dirge-agent 0.11.3

//! Context compression — structured summaries + session rotation.
//!
//! Faithful port of Hermes's `agent/context_compressor.py` and
//! `agent/conversation_compression.py`. When the conversation
//! approaches the model's context limit, the middle turns are
//! compressed into a structured summary by an auxiliary model,
//! and the session id rotates to enable lineage-based search.
//!
//! Algorithm (from Hermes):
//! 1. Check feasibility — prompt_tokens > 75% of context_window
//! 2. Prune old tool results in the middle section (cheap pre-pass)
//! 3. Determine boundaries — protect head + tail, compress middle
//! 4. Generate structured summary via auxiliary LLM call
//! 5. Assemble compressed: head + summary + tail
//! 6. Rotate session id (parent_session_id chain)
//!
//! WIRING (LOOP-9): Steps 1-3 (pruning + threshold) execute on every
//! fold. Step 4 fires when `LoopSpawnConfig::summarize_fn` is `Some`
//! (forwarded as the final argument to `run_agent_loop_with_summarizer`
//! / `run_loop_with_summarizer`) and there's still meaningful material
//! to summarize after pruning. The same path runs under the
//! `ExitWithSummary` defense-in-depth branch. Step 5 inserts the
//! summary as a system message at the head of
//! `current_context.messages` with the filter-safe `SUMMARY_PREFIX`.
//! Step 6 (actual `session.id` mutation + `Session::compactions` push +
//! `save_session` persistence) is delegated to the event consumer
//! side via the existing `LoopEvent::ContextCompacted` channel — see
//! the audit note in AUDIT_REPORT.md §8.

use serde_json::Value;
use std::pin::Pin;
use std::sync::Arc;

/// Async summarization callback. Receives the fully-built structured
/// prompt (Hermes-style — see `build_summary_prompt`) and returns the
/// summary body produced by the auxiliary model. Callers wire this
/// as a thin "LLM call" closure; the prompt assembly + summary
/// validation live in `run_compaction_pass`.
///
/// `run_agent_loop_with_summarizer` plugs an implementation built
/// from `AnyClient::compress_messages` (or any other one-shot LLM
/// call). `None` disables the LLM pass — the loop falls back to
/// pruning only.
pub type SummarizeFn = Arc<
    dyn Fn(String) -> Pin<Box<dyn Future<Output = anyhow::Result<String>> + Send>> + Send + Sync,
>;

/// Filter-safe preamble injected before the summary so the model
/// treats it as reference, not active instructions.
///
/// Leading sentinel of [`SUMMARY_PREFIX`]. Stable + filter-safe so other
/// subsystems can detect (and exclude) the compaction-summary block inside a
/// merged system prompt — notably the critic, which must NOT treat a summary's
/// already-completed `## Active Task` as an outstanding requirement
/// (`agent_loop::critic`). [`SUMMARY_PREFIX`] must keep starting with this.
pub(crate) const COMPACTION_MARKER: &str = "[CONTEXT COMPACTION — REFERENCE ONLY]";

/// Port of Hermes's SUMMARY_PREFIX (context_compressor.py:37-51).
const SUMMARY_PREFIX: &str = "\
[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted \
into the summary below. This is a handoff from a previous context \
window — treat it as background reference, NOT as active instructions. \
Do NOT answer questions or fulfill requests mentioned in this summary; \
they were already addressed. \
The work described here — INCLUDING the original task — may already be \
complete: the '## Completed Actions' and '## Active State' sections record \
what is already done, so do NOT redo it. \
Your current task is identified in the '## Active Task' section of the \
summary — resume exactly from there. \
Respond ONLY to the latest user message \
that appears AFTER this summary. The current session state (files, \
config, etc.) may reflect work described here — avoid repeating it:";

// Budget constants from Hermes (context_compressor.py:54-59).
const MIN_SUMMARY_TOKENS: u64 = 2000;
const SUMMARY_RATIO: f64 = 0.20;
const SUMMARY_TOKENS_CEILING: u64 = 12_000;

/// dirge-k6be: per-tool-result token cap applied at every
/// turn-end before the next model send. Port of Reasonix's
/// `TURN_END_RESULT_CAP_TOKENS` (docs/ARCHITECTURE.md §4.2).
/// 3000 tokens ≈ 12 KB at the 4-chars-per-token estimate;
/// the model that called the tool already received its full
/// result on the dispatch turn, subsequent turns see a
/// head+tail truncation with a count of dropped tokens so
/// the model can re-call if it needs more.
pub const TURN_END_RESULT_CAP_TOKENS: u64 = 3000;

/// When the estimated context exceeds this fraction of the model window,
/// switch the per-result cap to the tighter `AGGRESSIVE_RESULT_CAP_TOKENS`
/// to head off an overflow BEFORE the (reactive, post-response) fold
/// trigger fires. Intentionally below the 75% fold threshold so the
/// tighter cap has room to work first (IMPROVEMENTS_PLAN #3).
///
/// This is the lowest (0.60) rung of the context-budget ladder; the full
/// ladder is documented in
/// [`crate::agent::agent_loop::context_manager`].
pub const AGGRESSIVE_CAP_THRESHOLD: f64 = 0.60;

/// Per-result token cap in the aggressive tier — still enough to see an
/// error message + key output lines, tight enough that one `grep`/`find`
/// result can't eat ~10% of the window before a fold runs.
pub const AGGRESSIVE_RESULT_CAP_TOKENS: u64 = 1000;

/// Pick the per-result cap for `cap_oversized_tool_results` based on how
/// full the context already is: the tighter aggressive cap once
/// estimated usage crosses `AGGRESSIVE_CAP_THRESHOLD`, else the normal
/// cap (IMPROVEMENTS_PLAN #3). Pure so the tiering is unit-testable.
pub fn tiered_result_cap(estimate_tokens: u64, ctx_max: u64) -> u64 {
    let ratio = estimate_tokens as f64 / ctx_max.max(1) as f64;
    if ratio > AGGRESSIVE_CAP_THRESHOLD {
        AGGRESSIVE_RESULT_CAP_TOKENS
    } else {
        TURN_END_RESULT_CAP_TOKENS
    }
}

/// When the pre-send snip (`cap_oversized_tool_results`) frees at least
/// this fraction of the context window, a NORMAL post-response fold can
/// be skipped — the snip already bought enough headroom
/// (IMPROVEMENTS_PLAN #4).
pub const SNIP_SUFFICIENT_FRACTION: f64 = 0.10;

/// Whether a snip that freed `freed` tokens bought enough headroom to
/// skip a normal fold this turn. Aggressive / force-summary folds always
/// proceed regardless — at 80%+ you need the summary. Pure for testing.
pub fn snip_bought_enough(freed: u64, ctx_max: u64, aggressive: bool) -> bool {
    !aggressive && (freed as f64 / ctx_max.max(1) as f64) > SNIP_SUFFICIENT_FRACTION
}

/// Chars-per-token rough estimate. Port of Hermes's _CHARS_PER_TOKEN.
///
/// This is the project's single token estimator (see
/// [`estimate_messages_tokens`]); it backs the *pre-send* measurement point of
/// the context-budget ladder, while the *post-response* decision uses the
/// API's exact `prompt_tokens`. The two are different measurement points, not
/// two estimators — see [`crate::agent::agent_loop::context_manager`].
const CHARS_PER_TOKEN: u64 = 4;

/// Hard floor for a compression model's context window (64K).
#[allow(dead_code)]
const MINIMUM_CONTEXT_LENGTH: u64 = 64_000;

/// Default protected head (system prompt + first user/assistant turn)
/// and tail (recent live exchanges) message counts. Port of Hermes
/// `protect_head_size` and `protect_last_n` defaults.
pub const PROTECT_HEAD_DEFAULT: usize = 2;
pub const PROTECT_TAIL_DEFAULT: usize = 5;

// ── Public API ───────────────────────────────────────────

/// Should compression be attempted?
/// True when prompt_tokens exceeds the history-fold threshold fraction of
/// context_window. Shares [`HISTORY_FOLD_THRESHOLD`] with the post-usage fold
/// decision so the compression gate and the fold gate can't silently drift
/// apart (dirge-95gl) — both fire at the same point in the context window.
///
/// [`HISTORY_FOLD_THRESHOLD`]: crate::agent::agent_loop::context_manager::HISTORY_FOLD_THRESHOLD
pub fn should_compress(prompt_tokens: u64, context_window: u64) -> bool {
    should_compress_with_threshold(prompt_tokens, context_window, None)
}

/// As [`should_compress`], but honoring a configurable early-fold
/// threshold so the summarizer gate stays in lockstep with
/// [`decide_after_usage_with_threshold`] when an override is set.
///
/// [`decide_after_usage_with_threshold`]: crate::agent::agent_loop::context_manager::decide_after_usage_with_threshold
pub fn should_compress_with_threshold(
    prompt_tokens: u64,
    context_window: u64,
    fold_threshold_override: Option<f64>,
) -> bool {
    use crate::agent::agent_loop::context_manager::effective_fold_threshold;
    let threshold =
        (effective_fold_threshold(fold_threshold_override) * context_window as f64) as u64;
    prompt_tokens > threshold
}

