koda_core/

compact.rs

//! Session compaction — summarize old messages to reclaim context.
//!
//! When the conversation grows long, compaction replaces older messages with
//! a concise summary, freeing context window space for new work.
//!
//! ## How it works
//!
//! 1. **Trigger**: user types `/compact`, or auto-compact fires at ~80% context usage
//! 2. **Summarization**: a cheap model (Standard tier) generates a summary of old messages
//! 3. **Replacement**: old messages are archived in the DB, replaced by the summary
//! 4. **Result**: context usage drops, conversation continues with full history awareness
//!
//! ## Auto-compaction
//!
//! When context usage exceeds the threshold (configurable, default ~80%),
//! compaction runs automatically before the next inference call. The user
//! sees a brief "⚡ Compacting..." indicator.
//!
//! ## What's preserved
//!
//! - Summary of all prior conversation and decisions
//! - Progress tracking entries (survive compaction via DB metadata)
//! - Memory facts (injected from `MEMORY.md`, not from conversation)
//! - File ownership state (tracked in SQLite, not in messages)
//!
//! Pure logic, zero UI dependencies. Returns structured results
//! for the caller (TUI or headless) to render however it likes.
//!
//! Compaction uses a cheap model (Standard tier) when available,
//! falling back to the main model. Summarization is a simple task
//! that doesn't need frontier-class reasoning.

use crate::config::ModelSettings;
use crate::db::Database;
use crate::persistence::Persistence;
use crate::providers::{ChatMessage, LlmProvider};
use anyhow::{Result, bail};
use std::sync::Arc;
use std::sync::atomic::{AtomicU32, Ordering};
use tokio::sync::RwLock;

/// Minimum number of recent messages to keep verbatim during compaction.
pub const COMPACT_PRESERVE_COUNT: usize = 4;

/// Fraction of history to compact in partial mode (compact oldest half).
const PARTIAL_COMPACT_FRACTION: f64 = 0.5;

/// Below this message count, always do full compaction (partial is overhead).
const PARTIAL_COMPACT_THRESHOLD: usize = 12;

/// Stop auto-compacting after this many consecutive failures.
/// Prevents wasting an API call every turn when compaction is stuck
/// (e.g. history too large for the model, persistent API errors).
const MAX_CONSECUTIVE_FAILURES: u32 = 3;

/// Global consecutive failure counter. Shared across the session.
static CONSECUTIVE_FAILURES: AtomicU32 = AtomicU32::new(0);

/// Reset the failure counter (call after successful compaction or new session).
pub fn reset_compact_failures() {
    CONSECUTIVE_FAILURES.store(0, Ordering::Relaxed);
}

/// Check if the circuit breaker is tripped.
pub fn is_compact_circuit_broken() -> bool {
    CONSECUTIVE_FAILURES.load(Ordering::Relaxed) >= MAX_CONSECUTIVE_FAILURES
}

/// Record a compaction failure. Returns true if the circuit breaker just tripped.
pub fn record_compact_failure() -> bool {
    let prev = CONSECUTIVE_FAILURES.fetch_add(1, Ordering::Relaxed);
    prev + 1 >= MAX_CONSECUTIVE_FAILURES
}

/// Record a compaction success — resets the failure counter.
fn record_compact_success() {
    reset_compact_failures();
}

/// Maximum number of head-truncation retries when history is too large.
const MAX_TRUNCATION_RETRIES: usize = 3;

/// Fraction of messages to drop on each truncation attempt.
const TRUNCATION_DROP_FRACTION: f64 = 0.2;

/// Result of a successful compaction.
#[derive(Debug)]
pub struct CompactResult {
    /// Number of messages deleted from the database.
    pub deleted: usize,
    /// Estimated tokens in the summary.
    pub summary_tokens: usize,
}

/// Why compaction was skipped (not an error, just a precondition).
#[derive(Debug)]
pub enum CompactSkip {
    /// Session has unresolved tool calls — can't compact safely.
    PendingToolCalls,
    /// Session is too short to compact (contains N messages).
    TooShort(usize),
    /// History is too large for the current model to summarize without data loss.
    /// The user should switch to a model with a larger context window or start a new session.
    HistoryTooLarge,
}

