (spec.definition)();
            assert_eq!(
                def["function"]["name"].as_str(),
                Some(spec.name),
                "ToolSpec name {:?} != definition name",
                spec.name
            );
        }
    }

    /// #894: no built-in tool name is declared twice across the base array and
    /// the registry (a dup would double-advertise and confuse dispatch).
    #[test]
    fn builtin_tool_names_are_unique() {
        let mut seen = std::collections::HashSet::new();
        for name in ALL_TOOL_NAMES.iter() {
            assert!(seen.insert(*name), "duplicate built-in tool name: {name}");
        }
    }

    /// #894 anti-drift (the payoff): with EVERY gate on, the advertised set from
    /// `merged_tool_definitions` equals `ALL_TOOL_NAMES` in BOTH directions. This
    /// is the test that would have caught the `lifecycle` drift — a tool
    /// advertised/dispatched but missing from the real-name set (or vice versa)
    /// fails here.
    #[test]
    fn advertised_set_matches_all_tool_names_both_directions() {
        let all =
            merged_tool_definitions(&NoMcp, true, true, true, true, true, true, true, true, true);
        let advertised: std::collections::HashSet<&str> = all
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|d| d["function"]["name"].as_str())
            .collect();
        let names: std::collections::HashSet<&str> = ALL_TOOL_NAMES.iter().copied().collect();
        // Every advertised tool is a real (non-hallucinated) name...
        for a in &advertised {
            assert!(
                names.contains(a),
                "advertised but not in ALL_TOOL_NAMES: {a}"
            );
        }
        // ...and every real name is actually advertised when its gate is on.
        for n in &names {
            assert!(
                advertised.contains(n),
                "in ALL_TOOL_NAMES but never advertised: {n}"
            );
        }
    }

    /// #894: `BASE_TOOL_NAMES` mirrors the names inlined in `tool_definitions()`
    /// exactly and in order — the one hand-kept mirror, guarded here.
    #[test]
    fn base_tool_names_match_tool_definitions() {
        let defs = tool_definitions();
        let base: Vec<&str> = defs
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|d| d["function"]["name"].as_str())
            .collect();
        assert_eq!(base, BASE_TOOL_NAMES);
    }

    /// #894 regression for the concrete drift that motivated the registry: the
    /// `lifecycle` tool (#891) is advertised + dispatched, so it MUST be a real
    /// name — otherwise every legitimate `lifecycle` call is miscounted as a
    /// hallucination (inflating the anti-loop counter). Before the registry it
    /// was missing from `ALL_TOOL_NAMES`; the derivation makes that impossible.
    #[test]
    fn lifecycle_is_a_real_tool_name_not_a_hallucination() {
        assert!(
            ALL_TOOL_NAMES.contains(&"lifecycle"),
            "lifecycle must be a real tool name"
        );
        assert!(
            !is_hallucination("lifecycle", &serde_json::json!({"phase": "test"})),
            "a real lifecycle call must not be flagged as a hallucination"
        );
    }

    #[test]
    fn lifecycle_definition_enum_matches_phase_vocabulary() {
        // The schema's phase enum is built from `Phase::ALL`, so it can never
        // drift from the vocabulary the executor parses with `Phase::from_key`.
        let def = lifecycle_tool_definition();
        assert_eq!(def["function"]["name"], "lifecycle");
        let enum_vals: Vec<&str> = def["function"]["parameters"]["properties"]["phase"]["enum"]
            .as_array()
            .unwrap()
            .iter()
            .map(|v| v.as_str().unwrap())
            .collect();
        let vocab: Vec<&str> = crate::tooling::Phase::ALL
            .iter()
            .map(|p| p.as_str())
            .collect();
        assert_eq!(enum_vals, vocab);
    }

    #[test]
    fn run_phase_aliases_route_to_lifecycle() {
        for a in ["run_phase", "run_lifecycle", "lifecycle_run"] {
            assert!(
                matches!(
                    resolve_tool_alias(a),
                    Some(AliasOutcome::Rewrite("lifecycle"))
                ),
                "{a} should rewrite to lifecycle"
            );
        }
        // The canonical name is NOT an alias — it dispatches directly.
        assert!(resolve_tool_alias("lifecycle").is_none());
    }

    #[tokio::test]
    async fn lifecycle_unknown_phase_lists_valid_phases() {
        // An unknown phase returns before any fs/subprocess touch, so this is a
        // fully-mocked unit test.
        let caveats = crate::caveats::Caveats::top();
        let args = serde_json::json!({ "phase": "deploy" });
        let out = execute_tool(
            "lifecycle",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )
        .await;
        assert!(
            out.starts_with("error: unknown lifecycle phase 'deploy'"),
            "{out}"
        );
        assert!(out.contains("check"), "should name valid phases: {out}");
    }

    /// `save_note` is sink-gated: absent from the base `tool_definitions`
    /// (headless/eval callers see no memory tool) and from the merged set
    /// without a sink; present in the merged set when a sink exists.
    #[test]
    fn save_note_advertised_only_with_a_sink() {
        fn names(defs: &serde_json::Value) -> Vec<&str> {
            defs.as_array()
                .unwrap()
                .iter()
                .filter_map(|d| d["function"]["name"].as_str())
                .collect()
        }
        // Headless/eval callers see no memory tool in the base set …
        let base = tool_definitions();
        assert!(!names(&base).contains(&"save_note"), "got: {base}");
        // … nor in the merged set without a sink …
        let without = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        assert!(!names(&without).contains(&"save_note"));
        // … but a sink advertises it.
        let with = merged_tool_definitions(
            &NoMcp, true, false, false, false, false, false, false, false, false,
        );
        assert!(names(&with).contains(&"save_note"), "got: {with}");
    }

    /// `recall` is source-gated exactly like `save_note` is sink-gated
    /// (Step 17.5): absent from the base set and from the merged set
    /// without a source; present when one exists.
    #[test]
    fn recall_advertised_only_with_a_source() {
        fn names(defs: &serde_json::Value) -> Vec<&str> {
            defs.as_array()
                .unwrap()
                .iter()
                .filter_map(|d| d["function"]["name"].as_str())
                .collect()
        }
        let base = tool_definitions();
        assert!(!names(&base).contains(&"recall"), "got: {base}");
        let without = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        assert!(!names(&without).contains(&"recall"));
        let with = merged_tool_definitions(
            &NoMcp, false, true, false, false, false, false, false, false, false,
        );
        assert!(names(&with).contains(&"recall"), "got: {with}");
        // The two gates are independent: both on advertises both.
        let both = merged_tool_definitions(
            &NoMcp, true, true, false, false, false, false, false, false, false,
        );
        assert!(names(&both).contains(&"save_note"));
        assert!(names(&both).contains(&"recall"));
    }

    /// `memory_fetch` is source-gated exactly like `recall` (#319): absent
    /// from the base set and from the merged set without a `MemorySource`;
    /// present when one exists. The flag is independent of the others.
    #[test]
    fn memory_fetch_advertised_only_with_a_source() {
        fn names(defs: &serde_json::Value) -> Vec<&str> {
            defs.as_array()
                .unwrap()
                .iter()
                .filter_map(|d| d["function"]["name"].as_str())
                .collect()
        }
        let base = tool_definitions();
        assert!(!names(&base).contains(&"memory_fetch"), "got: {base}");
        // Flag off (every existing caller, the inert default) → not advertised.
        let without = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        assert!(!names(&without).contains(&"memory_fetch"));
        // Flag on → advertised.
        let with = merged_tool_definitions(
            &NoMcp, false, false, true, false, false, false, false, false, false,
        );
        assert!(names(&with).contains(&"memory_fetch"), "got: {with}");
        // Independent of the save_note / recall gates: all three on lists all.
        let all = merged_tool_definitions(
            &NoMcp, true, true, true, false, false, false, false, false, false,
        );
        assert!(names(&all).contains(&"save_note"));
        assert!(names(&all).contains(&"recall"));
        assert!(names(&all).contains(&"memory_fetch"));
    }

    /// `is_hallucination` correctly identifies tool-name-as-command and unknown
    /// tool names, and correctly skips MCP-namespaced tools.
    #[test]
    fn hallucination_detection_coverage() {
        // tool name passed to run_command → hallucination
        assert!(is_hallucination(
            "run_command",
            &serde_json::json!({"command": "list_dir ."})
        ));
        // normal shell command → not a hallucination
        assert!(!is_hallucination(
            "run_command",
            &serde_json::json!({"command": "cargo test"})
        ));
        // unknown tool → hallucination
        assert!(is_hallucination(
            "definitely_not_a_real_tool",
            &serde_json::json!({})
        ));
        // MCP-namespaced tool → not a hallucination
        assert!(!is_hallucination(
            "my_server__some_tool",
            &serde_json::json!({})
        ));
        // known direct tools → not hallucinations when called correctly
        for t in [
            "list_dir",
            "read_file",
            "write_file",
            "use_skill",
            "web_fetch",
            "save_note",
            "recall",
        ] {
            assert!(!is_hallucination(t, &serde_json::json!({"path": "."})));
        }
    }

    /// #898: `run_command_redirect` bounces LOCAL git ops (and other direct
    /// tools) to their embedded tool, but lets git NETWORK ops fall through to
    /// the shell — otherwise a model can never `git push` (the pushless embedded
    /// git tool + the blanket bounce were a hard dead-end).
    #[test]
    fn run_command_redirect_lets_git_network_ops_through() {
        // Network ops the embedded git tool cannot do → fall through (None).
        for cmd in [
            "git push origin fix/foo",
            "git push",
            "git fetch origin",
            "git pull",
            "git clone https://example.com/r.git",
        ] {
            assert_eq!(run_command_redirect(cmd), None, "{cmd} must fall through");
        }
        // Local ops the embedded git tool handles → still redirect.
        for cmd in [
            "git status",
            "git log --oneline",
            "git add .",
            "git commit -m x",
        ] {
            assert_eq!(
                run_command_redirect(cmd),
                Some("git"),
                "{cmd} must redirect"
            );
        }
        // Other direct tools still redirect; plain shell commands run as-is.
        assert_eq!(run_command_redirect("read_file foo.txt"), Some("read_file"));
        assert_eq!(run_command_redirect("list_dir ."), Some("list_dir"));
        assert_eq!(run_command_redirect("cargo test"), None);
        assert_eq!(run_command_redirect("gh pr create --fill"), None);
        assert_eq!(run_command_redirect(""), None);
    }

    /// #898 regression: a real `git push` at run_command must NOT be counted as
    /// a hallucination (it now runs), while a local `git status` still is.
    #[test]
    fn is_hallucination_allows_git_network_ops() {
        assert!(!is_hallucination(
            "run_command",
            &serde_json::json!({"command": "git push origin fix/foo"})
        ));
        assert!(!is_hallucination(
            "run_command",
            &serde_json::json!({"command": "git fetch"})
        ));
        assert!(is_hallucination(
            "run_command",
            &serde_json::json!({"command": "git status"})
        ));
    }

    /// #898: the forge PR/MR-creation URL is extracted from git's push output
    /// (GitHub and GitLab), and ordinary URLs do not false-positive.
    #[test]
    fn pr_creation_url_extracts_github_and_gitlab() {
        let github = "remote: Create a pull request for 'fix/foo' on GitHub by visiting:\n\
                      remote:      https://github.com/OWNER/REPO/pull/new/fix/foo\n";
        assert_eq!(
            pr_creation_url(github),
            Some("https://github.com/OWNER/REPO/pull/new/fix/foo")
        );
        let gitlab = "remote: To create a merge request for topic, visit:\n\
                      remote:   https://gitlab.com/g/p/-/merge_requests/new?x=topic\n";
        assert_eq!(
            pr_creation_url(gitlab),
            Some("https://gitlab.com/g/p/-/merge_requests/new?x=topic")
        );
        // No PR URL present → None (ordinary fetch/clone output, plain links).
        assert_eq!(pr_creation_url("Already up to date.\n"), None);
        assert_eq!(
            pr_creation_url("see https://github.com/OWNER/REPO/issues/1"),
            None
        );
    }

    /// #898: after a push whose output carries a PR-creation URL,
    /// `shell_envelope_output` appends the `gh pr create` next-step hint (and the
    /// URL survives), while ordinary command output is left untouched.
    #[test]
    fn shell_envelope_output_appends_pr_hint_on_push() {
        let push = serde_json::json!({
            "exit_code": 0,
            "stdout": "",
            "stderr": "remote: Create a pull request for 'fix/foo' on GitHub by visiting:\n\
                       remote:      https://github.com/OWNER/REPO/pull/new/fix/foo\n",
        });
        let out = shell_envelope_output(&push, 50, false, false, None);
        assert!(out.contains("gh pr create --fill"), "hint missing: {out}");
        assert!(
            out.contains("https://github.com/OWNER/REPO/pull/new/fix/foo"),
            "url dropped: {out}"
        );

        // Ordinary output: no hint, payload unchanged.
        let plain = serde_json::json!({ "exit_code": 0, "stdout": "hello\n", "stderr": "" });
        let out = shell_envelope_output(&plain, 50, false, false, None);
        assert!(!out.contains("gh pr create"), "spurious hint: {out}");
        assert_eq!(out, "hello\n");
    }

    #[test]
    fn shell_envelope_output_spills_full_output_before_head_tail_cap() {
        let full = format!(
            "HEAD_ONLY_MARKER\n{}\nMIDDLE_ONLY_MARKER\n{}\nTAIL_ONLY_MARKER\n",
            "alpha\n".repeat(10_000),
            "omega\n".repeat(10_000)
        );
        let envelope = serde_json::json!({
            "exit_code": 0,
            "stdout": full,
            "stderr": "",
        });
        let store = spill::SessionSpillStore::default();
        let out = shell_envelope_output(&envelope, 50, false, true, Some(&store));

        assert!(out.contains("HEAD_ONLY_MARKER"), "head dropped: {out}");
        assert!(out.contains("TAIL_ONLY_MARKER"), "tail dropped: {out}");
        assert!(
            out.contains("memory_fetch(\"spill:s0\")"),
            "spill handle missing: {out}"
        );
        assert!(
            out.contains("grep=\"<pattern>\""),
            "search affordance missing: {out}"
        );
        let stored = store.fetch("s0").expect("full output stored");
        assert!(
            stored.contains("MIDDLE_ONLY_MARKER"),
            "spilled payload was capped before storage"
        );
        assert!(stored.ends_with("TAIL_ONLY_MARKER\n"));
    }

    #[test]
    fn envelope_denied_reads_structured_flag_only() {
        assert!(envelope_denied(&serde_json::json!({"denied": true})));
        assert!(!envelope_denied(&serde_json::json!({"denied": false})));
        assert!(!envelope_denied(&serde_json::json!({})));
        // A non-bool `denied` is treated as not-denied, never a panic.
        assert!(!envelope_denied(&serde_json::json!({"denied": "yes"})));
    }

    #[test]
    fn envelope_denial_reason_joins_or_falls_back() {
        let multi = serde_json::json!({
            "denials": [
                {"kind": "exec", "target": "rm", "reason": "exec rm denied"},
                {"kind": "open", "target": "/etc/shadow", "reason": "open denied"}
            ]
        });
        assert_eq!(
            envelope_denial_reason(&multi),
            "exec rm denied; open denied"
        );
        // Missing or empty denials → the generic message, never a panic.
        let generic = "denied: the capability leash refused an operation";
        assert_eq!(envelope_denial_reason(&serde_json::json!({})), generic);
        assert_eq!(
            envelope_denial_reason(&serde_json::json!({"denials": []})),
            generic
        );
        // Entries without a string `reason` are skipped.
        assert_eq!(
            envelope_denial_reason(&serde_json::json!({"denials": [{"kind": "exec"}]})),
            generic
        );
    }