/// Estimate tokens for a slice of messages by summing content
/// lengths and dividing by CHARS_PER_TOKEN.
///
/// dirge-el3n: handles both content shapes:
/// - `content: "string"` (heal-on-load / OpenAI shape)
/// - `content: [{type: "text", text: "..."}, ...]` (Anthropic /
///   dirge's production tool-result shape). Non-text blocks
///   contribute zero — they reach the model as opaque references
///   (image SHA256, tool_use stubs), not raw text.
pub fn estimate_messages_tokens(messages: &[Value]) -> u64 {
    let total_chars: usize = messages
        .iter()
        .map(|m| content_chars(m.get("content")))
        .sum();
    (total_chars as u64).div_ceil(CHARS_PER_TOKEN)
}

fn content_chars(content: Option<&Value>) -> usize {
    match content {
        Some(Value::String(s)) => s.len(),
        Some(Value::Array(blocks)) => blocks.iter().filter_map(text_of_block).map(str::len).sum(),
        _ => 0,
    }
}

/// dirge-k6be: per-tool-result token cap. Truncates any
/// `role: "tool"` / `role: "toolResult"` message whose string
/// content exceeds `max_tokens`, replacing it with a head +
/// truncation marker + tail payload. Returns a new `Vec` —
/// the input slice is not mutated.
///
/// Port of Reasonix `shrinkOversizedToolResultsByTokens`
/// (`loop/shrink.ts:34-62`). Differences from
/// `prune_tool_outputs`:
/// - Token-bound, not chars (CHARS_PER_TOKEN estimator).
/// - Non-destructive: head + tail preserved with a marker,
///   no LLM summarization.
/// - No tail protection: applies to every position (the
///   model that called the tool already received the full
///   result on the dispatch turn).
/// - String-content only: structured-content messages are
///   skipped (out of scope; their truncation strategy
///   depends on block semantics).
///
/// Intended to run BEFORE every model send, idempotent so
/// repeat passes on already-capped results are no-ops.
pub fn cap_oversized_tool_results(messages: &[Value], max_tokens: u64) -> Vec<Value> {
    let max_chars = max_tokens.saturating_mul(CHARS_PER_TOKEN) as usize;
    if max_chars == 0 {
        return messages.to_vec();
    }
    messages
        .iter()
        .map(|msg| {
            let role = msg.get("role").and_then(|v| v.as_str()).unwrap_or("");
            if role != "tool" && role != "toolResult" {
                return msg.clone();
            }
            let Some(content) = msg.get("content") else {
                return msg.clone();
            };
            match content {
                // Heal-on-load shape: scalar string content.
                Value::String(s) => {
                    if s.len() <= max_chars {
                        return msg.clone();
                    }
                    let mut new_msg = msg.clone();
                    new_msg["content"] = Value::String(truncate_with_head_tail(s, max_chars));
                    new_msg
                }
                // Production shape: array of content blocks
                // (`[{type: "text", text: "..."}, ...]`). Sum
                // text-block lengths to compute the total the
                // model would see; cap each oversized text block
                // independently using a per-block budget that
                // shares the total fairly. Non-text blocks
                // (image, tool_use, etc.) pass through.
                Value::Array(blocks) => {
                    let total_text_len: usize = blocks
                        .iter()
                        .filter_map(text_of_block)
                        .map(|t| t.len())
                        .sum();
                    if total_text_len <= max_chars {
                        return msg.clone();
                    }
                    // Single-block fast path: cap directly to
                    // max_chars. (Common: tool result is one
                    // text block.)
                    let text_block_count =
                        blocks.iter().filter(|b| text_of_block(b).is_some()).count();
                    let per_block_budget = match max_chars.checked_div(text_block_count) {
                        Some(d) => std::cmp::max(d, MIN_PER_BLOCK_BUDGET),
                        None => return msg.clone(),
                    };
                    let new_blocks: Vec<Value> = blocks
                        .iter()
                        .map(|b| {
                            let Some(text) = text_of_block(b) else {
                                return b.clone();
                            };
                            if text.len() <= per_block_budget {
                                return b.clone();
                            }
                            let truncated = truncate_with_head_tail(text, per_block_budget);
                            let mut new_block = b.clone();
                            new_block["text"] = Value::String(truncated);
                            new_block
                        })
                        .collect();
                    let mut new_msg = msg.clone();
                    new_msg["content"] = Value::Array(new_blocks);
                    new_msg
                }
                _ => msg.clone(),
            }
        })
        .collect()
}

/// Like [`cap_oversized_tool_results`] but also reports how many tokens
/// the capping freed (IMPROVEMENTS_PLAN #4), measured with the same
/// estimator the fold decision uses — so the loop can tell whether the
/// snip already bought enough headroom to skip a fold. Delegates to the
/// unchanged capper so its behavior (and every existing caller) is
/// untouched.
pub fn cap_oversized_tool_results_counted(
    messages: &[Value],
    max_tokens: u64,
) -> (Vec<Value>, u64) {
    let before = estimate_messages_tokens(messages);
    let capped = cap_oversized_tool_results(messages, max_tokens);
    let after = estimate_messages_tokens(&capped);
    let freed = before.saturating_sub(after);
    (capped, freed)
}

/// Max working-set files re-injected after a fold, and the per-file
/// token budget. 5 × 5000 = 25k tokens worst case. This bounds the
/// restoration cost but does NOT by itself guarantee the window won't
/// re-cross the fold threshold near-full — the caller
/// (`restore_working_files`) enforces that separately via a post-fold
/// headroom guard before injecting (IMPROVEMENTS_PLAN #2).
pub const POST_COMPACT_MAX_FILES: usize = 5;
pub const POST_COMPACT_MAX_TOKENS_PER_FILE: u64 = 5_000;

/// Build `[Post-compaction file snapshot]` system messages for the
/// working-set files, capping the count (`POST_COMPACT_MAX_FILES`) and
/// per-file size (`POST_COMPACT_MAX_TOKENS_PER_FILE`, head+tail
/// truncated). Pure — the file reads happen in the caller — so the cap
/// + truncation are unit-testable (IMPROVEMENTS_PLAN #2).
pub fn build_post_compact_snapshots(files: &[(std::path::PathBuf, String)]) -> Vec<Value> {
    let per_file_chars = (POST_COMPACT_MAX_TOKENS_PER_FILE * CHARS_PER_TOKEN) as usize;
    files
        .iter()
        .take(POST_COMPACT_MAX_FILES)
        .map(|(path, content)| {
            let body = if content.len() > per_file_chars {
                truncate_with_head_tail(content, per_file_chars)
            } else {
                content.clone()
            };
            serde_json::json!({
                "role": "system",
                "content": format!(
                    "[Post-compaction file snapshot: {}]\n{}",
                    path.display(),
                    body
                ),
            })
        })
        .collect()
}

/// Minimum per-block content budget when splitting `max_chars`
/// across multiple text blocks. Ensures each block can hold at
/// least the marker payload — without this floor a
/// many-blocks message could produce empty truncations.
const MIN_PER_BLOCK_BUDGET: usize = 256;

/// Extract the `.text` field from a `{type: "text", text: "..."}`
/// content block. `None` for non-text blocks (image, tool_use…).
fn text_of_block(block: &Value) -> Option<&str> {
    let obj = block.as_object()?;
    if obj.get("type").and_then(|t| t.as_str())? != "text" {
        return None;
    }
    obj.get("text").and_then(|t| t.as_str())
}

/// Concatenate the text the model would actually see from a tool
/// result's `content` field, handling BOTH shapes: the scalar-string
/// (heal-on-load) shape and the production block-array shape
/// (`[{type:"text", text:"..."}, ...]`). dirge-u5ka: production tool
/// results are arrays, so a naive `.as_str()` saw nothing.
fn content_text(content: Option<&Value>) -> String {
    match content {
        Some(Value::String(s)) => s.clone(),
        Some(Value::Array(blocks)) => blocks
            .iter()
            .filter_map(text_of_block)
            .collect::<Vec<_>>()
            .join("\n"),
        _ => String::new(),
    }
}

/// Build a `head + marker + tail` payload sized so the
/// total length is `<= max_chars`. Tail gets 10% of the
/// remaining content budget (capped at 1024 chars to keep
/// deeply-nested file dumps from eating the whole tail
/// allotment). Port of Reasonix `truncateForModel`
/// (`mcp/registry.ts:254-262`).
///
/// Sized for idempotency: a second pass on the output is a
/// no-op because `output.len() <= max_chars` guarantees the
/// outer `content.len() <= max_chars` early-return fires.
fn truncate_with_head_tail(s: &str, max_chars: usize) -> String {
    // Reserve enough budget for the marker (with worst-case
    // 12-digit dropped count). The marker template stays
    // constant; the dropped-count is the only variable.
    const MARKER_OVERHEAD: usize = 160;
    if max_chars <= MARKER_OVERHEAD {
        // Cap too small for both content and a marker; emit
        // just the marker so downstream callers still see
        // "result was truncated".
        return format!(
            "[…truncated {} chars — call the tool with a narrower scope (filter, head, pagination) if you need more…]",
            s.len(),
        );
    }
    let content_budget = max_chars - MARKER_OVERHEAD;
    let tail_budget = std::cmp::min(1024, content_budget / 10);
    let head_budget = content_budget.saturating_sub(tail_budget);
    let head_end = crate::text::char_boundary_at_or_before(s, head_budget);
    let tail_start = crate::text::char_boundary_at_or_after(s, s.len().saturating_sub(tail_budget));
    let head = &s[..head_end];
    let tail = &s[tail_start..];
    let dropped = s.len().saturating_sub(head.len() + tail.len());
    format!(
        "{head}\n\n[…truncated {dropped} chars — call the tool with a narrower scope (filter, head, pagination) if you need more…]\n\n{tail}"
    )
}

