wcode_harness/

compaction.rs

//! Context compaction — deciding *when* to compact and *where* to cut.
//!
//! Compaction summarizes an old prefix of the conversation and keeps a recent
//! suffix. The prefix is discarded from what the model sees, so only the kept
//! suffix must be *valid* on the wire. The one invariant that makes a suffix
//! invalid is an orphaned `ToolResult`: a kept tool result whose matching
//! `ToolCall` (which lives in a preceding `Assistant` message) was summarized
//! away. Providers reject that as an unmatched tool result.
//!
//! [`safe_cut`] returns the earliest index at which the conversation may be
//! split so the kept suffix `ctx[cut..]` is self-consistent. It only ever moves
//! the cut *earlier* than requested — summarizing less rather than breaking the
//! pairing — and returns `0` when no boundary pairs everything.
//!
//! The reverse direction needs no check: the agent loop never leaves a
//! `ToolCall` without its `ToolResult`, and a `ToolResult` always follows its
//! `ToolCall`, so a kept call's results (later indices) are kept too.
//!
//! [`CompactionPolicy`] decides *when* to compact (the trigger) and *how much*
//! recent history to keep. Retention is absolute — a token budget plus a turn
//! floor — not a fraction of the window, so it means the same thing on an 8k
//! and a 1M model. The *trigger* uses the provider-reported input-token count
//! (never an estimate); only the cut boundary uses [`estimate_tokens`].

use std::collections::HashSet;

use futures::StreamExt as _;

use crate::event::LlmStreamEvent;
use crate::message::{AgentMessage, ContentBlock, Usage};
use crate::session::Session;
use crate::streamfn::{LlmOpts, StreamFn};

/// Earliest index at which `ctx` may be split so the kept suffix `ctx[cut..]`
/// has no orphaned tool result.
///
/// `initial_cut` is the caller's preferred split (e.g. `ctx.len() - keep`); the
/// result is always `<= initial_cut` (clamped to `ctx.len()`). Returns `0` —
/// keep everything — when walking all the way back still can't pair a kept
/// result with its call.
pub fn safe_cut(ctx: &[AgentMessage], initial_cut: usize) -> usize {
    let cut = initial_cut.min(ctx.len());

    let mut available: HashSet<&str> = HashSet::new();
    let mut missing: HashSet<&str> = HashSet::new();
    for msg in &ctx[cut..] {
        absorb(msg, &mut available, &mut missing);
    }
    if missing.is_empty() {
        return cut;
    }

    // Grow the kept suffix backward, one message at a time, until every kept
    // result has its call. Adding a call removes the corresponding missing
    // result; adding a result whose call isn't yet seen records a new one.
    for idx in (0..cut).rev() {
        absorb(&ctx[idx], &mut available, &mut missing);
        if missing.is_empty() {
            return idx;
        }
    }

    0
}

/// Fold one message's tool calls/results into the pairing trackers.
fn absorb<'a>(
    msg: &'a AgentMessage,
    available: &mut HashSet<&'a str>,
    missing: &mut HashSet<&'a str>,
) {
    match msg {
        AgentMessage::Assistant { content, .. } => {
            for block in content {
                if let ContentBlock::ToolCall { id, .. } = block {
                    missing.remove(id.as_str());
                    available.insert(id.as_str());
                }
            }
        }
        AgentMessage::ToolResult { tool_call_id, .. }
            if !available.contains(tool_call_id.as_str()) =>
        {
            missing.insert(tool_call_id.as_str());
        }
        _ => {}
    }
}

/// Tokens held free below the working ceiling by default: the trigger fires
/// once fewer than this many remain.
pub const DEFAULT_MIN_REMAINING: u64 = 16_384;

/// Recent context kept verbatim by default, in tokens — an absolute budget
/// (pi-style), independent of the model's window size.
pub const DEFAULT_KEEP_RECENT_TOKENS: u64 = 20_000;

/// Complete turns always kept by default, whatever their token size.
pub const DEFAULT_KEEP_RECENT_TURNS: usize = 2;

