supercode_harness/tools/

mod.rs

//! Tools the agent can call.
//!
//! A [`Tool`] is a named capability with a JSON-Schema input and an async
//! `execute`. The [`ToolRegistry`] holds the set offered to a model; built-ins
//! cover file read/write/edit, directory listing, glob, content search, and
//! shell execution. Every tool can be disabled or re-described per
//! [`crate::Config`], so the capability surface is entirely yours to shape.

pub(crate) mod builtins;
pub mod clock;
pub mod context_budget;
pub mod convert;
pub mod image_gen;
pub mod plan_mode;
pub mod question;
mod skill;
pub(crate) mod tiers;

use std::collections::HashMap;
use std::collections::HashSet;
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};

use async_trait::async_trait;

use crate::config::Config;
use crate::error::Result;
use crate::modules::ModuleId;

pub use builtins::{
    ApplyPatchTool, BashTool, EditFileTool, GlobTool, ListDirTool, PersistentShellTool,
    ReadFileTool, SearchTool, UpdatePlanTool, ViewImageTool, WebFetchTool, WebSearchTool,
    WriteFileTool, DEFAULT_WEB_SEARCH_URL, WEB_CACHE_DIR_ENV, WEB_SEARCH_URL_ENV,
};
// BP-3 (§2 modules 6/8 + the catalog's clock, context-budget and image rows):
// the new core tools. Each is a plain `Tool` registered by
// `ToolRegistry::from_config` under its own preset gate, so `supercode
// harness parity`'s `tool` evidence resolves against the real registry.
pub use clock::{CurrentTimeTool, SleepTool, CURRENT_TIME, MAX_SLEEP_SECS, SLEEP};
pub use context_budget::{
    ContextBudget, GetContextRemainingTool, NewContextRequest, NewContextTool,
    GET_CONTEXT_REMAINING, NEW_CONTEXT,
};
pub use image_gen::{ImageGenTool, IMAGE_GEN};
pub use plan_mode::{
    EnterPlanModeTool, ExitPlanModeTool, PlanModeState, ENTER_PLAN_MODE, EXIT_PLAN_MODE,
};
pub use question::{
    AskUserTool, Question, QuestionOption, UserQuestionHandler, ASK_USER, REQUEST_USER_INPUT,
};
// P5-1 F4: `crate::agent`'s permissions gate needs to check an
// `apply_patch` envelope's write surface against `protected_paths` — not
// part of the crate's public tool-registration API, so `pub(crate)` rather
// than folded into the `pub use` list above.
pub(crate) use builtins::patch_target_paths;
// P5-6 (§2 module 4 `tools.background`): `crate::agent::Agent`'s
// `background_exec` intrinsic reuses `BashTool`'s own sandboxed-spawn
// builder rather than duplicating it — see that function's doc comment.
pub(crate) use builtins::build_sandboxed_sh;

pub use skill::{SkillTool, SKILL_TOOL};
pub use tiers::{minify as minify_tool_schema, SchemaTier};
// `SandboxPolicy` and `ToolContext` are defined below in this module.

/// P4c (COMPOSABLE-HARNESS-DESIGN.md S1.2/S3.1 `core.tools.read_file
/// multimodal`, S1.2 `view_image`): the sentinel prefix a tool's plain
/// `String` result carries when it is actually an image data URL rather
/// than ordinary text — `Agent::run_loop` detects this prefix (before
/// `cap_tool_output` ever sees it) and builds a `content_parts` image
/// block instead of a plain-text tool result. Using a control character
/// (`\u{1}`, SOH) as part of the marker keeps a false-positive collision
/// with real tool output astronomically unlikely without requiring a new
/// `Tool::execute` return type across all ten built-ins (an L-sized
/// trait-signature change this S-sized catalog item does not call for).
pub const MULTIMODAL_IMAGE_MARKER: &str = "\u{1}SUPERCODE_IMAGE_DATA_URL\u{1}";

/// P4c (S1.2 `core.tools.read_file.multimodal` / `view_image`): recognized
/// image file extensions (lowercase, no dot) — the same set CC/pi treat as
/// "images" for multimodal read (catalog D1 row 2's `✓*`/`✓*` variants).
pub const IMAGE_EXTENSIONS: &[&str] = &["png", "jpg", "jpeg", "gif", "webp", "bmp"];

/// Whether `path`'s extension is a recognized image type (case-insensitive).
pub fn is_image_path(path: &Path) -> bool {
    path.extension()
        .and_then(|e| e.to_str())
        .map(|e| IMAGE_EXTENSIONS.contains(&e.to_ascii_lowercase().as_str()))
        .unwrap_or(false)
}

/// The `image/<subtype>` MIME type for a recognized image extension, for
/// the `data:` URL — falls back to `png` for anything [`is_image_path`]
/// didn't already gate (defensive; never actually hit through
/// [`is_image_path`]'s own extension list).
pub fn image_mime_for(path: &Path) -> &'static str {
    match path
        .extension()
        .and_then(|e| e.to_str())
        .map(|e| e.to_ascii_lowercase())
        .as_deref()
    {
        Some("jpg") | Some("jpeg") => "image/jpeg",
        Some("gif") => "image/gif",
        Some("webp") => "image/webp",
        Some("bmp") => "image/bmp",
        _ => "image/png",
    }
}

/// P4c (S1.2 `core.tools.edit_file.notebook_aware`): the extension that
/// gates `EditFileTool`'s Jupyter cell-surgery branch.
pub const NOTEBOOK_EXTENSION: &str = "ipynb";

/// P4c (S2 module 5 `tools.web`, S2.1 dep "network sandbox rules", S17):
/// the network-domain policy a caller (SDK embedder) may install on a
/// [`ToolContext`] so [`crate::tools::WebFetchTool`]/[`crate::tools::WebSearchTool`]
/// respect it — see [`ToolContext::check_network`]. `None` on the context
/// (the default) means no policy is configured, matching today's honest
/// gap (no P5 `capabilities.permissions.sandbox.network` engine exists
/// yet, C3 — tracked, not hidden).
#[derive(Debug, Clone, Default)]
pub struct NetworkPolicy {
    /// Whether the policy is enforced at all. `false` behaves exactly like
    /// `None` on the context.
    pub enabled: bool,
    /// If non-empty, only these hosts are allowed. BP-10: matched as
    /// `crate::config::glob_match` patterns through the one rule engine
    /// (`domain(<entry>)`), so a bare hostname still matches exactly as
    /// before and `*.example.com` now works too.
    pub allow_domains: Vec<String>,
    /// These hosts are always denied, even if also present in
    /// `allow_domains`. Same pattern treatment as [`Self::allow_domains`].
    pub deny_domains: Vec<String>,
}

