robit_agent/

context.rs

//! Context management — output truncation and history window management.

use async_openai::types::chat::{
    ChatCompletionRequestMessage, ChatCompletionRequestUserMessage,
    ChatCompletionRequestUserMessageContent,
};
use robit_ai::config::ContextConfig;

// ============================================================================
// Truncation result
// ============================================================================

/// Type of truncation action, determines how the caller should handle the result.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum TruncationAction {
    /// Generate a new summary segment from full conversation rounds.
    /// The removed messages in `TruncationResult` are the full rounds to summarize.
    NewSegment,
    /// Merge multiple existing summary segments into one.
    /// `summaries` contains the text of segments to merge (oldest first).
    /// `start_position` is the index of the first segment in history.
    /// `count` is how many consecutive segments to merge.
    MergeSegments {
        summaries: Vec<String>,
        start_position: usize,
        count: usize,
    },
    /// No compression needed — history was truncated but the removed content
    /// is too small to justify an LLM summary call.
    TruncateOnly,
}

/// Result of context truncation, used for async compression.
#[derive(Debug)]
pub struct TruncationResult {
    /// Number of conversation rounds removed.
    pub rounds_removed: usize,
    /// Number of individual messages removed.
    pub messages_removed: usize,
    /// The removed messages (for generating summary — only for NewSegment action).
    pub removed_messages: Vec<ChatCompletionRequestMessage>,
    /// Position where summary should be inserted / replaced.
    pub insert_position: usize,
    /// Whether compression is needed (token count exceeds threshold).
    pub needs_compression: bool,
    /// The type of truncation action taken.
    pub action: TruncationAction,
}

// ============================================================================
// Tool output truncation (Layer 1)
// ============================================================================

/// Truncate tool output based on line count and byte limits.
pub fn truncate_output(content: &str, max_lines: usize, max_bytes: usize) -> String {
    let lines: Vec<&str> = content.lines().collect();
    let total_lines = lines.len();
    let total_bytes = content.len();

    // Check if truncation is needed
    let line_truncated = total_lines > max_lines;
    let byte_truncated = total_bytes > max_bytes;

    if !line_truncated && !byte_truncated {
        return content.to_string();
    }

    let mut output = String::new();
    let mut byte_count = 0;
    let mut displayed_lines = 0;

    for (i, line) in lines.iter().enumerate() {
        if i >= max_lines {
            break;
        }
        let line_with_newline = if i < total_lines - 1 {
            format!("{}\n", line)
        } else {
            line.to_string()
        };

        if byte_count + line_with_newline.len() > max_bytes {
            break;
        }

        output.push_str(&line_with_newline);
        byte_count += line_with_newline.len();
        displayed_lines += 1;
    }

    if line_truncated {
        output.push_str(&format!(
            "\n... (Output truncated, {} lines total, showing first {}. Use offset/limit to read more)",
            total_lines, displayed_lines
        ));
    } else if byte_truncated {
        output.push_str(&format!(
            "\n... (Output truncated, {} bytes total, showing first {} bytes)",
            total_bytes, byte_count
        ));
    }

    output
}

// ============================================================================
// Token estimation
// ============================================================================

/// Estimate token count for a string.
///
/// Uses a more nuanced heuristic based on character type:
/// - ASCII letters/digits: ~3.5 chars/token (BPE tokenizer average)
/// - CJK characters: ~1.5 chars/token (most CJK chars are 1-2 tokens each)
/// - Whitespace: minimal token cost (usually merged with adjacent tokens)
/// - Punctuation/symbols: ~1 token per char (often individual tokens)
/// - Code (braces, operators): ~1 token per char
///
/// This is still an estimate; apply `token_safety_margin` at the message level.
pub fn estimate_tokens(text: &str) -> usize {
    if text.is_empty() {
        return 0;
    }

    let mut ascii_alnum = 0usize;
    let mut cjk = 0usize;
    let mut whitespace = 0usize;
    let mut other = 0usize; // punctuation, symbols, code characters