/// Attempt to compact a session.
///
/// Returns `Ok(Ok(result))` on success, `Ok(Err(skip))` if a
/// precondition prevented compaction, or `Err(e)` on failure.
pub async fn compact_session(
    db: &Database,
    session_id: &str,
    max_context_tokens: usize,
    model_settings: &crate::config::ModelSettings,
    provider: &Arc<RwLock<Box<dyn LlmProvider>>>,
) -> Result<std::result::Result<CompactResult, CompactSkip>> {
    let prov = provider.read().await;
    compact_session_with_provider(db, session_id, max_context_tokens, model_settings, &**prov).await
}

/// Core compaction logic — accepts `&dyn LlmProvider` directly.
///
/// Uses partial compaction for longer sessions (≥12 messages): only the oldest
/// half of messages are summarized and archived, preserving more recent context
/// verbatim. Short sessions fall back to full compaction (keep last 4).
///
/// Used by the inference loop for pre-flight compaction (where we already
/// have a `&dyn LlmProvider` and don't need the Arc<RwLock<>> wrapper).
pub async fn compact_session_with_provider(
    db: &Database,
    session_id: &str,
    max_context_tokens: usize,
    model_settings: &crate::config::ModelSettings,
    provider: &dyn LlmProvider,
) -> Result<std::result::Result<CompactResult, CompactSkip>> {
    // Check preconditions
    if db.has_pending_tool_calls(session_id).await.unwrap_or(false) {
        return Ok(Err(CompactSkip::PendingToolCalls));
    }

    let history = db.load_context(session_id).await?;

    if history.len() < 4 {
        return Ok(Err(CompactSkip::TooShort(history.len())));
    }

    // Decide how many messages to preserve (partial vs full compaction).
    // Partial: compact the oldest half, keep the newest half.
    // Full: compact everything except the last COMPACT_PRESERVE_COUNT.
    let preserve_count = compute_preserve_count(history.len());

    let compact_count = history.len().saturating_sub(preserve_count);
    if compact_count == 0 {
        return Ok(Err(CompactSkip::TooShort(history.len())));
    }

    // Only summarize the messages being compacted, not the ones we're keeping.
    let to_compact = &history[..compact_count];
    let conversation_text = build_conversation_text(to_compact);

    tracing::info!(
        "Compacting {compact_count}/{} messages (preserving {preserve_count})",
        history.len(),
    );

    // Check if the conversation text fits in the current model's context.
    // Reserve 4096 tokens for the summary output + overhead.
    let text_tokens = (conversation_text.len() as f64 / crate::inference_helpers::CHARS_PER_TOKEN)
        as usize
        + crate::inference_helpers::SYSTEM_PROMPT_OVERHEAD;
    let available = max_context_tokens.saturating_sub(4096);

    // If history fits, use it as-is. Otherwise, progressively truncate
    // the oldest messages until it fits (up to MAX_TRUNCATION_RETRIES).
    let final_text = if text_tokens <= available {
        conversation_text
    } else {
        match truncate_until_fits(to_compact, available) {
            Some(text) => text,
            None => return Ok(Err(CompactSkip::HistoryTooLarge)),
        }
    };

    let summary_prompt = build_summary_prompt(&final_text);

    let messages = vec![ChatMessage::text("user", &summary_prompt)];
    // Use reduced settings for compaction on the SAME model/provider.
    // The capacity check above guarantees the conversation text fits.
    // Savings come from disabling thinking/reasoning, not switching models.
    let compact_settings = ModelSettings {
        model: model_settings.model.clone(),
        max_tokens: Some(4096),
        temperature: Some(0.3),
        thinking_budget: None,
        reasoning_effort: None,
        max_context_tokens: model_settings.max_context_tokens,
    };
    let response = provider.chat(&messages, &[], &compact_settings).await?;

    let summary = match response.content {
        Some(text) if !text.trim().is_empty() => text,
        _ => bail!("LLM returned an empty summary"),
    };

    let summary = strip_analysis_block(&summary);
    let compact_message = format!("[Compacted conversation summary]\n\n{summary}");
    let deleted = db
        .compact_session(session_id, &compact_message, preserve_count)
        .await?;

    record_compact_success();

    Ok(Ok(CompactResult {
        deleted,
        summary_tokens: summary.len() / 4,
    }))
}

/// Compute how many messages to preserve during compaction.
///
/// - Short sessions (<12 messages): full compaction, keep last 4.
/// - Longer sessions: partial compaction, keep the newest ~50%.
///
/// This preserves more recent context verbatim in long sessions,
/// producing a smaller, more focused summary of just the oldest half.
fn compute_preserve_count(total: usize) -> usize {
    if total < PARTIAL_COMPACT_THRESHOLD {
        COMPACT_PRESERVE_COUNT
    } else {
        let keep = (total as f64 * (1.0 - PARTIAL_COMPACT_FRACTION)).ceil() as usize;
        keep.max(COMPACT_PRESERVE_COUNT)
    }
}

