hematite/agent/

compaction.rs

use crate::agent::inference::ChatMessage;
use std::collections::{BTreeSet, HashSet};
use std::fmt::Write as _;

/// Professional Compaction Configuration.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct CompactionConfig {
    pub preserve_recent_messages: usize,
    /// Token threshold before compaction fires. Set dynamically via `adaptive()`.
    pub max_estimated_tokens: usize,
}

impl Default for CompactionConfig {
    fn default() -> Self {
        Self {
            preserve_recent_messages: 10,
            max_estimated_tokens: 15_000,
        }
    }
}

impl CompactionConfig {
    /// Build a hardware-aware config that scales with the model's context window
    /// and current VRAM pressure.
    ///
    /// - `context_length`: tokens the loaded model can handle (from `/api/v0/models`)
    /// - `vram_ratio`: current VRAM usage 0.0–1.0 (from GpuState::ratio)
    ///
    /// Formula: threshold = ctx * 0.40 * (1 - vram * 0.5), clamped [4k, 60k].
    /// preserve_recent_messages scales with context: roughly 1 message per 3k tokens.
    pub fn adaptive(context_length: usize, vram_ratio: f64) -> Self {
        let vram = vram_ratio.clamp(0.0, 1.0);
        let effective = (context_length as f64 * 0.40 * (1.0 - vram * 0.5)) as usize;
        let max_estimated_tokens = effective.clamp(4_000, 60_000);
        let preserve_recent_messages = (context_length / 3_000).clamp(8, 20);
        Self {
            preserve_recent_messages,
            max_estimated_tokens,
        }
    }
}

pub struct CompactionResult {
    pub messages: Vec<ChatMessage>,
    pub summary: Option<String>,
}

const DEFAULT_MAX_SUMMARY_CHARS: usize = 2_000;
const DEFAULT_MAX_SUMMARY_LINES: usize = 40;
const DEFAULT_MAX_SUMMARY_LINE_CHARS: usize = 200;

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct SummaryCompressionBudget {
    pub max_chars: usize,
    pub max_lines: usize,
    pub max_line_chars: usize,
}

impl Default for SummaryCompressionBudget {
    fn default() -> Self {
        Self {
            max_chars: DEFAULT_MAX_SUMMARY_CHARS,
            max_lines: DEFAULT_MAX_SUMMARY_LINES,
            max_line_chars: DEFAULT_MAX_SUMMARY_LINE_CHARS,
        }
    }
}

#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SummaryCompressionResult {
    pub summary: String,
    pub original_chars: usize,
    pub compressed_chars: usize,
    pub original_lines: usize,
    pub compressed_lines: usize,
    pub removed_duplicate_lines: usize,
    pub omitted_lines: usize,
    pub truncated: bool,
}

pub fn compress_summary(
    summary: &str,
    budget: SummaryCompressionBudget,
) -> SummaryCompressionResult {
    let original_chars = summary.chars().count();
    let original_lines = summary.lines().count();
    let normalized = normalize_summary_lines(summary, budget.max_line_chars);

    if normalized.lines.is_empty() || budget.max_chars == 0 || budget.max_lines == 0 {
        return SummaryCompressionResult {
            summary: String::new(),
            original_chars,
            compressed_chars: 0,
            original_lines,
            compressed_lines: 0,
            removed_duplicate_lines: normalized.removed_duplicate_lines,
            omitted_lines: normalized.lines.len(),
            truncated: original_chars > 0,
        };
    }

    let selected = select_summary_line_indexes(&normalized.lines, budget);
    let mut compressed_lines = {
        let mut v = Vec::with_capacity(selected.len());
        for index in &selected {
            v.push(normalized.lines[*index].clone());
        }
        v
    };
    if compressed_lines.is_empty() {
        compressed_lines.push(truncate_summary_line(
            &normalized.lines[0],
            budget.max_chars,
        ));
    }
    let omitted_lines = normalized
        .lines
        .len()
        .saturating_sub(compressed_lines.len());
    if omitted_lines > 0 {
        push_summary_line_with_budget(
            &mut compressed_lines,
            format!("- ... {omitted_lines} additional line(s) omitted."),
            budget,
        );
    }

    let compressed_summary = compressed_lines.join("\n");
    SummaryCompressionResult {
        summary: compressed_summary.clone(),
        original_chars,
        compressed_chars: compressed_summary.chars().count(),
        original_lines,
        compressed_lines: compressed_lines.len(),
        removed_duplicate_lines: normalized.removed_duplicate_lines,
        omitted_lines,
        truncated: compressed_summary != summary.trim(),
    }
}

pub fn compress_summary_text(summary: &str) -> String {
    compress_summary(summary, SummaryCompressionBudget::default()).summary
}

const COMPACT_PREAMBLE: &str = "## CONTEXT SUMMARY (RECURSIVE CHAIN)\n\
    This session is being continued from a previous conversation. The summary below covers the earlier portion.\n\n";
const COMPACT_INSTRUCTION: &str = "\n\nIMPORTANT: Resume directly from the last message. Do not recap or acknowledge this summary.";