/// Prune large tool outputs in the middle section before
/// summarization. Replaces tool-result content > 500 chars
/// with a 1-line summary of what the tool did.
/// Port of Hermes's _prune_old_tool_results (context_compressor.py).
///
/// LOOP-7: matches both `role: "tool"` (heal/legacy shape) and
/// `role: "toolResult"` (loop transcript shape). Also reads both
/// `"tool_name"` (snake_case) and `"toolName"` (camelCase)
/// for the tool name field.
pub fn prune_tool_outputs(messages: &[Value], protect_tail: usize) -> Vec<Value> {
    let n = messages.len();
    if n <= protect_tail {
        return messages.to_vec();
    }
    let end = n.saturating_sub(protect_tail);
    let mut pruned = 0usize;

    messages
        .iter()
        .enumerate()
        .map(|(i, msg)| {
            if i >= end {
                return msg.clone();
            }
            let role = msg.get("role").and_then(|v| v.as_str()).unwrap_or("");
            if role != "tool" && role != "toolResult" {
                return msg.clone();
            }
            // dirge-u5ka: read text from BOTH the scalar-string and the
            // production block-array shapes — previously only `.as_str()`
            // was checked, so live tool results (always arrays) were never
            // pruned and this pass was a silent no-op.
            let content = msg.get("content");
            let text = content_text(content);
            if text.len() <= 500 {
                return msg.clone();
            }
            // Summarize: 1-line tool result.
            let tool_name = msg
                .get("tool_name")
                .or_else(|| msg.get("toolName"))
                .and_then(|v| v.as_str())
                .unwrap_or("tool");
            pruned += 1;
            let summary = summarize_tool_result(tool_name, &text);
            let mut new_msg = msg.clone();
            // Preserve the original content SHAPE: an array stays a single
            // text block (so the LLM-API contract is unchanged), a string
            // stays a string.
            new_msg["content"] = match content {
                Some(Value::Array(_)) => {
                    Value::Array(vec![serde_json::json!({"type": "text", "text": summary})])
                }
                _ => Value::String(summary),
            };
            new_msg
        })
        .collect()
}

fn fmt_count(n: usize) -> String {
    if n < 1000 {
        return n.to_string();
    }
    let s = n.to_string();
    let mut result = String::new();
    let len = s.len();
    for (i, ch) in s.chars().enumerate() {
        if i > 0 && (len - i) % 3 == 0 {
            result.push(',');
        }
        result.push(ch);
    }
    result
}

/// Produce a 1-line summary of a tool result for the pruning pass.
/// Port of Hermes's _summarize_tool_result (context_compressor.py:332).
fn summarize_tool_result(tool_name: &str, content: &str) -> String {
    let content_len = content.len();
    let line_count = content.lines().count();
    let clen = fmt_count(content_len);
    let lc = line_count;

    match tool_name {
        "bash" => {
            let cmd = content
                .lines()
                .next()
                .map(|l| l.trim_start_matches("$ ").trim_start_matches("> "))
                .unwrap_or("?");
            let cmd_short = if cmd.len() > 80 {
                format!("{}…", &cmd[..77])
            } else {
                cmd.to_string()
            };
            format!("[bash] ran `{cmd_short}` -> {lc} lines, {clen} chars")
        }
        "read" => {
            format!("[read] {clen} chars, {lc} lines")
        }
        "write" => {
            format!("[write] wrote {clen} chars")
        }
        "edit" => {
            format!("[edit] patched {clen} chars")
        }
        "grep" => {
            format!("[grep] {lc} matches, {clen} chars")
        }
        "glob" | "find_files" | "list_dir" => {
            let first_line = content.lines().next().unwrap_or("");
            format!("[{}] {first_line}", tool_name)
        }
        "task" | "task_status" => {
            format!("[{tool_name}] {clen} chars result")
        }
        _ => {
            let preview: String = content.chars().take(80).collect();
            format!(
                "[{tool_name}] {preview}{} ({clen} chars)",
                if content.len() > 80 { "…" } else { "" }
            )
        }
    }
}

/// Build the structured summary prompt for the auxiliary model.
/// Port of Hermes's _generate_summary prompt (context_compressor.py:960-1046).
pub fn build_summary_prompt(
    turns_to_summarize: &[Value],
    summary_budget: u64,
    previous_summary: Option<&str>,
    focus_topic: Option<&str>,
) -> String {
    let _summarizer_preamble = "\
You are a summarization agent creating a context checkpoint. \
Treat the conversation turns below as source material for a \
compact record of prior work. \
Produce only the structured summary; do not add a greeting, \
preamble, or prefix. \
Write the summary in the same language the user was using in the \
conversation — do not translate or switch to English.";

    // /compress <focus> argument. When the caller supplies a focus
    // topic, ask the model to allocate ~60-70% of its budget to
    // content related to that topic. Verbatim port of Hermes's
    // FOCUS TOPIC framing (context_compressor.py:1050-1054). Empty
    // / whitespace-only topics are ignored.
    let focus_block: String = match focus_topic.map(|t| t.trim()).filter(|t| !t.is_empty()) {
        Some(topic) => format!(
            "\n\nFOCUS TOPIC: \"{topic}\"\nThe user has requested that this \
            compaction PRIORITISE preserving all information related to the focus \
            topic above. For content related to \"{topic}\", include full detail — \
            exact values, file paths, command outputs, error messages, and \
            decisions. For content NOT related to the focus topic, summarise more \
            aggressively (brief one-liners or omit if truly irrelevant). The focus \
            topic sections should receive roughly 60-70% of the summary token \
            budget. Even for the focus topic, NEVER preserve API keys, tokens, \
            passwords, or credentials — use [REDACTED]."
        ),
        None => String::new(),
    };

    let _template_sections = format!(
        "## Active Task\n\
[THE SINGLE MOST IMPORTANT FIELD. State what should happen NEXT — the\n\
immediate piece of work in flight right now, in plain terms. This is NOT\n\
necessarily the user's original wording: the current work is often an\n\
emergent follow-up (e.g. debugging a failing test) that arose mid-session\n\
and was never an explicit user request — capture THAT, not the original\n\
assignment, when that is what is actually underway. If the user's original\n\
request is already COMPLETE and only follow-up work remains, the Active\n\
Task IS that follow-up; state plainly that the original request is already\n\
done so the next context does not redo it. If multiple tasks were requested\n\
and only some are done, list only the ones NOT yet completed. If nothing is\n\
outstanding, write \"None.\"]\n\
\n\
## Goal\n\
[What the user is trying to accomplish overall]\n\
\n\
## Constraints & Preferences\n\
[User preferences, coding style, constraints, important decisions]\n\
\n\
## Completed Actions\n\
[Numbered list of concrete actions taken — include tool used, target, and outcome.]\n\
\n\
## Active State\n\
[Current working state — directory, branch, modified files, test status]\n\
\n\
## In Progress\n\
[Work currently underway — what was being done when compaction fired]\n\
\n\
## Blocked\n\
[Any blockers, errors, or issues not yet resolved. Include exact error messages.]\n\
\n\
## Key Decisions\n\
[Important technical decisions and WHY they were made]\n\
\n\
## Resolved Questions\n\
[Questions already answered — include the answer]\n\
\n\
## Pending User Asks\n\
[Questions or requests NOT yet answered. If none, write \"None.\"]\n\
\n\
## Relevant Files\n\
[Files read, modified, or created — with brief note on each]\n\
\n\
## Remaining Work\n\
[What remains to be done — framed as context, not instructions]\n\
\n\
## Critical Context\n\
[Specific values, error messages, config details that would be lost\n\
without explicit preservation]\n\
\n\
Target ~{summary_budget} tokens. Be CONCRETE — include file paths,\n\
command outputs, error messages, line numbers, and specific values.\n\
Write only the summary body. Do not include any preamble or prefix."
    );

    let serialized = serialize_turns_for_summary(turns_to_summarize);

    if let Some(prev) = previous_summary {
        format!(
            "{_summarizer_preamble}\n\n\
You are updating a context compaction summary. A previous compaction \
produced the summary below. New conversation turns have occurred since \
then and need to be incorporated.\n\n\
PREVIOUS SUMMARY:\n{prev}\n\n\
NEW TURNS TO INCORPORATE:\n{serialized}{focus_block}\n\n\
Update the summary using this exact structure. PRESERVE all existing \
information that is still relevant. CRITICAL: Update \"## Active Task\" \
to reflect the user's most recent unfulfilled request.\n\n\
{_template_sections}"
        )
    } else {
        format!(
            "{_summarizer_preamble}\n\n\
Create a structured checkpoint summary for the conversation after earlier \
turns are compacted. The summary should preserve enough detail for \
continuity without re-reading the original turns.\n\n\
TURNS TO SUMMARIZE:\n{serialized}{focus_block}\n\n\
Use this exact structure:\n\n\
{_template_sections}"
        )
    }
}