/// Build the 9-section summarization prompt.
///
/// Adapted from CC's compaction prompt. Uses an `<analysis>` scratchpad block
/// (stripped before storing) that demonstrably improves summary quality.
fn build_summary_prompt(conversation_text: &str) -> String {
    format!(
        "CRITICAL: Respond with TEXT ONLY. Do NOT call any tools.\n\
         Tool calls will be REJECTED and will waste your only turn.\n\
         Your entire response must be plain text: an <analysis> block followed by a <summary> block.\n\
         \n\
         Your task is to create a detailed summary of the conversation so far, paying close \n\
         attention to the user's explicit requests and your previous actions.\n\
         This summary should be thorough in capturing technical details, code patterns, and \n\
         architectural decisions that would be essential for continuing development work \n\
         without losing context.\n\
         \n\
         Before providing your final summary, wrap your analysis in <analysis> tags to \n\
         organize your thoughts and ensure you've covered all necessary points. In your analysis:\n\
         \n\
         1. Chronologically analyze each message. For each section thoroughly identify:\n\
            - The user's explicit requests and intents\n\
            - Your approach to addressing them\n\
            - Key decisions, technical concepts and code patterns\n\
            - Specific details: file names, code snippets, function signatures, file edits\n\
            - Errors encountered and how they were fixed\n\
            - Specific user feedback, especially corrections\n\
         2. Double-check for technical accuracy and completeness.\n\
         \n\
         Your summary should include these sections:\n\
         \n\
         1. **Primary Request and Intent**: Capture ALL of the user's explicit requests in detail.\n\
         2. **Key Technical Concepts**: List all important technologies and frameworks discussed.\n\
         3. **Files and Code Sections**: Enumerate specific files examined, modified, or created. \n\
            Include code snippets where applicable and a summary of why each file matters.\n\
            **Be exhaustive about file paths** — once compaction runs, the only record of\n\
            files touched in the compacted range is this summary, so missing a path means\n\
            losing it for the rest of the session. Group as: created / modified / deleted.\n\
         4. **Errors and Fixes**: List all errors and how they were resolved. Note user feedback.\n\
         5. **Problem Solving**: Document problems solved and ongoing troubleshooting.\n\
         6. **All User Messages**: List ALL user messages (not tool results). Critical for \n\
            preserving feedback and changing intent.\n\
         7. **Pending Tasks**: Outline anything unfinished or deferred.\n\
            **Preserve every outstanding TodoWrite item verbatim** with its current status\n\
            (pending / in_progress). Compaction is the only mechanism that defends plan\n\
            continuity across context-window pressure (`DESIGN.md § Progress Tracking:\n\
            Model-Owned, History-Persisted, Engine-Surfaced`) — the system prompt does NOT\n\
            re-inject the todo list, so anything dropped here is gone.\n\
         8. **Current Work**: Describe precisely what was being worked on immediately before \n\
            this summary. Include file names and code snippets.\n\
         9. **Optional Next Step**: Only if directly in line with the user's most recent \n\
            explicit request. Include direct quotes from the conversation to prevent drift.\n\
         \n\
         Format your response as:\n\
         \n\
         <analysis>\n\
         [Your thought process ensuring all points are covered]\n\
         </analysis>\n\
         \n\
         <summary>\n\
         1. Primary Request and Intent:\n\
            [Detailed description]\n\
         ...\n\
         </summary>\n\
         \n\
         REMINDER: Do NOT call any tools. Respond with plain text only.\n\
         \n\
         ---\n\n{conversation_text}"
    )
}

/// Strip the `<analysis>` scratchpad block from the summary.
///
/// The analysis block improves summary quality (model "thinks" before writing)
/// but has no informational value once the summary is written. Stripping it
/// saves tokens in the ongoing context.
///
/// # Examples
///
/// ```
/// use koda_core::compact::strip_analysis_block;
///
/// let input = "<analysis>\nthinking...\n</analysis>\n\n<summary>\nThe result.\n</summary>";
/// let result = strip_analysis_block(input);
/// assert_eq!(result, "The result.");
///
/// // Plain text without tags passes through unchanged:
/// assert_eq!(strip_analysis_block("just text"), "just text");
/// ```
pub fn strip_analysis_block(summary: &str) -> String {
    // Remove <analysis>...</analysis> including the tags
    let stripped = if let Some(start) = summary.find("<analysis>") {
        if let Some(end) = summary.find("</analysis>") {
            let after = end + "</analysis>".len();
            format!("{}{}", &summary[..start], &summary[after..])
        } else {
            summary.to_string()
        }
    } else {
        summary.to_string()
    };