/// Layer 6: Structured Session Memory.
/// Preserves the "Mission Context" across compactions.
#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
pub struct SessionCheckpoint {
    pub state: String,
    pub summary: String,
}

#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
pub struct SessionVerification {
    pub successful: bool,
    pub summary: String,
}

#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
pub struct SessionCompactionLedger {
    pub count: u32,
    pub removed_message_count: usize,
    pub summary: String,
}

#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
pub struct SessionMemory {
    pub current_task: String,
    pub working_set: std::collections::HashSet<String>,
    pub learnings: Vec<String>,
    #[serde(default)]
    pub current_plan: Option<crate::tools::plan::PlanHandoff>,
    #[serde(default)]
    pub last_checkpoint: Option<SessionCheckpoint>,
    #[serde(default)]
    pub last_blocker: Option<SessionCheckpoint>,
    #[serde(default)]
    pub last_recovery: Option<SessionCheckpoint>,
    #[serde(default)]
    pub last_verification: Option<SessionVerification>,
    #[serde(default)]
    pub last_compaction: Option<SessionCompactionLedger>,
}

impl SessionMemory {
    pub fn has_signal(&self) -> bool {
        let task = self.current_task.trim();
        (!task.is_empty() && task != "Ready for new mission.")
            || !self.working_set.is_empty()
            || !self.learnings.is_empty()
            || self.last_checkpoint.is_some()
            || self.last_blocker.is_some()
            || self.last_recovery.is_some()
            || self.last_verification.is_some()
            || self.last_compaction.is_some()
            || self
                .current_plan
                .as_ref()
                .map(|plan| plan.has_signal())
                .unwrap_or(false)
    }

    pub fn to_prompt(&self) -> String {
        let mut s = format!("- **Active Task**: {}\n", self.current_task);
        if let Some(plan) = &self.current_plan {
            if plan.has_signal() {
                s.push_str("- **Active Plan Handoff**:\n");
                s.push_str(&plan.to_prompt());
            }
        }
        if !self.working_set.is_empty() {
            s.push_str("- **Working Set**: ");
            let mut first = true;
            for f in &self.working_set {
                if !first {
                    s.push_str(", ");
                }
                s.push_str(f);
                first = false;
            }
            s.push('\n');
        }
        if !self.learnings.is_empty() {
            s.push_str("- **Key Learnings**:\n");
            for l in &self.learnings {
                let _ = writeln!(s, "  - {l}");
            }
        }
        if let Some(checkpoint) = &self.last_checkpoint {
            if checkpoint.summary.trim().is_empty() {
                let _ = writeln!(s, "- **Latest Checkpoint**: {}", checkpoint.state);
            } else {
                let _ = writeln!(
                    s,
                    "- **Latest Checkpoint**: {} - {}",
                    checkpoint.state, checkpoint.summary
                );
            }
        }
        if let Some(blocker) = &self.last_blocker {
            if blocker.summary.trim().is_empty() {
                let _ = writeln!(s, "- **Latest Blocker**: {}", blocker.state);
            } else {
                let _ = writeln!(
                    s,
                    "- **Latest Blocker**: {} - {}",
                    blocker.state, blocker.summary
                );
            }
        }
        if let Some(recovery) = &self.last_recovery {
            if recovery.summary.trim().is_empty() {
                let _ = writeln!(s, "- **Latest Recovery**: {}", recovery.state);
            } else {
                let _ = writeln!(
                    s,
                    "- **Latest Recovery**: {} - {}",
                    recovery.state, recovery.summary
                );
            }
        }
        if let Some(verification) = &self.last_verification {
            let status = if verification.successful {
                "passed"
            } else {
                "failed"
            };
            let _ = writeln!(
                s,
                "- **Latest Verification**: {} - {}",
                status, verification.summary
            );
        }
        if let Some(compaction) = &self.last_compaction {
            let _ = writeln!(
                s,
                "- **Latest Compaction**: pass {} removed {} message(s) - {}",
                compaction.count, compaction.removed_message_count, compaction.summary
            );
        }
        s
    }