/// When to compact and how much recent context to retain.
///
/// The model window is not stored here: it comes from [`crate::limits`] (the
/// model's advertised context) or an override, and is passed to
/// [`CompactionPolicy::should_compact`]. `budget` is an optional working
/// ceiling (cost control) that defaults to the window; `min_remaining` is the
/// "minimum remaining context" trigger; the `keep_recent_*` fields are the
/// absolute recent history retained after compacting.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct CompactionPolicy {
    /// Context-window override, consulted only when [`crate::limits`] doesn't
    /// know the model (see [`CompactionPolicy::context_window`]). `None` =
    /// derive it from the catalog, or fall back to a default.
    pub window: Option<u64>,
    /// Working ceiling in tokens. `None` = the model window. Capped to the
    /// window at trigger time, so it can never disable window safety.
    pub budget: Option<u64>,
    /// Compact once fewer than this many tokens remain below the ceiling
    /// (`used >= ceiling - min_remaining`).
    pub min_remaining: u64,
    /// Recent context kept verbatim after compacting, in tokens.
    pub keep_recent_tokens: u64,
    /// Complete turns always kept, whatever their token size.
    pub keep_recent_turns: usize,
}

impl Default for CompactionPolicy {
    fn default() -> Self {
        Self {
            window: None,
            budget: None,
            min_remaining: DEFAULT_MIN_REMAINING,
            keep_recent_tokens: DEFAULT_KEEP_RECENT_TOKENS,
            keep_recent_turns: DEFAULT_KEEP_RECENT_TURNS,
        }
    }
}

impl CompactionPolicy {
    /// The working ceiling: the configured budget, capped to the model window.
    pub fn ceiling(&self, window: u64) -> u64 {
        self.budget.map_or(window, |b| b.min(window))
    }

    /// The context window for `model` served from `base_url`: the model's
    /// advertised context (via [`crate::limits`]) when known, else the
    /// configured [`CompactionPolicy::window`], else
    /// [`crate::limits::DEFAULT_CONTEXT_WINDOW`].
    pub fn context_window(&self, base_url: Option<&str>, model: &str) -> u64 {
        crate::limits::model_limit(base_url, model)
            .map(|l| l.context)
            .or(self.window)
            .unwrap_or(crate::limits::DEFAULT_CONTEXT_WINDOW)
    }

    /// True once the provider-reported `used` input tokens leave less than
    /// `min_remaining` free below the ceiling.
    pub fn should_compact(&self, window: u64, used: u64) -> bool {
        used >= self.ceiling(window).saturating_sub(self.min_remaining)
    }
}

/// Rough per-message token estimate (~4 chars/token plus a small fixed
/// overhead). Used only to place the cut boundary; the compaction *trigger*
/// uses provider-reported token counts, never this.
pub fn estimate_tokens(msg: &AgentMessage) -> u64 {
    let bytes = match msg {
        AgentMessage::User { content } | AgentMessage::Assistant { content, .. } => {
            content.iter().map(block_bytes).sum()
        }
        AgentMessage::ToolResult {
            output,
            tool_call_id,
            name,
            ..
        } => (output.len() + tool_call_id.len() + name.len()) as u64,
    };
    bytes / 4 + 4
}

fn block_bytes(block: &ContentBlock) -> u64 {
    match block {
        ContentBlock::Text { text } | ContentBlock::Thinking { text } => text.len() as u64,
        ContentBlock::ToolCall {
            name, arguments, ..
        } => (name.len() + arguments.to_string().len()) as u64,
    }
}

/// Oldest index to keep when retaining `keep_budget` tokens of the most recent
/// history, adjusted by [`safe_cut`] so the kept suffix never orphans a tool
/// result.
#[cfg(test)]
fn cut_for_budget(ctx: &[AgentMessage], keep_budget: u64) -> usize {
    safe_cut(ctx, token_cut(ctx, keep_budget))
}