    // Extract content from <summary> tags if present
    let stripped = if let Some(start) = stripped.find("<summary>") {
        if let Some(end) = stripped.find("</summary>") {
            let content_start = start + "<summary>".len();
            stripped[content_start..end].trim().to_string()
        } else {
            stripped
        }
    } else {
        stripped
    };

    // Clean up extra whitespace
    let mut result = String::new();
    let mut prev_empty = false;
    for line in stripped.lines() {
        let is_empty = line.trim().is_empty();
        if is_empty && prev_empty {
            continue;
        }
        if !result.is_empty() {
            result.push('\n');
        }
        result.push_str(line);
        prev_empty = is_empty;
    }
    result.trim().to_string()
}

/// Progressively drop oldest messages until the conversation text fits
/// in the available token budget. Keeps at least `COMPACT_PRESERVE_COUNT`
/// recent messages. Returns `None` if it can't fit after max retries.
fn truncate_until_fits(history: &[crate::db::Message], available_tokens: usize) -> Option<String> {
    let total = history.len();
    // Minimum messages to keep: the preserved tail + at least 1 to summarize
    let min_keep = COMPACT_PRESERVE_COUNT + 1;
    if total <= min_keep {
        return None;
    }

    let mut drop_count = 0usize;
    for attempt in 0..MAX_TRUNCATION_RETRIES {
        // Drop 20% of remaining summarizable messages each attempt
        let summarizable = total.saturating_sub(drop_count);
        let to_drop = (summarizable as f64 * TRUNCATION_DROP_FRACTION).ceil() as usize;
        drop_count += to_drop.max(1); // always drop at least 1

        // Never drop so many that we have fewer than min_keep
        if total.saturating_sub(drop_count) < min_keep {
            drop_count = total - min_keep;
        }

        let truncated = &history[drop_count..];
        let text = build_conversation_text(truncated);
        let text_tokens = (text.len() as f64 / crate::inference_helpers::CHARS_PER_TOKEN) as usize
            + crate::inference_helpers::SYSTEM_PROMPT_OVERHEAD;

        tracing::info!(
            "Truncation attempt {}: dropped {drop_count}/{total} messages, \
             ~{text_tokens} tokens (budget: {available_tokens})",
            attempt + 1,
        );

        if text_tokens <= available_tokens {
            return Some(text);
        }
    }

    None
}

/// Format conversation history into a single string for the summarizer.
///
/// Per-message content is truncated to 2000 chars (individual tool outputs
/// can be huge but add little summarization value beyond a preview).
/// No total cap — the capacity check in `compact_session_with_provider`
/// guarantees the result fits in the model's context window.
fn build_conversation_text(history: &[crate::db::Message]) -> String {
    let mut text = String::new();
    for msg in history {
        let role = msg.role.as_str();
        if let Some(ref content) = msg.content {
            let truncated: String = content.chars().take(2000).collect();
            text.push_str(&format!("[{role}]: {truncated}\n\n"));
        }
        if let Some(ref tool_calls) = msg.tool_calls {
            let truncated: String = tool_calls.chars().take(500).collect();
            text.push_str(&format!("[{role} tool_calls]: {truncated}\n\n"));
        }
    }
    text
}

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

    fn make_msg(role: &str, content: Option<&str>, tool_calls: Option<&str>) -> Message {
        Message {
            id: 0,
            session_id: String::new(),
            role: role.parse().unwrap_or(crate::db::Role::User),
            content: content.map(String::from),
            full_content: None,
            tool_calls: tool_calls.map(String::from),
            tool_call_id: None,
            prompt_tokens: None,
            completion_tokens: None,
            cache_read_tokens: None,
            cache_creation_tokens: None,
            thinking_tokens: None,
            thinking_content: None,
            created_at: None,
        }
    }

    #[test]
    fn test_circuit_breaker() {
        reset_compact_failures();
        assert!(!is_compact_circuit_broken());

        assert!(!record_compact_failure()); // 1st
        assert!(!is_compact_circuit_broken());

        assert!(!record_compact_failure()); // 2nd
        assert!(!is_compact_circuit_broken());

        assert!(record_compact_failure()); // 3rd — trips
        assert!(is_compact_circuit_broken());