impl NetworkPolicy {
    /// BP-10 (catalog row "Allow/ask/deny rule language", the DOMAIN
    /// subject): this policy's two lists expressed IN the rule algebra —
    /// a [`crate::permissions::RuleSet`] of `domain(...)` patterns plus
    /// the baseline [`crate::permissions::Decision`] a host matching
    /// nothing gets.
    ///
    /// An allowlist is not a deny rule: in a deny→ask→allow FIRST-MATCH
    /// engine "only these hosts" is expressed by the BASELINE being
    /// `Deny`, with each allowed host in the `allow` tier — a
    /// `domain(*)` deny rule would (correctly, per tier priority) also
    /// swallow the allowlist. Empty `allow_domains` keeps the baseline
    /// `Allow`, which is why a pure denylist behaves exactly as it did
    /// before this translation existed.
    pub fn domain_rule_set(&self) -> (crate::permissions::RuleSet, crate::permissions::Decision) {
        let pattern = |d: &String| format!("domain({d})");
        let default = if self.allow_domains.is_empty() {
            crate::permissions::Decision::Allow
        } else {
            crate::permissions::Decision::Deny
        };
        (
            crate::permissions::RuleSet {
                deny: self.deny_domains.iter().map(pattern).collect(),
                ask: Vec::new(),
                allow: self.allow_domains.iter().map(pattern).collect(),
            },
            default,
        )
    }
}

/// BP-2 (catalog:32 "Edit refuses unless file was read (and unchanged)
/// this conversation"): how a path stands relative to the model's own most
/// recent read of it — see [`ToolContext::read_state`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ReadState {
    /// No read of this path has been recorded this conversation.
    NeverRead,
    /// Read, but the file's bytes have changed since (or it is no longer
    /// readable): the model's view is stale.
    Stale,
    /// Read, and the file is byte-identical to what the model saw.
    Fresh,
}

/// BP-2: the content stamp behind [`ReadState`] — blake3 (already a
/// dependency) truncated to 64 bits, which is a staleness detector, not a
/// security boundary: an adversary who can rewrite the file can rewrite the
/// edit too, so collision resistance beyond "different content looks
/// different" buys nothing here.
fn content_hash(bytes: &[u8]) -> u64 {
    let digest = blake3::hash(bytes);
    let b = digest.as_bytes();
    u64::from_le_bytes([b[0], b[1], b[2], b[3], b[4], b[5], b[6], b[7]])
}

/// Filesystem confinement applied to write-capable tools — the analog of
/// Codex's `read-only` / `workspace-write` / `danger-full-access` sandbox modes.
///
/// Enforced at the tool layer for file operations (`write_file`, `edit_file`,
/// `apply_patch`). Note: this confines the *file tools*; it does not OS-sandbox
/// arbitrary subprocesses (`bash`/`shell`) — true process isolation needs
/// platform primitives (seatbelt/landlock) and is a separate concern. Use
/// [`shell_sandbox_unenforceable`] as the runtime check for whether that gap
/// applies to the current platform and enabled tools.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum SandboxPolicy {
    /// No file writes are permitted by the file tools.
    ReadOnly,
    /// Writes are permitted only inside the working directory.
    WorkspaceWrite,
    /// No confinement (default — preserves prior behavior).
    #[default]
    DangerFullAccess,
}

/// P5-9 (design §2 module 20 `checkpoint`, §2.1 D-5 "write-path
/// interception seam shared with `formatters`"): the ONE well-defined
/// interception point around every file-mutating built-in tool
/// (`write_file`/`edit_file`/`apply_patch`) — installed on
/// [`ToolContext::write_observer`], `None` by default. Both hooks fire
/// AFTER [`ToolContext::check_write`] has already approved the call (so an
/// observer never sees a write the sandbox itself refused) and BEFORE/AFTER
/// the actual mutation:
/// - [`Self::before_write`] — pre-image capture. `crate::checkpoint`'s
///   [`crate::checkpoint::CheckpointObserver`] is the only implementation
///   today: it snapshots `path`'s current on-disk content (or records "did
///   not exist") so a later `checkpoint restore` can undo the write.
/// - [`Self::after_write`] — post-write. A true no-op in every
///   implementation shipped so far; reserved for `formatters` (P5-11,
///   design line 510 "shared seam with checkpoint") to run format-on-write
///   from, without needing a SECOND interception point wired through the
///   same three tools.
///
/// `None` (the default — `[capabilities.checkpoint]` off and no formatters
/// module yet) means neither hook is ever consulted: every write-tool
/// call-site's observer check is `if let Some(obs) = &ctx.write_observer`,
/// a branch that's simply never taken, so behavior is byte-identical to
/// before this seam existed.
///
/// P5-11 (§2 modules 28/29 `lsp`/`formatters`, C10): `async_trait` (rather
/// than the plain sync methods P5-9 originally shipped) because BOTH new
/// observers need real async I/O in `after_write` — `formatters` spawns and
/// awaits a subprocess, `lsp` writes/reads framed JSON-RPC over a child's
/// stdio — and neither can block the tokio runtime thread the way a
/// synchronous call from inside an already-`async fn execute()` would.
/// `CheckpointObserver`'s own hooks stay synchronous *internally* (plain
/// blocking `std::fs` calls); wrapping them in `async fn` changes nothing
/// observable for it, since that blocking work already ran on the calling
/// task before this signature changed. `after_write` now RETURNS
/// `Option<String>` — an annotation to append to the calling tool's result
/// string (formatter diff-back content, or LSP diagnostics) — `None` when
/// the observer has nothing to report, which is the only value
/// `CheckpointObserver::after_write` (still a no-op) ever returns, keeping
/// today's tool-result text byte-identical whenever checkpoint is the only
/// observer installed.
#[async_trait]
pub trait WriteObserver: Send + Sync + std::fmt::Debug {
    /// `path` (already resolved + sandbox-checked) is about to be
    /// created/overwritten/deleted. Implementations must be fast and must
    /// never propagate a failure as a tool error — a capture failure should
    /// degrade the OBSERVER (e.g. disable itself with a one-time warning),
    /// never block or fail the user's actual edit.
    async fn before_write(&self, path: &Path);
    /// `path` was just written/deleted successfully. Not called when the
    /// tool call itself failed (e.g. the write errored before completing).
    /// Returns an optional annotation for the calling tool's result text —
    /// see the trait doc comment above.
    async fn after_write(&self, path: &Path) -> Option<String>;
}

/// P5-11 (§2 modules 28/29, D-5 "shared write-path interception seam"): an
/// ORDERED chain of [`WriteObserver`]s installed as a single
/// `ToolContext::write_observer`, so the ONE seam P5-9 built keeps
/// supporting exactly one call site per tool while now composing multiple
/// concerns. Order is caller-determined (`crate::agent::build_tool_context`
/// builds it `checkpoint → formatters → lsp`, design's own required
/// ordering: checkpoint must capture the PRE-image before anything mutates
/// the file; formatters must run before lsp so diagnostics reflect the
/// FINAL, formatted file, not the model's pre-format draft).
/// `before_write` runs every observer in order; `after_write` runs every
/// observer in order too and joins any non-empty annotations with a blank
/// line, so a formatter's diff-back and an LSP diagnostics block can both
/// appear in one tool result without one silently discarding the other.
#[derive(Debug)]
pub struct WriteObserverChain(Vec<Arc<dyn WriteObserver>>);

impl WriteObserverChain {
    /// Build a chain that runs `observers` in order for both hooks.
    pub fn new(observers: Vec<Arc<dyn WriteObserver>>) -> Self {
        WriteObserverChain(observers)
    }
}