/// Newest-first accumulation of [`estimate_tokens`] until `budget` is met:
/// the rough index where the retained verbatim tail begins.
fn token_cut(ctx: &[AgentMessage], budget: u64) -> usize {
    let mut acc = 0u64;
    let mut idx = ctx.len();
    while idx > 0 && acc < budget {
        idx -= 1;
        acc += estimate_tokens(&ctx[idx]);
    }
    idx
}

/// Index of the user message that begins the `turns`-th-most-recent turn, so
/// keeping from here retains at least `turns` complete turns. `0` when the
/// conversation has fewer than `turns` user messages; `ctx.len()` (no floor)
/// when `turns` is 0.
fn turn_start_cut(ctx: &[AgentMessage], turns: usize) -> usize {
    if turns == 0 {
        return ctx.len();
    }
    let mut seen = 0usize;
    for idx in (0..ctx.len()).rev() {
        if matches!(&ctx[idx], AgentMessage::User { .. }) {
            seen += 1;
            if seen == turns {
                return idx;
            }
        }
    }
    0
}

/// Cut for `policy`: keep the newest messages up to `keep_recent_tokens` but
/// never fewer than `keep_recent_turns` turns, made tool-safe by [`safe_cut`].
pub fn cut_for_policy(ctx: &[AgentMessage], policy: &CompactionPolicy) -> usize {
    let by_tokens = token_cut(ctx, policy.keep_recent_tokens);
    let by_turns = turn_start_cut(ctx, policy.keep_recent_turns);
    safe_cut(ctx, by_tokens.min(by_turns))
}

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

    fn user(text: &str) -> AgentMessage {
        AgentMessage::user_text(text)
    }

    fn assistant_text(text: &str) -> AgentMessage {
        AgentMessage::Assistant {
            content: vec![ContentBlock::Text { text: text.into() }],
            stop_reason: StopReason::Stop,
            usage: None,
            model: None,
        }
    }

    fn assistant_calls(ids: &[&str]) -> AgentMessage {
        AgentMessage::Assistant {
            content: ids
                .iter()
                .map(|id| ContentBlock::ToolCall {
                    id: (*id).into(),
                    name: "t".into(),
                    arguments: serde_json::json!({}),
                })
                .collect(),
            stop_reason: StopReason::ToolUse,
            usage: None,
            model: None,
        }
    }

    fn tool_result(id: &str) -> AgentMessage {
        AgentMessage::ToolResult {
            tool_call_id: id.into(),
            name: "t".into(),
            output: "o".into(),
            is_error: false,
        }
    }

    #[test]
    fn empty_and_out_of_range_are_clamped() {
        assert_eq!(safe_cut(&[], 0), 0);
        assert_eq!(safe_cut(&[], 5), 0);
        let ctx = vec![user("a")];
        assert_eq!(safe_cut(&ctx, 0), 0);
        assert_eq!(safe_cut(&ctx, 1), 1);
        assert_eq!(safe_cut(&ctx, 9), 1, "clamps to ctx.len()");
    }

    #[test]
    fn tool_free_context_keeps_initial_cut() {
        let ctx = vec![user("a"), assistant_text("x"), user("b"), assistant_text("y")];
        assert_eq!(safe_cut(&ctx, 2), 2);
        assert_eq!(safe_cut(&ctx, 3), 3);
    }

    #[test]
    fn cut_inside_a_tool_exchange_pulls_in_the_call() {
        // [user, assistant(c1), result(c1), assistant]
        let ctx = vec![
            user("q"),
            assistant_calls(&["c1"]),
            tool_result("c1"),
            assistant_text("done"),
        ];
        // Cutting at the result (2) would orphan it -> extend back to the call.
        assert_eq!(safe_cut(&ctx, 2), 1);
        assert_eq!(safe_cut(&ctx, 3), 3, "cut past the exchange drops it whole");
        // A cut already at the call boundary is safe as-is.
        assert_eq!(safe_cut(&ctx, 1), 1);
        // A cut at the trailing boundary keeps only the final assistant.
        assert_eq!(safe_cut(&ctx, 4), 4);
    }