/// Serialize turns for the summarizer prompt. Each turn gets
/// role + content (text fields only, tool results truncated).
fn serialize_turns_for_summary(turns: &[Value]) -> String {
    let mut out = String::new();
    for (i, turn) in turns.iter().enumerate() {
        let role = turn.get("role").and_then(|v| v.as_str()).unwrap_or("?");
        let content = turn.get("content").and_then(|v| v.as_str()).unwrap_or("");
        out.push_str(&format!("[{i}] {role}: "));
        if content.len() > 2000 {
            let truncated: String = content.chars().take(2000).collect();
            out.push_str(&format!(
                "{truncated}… [truncated, {} total chars]\n",
                content.len()
            ));
        } else {
            out.push_str(content);
            out.push('\n');
        }
    }
    out
}

/// Compute the summary budget from the compressed token count.
/// Port of Hermes's _compute_summary_budget.
pub fn summary_budget(compressed_tokens: u64) -> u64 {
    let ratio_budget = (SUMMARY_RATIO * compressed_tokens as f64) as u64;
    ratio_budget.clamp(MIN_SUMMARY_TOKENS, SUMMARY_TOKENS_CEILING)
}

/// Validate that a summary contains the expected sections.
/// At minimum it should mention Active Task and have some structure.
pub fn validate_summary(summary: &str) -> bool {
    if summary.is_empty() {
        return false;
    }
    // Must contain at least one of the expected section headers.
    let required = ["Active Task", "Goal", "Completed Actions", "Remaining Work"];
    required.iter().any(|s| summary.contains(s))
}

/// Find the latest context summary marker in the message list.
/// Returns (index, body) of the last system message containing
/// SUMMARY_PREFIX, or None.
pub fn find_previous_summary(messages: &[Value]) -> Option<(usize, String)> {
    messages.iter().enumerate().rev().find_map(|(i, m)| {
        let role = m.get("role").and_then(|v| v.as_str()).unwrap_or("");
        if role != "system" {
            return None;
        }
        let content = m.get("content").and_then(|v| v.as_str()).unwrap_or("");
        if content.starts_with(SUMMARY_PREFIX) {
            let body = content
                .strip_prefix(SUMMARY_PREFIX)
                .unwrap_or("")
                .trim()
                .to_string();
            Some((i, body))
        } else {
            None
        }
    })
}

/// Replace the middle section of `messages` with a single
/// system-summary message. Returns the new messages list.
/// Port of Hermes `compress` phase 4 (context_compressor.py:1632-1714).
///
/// `compress_start..compress_end` is dropped and replaced with one
/// system message carrying `SUMMARY_PREFIX + summary`. Messages
/// before `compress_start` (protected head — system prompt + first
/// exchange) and at or after `compress_end` (protected tail) are
/// preserved verbatim.
pub fn apply_summary(
    messages: &[Value],
    summary: &str,
    compress_start: usize,
    compress_end: usize,
) -> Vec<Value> {
    let n = messages.len();
    let compress_start = compress_start.min(n);
    let compress_end = compress_end.min(n).max(compress_start);

    let mut out: Vec<Value> =
        Vec::with_capacity(n.saturating_sub(compress_end - compress_start) + 1);
    // Protected head — copy verbatim.
    for msg in messages.iter().take(compress_start) {
        out.push(msg.clone());
    }
    // Summary marker — filter-safe prefix + body.
    let summary_msg = serde_json::json!({
        "role": "system",
        "content": format!("{}{}", SUMMARY_PREFIX, summary),
    });
    out.push(summary_msg);
    // Protected tail — copy verbatim.
    for msg in messages.iter().skip(compress_end) {
        out.push(msg.clone());
    }
    out
}

/// Fold using a precomputed running summary that already covers
/// `messages[0..boundary]` (the background incremental checkpoint). The
/// covered prefix is replaced with one summary marker; everything from the
/// snapped tail-cut onward is kept verbatim. This is the FAST fold path:
/// no inline summarizer call, because the expensive summarization already
/// ran off the loop.
///
/// Returns `Some((new_messages, first_kept_index))` on success, or `None`
/// when reuse wouldn't be safe or useful:
/// - `boundary` is 0 or past the end (nothing to fold, or stale),
/// - the user-boundary snap collapses the cut to 0 (no whole turn to fold).
///
/// The cut snaps BACKWARD to a user-message boundary so the kept tail never
/// begins with an orphaned tool_result (which would 400 the next request).
/// Snapping backward only ever keeps MORE verbatim than the summary covers
/// — it never drops an un-summarized message — so it can't lose data; the
/// few messages in `[cut..boundary]` are simply carried both in the summary
/// and verbatim.
pub fn apply_checkpoint_summary(
    messages: &[Value],
    summary: &str,
    boundary: usize,
) -> Option<(Vec<Value>, usize)> {
    let n = messages.len();
    if boundary == 0 || boundary > n {
        return None;
    }
    let cut = snap_backward_to_user(messages, boundary);
    if cut == 0 {
        return None;
    }
    let out = apply_summary(messages, summary, 0, cut);
    Some((out, cut))
}

/// Compute the boundary `(compress_start, compress_end)` for the
/// middle section to summarize. Port of Hermes's
/// `_protect_head_size` + `_find_tail_cut_by_tokens`.
///
/// `protect_head` and `protect_tail` are message counts. Returns
/// `(0, 0)` to signal "nothing to compress" when the message list
/// is too short to safely partition.
pub fn compute_compress_window(
    messages: &[Value],
    protect_head: usize,
    protect_tail: usize,
) -> (usize, usize) {
    let n = messages.len();
    if n < protect_head + protect_tail + 1 {
        return (0, 0);
    }
    let raw_start = protect_head;
    let raw_end = n.saturating_sub(protect_tail);
    if raw_start >= raw_end {
        return (0, 0);
    }
    // dirge-89fm: snap both cuts to USER-message boundaries. A user
    // message never carries a dangling tool_use (its results follow as
    // separate messages) and is never itself an orphaned tool_result, so
    // cutting there guarantees the dropped middle removes whole turns —
    // neither the protected head (ends just before a user turn) nor the
    // protected tail (starts on a user turn) is left holding half of a
    // tool_use↔tool_result pair, which would 400 the next API request.
    // Same discipline as the `/compress` path's `align_cut_to_user_boundary`.
    // Head snaps FORWARD (protects ≥ protect_head), tail snaps BACKWARD
    // (protects ≥ protect_tail); both only ever protect more, never less.
    let start = snap_forward_to_user(messages, raw_start);
    let end = snap_backward_to_user(messages, raw_end);
    if start >= end {
        return (0, 0);
    }
    (start, end)
}

/// True when `msg` is a user-role turn.
fn is_user_msg(msg: &Value) -> bool {
    msg.get("role").and_then(|r| r.as_str()) == Some("user")
}

/// Smallest index `>= idx` whose message is a user turn, clamped to
/// `messages.len()` when none is found at or after `idx`.
fn snap_forward_to_user(messages: &[Value], idx: usize) -> usize {
    let n = messages.len();
    let mut i = idx.min(n);
    while i < n && !is_user_msg(&messages[i]) {
        i += 1;
    }
    i
}

/// Largest index `<= idx` whose message is a user turn, or `0` when none
/// is found at or before `idx`.
fn snap_backward_to_user(messages: &[Value], idx: usize) -> usize {
    let mut i = idx.min(messages.len().saturating_sub(1));
    loop {
        if is_user_msg(&messages[i]) {
            return i;
        }
        if i == 0 {
            return 0;
        }
        i -= 1;
    }
}

/// Generate a new session id with a `compacted-` prefix to
/// disambiguate from fresh sessions. Port of Hermes's
/// `parent_session_id` rotation pattern (conversation_compression.py:383).
pub fn rotate_session_id() -> String {
    format!(
        "compacted-{}",
        uuid::Uuid::new_v4()
            .to_string()
            .chars()
            .take(8)
            .collect::<String>()
    )
}

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

    /// `COMPACTION_MARKER` is the sentinel other subsystems (the critic) use to
    /// detect + strip the summary block. It MUST stay the prefix of the full
    /// `SUMMARY_PREFIX`, or those strippers silently stop working.
    #[test]
    fn summary_prefix_starts_with_compaction_marker() {
        assert!(SUMMARY_PREFIX.starts_with(COMPACTION_MARKER));
    }

    /// #443: the prefix must keep the marker AND warn that work described in
    /// the summary (including the original task) may already be complete, so a
    /// resumed context doesn't redo finished work.
    #[test]
    fn summary_prefix_warns_original_task_may_be_complete() {
        assert!(SUMMARY_PREFIX.starts_with(COMPACTION_MARKER));
        // Points the model at the already-done record.
        assert!(SUMMARY_PREFIX.contains("Completed Actions"));
        // Says the work — including the original task — may already be complete.
        let lower = SUMMARY_PREFIX.to_lowercase();
        assert!(lower.contains("already"));
        assert!(lower.contains("complete") || lower.contains("done"));
        assert!(lower.contains("do not redo") || lower.contains("not redo"));
    }