    for ch in text.chars() {
        if ch.is_whitespace() {
            whitespace += 1;
        } else if ch.is_ascii_alphanumeric() {
            ascii_alnum += 1;
        } else {
            let cp = ch as u32;
            // CJK Unified Ideographs + extensions + fullwidth forms
            // + Hiragana, Katakana, Hangul, CJK punctuation
            if (0x4E00..=0x9FFF).contains(&cp)
                || (0x3400..=0x4DBF).contains(&cp)
                || (0xF900..=0xFAFF).contains(&cp)
                || (0xFF00..=0xFFEF).contains(&cp)
                || (0x3000..=0x303F).contains(&cp)
                || (0x3040..=0x309F).contains(&cp)
                || (0x30A0..=0x30FF).contains(&cp)
                || (0xAC00..=0xD7AF).contains(&cp)
            {
                cjk += 1;
            } else {
                other += 1;
            }
        }
    }

    // BPE tokenizer averages:
    // - ASCII alphanumeric: ~3.5 chars per token
    // - CJK: ~1.5 chars per token (most are 1 token each, some pairs)
    // - Whitespace: negligible (merged with adjacent tokens)
    // - Other (punctuation/code): ~1 char per token
    let ascii_tokens = (ascii_alnum as f64 / 3.5).ceil() as usize;
    let cjk_tokens = (cjk as f64 / 1.5).ceil() as usize;
    let whitespace_tokens = (whitespace as f64 / 10.0).ceil() as usize;
    let other_tokens = other; // ~1:1

    ascii_tokens + cjk_tokens + whitespace_tokens + other_tokens
}

/// Estimate tokens for a list of messages.
pub fn estimate_messages_tokens(messages: &[ChatCompletionRequestMessage]) -> usize {
    let mut total = 0;
    for msg in messages {
        // Each message has ~4 tokens of overhead (role, delimiters)
        total += 4;
        total += estimate_message_content_tokens(msg);
    }
    total
}

/// Estimate tokens for messages, applying the configured safety margin.
pub fn estimate_messages_tokens_with_margin(
    messages: &[ChatCompletionRequestMessage],
    safety_margin: f32,
) -> usize {
    let raw = estimate_messages_tokens(messages);
    (raw as f32 * safety_margin).ceil() as usize
}

/// Estimate tokens for a single message's content.
fn estimate_message_content_tokens(msg: &ChatCompletionRequestMessage) -> usize {
    use async_openai::types::chat::ChatCompletionRequestUserMessageContentPart;

    // For user messages with multimodal (array) content, estimate each part
    // separately. Image base64 data URLs must NOT be counted by string length:
    // a 2K image is ~10MB of base64 but only ~1-2k tokens to the vision API.
    // Counting the raw base64 wildly overestimates tokens and triggers endless
    // truncation loops.
    if let ChatCompletionRequestMessage::User(user_msg) = msg {
        if let ChatCompletionRequestUserMessageContent::Array(parts) = &user_msg.content {
            let mut total = 0;
            for part in parts {
                match part {
                    ChatCompletionRequestUserMessageContentPart::Text(t) => {
                        total += estimate_tokens(&t.text);
                    }
                    ChatCompletionRequestUserMessageContentPart::ImageUrl(_) => {
                        // Vision models count image tokens by resolution,
                        // typically ~765-2000 tokens per image. Use a
                        // conservative flat estimate.
                        total += 2000;
                    }
                    _ => {
                        // InputAudio, File, etc. - not used in this codebase.
                    }
                }
            }
            return total;
        }
    }

    // Fallback: text-only messages - estimate from the JSON serialization.
    match serde_json::to_string(msg) {
        Ok(json) => estimate_tokens(&json),
        Err(_) => 0,
    }
}

// ============================================================================
// Context manager (Layer 2: history truncation)
// ============================================================================