    pub fn inherit_runtime_ledger_from(&mut self, other: &Self) {
        self.last_checkpoint = other.last_checkpoint.clone();
        self.last_blocker = other.last_blocker.clone();
        self.last_recovery = other.last_recovery.clone();
        self.last_verification = other.last_verification.clone();
        self.last_compaction = other.last_compaction.clone();
    }

    pub fn record_checkpoint(&mut self, state: impl Into<String>, summary: impl Into<String>) {
        let checkpoint = SessionCheckpoint {
            state: state.into(),
            summary: summary.into(),
        };
        let state_name = checkpoint.state.as_str();
        if state_name == "recovering_provider" {
            self.last_recovery = Some(checkpoint.clone());
        }
        if state_name.starts_with("blocked_") {
            self.last_blocker = Some(checkpoint.clone());
        }
        self.last_checkpoint = Some(checkpoint);
    }

    pub fn record_verification(&mut self, successful: bool, summary: impl Into<String>) {
        self.last_verification = Some(SessionVerification {
            successful,
            summary: summary.into(),
        });
    }

    pub fn record_recovery(&mut self, state: impl Into<String>, summary: impl Into<String>) {
        let checkpoint = SessionCheckpoint {
            state: state.into(),
            summary: summary.into(),
        };
        self.last_recovery = Some(checkpoint.clone());
        self.last_checkpoint = Some(checkpoint);
    }

    pub fn record_compaction(&mut self, removed_message_count: usize, summary: impl Into<String>) {
        let count = self
            .last_compaction
            .as_ref()
            .map_or(1, |entry| entry.count.saturating_add(1));
        self.last_compaction = Some(SessionCompactionLedger {
            count,
            removed_message_count,
            summary: summary.into(),
        });
    }

    pub fn clear(&mut self) {
        self.current_task = "Ready for new mission.".to_string();
        self.working_set.clear();
        self.learnings.clear();
        self.current_plan = None;
        self.last_checkpoint = None;
        self.last_blocker = None;
        self.last_recovery = None;
        self.last_verification = None;
        self.last_compaction = None;
    }
}

/// Returns true when history is large enough to warrant compaction.
/// Pass the model's context_length and current vram_ratio for adaptive thresholds.
pub fn should_compact(history: &[ChatMessage], context_length: usize, vram_ratio: f64) -> bool {
    let config = CompactionConfig::adaptive(context_length, vram_ratio);
    history.len().saturating_sub(1) > config.preserve_recent_messages + 5
        || estimate_compactable_tokens(history) > config.max_estimated_tokens
}

pub fn compact_history(
    history: &[ChatMessage],
    existing_summary: Option<&str>,
    config: CompactionConfig,
    // The index of the user message that started the CURRENT turn.
    // We must NEVER summarize past this index if we are in the middle of a turn.
    anchor_index: Option<usize>,
) -> CompactionResult {
    if history.len() <= config.preserve_recent_messages + 5 {
        return CompactionResult {
            messages: history.to_vec(),
            summary: existing_summary.map(|s| s.to_string()),
        };
    }

    // Triple-Slicer Strategy:
    // 1. [SYSTEM] (Index 0)
    // 2. [PAST TURNS] (Index 1 .. Anchor) -> Folded into summary.
    // 3. [ENTRY PROMPT] (Index Anchor) -> Kept verbatim for Jinja alignment.
    // 4. [MIDDLE OF TURN] (Index Anchor+1 .. End - Preserve) -> Folded into summary.
    // 5. [RECENT WORK] (End - Preserve .. End) -> Kept verbatim.

    // The anchor MUST be at least 1 (to avoid 1..0 slice panics) and
    // capped at history.len() - 1.
    let anchor = anchor_index.unwrap_or(1).max(1).min(history.len() - 1);
    let keep_from = history
        .len()
        .saturating_sub(config.preserve_recent_messages);

    let mut messages_to_summarize = Vec::with_capacity(history.len());
    let mut preserved_messages = Vec::with_capacity(history.len());

    // Preserve the Turn Entry User Prompt as the primary anchor.
    // Everything before it is permanently summarized.
    if anchor > 1 {
        messages_to_summarize.extend(history[1..anchor].iter().cloned());
    }
    preserved_messages.push(history[anchor].clone());