#[async_trait]
impl WriteObserver for WriteObserverChain {
    async fn before_write(&self, path: &Path) {
        for obs in &self.0 {
            obs.before_write(path).await;
        }
    }
    async fn after_write(&self, path: &Path) -> Option<String> {
        let mut notes: Vec<String> = Vec::new();
        for obs in &self.0 {
            if let Some(note) = obs.after_write(path).await {
                if !note.is_empty() {
                    notes.push(note);
                }
            }
        }
        if notes.is_empty() {
            None
        } else {
            Some(notes.join("\n\n"))
        }
    }
}

/// Ambient context passed to every tool invocation.
#[derive(Debug, Clone)]
pub struct ToolContext {
    /// The working directory tools resolve relative paths against.
    pub cwd: PathBuf,
    /// BP-10 (catalog row "Additional working directories", cc/cx
    /// `--add-dir`): extra roots granted BEYOND [`Self::cwd`], from
    /// `core.additional_dirs`/`--add-dir`. These are real grants, not
    /// discovery hints: [`Self::check_write`] treats a path under one of
    /// them as inside the workspace, the OS backstop adds each to the
    /// subprocess's writable set (`crate::sandbox::apply_linux_confinement`
    /// on Linux, the seatbelt profile on macOS), and the permissions
    /// engine's path rules are evaluated relative to each root as well as
    /// to `cwd` (so a `write(.git/**)` floor still covers an extra root's
    /// own `.git`). Empty (the default) is byte-identical to confining
    /// everything to `cwd` alone.
    pub extra_roots: Vec<PathBuf>,
    /// Filesystem confinement for write-capable tools.
    pub sandbox: SandboxPolicy,
    /// P4c (S1.2 `core.tools.read_file.multimodal`): whether `read_file`
    /// (and `view_image`, unconditionally) returns a recognized image file
    /// as a model-visible image content block. `false` (the default) is
    /// byte-identical to today's UTF-8-lossy-decode behavior.
    pub multimodal_read: bool,
    /// BP-2 (S1.2 `core.tools.read_file.line_numbers`, catalog:26): whether
    /// `read_file` prefixes every returned line with its 1-based file line
    /// number and a tab (`cat -n`), numbered from the requested `offset`.
    /// `false` (the default) returns the raw slice, as today.
    pub read_line_numbers: bool,
    /// P4c (S1.2 `core.tools.edit_file.require_read_before_edit`, UNIQUE CC
    /// row): whether `edit_file` refuses a path not yet read this
    /// conversation. `false` (the default) is byte-identical to today's
    /// behavior — [`Self::read_paths`] is simply never consulted.
    pub require_read_before_edit: bool,
    /// P4c: canonicalized paths `read_file` has successfully read so far
    /// this conversation — shared (via `Arc<Mutex<_>>`) across every clone
    /// of this context, since `Agent` constructs one `ToolContext` at
    /// startup and reuses it for every tool call. Consulted by `EditFileTool`
    /// only when [`Self::require_read_before_edit`] is `true`.
    ///
    /// BP-2 (catalog:32 "Edit refuses unless file was read **and
    /// unchanged** this conversation"): the value is the content hash AT
    /// READ TIME, so a file modified behind the model's back after its read
    /// is detected as STALE instead of editing cleanly against a view that
    /// no longer exists — see [`Self::read_state`].
    pub read_paths: Arc<Mutex<HashMap<PathBuf, u64>>>,
    /// P4c (S1.2 `core.tools.edit_file.notebook_aware`, UNIQUE CC row
    /// "NotebookEdit"): whether `edit_file` accepts Jupyter cell
    /// replace/insert/delete operations against a `.ipynb` target. `false`
    /// (the default) is byte-identical to today's exact-string-replace-only
    /// behavior.
    pub notebook_aware: bool,
    /// P4c (S1.2 `core.shell_env_snapshot`): the user's captured
    /// interactive-shell environment, if [`crate::Config::shell_env_snapshot`]
    /// is on — `BashTool`/`PersistentShellTool` merge this into the spawned
    /// process's environment. `None` (the default) is byte-identical to
    /// today's behavior: no extra environment is injected.
    pub shell_env: Option<Arc<HashMap<String, String>>>,
    /// P4c (S1.4 `core.nested_instructions`, deferred from P4b): whether a
    /// file-touching tool injects an as-yet-unseen subdirectory's own
    /// `CLAUDE.md`/`AGENTS.md` into its result the first time a path under
    /// it is touched. `false` (the default) is byte-identical to today's
    /// behavior.
    pub nested_instructions: bool,
    /// P4c: subdirectories (relative to [`Self::cwd`]) whose nested
    /// instructions have already been injected this conversation — shared
    /// across clones, same rationale as [`Self::read_paths`]. Consulted only
    /// when [`Self::nested_instructions`] is `true`.
    pub injected_instruction_dirs: Arc<Mutex<HashSet<PathBuf>>>,
    /// BP-5 (catalog D2 "Path-scoped rules", cc§2 `.claude/rules` `paths:`):
    /// the rule files this config loaded whose `paths:` selector holds them
    /// back until a matching file is touched. Empty (and inert) unless
    /// `[core.path_rules]` is on.
    pub path_rules: Arc<Vec<crate::path_rules::RuleFile>>,
    /// BP-5: which of [`Self::path_rules`] have already been injected this
    /// conversation — one injection per rule, the same de-duplication
    /// [`Self::injected_instruction_dirs`] gives nested instructions.
    pub injected_rule_files: Arc<Mutex<HashSet<PathBuf>>>,
    /// P4c (S2 module 5 `tools.web`, S17): the network-domain policy
    /// `web_fetch`/`web_search` must respect, if one is configured. `None`
    /// (the default) means no policy is enforced — see [`NetworkPolicy`]'s
    /// doc comment for the honest-gap rationale.
    pub network_policy: Option<NetworkPolicy>,
    /// BP-10 (catalog row "Allow/ask/deny rule language"): the
    /// CONFIG-DECLARED rule set (`capabilities.permissions.rules.*`, the
    /// same arrays `crate::agent::Agent`'s dispatch gate evaluates), so a
    /// `domain(...)` rule written there is enforced by the ONE engine at
    /// the network surface too — see [`Self::check_network`]. `None` (the
    /// default, and whenever `capabilities.permissions` is off) leaves the
    /// network check reading `network_policy`'s own two lists alone,
    /// byte-identical to before.
    pub permission_rules: Option<Arc<crate::permissions::RuleSet>>,
    /// P4e (S3.1 `core.tools.bash.timeout_secs`, S14): the DEFAULT
    /// execution timeout (seconds) `BashTool::execute` falls back to when a
    /// model-issued call carries no `timeout_ms` argument of its own -- see
    /// `crate::config::ToolOverride::timeout_secs`. `None` (the default) is
    /// byte-identical to today's behavior: `BashTool`'s built-in
    /// `DEFAULT_BASH_TIMEOUT_MS` (120s) stands.
    pub bash_timeout_secs: Option<u64>,
    /// P5-9 (§2 module 20, D-5 shared write-path interception seam) — see
    /// [`WriteObserver`]'s doc comment. `None` (the default) is a true
    /// no-op: every write-tool call site's `if let Some(obs) = ...` branch
    /// is simply never taken.
    pub write_observer: Option<Arc<dyn WriteObserver>>,
    /// P5-10 (§2 module 12 `permissions.sandbox`): whether the OS-level
    /// backstop (Landlock/seatbelt) is engaged for the `bash`/`shell`
    /// subprocess — see `crate::sandbox::os_sandbox_active`. `None` (the
    /// default) preserves the pre-P5-10 trigger (confine whenever
    /// [`Self::sandbox`] isn't [`SandboxPolicy::DangerFullAccess`]).
    pub sandbox_os_enabled: Option<bool>,
    /// P5-10 (§2 module 12, `escalation`): what to do when a confining fs
    /// tier can't actually be enforced on this platform/kernel — see
    /// `crate::sandbox::SandboxEscalation`. Defaults to `Deny`
    /// (fail-closed).
    pub sandbox_escalation: crate::sandbox::SandboxEscalation,
    /// P5-10 (§2 module 12, `env_policy`): child-process environment
    /// sanitization for the spawned subprocess — see
    /// `crate::sandbox::SandboxEnvPolicy`. Defaults to `Inherit`
    /// (byte-identical to pre-P5-10 behavior).
    pub sandbox_env_policy: crate::sandbox::SandboxEnvPolicy,
    /// P5-10 (§2 module 12, `escalation = "ask"` → `permissions.approvals`,
    /// P5-1): the ambient handler `crate::sandbox::decide_fs` consults for
    /// an `ask`-tier sandbox-unenforceable decision. `None` (the default —
    /// no handler installed) is fail-closed, same posture as the P5-1 rule
    /// engine's own `Ask` tier with no handler.
    pub sandbox_approval_handler: Option<crate::sandbox::SandboxApprovalHandler>,
    /// BP-3 (§2 module 6 `tools.question`): the door `ask_user` asks the
    /// human through — the SAME `elicitation/create` handler the design
    /// names as "the `tools.question` surface's PROTOCOL side"
    /// (`crate::mcp::McpElicitationHandler`), installed by
    /// `crate::agent::Agent::set_user_question_handler`. `None` (the
    /// default) means no interactive frontend is attached, and the tool
    /// says so rather than blocking on an answer nobody can give.
    pub question_handler: Option<question::UserQuestionHandler>,
    /// BP-3 (§2 module 8 `plan_mode`): the approval door `exit_plan_mode`
    /// presents the plan on — the same
    /// `crate::permissions::PermissionsApprovalHandler`
    /// `Agent::set_permissions_approval_handler` installs (under an
    /// SDK-owned runtime, that is the frontend request broker). `None` is
    /// fail-closed: the plan cannot be approved, so plan mode stays on.
    pub approval_handler: Option<ToolApprovalHandler>,
    /// BP-3 (§2 module 8): the shared plan-mode state — read by the agent's
    /// permission gate ([`plan_mode::deny_rules`]), written by
    /// `enter_plan_mode`/`exit_plan_mode` and the REPL's `/plan`. Inactive
    /// by default, and an inactive state contributes no rules at all.
    pub plan_mode: Arc<plan_mode::PlanModeState>,
    /// BP-3 (catalog row "Context-budget tools"): the shared token
    /// accounting `get_context_remaining` reads and `new_context` parks its
    /// request on. The agent publishes onto it; nothing is published until
    /// a turn has actually run.
    pub context_budget: Arc<context_budget::ContextBudget>,
    /// BP-8 (§2 module `todos` `persist`, catalog:156 "Todos/plan persisted
    /// per session"): the session's `update_plan` checklist. It lives HERE,
    /// on the context the agent owns and shares with every clone, rather
    /// than inside `UpdatePlanTool` — a plan the agent cannot read is a
    /// plan it cannot persist, which is exactly the residue the ledger row
    /// named. Empty by default, at zero cost.
    pub plan: Arc<Mutex<Vec<crate::session_journal::PlanEntry>>>,
}