/// Manages the context window, truncating history when approaching token limits.
pub struct ContextManager {
    /// Model's context window size in tokens.
    pub max_tokens: usize,
    /// Ratio of context window to reserve for LLM response (default 0.2 = 20%).
    pub reserve_ratio: f32,
    /// Fraction of max_tokens at which truncation triggers (default 0.7).
    pub truncation_ratio: f32,
    /// Minimum conversation rounds to keep after truncation (default 3).
    pub min_keep_rounds: usize,
    /// Safety multiplier for token estimates (default 1.3).
    pub token_safety_margin: f32,
    /// Max output lines for tool results.
    pub max_output_lines: usize,
    /// Max output bytes for tool results.
    pub max_output_bytes: usize,
    /// Token threshold for triggering compression.
    pub compression_token_threshold: usize,
    /// Whether compression is enabled.
    pub compression_enabled: bool,
    /// Maximum tool calls per turn (default 30).
    pub max_tool_calls_per_turn: usize,
    /// Whether progressive segmented compression is enabled (default true).
    pub progressive_compression: bool,
    /// Number of full rounds per summary segment (default 3).
    pub rounds_per_summary: usize,
    /// Maximum number of summary segments to keep (default 5).
    pub max_summary_segments: usize,
    /// Number of segments to merge at a time (default 2).
    pub merge_count: usize,
    /// Maximum merges per segment before discarding (default 2).
    pub max_merges_per_segment: usize,
    /// Max dimension (longest side, px) for images encoded into the context
    /// (default 1024). Larger images are downscaled and re-encoded as JPEG.
    /// 0 disables compression.
    pub max_image_dimension: u32,
}

impl ContextManager {
    pub fn new(context_window: Option<u64>, config: Option<&ContextConfig>) -> Self {
        let max_tokens = context_window.unwrap_or(65536) as usize;

        let (
            max_output_lines,
            max_output_bytes,
            reserve_ratio,
            truncation_ratio,
            min_keep_rounds,
            token_safety_margin,
            compression_token_threshold,
            compression_enabled,
            max_tool_calls_per_turn,
            progressive_compression,
            rounds_per_summary,
            max_summary_segments,
            merge_count,
            max_merges_per_segment,
            max_image_dimension,
        ) = match config {
            Some(c) => (
                c.max_output_lines.unwrap_or(500),
                c.max_output_bytes.unwrap_or(51200),
                c.reserve_ratio.unwrap_or(0.2),
                c.truncation_ratio.unwrap_or(0.7),
                c.min_keep_rounds.unwrap_or(3),
                c.token_safety_margin.unwrap_or(1.3),
                c.compression_token_threshold.unwrap_or(5000),
                c.compression_enabled.unwrap_or(true),
                c.max_tool_calls_per_turn.unwrap_or(30),
                c.progressive_compression.unwrap_or(true),
                c.rounds_per_summary.unwrap_or(3),
                c.max_summary_segments.unwrap_or(5),
                c.merge_count.unwrap_or(2),
                c.max_merges_per_segment.unwrap_or(2),
                c.max_image_dimension.unwrap_or(1024),
            ),
            None => (500, 51200, 0.2, 0.7, 3, 1.3, 5000, true, 30, true, 3, 5, 2, 2, 1024),
        };

        Self {
            max_tokens,
            reserve_ratio,
            truncation_ratio,
            min_keep_rounds,
            token_safety_margin,
            max_output_lines,
            max_output_bytes,
            compression_token_threshold,
            compression_enabled,
            max_tool_calls_per_turn,
            progressive_compression,
            rounds_per_summary,
            max_summary_segments,
            merge_count,
            max_merges_per_segment,
            max_image_dimension,
        }
    }

    /// Maximum tokens at which truncation is triggered.
    /// Uses `truncation_ratio` (default 0.7) to trigger earlier than the
    /// absolute limit, leaving headroom for estimation errors and LLM response.
    pub fn truncation_threshold(&self) -> usize {
        (self.max_tokens as f32 * self.truncation_ratio) as usize
    }

    /// Maximum tokens available for input (total - reserved for response).
    /// Note: this is the absolute upper bound; truncation actually triggers
    /// earlier via `truncation_threshold()`.
    pub fn available_tokens(&self) -> usize {
        (self.max_tokens as f32 * (1.0 - self.reserve_ratio)) as usize
    }

    /// Truncate tool output using the configured limits.
    pub fn truncate_tool_output(&self, content: &str) -> String {
        truncate_output(content, self.max_output_lines, self.max_output_bytes)
    }