        // Reset should untrip
        reset_compact_failures();
        assert!(!is_compact_circuit_broken());
    }

    #[test]
    fn test_empty_history() {
        assert_eq!(build_conversation_text(&[]), "");
    }

    #[test]
    fn test_basic_conversation() {
        let msgs = vec![
            make_msg("user", Some("hello"), None),
            make_msg("assistant", Some("hi"), None),
        ];
        let text = build_conversation_text(&msgs);
        assert!(text.contains("[user]: hello"));
        assert!(text.contains("[assistant]: hi"));
    }

    #[test]
    fn test_truncates_long_content_per_message() {
        let long = "x".repeat(3000);
        let msgs = vec![make_msg("user", Some(&long), None)];
        let text = build_conversation_text(&msgs);
        // Each msg content capped at 2000 chars
        assert!(text.len() < 2100);
    }

    #[test]
    fn test_no_total_cap() {
        // 50 messages × 500 chars each = 25K chars — no cap applied
        let content = "y".repeat(500);
        let msgs: Vec<_> = (0..50)
            .map(|_| make_msg("user", Some(&content), None))
            .collect();
        let text = build_conversation_text(&msgs);
        // All 50 messages should be included (no 20K cap)
        assert!(text.len() > 20_000);
        assert!(!text.contains("truncated"));
    }

    #[test]
    fn test_multibyte_boundary_safe() {
        // Put emoji right at the 2000-char boundary
        let mut content = "a".repeat(1999);
        content.push('\u{1f43b}'); // bear emoji (4 bytes)
        content.push_str("after");
        let msgs = vec![make_msg("user", Some(&content), None)];
        let text = build_conversation_text(&msgs);
        // Should not panic on char boundary
        assert!(text.contains("\u{1f43b}") || !text.contains("after"));
    }

    #[test]
    fn test_tool_calls_included() {
        let msgs = vec![make_msg("assistant", None, Some("{\"name\": \"Read\"}"))];
        let text = build_conversation_text(&msgs);
        assert!(text.contains("tool_calls"));
        assert!(text.contains("Read"));
    }

    #[test]
    fn test_none_content_skipped() {
        let msgs = vec![make_msg("tool", None, None)];
        let text = build_conversation_text(&msgs);
        assert_eq!(text, "");
    }

    #[test]
    fn test_strip_analysis_block() {
        let input = "<analysis>\nthinking here\n</analysis>\n\n<summary>\n1. Primary Request:\n   Build a thing\n</summary>";
        let result = strip_analysis_block(input);
        assert!(result.contains("Primary Request"));
        assert!(!result.contains("<analysis>"));
        assert!(!result.contains("thinking here"));
        assert!(!result.contains("<summary>"));
    }

    #[test]
    fn test_strip_analysis_no_tags() {
        let input = "Just a plain summary";
        assert_eq!(strip_analysis_block(input), "Just a plain summary");
    }

    #[test]
    fn test_strip_analysis_only_summary_tags() {
        let input = "<summary>\nThe good stuff\n</summary>";
        let result = strip_analysis_block(input);
        assert_eq!(result, "The good stuff");
    }

    #[test]
    fn test_truncate_until_fits_drops_oldest() {
        // 20 messages, each ~50 chars
        let msgs: Vec<_> = (0..20)
            .map(|i| {
                make_msg(
                    "user",
                    Some(&format!("Message number {i} with some padding text here")),
                    None,
                )
            })
            .collect();

        // Budget: fits ~half the messages but not all 20
        // 20 msgs × ~50 chars / 3.5 ≈ 286 tokens + 100 overhead ≈ 386
        // Want to force truncation: set budget to ~250 tokens
        let result = truncate_until_fits(&msgs, 250);
        assert!(result.is_some(), "should succeed after truncation");
        let text = result.unwrap();
        // Should contain the last messages but not the first
        assert!(text.contains("Message number 19"));
        assert!(!text.contains("Message number 0"));
    }

    #[test]
    fn test_truncate_until_fits_too_few_messages() {
        // Only COMPACT_PRESERVE_COUNT + 1 = 5 messages, can't drop any
        let msgs: Vec<_> = (0..5)
            .map(|_| make_msg("user", Some(&"x".repeat(10_000)), None))
            .collect();
        // Tiny budget that can't fit even 5 messages
        let result = truncate_until_fits(&msgs, 10);
        assert!(result.is_none());
    }