    // Evaluate the Middle of the Turn.
    if keep_from > anchor + 1 {
        // We have enough bulk in the current turn to justify a "Partial Turn" summary.
        messages_to_summarize.extend(history[anchor + 1..keep_from].iter().cloned());
        preserved_messages.extend(history[keep_from..].iter().cloned());
    } else {
        // Not enough bulk inside the turn yet; just preserve the rest.
        preserved_messages.extend(history[anchor + 1..].iter().cloned());
    }

    let new_summary_txt = build_technical_summary(&messages_to_summarize);
    let merged_summary = match existing_summary {
        Some(existing) => merge_summaries(existing, &new_summary_txt),
        None => new_summary_txt,
    };

    let summary_content = format!(
        "{}{}{}",
        COMPACT_PREAMBLE, merged_summary, COMPACT_INSTRUCTION
    );
    let summary_msg = ChatMessage::system(&summary_content);

    let mut new_history = vec![history[0].clone()];
    new_history.push(summary_msg);
    new_history.extend(preserved_messages);

    CompactionResult {
        messages: new_history,
        summary: Some(merged_summary),
    }
}

/// Heuristic extraction of "The Mission" from a set of messages.
pub fn extract_memory(messages: &[ChatMessage]) -> SessionMemory {
    let mut mem = SessionMemory::default();

    // Use the most recent user message for current_task to avoid topic pollution,
    // but scan ALL turns for working_set so files touched earlier in the session
    // are not silently dropped after the first compaction pass.
    let last_user_idx = messages.iter().rposition(|m| m.role == "user");

    if let Some(idx) = last_user_idx {
        let m = &messages[idx];
        let content_str = m.content.as_str();
        let limit = 250;
        mem.current_task = content_str.chars().take(limit).collect();
        if content_str.len() > limit {
            mem.current_task.push_str("...");
        }
    }

    // Working set: collect path args from every tool call across all turns,
    // giving higher priority to tool calls in the most recent user turn.
    // Cap at 12 files; most-recently-touched files survive longest.
    let mut all_files: Vec<String> = Vec::with_capacity(messages.len());
    for msg in messages {
        if let Some(calls) = &msg.tool_calls {
            for call in calls {
                let args = call.function.arguments.clone();
                if let Some(path) = args.get("path").and_then(|v| v.as_str()) {
                    // Push in traversal order so later (more recent) entries
                    // win deduplication when we reverse below.
                    all_files.push(path.to_string());
                }
            }
        }
    }
    // Keep unique files, most-recent first, capped at 12.
    let mut seen = HashSet::new();
    for path in all_files.into_iter().rev() {
        if seen.insert(path.clone()) {
            mem.working_set.insert(path);
            if mem.working_set.len() >= 12 {
                break;
            }
        }
    }

    // Learnings: scan the most recent user turn only to avoid stale signals.
    if let Some(idx) = last_user_idx {
        for turn_msg in &messages[idx..] {
            if turn_msg.role == "tool" {
                let content_str = turn_msg.content.as_str();
                if content_str.contains("Error:")
                    || content_str.contains("Finished")
                    || content_str.contains("Complete")
                {
                    let lines: Vec<_> = content_str.lines().take(2).collect();
                    mem.learnings.push(lines.join(" "));
                }
            }
        }
    }

    // De-duplicate and cap learnings.
    mem.learnings.dedup();
    if mem.learnings.len() > 5 {
        mem.learnings.truncate(5);
    }

    mem
}

pub fn estimate_tokens(messages: &[ChatMessage]) -> usize {
    messages
        .iter()
        .map(|m| m.content.as_str().len() / 4 + 1)
        .sum()
}

pub fn estimate_compactable_tokens(history: &[ChatMessage]) -> usize {
    if history.len() <= 1 {
        0
    } else {
        estimate_tokens(&history[1..])
    }
}

fn build_technical_summary(messages: &[ChatMessage]) -> String {
    let mut lines = vec![format!(
        "- Scope: {} earlier turns compacted.",
        messages.len()
    )];

    // 1. Extract files from tool-call path args (precise) then fall back to
    //    word-scan for any path-like tokens not captured that way.
    let mut files: IndexedSet = IndexedSet::default();
    let mut tools: HashSet<String> = HashSet::new();
    let mut requests: Vec<String> = Vec::with_capacity(messages.len().min(16));
    let mut assistant_notes: Vec<String> = Vec::with_capacity(messages.len().min(8));
    // Tool results: verify_build, edit outcomes, notable errors.
    let mut verify_outcome: Option<bool> = None;
    let mut error_snippets: Vec<String> = Vec::with_capacity(4);

    for m in messages {
        // Precise file extraction from tool call arguments.
        if let Some(calls) = &m.tool_calls {
            for call in calls {
                tools.insert(call.function.name.clone());
                let args = call.function.arguments.clone();
                if let Some(path) = args.get("path").and_then(|v| v.as_str()) {
                    files.insert(path.to_string());
                }
            }
        }