    /// #443: the `## Active Task` instruction must frame the active task as the
    /// immediate next/follow-up work, not a verbatim copy of the user's
    /// original request.
    #[test]
    fn active_task_section_frames_followup_not_verbatim() {
        let turns: Vec<Value> = vec![serde_json::json!({
            "role": "user",
            "content": "convert this to stdlib"
        })];
        let prompt = build_summary_prompt(&turns, 2000, None, None);

        // New framing is present.
        assert!(prompt.contains("## Active Task"));
        assert!(prompt.contains("follow-up"));
        let lower = prompt.to_lowercase();
        assert!(lower.contains("already") || lower.contains("complete"));

        // Old verbatim-copy framing is gone.
        assert!(!prompt.contains("verbatim — the exact words they used"));
    }

    // IMPROVEMENTS_PLAN #3: the per-result cap tightens above 60% ctx.
    #[test]
    fn tiered_result_cap_switches_at_threshold() {
        let ctx = 128_000u64;
        // Well below 60% → normal cap.
        assert_eq!(
            tiered_result_cap(ctx / 4, ctx), // 25%
            TURN_END_RESULT_CAP_TOKENS
        );
        // Exactly at 60% is NOT over the threshold (strict `>`).
        assert_eq!(
            tiered_result_cap((ctx as f64 * 0.60) as u64, ctx),
            TURN_END_RESULT_CAP_TOKENS
        );
        // Above 60% → aggressive cap.
        assert_eq!(
            tiered_result_cap((ctx as f64 * 0.70) as u64, ctx),
            AGGRESSIVE_RESULT_CAP_TOKENS
        );
        // Degenerate ctx_max doesn't panic (div-by-zero guard).
        let _ = tiered_result_cap(100, 0);
    }

    // IMPROVEMENTS_PLAN #4: snip feedback loop.
    #[test]
    fn snip_bought_enough_gates_normal_folds_only() {
        let ctx = 128_000u64;
        // > 10% freed, normal fold → enough.
        assert!(snip_bought_enough((ctx as f64 * 0.11) as u64, ctx, false));
        // < 10% freed → not enough.
        assert!(!snip_bought_enough((ctx as f64 * 0.05) as u64, ctx, false));
        // Aggressive fold always proceeds, even if a lot was freed.
        assert!(!snip_bought_enough(ctx, ctx, true));
        // div-by-zero guard.
        assert!(!snip_bought_enough(0, 0, false));
    }

    #[test]
    fn cap_counted_reports_freed_tokens() {
        // One oversized tool result.
        let big = "x".repeat(40_000); // ~10k tokens
        let msgs = vec![serde_json::json!({"role": "tool", "content": big})];
        let (capped, freed) = cap_oversized_tool_results_counted(&msgs, 1000);
        // Same result as the uncounted capper.
        assert_eq!(capped, cap_oversized_tool_results(&msgs, 1000));
        // It freed a meaningful number of tokens (oversized → trimmed).
        assert!(
            freed > 5_000,
            "expected substantial freed tokens, got {freed}"
        );

        // A result already under the cap frees nothing.
        let small = vec![serde_json::json!({"role": "tool", "content": "ok"})];
        let (_, freed0) = cap_oversized_tool_results_counted(&small, 1000);
        assert_eq!(freed0, 0);
    }

    // IMPROVEMENTS_PLAN #2: post-compaction working-set snapshots.
    #[test]
    fn build_post_compact_snapshots_caps_count_and_truncates() {
        use std::path::PathBuf;
        // More files than the cap → capped, in order.
        let files: Vec<(PathBuf, String)> = (0..8)
            .map(|i| (PathBuf::from(format!("f{i}.rs")), "x".repeat(100)))
            .collect();
        let snaps = build_post_compact_snapshots(&files);
        assert_eq!(snaps.len(), POST_COMPACT_MAX_FILES, "capped at MAX_FILES");
        for (i, s) in snaps.iter().enumerate() {
            assert_eq!(s["role"], "system");
            let c = s["content"].as_str().unwrap();
            assert!(
                c.contains(&format!("[Post-compaction file snapshot: f{i}.rs]")),
                "snapshot marker + path missing: {c}"
            );
        }

        // An oversized file is truncated to roughly the per-file budget.
        let per_file_chars = (POST_COMPACT_MAX_TOKENS_PER_FILE * CHARS_PER_TOKEN) as usize;
        let big = (PathBuf::from("big.rs"), "y".repeat(per_file_chars * 4));
        let snaps = build_post_compact_snapshots(std::slice::from_ref(&big));
        let c = snaps[0]["content"].as_str().unwrap();
        assert!(
            c.len() < per_file_chars + 1_000,
            "oversized file must be truncated to ~budget; got {} chars",
            c.len()
        );
    }

    // ── should_compress ─────────────────────────────────

    #[test]
    fn below_75pct_no_compress() {
        // 50K tokens in 128K window = 39% — no compression.
        assert!(!should_compress(50_000, 128_000));
    }

    #[test]
    fn at_threshold_no_compress() {
        // Exactly 75% — NOT compressed (must EXCEED threshold).
        assert!(!should_compress(96_000, 128_000));
    }

    #[test]
    fn above_threshold_compress() {
        // Just above 75% threshold.
        assert!(should_compress(96_001, 128_000));
    }

    /// dirge-95gl: the compression gate tracks the post-usage fold constant —
    /// they must flip at the same fraction so the two decisions can't drift.
    #[test]
    fn should_compress_tracks_history_fold_threshold() {
        use crate::agent::agent_loop::context_manager::HISTORY_FOLD_THRESHOLD;
        let win = 200_000u64;
        let at = (HISTORY_FOLD_THRESHOLD * win as f64) as u64;
        assert!(!should_compress(at, win)); // exactly at the shared threshold → no
        assert!(should_compress(at + 1, win)); // just over → yes
    }

    #[test]
    fn exactly_at_threshold_edge() {
        // 75% of 128000 = 96000
        assert!(!should_compress(96_000, 128_000));
        assert!(should_compress(96_001, 128_000));
    }

    // ── summary_budget ──────────────────────────────────

    #[test]
    fn budget_minimum() {
        // Small compressed content → minimum budget.
        assert_eq!(summary_budget(1_000), MIN_SUMMARY_TOKENS);
    }

    #[test]
    fn budget_proportional() {
        // 50K compressed → 10K budget (20%).
        assert_eq!(summary_budget(50_000), 10_000);
    }

    #[test]
    fn budget_ceiling() {
        // Very large compressed content → ceiling.
        assert_eq!(summary_budget(500_000), SUMMARY_TOKENS_CEILING);
    }

    #[test]
    fn budget_clamp() {
        assert_eq!(summary_budget(0), MIN_SUMMARY_TOKENS);
        assert_eq!(summary_budget(1_000_000), SUMMARY_TOKENS_CEILING);
    }

    // ── prune_tool_outputs ──────────────────────────────

    #[test]
    fn prune_large_tool_results() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello"}),
            serde_json::json!({"role": "assistant", "content": "hi"}),
            serde_json::json!({"role": "tool", "content": "x".repeat(1000), "tool_name": "read"}),
            serde_json::json!({"role": "tool", "content": "small", "tool_name": "grep"}),
            serde_json::json!({"role": "user", "content": "tail"}),
        ];

        let pruned = prune_tool_outputs(&msgs, 2);
        // Large tool result should be summarized.
        let tool1 = &pruned[2];
        assert!(tool1["content"].as_str().unwrap().contains("[read]"));
        assert!(!tool1["content"].as_str().unwrap().contains("xxxxx"));

        // Small tool result unchanged.
        assert_eq!(pruned[3]["content"].as_str().unwrap(), "small");

        // Tail protected.
        assert_eq!(pruned[4]["content"].as_str().unwrap(), "tail");
    }

    /// LOOP-7: loop transcripts use `role: "toolResult"` and
    /// `"toolName"` (camelCase), not `role: "tool"` and `"tool_name"`.
    /// Pruning must recognize both formats.
    #[test]
    fn prune_handles_tool_result_role_and_camelcase_toolname() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello"}),
            serde_json::json!({"role": "toolResult", "content": "x".repeat(1000), "toolName": "bash"}),
            serde_json::json!({"role": "toolResult", "content": "small", "toolName": "grep"}),
            serde_json::json!({"role": "user", "content": "tail"}),
        ];

        let pruned = prune_tool_outputs(&msgs, 2);
        // The large toolResult should be summarized now (contains "[bash]" marker).
        let summary = pruned[1]["content"].as_str().unwrap();
        assert!(
            summary.contains("[bash]"),
            "should summarize bash tool result: {summary}"
        );
        // The summary should be MUCH shorter than the original 1000 chars
        // (it contains the escaped command + metadata, but the 1000 x's are truncated).
        assert!(
            summary.len() < 500,
            "summary should be under 500 chars: {}",
            summary.len()
        );
        // Small result in tail should be untouched.
        assert_eq!(pruned[2]["content"].as_str().unwrap(), "small");
    }