    #[test]
    fn exec_allowlist_name_takes_basename() {
        assert_eq!(exec_allowlist_name("env"), "env");
        assert_eq!(exec_allowlist_name("/usr/bin/env"), "env");
        assert_eq!(exec_allowlist_name("/usr/bin/"), "bin");
        assert_eq!(exec_allowlist_name("C:\\tools\\env.exe"), "env.exe");
    }

    /// #775 (§2.5): the human exec-denial NOTICE shows the BARE command
    /// target(s), never the reason sentence — restoring [`print_denied`]'s
    /// `'{target}'` contract. Stuffing the full reason there produced the
    /// field-report garble `capability denied: exec does not permit '<whole
    /// reason sentence>'`.
    #[test]
    fn exec_denial_target_label_is_the_bare_command_not_the_reason() {
        let one = serde_json::json!({
            "denied": true,
            "denials": [{
                "kind": "exec",
                "target": "export",
                "reason": "exec of \"export\" is not within the granted authority"
            }]
        });
        let label = exec_denial_target_label(&one);
        assert_eq!(label, "export");
        // It is the bare command — NEVER the reason sentence (which, in the
        // `'{target}'` slot, was the nested garble).
        assert!(!label.contains("is not within the granted authority"));
        // Multiple targets join cleanly; an envelope with no target falls back
        // to a generic label so the notice still prints one clean line.
        let multi = serde_json::json!({
            "denials": [
                {"kind": "exec", "target": "export", "reason": "r"},
                {"kind": "exec", "target": "set", "reason": "r"}
            ]
        });
        assert_eq!(exec_denial_target_label(&multi), "export, set");
        assert_eq!(
            exec_denial_target_label(&serde_json::json!({})),
            "a command"
        );
    }

    #[test]
    fn host_of_url_extracts_hosts_conservatively() {
        assert_eq!(host_of_url("https://docs.rs/serde"), Some("docs.rs".into()));
        assert_eq!(host_of_url("http://Docs.RS"), Some("docs.rs".into()));
        assert_eq!(
            host_of_url("https://user:pw@example.com:8443/p?q#f"),
            Some("example.com".into())
        );
        assert_eq!(host_of_url("https://[::1]:8080/x"), Some("::1".into()));
        // Unparseable / non-http inputs skip the pre-check (None) rather
        // than guessing — enforcement stays with the leash either way.
        assert_eq!(host_of_url("not a url"), None);
        assert_eq!(host_of_url("ftp://example.com"), None);
        assert_eq!(host_of_url("https:///path-only"), None);
    }

    #[test]
    fn exec_denial_requests_lifts_only_pure_exec_envelopes() {
        // The promptable case: every entry is an exec denial with a target;
        // the request target is the allowlist basename (the grantable name).
        let exec_only = serde_json::json!({
            "denied": true,
            "denials": [
                {"kind": "exec", "target": "/usr/bin/npm", "reason": "exec npm denied"},
                {"kind": "exec", "target": "node", "reason": "exec node denied"}
            ]
        });
        let reqs = exec_denial_requests(&exec_only).expect("promptable");
        assert_eq!(reqs.len(), 2);
        assert_eq!(reqs[0].tool, "run_command");
        assert_eq!(reqs[0].kind, DenialKind::Exec);
        assert_eq!(
            reqs[0].target, "npm",
            "basename, same rule as the config hint"
        );
        assert_eq!(reqs[0].reason, "exec npm denied");
        assert_eq!(reqs[1].target, "node");

        // A non-exec entry anywhere keeps the standard denial: mapping an
        // opaque `open` onto an fs axis would over-grant.
        let mixed = serde_json::json!({
            "denials": [
                {"kind": "exec", "target": "npm", "reason": "r"},
                {"kind": "open", "target": "/etc/shadow", "reason": "r"}
            ]
        });
        assert!(exec_denial_requests(&mixed).is_none());

        // Missing/empty pieces are never promptable.
        assert!(exec_denial_requests(&serde_json::json!({})).is_none());
        assert!(exec_denial_requests(&serde_json::json!({"denials": []})).is_none());
        let empty_target = serde_json::json!({
            "denials": [{"kind": "exec", "target": "", "reason": "r"}]
        });
        assert!(exec_denial_requests(&empty_target).is_none());
        let no_target = serde_json::json!({
            "denials": [{"kind": "exec", "reason": "r"}]
        });
        assert!(exec_denial_requests(&no_target).is_none());
    }

    /// #905: a NET denial envelope (agent-bridle #196 shape) lifts to a per-host
    /// net PermissionRequest; the target is the CONNECT host verbatim (no
    /// basename mangling). Non-net / mixed / empty batches stay flat.
    #[test]
    fn net_denial_requests_lifts_only_pure_net_envelopes() {
        let net_only = serde_json::json!({
            "denied": true,
            "denials": [
                {"kind": "net", "target": "github.com", "reason": "net does not permit 'github.com'"},
                {"kind": "net", "target": "api.github.com", "reason": "net does not permit 'api.github.com'"}
            ]
        });
        let reqs = net_denial_requests(&net_only).expect("promptable");
        assert_eq!(reqs.len(), 2);
        assert_eq!(reqs[0].tool, "run_command");
        assert_eq!(reqs[0].kind, DenialKind::Net);
        assert_eq!(
            reqs[0].target, "github.com",
            "host verbatim, not a basename"
        );
        assert_eq!(reqs[0].reason, "net does not permit 'github.com'");
        assert_eq!(reqs[1].target, "api.github.com");

        // A non-net entry anywhere → not net-promptable (exec lifter handles exec).
        let mixed = serde_json::json!({
            "denials": [
                {"kind": "net", "target": "github.com", "reason": "r"},
                {"kind": "exec", "target": "npm", "reason": "r"}
            ]
        });
        assert!(net_denial_requests(&mixed).is_none());
        // Exec-only is not net-promptable; empty/missing targets never are.
        let exec_only = serde_json::json!({"denials": [{"kind": "exec", "target": "npm"}]});
        assert!(net_denial_requests(&exec_only).is_none());
        assert!(net_denial_requests(&serde_json::json!({"denials": []})).is_none());
        let empty_target = serde_json::json!({"denials": [{"kind": "net", "target": ""}]});
        assert!(net_denial_requests(&empty_target).is_none());
    }

    /// #905: the human denial NOTICE labels a pure-net refusal `net` (not `exec`),
    /// so it never reads "exec does not permit '<host>'". Exec / mixed stay `exec`.
    #[test]
    fn denial_axis_label_is_net_only_for_pure_net() {
        let net = serde_json::json!({"denials": [{"kind": "net", "target": "github.com"}]});
        assert_eq!(denial_axis_label(&net), "net");
        let exec = serde_json::json!({"denials": [{"kind": "exec", "target": "rm"}]});
        assert_eq!(denial_axis_label(&exec), "exec");
        let mixed = serde_json::json!({
            "denials": [{"kind": "net", "target": "h"}, {"kind": "exec", "target": "rm"}]
        });
        assert_eq!(denial_axis_label(&mixed), "exec", "mixed defaults to exec");
        assert_eq!(denial_axis_label(&serde_json::json!({})), "exec");
    }

    #[test]
    fn tui_permits_path_prefix_semantics() {
        use crate::caveats::Scope;
        assert!(tui_permits_path(&Scope::All, "/anything/at/all"));
        assert!(!tui_permits_path(&Scope::<String>::none(), "/ws/file"));
        let only = Scope::only(["/ws".to_string()]);
        assert!(tui_permits_path(&only, "/ws/sub/file.rs"));
        assert!(tui_permits_path(&only, "/ws"), "the workspace root itself");
        assert!(!tui_permits_path(&only, "/elsewhere/file.rs"));
        // `..` traversal must NOT escape: a path that lexically resolves outside
        // the workspace is denied even though it textually begins with it.
        assert!(
            !tui_permits_path(&only, "/ws/../etc/passwd"),
            "`..` traversal escapes the workspace"
        );
        assert!(
            !tui_permits_path(&only, "/ws/../../etc/passwd"),
            "repeated `..` traversal escapes the workspace"
        );
        // A sibling dir that merely shares the string prefix is not under /ws.
        assert!(
            !tui_permits_path(&only, "/ws-secret/file.rs"),
            "sibling-prefix collision escapes the workspace"
        );
        // A `..` that stays inside the workspace is still permitted.
        assert!(tui_permits_path(&only, "/ws/sub/../file.rs"));
    }

    /// Ratchet for the OPEN `fs-canonical-containment` deviation (issue #522,
    /// `docs/security/ocap-deviations.md`). `tui_permits_path` is string-lexical:
    /// it collapses `..` but does NOT resolve symlinks, so a link *inside* the
    /// workspace pointing OUT is permitted even though the OS would read the
    /// outside target. This test builds the path the call sites do
    /// (`workspace.join(model_path)`) over a REAL symlink and PINS that residual.
    ///
    /// When canonicalize-then-contain lands (the deviation's closure criterion),
    /// the gate will deny the symlinked path and this assertion MUST flip to
    /// `!tui_permits_path(...)` — that break is the signal to close the deviation.
    /// Unix-only: Windows symlinks need privileges (mirrors
    /// `find_does_not_follow_symlinks_out_of_workspace`).
    #[cfg(unix)]
    #[test]
    fn tui_permits_path_symlink_escape_is_the_known_residual() {
        use crate::caveats::Scope;
        let outside = tempfile::TempDir::new().unwrap();
        std::fs::write(outside.path().join("secret"), b"x").unwrap();
        let ws = tempfile::TempDir::new().unwrap();
        // A symlink under the workspace whose target is OUTSIDE it.
        std::os::unix::fs::symlink(outside.path(), ws.path().join("link")).unwrap();

        let only = Scope::only([ws.path().to_string_lossy().into_owned()]);

        // What the read/write call sites feed the gate for model path "link/secret".
        let via_link = ws.path().join("link").join("secret");
        // RESIDUAL: permitted today — the gate can't see through the symlink.
        // Flip to `!` when the gate canonicalizes (closes fs-canonical-containment).
        assert!(
            tui_permits_path(&only, &via_link.to_string_lossy()),
            "string gate permits a symlinked escape — known residual (#522)"
        );

        // Contrast: a plain `..` escape through the SAME root is already denied
        // (lexical containment, the part #502 did fix) — so this isn't a blanket
        // hole, only the symlink-resolution gap.
        let dotdot = ws.path().join("..").join("etc").join("passwd");
        assert!(
            !tui_permits_path(&only, &dotdot.to_string_lossy()),
            "`..` escape is denied even though symlink escape is not"
        );
    }

    // --- PR4: the `git` tool is presence-gated -----------------------------

    #[test]
    fn git_tool_advertised_only_with_the_presence_gate() {
        fn names(defs: &serde_json::Value) -> Vec<&str> {
            defs.as_array()
                .unwrap()
                .iter()
                .filter_map(|d| d["function"]["name"].as_str())
                .collect()
        }
        let with = merged_tool_definitions(
            &NoMcp, false, false, false, true, false, false, false, false, false,
        );
        assert!(names(&with).contains(&"git"), "with_git advertises git");
        let without = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        assert!(!names(&without).contains(&"git"), "no git without the gate");
        // #479: the /team toggle advertises both crew tools, and only then.
        let team = merged_tool_definitions(
            &NoMcp, false, false, false, false, true, false, false, false, false,
        );
        assert!(
            names(&team).contains(&"crew") && names(&team).contains(&"compose_roster"),
            "with_team advertises crew + compose_roster"
        );
        assert!(
            !names(&without).contains(&"crew"),
            "no crew without the gate"
        );
        // Step 26.4 (#583): the scratchpad state tools, only with the gate on.
        let scratch = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, true, false, false, false,
        );
        for t in ["state_set", "state_get", "state_clear"] {
            assert!(
                names(&scratch).contains(&t),
                "{t} advertised with_scratchpad"
            );
            assert!(!names(&without).contains(&t), "{t} hidden without the gate");
            assert!(
                !is_hallucination(t, &serde_json::json!({})),
                "{t} is a real tool"
            );
        }
        // Step 26.5.5 (#582): the code_search tool, only with its gate on.
        let code = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, true, false, false,
        );
        assert!(
            names(&code).contains(&"code_search"),
            "code_search advertised"
        );
        assert!(
            !names(&without).contains(&"code_search"),
            "code_search hidden without the gate"
        );
        assert!(!is_hallucination("code_search", &serde_json::json!({})));
        // Step 26.6a (#585): the experiential record/recall tools, only with the gate.
        let exp = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, true, false,
        );
        for t in ["experience_record", "experience_recall"] {
            assert!(names(&exp).contains(&t), "{t} advertised with_experiential");
            assert!(!names(&without).contains(&t), "{t} hidden without the gate");
            assert!(
                !is_hallucination(t, &serde_json::json!({})),
                "{t} is a real tool"
            );
        }
        // Step 26.6b (#586) / #715 PR2: the scheduled update_plan + plan_get tools,
        // only with the gate (plan_set/plan_advance collapsed into update_plan).
        let sched = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, true,
        );
        for t in ["update_plan", "plan_get"] {
            assert!(names(&sched).contains(&t), "{t} advertised with_scheduled");
            assert!(!names(&without).contains(&t), "{t} hidden without the gate");
            assert!(
                !is_hallucination(t, &serde_json::json!({})),
                "{t} is a real tool"
            );
        }
    }

    #[tokio::test]
    async fn state_tools_dispatch_only_with_a_store() {
        use crate::agentic::scratchpad::{ScratchpadStore, SessionScratchpadStore};
        let caveats = crate::caveats::Caveats::top();
        let args = serde_json::json!({ "key": "k", "value": "v" });
        // Step 26.4: without a store the tool was never advertised → unknown.
        let none = execute_tool(
            "state_set",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )
        .await;
        assert!(none.starts_with("unknown tool: state_set"), "{none}");
        // With a store → routes to the executor and mutates it.
        let store = SessionScratchpadStore::default();
        let set = execute_tool(
            "state_set",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            Some(&store as &dyn ScratchpadStore),
            None,
            None,
            None,
        )
        .await;
        assert_eq!(set, "stored: k");
        assert_eq!(store.get("k").as_deref(), Some("v"));
    }

    #[tokio::test]
    async fn code_search_dispatch_only_with_a_searcher() {
        use crate::agentic::semantic::{CodeSearch, Embedder, SessionSemanticIndex};
        struct E;
        #[async_trait::async_trait]
        impl Embedder for E {
            async fn embed(&self, _t: &str) -> anyhow::Result<Vec<f32>> {
                Ok(vec![1.0])
            }
        }
        let caveats = crate::caveats::Caveats::top();
        let args = serde_json::json!({ "query": "find it" });
        // Step 26.5.5: no searcher → unknown tool (presence-gate parity).
        let none = execute_tool(
            "code_search",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )
        .await;
        assert!(none.starts_with("unknown tool: code_search"), "{none}");
        // with a searcher (empty index) → routes to the executor (labelled no-match).
        let idx = SessionSemanticIndex::default();
        let search = CodeSearch {
            embedder: &E,
            index: &idx,
            top_k: 1,
        };
        let out = execute_tool(
            "code_search",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            Some(search),
            None,
            None,
        )
        .await;
        assert!(out.contains("no code matched"), "{out}");
    }