    /// Hybrid token estimation: precise API anchor + incremental estimation.
    ///
    /// When `last_known_prompt_tokens` is `Some` and history hasn't been truncated
    /// (i.e. `messages.len() >= snapshot_message_count`), uses the API-reported
    /// `prompt_tokens` as an exact baseline and adds estimated tokens for only
    /// the new messages appended since the snapshot. This is far more accurate
    /// than full heuristic estimation because the baseline is from the model's
    /// own tokenizer.
    ///
    /// Falls back to full heuristic estimation with `token_safety_margin` when:
    /// - No calibration data yet (first call, before any API response)
    /// - History was truncated/compressed (messages.len() < snapshot)
    pub fn estimate_context_tokens(
        &self,
        messages: &[ChatCompletionRequestMessage],
        last_known_prompt_tokens: Option<u32>,
        snapshot_message_count: usize,
    ) -> usize {
        match last_known_prompt_tokens {
            Some(known) if messages.len() >= snapshot_message_count => {
                // Precise baseline + incremental estimate for new messages only.
                // No safety_margin needed: the baseline is exact, and incremental
                // error is small (only 1-3 new messages).
                let delta = estimate_messages_tokens(&messages[snapshot_message_count..]);
                let total = known as usize + delta;
                tracing::trace!(
                    "estimate_context_tokens: calibrated (baseline={}, snapshot={}, new={}, delta={}, total={})",
                    known, snapshot_message_count, messages.len() - snapshot_message_count, delta, total
                );
                total
            }
            _ => {
                // Fallback: full heuristic estimation with safety margin
                let estimated = estimate_messages_tokens_with_margin(messages, self.token_safety_margin);
                tracing::trace!(
                    "estimate_context_tokens: fallback heuristic ({} messages, {} tokens with {:.1}x margin)",
                    messages.len(), estimated, self.token_safety_margin
                );
                estimated
            }
        }
    }

    /// Check if history needs truncation and perform it if necessary.
    /// Returns `TruncationResult` with removed messages for async compression.
    ///
    /// Strategy (progressive compression):
    /// 1. Uses `truncation_threshold()` (default 70% of max_tokens) as trigger point.
    /// 2. When over threshold, performs exactly one compression action per call:
    ///    - Priority 1: Compress the oldest `rounds_per_summary` full rounds into a new summary segment.
    ///    - Priority 2: Merge the oldest `merge_count` summary segments into one.
    ///    - Priority 3: Discard the oldest summary segment (if merge limit reached).
    ///    - Priority 4: Fall back to aggressive truncation (old behavior).
    /// 3. If `progressive_compression` is disabled, falls back to single-shot truncation.
    pub fn maybe_truncate(
        &self,
        messages: &mut Vec<ChatCompletionRequestMessage>,
        last_known_prompt_tokens: Option<u32>,
        snapshot_message_count: usize,
    ) -> TruncationResult {
        let estimated = self.estimate_context_tokens(messages, last_known_prompt_tokens, snapshot_message_count);
        let threshold = self.truncation_threshold();

        if estimated <= threshold {
            // The overwhelmingly common path — keep at trace to avoid log spam
            // (this check runs twice per agent step).
            tracing::trace!("maybe_truncate: no truncation needed ({} messages, {} <= {} tokens)",
                messages.len(), estimated, threshold);
            return TruncationResult {
                rounds_removed: 0,
                messages_removed: 0,
                removed_messages: Vec::new(),
                insert_position: 0,
                needs_compression: false,
                action: TruncationAction::TruncateOnly,
            };
        }

        tracing::info!(
            "=== Context truncation triggered ==="
        );
        tracing::info!(
            "Estimated: {} tokens (with {:.1}x margin), Threshold: {} tokens, Max: {} tokens",
            estimated,
            self.token_safety_margin,
            threshold,
            self.max_tokens
        );
        tracing::info!("Compression enabled: {}, Progressive: {}",
            self.compression_enabled, self.progressive_compression);

        // Fall back to legacy single-shot truncation if progressive is disabled
        if !self.progressive_compression {
            tracing::info!("Progressive compression disabled, using legacy single-shot truncation");
            return self.legacy_truncate(messages, threshold);
        }