    /// dirge-u5ka: production tool results carry block-ARRAY content
    /// (`[{type:"text", text:"..."}]`), not a scalar string. Pruning must
    /// see through the array — previously `.as_str()` returned None so
    /// these were never pruned and the pass was a silent no-op.
    #[test]
    fn prune_handles_block_array_content() {
        let big = "y".repeat(1000);
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello"}),
            serde_json::json!({
                "role": "toolResult",
                "toolName": "read",
                "content": [{"type": "text", "text": big}],
            }),
            serde_json::json!({"role": "user", "content": "tail"}),
        ];
        let pruned = prune_tool_outputs(&msgs, 1);
        let content = &pruned[1]["content"];
        // Shape preserved: still an array of one text block...
        let blocks = content.as_array().expect("content stays a block array");
        assert_eq!(blocks.len(), 1);
        let text = blocks[0]["text"].as_str().unwrap();
        // ...but summarized (carries the [read] marker, drops the 1000 y's).
        assert!(text.contains("[read]"), "summarized: {text}");
        assert!(!text.contains("yyyyy"), "raw content dropped: {text}");
        assert!(text.len() < 500);
    }

    /// A small block-array tool result is left untouched.
    #[test]
    fn prune_leaves_small_block_array_untouched() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello"}),
            serde_json::json!({
                "role": "toolResult",
                "toolName": "grep",
                "content": [{"type": "text", "text": "two matches"}],
            }),
            serde_json::json!({"role": "user", "content": "tail"}),
        ];
        let pruned = prune_tool_outputs(&msgs, 1);
        assert_eq!(pruned[1], msgs[1], "small block-array result is unchanged");
    }

    // ── cap_oversized_tool_results (dirge-k6be) ─────────

    /// Under the cap → message passes through unchanged.
    #[test]
    fn cap_passes_small_tool_results_through() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello"}),
            serde_json::json!({"role": "toolResult", "content": "tiny", "toolName": "read"}),
        ];
        let capped = cap_oversized_tool_results(&msgs, 1000);
        assert_eq!(capped, msgs, "no message should change under the cap");
    }

    /// Over the cap → content gets truncated with a head + tail +
    /// marker payload that mentions the dropped count. Original
    /// content NOT preserved verbatim.
    #[test]
    fn cap_truncates_oversized_tool_result_with_head_tail_marker() {
        // 40 KB of 'x' chars ≈ 10000 tokens — well over a 100-token cap.
        let big = "x".repeat(40_000);
        let msgs = vec![
            serde_json::json!({"role": "toolResult", "content": big.clone(), "toolName": "read"}),
        ];
        let capped = cap_oversized_tool_results(&msgs, 100);
        let content = capped[0]["content"].as_str().unwrap();
        // Must be smaller than the input.
        assert!(
            content.len() < big.len(),
            "capped content must be shorter: got {} vs {}",
            content.len(),
            big.len(),
        );
        // Truncation marker must mention the dropped count.
        assert!(
            content.contains("truncated"),
            "must mention truncation: {content:?}",
        );
        // Both head and tail of the original must be present (x's).
        assert!(content.starts_with('x'), "head preserved: {content:?}");
        assert!(content.ends_with('x'), "tail preserved: {content:?}");
    }

    /// Cap respects both `role: "tool"` and `role: "toolResult"`.
    #[test]
    fn cap_handles_both_tool_role_shapes() {
        let big = "y".repeat(40_000);
        let msgs = vec![
            serde_json::json!({"role": "tool", "content": big.clone(), "tool_name": "bash"}),
            serde_json::json!({"role": "toolResult", "content": big.clone(), "toolName": "bash"}),
        ];
        let capped = cap_oversized_tool_results(&msgs, 100);
        for (i, msg) in capped.iter().enumerate() {
            let content = msg["content"].as_str().unwrap();
            assert!(
                content.len() < big.len(),
                "message {i} must be capped: len={}",
                content.len()
            );
            assert!(content.contains("truncated"), "message {i} missing marker");
        }
    }

    /// Non-tool roles (`user`, `assistant`, `system`) are NEVER
    /// touched, even when their content is huge. Truncating a
    /// user prompt would corrupt authored intent
    /// (Reasonix `shrink.ts:17`).
    #[test]
    fn cap_never_touches_non_tool_messages() {
        let big = "z".repeat(40_000);
        let msgs = vec![
            serde_json::json!({"role": "user", "content": big.clone()}),
            serde_json::json!({"role": "assistant", "content": big.clone()}),
            serde_json::json!({"role": "system", "content": big.clone()}),
        ];
        let capped = cap_oversized_tool_results(&msgs, 100);
        assert_eq!(capped, msgs, "non-tool messages must pass through verbatim");
    }

    /// Idempotent: capping an already-capped result is a no-op
    /// (the marker payload itself is under any reasonable cap).
    #[test]
    fn cap_is_idempotent_on_already_capped_results() {
        let big = "a".repeat(40_000);
        let msgs =
            vec![serde_json::json!({"role": "toolResult", "content": big, "toolName": "read"})];
        let first = cap_oversized_tool_results(&msgs, 100);
        let second = cap_oversized_tool_results(&first, 100);
        assert_eq!(
            first, second,
            "second pass must produce no change: first={first:?} second={second:?}",
        );
    }

    /// Cap applies to ALL tool results regardless of position.
    /// Reasonix `shrink.ts:23-31` has no tail-protection.
    /// (Unlike `prune_tool_outputs` which protects the tail —
    /// that's a different pass for a different purpose.)
    #[test]
    fn cap_applies_to_every_position_including_last() {
        let big = "b".repeat(40_000);
        let msgs = vec![
            serde_json::json!({"role": "toolResult", "content": big.clone(), "toolName": "read"}),
            serde_json::json!({"role": "user", "content": "next"}),
            serde_json::json!({"role": "toolResult", "content": big.clone(), "toolName": "read"}),
        ];
        let capped = cap_oversized_tool_results(&msgs, 100);
        for i in [0, 2] {
            let content = capped[i]["content"].as_str().unwrap();
            assert!(
                content.len() < big.len(),
                "tool result at index {i} must be capped",
            );
        }
    }

    /// Content that's borderline — slightly over the cap —
    /// still truncates (no off-by-one slack).
    #[test]
    fn cap_truncates_borderline_oversized_content() {
        // Cap = 50 tokens ≈ 200 chars. 250 chars is over.
        let content = "c".repeat(250);
        let msgs =
            vec![serde_json::json!({"role": "toolResult", "content": content, "toolName": "read"})];
        let capped = cap_oversized_tool_results(&msgs, 50);
        let s = capped[0]["content"].as_str().unwrap();
        assert!(
            s.contains("truncated"),
            "borderline content must trigger cap: {s}"
        );
    }

    /// Production shape: `content` is an array of text blocks.
    /// Oversized text gets capped inside the block. This is
    /// the shape `tool_result_to_value` (run.rs) actually
    /// produces in the live loop.
    #[test]
    fn cap_truncates_oversized_text_inside_block_array() {
        let big = "d".repeat(40_000);
        let msgs = vec![serde_json::json!({
            "role": "toolResult",
            "content": [{"type": "text", "text": big}],
            "toolName": "read",
        })];
        let capped = cap_oversized_tool_results(&msgs, 100);
        let text = capped[0]["content"][0]["text"].as_str().unwrap();
        assert!(
            text.len() < 40_000,
            "block text must be capped: {}",
            text.len()
        );
        assert!(
            text.contains("truncated"),
            "marker must be present: {text:?}"
        );
        // Block type preserved.
        assert_eq!(capped[0]["content"][0]["type"].as_str(), Some("text"));
    }

    /// Multi-block: cap distributes the budget across text
    /// blocks. Non-text blocks pass through.
    #[test]
    fn cap_handles_multi_block_content_with_mixed_types() {
        let big = "e".repeat(20_000);
        let msgs = vec![serde_json::json!({
            "role": "toolResult",
            "content": [
                {"type": "text", "text": big.clone()},
                {"type": "image", "source": "ignored"},
                {"type": "text", "text": big.clone()},
            ],
            "toolName": "bash",
        })];
        let capped = cap_oversized_tool_results(&msgs, 500);
        let blocks = capped[0]["content"].as_array().unwrap();
        // Image block passes through.
        assert_eq!(blocks[1]["type"].as_str(), Some("image"));
        assert_eq!(blocks[1]["source"].as_str(), Some("ignored"));
        // Both text blocks capped.
        assert!(blocks[0]["text"].as_str().unwrap().len() < 20_000);
        assert!(blocks[2]["text"].as_str().unwrap().len() < 20_000);
    }

    /// Array shape under the cap is a no-op.
    #[test]
    fn cap_passes_small_block_arrays_through() {
        let msgs = vec![serde_json::json!({
            "role": "toolResult",
            "content": [{"type": "text", "text": "small"}],
            "toolName": "read",
        })];
        let capped = cap_oversized_tool_results(&msgs, 100);
        assert_eq!(capped, msgs);
    }

    #[test]
    fn prune_protects_tail() {
        let msgs = vec![
            serde_json::json!({"role": "tool", "content": "x".repeat(1000), "tool_name": "bash"}),
            serde_json::json!({"role": "tool", "content": "y".repeat(1000), "tool_name": "read"}),
            serde_json::json!({"role": "user", "content": "protected"}),
            serde_json::json!({"role": "assistant", "content": "protected"}),
        ];