/// BP-3: an approval handler reachable from inside a `Tool::execute`
/// (today, `exit_plan_mode`'s plan approval). A newtype purely so
/// [`ToolContext`] can stay `Debug` — the same shape, and the same reason,
/// as [`crate::sandbox::SandboxApprovalHandler`].
#[derive(Clone)]
pub struct ToolApprovalHandler(pub Arc<dyn crate::permissions::PermissionsApprovalHandler>);

impl std::fmt::Debug for ToolApprovalHandler {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_str("ToolApprovalHandler(..)")
    }
}

impl std::ops::Deref for ToolApprovalHandler {
    type Target = dyn crate::permissions::PermissionsApprovalHandler;
    fn deref(&self) -> &Self::Target {
        &*self.0
    }
}

impl ToolContext {
    /// A context rooted at `cwd` with no confinement.
    pub fn new(cwd: impl Into<PathBuf>) -> Self {
        ToolContext {
            cwd: cwd.into(),
            extra_roots: Vec::new(),
            sandbox: SandboxPolicy::DangerFullAccess,
            multimodal_read: false,
            read_line_numbers: false,
            require_read_before_edit: false,
            read_paths: Arc::new(Mutex::new(HashMap::new())),
            notebook_aware: false,
            shell_env: None,
            nested_instructions: false,
            injected_instruction_dirs: Arc::new(Mutex::new(HashSet::new())),
            path_rules: Arc::new(Vec::new()),
            injected_rule_files: Arc::new(Mutex::new(HashSet::new())),
            network_policy: None,
            permission_rules: None,
            bash_timeout_secs: None,
            write_observer: None,
            sandbox_os_enabled: None,
            sandbox_escalation: crate::sandbox::SandboxEscalation::default(),
            sandbox_env_policy: crate::sandbox::SandboxEnvPolicy::default(),
            sandbox_approval_handler: None,
            question_handler: None,
            approval_handler: None,
            plan_mode: Arc::new(plan_mode::PlanModeState::new()),
            context_budget: Arc::new(context_budget::ContextBudget::new()),
            plan: Arc::new(Mutex::new(Vec::new())),
        }
    }

    /// BP-8: the current plan, as `(step, status)` pairs.
    pub fn plan_snapshot(&self) -> Vec<crate::session_journal::PlanEntry> {
        self.plan.lock().map(|p| p.clone()).unwrap_or_default()
    }