    #[test]
    fn multi_call_assistant_is_pulled_in_whole() {
        // One assistant issuing two calls; results follow in order.
        let ctx = vec![
            user("q"),
            assistant_calls(&["c1", "c2"]),
            tool_result("c1"),
            tool_result("c2"),
            assistant_text("done"),
        ];
        // Cutting between the two results pulls the shared assistant back.
        assert_eq!(safe_cut(&ctx, 3), 1);
        assert_eq!(safe_cut(&ctx, 2), 1);
        // Cutting at a clean turn boundary stays put.
        assert_eq!(safe_cut(&ctx, 5), 5);
    }

    #[test]
    fn cut_at_a_user_boundary_after_a_completed_turn_is_kept() {
        let ctx = vec![
            user("q1"),
            assistant_calls(&["c1"]),
            tool_result("c1"),
            user("q2"),
            assistant_text("done"),
        ];
        assert_eq!(safe_cut(&ctx, 3), 3, "the q2 boundary needs no adjustment");
    }

    #[test]
    fn unpaired_result_walks_all_the_way_back() {
        // A result whose call is absent from the whole context cannot be
        // paired: a cut that keeps it gives up, one that drops it is fine.
        let ctx = vec![user("q"), tool_result("ghost"), assistant_text("x")];
        assert_eq!(safe_cut(&ctx, 1), 0);
        assert_eq!(safe_cut(&ctx, 2), 2, "dropping the orphan is fine");
    }

    #[test]
    fn kept_call_with_its_results_needs_no_extension() {
        // A kept call whose results are also kept (they follow it) is valid.
        let ctx = vec![
            user("q"),
            assistant_calls(&["c1"]),
            tool_result("c1"),
            user("next"),
        ];
        assert_eq!(safe_cut(&ctx, 3), 3, "keeps only the trailing user message");
        assert_eq!(safe_cut(&ctx, 1), 1, "keeps the whole tool exchange");
    }
}

#[cfg(test)]
mod policy_tests {
    use super::*;
    use crate::message::StopReason;

    fn user(text: &str) -> AgentMessage {
        AgentMessage::user_text(text)
    }

    fn assistant_text(text: &str) -> AgentMessage {
        AgentMessage::Assistant {
            content: vec![ContentBlock::Text { text: text.into() }],
            stop_reason: StopReason::Stop,
            usage: None,
            model: None,
        }
    }

    fn assistant_tool_call(id: &str, args: String) -> AgentMessage {
        AgentMessage::Assistant {
            content: vec![ContentBlock::ToolCall {
                id: id.into(),
                name: "t".into(),
                arguments: serde_json::json!({ "d": args }),
            }],
            stop_reason: StopReason::ToolUse,
            usage: None,
            model: None,
        }
    }

    fn tool_result(id: &str, output: String) -> AgentMessage {
        AgentMessage::ToolResult {
            tool_call_id: id.into(),
            name: "t".into(),
            output,
            is_error: false,
        }
    }

    #[test]
    fn trigger_uses_min_remaining_and_budget() {
        // No budget: the ceiling is the model window.
        let p = CompactionPolicy {
            min_remaining: 100,
            ..Default::default()
        };
        assert!(!p.should_compact(1000, 899));
        assert!(p.should_compact(1000, 900), "window - min_remaining");
        assert!(p.should_compact(1000, 5000), "past the window still triggers");

        // A working budget below the window triggers far earlier.
        let p = CompactionPolicy {
            budget: Some(500),
            min_remaining: 100,
            ..Default::default()
        };
        assert!(!p.should_compact(1_000_000, 399));
        assert!(p.should_compact(1_000_000, 400));

        // A budget above the window is capped to the window.
        let p = CompactionPolicy {
            budget: Some(2_000_000),
            min_remaining: 100,
            ..Default::default()
        };
        assert!(!p.should_compact(1000, 899));
        assert!(p.should_compact(1000, 900));
    }