        // Tool result signals.
        if m.role == "tool" {
            let text = m.content.as_str();
            // verify_build result — last one wins.
            if text.contains("BUILD OK") || text.contains("BUILD SUCCESS") {
                verify_outcome = Some(true);
            } else if text.contains("BUILD FAIL") || text.contains("error[") {
                verify_outcome = Some(false);
            }
            // Capture first error line from any tool result.
            if text.contains("Error:") || text.contains("error:") {
                if let Some(err_line) = text.lines().find(|l| {
                    l.trim_start().starts_with("Error:") || l.trim_start().starts_with("error:")
                }) {
                    let snippet: String = err_line.chars().take(100).collect();
                    error_snippets.push(snippet);
                }
            }
        }

        // User requests (up to 4, most recent last).
        if m.role == "user" && !m.content.as_str().trim().is_empty() && requests.len() < 4 {
            let text = m
                .content
                .as_str()
                .trim_start_matches("/think\n")
                .trim_start_matches("/no_think\n")
                .trim();
            requests.push(truncate_summary_line(
                &collapse_inline_whitespace(text),
                140,
            ));
        }

        // Assistant prose (up to 3) — capture decisions and explanations made.
        if m.role == "assistant"
            && !m.content.as_str().trim().is_empty()
            && m.tool_calls.as_ref().is_none_or(|tc| tc.is_empty())
            && assistant_notes.len() < 3
        {
            let text = m.content.as_str().trim();
            if text.len() > 20 {
                assistant_notes.push(truncate_summary_line(
                    &collapse_inline_whitespace(text),
                    120,
                ));
            }
        }

        // Word-scan fallback for path-like tokens not in tool args.
        for word in m.content.as_str().split_whitespace() {
            let clean = word.trim_matches(|c: char| {
                matches!(c, ',' | '.' | ':' | ';' | ')' | '(' | '"' | '\'' | '`')
            });
            if clean.len() > 4
                && clean.contains('.')
                && (clean.contains('/') || clean.contains('\\'))
            {
                files.insert(clean.to_string());
            }
        }
    }

    if !files.0.is_empty() {
        let list: Vec<String> = files.0.into_iter().take(10).collect();
        lines.push(format!("- Key files: {}.", list.join(", ")));
    }
    if !tools.is_empty() {
        let list: Vec<String> = tools.into_iter().take(8).collect();
        lines.push(format!("- Tools used: {}.", list.join(", ")));
    }
    if let Some(ok) = verify_outcome {
        lines.push(format!(
            "- Last verify_build: {}.",
            if ok { "BUILD OK" } else { "BUILD FAILED" }
        ));
    }
    // Include up to 2 unique error snippets.
    error_snippets.dedup();
    for snippet in error_snippets.into_iter().take(2) {
        lines.push(format!("- Error seen: {}", snippet));
    }
    if !assistant_notes.is_empty() {
        lines.push("- Assistant decisions/responses (oldest→newest):".to_string());
        for note in &assistant_notes {
            lines.push(format!("  - {}", note));
        }
    }
    if !requests.is_empty() {
        lines.push("- User requests (oldest→newest):".to_string());
        for request in &requests {
            lines.push(format!("  - {}", request));
        }
    }

    // 2. Timeline: last 6 messages for context.
    lines.push("- Compacted context:".to_string());
    for m in messages.iter().rev().take(6).rev() {
        let content_str = m.content.as_str();
        let preview = if content_str.len() > 120 {
            let mut s: String = content_str.chars().take(117).collect();
            s.push_str("...");
            s
        } else if content_str.is_empty()
            && m.tool_calls
                .as_ref()
                .map(|c| !c.is_empty())
                .unwrap_or(false)
        {
            {
                let calls = m.tool_calls.as_ref().unwrap();
                let mut s = String::with_capacity(
                    14 + calls.iter().map(|c| c.function.name.len()).sum::<usize>()
                        + calls.len().saturating_sub(1) * 2,
                );
                s.push_str("Executing: ");
                for (i, c) in calls.iter().enumerate() {
                    if i > 0 {
                        s.push_str(", ");
                    }
                    s.push_str(&c.function.name);
                }
                s
            }
        } else {
            content_str.to_string()
        };
        lines.push(format!(
            "  - {}: {}",
            m.role,
            preview.replace('\n', " ").trim()
        ));
    }