    /// BP-8: replace the plan wholesale (`update_plan` replaces; a resume
    /// restores).
    pub fn set_plan(&self, steps: Vec<crate::session_journal::PlanEntry>) {
        if let Ok(mut p) = self.plan.lock() {
            *p = steps;
        }
    }

    /// BP-10: whether the permissions ENGINE adjudicated this call —
    /// i.e. `capabilities.permissions.enabled` was on when this context
    /// was built, so `crate::agent::Agent`'s dispatch gate ran the rule
    /// algebra (and any `Ask` tier) before the tool was invoked.
    ///
    /// The one consumer is the sandbox-escalation path
    /// (`builtins::escalation_requested`): a model-issued
    /// `with_escalated_permissions` is only honored where a gate exists to
    /// have approved it.
    pub fn permissions_engine_active(&self) -> bool {
        self.permission_rules.is_some()
    }

    /// P5-10: whether the OS-level backstop is active for this context —
    /// thin wrapper over `crate::sandbox::os_sandbox_active`.
    pub fn os_sandbox_active(&self) -> bool {
        crate::sandbox::os_sandbox_active(self.sandbox, self.sandbox_os_enabled)
    }

    /// P4c: record `path` (canonicalized if possible, else the resolved
    /// path as-is) as having been read this conversation — called by
    /// `ReadFileTool` on every successful read, unconditionally (cheap; the
    /// map is only ever CONSULTED when [`Self::require_read_before_edit`] is
    /// on, but recording it unconditionally means turning the knob on
    /// mid-conversation sees every read that already happened).
    ///
    /// BP-2: reads the file's CURRENT bytes to stamp the content hash. Use
    /// [`Self::mark_read_bytes`] from a caller that already holds them.
    pub fn mark_read(&self, path: &Path) {
        let bytes = std::fs::read(path).unwrap_or_default();
        self.mark_read_bytes(path, &bytes);
    }

    /// BP-2: the same record, stamped from bytes the caller just read.
    pub fn mark_read_bytes(&self, path: &Path, bytes: &[u8]) {
        let key = std::fs::canonicalize(path).unwrap_or_else(|_| path.to_path_buf());
        if let Ok(mut map) = self.read_paths.lock() {
            map.insert(key, content_hash(bytes));
        }
    }

    /// P4c: whether `path` was previously recorded via [`Self::mark_read`],
    /// ignoring whether it has changed since.
    pub fn was_read(&self, path: &Path) -> bool {
        !matches!(self.read_state(path), ReadState::NeverRead)
    }

    /// BP-2 (catalog:32): what `edit_file` needs to know before accepting
    /// an edit — never read, read but changed on disk since, or read and
    /// still identical.
    pub fn read_state(&self, path: &Path) -> ReadState {
        let key = std::fs::canonicalize(path).unwrap_or_else(|_| path.to_path_buf());
        let Some(recorded) = self
            .read_paths
            .lock()
            .ok()
            .and_then(|map| map.get(&key).copied())
        else {
            return ReadState::NeverRead;
        };
        match std::fs::read(&key) {
            Ok(bytes) if content_hash(&bytes) == recorded => ReadState::Fresh,
            // Unreadable now (deleted/permissions) counts as changed: the
            // model's view is provably not the file's current state.
            _ => ReadState::Stale,
        }
    }

    /// P4c (S2.1 S17): does `url` pass `Self::network_policy`, if one is
    /// configured? `Ok(())` when no policy is set (the honest-gap default)
    /// or the policy is present-but-disabled; `Err` names the reason
    /// otherwise. A URL with no parseable host is denied whenever a policy
    /// is actively enforced (fail closed — an unparseable host can't be
    /// matched against an allowlist).
    pub fn check_network(&self, url: &str) -> Result<()> {
        check_network_policy(
            self.network_policy.as_ref(),
            self.permission_rules.as_deref(),
            self.approval_handler.as_deref(),
            url,
        )
    }

    /// Resolve a possibly-relative path against the working directory.
    pub fn resolve(&self, path: &str) -> PathBuf {
        let p = PathBuf::from(path);
        if p.is_absolute() {
            p
        } else {
            self.cwd.join(p)
        }
    }

    /// BP-10: every root a `WorkspaceWrite` call may write under — `cwd`
    /// first, then each [`Self::extra_roots`] entry. The ONE list
    /// [`Self::check_write`], the seatbelt profile, and the Landlock
    /// writable set all read, so a grant can never be honored by one and
    /// missed by another.
    pub fn write_roots(&self) -> Vec<PathBuf> {
        let mut roots = Vec::with_capacity(1 + self.extra_roots.len());
        roots.push(self.cwd.clone());
        roots.extend(self.extra_roots.iter().cloned());
        roots
    }

    /// Enforce the sandbox policy for a write to `path`. `Err` if denied.
    pub fn check_write(&self, path: &Path) -> Result<()> {
        match self.sandbox {
            SandboxPolicy::DangerFullAccess => Ok(()),
            SandboxPolicy::ReadOnly => Err(crate::error::Error::tool(
                "sandbox",
                "write denied: sandbox is read-only",
            )),
            SandboxPolicy::WorkspaceWrite => {
                // BP-10: an `--add-dir` root is a real grant — a write
                // under one is inside the workspace, exactly as a write
                // under `cwd` is. Empty `extra_roots` (the default) makes
                // this the same single `cwd` check as before.
                if self.write_roots().iter().any(|r| path_within(r, path)) {
                    Ok(())
                } else {
                    Err(crate::error::Error::tool(
                        "sandbox",
                        format!(
                            "write denied: {} is outside the workspace {} (and its {} \
                             additional root(s))",
                            path.display(),
                            self.cwd.display(),
                            self.extra_roots.len()
                        ),
                    ))
                }
            }
        }
    }
}

/// P5-2 (§2 module 15, security note "remote MCP over http/sse: respect the
/// NetworkPolicy from P5-1 if one is active"): the same policy-and-url check
/// [`ToolContext::check_network`] performs, factored out to a free function
/// so `crate::mcp::McpClient::connect_http`/`connect_sse` can enforce the
/// identical allow/deny/SSRF floor a `web_fetch` call would get — one
/// enforcement point, not a second parallel one that could silently drift
/// from it.
pub(crate) fn check_network_policy(
    policy: Option<&NetworkPolicy>,
    rules: Option<&crate::permissions::RuleSet>,
    approval: Option<&dyn crate::permissions::PermissionsApprovalHandler>,
    url: &str,
) -> Result<()> {
    let ctx_like = domain_tier_of(policy, rules);
    check_host_against_tier(&ctx_like, approval, url_host(url).as_deref())
}