    #[test]
    fn ceiling_is_budget_capped_to_window() {
        let p = CompactionPolicy {
            budget: Some(500),
            ..Default::default()
        };
        assert_eq!(p.ceiling(1_000_000), 500);
        let p = CompactionPolicy {
            budget: Some(2_000_000),
            ..Default::default()
        };
        assert_eq!(p.ceiling(1_000_000), 1_000_000);
        let p = CompactionPolicy::default();
        assert_eq!(p.ceiling(1234), 1234);
    }

    #[test]
    fn estimate_grows_with_content_and_counts_tool_output() {
        assert!(estimate_tokens(&user("hello there")) > estimate_tokens(&user("hi")));
        let big = AgentMessage::ToolResult {
            tool_call_id: "c".into(),
            name: "t".into(),
            output: "x".repeat(400),
            is_error: false,
        };
        assert!(estimate_tokens(&big) >= 100, "~400 chars -> ~100 tokens");
    }

    #[test]
    fn cut_keeps_the_newest_within_budget() {
        // Six ~54-token messages; a 120-token budget keeps the last three.
        let ctx = vec![
            user(&"a".repeat(200)),
            assistant_text(&"b".repeat(200)),
            user(&"c".repeat(200)),
            assistant_text(&"d".repeat(200)),
            user(&"e".repeat(200)),
            assistant_text(&"f".repeat(200)),
        ];
        assert_eq!(cut_for_budget(&ctx, 120), 3, "keeps the newest that fit");
    }

    #[test]
    fn cut_boundary_is_made_tool_safe() {
        // Budget lands on a tool result, so safe_cut pulls its call in.
        let ctx = vec![
            user("q"),
            assistant_tool_call("c1", "x".repeat(400)),
            tool_result("c1", "y".repeat(400)),
            assistant_text("done"),
        ];
        // A tiny budget keeps only the trailing assistant (no adjustment).
        assert_eq!(cut_for_budget(&ctx, 1), ctx.len() - 1);
        // A budget that reaches the result extends back to the call.
        assert_eq!(cut_for_budget(&ctx, 120), 1, "the call joins the suffix");
    }

    #[test]
    fn cut_for_policy_keeps_recent_tokens() {
        let ctx = vec![
            user(&"a".repeat(200)),
            assistant_text(&"b".repeat(200)),
            user(&"c".repeat(200)),
            assistant_text(&"d".repeat(200)),
        ];
        // ~54 tokens each; a 100-token budget keeps the newest two.
        let p = CompactionPolicy {
            keep_recent_tokens: 100,
            keep_recent_turns: 0,
            ..Default::default()
        };
        assert_eq!(cut_for_policy(&ctx, &p), 2);
    }

    #[test]
    fn cut_for_policy_keeps_at_least_the_turn_floor() {
        // A tiny token budget must not drop below the last two turns.
        let ctx = vec![
            user("u1"),
            assistant_text("a1"),
            user("u2"),
            assistant_text("a2"),
            user("u3"),
            assistant_text("a3"),
        ];
        let p = CompactionPolicy {
            keep_recent_tokens: 0,
            keep_recent_turns: 2,
            ..Default::default()
        };
        assert_eq!(cut_for_policy(&ctx, &p), 2, "last two turns start at u2");
    }
}

/// System prompt for the summarizer.
const SUMMARY_PROMPT: &str = "\
You are summarizing a coding-agent conversation so the agent can continue the \
work later with this summary in place of the summarized messages. Write \
concise plain prose under these headings, preserving concrete details (file \
paths, names, decisions, errors) and omitting pleasantries:
- Context: what we are working on and why.
- What we did: key actions taken, files changed, problems solved.
- Current state: what works, what is broken, what is next.
- User preferences: requirements or decisions the user made. The turns you \
summarize are not lost: they remain in this session's transcript and can be \
searched later with the session_search tool (use scope=\"current\").";