    compress_summary_text(&lines.join("\n"))
}

/// Insertion-ordered set: tracks insertion order for deterministic file output
/// while deduplicating entries.
#[derive(Default)]
struct IndexedSet(Vec<String>);

impl IndexedSet {
    fn insert(&mut self, s: String) {
        if !self.0.contains(&s) {
            self.0.push(s);
        }
    }
}

fn merge_summaries(existing: &str, new: &str) -> String {
    compress_summary_text(&format!(
        "Conversation summary:\n- Previously compacted context:\n{}\n- Newly compacted context:\n{}",
        existing.trim(),
        new.trim()
    ))
}

#[derive(Debug, Default)]
struct NormalizedSummary {
    lines: Vec<String>,
    removed_duplicate_lines: usize,
}

fn normalize_summary_lines(summary: &str, max_line_chars: usize) -> NormalizedSummary {
    let mut seen = BTreeSet::new();
    let mut lines = Vec::new();
    let mut removed_duplicate_lines = 0;

    for raw_line in summary.lines() {
        let normalized = collapse_inline_whitespace(raw_line);
        if normalized.is_empty() {
            continue;
        }
        let truncated = truncate_summary_line(&normalized, max_line_chars);
        let dedupe_key = truncated.to_ascii_lowercase();
        if !seen.insert(dedupe_key) {
            removed_duplicate_lines += 1;
            continue;
        }
        lines.push(truncated);
    }

    NormalizedSummary {
        lines,
        removed_duplicate_lines,
    }
}

fn select_summary_line_indexes(lines: &[String], budget: SummaryCompressionBudget) -> Vec<usize> {
    let mut selected = BTreeSet::<usize>::new();

    for priority in 0..=3 {
        for (index, line) in lines.iter().enumerate() {
            if selected.contains(&index) || summary_line_priority(line) != priority {
                continue;
            }
            let new_len = selected.len() + 1;
            if new_len > budget.max_lines {
                continue;
            }
            let char_count: usize = selected
                .iter()
                .map(|si| lines[*si].chars().count())
                .sum::<usize>()
                + line.chars().count()
                + new_len.saturating_sub(1);
            if char_count > budget.max_chars {
                continue;
            }
            selected.insert(index);
        }
    }

    selected.into_iter().collect()
}

fn push_summary_line_with_budget(
    lines: &mut Vec<String>,
    line: String,
    budget: SummaryCompressionBudget,
) {
    let new_len = lines.len() + 1;
    if new_len > budget.max_lines {
        return;
    }
    let char_count: usize = lines.iter().map(|l| l.chars().count()).sum::<usize>()
        + line.chars().count()
        + new_len.saturating_sub(1);
    if char_count <= budget.max_chars {
        lines.push(line);
    }
}

fn summary_line_priority(line: &str) -> usize {
    if line == "Conversation summary:" || is_core_summary_detail(line) {
        0
    } else if line.ends_with(':') {
        1
    } else if line.starts_with("- ") || line.starts_with("  - ") {
        2
    } else {
        3
    }
}

fn is_core_summary_detail(line: &str) -> bool {
    [
        "- Scope:",
        "- Key files referenced:",
        "- Tools mentioned:",
        "- Recent user requests:",
        "- Previously compacted context:",
        "- Newly compacted context:",
    ]
    .iter()
    .any(|prefix| line.starts_with(prefix))
}

fn collapse_inline_whitespace(line: &str) -> String {
    let mut out = String::with_capacity(line.len());
    for word in line.split_whitespace() {
        if !out.is_empty() {
            out.push(' ');
        }
        out.push_str(word);
    }
    out
}

fn truncate_summary_line(line: &str, max_chars: usize) -> String {
    if max_chars == 0 {
        return String::new();
    }
    if max_chars == 1 {
        return ".".to_string();
    }
    let take_n = max_chars.saturating_sub(3);
    let mut chars = line.chars();
    let mut head = String::with_capacity(take_n);
    head.extend(chars.by_ref().take(take_n));
    // If 4+ chars remain, total > max_chars — truncate. Otherwise return original.
    if chars.take(4).count() >= 4 {
        head.push_str("...");
        head
    } else {
        line.to_string()
    }
}