        // Try progressive compression actions in priority order
        if let Some(result) = self.try_new_segment(messages) {
            tracing::info!("Progressive action: NewSegment ({} rounds)", result.rounds_removed);
            return result;
        }

        if let Some(result) = self.try_merge_segments(messages) {
            tracing::info!("Progressive action: MergeSegments");
            return result;
        }

        if let Some(result) = self.try_discard_oldest_segment(messages) {
            tracing::info!("Progressive action: Discard oldest segment");
            return result;
        }

        // Fallback: aggressive single-shot truncation
        tracing::warn!("All progressive actions exhausted, falling back to legacy truncation");
        self.legacy_truncate(messages, threshold)
    }

    // ------------------------------------------------------------------------
    // Progressive compression: Priority 1 — new summary segment
    // ------------------------------------------------------------------------

    fn try_new_segment(
        &self,
        messages: &mut Vec<ChatCompletionRequestMessage>,
    ) -> Option<TruncationResult> {
        if !self.compression_enabled {
            return None;
        }

        // Find all full (non-summary, non-system) user rounds,
        // skipping summary segments and the discard notice.
        let round_starts: Vec<usize> = messages
            .iter()
            .enumerate()
            .filter(|(_, m)| is_user_message(m) && !is_summary_segment(m) && !is_discard_notice(m))
            .map(|(i, _)| i)
            .collect();

        if round_starts.len() <= self.min_keep_rounds + self.rounds_per_summary {
            tracing::debug!("Not enough full rounds to compress (have {}, need min_keep + rounds_per_summary = {})",
                round_starts.len(), self.min_keep_rounds + self.rounds_per_summary);
            return None;
        }

        // Take the oldest `rounds_per_summary` rounds
        let take_rounds = self.rounds_per_summary.min(round_starts.len() - self.min_keep_rounds);
        if take_rounds == 0 {
            return None;
        }

        let start_idx = round_starts[0];
        let end_idx = if take_rounds < round_starts.len() {
            round_starts[take_rounds]
        } else {
            messages.len()
        };

        let removed_messages: Vec<ChatCompletionRequestMessage> =
            messages[start_idx..end_idx].to_vec();
        let messages_removed = removed_messages.len();
        let removed_tokens = estimate_messages_tokens(&removed_messages);

        // Need enough tokens to justify compression
        if removed_tokens < self.compression_token_threshold {
            tracing::debug!("Removed tokens ({}) below compression threshold ({})",
                removed_tokens, self.compression_token_threshold);
            return None;
        }

        // Remove the rounds
        messages.drain(start_idx..end_idx);

        // Insert placeholder after system messages + discard notice (if any)
        // i.e. before any other summary segments
        let system_msg_count = messages.iter().take_while(|m| is_system_message(m)).count();
        let has_discard = find_discard_notice_pos(messages).is_some();
        let insert_pos = system_msg_count + if has_discard { 1 } else { 0 };

        let placeholder = make_summary_placeholder(take_rounds);
        messages.insert(insert_pos, placeholder);

        tracing::info!(
            "New summary segment: removed {} rounds ({} messages, ~{} tokens), insert at {}",
            take_rounds, messages_removed, removed_tokens, insert_pos
        );

        Some(TruncationResult {
            rounds_removed: take_rounds,
            messages_removed,
            removed_messages,
            insert_position: insert_pos,
            needs_compression: true,
            action: TruncationAction::NewSegment,
        })
    }

    // ------------------------------------------------------------------------
    // Progressive compression: Priority 2 — merge summary segments
    // ------------------------------------------------------------------------

    fn try_merge_segments(
        &self,
        messages: &mut Vec<ChatCompletionRequestMessage>,
    ) -> Option<TruncationResult> {
        if !self.compression_enabled {
            return None;
        }

        let segments = find_summary_segments(messages);
        if segments.len() <= self.max_summary_segments {
            tracing::debug!("Segment count ({}) within limit ({})", segments.len(), self.max_summary_segments);
            return None;
        }

        // Check if the oldest segment can still be merged
        let oldest = segments.first()?;
        if oldest.merge_level >= self.max_merges_per_segment {
            tracing::debug!("Oldest segment at merge level {} >= max {}, will discard instead",
                oldest.merge_level, self.max_merges_per_segment);
            return None;
        }