    #[tokio::test]
    async fn experiential_dispatch_only_with_a_store() {
        use crate::agentic::experiential::{ExperienceStore, SessionExperienceStore};
        let caveats = crate::caveats::Caveats::top();
        let args = serde_json::json!({
            "task": "ci flake", "outcome": "fixed", "lesson": "pin the seed for the fuzz test"
        });
        // Step 26.6a: no store → unknown tool for BOTH arms (presence-gate parity).
        for name in ["experience_record", "experience_recall"] {
            let out = execute_tool(
                name, &args, ".", false, 20, &caveats, &mut NoMcp, None, None, None, None, None,
                None, None, None, None, None, None, None,
            )
            .await;
            assert!(out.starts_with(&format!("unknown tool: {name}")), "{out}");
        }
        // with a store → record routes to the executor and mutates it.
        let store = SessionExperienceStore::default();
        let out = execute_tool(
            "experience_record",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            Some(&store as &dyn ExperienceStore),
            None,
        )
        .await;
        assert_eq!(out, "recorded experience");
        assert_eq!(store.count(), 1);
    }

    #[tokio::test]
    async fn scheduled_dispatch_only_with_a_ledger() {
        use crate::agentic::scheduled::{SessionStepLedger, StepLedger};
        let caveats = crate::caveats::Caveats::top();
        let args = serde_json::json!({ "plan": [
            { "step": "a", "status": "in_progress" },
            { "step": "b", "status": "pending" },
        ] });
        // Step 26.6b / #716 / #715 PR2: no ledger → unknown tool for ALL plan arms
        // (presence-gate parity, including the read-only plan_get).
        for name in ["update_plan", "plan_get"] {
            let out = execute_tool(
                name, &args, ".", false, 20, &caveats, &mut NoMcp, None, None, None, None, None,
                None, None, None, None, None, None, None,
            )
            .await;
            assert!(out.starts_with(&format!("unknown tool: {name}")), "{out}");
        }
        // with a ledger → update_plan routes to the executor and mutates it.
        let ledger = SessionStepLedger::default();
        let out = execute_tool(
            "update_plan",
            &args,
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            Some(&ledger as &dyn StepLedger),
        )
        .await;
        assert!(out.starts_with("<plan>\n"), "{out}");
        assert_eq!(ledger.count(), 2);
        // #716: plan_get with a ledger renders the <plan> block, read-only.
        let got = execute_tool(
            "plan_get",
            &serde_json::json!({}),
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
            Some(&ledger as &dyn StepLedger),
        )
        .await;
        assert!(got.starts_with("<plan>\n"), "{got}");
        assert_eq!(ledger.count(), 2, "plan_get is read-only");
    }

    #[tokio::test]
    async fn resume_context_dispatch_degrades_without_a_recall_source() {
        // #714: advertised ALWAYS, so dispatch never reports "unknown tool" —
        // with no recall_source (headless) it returns the clear no-history line.
        let caveats = crate::caveats::Caveats::top();
        let out = execute_tool(
            "resume_context",
            &serde_json::json!({}),
            ".",
            false,
            20,
            &caveats,
            &mut NoMcp,
            None, // build_check_cmd
            None, // note_sink
            None, // recall_source
            None, // memory_source
            None, // permission_gate
            None, // exec_floor
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert!(
            out.contains("no conversation history available this session"),
            "{out}"
        );
        assert!(!out.starts_with("unknown tool"), "{out}");
    }

    #[test]
    fn run_build_check_reports_pass_fail_and_spawn_error() {
        let ws = tempfile::TempDir::new().unwrap();
        let ws_str = ws.path().to_string_lossy();
        assert_eq!(
            run_build_check(passing_build_check_cmd(), &ws_str),
            "  ✓ build check passed"
        );
        let failed = run_build_check(&failing_build_check_cmd("boom"), &ws_str);
        assert!(failed.contains("✗ build check failed"), "got: {failed}");
        assert!(failed.contains("boom"), "stderr excerpt shown: {failed}");
        // A nonexistent workspace dir → the command can't even spawn.
        let err = run_build_check(passing_build_check_cmd(), "/definitely/not/a/dir");
        assert!(err.contains("⚠ build check could not run"), "got: {err}");
    }
}

// ---------------------------------------------------------------------------
// execute_tool branch tests — edit_file / shrink guard / denial paths
// ---------------------------------------------------------------------------

#[cfg(test)]
mod execute_tool_branch_tests {
    use super::super::NoMcp;
    use super::*;
    use crate::caveats::{Caveats, CountBound, Scope};

    /// fs read everywhere, fs write scoped to the workspace (skips the y/N
    /// confirm — the scoped preset is the consent), nothing else.
    fn caveats_rw(ws: &std::path::Path) -> Caveats {
        Caveats {
            fs_read: Scope::All,
            fs_write: Scope::only([ws.to_string_lossy().into_owned()]),
            exec: Scope::none(),
            net: Scope::none(),
            max_calls: CountBound::Unlimited,
            valid_for_generation: Scope::All,
        }
    }

    // --- PR4: the `git` tool arm in execute_tool ---------------------------

    /// A stub GitTool: echoes the op, and refuses `commit` when the projected
    /// GitCaveats deny it — exercises the arm's caveat projection without a repo.
    struct StubGit;
    impl crate::agentic::GitTool for StubGit {
        fn dispatch(
            &self,
            op: &str,
            _args: &serde_json::Value,
            caps: &crate::git_caveats::GitCaveats,
        ) -> Result<String, String> {
            match op {
                "status" => Ok("on branch main (HEAD abc123)".to_string()),
                "commit" if !caps.permits_commit() => {
                    Err("capability denied: git commit not permitted".to_string())
                }
                "commit" => Ok("committed abc123: msg".to_string()),
                other => Err(format!("unknown git op '{other}'")),
            }
        }
    }

    async fn run_git(
        op: &str,
        caveats: &Caveats,
        git: Option<&dyn crate::agentic::GitTool>,
    ) -> String {
        let ws = tempfile::TempDir::new().unwrap();
        execute_tool(
            "git",
            &serde_json::json!({ "op": op }),
            &ws.path().to_string_lossy(),
            false,
            20,
            caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None,
            None,
            None,
            git,
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await
    }

    #[tokio::test]
    async fn git_arm_dispatches_when_injected() {
        let ws = tempfile::TempDir::new().unwrap();
        let out = run_git("status", &caveats_rw(ws.path()), Some(&StubGit)).await;
        assert!(out.contains("on branch main"), "got: {out}");
    }

    #[tokio::test]
    async fn git_arm_surfaces_denials_from_projected_caveats() {
        // A session with no fs_write → from_session denies commit_local.
        let ws = tempfile::TempDir::new().unwrap();
        let read_only = Caveats {
            fs_write: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let out = run_git("commit", &read_only, Some(&StubGit)).await;
        assert!(
            out.contains("error:") && out.contains("commit"),
            "got: {out}"
        );
        // The same session can still run a read op.
        let out = run_git("status", &read_only, Some(&StubGit)).await;
        assert!(out.contains("on branch main"), "got: {out}");
    }

    #[tokio::test]
    async fn git_arm_unknown_op_is_an_error_not_a_panic() {
        let ws = tempfile::TempDir::new().unwrap();
        let out = run_git("frobnicate", &caveats_rw(ws.path()), Some(&StubGit)).await;
        assert!(
            out.contains("error:") && out.contains("unknown git op"),
            "got: {out}"
        );
    }

    #[tokio::test]
    async fn git_arm_without_injection_is_unknown_tool() {
        let ws = tempfile::TempDir::new().unwrap();
        let out = run_git("status", &caveats_rw(ws.path()), None).await;
        assert!(out.contains("unknown tool: git"), "got: {out}");
    }

    // #479: the agent-callable crew/compose_roster tools route through the
    // injected CrewRunner — same presence-gating + dispatch shape as `git`.
    struct StubCrew;
    #[async_trait::async_trait]
    impl crate::agentic::CrewRunner for StubCrew {
        async fn dispatch(
            &self,
            op: &str,
            _args: &serde_json::Value,
            _caveats: &Caveats,
        ) -> Result<String, String> {
            match op {
                "compose_roster" => Ok("proposed roster: planner <- qwen3-coder:30b".to_string()),
                "crew" => Ok("crew ran: diff +1/-0, status PASS".to_string()),
                other => Err(format!("unknown op: {other}")),
            }
        }
    }

    async fn run_crew_tool(
        name: &str,
        args: serde_json::Value,
        crew: Option<&dyn crate::agentic::CrewRunner>,
    ) -> String {
        let ws = tempfile::TempDir::new().unwrap();
        execute_tool(
            name,
            &args,
            &ws.path().to_string_lossy(),
            false,
            20,
            &caveats_rw(ws.path()),
            &mut NoMcp,
            None, // build_check_cmd
            None, // note_sink
            None, // recall_source
            None, // memory_source
            None, // permission_gate
            None, // exec_floor
            None, // git_tool
            crew,
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await
    }

    #[tokio::test]
    async fn crew_arm_dispatches_when_injected() {
        let out = run_crew_tool(
            "crew",
            serde_json::json!({ "task": "do X" }),
            Some(&StubCrew),
        )
        .await;
        assert!(
            out.contains("crew ran") && out.contains("PASS"),
            "got: {out}"
        );
        let out = run_crew_tool(
            "compose_roster",
            serde_json::json!({ "mode": "crew" }),
            Some(&StubCrew),
        )
        .await;
        assert!(out.contains("proposed roster"), "got: {out}");
    }

    /// #479 (G4): with no `CrewRunner` injected (the OFF default), the dispatch
    /// arm coaches recovery instead of the old flat `unknown tool` dead-end — it
    /// names the operator gesture (`NEWT_TEAM`) and a real solo alternative, and
    /// must NOT read as "unknown tool".
    #[tokio::test]
    async fn crew_arm_without_injection_coaches_recovery() {
        for name in ["crew", "compose_roster"] {
            let out = run_crew_tool(name, serde_json::json!({ "task": "x" }), None).await;
            assert!(out.contains("NEWT_TEAM"), "{name}: {out}");
            assert!(out.contains("read_file"), "{name}: {out}");
            assert!(!out.contains("unknown tool"), "{name}: {out}");
        }
    }

    /// #479 (G4): the factored coach helper names the gate + a real alternative
    /// and never reads as "unknown tool" — the regression point for the wording.
    #[test]
    fn crew_off_recovery_result_names_gate_and_alternative() {
        let out = crew_off_recovery_result("crew");
        assert!(out.contains("'crew'"), "{out}");
        assert!(out.contains("NEWT_TEAM"), "{out}");
        // A real, always-available solo alternative is offered.
        assert!(out.contains("write_file"), "{out}");
        assert!(!out.contains("unknown tool"), "{out}");
    }

    /// #479 (G4): the gated-off telemetry seam. A `crew`/`compose_roster` reach
    /// with the surface OFF records a `GatedOff` phantom; the same names with the
    /// surface ON record nothing (they dispatch normally), and a non-crew name is
    /// never gated-off.
    #[test]
    fn classify_gated_off_reach_only_fires_for_off_crew_names() {
        for name in ["crew", "compose_roster"] {
            assert_eq!(
                classify_gated_off_reach(name, false),
                Some(crate::PhantomResolution::GatedOff(
                    "crew/team surface off (NEWT_TEAM)".into()
                )),
                "{name} OFF should record GatedOff"
            );
            assert_eq!(
                classify_gated_off_reach(name, true),
                None,
                "{name} ON dispatches normally — no phantom"
            );
        }
        // A non-crew tool is never a gated-off reach, OFF or ON.
        assert_eq!(classify_gated_off_reach("read_file", false), None);
        assert_eq!(classify_gated_off_reach("read_file", true), None);
    }

    /// #479 (G4) guard: the OFF-state changes do not touch `is_hallucination`
    /// (crew/compose_roster stay real names) or `classify_phantom_reach` for the
    /// crew names — both kept exactly so the ON path stays a normal dispatch.
    #[test]
    fn crew_names_stay_real_and_unflagged_by_existing_seams() {
        for name in ["crew", "compose_roster"] {
            assert!(
                !is_hallucination(name, &serde_json::json!({ "task": "x" })),
                "{name} must stay a real tool name"
            );
            assert_eq!(
                classify_phantom_reach(name, &serde_json::json!({ "task": "x" }), "ok", true),
                None,
                "{name} must not be flagged by classify_phantom_reach"
            );
        }
    }

    // --- #496: the embedded `find` tool -----------------------------------

    /// Convenience for `find` calls through the real dispatch under a
    /// read-everything session.
    async fn run_find(args: serde_json::Value, ws: &std::path::Path) -> String {
        run_tool("find", args, ws, &caveats_rw(ws), None).await
    }

    fn touch(root: &std::path::Path, rel: &str) {
        let p = root.join(rel);
        std::fs::create_dir_all(p.parent().unwrap()).unwrap();
        std::fs::write(p, b"x").unwrap();
    }

    /// Regression for #496: an agent needed `find . -name pyo3_module.rs` but
    /// the build's shell tool was unavailable. The embedded tool must locate the
    /// file by basename, ignoring decoys, and return its workspace-relative path
    /// (no shell, no `| sort`). Fails before this tool existed (`unknown tool:
    /// find`).
    #[tokio::test]
    async fn find_locates_file_by_name_issue_496() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "newt-core/src/pyo3_module.rs");
        touch(ws.path(), "newt-data/src/other.rs");
        touch(ws.path(), "docs/pyo3_module.md"); // decoy: wrong extension
        let out = run_find(serde_json::json!({ "name": "pyo3_module.rs" }), ws.path()).await;
        assert_eq!(out, "newt-core/src/pyo3_module.rs", "got: {out}");
    }

    /// The other call the blocked agent reached for:
    /// `find examples -maxdepth 2 -type f -name '*.py'`. Exercises glob + type
    /// filter + max_depth together, and confirms output is pre-sorted.
    #[tokio::test]
    async fn find_glob_type_and_maxdepth_together() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "examples/a.py"); // depth 1 — match
        touch(ws.path(), "examples/sub/b.py"); // depth 2 — match
        touch(ws.path(), "examples/sub/deep/c.py"); // depth 3 — too deep
        touch(ws.path(), "examples/readme.md"); // wrong extension
        std::fs::create_dir_all(ws.path().join("examples/empty_dir")).unwrap();
        let out = run_find(
            serde_json::json!({
                "path": "examples", "name": "*.py", "type": "f", "max_depth": 2
            }),
            ws.path(),
        )
        .await;
        // Pre-sorted, exactly the two in-depth .py files, no dir, no .md, no
        // depth-3 file — and no shell `| sort` needed.
        assert_eq!(out, "examples/a.py\nexamples/sub/b.py", "got: {out}");
    }

    /// Output is sorted ascending regardless of filesystem/creation order.
    #[tokio::test]
    async fn find_output_is_sorted() {
        let ws = tempfile::TempDir::new().unwrap();
        for f in ["m.txt", "a.txt", "z.txt", "c.txt"] {
            touch(ws.path(), f);
        }
        let out = run_find(serde_json::json!({ "name": "*.txt" }), ws.path()).await;
        let lines: Vec<&str> = out.lines().collect();
        assert_eq!(
            lines,
            vec!["a.txt", "c.txt", "m.txt", "z.txt"],
            "got: {out}"
        );
    }

    /// `type` restricts to files or directories.
    #[tokio::test]
    async fn find_type_filter() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "pkg/file.rs");
        std::fs::create_dir_all(ws.path().join("pkg/sub")).unwrap();
        let dirs = run_find(serde_json::json!({ "type": "d" }), ws.path()).await;
        assert!(
            dirs.contains("pkg") && dirs.contains("pkg/sub"),
            "got: {dirs}"
        );
        assert!(!dirs.contains("file.rs"), "dirs-only leaked a file: {dirs}");
        let files = run_find(serde_json::json!({ "type": "f" }), ws.path()).await;
        assert!(files.contains("pkg/file.rs"), "got: {files}");
        assert!(
            !files.lines().any(|l| l == "pkg" || l == "pkg/sub"),
            "files-only leaked a dir: {files}"
        );
    }