/// BP-10: [`ToolContext::domain_tier`]'s body, as a free function, so the
/// non-`ToolContext` caller (`crate::mcp::McpClient::connect_http`) folds
/// the SAME two sources in the SAME order rather than a second, drifting
/// copy. See that method's doc comment for the two sources.
pub(crate) fn domain_tier_of(
    policy: Option<&NetworkPolicy>,
    rules: Option<&crate::permissions::RuleSet>,
) -> (crate::permissions::RuleSet, crate::permissions::Decision) {
    let mut out = crate::permissions::RuleSet::default();
    let mut default = crate::permissions::Decision::Allow;
    if let Some(policy) = policy {
        if policy.enabled {
            let (list_rules, list_default) = policy.domain_rule_set();
            out.deny.extend(list_rules.deny);
            out.allow.extend(list_rules.allow);
            default = list_default;
        }
    }
    if let Some(config_rules) = rules {
        out.deny.extend(config_rules.deny.iter().cloned());
        out.ask.extend(config_rules.ask.iter().cloned());
        out.allow.extend(config_rules.allow.iter().cloned());
    }
    (out, default)
}

/// P4c-review (MEDIUM/LOW follow-up, dep 8's neighboring `tools.web` SSRF
/// gap): the SAME allow/deny decision [`ToolContext::check_network`] applies
/// to the INITIAL url, factored out so [`network_checked_redirect_policy`]
/// can apply it to every REDIRECT hop too. Without this, `check_network`
/// validated only the url the caller passed in — once a real network policy
/// is wired up (P5), a denied host reachable only via an allowed host's HTTP
/// redirect (reqwest follows up to 10 by default) bypassed the check
/// entirely. `host: None` (unparseable/absent) fails closed, exactly like
/// `check_network`'s own prior inline behavior.
fn check_host_against_tier(
    tier: &(crate::permissions::RuleSet, crate::permissions::Decision),
    approval: Option<&dyn crate::permissions::PermissionsApprovalHandler>,
    host: Option<&str>,
) -> Result<()> {
    use crate::permissions::{Decision, RuleSet};
    let (rules, default): (&RuleSet, Decision) = (&tier.0, tier.1);
    // Nothing to enforce: no domain rule from either source. Byte-identical
    // to "no policy configured" — this is the common path.
    if rules.is_empty() && default == Decision::Allow {
        return Ok(());
    }
    let Some(host_str) = host else {
        return Err(crate::error::Error::tool(
            "network",
            "cannot determine host from url; denied under an active network policy",
        ));
    };
    let host = host_str.to_ascii_lowercase();
    match crate::permissions::evaluate_domain(rules, Some(&host), default) {
        Decision::Allow => Ok(()),
        Decision::Ask => {
            // BP-10: the `domain(...)` ASK tier resolves on the SAME door
            // every other `Ask` in this engine uses. No door installed
            // denies, the fail-closed posture
            // `PermissionsApprovalHandler`'s own doc comment documents.
            let raw_args = serde_json::json!({ "host": host });
            let req = crate::permissions::ApprovalRequest {
                tool: "domain",
                subject: Some(&host),
                raw_args: &raw_args,
            };
            match approval.map(|h| h.ask(&req)) {
                Some(crate::permissions::ApprovalOutcome::Allow)
                | Some(crate::permissions::ApprovalOutcome::AllowForSession) => Ok(()),
                _ => Err(crate::error::Error::tool(
                    "network",
                    format!("host `{host}` requires approval and none was given"),
                )),
            }
        }
        Decision::Deny => {
            if crate::permissions::domain_denied_explicitly(rules, &host) {
                Err(crate::error::Error::tool(
                    "network",
                    format!("host `{host}` is denied by the active network policy"),
                ))
            } else {
                Err(crate::error::Error::tool(
                    "network",
                    format!("host `{host}` is not on the network policy's allowlist"),
                ))
            }
        }
    }
}

/// P4c-review (MEDIUM/LOW follow-up): a `reqwest::redirect::Policy` for
/// `WebFetchTool`/`WebSearchTool`'s client that re-runs
/// [`check_host_against_policy`] (the exact same check
/// [`ToolContext::check_network`] applies to the initial url) against every
/// redirect hop's target host, refusing to follow one that a network policy
/// denies. `policy: None` (no policy configured) or a present-but-disabled
/// one behaves like reqwest's own default policy — follow, capped at the
/// same 10-hop limit `redirect::Policy::default()` uses (the crate's `custom`
/// variant does NOT enforce a redirect cap on its own — see its doc comment
/// — so this reimplements that cap by hand).
pub(crate) fn network_checked_redirect_policy(
    policy: Option<NetworkPolicy>,
    rules: Option<Arc<crate::permissions::RuleSet>>,
) -> reqwest::redirect::Policy {
    const MAX_REDIRECTS: usize = 10; // matches reqwest::redirect::Policy::default()
    let tier = domain_tier_of(policy.as_ref(), rules.as_deref());
    reqwest::redirect::Policy::custom(move |attempt| {
        if attempt.previous().len() >= MAX_REDIRECTS {
            return attempt.error("too many redirects");
        }
        // BP-10: a redirect hop gets the SAME domain tier as the initial
        // url, but never an interactive prompt — an `Ask` mid-flight has
        // no user-visible action to describe, so it fails closed here
        // (`approval: None`) rather than blocking a redirect chain on a
        // question about a host the user never typed.
        if let Err(e) = check_host_against_tier(&tier, None, attempt.url().host_str()) {
            return attempt.error(e.to_string());
        }
        attempt.follow()
    })
}

/// P4c (S2.1 S17): extract the host from an `http(s)://` URL — the smallest
/// parser that satisfies [`ToolContext::check_network`]'s needs without a
/// new `url`-crate dependency (matches this crate's existing `glob_match`
/// precedent of hand-rolling a small parser rather than reaching for a
/// dependency for an S-sized need). Returns `None` for anything that isn't
/// `http://`/`https://` or has an empty host component.
fn url_host(url: &str) -> Option<String> {
    let rest = url
        .strip_prefix("https://")
        .or_else(|| url.strip_prefix("http://"))?;
    let end = rest.find(['/', '?', '#']).unwrap_or(rest.len());
    let authority = &rest[..end];
    // Strip a `user:pass@` prefix and a `:port` suffix, keeping the host.
    let host_and_port = authority.rsplit('@').next().unwrap_or(authority);
    let host = host_and_port.split(':').next().unwrap_or(host_and_port);
    if host.is_empty() {
        None
    } else {
        Some(host.to_ascii_lowercase())
    }
}