        // Take the oldest `merge_count` segments
        let take_count = self.merge_count.min(segments.len());
        if take_count < 2 {
            return None;
        }

        let start_pos = segments[0].index;
        let end_pos = segments[take_count - 1].index + 1;

        let summaries: Vec<String> = segments[..take_count]
            .iter()
            .map(|s| s.content.clone())
            .collect();

        let max_level = segments[..take_count]
            .iter()
            .map(|s| s.merge_level)
            .max()
            .unwrap_or(0);
        let new_level = max_level + 1;

        // Remove the old segments
        messages.drain(start_pos..end_pos);

        // Insert merged placeholder
        let placeholder = make_merge_placeholder(new_level, take_count);
        messages.insert(start_pos, placeholder);

        tracing::info!(
            "Merging {} summary segments into one (level {}), start pos {}",
            take_count, new_level, start_pos
        );

        Some(TruncationResult {
            rounds_removed: 0,
            messages_removed: take_count,
            removed_messages: Vec::new(),
            insert_position: start_pos,
            needs_compression: true,
            action: TruncationAction::MergeSegments {
                summaries,
                start_position: start_pos,
                count: take_count,
            },
        })
    }

    // ------------------------------------------------------------------------
    // Progressive compression: Priority 3 — discard oldest summary segment
    // ------------------------------------------------------------------------

    fn try_discard_oldest_segment(
        &self,
        messages: &mut Vec<ChatCompletionRequestMessage>,
    ) -> Option<TruncationResult> {
        let segments = find_summary_segments(messages);
        if segments.len() <= self.max_summary_segments {
            return None;
        }

        let oldest = segments.first()?;
        if oldest.merge_level < self.max_merges_per_segment {
            // Should have been handled by try_merge_segments
            return None;
        }

        // Remove the oldest segment
        let pos = oldest.index;
        messages.remove(pos);

        tracing::info!("Discarded oldest summary segment at position {}", pos);

        // Ensure discard notice exists
        let system_msg_count = messages.iter().take_while(|m| is_system_message(m)).count();
        if find_discard_notice_pos(messages).is_none() {
            let notice = make_discard_notice();
            messages.insert(system_msg_count, notice);
            tracing::debug!("Added discard notice at position {}", system_msg_count);
        }

        Some(TruncationResult {
            rounds_removed: 0,
            messages_removed: 1,
            removed_messages: Vec::new(),
            insert_position: pos,
            needs_compression: false,
            action: TruncationAction::TruncateOnly,
        })
    }

    // ------------------------------------------------------------------------
    // Legacy single-shot truncation (fallback)
    // ------------------------------------------------------------------------

    fn legacy_truncate(
        &self,
        messages: &mut Vec<ChatCompletionRequestMessage>,
        threshold: usize,
    ) -> TruncationResult {
        // Find round boundaries: a round starts with a User message
        let mut round_starts: Vec<usize> = Vec::new();
        for (i, msg) in messages.iter().enumerate() {
            if is_user_message(msg) {
                round_starts.push(i);
            }
        }
        tracing::debug!("Found {} user message round boundaries", round_starts.len());

        if round_starts.is_empty() {
            tracing::debug!("No user messages found, no truncation performed");
            return TruncationResult {
                rounds_removed: 0,
                messages_removed: 0,
                removed_messages: Vec::new(),
                insert_position: 0,
                needs_compression: false,
                action: TruncationAction::TruncateOnly,
            };
        }

        let total_rounds = round_starts.len();
        let must_keep = self.min_keep_rounds.min(total_rounds);
        tracing::debug!("Total rounds: {}, Must keep at least: {} rounds", total_rounds, must_keep);

        let mut removed_messages: Vec<ChatCompletionRequestMessage> = Vec::new();
        let mut rounds_removed = 0;
        let mut messages_removed = 0;

        while round_starts.len() > must_keep
            && estimate_messages_tokens_with_margin(messages, self.token_safety_margin) > threshold
        {
            let start_idx = round_starts[0];
            let end_idx = if round_starts.len() > 1 {
                round_starts[1]
            } else {
                messages.len()
            };