        // Protect last 3 → only the first tool result is pruned.
        let pruned = prune_tool_outputs(&msgs, 3);
        assert!(pruned[0]["content"].as_str().unwrap().contains("[bash]"));
        // Second tool result is in the tail (index 1, n=4, protect 3 → end=1).
        // Index 1 is protected if n - protect_tail = 4 - 3 = 1, end=1,
        // so indices 0..0 are pruned, index 1 is protected.
        assert!(pruned[1]["content"].as_str().unwrap().contains("yyyy"));
    }

    // ── estimate_messages_tokens ────────────────────────

    #[test]
    fn estimate_tokens_from_content() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello world"}),
            serde_json::json!({"role": "assistant", "content": "0123456789012345"}),
        ];
        // "hello world" = 11 chars, "0123456789012345" = 16 chars, total = 27
        // 27 / 4 = 6.75 → ceil = 7
        assert_eq!(estimate_messages_tokens(&msgs), 7);
    }

    #[test]
    fn estimate_tokens_handles_missing_content() {
        let msgs = vec![serde_json::json!({"role": "system"})];
        assert_eq!(estimate_messages_tokens(&msgs), 0);
    }

    /// dirge-el3n: the estimator must count block-shaped content
    /// (production tool-result shape), not just scalar strings.
    /// Without this, a turn that's 95% tool-result-blocks looks
    /// like 0 tokens and the proactive fold never fires.
    #[test]
    fn estimate_tokens_counts_text_inside_block_arrays() {
        let big = "x".repeat(40);
        let msgs = vec![serde_json::json!({
            "role": "toolResult",
            "content": [{"type": "text", "text": big.clone()}],
            "toolName": "read",
        })];
        // 40 chars / 4 = 10 tokens.
        assert_eq!(estimate_messages_tokens(&msgs), 10);
    }

    /// Multi-block content sums across all text blocks; non-text
    /// blocks (image, tool_use) contribute zero.
    #[test]
    fn estimate_tokens_sums_multi_block_skipping_non_text() {
        let msgs = vec![serde_json::json!({
            "role": "toolResult",
            "content": [
                {"type": "text", "text": "a".repeat(20)},
                {"type": "image", "source": "ignored"},
                {"type": "text", "text": "b".repeat(20)},
            ],
            "toolName": "bash",
        })];
        // 40 chars / 4 = 10 tokens.
        assert_eq!(estimate_messages_tokens(&msgs), 10);
    }

    /// Mix of string-content + block-content messages — both
    /// shapes contribute. (Realistic transcript: user/assistant
    /// strings + tool-result blocks.)
    #[test]
    fn estimate_tokens_mixed_string_and_block_messages() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "hello"}), // 5 chars
            serde_json::json!({
                "role": "toolResult",
                "content": [{"type": "text", "text": "x".repeat(11)}],
                "toolName": "read",
            }), // 11 chars
        ];
        // (5 + 11) / 4 = 4 tokens.
        assert_eq!(estimate_messages_tokens(&msgs), 4);
    }

    // ── validate_summary ────────────────────────────────

    #[test]
    fn valid_summary_passes() {
        assert!(validate_summary(
            "## Active Task\nRefactor auth module\n\n## Completed Actions\n1. READ config.py"
        ));
    }

    #[test]
    fn empty_summary_fails() {
        assert!(!validate_summary(""));
    }

    #[test]
    fn irrelevant_text_fails() {
        assert!(!validate_summary("just some random text with no structure"));
    }

    // ── build_summary_prompt ────────────────────────────

    #[test]
    fn prompt_contains_filter_safe_preamble() {
        let turns = vec![
            serde_json::json!({"role": "user", "content": "fix the bug"}),
            serde_json::json!({"role": "assistant", "content": "ok let me read the file"}),
        ];
        let prompt = build_summary_prompt(&turns, 2000, None, None);
        assert!(prompt.contains("summarization agent"));
        assert!(prompt.contains("TURNS TO SUMMARIZE"));
        assert!(prompt.contains("## Active Task"));
        assert!(prompt.contains("## Remaining Work"));
        assert!(prompt.contains("fix the bug"));
        assert!(prompt.contains("ok let me read the file"));
    }

    #[test]
    fn iterative_prompt_includes_previous_summary() {
        let turns = vec![serde_json::json!({"role": "user", "content": "new stuff"})];
        let prompt = build_summary_prompt(&turns, 2000, Some("Old summary"), None);
        assert!(prompt.contains("PREVIOUS SUMMARY"));
        assert!(prompt.contains("Old summary"));
        assert!(prompt.contains("NEW TURNS TO INCORPORATE"));
    }

    #[test]
    fn prompt_truncates_long_content() {
        let long = "x".repeat(3000);
        let turns = vec![serde_json::json!({"role": "assistant", "content": long})];
        let prompt = build_summary_prompt(&turns, 2000, None, None);
        assert!(prompt.contains("truncated"));
        // The prompt includes template text + truncated content, so it'll be
        // under a reasonable size but longer than the content alone.
        assert!(prompt.len() < 10_000, "prompt should be under 10K chars");
    }

    // ── find_previous_summary ───────────────────────────

    #[test]
    fn finds_latest_summary() {
        let msgs = vec![
            serde_json::json!({"role": "system", "content": "system prompt"}),
            serde_json::json!({"role": "user", "content": "hello"}),
            serde_json::json!({"role": "system", "content": format!("{}## Active Task\nfix the bug", SUMMARY_PREFIX)}),
        ];
        let found = find_previous_summary(&msgs);
        assert!(found.is_some());
        let (_idx, body) = found.unwrap();
        assert!(body.contains("fix the bug"));
    }

    #[test]
    fn no_summary_returns_none() {
        let msgs = vec![
            serde_json::json!({"role": "system", "content": "system prompt"}),
            serde_json::json!({"role": "user", "content": "hello"}),
        ];
        assert!(find_previous_summary(&msgs).is_none());
    }

    // ── apply_summary / compute_compress_window ─────────

    #[test]
    fn apply_summary_inserts_system_message_with_prefix() {
        let msgs = vec![
            serde_json::json!({"role": "system", "content": "you are an agent"}),
            serde_json::json!({"role": "user", "content": "first user msg"}),
            serde_json::json!({"role": "assistant", "content": "old assistant"}),
            serde_json::json!({"role": "user", "content": "old user"}),
            serde_json::json!({"role": "assistant", "content": "old assistant 2"}),
            serde_json::json!({"role": "user", "content": "recent user"}),
            serde_json::json!({"role": "assistant", "content": "recent assistant"}),
        ];
        let summary = "## Active Task\nfix the bug\n## Remaining Work\nrun tests";
        let out = apply_summary(&msgs, summary, 2, 5);
        // Head preserved (2 messages) + 1 summary + tail (2 messages) = 5.
        assert_eq!(out.len(), 5);
        assert_eq!(out[0]["content"].as_str().unwrap(), "you are an agent");
        assert_eq!(out[1]["content"].as_str().unwrap(), "first user msg");
        // Summary message at index 2.
        assert_eq!(out[2]["role"].as_str().unwrap(), "system");
        let s = out[2]["content"].as_str().unwrap();
        assert!(
            s.starts_with(SUMMARY_PREFIX),
            "summary should start with prefix"
        );
        assert!(s.contains("## Active Task"));
        assert!(s.contains("fix the bug"));
        // Tail.
        assert_eq!(out[3]["content"].as_str().unwrap(), "recent user");
        assert_eq!(out[4]["content"].as_str().unwrap(), "recent assistant");
    }

    #[test]
    fn checkpoint_reuse_folds_covered_prefix_and_keeps_tail() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "u0"}),
            serde_json::json!({"role": "assistant", "content": "a0"}),
            serde_json::json!({"role": "user", "content": "u1"}),
            serde_json::json!({"role": "assistant", "content": "a1"}),
            // boundary = 4: checkpoint covers u0..a1; tail starts here.
            serde_json::json!({"role": "user", "content": "u2"}),
            serde_json::json!({"role": "assistant", "content": "a2"}),
        ];
        let summary = "## Active Task\nport the loop\n## Remaining Work\nwire it";
        let (out, first_kept) = apply_checkpoint_summary(&msgs, summary, 4).unwrap();
        // index 4 ("u2") is already a user boundary, so cut == boundary.
        assert_eq!(first_kept, 4);
        // [summary] + tail(u2, a2) = 3 messages.
        assert_eq!(out.len(), 3);
        assert_eq!(out[0]["role"].as_str().unwrap(), "system");
        assert!(
            out[0]["content"]
                .as_str()
                .unwrap()
                .starts_with(SUMMARY_PREFIX)
        );
        assert!(
            out[0]["content"]
                .as_str()
                .unwrap()
                .contains("port the loop")
        );
        assert_eq!(out[1]["content"].as_str().unwrap(), "u2");
        assert_eq!(out[2]["content"].as_str().unwrap(), "a2");
    }