/// Per-message/-block character cap when rendering a conversation for the
/// summarizer, so one giant tool result can't crowd out the rest.
const MAX_BLOCK_CHARS: usize = 4_000;

/// Total rendered-conversation cap (~100k tokens); the middle is elided when
/// exceeded, keeping the start and the most recent end.
const MAX_CONVERSATION_CHARS: usize = 400_000;

/// A produced summary and the usage that generated it.
#[derive(Debug, Clone)]
pub struct Summary {
    pub text: String,
    pub usage: Option<Usage>,
}

/// What a compaction pass did.
#[derive(Debug, Clone)]
pub enum CompactOutcome {
    /// Not enough history to bother summarizing.
    NothingToDo,
    /// Summarized `summarized` messages; `kept` were retained; a summary
    /// message now stands in their place.
    Done {
        summarized: usize,
        kept: usize,
        usage: Option<Usage>,
    },
}

/// Summarize the older prefix of `ctx` in place: keep the newest messages per
/// `policy`, replace the rest with a summary message, and — when a session is
/// open — persist the boundary so a resume rebuilds the same view.
///
/// Ordered so a failure leaves `ctx` untouched: the session write happens
/// *before* the in-memory splice. `Ok(NothingToDo)` when the cut keeps
/// everything (nothing worth summarizing).
pub async fn compact_ctx(
    ctx: &mut Vec<AgentMessage>,
    policy: &CompactionPolicy,
    stream_fn: &StreamFn,
    opts: &LlmOpts,
    instructions: Option<&str>,
    session: Option<&mut Session>,
) -> Result<CompactOutcome, String> {
    let cut = cut_for_policy(ctx, policy);
    if cut == 0 {
        return Ok(CompactOutcome::NothingToDo);
    }
    let prefix = ctx[..cut].to_vec();
    let summary = summarize(stream_fn, opts, &prefix, instructions).await?;
    let kept = ctx.len() - cut;

    if let Some(session) = session {
        let tokens_before = ctx.iter().rev().find_map(|m| match m {
            AgentMessage::Assistant { usage: Some(u), .. } => Some(u.input_tokens),
            _ => None,
        });
        session
            .record_compaction(&summary.text, kept, tokens_before, summary.usage)
            .map_err(|e| e.to_string())?;
    }

    let mut next = Vec::with_capacity(kept + 1);
    next.push(summary_message(&summary.text));
    next.extend_from_slice(&ctx[cut..]);
    *ctx = next;

    Ok(CompactOutcome::Done {
        summarized: cut,
        kept,
        usage: summary.usage,
    })
}

/// The message injected in place of a summarized prefix.
pub fn summary_message(summary: &str) -> AgentMessage {
    AgentMessage::summary(summary)
}

/// Summarize `conversation` with the configured LLM, optionally focused by
/// `instructions` (the `/compact <prompt>` argument). Drives the same
/// [`StreamFn`] the agent uses, collecting the streamed text. Returns an error
/// string on a stream error, or when the summarizer produces no text.
pub async fn summarize(
    stream_fn: &StreamFn,
    opts: &LlmOpts,
    conversation: &[AgentMessage],
    instructions: Option<&str>,
) -> Result<Summary, String> {
    let mut system = String::from(SUMMARY_PROMPT);
    if let Some(extra) = instructions.map(str::trim).filter(|s| !s.is_empty()) {
        system.push_str("\n\nAdditional focus for this summary: ");
        system.push_str(extra);
    }
    let user = format!(
        "Summarize this conversation:\n\n{}",
        render_conversation(conversation)
    );

    let mut stream = stream_fn(&[AgentMessage::user_text(user)], &system, &[], opts);
    let mut text = String::new();
    let mut usage = None;
    while let Some(event) = stream.next().await {
        match event {
            LlmStreamEvent::TextDelta(delta) => text.push_str(&delta),
            LlmStreamEvent::Done { usage: u, .. } => {
                usage = u;
                break;
            }
            LlmStreamEvent::Error { message, .. } => return Err(message),
            _ => {}
        }
    }

    let text = text.trim().to_string();
    if text.is_empty() {
        return Err("summarizer produced no text".to_string());
    }
    Ok(Summary { text, usage })
}