            if self.compression_enabled {
                removed_messages.extend(messages[start_idx..end_idx].to_vec());
            }

            let count = end_idx - start_idx;
            messages.drain(start_idx..end_idx);

            round_starts.remove(0);
            for idx in round_starts.iter_mut() {
                *idx = idx.saturating_sub(count);
            }

            rounds_removed += 1;
            messages_removed += count;
        }

        if rounds_removed == 0 {
            tracing::debug!("No rounds removed after checks");
            return TruncationResult {
                rounds_removed: 0,
                messages_removed: 0,
                removed_messages: Vec::new(),
                insert_position: 0,
                needs_compression: false,
                action: TruncationAction::TruncateOnly,
            };
        }

        let removed_tokens = estimate_messages_tokens(&removed_messages);
        let needs_compression =
            self.compression_enabled && removed_tokens >= self.compression_token_threshold;

        let system_msg_count = messages
            .iter()
            .take_while(|m| is_system_message(m))
            .count();

        let notice = if needs_compression {
            format!(
                "[Context compressed: {} earlier rounds ({} messages, ~{} tokens) have been summarized. {} most recent rounds preserved.]",
                rounds_removed, messages_removed, removed_tokens, round_starts.len()
            )
        } else {
            format!(
                "[Context truncated: {} earlier rounds ({} messages) removed to stay within token limit. {} most recent rounds preserved.]",
                rounds_removed, messages_removed, round_starts.len()
            )
        };

        let notice_msg = ChatCompletionRequestMessage::User(
            async_openai::types::chat::ChatCompletionRequestUserMessage {
                content: notice.into(),
                name: Some("system_notice".to_string()),
            }
            .into(),
        );

        messages.insert(system_msg_count, notice_msg);

        tracing::info!(
            "Legacy truncation: removed {} rounds ({} messages), kept {} rounds",
            rounds_removed, messages_removed, round_starts.len()
        );

        TruncationResult {
            rounds_removed,
            messages_removed,
            removed_messages,
            insert_position: system_msg_count,
            needs_compression,
            action: if needs_compression {
                // For legacy mode, treat single-shot summary as NewSegment
                // (one summary from many rounds)
                TruncationAction::NewSegment
            } else {
                TruncationAction::TruncateOnly
            },
        }
    }
}

fn is_user_message(msg: &ChatCompletionRequestMessage) -> bool {
    matches!(msg, ChatCompletionRequestMessage::User(_))
}

fn is_system_message(msg: &ChatCompletionRequestMessage) -> bool {
    matches!(msg, ChatCompletionRequestMessage::System(_))
}

// ============================================================================
// Summary segment helpers (progressive compression)
// ============================================================================

const SUMMARY_SEGMENT_PREFIX: &str = "summary_segment";
const DISCARD_NOTICE_NAME: &str = "discard_notice";
const LEGACY_NOTICE_NAME: &str = "system_notice";

/// Returns the `name` field of a User message, if any.
fn user_message_name(msg: &ChatCompletionRequestMessage) -> Option<&str> {
    match msg {
        ChatCompletionRequestMessage::User(u) => u.name.as_deref(),
        _ => None,
    }
}

/// Returns the text content of a User message, empty if not text.
fn user_message_text(msg: &ChatCompletionRequestMessage) -> String {
    match msg {
        ChatCompletionRequestMessage::User(u) => match &u.content {
            ChatCompletionRequestUserMessageContent::Text(t) => t.clone(),
            ChatCompletionRequestUserMessageContent::Array(parts) => parts
                .iter()
                .filter_map(|p| match p {
                    async_openai::types::chat::ChatCompletionRequestUserMessageContentPart::Text(t) => Some(t.text.as_str()),
                    _ => None,
                })
                .collect::<Vec<_>>()
                .join(" "),
        },
        _ => String::new(),
    }
}

/// Check whether a message is a summary segment (any merge level, including legacy notice).
pub fn is_summary_segment(msg: &ChatCompletionRequestMessage) -> bool {
    match user_message_name(msg) {
        Some(name) => {
            name.starts_with(SUMMARY_SEGMENT_PREFIX) || name == LEGACY_NOTICE_NAME
        }
        None => false,
    }
}