    /// .gitignore + the default build/dep skips are honoured by default and
    /// can be disabled with `respect_gitignore=false`.
    #[tokio::test]
    async fn find_gitignore_and_default_skips() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join(".gitignore"), "ignored.txt\n").unwrap();
        touch(ws.path(), "kept.txt");
        touch(ws.path(), "ignored.txt");
        touch(ws.path(), "target/build_artifact.txt");
        touch(ws.path(), "node_modules/dep.txt");

        let on = run_find(serde_json::json!({ "name": "*.txt" }), ws.path()).await;
        assert!(on.contains("kept.txt"), "got: {on}");
        assert!(!on.contains("ignored.txt"), "gitignore not honoured: {on}");
        assert!(!on.contains("target/"), "target not skipped: {on}");
        assert!(
            !on.contains("node_modules/"),
            "node_modules not skipped: {on}"
        );

        let off = run_find(
            serde_json::json!({ "name": "*.txt", "respect_gitignore": false }),
            ws.path(),
        )
        .await;
        assert!(off.contains("ignored.txt"), "opt-out should show it: {off}");
        assert!(off.contains("target/build_artifact.txt"), "got: {off}");
    }

    /// `max_results` caps output and the result notes the truncation.
    #[tokio::test]
    async fn find_max_results_caps_and_notes_truncation() {
        let ws = tempfile::TempDir::new().unwrap();
        for i in 0..10 {
            touch(ws.path(), &format!("f{i}.txt"));
        }
        let out = run_find(
            serde_json::json!({ "name": "*.txt", "max_results": 3 }),
            ws.path(),
        )
        .await;
        let body: Vec<&str> = out.lines().filter(|l| l.ends_with(".txt")).collect();
        assert_eq!(body.len(), 3, "should cap at 3: {out}");
        assert!(out.contains("truncated at 3"), "got: {out}");
    }

    /// A missing root is a clear error, and an empty match set says so.
    #[tokio::test]
    async fn find_missing_root_and_no_matches() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "a.txt");
        let missing = run_find(serde_json::json!({ "path": "does/not/exist" }), ws.path()).await;
        assert!(missing.starts_with("error:"), "got: {missing}");
        let empty = run_find(serde_json::json!({ "name": "*.nope" }), ws.path()).await;
        assert_eq!(empty, "no matches", "got: {empty}");
    }

    /// fs_read denial: no scope + no prompt gate ⇒ capability denied (same UX
    /// as list_dir/read_file).
    #[tokio::test]
    async fn find_denied_without_fs_read() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "secret.txt");
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let out = run_tool(
            "find",
            serde_json::json!({ "name": "*" }),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert!(out.starts_with("capability denied"), "got: {out}");
    }

    /// A `..` root that escapes the workspace is refused even when the session
    /// grants fs_read everywhere (defence-in-depth for a recursive read).
    #[tokio::test]
    async fn find_refuses_root_outside_workspace() {
        let parent = tempfile::TempDir::new().unwrap();
        std::fs::write(parent.path().join("outside.txt"), b"x").unwrap();
        let ws = parent.path().join("ws");
        std::fs::create_dir_all(&ws).unwrap();
        // fs_read: All, so the only thing that can stop the escape is the
        // canonical-root containment check.
        let out = run_find(serde_json::json!({ "path": ".." }), &ws).await;
        assert!(out.starts_with("capability denied"), "got: {out}");
    }

    /// An empty `name` is treated as "match everything" (the `!g.is_empty()`
    /// guard routes `Some("")` to the no-filter path; without it the glob would
    /// compile to `^$` and match nothing).
    #[tokio::test]
    async fn find_empty_name_matches_everything() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "a.txt");
        touch(ws.path(), "sub/b.rs");
        let out = run_find(serde_json::json!({ "name": "" }), ws.path()).await;
        for expected in ["a.txt", "sub", "sub/b.rs"] {
            assert!(
                out.lines().any(|l| l == expected),
                "empty name should match `{expected}`: {out}"
            );
        }
    }

    /// Hidden entries (dotfiles / dotdirs) are pruned by default and surface
    /// only when `respect_gitignore=false` — relevant because dotfiles can hold
    /// secrets (.env, .ssh). Pins the `.hidden(respect_gitignore)` branch.
    #[tokio::test]
    async fn find_hidden_entries_gated_by_respect_gitignore() {
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "visible.txt");
        touch(ws.path(), ".hidden.txt");
        touch(ws.path(), ".config/secret.txt");

        let default = run_find(serde_json::json!({ "name": "*" }), ws.path()).await;
        assert!(
            default.lines().any(|l| l == "visible.txt"),
            "got: {default}"
        );
        assert!(
            !default.contains(".hidden") && !default.contains(".config"),
            "hidden entries must be skipped by default: {default}"
        );

        let all = run_find(
            serde_json::json!({ "name": "*", "respect_gitignore": false }),
            ws.path(),
        )
        .await;
        assert!(all.contains(".hidden.txt"), "opt-out should show it: {all}");
        assert!(all.contains(".config/secret.txt"), "got: {all}");
    }

    /// Security boundary: `find` never follows symlinked directories, so a link
    /// pointing outside the workspace cannot leak the target's contents (pins
    /// `.follow_links(false)`). Unix-only — Windows symlinks need privileges.
    #[cfg(unix)]
    #[tokio::test]
    async fn find_does_not_follow_symlinks_out_of_workspace() {
        let outside = tempfile::TempDir::new().unwrap();
        std::fs::write(outside.path().join("secret.txt"), b"x").unwrap();
        let ws = tempfile::TempDir::new().unwrap();
        touch(ws.path(), "inside.txt");
        std::os::unix::fs::symlink(outside.path(), ws.path().join("link")).unwrap();

        // The symlink is present but is NOT descended into.
        let leaked = run_find(serde_json::json!({ "name": "secret.txt" }), ws.path()).await;
        assert_eq!(
            leaked, "no matches",
            "symlink was followed out of ws: {leaked}"
        );
        // Sanity: a real in-workspace file is still found.
        let found = run_find(serde_json::json!({ "name": "inside.txt" }), ws.path()).await;
        assert_eq!(found, "inside.txt", "got: {found}");
    }

    #[test]
    fn glob_to_regex_anchors_and_escapes() {
        // '*' is a wildcard; '.' is literal (not "any char").
        let re = glob_to_regex("*.py", true).unwrap();
        assert!(re.is_match("foo.py"));
        assert!(!re.is_match("foo.pyc")); // anchored at end
        assert!(!re.is_match("fooxpy")); // '.' is literal
                                         // Exact basename, '?' = single char, case-sensitivity honoured.
        assert!(glob_to_regex("a?c", true).unwrap().is_match("abc"));
        assert!(!glob_to_regex("a?c", true).unwrap().is_match("ac"));
        assert!(glob_to_regex("readme.md", false)
            .unwrap()
            .is_match("README.MD"));
        assert!(!glob_to_regex("readme.md", true)
            .unwrap()
            .is_match("README.MD"));
    }

    async fn run_tool(
        name: &str,
        args: serde_json::Value,
        ws: &std::path::Path,
        caveats: &Caveats,
        build_check: Option<&str>,
    ) -> String {
        execute_tool(
            name,
            &args,
            &ws.to_string_lossy(),
            false,
            20,
            caveats,
            &mut NoMcp,
            build_check,
            None,
            None,
            None, // memory_source
            None,
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await
    }

    #[tokio::test]
    async fn edit_file_replaces_unique_match_and_reports_delta() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("f.txt"), "hello world\nsecond line\n").unwrap();
        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "edit_file",
            serde_json::json!({
                "path": "f.txt",
                "old_string": "world",
                "new_string": "rust\nand more"
            }),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.starts_with("edited f.txt (+1 lines"), "got: {out}");
        assert_eq!(
            std::fs::read_to_string(ws.path().join("f.txt")).unwrap(),
            "hello rust\nand more\nsecond line\n"
        );
    }

    #[tokio::test]
    async fn edit_file_rejects_empty_missing_and_ambiguous_old_string() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("f.txt"), "dup\ndup\n").unwrap();
        let caveats = caveats_rw(ws.path());

        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "", "new_string": "x"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.contains("old_string must not be empty"), "got: {out}");

        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "absent", "new_string": "x"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.contains("old_string not found in f.txt"), "got: {out}");
        // The miss error now shows the file's actual contents so the model can
        // copy the exact text instead of blind-guessing old_string again.
        assert!(out.contains("do not guess again"), "got: {out}");
        assert!(
            out.contains("dup"),
            "miss error must include the file content: {out}"
        );

        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "dup", "new_string": "x"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.contains("matches 2 locations"), "got: {out}");
        // The ambiguous edit must NOT have touched the file.
        assert_eq!(
            std::fs::read_to_string(ws.path().join("f.txt")).unwrap(),
            "dup\ndup\n"
        );
    }

    #[tokio::test]
    async fn edit_file_denied_outside_fs_write_scope_and_missing_file() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = Caveats {
            fs_write: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "a", "new_string": "b"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(
            out.contains("capability denied: fs_write"),
            "denied before any fs access, got: {out}"
        );

        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "missing.txt", "old_string": "a", "new_string": "b"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.contains("error reading missing.txt"), "got: {out}");
    }

    #[tokio::test]
    async fn edit_file_appends_build_check_result() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("f.txt"), "old\n").unwrap();
        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "old", "new_string": "new"}),
            ws.path(),
            &caveats,
            Some(passing_build_check_cmd()),
        )
        .await;
        assert!(out.contains("✓ build check passed"), "got: {out}");

        let failing_check = failing_build_check_cmd("broke");
        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "new", "new_string": "newer"}),
            ws.path(),
            &caveats,
            Some(&failing_check),
        )
        .await;
        assert!(out.contains("✗ build check failed"), "got: {out}");
        assert!(out.contains("broke"), "model sees the failure text: {out}");
    }

    #[tokio::test]
    async fn write_file_shrink_guard_refuses_large_deletion() {
        let ws = tempfile::TempDir::new().unwrap();
        let big: String = (0..100).map(|i| format!("line {i}\n")).collect();
        std::fs::write(ws.path().join("big.txt"), &big).unwrap();
        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "write_file",
            serde_json::json!({"path": "big.txt", "content": "tiny\n"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(
            out.contains("would shrink big.txt from 100 → 1 lines"),
            "got: {out}"
        );
        assert!(out.contains("edit_file"), "points at the safer tool: {out}");
        // The guard refused — the original file must be intact.
        assert_eq!(
            std::fs::read_to_string(ws.path().join("big.txt")).unwrap(),
            big
        );
    }

    #[tokio::test]
    async fn write_file_creates_parent_directories() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "write_file",
            serde_json::json!({"path": "a/b/c.txt", "content": "nested"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.starts_with("wrote a/b/c.txt"), "got: {out}");
        assert_eq!(
            std::fs::read_to_string(ws.path().join("a/b/c.txt")).unwrap(),
            "nested"
        );
    }

    #[tokio::test]
    async fn read_file_denial_and_missing_file_errors() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("secret.txt"), "x").unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let out = run_tool(
            "read_file",
            serde_json::json!({"path": "secret.txt"}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert!(out.contains("capability denied: fs_read"), "got: {out}");

        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "read_file",
            serde_json::json!({"path": "nope.txt"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.contains("error reading nope.txt"), "got: {out}");
    }

    #[tokio::test]
    async fn list_dir_denial_and_missing_dir_errors() {
        let ws = tempfile::TempDir::new().unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let out = run_tool(
            "list_dir",
            serde_json::json!({"path": "."}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert!(out.contains("capability denied: fs_read"), "got: {out}");

        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "list_dir",
            serde_json::json!({"path": "not-a-dir"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.starts_with("error:"), "got: {out}");
    }

    #[tokio::test]
    async fn unknown_tool_name_is_reported_not_executed() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "definitely_not_a_tool",
            serde_json::json!({}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        // Step 27.1: the bare "unknown tool: X" is now a corrective message that
        // still leads with the same prefix but also names the real catalog.
        assert!(
            out.starts_with("unknown tool: definitely_not_a_tool"),
            "got: {out}"
        );
        assert!(out.contains("Available tools include:"), "got: {out}");
    }

    // -- Step 27.1: tool-alias resolution + corrective feedback -------------

    #[test]
    fn alias_rewrites_shell_names_to_run_command() {
        for n in [
            "execute",
            "exec",
            "bash",
            "shell",
            "sh",
            "zsh",
            "terminal",
            "run_shell_command",
            "shell_command",
            "system",
        ] {
            assert!(
                matches!(
                    resolve_tool_alias(n),
                    Some(AliasOutcome::Rewrite("run_command"))
                ),
                "{n} should rewrite to run_command"
            );
        }
    }

    #[test]
    fn alias_corrects_edit_and_create_names() {
        for n in [
            "str_replace_editor",
            "str_replace",
            "apply_patch",
            "edit",
            "replace_in_file",
        ] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            assert!(msg.contains("edit_file"), "{n}: {msg}");
            assert!(msg.contains("write_file"), "{n}: {msg}");
        }
        for n in ["create_file", "new_file", "touch"] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            assert!(msg.contains("write_file"), "{n}: {msg}");
        }
    }

    #[test]
    fn alias_coaches_mkdir_to_write_file() {
        // #721: newt has no directory-creation tool — coach to write_file, which
        // does create_dir_all on the parent. Turns the issue's `mkdir -p …/src`
        // dead-end into a self-correcting tool call.
        for n in [
            "mkdir",
            "make_dir",
            "makedirs",
            "mkdirs",
            "create_dir",
            "create_directory",
        ] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            assert!(msg.contains("write_file"), "{n}: {msg}");
            assert!(msg.contains("create_dir_all"), "{n}: {msg}");
        }
        // `touch` is intentionally NOT in the mkdir arm — it stays a create-file
        // alias (→ write_file), so there is no duplicate match arm / collision.
        let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias("touch") else {
            panic!("touch should still be a create-file Correct outcome");
        };
        assert!(msg.contains("write_file"), "touch: {msg}");
    }

    #[test]
    fn alias_passes_through_real_and_mcp_names() {
        for n in [
            "run_command",
            "read_file",
            "write_file",
            "edit_file",
            "git",
            "update_plan",
            "plan_get",
            "server__do_thing",
        ] {
            assert!(
                resolve_tool_alias(n).is_none(),
                "{n} must dispatch unchanged"
            );
        }
    }

    // -- #716: plan / plan-read / crew / workflow alias families --------------

    #[test]
    fn alias_corrects_plan_names_to_update_plan() {
        for n in [
            "enter_plan",
            "enter_plan_mode",
            "plan_mode",
            "start_plan",
            "begin_plan",
            "make_plan",
            "create_plan",
            "plan",
            "planning",
            "todo",
            "todos",
            "todo_write",
        ] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            assert!(msg.contains("update_plan"), "{n}: {msg}");
        }
        // #715 PR2: the advance-ish verbs coach update_plan + "completed" too.
        for n in [
            "next_step",
            "complete_step",
            "finish_step",
            "mark_done",
            "step_done",
        ] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            assert!(msg.contains("update_plan"), "{n}: {msg}");
            assert!(msg.contains("completed"), "{n}: {msg}");
        }
        // #715 PR2: update_plan is the REAL tool now → not an alias (returns None),
        // exactly like the resume_context fix; the old set_plan name is gone too.
        assert!(
            resolve_tool_alias("update_plan").is_none(),
            "update_plan must dispatch as the real tool, not a self-alias"
        );
    }