/// Render a conversation to plain text for the summarizer.
fn render_conversation(messages: &[AgentMessage]) -> String {
    let mut out = String::new();
    for msg in messages {
        match msg {
            AgentMessage::User { content } => {
                out.push_str("USER:\n");
                render_blocks(&mut out, content);
            }
            AgentMessage::Assistant { content, .. } => {
                out.push_str("ASSISTANT:\n");
                render_blocks(&mut out, content);
            }
            AgentMessage::ToolResult {
                tool_call_id,
                name,
                output,
                is_error,
            } => {
                out.push_str("TOOL RESULT ");
                out.push_str(name);
                out.push_str(" (");
                out.push_str(tool_call_id);
                out.push(')');
                if *is_error {
                    out.push_str(" [error]");
                }
                out.push_str(":\n");
                push_capped(&mut out, output, MAX_BLOCK_CHARS);
                out.push('\n');
            }
        }
        out.push('\n');
    }
    cap_total(out)
}

fn render_blocks(out: &mut String, content: &[ContentBlock]) {
    for block in content {
        match block {
            ContentBlock::Text { text } => {
                push_capped(out, text, MAX_BLOCK_CHARS);
                out.push('\n');
            }
            ContentBlock::Thinking { text } => {
                out.push_str("[thinking] ");
                push_capped(out, text, MAX_BLOCK_CHARS);
                out.push('\n');
            }
            ContentBlock::ToolCall {
                id,
                name,
                arguments,
            } => {
                out.push_str("[tool call ");
                out.push_str(name);
                out.push_str(" (");
                out.push_str(id);
                out.push_str(") args: ");
                push_capped(out, &arguments.to_string(), MAX_BLOCK_CHARS);
                out.push_str("]\n");
            }
        }
    }
}

/// Append `s`, truncated on a char boundary to `max` bytes if longer.
fn push_capped(out: &mut String, s: &str, max: usize) {
    if s.len() <= max {
        out.push_str(s);
        return;
    }
    let mut end = max;
    while end > 0 && !s.is_char_boundary(end) {
        end -= 1;
    }
    out.push_str(&s[..end]);
    out.push_str(&format!("\n… [{} chars truncated]", s.len() - end));
}

/// Keep the head and tail of an over-long rendering, eliding the middle.
fn cap_total(s: String) -> String {
    if s.len() <= MAX_CONVERSATION_CHARS {
        return s;
    }
    let half = MAX_CONVERSATION_CHARS / 2;
    let mut head_end = half;
    while head_end > 0 && !s.is_char_boundary(head_end) {
        head_end -= 1;
    }
    let mut tail_start = s.len() - half;
    while tail_start < s.len() && !s.is_char_boundary(tail_start) {
        tail_start += 1;
    }
    format!(
        "{}\n… [conversation truncated to fit the summarizer] …\n{}",
        &s[..head_end],
        &s[tail_start..]
    )
}

#[cfg(test)]
mod summarize_tests {
    use super::*;
    use crate::message::StopReason;
    use crate::streamfn::LlmStream;
    use std::sync::{Arc, Mutex};

    /// A StreamFn that records the (system, user) prompt it was handed and
    /// replays a fixed script of events.
    fn scripted(events: Vec<LlmStreamEvent>, seen: Arc<Mutex<Vec<String>>>) -> StreamFn {
        Arc::new(move |ctx, system, _tools, _opts| {
            seen.lock()
                .unwrap()
                .push(format!("{system}\n---\n{}", ctx[0].as_text()));
            Box::pin(futures::stream::iter(events.clone())) as LlmStream
        })
    }

    fn done() -> LlmStreamEvent {
        LlmStreamEvent::Done {
            stop_reason: StopReason::Stop,
            usage: Some(Usage::default()),
        }
    }