    #[test]
    fn checkpoint_reuse_snaps_cut_back_to_user_boundary() {
        // boundary lands mid-turn (on an assistant/tool message); the cut
        // must snap BACK to the preceding user turn so the kept tail never
        // starts with an orphaned tool_result.
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "u0"}),
            serde_json::json!({"role": "assistant", "content": "a0"}),
            serde_json::json!({"role": "user", "content": "u1"}),
            serde_json::json!({"role": "assistant", "content": "a1"}),
            serde_json::json!({"role": "tool", "content": "t1"}),
        ];
        let summary = "## Goal\nx";
        // boundary = 5 (end); snap_backward finds u1 at index 2.
        let (out, first_kept) = apply_checkpoint_summary(&msgs, summary, 5).unwrap();
        assert_eq!(first_kept, 2, "cut snaps back to the user turn at index 2");
        // [summary] + u1 + a1 + t1.
        assert_eq!(out.len(), 4);
        assert_eq!(out[1]["content"].as_str().unwrap(), "u1");
    }

    #[test]
    fn checkpoint_reuse_rejects_out_of_range_or_zero_boundary() {
        let msgs = vec![
            serde_json::json!({"role": "user", "content": "u0"}),
            serde_json::json!({"role": "assistant", "content": "a0"}),
        ];
        assert!(apply_checkpoint_summary(&msgs, "## Goal\nx", 0).is_none());
        assert!(apply_checkpoint_summary(&msgs, "## Goal\nx", 99).is_none());
    }

    #[test]
    fn checkpoint_reuse_rejects_when_snap_collapses_to_zero() {
        // No user turn at/before the boundary → snap_backward returns 0 →
        // None (nothing whole to fold).
        let msgs = vec![
            serde_json::json!({"role": "assistant", "content": "a0"}),
            serde_json::json!({"role": "assistant", "content": "a1"}),
            serde_json::json!({"role": "assistant", "content": "a2"}),
        ];
        assert!(apply_checkpoint_summary(&msgs, "## Goal\nx", 2).is_none());
    }

    #[test]
    fn compute_window_partitions_correctly() {
        let msgs: Vec<Value> = (0..10)
            .map(|i| serde_json::json!({"role": "user", "content": format!("msg {i}")}))
            .collect();
        let (start, end) = compute_compress_window(&msgs, 2, 3);
        assert_eq!(start, 2);
        assert_eq!(end, 7);
    }

    /// dirge-89fm: with tool_use/tool_result pairs straddling the raw
    /// count-based cut, the window must snap to user boundaries so neither
    /// the head nor the tail is left holding half a pair.
    #[test]
    fn compute_window_snaps_off_tool_pairs() {
        // 0 system, 1 user, 2 assistant(tool_call), 3 toolResult,
        // 4 assistant(final), 5 user, 6 assistant(tool_call), 7 toolResult,
        // 8 assistant(final), 9 user, 10 assistant, 11 user(latest)
        let msgs = vec![
            serde_json::json!({"role": "system", "content": "s"}),
            serde_json::json!({"role": "user", "content": "u0"}),
            serde_json::json!({"role": "assistant", "content": "a0", "tool_calls": [{"id": "c0"}]}),
            serde_json::json!({"role": "toolResult", "toolCallId": "c0", "content": "t0"}),
            serde_json::json!({"role": "assistant", "content": "a0-final"}),
            serde_json::json!({"role": "user", "content": "u1"}),
            serde_json::json!({"role": "assistant", "content": "a1", "tool_calls": [{"id": "c1"}]}),
            serde_json::json!({"role": "toolResult", "toolCallId": "c1", "content": "t1"}),
            serde_json::json!({"role": "assistant", "content": "a1-final"}),
            serde_json::json!({"role": "user", "content": "u2"}),
            serde_json::json!({"role": "assistant", "content": "a2"}),
            serde_json::json!({"role": "user", "content": "u3 latest"}),
        ];
        let (start, end) = compute_compress_window(&msgs, 2, 2);
        // Both cuts land on user messages → no split pair.
        assert!(start < end);
        assert_eq!(msgs[start]["role"].as_str().unwrap(), "user");
        assert_eq!(msgs[end]["role"].as_str().unwrap(), "user");
        // Apply and verify the result has NO orphaned toolResult: every
        // toolResult is immediately preceded by an assistant message.
        let out = apply_summary(&msgs, "S", start, end);
        for (i, m) in out.iter().enumerate() {
            if m["role"].as_str() == Some("toolResult") {
                assert_eq!(
                    out[i - 1]["role"].as_str(),
                    Some("assistant"),
                    "toolResult at {i} must follow an assistant: {out:?}"
                );
            }
        }
        // And the message right before the summary is never a dangling
        // assistant tool_call (it precedes a user turn in the source).
        let summary_idx = out
            .iter()
            .position(|m| {
                m["content"]
                    .as_str()
                    .is_some_and(|c| c.starts_with(SUMMARY_PREFIX))
            })
            .unwrap();
        assert!(out[summary_idx - 1]["tool_calls"].is_null());
    }

    #[test]
    fn compute_window_short_list_returns_zero() {
        let msgs: Vec<Value> = (0..3)
            .map(|i| serde_json::json!({"role": "user", "content": format!("msg {i}")}))
            .collect();
        // 3 messages with head=2, tail=3 — too short.
        assert_eq!(compute_compress_window(&msgs, 2, 3), (0, 0));
    }

    #[test]
    fn rotate_session_id_prefix_and_length() {
        let id = rotate_session_id();
        assert!(id.starts_with("compacted-"));
        // "compacted-" (10) + 8 hex chars = 18.
        assert_eq!(id.len(), 18);
    }

    // ── full-wire integration: prompt → mock summarizer → applied ──

    /// LOOP-9: integration-style test exercising the full compaction
    /// wire end-to-end. Builds a long conversation, calls the prompt
    /// builder, runs a mock summarizer (no LLM), applies the result.
    /// Asserts the summary lands as a system message and the older
    /// turns are gone.
    #[tokio::test]
    async fn full_compaction_wire_with_mock_summarizer() {
        // Build a long conversation: system + 20 turns.
        let mut msgs: Vec<Value> = vec![
            serde_json::json!({"role": "system", "content": "you are an agent"}),
            serde_json::json!({"role": "user", "content": "initial task"}),
        ];
        for i in 0..18 {
            let role = if i % 2 == 0 { "assistant" } else { "user" };
            msgs.push(serde_json::json!({
                "role": role,
                "content": format!("turn {i} content with some length to make tokens"),
            }));
        }
        msgs.push(serde_json::json!({"role": "user", "content": "latest user request"}));

        let n_before = msgs.len();

        // 1. should_compress at the threshold.
        let tokens = estimate_messages_tokens(&msgs);
        // With small messages this is well under 75% — bypass via direct call.
        let _ = tokens;

        // 2. compute window.
        let (start, end) =
            compute_compress_window(&msgs, PROTECT_HEAD_DEFAULT, PROTECT_TAIL_DEFAULT);
        assert!(start < end);
        let middle = &msgs[start..end];
        assert!(!middle.is_empty());

        // 3. build prompt.
        let prompt = build_summary_prompt(
            middle,
            summary_budget(estimate_messages_tokens(middle)),
            None,
            None,
        );
        assert!(prompt.contains("TURNS TO SUMMARIZE"));
        // The window snaps to user boundaries (dirge-89fm), so it begins a
        // little after the raw head cut; assert it carries the mid
        // conversation rather than a fixed "turn 0".
        assert!(
            prompt.contains("turn "),
            "prompt should include the middle turns: {prompt}"
        );

        // 4. mock summarizer — implements SummarizeFn shape.
        let summarizer: SummarizeFn = Arc::new(|_prompt| {
            Box::pin(async move {
                Ok("## Active Task\nlatest user request\n\n\
                    ## Completed Actions\n1. turn 0\n2. turn 1\n\n\
                    ## Remaining Work\nfinish the task"
                    .to_string())
            })
        });
        let summary = summarizer(prompt.clone()).await.expect("summarizer ok");

        // 5. validate.
        assert!(validate_summary(&summary));

        // 6. apply.
        let compressed = apply_summary(&msgs, &summary, start, end);

        // dirge-89fm: the window now snaps to user-message boundaries, so
        // the protected head/tail can be a little larger than the raw
        // PROTECT_* counts. Assert the structure (head + 1 summary + tail =
        // start + 1 + (n - end)) rather than the exact pre-snap size.
        assert_eq!(compressed.len(), start + 1 + (n_before - end));
        // Original was much longer.
        assert!(compressed.len() < n_before);
        // The single summary message sits at `start`, carrying SUMMARY_PREFIX.
        let summary_msg = &compressed[start];
        assert_eq!(summary_msg["role"].as_str().unwrap(), "system");
        let body = summary_msg["content"].as_str().unwrap();
        assert!(body.starts_with(SUMMARY_PREFIX));
        assert!(body.contains("## Active Task"));
        // Exactly one summary marker (no duplication).
        assert_eq!(
            compressed
                .iter()
                .filter(|m| m["content"]
                    .as_str()
                    .is_some_and(|c| c.starts_with(SUMMARY_PREFIX)))
                .count(),
            1
        );
        // The latest user message is preserved in the tail.
        let last = compressed.last().unwrap();
        assert_eq!(last["content"].as_str().unwrap(), "latest user request");
        // Session id rotates.
        let new_id = rotate_session_id();
        assert!(new_id.starts_with("compacted-"));
    }
}