    #[test]
    fn alias_rewrites_plan_read_names_to_plan_get() {
        for n in [
            "get_plan",
            "show_plan",
            "read_plan",
            "current_plan",
            "what_was_i_doing",
        ] {
            assert!(
                matches!(
                    resolve_tool_alias(n),
                    Some(AliasOutcome::Rewrite("plan_get"))
                ),
                "{n} should rewrite to plan_get"
            );
        }
    }

    #[test]
    fn alias_rewrites_resume_reaches_to_resume_context() {
        // #714: the instinctive "where did we leave off" reaches redirect to the
        // self-recovery tool, not plan_get.
        for n in [
            "resume",
            "where_were_we",
            "where_did_we_leave_off",
            "catch_me_up",
            "recap",
        ] {
            assert!(
                matches!(
                    resolve_tool_alias(n),
                    Some(AliasOutcome::Rewrite("resume_context"))
                ),
                "{n} should rewrite to resume_context"
            );
        }
        // The REAL tool name is not an alias: it returns None so a direct
        // resume_context call dispatches as a real tool and is NOT logged as a
        // phantom Rewrite by #717 telemetry (real names must return None).
        assert!(
            resolve_tool_alias("resume_context").is_none(),
            "the real tool name must return None, not a self-Rewrite"
        );
        // No regression: `what_was_i_doing` still asks specifically for the plan.
        assert!(
            matches!(
                resolve_tool_alias("what_was_i_doing"),
                Some(AliasOutcome::Rewrite("plan_get"))
            ),
            "what_was_i_doing must stay → plan_get"
        );
    }

    #[test]
    fn alias_corrects_crew_names_and_flags_team_gating() {
        for n in [
            "delegate",
            "spawn_agent",
            "subagent",
            "sub_agent",
            "crew_dispatch",
            "run_crew",
            "dispatch_crew",
            "fork_agent",
            "assign",
            "team",
        ] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            // Names the real targets...
            assert!(msg.contains("compose_roster"), "{n}: {msg}");
            assert!(msg.contains("crew"), "{n}: {msg}");
            // ...but makes clear the model cannot self-enable the /team surface.
            assert!(msg.contains("/team"), "{n}: {msg}");
            assert!(
                msg.contains("human enables") || msg.contains("cannot turn it on yourself"),
                "crew correction must not imply the model can invoke it: {msg}"
            );
        }
    }

    #[test]
    fn alias_corrects_workflow_names_to_plan_plus_crew() {
        for n in ["workflow", "run_workflow", "start_workflow", "pipeline"] {
            let Some(AliasOutcome::Correct(msg)) = resolve_tool_alias(n) else {
                panic!("{n} should produce a Correct outcome");
            };
            assert!(msg.contains("no workflow tool"), "{n}: {msg}");
            assert!(msg.contains("update_plan"), "{n}: {msg}");
        }
    }

    #[test]
    fn levenshtein_matches_known_distances() {
        assert_eq!(levenshtein("kitten", "sitting"), 3);
        assert_eq!(levenshtein("read_file", "read_file"), 0);
        assert_eq!(levenshtein("read_fil", "read_file"), 1);
        assert_eq!(levenshtein("", "abc"), 3);
    }

    #[test]
    fn nearest_tool_name_suggests_close_only() {
        assert_eq!(nearest_tool_name("read_fil"), Some("read_file"));
        assert_eq!(nearest_tool_name("edit_fil"), Some("edit_file"));
        assert_eq!(nearest_tool_name("memory_fetchh"), Some("memory_fetch"));
        assert_eq!(nearest_tool_name("definitely_not_a_tool"), None);
    }

    #[test]
    fn unknown_tool_message_names_catalog_and_suggestion() {
        let m = unknown_tool_message("read_fil");
        assert!(m.starts_with("unknown tool: read_fil"), "{m}");
        assert!(m.contains("Did you mean 'read_file'"), "{m}");
        assert!(m.contains("Available tools include:"), "{m}");

        let m2 = unknown_tool_message("zzzzzzzzzzzz");
        assert!(m2.starts_with("unknown tool: zzzzzzzzzzzz"), "{m2}");
        assert!(!m2.contains("Did you mean"), "{m2}");
        assert!(m2.contains("Available tools include:"), "{m2}");
    }

    /// An incompatible-arg alias is corrected (not dead-ended) by execute_tool:
    /// a model that emits `str_replace_editor` is told to use edit_file. The
    /// correction returns before any fs/caveat work, so this is deterministic.
    #[tokio::test]
    async fn execute_tool_corrects_str_replace_editor_alias() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let out = run_tool(
            "str_replace_editor",
            serde_json::json!({"command": "str_replace", "path": "f.txt"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.contains("edit_file"), "got: {out}");
        assert!(!out.starts_with("unknown tool"), "got: {out}");
    }

    // -- #263 prompted permission grants through execute_tool ---------------

    /// Scripted gate: records every request it is asked about and answers
    /// allow (with caveats widened by exactly the requested grants) or deny.
    struct MockGate {
        allow: bool,
        base: Caveats,
        asks: Vec<(String, String)>,
    }

    impl MockGate {
        fn new(allow: bool, base: &Caveats) -> Self {
            Self {
                allow,
                base: base.clone(),
                asks: Vec::new(),
            }
        }
    }

    impl super::PermissionGate for MockGate {
        fn ask(&mut self, requests: &[super::PermissionRequest]) -> super::PermissionDecision {
            for r in requests {
                self.asks
                    .push((r.tool.clone(), format!("{}:{}", r.kind.as_str(), r.target)));
            }
            if self.allow {
                let grants: Vec<_> = requests
                    .iter()
                    .map(|r| (r.kind, r.target.clone()))
                    .collect();
                super::PermissionDecision::Allow(crate::agentic::widen_caveats(&self.base, &grants))
            } else {
                super::PermissionDecision::Deny
            }
        }
        // #728: this gate exercises the GRANT path only; it has no human to
        // answer free-text questions, so it returns None (a trivial impl).
        fn ask_question(&mut self, _question: &str) -> Option<String> {
            None
        }
    }

    #[allow(clippy::too_many_arguments)]
    async fn run_tool_gated(
        name: &str,
        args: serde_json::Value,
        ws: &std::path::Path,
        caveats: &Caveats,
        gate: &mut MockGate,
    ) -> String {
        execute_tool(
            name,
            &args,
            &ws.to_string_lossy(),
            false,
            20,
            caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None, // memory_source
            Some(gate),
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await
    }

    // -- #721 recoverable denials + request_permissions ---------------------

    #[test]
    fn exec_denial_is_recoverable_not_a_dead_end() {
        // #721 + #775: the exec denial the MODEL sees is ONE clean level —
        // `capability denied: <bare reason>. <recovery hint>` — leading to the
        // model-actionable request_permissions path, NOT the stale `extra_exec`
        // config edit (which #721 superseded and the model cannot perform
        // mid-turn).
        let envelope = serde_json::json!({
            "denied": true,
            "denials": [{
                "kind": "exec",
                "target": "mkdir",
                "reason": "exec of \"mkdir\" is not within the granted authority"
            }]
        });
        let out = denied_run_command_result(&envelope, false);
        assert!(out.starts_with("capability denied:"), "got: {out}");
        assert!(out.contains("request_permissions"), "got: {out}");
        // #775: the stale `extra_exec` config hint is GONE from the model-facing
        // message (it leaked in before; the human transcript carries the bare
        // command via print_denied instead).
        assert!(
            !out.contains("extra_exec"),
            "the model message must not carry the stale config hint: {out}"
        );
    }

    /// #775 (§2.5) regression: the model-facing `run_command` denial is ONE
    /// clean level and never a denial sentence NESTED inside another. Before
    /// the fix, `denied_run_command_result` appended the `extra_exec` config
    /// hint to the reason (and `print_denied` stuffed that whole sentence into
    /// its bare `'{target}'` slot), yielding `capability denied: exec does not
    /// permit '<reason> - add it via …>'`. The model-facing return now carries
    /// exactly one `capability denied:`, the bare reason, and the recovery hint.
    #[test]
    fn run_command_denial_is_single_level_not_nested() {
        let envelope = serde_json::json!({
            "denied": true,
            "denials": [{
                "kind": "exec",
                "target": "export",
                "reason": "exec of \"export\" is not within the granted authority"
            }]
        });
        let out = denied_run_command_result(&envelope, false);
        // Exactly one denial prefix — never a `capability denied:` inside another.
        assert_eq!(
            out.matches("capability denied:").count(),
            1,
            "exactly one denial level: {out}"
        );
        // RED on today: the stale config hint was glued onto the model message.
        assert!(!out.contains("add it via"), "stale config hint: {out}");
        assert!(!out.contains("extra_exec"), "stale config hint: {out}");
        // No reason sentence nested inside a `does not permit '…'` slot.
        assert!(
            !out.contains("does not permit 'exec of"),
            "nested denial sentence: {out}"
        );
        // The bare reason and the #721 recovery hint are both present.
        assert!(
            out.contains("exec of \"export\" is not within the granted authority"),
            "got: {out}"
        );
        assert!(out.contains("request_permissions"), "got: {out}");
    }

    #[test]
    fn parse_capability_maps_synonyms_and_rejects_unknown() {
        assert_eq!(parse_capability("exec"), Some(DenialKind::Exec));
        assert_eq!(parse_capability("shell"), Some(DenialKind::Exec));
        assert_eq!(parse_capability("FS_READ"), Some(DenialKind::FsRead));
        assert_eq!(parse_capability("write"), Some(DenialKind::FsWrite));
        assert_eq!(parse_capability("network"), Some(DenialKind::Net));
        assert_eq!(parse_capability("gpu"), None);
        assert_eq!(parse_capability(""), None);
    }

    #[test]
    fn request_permissions_grant_deny_and_no_gate() {
        let base = Caveats::top();

        // Mock gate ALLOWS → "granted" + the retry coaching; the gate was asked
        // with the parsed axis + target.
        let mut gate = MockGate::new(true, &base);
        let out = execute_request_permissions(
            &serde_json::json!({"capability": "exec", "target": "mkdir", "reason": "make a dir"}),
            Some(&mut gate),
            false,
            20,
        );
        assert!(out.starts_with("granted:"), "got: {out}");
        assert!(out.contains("Retry the original operation"), "got: {out}");
        assert_eq!(gate.asks.len(), 1);
        assert_eq!(
            gate.asks[0],
            ("request_permissions".to_string(), "exec:mkdir".to_string())
        );

        // Mock gate DENIES → "denied" + don't-retry coaching.
        let mut gate = MockGate::new(false, &base);
        let out = execute_request_permissions(
            &serde_json::json!({"capability": "fs_write", "target": "/tmp/x", "reason": "w"}),
            Some(&mut gate),
            false,
            20,
        );
        assert!(out.starts_with("denied:"), "got: {out}");
        assert!(out.contains("different approach"), "got: {out}");

        // NO gate (headless / eval) → "no operator available" — recoverable,
        // never a hang or a config-only dead end.
        let out = execute_request_permissions(
            &serde_json::json!({"capability": "net", "target": "docs.rs", "reason": "fetch"}),
            None,
            false,
            20,
        );
        assert!(out.contains("no operator available"), "got: {out}");
    }

    #[test]
    fn request_permissions_coaches_bad_inputs() {
        // Unknown capability → coach listing the valid axes (no gate consulted).
        let out = execute_request_permissions(
            &serde_json::json!({"capability": "gpu", "target": "x", "reason": "y"}),
            None,
            false,
            20,
        );
        assert!(out.contains("unknown capability"), "got: {out}");
        assert!(out.contains("fs_read"), "got: {out}");
        // Missing target → coach.
        let out = execute_request_permissions(
            &serde_json::json!({"capability": "exec", "reason": "y"}),
            None,
            false,
            20,
        );
        assert!(out.contains("'target' is required"), "got: {out}");
    }

    #[test]
    fn request_permissions_is_a_real_tool_not_a_phantom() {
        // #721: a real, always-advertised tool — never an alias / hallucination.
        assert!(resolve_tool_alias("request_permissions").is_none());
        assert!(ALL_TOOL_NAMES.contains(&"request_permissions"));
        assert!(classify_phantom_reach(
            "request_permissions",
            &serde_json::json!({"capability": "exec", "target": "mkdir", "reason": "r"}),
            "granted: the operator allowed exec for 'mkdir'.",
            true,
        )
        .is_none());
    }

    // -- #728 request_user_input (generic ask-the-human) --------------------

    /// A gate that answers a free-text question with a scripted answer (or None
    /// for "no human"). Its grant path (`ask`) is irrelevant here — it denies.
    struct AskGate {
        answer: Option<String>,
        asked: Vec<String>,
    }
    impl AskGate {
        fn new(answer: Option<&str>) -> Self {
            Self {
                answer: answer.map(str::to_string),
                asked: Vec::new(),
            }
        }
    }
    impl super::PermissionGate for AskGate {
        fn ask(&mut self, _requests: &[super::PermissionRequest]) -> super::PermissionDecision {
            super::PermissionDecision::Deny
        }
        fn ask_question(&mut self, question: &str) -> Option<String> {
            self.asked.push(question.to_string());
            self.answer.clone()
        }
    }

    #[test]
    fn request_user_input_returns_the_human_answer() {
        // A gate whose ask_question returns Some(answer) → the tool returns that
        // answer verbatim, and the gate was asked the exact question.
        let mut gate = AskGate::new(Some("postgres"));
        let out = execute_request_user_input(
            &serde_json::json!({"question": "which database should I target?"}),
            Some(&mut gate),
            false,
            20,
        );
        assert_eq!(out, "postgres");
        assert_eq!(
            gate.asked,
            vec!["which database should I target?".to_string()]
        );
    }

    #[test]
    fn request_user_input_no_gate_reports_headless_never_hangs() {
        // No gate (headless / eval / ACP) → the recoverable "no human available"
        // message — never a hang. (This test completing IS the no-hang proof: it
        // touches no real stdin.)
        let out = execute_request_user_input(
            &serde_json::json!({"question": "are you sure?"}),
            None,
            false,
            20,
        );
        assert_eq!(out, HEADLESS_NO_HUMAN);
        assert!(out.contains("no human available"), "got: {out}");
    }

    #[test]
    fn request_user_input_gate_with_no_human_reports_headless() {
        // A gate present but with no human to consult (ask_question → None) →
        // the SAME headless message, not a hang or an empty answer.
        let mut gate = AskGate::new(None);
        let out = execute_request_user_input(
            &serde_json::json!({"question": "pick one"}),
            Some(&mut gate),
            false,
            20,
        );
        assert_eq!(out, HEADLESS_NO_HUMAN);
    }

    #[test]
    fn request_user_input_requires_a_question() {
        // Missing / blank question → coach; the gate is never consulted.
        let mut gate = AskGate::new(Some("unused"));
        let out = execute_request_user_input(
            &serde_json::json!({"question": "   "}),
            Some(&mut gate),
            false,
            20,
        );
        assert!(out.contains("'question' is required"), "got: {out}");
        assert!(
            gate.asked.is_empty(),
            "gate not consulted for a blank question"
        );
    }