    #[tokio::test]
    async fn summarize_collects_text_usage_and_rendered_conversation() {
        let seen = Arc::new(Mutex::new(Vec::new()));
        let stream_fn = scripted(
            vec![
                LlmStreamEvent::TextDelta("Context: ".into()),
                LlmStreamEvent::TextDelta("do X".into()),
                done(),
            ],
            seen.clone(),
        );
        let convo = vec![
            AgentMessage::user_text("please do X"),
            AgentMessage::user_text("and Y"),
        ];
        let s = summarize(&stream_fn, &LlmOpts::default(), &convo, None)
            .await
            .unwrap();
        assert_eq!(s.text, "Context: do X");
        assert!(s.usage.is_some());
        let prompt = seen.lock().unwrap()[0].clone();
        assert!(prompt.contains("USER:"), "conversation reached the prompt");
        assert!(prompt.contains("please do X"));
    }

    #[tokio::test]
    async fn summarize_folds_instructions_into_the_system_prompt() {
        let seen = Arc::new(Mutex::new(Vec::new()));
        let stream_fn = scripted(vec![LlmStreamEvent::TextDelta("x".into()), done()], seen.clone());
        summarize(
            &stream_fn,
            &LlmOpts::default(),
            &[AgentMessage::user_text("hi")],
            Some("focus on the API design"),
        )
        .await
        .unwrap();
        let prompt = seen.lock().unwrap()[0].clone();
        assert!(prompt.contains("Additional focus for this summary: focus on the API design"));
    }

    #[tokio::test]
    async fn summarize_surfaces_stream_error() {
        let stream_fn = scripted(
            vec![LlmStreamEvent::Error {
                message: "boom".into(),
                fatal: true,
            }],
            Arc::new(Mutex::new(Vec::new())),
        );
        let err = summarize(
            &stream_fn,
            &LlmOpts::default(),
            &[AgentMessage::user_text("x")],
            None,
        )
        .await
        .unwrap_err();
        assert_eq!(err, "boom");
    }

    #[tokio::test]
    async fn summarize_rejects_empty_output() {
        let stream_fn = scripted(vec![done()], Arc::new(Mutex::new(Vec::new())));
        let err = summarize(
            &stream_fn,
            &LlmOpts::default(),
            &[AgentMessage::user_text("x")],
            None,
        )
        .await
        .unwrap_err();
        assert!(err.contains("no text"), "got: {err}");
    }

    #[test]
    fn render_truncates_a_giant_tool_output() {
        let big = AgentMessage::ToolResult {
            tool_call_id: "c".into(),
            name: "t".into(),
            output: "x".repeat(10_000),
            is_error: false,
        };
        let rendered = render_conversation(&[big]);
        assert!(rendered.contains("chars truncated"));
        assert!(rendered.len() < 5_000, "len was {}", rendered.len());
    }

    #[test]
    fn summary_message_is_a_user_message_with_a_header() {
        let m = summary_message("did things");
        assert!(matches!(m, AgentMessage::User { .. }));
        assert!(m.as_text().contains("did things"));
        assert!(m.as_text().starts_with("Summary"));
    }
}

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

    #[test]
    fn context_window_prefers_catalog_then_override_then_default() {
        // A known opencode-go model: the catalog wins over the override.
        let p = CompactionPolicy {
            window: Some(999),
            ..Default::default()
        };
        assert_eq!(
            p.context_window(Some("https://opencode.ai/zen/go/v1"), "deepseek-v4.1-flash"),
            1_000_000
        );

        // An unknown model: the configured override is used.
        let p = CompactionPolicy {
            window: Some(999),
            ..Default::default()
        };
        assert_eq!(p.context_window(Some("https://api.example.com/v1"), "who"), 999);

        // Unknown model, no override: the conservative default.
        let p = CompactionPolicy::default();
        assert_eq!(
            p.context_window(None, "who"),
            crate::limits::DEFAULT_CONTEXT_WINDOW
        );
    }
}