/// True when the requested sandbox policy cannot be enforced for shell
/// subprocesses: a confining policy, a platform without an OS sandbox
/// primitive wired up (only macOS/seatbelt is, via `sandbox-exec`), and at
/// least one shell tool (`"bash"` or `"shell"`) enabled.
///
/// This is a pure function so it's mechanically testable on any host OS:
/// callers pass the platform (typically `std::env::consts::OS`) and the set
/// of enabled tool names rather than relying on `cfg!`/`target_os`. It does
/// not itself sandbox anything — it only tells embedders/CLIs whether the
/// gap documented on [`SandboxPolicy`] applies right now, so they can warn.
/// P5-10 (§2 module 12): `landlock_available` is the caller's REAL Linux
/// Landlock-availability probe (`crate::sandbox::landlock_available()`,
/// typically) — a PARAMETER, not an internal `cfg!`/probe call, same "pure,
/// mechanically testable" contract this function already had. Before
/// P5-10, `platform == "linux"` always meant "unenforceable" (no OS
/// primitive existed yet); now it means "unenforceable UNLESS Landlock is
/// actually available on this kernel" — a confining tier on a
/// Landlock-capable Linux box is REAL enforcement, not a gap, so this must
/// say `false` for it (never claim a gap that no longer exists).
/// `platform == "macos"` is unconditionally `false` regardless of this
/// parameter (seatbelt, a separate primitive, always exists there); every
/// other platform (including `platform == "linux"` with
/// `landlock_available == false`) is unaffected by this parameter and
/// keeps the pre-P5-10 "no primitive" answer.
pub fn shell_sandbox_unenforceable(
    policy: SandboxPolicy,
    platform: &str,
    tools_enabled: &[&str],
    landlock_available: bool,
) -> bool {
    policy != SandboxPolicy::DangerFullAccess
        && platform != "macos"
        && !(platform == "linux" && landlock_available)
        && tools_enabled.iter().any(|t| *t == "bash" || *t == "shell")
}

/// Whether `path` is inside `root`. SECURITY (safe-path consolidation,
/// LOWER-URGENCY fix folded into the permissions-gate CRITICAL fix): this
/// used to compare only LEXICALLY-normalized paths (`..` traversal caught,
/// but a pre-existing in-workspace symlink pointing outside `root` was NOT —
/// `link -> /etc` plus a write to `link/passwd` lexically normalizes to
/// `<root>/link/passwd`, which "starts with" `root` even though it actually
/// resolves outside it). Now delegates to `crate::safe_path::contained`,
/// which ALSO resolves symlinks along the longest existing ancestor (the
/// same proven dual lexical+resolved check `crate::checkpoint`'s P5-9 fix
/// uses), so a symlink escape is caught here too. A non-existent target
/// (e.g. a file about to be created) is still handled correctly.
fn path_within(root: &Path, path: &Path) -> bool {
    crate::safe_path::contained(root, path)
}

/// `pub(crate)`: also the lexical-`..`-collapse step
/// [`crate::checkpoint`]'s containment check builds on (P5-9) — one
/// normalizer, not a second hand-rolled one.
pub(crate) fn normalize(path: &Path) -> Option<PathBuf> {
    use std::path::Component;
    // Make absolute against CWD if needed (paths are already joined to cwd by
    // resolve(), but be defensive).
    let abs = if path.is_absolute() {
        path.to_path_buf()
    } else {
        std::env::current_dir().ok()?.join(path)
    };
    let mut out = PathBuf::new();
    for c in abs.components() {
        match c {
            Component::ParentDir => {
                out.pop();
            }
            Component::CurDir => {}
            other => out.push(other.as_os_str()),
        }
    }
    Some(out)
}

/// A callable capability.
#[async_trait]
pub trait Tool: Send + Sync {
    /// Stable, unique tool name (what the model calls).
    fn name(&self) -> &str;

    /// The built-in description. May be overridden via [`crate::Config`].
    fn description(&self) -> &str;

    /// JSON Schema describing the tool's input object.
    fn parameters(&self) -> serde_json::Value;

    /// Whether a successful textual result is also a complete JSON value
    /// that protocol adapters should expose as structured output. Text
    /// remains the model-facing representation, preserving compatibility.
    fn structured_output(&self) -> bool {
        false
    }

    /// Run the tool. Returns text to feed back to the model.
    async fn execute(&self, args: serde_json::Value, ctx: &ToolContext) -> Result<String>;
}

/// An ordered set of tools offered to the model.
#[derive(Default)]
pub struct ToolRegistry {
    tools: Vec<Box<dyn Tool>>,
}

impl ToolRegistry {
    /// An empty registry.
    pub fn new() -> Self {
        ToolRegistry::default()
    }

    /// A registry pre-populated with all built-in tools.
    pub fn with_builtins() -> Self {
        let mut r = ToolRegistry::new();
        r.register(ReadFileTool);
        r.register(WriteFileTool);
        r.register(EditFileTool);
        r.register(ListDirTool);
        r.register(GlobTool);
        r.register(SearchTool);
        r.register(ApplyPatchTool);
        r.register(BashTool::default());
        r.register(PersistentShellTool::default());
        r.register(UpdatePlanTool::default());
        r
    }

    /// P3 (COMPOSABLE-HARNESS-DESIGN.md §5.2 phase P3): build a registry
    /// from a resolved [`Config`]'s module-activation set, the intended
    /// replacement for unconditional [`Self::with_builtins`] call sites.
    ///
    /// **BP-1: this is now the default path.** Every `Config` materialized
    /// by [`crate::configfile::resolve`] carries
    /// [`Config::module_registry`] `= true`, so a preset's
    /// `[capabilities.*]` and `[core.tools] enabled` actually shape the
    /// registry. `[experimental] module_registry = false` is the explicit
    /// OPT-OUT that pins a resolved config back to the unfiltered stack,
    /// and a hand-built [`Config::default`] (which never went through the
    /// resolver) still has the flag `false`. In that `false` state this
    /// returns EXACTLY [`Self::with_builtins`] — same 10 tools, same
    /// order, zero behavior change. When it is on,
    /// [`Config::module_activation`]/[`Config::core_tools_enabled`]
    /// shape which tool objects get registered AT ALL: a disabled module
    /// contributes no tool (never registered, so never advertised and never
    /// mentioned anywhere) — e.g. `todos` off means `update_plan` is not in
    /// this registry; `tools_search` off (or its `list_dir`/`glob`/
    /// `content_search` sub-flags off) means the corresponding tool is
    /// absent too.
    pub fn from_config(config: &Config) -> Self {
        // The opt-out (or a Config that never met the resolver at all).
        if !config.module_registry {
            return Self::with_builtins();
        }
        let mut r = ToolRegistry::new();
        let act = &config.module_activation;

        // BP-13 (catalog D9 "Per-model capability bits drive tools", cx§9
        // "Model catalog"): the model's own capability bits, resolved out of
        // the SAME `Config::model_routing` table every other routing
        // decision reads. They shape THIS selection rather than a parallel
        // registry — the only thing they can do is decide which write
        // surface a model is offered.
        //
        // Armed only under `[capabilities.tools_apply_patch] per_model =
        // true` (the flag cx-parity already sets, and whose only previous
        // reader was the resolver's C1 warning suppression) AND only when a
        // rule actually matches this model. With no matching rule the
        // selection below is byte-identical to pre-BP-13: `core.tools`
        // decides edit/write, the module decides apply_patch.
        let bits = if act.is_active(ModuleId::ToolsApplyPatch) && act.tools_apply_patch_per_model {
            config.model_routing.rules_for(&config.model)
        } else {
            crate::model_catalog::ModelRules::default()
        };
        let core_has = |name: &str| match (name, bits.apply_patch) {
            // A model the catalog marks as NOT taking the freeform
            // apply_patch envelope gets the edit/write pair instead, even
            // where `core.tools` lists neither — that swap IS the row.
            ("edit_file" | "write_file", Some(false)) => true,
            // …and the converse: a model that DOES take apply_patch is not
            // also handed the pair, so the two write formats are never
            // co-advertised to it (§2.2 C1's whole point).
            ("edit_file" | "write_file", Some(true)) => false,
            _ => config.core_tools_enabled.iter().any(|t| t == name),
        };