    #[test]
    fn request_user_input_is_a_real_tool_not_a_phantom() {
        // #728: a real, always-advertised tool — never an alias of itself or a
        // hallucination.
        assert!(resolve_tool_alias("request_user_input").is_none());
        assert!(ALL_TOOL_NAMES.contains(&"request_user_input"));
        assert!(classify_phantom_reach(
            "request_user_input",
            &serde_json::json!({"question": "which db?"}),
            "postgres",
            true,
        )
        .is_none());
        // The always-advertised def rides in every session (empty MCP).
        let defs = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        let names: Vec<&str> = defs
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|d| d["function"]["name"].as_str())
            .collect();
        assert!(names.contains(&"request_user_input"), "got: {names:?}");
    }

    #[test]
    fn ask_verbs_rewrite_to_request_user_input() {
        // #728: the instinctive ask-the-human verbs resolve to the real tool.
        for verb in [
            "ask_user",
            "ask_human",
            "prompt_user",
            "get_user_input",
            "ask_question",
            "clarify",
            "ask",
        ] {
            match resolve_tool_alias(verb) {
                Some(AliasOutcome::Rewrite(c)) => {
                    assert_eq!(c, "request_user_input", "verb: {verb}");
                }
                _ => panic!("expected Rewrite(request_user_input) for {verb}"),
            }
        }
    }

    #[tokio::test]
    async fn request_user_input_dispatches_through_execute_tool() {
        // End-to-end through the dispatcher: the question reaches the gate and
        // the answer flows back. Fully mocked (AskGate, no real stdin).
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = Caveats::top();
        let mut gate = AskGate::new(Some("the answer"));
        let out = execute_tool(
            "request_user_input",
            &serde_json::json!({"question": "what now?"}),
            &ws.path().to_string_lossy(),
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None, // memory_source
            Some(&mut gate),
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert_eq!(out, "the answer");
        assert_eq!(gate.asked, vec!["what now?".to_string()]);
    }

    #[test]
    fn get_context_remaining_is_a_real_tool_not_a_phantom() {
        // #727: real, always-advertised, no-arg budget read — never treated as
        // an alias of itself or a hallucination.
        assert!(resolve_tool_alias("get_context_remaining").is_none());
        assert!(ALL_TOOL_NAMES.contains(&"get_context_remaining"));
        assert!(classify_phantom_reach(
            "get_context_remaining",
            &serde_json::json!({}),
            "Context budget: ~10 tokens used of an input ceiling of ~80 (80% of num_ctx 100).",
            true,
        )
        .is_none());
        // The always-advertised def rides in every session (empty MCP).
        let defs = merged_tool_definitions(
            &NoMcp, false, false, false, false, false, false, false, false, false,
        );
        assert!(defs
            .as_array()
            .unwrap()
            .iter()
            .any(|d| d["function"]["name"] == "get_context_remaining"));
    }

    #[test]
    fn budget_verbs_rewrite_to_get_context_remaining() {
        // #727: the instinctive "how much context is left" reaches all resolve
        // to the canonical no-arg read (safe silent Rewrite — matching arg shape).
        for n in [
            "context_remaining",
            "tokens_left",
            "remaining_tokens",
            "budget",
            "how_much_context",
            "context_budget",
            "token_budget",
        ] {
            assert!(
                matches!(
                    resolve_tool_alias(n),
                    Some(AliasOutcome::Rewrite("get_context_remaining"))
                ),
                "{n} must rewrite to get_context_remaining"
            );
            // A Rewrite alias is mined by the #717 telemetry as a Rewrite.
            assert!(
                is_context_remaining_call(n),
                "{n} must be recognized as a budget call by the loop"
            );
        }
        // The canonical name is recognized by the loop but is NOT an alias.
        assert!(is_context_remaining_call("get_context_remaining"));
        assert!(resolve_tool_alias("get_context_remaining").is_none());
        // An unrelated name is neither.
        assert!(!is_context_remaining_call("read_file"));
    }

    /// FLAG OFF (no gate): the denial is deterministic and still DENIES every
    /// fs op (the #263 default-deny posture is intact) — now in the #721
    /// recoverable form (`denied_fs_result`, carrying the request_permissions
    /// path), pinned via the shared helper so the wording can't drift.
    #[tokio::test]
    async fn no_gate_denials_are_bit_for_bit_unchanged() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("secret.txt"), "x").unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            fs_write: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let out = run_tool(
            "read_file",
            serde_json::json!({"path": "secret.txt"}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_read", "secret.txt"));
        let out = run_tool(
            "list_dir",
            serde_json::json!({"path": "."}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_read", "."));
        let out = run_tool(
            "write_file",
            serde_json::json!({"path": "a.txt", "content": "c"}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_write", "a.txt"));
        let out = run_tool(
            "edit_file",
            serde_json::json!({"path": "a.txt", "old_string": "a", "new_string": "b"}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_write", "a.txt"));
        // #721: every fs denial now carries the model-actionable recovery path.
        assert!(out.contains("request_permissions"), "got: {out}");
    }

    /// Gate allows an fs_read denial → the read proceeds and returns the
    /// real contents; the gate was consulted with the tool + axis + full
    /// path it would be granting.
    #[tokio::test]
    async fn gate_allow_turns_fs_read_denial_into_the_real_result() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("secret.txt"), "the contents").unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let mut gate = MockGate::new(true, &denied);
        let out = run_tool_gated(
            "read_file",
            serde_json::json!({"path": "secret.txt"}),
            ws.path(),
            &denied,
            &mut gate,
        )
        .await;
        assert_eq!(out, "the contents");
        let full = ws.path().join("secret.txt").to_string_lossy().into_owned();
        assert_eq!(
            gate.asks,
            vec![("read_file".to_string(), format!("fs_read:{full}"))]
        );
    }

    /// Gate denies → the result is the standard denial, bit-for-bit equal to
    /// the no-gate path (#263: deny = the current denial result).
    #[tokio::test]
    async fn gate_deny_keeps_the_standard_denial_bit_for_bit() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("secret.txt"), "x").unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let mut gate = MockGate::new(false, &denied);
        let gated = run_tool_gated(
            "read_file",
            serde_json::json!({"path": "secret.txt"}),
            ws.path(),
            &denied,
            &mut gate,
        )
        .await;
        let ungated = run_tool(
            "read_file",
            serde_json::json!({"path": "secret.txt"}),
            ws.path(),
            &denied,
            None,
        )
        .await;
        assert_eq!(gated, ungated);
        assert_eq!(gated, denied_fs_result("fs_read", "secret.txt"));
        assert_eq!(gate.asks.len(), 1, "the human was asked exactly once");
    }

    /// Gate allows fs_write denials → write_file and edit_file proceed.
    #[tokio::test]
    async fn gate_allow_turns_fs_write_denials_into_real_writes() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("f.txt"), "old\n").unwrap();
        let denied = Caveats {
            fs_write: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let mut gate = MockGate::new(true, &denied);
        let out = run_tool_gated(
            "write_file",
            serde_json::json!({"path": "new.txt", "content": "fresh"}),
            ws.path(),
            &denied,
            &mut gate,
        )
        .await;
        assert!(out.starts_with("wrote new.txt"), "got: {out}");
        assert_eq!(
            std::fs::read_to_string(ws.path().join("new.txt")).unwrap(),
            "fresh"
        );
        let out = run_tool_gated(
            "edit_file",
            serde_json::json!({"path": "f.txt", "old_string": "old", "new_string": "new"}),
            ws.path(),
            &denied,
            &mut gate,
        )
        .await;
        assert!(out.starts_with("edited f.txt"), "got: {out}");
        assert_eq!(gate.asks.len(), 2);
        assert_eq!(gate.asks[0].0, "write_file");
        assert!(
            gate.asks[1].1.starts_with("fs_write:"),
            "got: {:?}",
            gate.asks[1]
        );
    }

    /// list_dir consults the gate on an fs_read denial like read_file does.
    #[tokio::test]
    async fn gate_allow_turns_list_dir_denial_into_the_listing() {
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("seen.txt"), "x").unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let mut gate = MockGate::new(true, &denied);
        let out = run_tool_gated(
            "list_dir",
            serde_json::json!({"path": "."}),
            ws.path(),
            &denied,
            &mut gate,
        )
        .await;
        assert!(out.contains("seen.txt"), "got: {out}");
    }

    /// A buggy/hostile gate answering Allow with caveats that STILL don't
    /// cover the path must not bypass enforcement: the widened authority is
    /// re-checked, never assumed (fs_gate_allows' re-check).
    #[tokio::test]
    async fn gate_allow_without_real_coverage_is_still_denied() {
        struct LyingGate;
        impl super::PermissionGate for LyingGate {
            fn ask(&mut self, _requests: &[super::PermissionRequest]) -> super::PermissionDecision {
                // "Allow", but the caveats grant nothing at all.
                super::PermissionDecision::Allow(Caveats {
                    fs_read: Scope::none(),
                    fs_write: Scope::none(),
                    exec: Scope::none(),
                    net: Scope::none(),
                    max_calls: CountBound::Unlimited,
                    valid_for_generation: Scope::All,
                })
            }
            fn ask_question(&mut self, _question: &str) -> Option<String> {
                None
            }
        }
        let ws = tempfile::TempDir::new().unwrap();
        std::fs::write(ws.path().join("secret.txt"), "x").unwrap();
        let denied = Caveats {
            fs_read: Scope::none(),
            ..caveats_rw(ws.path())
        };
        let mut gate = LyingGate;
        let out = execute_tool(
            "read_file",
            &serde_json::json!({"path": "secret.txt"}),
            &ws.path().to_string_lossy(),
            false,
            20,
            &denied,
            &mut NoMcp,
            None,
            None,
            None,
            None, // memory_source
            Some(&mut gate),
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_read", "secret.txt"));
    }

    /// web_fetch with a gate: an out-of-allowlist host consults the gate
    /// with the parsed host; on deny the dispatch runs under the ORIGINAL
    /// caveats, so the leash produces today's denial (an `error:` result —
    /// nothing is fetched).
    #[tokio::test]
    async fn web_fetch_gate_deny_dispatches_under_original_caveats() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path()); // net: Scope::none()
        let mut gate = MockGate::new(false, &caveats);
        let out = run_tool_gated(
            "web_fetch",
            serde_json::json!({"url": "https://denied.example.com:8443/page"}),
            ws.path(),
            &caveats,
            &mut gate,
        )
        .await;
        assert!(out.starts_with("error:"), "leash denial surfaces: {out}");
        assert_eq!(
            gate.asks,
            vec![(
                "web_fetch".to_string(),
                "net:denied.example.com".to_string()
            )]
        );
    }

    /// Regression for the field report: github.com is outside the default net
    /// scope, so a TUI-provided gate must be consulted before the bridle leash
    /// returns the denial to the model.
    #[tokio::test]
    async fn web_fetch_github_denial_consults_permission_gate() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path()); // net: Scope::none()
        let mut gate = MockGate::new(false, &caveats);
        let out = run_tool_gated(
            "web_fetch",
            serde_json::json!({"url": "https://github.com/openai/codex"}),
            ws.path(),
            &caveats,
            &mut gate,
        )
        .await;
        assert!(out.starts_with("error:"), "leash denial surfaces: {out}");
        assert_eq!(
            gate.asks,
            vec![("web_fetch".to_string(), "net:github.com".to_string())]
        );
    }

    /// An unparseable URL skips the net pre-check entirely — the gate is
    /// never consulted and the dispatch (with the original caveats) answers.
    #[tokio::test]
    async fn web_fetch_unparseable_url_never_prompts() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let mut gate = MockGate::new(true, &caveats);
        let out = run_tool_gated(
            "web_fetch",
            serde_json::json!({"url": "not-a-url"}),
            ws.path(),
            &caveats,
            &mut gate,
        )
        .await;
        assert!(out.starts_with("error:"), "got: {out}");
        assert!(gate.asks.is_empty(), "no prompt for an unparseable URL");
    }

    // -- save_note dispatch through execute_tool (Step 19.3) ----------------

    #[tokio::test]
    async fn save_note_without_sink_is_unknown_tool() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        // run_tool passes note_sink: None — the no-sink (headless) shape.
        let out = run_tool(
            "save_note",
            serde_json::json!({"action": "add", "text": "a fact"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.starts_with("unknown tool: save_note"), "got: {out}");
    }

    #[tokio::test]
    async fn save_note_with_sink_routes_through_execute_tool() {
        use crate::agentic::note_sink::tests::MockSink;
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let mut sink = MockSink::default();
        let out = execute_tool(
            "save_note",
            &serde_json::json!({"action": "add", "text": "workspace builds with just check"}),
            &ws.path().to_string_lossy(),
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            Some(&mut sink),
            None,
            None, // memory_source
            None,
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert_eq!(sink.calls, vec!["add:workspace builds with just check"]);
        assert!(
            out.starts_with("note saved: workspace builds"),
            "got: {out}"
        );
    }

    // -- recall dispatch through execute_tool (Step 17.5) -------------------

    #[tokio::test]
    async fn recall_without_source_is_unknown_tool() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        // run_tool passes recall_source: None — the no-store (headless) shape.
        let out = run_tool(
            "recall",
            serde_json::json!({"query": "tokio panic"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.starts_with("unknown tool: recall"), "got: {out}");
    }

    #[tokio::test]
    async fn recall_with_source_routes_through_execute_tool() {
        use crate::agentic::recall::tests::{hit, MockSource};
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let source = MockSource {
            hits: vec![hit(
                "123456789012-abcd",
                "past work",
                3,
                ">>>tokio<<< panic",
            )],
            ..Default::default()
        };
        let out = execute_tool(
            "recall",
            &serde_json::json!({"query": "tokio panic"}),
            &ws.path().to_string_lossy(),
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            Some(&source),
            None, // memory_source
            None,
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert_eq!(
            *source.calls.lock().unwrap(),
            vec![("tokio panic".to_string(), 5)]
        );
        assert!(out.contains("«tokio» panic"), "got: {out}");
        assert!(out.contains("past work"), "got: {out}");
    }

    // -- memory_fetch dispatch through execute_tool (#319) ------------------

    /// FLAG OFF (no source): a `memory_fetch` call is treated like any unknown
    /// tool — the inert-by-default shape (the tool was never advertised, so a
    /// call here is a hallucination). Mirrors `recall_without_source`.
    #[tokio::test]
    async fn memory_fetch_without_source_is_unknown_tool() {
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        // run_tool passes memory_source: None — the no-source (headless) shape.
        let out = run_tool(
            "memory_fetch",
            serde_json::json!({"address": "note:1"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert!(out.starts_with("unknown tool: memory_fetch"), "got: {out}");
    }

    /// FLAG ON (source present): a `memory_fetch` call routes through the
    /// injected `MemorySource` and returns its body. Mirrors
    /// `recall_with_source_routes_through_execute_tool`.
    #[tokio::test]
    async fn memory_fetch_with_source_routes_through_execute_tool() {
        use crate::agentic::memory_fetch::tests::MockSource;
        use crate::agentic::MemAddr;
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_rw(ws.path());
        let source = MockSource {
            body: Some("the exact note body".to_string()),
            ..Default::default()
        };
        let out = execute_tool(
            "memory_fetch",
            &serde_json::json!({"address": "note:1"}),
            &ws.path().to_string_lossy(),
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            Some(&source),
            None,
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert_eq!(out, "the exact note body");
        assert_eq!(
            *source.calls.lock().unwrap(),
            vec![MemAddr::Note { id: "1".into() }]
        );
    }
}

// ---------------------------------------------------------------------------
// INTERIM (#297) --disable-ocap / --yolo tests — the exec escape hatch.
// Removed with the bypass when brush upstreams CommandInterceptor
// (agent-bridle#20).
// ---------------------------------------------------------------------------

#[cfg(test)]
mod disable_ocap_tests {
    use super::super::NoMcp;
    use super::*;
    use crate::caveats::{Caveats, CountBound, Scope};
    use tokio::sync::{Mutex, MutexGuard};