    #[test]
    fn test_truncate_until_fits_already_fits() {
        let msgs: Vec<_> = (0..10)
            .map(|i| make_msg("user", Some(&format!("Short {i}")), None))
            .collect();
        // Huge budget
        let result = truncate_until_fits(&msgs, 100_000);
        assert!(result.is_some());
        let text = result.unwrap();
        // First attempt drops 20% but still fits, so it drops
        // We just check it returns something valid
        assert!(text.contains("Short 9"));
    }

    #[test]
    fn test_compute_preserve_count_short_sessions() {
        // Below threshold: always keep COMPACT_PRESERVE_COUNT (4)
        assert_eq!(compute_preserve_count(4), 4);
        assert_eq!(compute_preserve_count(8), 4);
        assert_eq!(compute_preserve_count(11), 4);
    }

    #[test]
    fn test_compute_preserve_count_partial() {
        // At threshold (12): keep ceil(12 * 0.5) = 6
        assert_eq!(compute_preserve_count(12), 6);
        // 20 messages: keep ceil(20 * 0.5) = 10
        assert_eq!(compute_preserve_count(20), 10);
        // 50 messages: keep 25
        assert_eq!(compute_preserve_count(50), 25);
        // 100 messages: keep 50
        assert_eq!(compute_preserve_count(100), 50);
    }

    #[test]
    fn test_compute_preserve_count_never_below_minimum() {
        // Even at threshold, result must be >= COMPACT_PRESERVE_COUNT
        for n in 0..200 {
            assert!(compute_preserve_count(n) >= COMPACT_PRESERVE_COUNT);
        }
    }

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

    #[test]
    fn test_build_summary_prompt_embeds_conversation() {
        let text = build_summary_prompt("[user]: hello\n\n[assistant]: hi\n\n");
        assert!(
            text.contains("[user]: hello"),
            "prompt should embed the conversation text verbatim"
        );
        assert!(text.contains("[assistant]: hi"));
    }

    #[test]
    fn test_build_summary_prompt_instructs_no_tool_calls() {
        let text = build_summary_prompt("some conversation");
        assert!(
            text.contains("Do NOT call any tools"),
            "prompt must forbid tool calls"
        );
        assert!(text.contains("CRITICAL"));
    }

    #[test]
    fn test_build_summary_prompt_requests_analysis_and_summary_tags() {
        let text = build_summary_prompt("some conversation");
        assert!(
            text.contains("<analysis>"),
            "prompt should ask for <analysis> block"
        );
        assert!(
            text.contains("<summary>"),
            "prompt should ask for <summary> block"
        );
    }

    // ── build_conversation_text edge cases ────────────────────────────

    #[test]
    fn test_build_conversation_text_tool_calls_truncated_at_500() {
        let long_tc = "T".repeat(600);
        let msgs = vec![make_msg("assistant", None, Some(&long_tc))];
        let text = build_conversation_text(&msgs);
        // 500 char cap on tool_calls
        assert!(
            text.len() <= 550,
            "tool_calls should be capped at 500 chars"
        );
    }

    #[test]
    fn test_build_conversation_text_both_content_and_tool_calls() {
        let msgs = vec![make_msg(
            "assistant",
            Some("I will read the file"),
            Some("{\"name\": \"Read\"}"),
        )];
        let text = build_conversation_text(&msgs);
        assert!(text.contains("I will read the file"));
        assert!(text.contains("tool_calls"));
    }

    // ── strip_analysis_block edge cases ──────────────────────────────

    #[test]
    fn test_strip_analysis_unclosed_tag_passthrough() {
        // If <analysis> has no closing tag, leave the text alone.
        let input = "<analysis>\nthinking...\n1. Primary Request: build a thing";
        let result = strip_analysis_block(input);
        assert!(
            result.contains("thinking"),
            "unclosed analysis tag should leave text intact"
        );
    }

    #[test]
    fn test_strip_analysis_trims_extra_whitespace() {
        let input = "<analysis>\nthink\n</analysis>\n\n\n\n<summary>\nClean content\n</summary>";
        let result = strip_analysis_block(input);
        // Collapsed blank lines, no leading/trailing whitespace
        assert!(!result.starts_with('\n'));
        assert!(!result.ends_with('\n'));
        assert_eq!(result, "Clean content");
    }

    // ── circuit breaker ───────────────────────────────────────────────────────
    // Covered by test_circuit_breaker above. Removed duplicate tests
    // that raced on the global CONSECUTIVE_FAILURES AtomicU32 when
    // running in parallel.
}