        // Same relative order as `with_builtins()` for everything both paths
        // can register, so a partial activation set stays predictable.
        if core_has("read_file") {
            r.register(ReadFileTool);
        }
        if core_has("write_file") {
            r.register(WriteFileTool);
        }
        if core_has("edit_file") {
            r.register(EditFileTool);
        }
        // P4c (S1.2 `view_image`, S12): a fifth OPTIONAL default-tool name —
        // "recognized alongside read_file/bash/edit_file/write_file as a
        // fifth optional default-tool name, not a new module" — so it's
        // read from the SAME `core_tools_enabled` list as the other four,
        // not a `ModuleId`. Absent from the list by default (today's
        // `["read_file","bash","edit_file","write_file"]` default), so this
        // is a no-op unless a caller explicitly adds `"view_image"`.
        if core_has("view_image") {
            r.register(ViewImageTool);
        }
        // BP-13: `search_tool = false` withdraws the dedicated search tool
        // for a model the catalog says cannot use it (cx§9
        // `supports_search_tool`); unset leaves the module's own sub-flags
        // in sole charge, exactly as before.
        if act.is_active(ModuleId::ToolsSearch) && bits.search_tool != Some(false) {
            if act.tools_search_list_dir {
                r.register(ListDirTool);
            }
            if act.tools_search_glob {
                r.register(GlobTool);
            }
            if act.tools_search_content_search {
                r.register(SearchTool);
            }
        }
        // BP-13: with per-model bits armed, a model the catalog says cannot
        // take the freeform envelope is not offered it.
        if act.is_active(ModuleId::ToolsApplyPatch) && bits.apply_patch != Some(false) {
            r.register(ApplyPatchTool);
        }
        if core_has("bash") {
            r.register(BashTool::default());
        }
        if act.is_active(ModuleId::ToolsPersistentShell) {
            r.register(PersistentShellTool::default());
        }
        if act.is_active(ModuleId::Todos) {
            r.register(UpdatePlanTool::default());
        }
        // P4c (S2 module 5 `tools.web`, S4a "trivially addable"): single
        // tool each, gated by the module's own `fetch`/`search` sub-flags
        // (S3.1: `[capabilities.tools_web] { enabled = false, fetch = true,
        // search = true }`) exactly like `tools_search`'s three sub-flags.
        if act.is_active(ModuleId::ToolsWeb) {
            if act.tools_web_fetch {
                r.register(WebFetchTool);
            }
            if act.tools_web_search {
                r.register(WebSearchTool);
            }
        }
        // BP-3 (§2 module 6 `tools.question`): the module contributes the
        // question tool. `cx-parity` additionally names Codex's own
        // experimental spelling in `[core.tools] enabled`, so a continued
        // Codex session's `request_user_input` calls keep resolving — the
        // SAME tool object under a second registered name, never a second
        // implementation.
        if act.is_active(ModuleId::ToolsQuestion) {
            r.register(AskUserTool::new(question::ASK_USER));
        }
        if core_has(question::REQUEST_USER_INPUT) {
            r.register(AskUserTool::new(question::REQUEST_USER_INPUT));
        }
        // BP-3 (§2 module 8 `plan_mode`): the two tools the module is
        // defined by. The restriction they establish is enforced by the
        // permissions engine (`plan_mode::deny_rules`, folded into its deny
        // tier by `crate::agent`'s gate), which is why the module's §2.1
        // dependency edge is `plan_mode → permissions.rules|sandbox`.
        if act.is_active(ModuleId::PlanMode) {
            r.register(EnterPlanModeTool);
            r.register(ExitPlanModeTool);
        }
        // BP-3 (catalog rows "Clock / sleep tools", "Context-budget tools",
        // "Image generation tool"): four more OPTIONAL default-tool names,
        // read from the same `[core.tools] enabled` list as `view_image`
        // (§1.2's "fifth optional default-tool name, not a new module"
        // precedent) — absent from the default four, so a config that does
        // not name them gets byte-identical tools to before.
        if core_has(clock::CURRENT_TIME) {
            r.register(CurrentTimeTool);
        }
        if core_has(clock::SLEEP) {
            r.register(SleepTool);
        }
        if core_has(context_budget::GET_CONTEXT_REMAINING) {
            r.register(GetContextRemainingTool);
        }
        if core_has(context_budget::NEW_CONTEXT) {
            r.register(NewContextTool);
        }
        if core_has(image_gen::IMAGE_GEN) {
            r.register(ImageGenTool::new(
                config.base_url.clone(),
                config.api_key.clone(),
                config.api_key_env.clone(),
            ));
        }
        // BP-6 (catalog D1 "Skill-invocation surface"): the one tool every
        // skill is invoked through (CC's `Skill` shape). The tool IS the
        // on-demand read pathway D-7 is about, so it does not itself depend
        // on `read_file`/`bash` being present (cx-parity has neither for
        // files). Discovery runs here, at registry construction, and reads
        // only frontmatter: no body is opened until the model calls it.
        //
        // Registered exactly when the config READS a skill-root table —
        // `[core.skills] enabled` plus a `harness` naming whose roots to
        // walk, the same precondition `load_for_config` applies. A config
        // that enables skills without naming a table discovers nothing, so
        // the tool would have nothing to load; leaving it out keeps every
        // pre-BP-6 config's tool set byte-identical.
        if config.skills_enabled && config.skills_harness.is_some() {
            r.register(
                SkillTool::new(crate::skills::load_for_config(config))
                    // BP-5: the tool door loads bodies under the same
                    // permission-engine authorization every other
                    // invocation door uses.
                    .with_shell(crate::skills::ShellInjection::from_config(config)),
            );
        }
        r
    }

    /// Add a tool. A later registration with the same name shadows the earlier.
    pub fn register(&mut self, tool: impl Tool + 'static) {
        self.tools.push(Box::new(tool));
    }

    /// Look up a tool by name (last registration wins).
    pub fn get(&self, name: &str) -> Option<&dyn Tool> {
        self.tools
            .iter()
            .rev()
            .find(|t| t.name() == name)
            .map(|b| b.as_ref())
    }

    /// Iterate all tools.
    pub fn iter(&self) -> impl Iterator<Item = &dyn Tool> {
        self.tools.iter().map(|b| b.as_ref())
    }

    /// Number of registered tools.
    pub fn len(&self) -> usize {
        self.tools.len()
    }

    /// Whether the registry is empty.
    pub fn is_empty(&self) -> bool {
        self.tools.is_empty()
    }
}