    /// Serializes every test that reads or writes `NEWT_DISABLE_OCAP` (and
    /// the venv vars the bypass forwards): the process environment is shared
    /// across the parallel test runner. Async-aware (tokio) so the guard may
    /// be held across the `execute_tool` awaits; no poisoning — the `EnvVar`
    /// guards below restore the environment even on panic.
    static ENV_LOCK: Mutex<()> = Mutex::const_new(());

    async fn env_lock() -> MutexGuard<'static, ()> {
        ENV_LOCK.lock().await
    }

    /// RAII env override: set/unset `key` for the test body, restore the
    /// previous value on drop — including on a failed assertion, so yolo can
    /// never leak into a neighboring test.
    struct EnvVar {
        key: &'static str,
        saved: Option<String>,
    }

    impl EnvVar {
        fn set(key: &'static str, value: &str) -> Self {
            let saved = std::env::var(key).ok();
            std::env::set_var(key, value);
            Self { key, saved }
        }

        fn unset(key: &'static str) -> Self {
            let saved = std::env::var(key).ok();
            std::env::remove_var(key);
            Self { key, saved }
        }
    }

    impl Drop for EnvVar {
        fn drop(&mut self) {
            match self.saved.take() {
                Some(v) => std::env::set_var(self.key, v),
                None => std::env::remove_var(self.key),
            }
        }
    }

    /// Workspace-fenced fs, NO exec, NO net — the shape under which the
    /// confined shell denies (real build) or fails closed (stub build).
    fn caveats_no_exec(ws: &std::path::Path) -> Caveats {
        Caveats {
            fs_read: Scope::only([ws.to_string_lossy().into_owned()]),
            fs_write: Scope::only([ws.to_string_lossy().into_owned()]),
            exec: Scope::none(),
            net: Scope::none(),
            max_calls: CountBound::Unlimited,
            valid_for_generation: Scope::All,
        }
    }

    async fn run_tool(
        name: &str,
        args: serde_json::Value,
        ws: &std::path::Path,
        caveats: &Caveats,
    ) -> String {
        run_tool_with_floor(name, args, ws, caveats, None).await
    }

    /// #307: like [`run_tool`] but with an explicit exec FLOOR (the active
    /// named-permission-preset clamp). `Some(scope)` makes the `--disable-ocap`
    /// bypass conditional on the floor permitting the command; `None` is the
    /// pre-#307 behavior.
    async fn run_tool_with_floor(
        name: &str,
        args: serde_json::Value,
        ws: &std::path::Path,
        caveats: &Caveats,
        exec_floor: Option<&Scope<String>>,
    ) -> String {
        execute_tool(
            name,
            &args,
            &ws.to_string_lossy(),
            false,
            20,
            caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None, // memory_source
            None,
            exec_floor,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await
    }

    /// The switch reads fail-closed: ONLY the exact value `1` (the value the
    /// CLI exports and the issue documents) asserts the bypass. This is also
    /// the env-var-equivalence half of the #297 test list — the flag and the
    /// env var are one mechanism (`--disable-ocap` just exports the var).
    #[test]
    fn ocap_disabled_requires_exactly_1() {
        let _l = ENV_LOCK.blocking_lock();
        {
            let _unset = EnvVar::unset("NEWT_DISABLE_OCAP");
            assert!(!ocap_disabled(), "absent ⇒ confinement stays on");
        }
        for (value, expected) in [
            ("1", true),
            ("0", false),
            ("", false),
            ("true", false),
            ("yes", false),
            ("YOLO", false),
        ] {
            let _set = EnvVar::set("NEWT_DISABLE_OCAP", value);
            assert_eq!(
                ocap_disabled(),
                expected,
                "NEWT_DISABLE_OCAP={value:?} must read as {expected}"
            );
        }
    }

    /// Same fail-closed contract for the `--full-access` preset override:
    /// ONLY the exact value `1` asserts it (the flag and the env var are one
    /// mechanism — `--full-access` just exports the var).
    #[test]
    fn full_access_requested_requires_exactly_1() {
        let _l = ENV_LOCK.blocking_lock();
        {
            let _unset = EnvVar::unset("NEWT_FULL_ACCESS");
            assert!(!full_access_requested(), "absent ⇒ configured preset rules");
        }
        for (value, expected) in [
            ("1", true),
            ("0", false),
            ("", false),
            ("true", false),
            ("yes", false),
            ("FULL", false),
        ] {
            let _set = EnvVar::set("NEWT_FULL_ACCESS", value);
            assert_eq!(
                full_access_requested(),
                expected,
                "NEWT_FULL_ACCESS={value:?} must read as {expected}"
            );
        }
    }

    /// FLAG OFF ⇒ the command goes to the confined dispatch, which governs it.
    /// Built against the agent-bridle env-seam branch (#783), the bridle ships
    /// the REAL safe-subset shell (not the old fail-closed stub), so an
    /// ungranted `echo` under `exec = none` is DENIED by the L3 boundary. This
    /// is the "when the real shell returns" case the prior stub-build note
    /// anticipated; the "unavailable in this build" stub error is retired.
    #[tokio::test]
    async fn flag_off_run_command_keeps_the_confined_dispatch_verbatim() {
        let _l = env_lock().await;
        let _off = EnvVar::unset("NEWT_DISABLE_OCAP");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_tool(
            "run_command",
            serde_json::json!({"command": "echo hi"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert!(
            out.contains("capability denied"),
            "flag off ⇒ the confined dispatch must govern (deny) the command, got: {out}"
        );
    }

    /// FLAG ON: a command the confined shell fails closed on now runs on the
    /// host shell and returns its real output through the SAME envelope
    /// formatter (`shell_envelope_output`).
    #[cfg(unix)]
    #[tokio::test]
    async fn yolo_runs_the_denied_command_on_the_host_shell() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_tool(
            "run_command",
            serde_json::json!({"command": "echo yolo-ok"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert_eq!(out, "yolo-ok\n");

        // No output ⇒ the same `(exit N)` shape the bridle path produces.
        let out = run_tool(
            "run_command",
            serde_json::json!({"command": "exit 3"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert_eq!(out, "(exit 3)");
    }

    /// #726/#945: a verbose `run_command` MUST NOT flood the model's context
    /// window, but MUST still surface both ends of the output — a command's
    /// summary/failure/exit status lives at the TAIL, and #726's original
    /// head-only cap silently dropped exactly that (the gap #945 closed). Runs
    /// through the real host shell (yolo path) so it exercises the actual
    /// `shell_envelope_output` → cap composition. The global budget is the
    /// default 10k in the test binary (nothing raises it above default), so
    /// the assertions are upper-bounded and robust regardless of a smaller
    /// racing value. This test goes through the legacy `execute_tool` path
    /// (`run_tool`/`run_tool_with_floor` below), which has no spill store —
    /// the no-spill-id elision marker branch, not the `spill:<id>` one.
    #[cfg(unix)]
    #[tokio::test]
    async fn run_command_output_over_budget_is_token_capped() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        // ~350k chars of output — well over the default ~40k-char budget.
        let out = run_tool(
            "run_command",
            serde_json::json!({"command": "seq 1 60000"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert!(
            out.len() < 41_500,
            "model-facing output capped near the ~40k-char budget, got {} bytes",
            out.len()
        );
        assert!(
            out.contains("chars elided (head+tail shown"),
            "carries the head+tail elision marker: {:?}",
            &out[..out.len().min(400)]
        );
        // #945: the HEAD survives — the earliest lines are still visible.
        assert!(
            out.starts_with("1\n2\n3\n"),
            "head preserved: {:?}",
            &out[..out.len().min(160)]
        );
        // #945 (the regression this test now guards): the TAIL survives too —
        // under the old head-only cap this was the first assertion to break
        // (it asserted the OPPOSITE: `!out.contains("60000")`).
        assert!(
            out.trim_end().ends_with("60000"),
            "tail preserved, not dropped by the cap: {:?}",
            &out[out.len().saturating_sub(160)..]
        );
    }

    // --- #307 floor property: preset clamp WINS over --disable-ocap -------

    /// Unit-cover every branch of the bypass-floor predicate.
    #[test]
    fn exec_floor_permits_covers_each_branch() {
        use crate::caveats::Scope;
        // No floor ⇒ always permit (bit-for-bit pre-#307).
        assert!(exec_floor_permits(None, "rm -rf /"));
        // Empty command ⇒ let it through to the normal path.
        let only_echo = Scope::only(["echo".to_string()]);
        assert!(exec_floor_permits(Some(&only_echo), ""));
        // In-floor simple command ⇒ permitted.
        assert!(exec_floor_permits(Some(&only_echo), "echo hi"));
        // Out-of-floor program ⇒ refused.
        assert!(!exec_floor_permits(Some(&only_echo), "rm hi"));
        // Compound command ⇒ refused even with an allow-listed leading token.
        assert!(!exec_floor_permits(Some(&only_echo), "echo hi && rm x"));
        assert!(!exec_floor_permits(Some(&only_echo), "echo a | tee b"));
        assert!(!exec_floor_permits(Some(&only_echo), "echo $(rm x)"));
        // `Scope::All` floor permits any simple command.
        let all: Scope<String> = Scope::All;
        assert!(exec_floor_permits(Some(&all), "anything goes"));
        assert!(!exec_floor_permits(Some(&all), "anything; sneaky"));
    }

    /// ADVERSARIAL PROBE (review #312): exhaustively attack `exec_floor_permits`
    /// with EVERY shell injection / compound form so the floor is proven against
    /// more than just `&&`. An `echo`-only floor must refuse to bypass for any
    /// form that could chain or substitute a second program.
    #[test]
    fn exec_floor_refuses_every_metacharacter_form() {
        use crate::caveats::Scope;
        let echo = Scope::only(["echo".to_string()]);
        // Each of these begins with the allow-listed `echo` but smuggles or
        // could smuggle a second program. None may bypass.
        let attacks = [
            "echo ok && rm -rf /tmp/x", // && and
            "echo ok || rm -rf /tmp/x", // || or
            "echo ok ; rm -rf /tmp/x",  // ; sequence
            "echo ok | sh",             // | pipe
            "echo ok|sh",               // | no spaces
            "echo $(rm x)",             // $() command substitution
            "echo ${IFS}rm",            // ${} parameter expansion
            "echo `rm x`",              // backtick substitution
            "echo ok & rm x",           // & background
            "echo ok > /etc/passwd",    // > redirect out
            "echo ok >> /etc/passwd",   // >> append
            "echo < /etc/shadow",       // < redirect in
            "echo ok 2> err",           // 2> fd redirect (contains >)
            "(rm x)",                   // ( subshell
            "echo ok\nrm -rf /tmp/x",   // newline-separated
            "echo ok\nrm x\n",          // trailing newline
        ];
        for a in attacks {
            assert!(
                !exec_floor_permits(Some(&echo), a),
                "metacharacter form must NOT bypass the floor: {a:?}"
            );
        }
        // Forms with NO shell metacharacter that should still be refused because
        // the LEADING TOKEN is not the allow-listed program:
        let leading_token_attacks = [
            "rm -rf /tmp/x", // plain out-of-floor program
            "FOO=bar rm x",  // env-prefix: leading token `FOO=bar` ∉ floor
            "/bin/echo ok",  // path form: `/bin/echo` ≠ `echo` (exact match)
            "  rm x",        // leading whitespace, still `rm`
            "env rm x",      // `env` wrapper, leading token `env` ∉ floor
            "bash -c rm",    // `bash` ∉ floor
        ];
        for a in leading_token_attacks {
            assert!(
                !exec_floor_permits(Some(&echo), a),
                "out-of-floor leading token must be refused: {a:?}"
            );
        }
        // Sanity: a bare in-floor command with only a benign arg DOES bypass —
        // the floor is a ceiling, not a blanket off-switch. (A dangerous arg to
        // a permitted program is the user's accepted risk: they allow-listed it.)
        assert!(exec_floor_permits(Some(&echo), "echo hello world"));
        assert!(exec_floor_permits(Some(&echo), "echo -n trailing"));
    }

    /// FLOOR TEST (a) — the security contract: with `--disable-ocap` asserted,
    /// an exec FLOOR that denies the command must STOP the unconfined bypass.
    /// `echo` is outside a readonly floor (`exec = none`), so even with yolo on
    /// it does NOT run on the host shell — it falls through to the confined
    /// dispatch, which (env-seam real shell) DENIES it. A deliberately
    /// restricted triage mode is NOT un-clamped by `--yolo`.
    #[tokio::test]
    async fn floor_blocks_disable_ocap_for_a_denied_exec() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        // A readonly-triage preset clamp: exec denies everything.
        let floor = crate::NamedPermissionPreset {
            readonly: true,
            ..Default::default()
        }
        .clamp();
        let out = run_tool_with_floor(
            "run_command",
            serde_json::json!({"command": "echo should-not-run"}),
            ws.path(),
            &caveats,
            Some(&floor.exec),
        )
        .await;
        // The bypass did NOT fire: the command never reached the host shell, so
        // it fell to the confined dispatch and was denied, not `should-not-run\n`.
        assert_ne!(out, "should-not-run\n", "the floor must block the bypass");
        assert!(
            out.contains("capability denied"),
            "fell to confined dispatch and was denied, got: {out}"
        );
    }

    /// FLOOR TEST (a, positive) — a command INSIDE the floor still takes the
    /// fast unconfined path under `--disable-ocap`. The floor is a ceiling, not
    /// a blanket off-switch: an explicitly allow-listed command runs.
    #[cfg(unix)]
    #[tokio::test]
    async fn floor_allows_disable_ocap_for_an_in_floor_exec() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        // A triage preset that allow-lists `echo`.
        let floor = crate::NamedPermissionPreset {
            readonly: true,
            exec_allow: vec!["echo".to_string()],
            ..Default::default()
        }
        .clamp();
        let out = run_tool_with_floor(
            "run_command",
            serde_json::json!({"command": "echo in-floor-ok"}),
            ws.path(),
            &caveats,
            Some(&floor.exec),
        )
        .await;
        assert_eq!(out, "in-floor-ok\n", "in-floor command runs unconfined");
    }

    /// FLOOR conservatism — a COMPOUND command never bypasses under an active
    /// floor, even if its leading token is allow-listed: `echo ok && rm -rf /`
    /// must not smuggle `rm` past an `echo` grant. It falls to the confined
    /// shell (env-seam real shell ⇒ denied), which gates each spawn.
    #[tokio::test]
    async fn floor_refuses_bypass_for_a_compound_command() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        // `echo` is allow-listed, but the `&&` chains an unlisted `rm`.
        let floor = crate::NamedPermissionPreset {
            readonly: true,
            exec_allow: vec!["echo".to_string()],
            ..Default::default()
        }
        .clamp();
        let out = run_tool_with_floor(
            "run_command",
            serde_json::json!({"command": "echo ok && rm -rf /tmp/x"}),
            ws.path(),
            &caveats,
            Some(&floor.exec),
        )
        .await;
        assert_ne!(out, "ok\n", "a compound command must not bypass the floor");
        // Compound ⇒ never bypasses; it falls to the confined shell, which
        // (env-seam real shell) denies the ungranted command under `exec = none`.
        assert!(
            out.contains("capability denied"),
            "fell to confined dispatch and was denied, got: {out}"
        );
    }

    /// FLOOR TEST (c) — `None` floor is bit-for-bit the pre-#307 bypass: a
    /// denied-by-caveats command still runs unconfined under `--disable-ocap`,
    /// proving the floor is opt-in and the no-preset case is unchanged.
    #[cfg(unix)]
    #[tokio::test]
    async fn no_floor_keeps_disable_ocap_bit_for_bit() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_tool_with_floor(
            "run_command",
            serde_json::json!({"command": "echo no-floor-ok"}),
            ws.path(),
            &caveats,
            None,
        )
        .await;
        assert_eq!(out, "no-floor-ok\n", "no floor ⇒ bypass unchanged");
    }

    /// Envelope parity (#297): the host-shell envelope is structurally
    /// identical to the bridle one — `exit_code` / `stdout` / `stderr` /
    /// `sandbox_kind`, `denied`/`denials` omitted (⇒ not denied) — so the
    /// existing envelope readers apply unchanged.
    #[cfg(unix)]
    #[tokio::test]
    async fn host_shell_envelope_matches_the_bridle_shape() {
        let ws = tempfile::TempDir::new().unwrap();
        let envelope = host_shell_dispatch(
            "echo out; echo err >&2; exit 3",
            &ws.path().to_string_lossy(),
        )
        .await
        .expect("host shell runs");
        assert_eq!(envelope["exit_code"], 3);
        assert_eq!(envelope["stdout"], "out\n");
        assert_eq!(envelope["stderr"], "err\n");
        assert_eq!(envelope["sandbox_kind"], "none");
        // Omitted exactly as the bridle envelope omits them on the
        // nothing-was-denied path — `envelope_denied` reads it natively.
        assert!(envelope.get("denied").is_none(), "got: {envelope}");
        assert!(envelope.get("denials").is_none(), "got: {envelope}");
        assert!(!envelope_denied(&envelope));
        // And the shared formatter renders it like any confined result.
        assert_eq!(
            shell_envelope_output(&envelope, 20, false, false, None),
            "out\nerr\n"
        );
    }

    #[test]
    fn decode_shell_stream_preserves_valid_utf8() {
        let text = "// ── Model — test ──\n";
        assert_eq!(decode_shell_stream(text.as_bytes()), text);
    }

    #[test]
    fn decode_shell_stream_repairs_bsd_cat_v_utf8_notation() {
        // This is what macOS/BSD `cat -v` emits for "─ —\n" in a UTF-8
        // locale: the leading e2 byte is raw, while continuation bytes are
        // rendered as M-^T/M-^@ etc. A lossy decode would display
        // "�M-^TM-^@ �M-^@M-^T".
        let cat_v = b"\xe2M-^TM-^@ \xe2M-^@M-^T\n";
        assert_eq!(decode_shell_stream(cat_v), "─ —\n");
    }

    #[test]
    fn decode_shell_stream_repairs_two_byte_bsd_cat_v_notation() {
        // "é" is c3 a9; BSD `cat -v` leaves c3 raw and renders a9 as M-).
        let cat_v = b"caf\xc3M-)\n";
        assert_eq!(decode_shell_stream(cat_v), "café\n");
    }

    /// The venv/PATH prefix logic rides the HOST-BYPASS path unchanged: the
    /// `export VIRTUAL_ENV=…; export PATH=…;` prefix is prepended to the
    /// `--yolo` command, which runs on a real `/bin/sh` where `export` works.
    /// (The confined path no longer gets the prefix — it uses the env seam;
    /// see `confined_dispatch_uses_env_seam_not_export_prefix_783`.)
    #[cfg(unix)]
    #[tokio::test]
    async fn yolo_keeps_the_venv_prefix_logic() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let _venv = EnvVar::set("NEWT_VENV", "/opt/fake-venv");
        let _virtual = EnvVar::unset("VIRTUAL_ENV");
        let _paths = EnvVar::unset("NEWT_EXEC_PATHS");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_tool(
            "run_command",
            serde_json::json!({"command": "echo \"$VIRTUAL_ENV\""}),
            ws.path(),
            &caveats,
        )
        .await;
        assert_eq!(out, "/opt/fake-venv\n");
    }

    /// #783 regression (Bug A): the confined-shell dispatch carries the RAW
    /// user command and the venv via agent-bridle's structured `env` seam — NOT
    /// an `export …;` prefix on `cmd`. The old code built
    /// `{ "cmd": cmd_with_venv, "cwd": … }` (the prefixed form), and the
    /// confined safe-subset engine refuses an `export` builtin on a compound
    /// command, which is the bug. Pure: builds the dispatch args only, no spawn.
    #[tokio::test]
    async fn confined_dispatch_uses_env_seam_not_export_prefix_783() {
        let _l = env_lock().await;
        let _venv = EnvVar::set("NEWT_VENV", "/opt/fake-venv");
        let _virtual = EnvVar::unset("VIRTUAL_ENV");
        let _paths = EnvVar::unset("NEWT_EXEC_PATHS");

        // The literal failing case from #783.
        let cmd = "hostname; sw_vers 2>/dev/null | head -1; uname -s";
        let args = confined_dispatch_args(cmd, "/work/dir");

        // The command is passed RAW — no `export …;` prefix smuggled in.
        assert_eq!(args["cmd"], cmd);
        assert!(
            !args["cmd"]
                .as_str()
                .expect("cmd is a string")
                .contains("export "),
            "confined cmd must not carry an export prefix: {args}"
        );
        assert_eq!(args["cwd"], "/work/dir");

        // The venv rides the env seam: VIRTUAL_ENV + venv bin prepended to PATH.
        assert_eq!(args["env"]["VIRTUAL_ENV"], "/opt/fake-venv");
        let path = args["env"]["PATH"].as_str().expect("PATH in the env seam");
        assert!(
            path.starts_with("/opt/fake-venv/bin"),
            "venv bin must be prepended to PATH: {path}"
        );
    }

    /// #783: with neither venv input set, the env seam is empty (no spurious
    /// VIRTUAL_ENV / PATH keys) — the no-venv invocation is unaffected.
    #[tokio::test]
    async fn confined_dispatch_env_seam_empty_without_venv_783() {
        let _l = env_lock().await;
        let _venv = EnvVar::unset("NEWT_VENV");
        let _virtual = EnvVar::unset("VIRTUAL_ENV");
        let _paths = EnvVar::unset("NEWT_EXEC_PATHS");

        let args = confined_dispatch_args("ls -la", "/work/dir");
        assert_eq!(args["cmd"], "ls -la");
        assert_eq!(
            args["env"],
            serde_json::json!({}),
            "no venv inputs ⇒ empty env map: {args}"
        );
    }

    /// fs fence under yolo (#297): the newt-native workspace fence is NOT
    /// bypassed — a write/read outside the granted scope keeps the standard
    /// denial bit-for-bit. Yolo is unconfined exec, never authority-off.
    #[tokio::test]
    async fn yolo_keeps_the_fs_workspace_fence() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let escape = "/definitely-outside-the-fence/escape.txt";
        let out = run_tool(
            "write_file",
            serde_json::json!({"path": escape, "content": "nope"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_write", escape));
        assert!(!std::path::Path::new(escape).exists());

        let out = run_tool(
            "read_file",
            serde_json::json!({"path": "/etc/hostname"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert_eq!(out, denied_fs_result("fs_read", "/etc/hostname"));
    }

    /// Precedence (#297): with both `--disable-ocap` and a #263 gate present,
    /// exec never prompts — nothing is denied, so the gate is structurally
    /// unreachable for run_command. (fs prompting stays live; the fs-fence
    /// test above and the #263 suite cover that axis.)
    #[cfg(unix)]
    #[tokio::test]
    async fn yolo_never_consults_the_permission_gate_for_exec() {
        struct PanicGate;
        impl super::PermissionGate for PanicGate {
            fn ask(&mut self, requests: &[super::PermissionRequest]) -> super::PermissionDecision {
                panic!("yolo exec must never prompt, but the gate was asked: {requests:?}");
            }
            fn ask_question(&mut self, question: &str) -> Option<String> {
                panic!("yolo exec must never prompt, but the gate was asked: {question:?}");
            }
        }
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let mut gate = PanicGate;
        let out = execute_tool(
            "run_command",
            &serde_json::json!({"command": "echo no-prompt"}),
            &ws.path().to_string_lossy(),
            false,
            20,
            &caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None, // memory_source
            Some(&mut gate),
            None,
            None, // git_tool
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await;
        assert_eq!(out, "no-prompt\n");
    }

    /// The corrective tool-name guard still answers BEFORE the bypass: yolo
    /// changes where commands run, not what counts as a command.
    #[tokio::test]
    async fn yolo_keeps_the_tool_name_corrective_guard() {
        let _l = env_lock().await;
        let _on = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_tool(
            "run_command",
            serde_json::json!({"command": "read_file foo.txt"}),
            ws.path(),
            &caveats,
        )
        .await;
        assert!(out.contains("is a tool, not a shell command"), "got: {out}");
    }

    // --- facade P4 (#780): hidden tool-call routing dispatch ---------------

    /// A git stub that proves *which path served the call*: a routed
    /// `git status` lands here as op `status`; a routed write op would surface
    /// the unexpected-op error (so a test can assert it was NOT routed).
    struct RoutingStubGit;
    impl crate::agentic::GitTool for RoutingStubGit {
        fn dispatch(
            &self,
            op: &str,
            _args: &serde_json::Value,
            _caps: &crate::git_caveats::GitCaveats,
        ) -> Result<String, String> {
            match op {
                "status" => Ok("on branch main (routed via git built-in)".to_string()),
                other => Err(format!("unexpected routed git op '{other}'")),
            }
        }
    }

    async fn run_routed_with_git(command: &str, ws: &std::path::Path, caveats: &Caveats) -> String {
        execute_tool(
            "run_command",
            &serde_json::json!({ "command": command }),
            &ws.to_string_lossy(),
            false,
            20,
            caveats,
            &mut NoMcp,
            None,
            None,
            None,
            None, // memory_source
            None, // permission_gate
            None, // exec_floor
            Some(&RoutingStubGit as &dyn crate::agentic::GitTool),
            None, // crew_runner
            None, // scratchpad_store
            None, // code_search
            None, // experience_store
            None, // step_ledger
        )
        .await
    }

    /// The routing switch reads fail-closed (only the exact `1`), and it is a
    /// DISTINCT mechanism from `ocap_disabled` (§7-F5): asserting `NEWT_NO_ROUTE`
    /// never moves `ocap_disabled`, and asserting `NEWT_DISABLE_OCAP` never moves
    /// `routing_disabled`. The two switches can never alias.
    #[test]
    fn routing_disabled_requires_exactly_1_and_is_independent_of_ocap() {
        let _l = ENV_LOCK.blocking_lock();
        let _no_ocap = EnvVar::unset("NEWT_DISABLE_OCAP");
        {
            let _unset = EnvVar::unset("NEWT_NO_ROUTE");
            assert!(!routing_disabled(), "absent ⇒ routing stays on");
        }
        for (value, expected) in [("1", true), ("0", false), ("", false), ("true", false)] {
            let _set = EnvVar::set("NEWT_NO_ROUTE", value);
            assert_eq!(routing_disabled(), expected, "NEWT_NO_ROUTE={value:?}");
            // F5: turning routing off NEVER turns on the L3-off unconfine.
            assert!(
                !ocap_disabled(),
                "NEWT_NO_ROUTE must not imply --disable-ocap"
            );
        }
        // And the inverse: --disable-ocap must not imply --no-route.
        let _unset_route = EnvVar::unset("NEWT_NO_ROUTE");
        let _on_ocap = EnvVar::set("NEWT_DISABLE_OCAP", "1");
        assert!(ocap_disabled());
        assert!(
            !routing_disabled(),
            "--disable-ocap must not imply --no-route"
        );
    }

    /// TDD: a routed read goes through the SAME fs floor — routing is NOT a
    /// bypass. An out-of-scope `cat /etc/shadow` routes to `read_file` and is
    /// denied by `fs_read` exactly as a direct `read_file` would be (the denial
    /// short-circuits before any real fs access). The `fs_read` denial wording
    /// also proves it reached the `read_file` arm (vs. the exec/shell path).
    #[tokio::test]
    async fn routed_cat_goes_through_the_fs_floor_not_a_bypass() {
        let _l = env_lock().await;
        let _route_on = EnvVar::unset("NEWT_NO_ROUTE");
        let _ocap_off = EnvVar::unset("NEWT_DISABLE_OCAP");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path()); // fs_read scoped to ws only
        let out = run_tool(
            "run_command",
            serde_json::json!({ "command": "cat /etc/shadow" }),
            ws.path(),
            &caveats,
        )
        .await;
        assert!(
            out.contains("capability denied: fs_read does not permit")
                && out.contains("/etc/shadow"),
            "routed cat must hit the fs floor, not run unconfined; got: {out}"
        );
    }

    /// TDD: read-only `git status` is silently routed to the governed `git`
    /// built-in (the stub proves the built-in served it). Revert the routing
    /// promotion and this is red — the command would instead hit the run_command
    /// corrective guard.
    #[tokio::test]
    async fn routed_git_status_dispatches_through_the_git_builtin() {
        let _l = env_lock().await;
        let _route_on = EnvVar::unset("NEWT_NO_ROUTE");
        let _ocap_off = EnvVar::unset("NEWT_DISABLE_OCAP");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_routed_with_git("git status", ws.path(), &caveats).await;
        assert!(
            out.contains("routed via git built-in"),
            "git status must route to the governed git built-in; got: {out}"
        );
    }

    /// TDD: state-modifying `git add` is GATED as exec — NOT silently routed
    /// (owner decision 2). It never reaches the git built-in (no unexpected-op
    /// error from the stub); it falls through to the normal run_command path.
    #[tokio::test]
    async fn state_modifying_git_add_is_not_routed() {
        let _l = env_lock().await;
        let _route_on = EnvVar::unset("NEWT_NO_ROUTE");
        let _ocap_off = EnvVar::unset("NEWT_DISABLE_OCAP");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_routed_with_git("git add a.txt", ws.path(), &caveats).await;
        assert!(
            !out.contains("routed"),
            "git add must NOT route to the git built-in; got: {out}"
        );
        // It falls through to the run_command path (git ∈ DIRECT_TOOL_NAMES ⇒
        // the existing corrective guard), never silently routed.
        assert!(out.contains("is a tool, not a shell command"), "got: {out}");
    }

    /// F5 (§7-F5): `--no-route` bypasses routing but NEVER disables L3. With
    /// `NEWT_NO_ROUTE=1`, the same out-of-bounds `cat` is no longer routed to
    /// `read_file` (no `fs_read` denial), yet it does NOT run unconfined — it
    /// falls to the confined shell (env-seam real shell ⇒ denied), and
    /// `ocap_disabled()` stays false. The boundary holds.
    #[tokio::test]
    async fn no_route_bypasses_routing_but_keeps_l3() {
        let _l = env_lock().await;
        let _route_off = EnvVar::set("NEWT_NO_ROUTE", "1");
        let _ocap_off = EnvVar::unset("NEWT_DISABLE_OCAP");
        assert!(routing_disabled() && !ocap_disabled(), "L2 off, L3 on");
        let ws = tempfile::TempDir::new().unwrap();
        let caveats = caveats_no_exec(ws.path());
        let out = run_tool(
            "run_command",
            serde_json::json!({ "command": "cat /etc/shadow" }),
            ws.path(),
            &caveats,
        )
        .await;
        // Routing was OFF ⇒ NOT rewritten to read_file (no fs_read denial)…
        assert!(
            !out.contains("fs_read does not permit"),
            "--no-route must not route to read_file; got: {out}"
        );
        // …and the command did NOT run unconfined: it took the confined shell
        // (env-seam real shell ⇒ denied — the L3 boundary held).
        assert!(
            out.contains("capability denied"),
            "the L3 confined dispatch must still gate the command; got: {out}"
        );
    }
}