/// Get the merge level (number of times this segment has been merged).
/// 0 = fresh segment, 1 = merged once, etc.
pub fn get_merge_level(msg: &ChatCompletionRequestMessage) -> usize {
    match user_message_name(msg) {
        Some(name) => {
            if name == LEGACY_NOTICE_NAME {
                0
            } else if let Some(suffix) = name.strip_prefix(SUMMARY_SEGMENT_PREFIX) {
                if suffix.is_empty() {
                    0
                } else if let Some(num_str) = suffix.strip_prefix("_m") {
                    num_str.parse::<usize>().unwrap_or(0)
                } else {
                    0
                }
            } else {
                0
            }
        }
        None => 0,
    }
}

/// Check if a message is the discard notice.
fn is_discard_notice(msg: &ChatCompletionRequestMessage) -> bool {
    matches!(user_message_name(msg), Some(name) if name == DISCARD_NOTICE_NAME)
}

/// Build the `name` value for a summary segment at the given merge level.
fn summary_segment_name(merge_level: usize) -> String {
    if merge_level == 0 {
        SUMMARY_SEGMENT_PREFIX.to_string()
    } else {
        format!("{}_m{}", SUMMARY_SEGMENT_PREFIX, merge_level)
    }
}

/// Build a placeholder summary segment message (pending LLM generation).
fn make_summary_placeholder(rounds_removed: usize) -> ChatCompletionRequestMessage {
    let text = format!(
        "[Compressing {} earlier conversation rounds into a summary...]",
        rounds_removed
    );
    ChatCompletionRequestMessage::User(
        ChatCompletionRequestUserMessage {
            content: text.into(),
            name: Some(summary_segment_name(0)),
        }
        .into(),
    )
}

/// Build a placeholder for merged segments (pending LLM generation).
fn make_merge_placeholder(merge_level: usize, count: usize) -> ChatCompletionRequestMessage {
    let text = format!(
        "[Merging {} earlier summary segments...]",
        count
    );
    ChatCompletionRequestMessage::User(
        ChatCompletionRequestUserMessage {
            content: text.into(),
            name: Some(summary_segment_name(merge_level)),
        }
        .into(),
    )
}

/// Build the discard notice message.
fn make_discard_notice() -> ChatCompletionRequestMessage {
    let text = "[Note: Earlier conversation history beyond the earliest summary has been discarded to save context space.]";
    ChatCompletionRequestMessage::User(
        ChatCompletionRequestUserMessage {
            content: text.into(),
            name: Some(DISCARD_NOTICE_NAME.to_string()),
        }
        .into(),
    )
}

/// Information about a summary segment found in history.
#[derive(Debug, Clone)]
struct SummarySegmentInfo {
    index: usize,
    merge_level: usize,
    content: String,
}

/// Scan message history and collect all summary segments, ordered from oldest to newest.
fn find_summary_segments(messages: &[ChatCompletionRequestMessage]) -> Vec<SummarySegmentInfo> {
    let mut segments = Vec::new();
    for (i, msg) in messages.iter().enumerate() {
        if is_summary_segment(msg) {
            segments.push(SummarySegmentInfo {
                index: i,
                merge_level: get_merge_level(msg),
                content: user_message_text(msg),
            });
        }
    }
    segments
}

/// Find the position of the discard notice, if any.
fn find_discard_notice_pos(messages: &[ChatCompletionRequestMessage]) -> Option<usize> {
    messages.iter().position(is_discard_notice)
}

// ============================================================================
// Transcript formatting for summary compression
// ============================================================================

/// Format removed messages into a compact transcript for summary generation.
/// Extracts user messages, assistant text, and tool call names only.
/// Truncates each message to keep the transcript concise.
pub fn format_removed_messages_as_transcript(
    messages: &[ChatCompletionRequestMessage],
) -> String {
    let mut transcript = String::new();

    for msg in messages {
        match msg {
            ChatCompletionRequestMessage::User(user_msg) => {
                let text = match &user_msg.content {
                    async_openai::types::chat::ChatCompletionRequestUserMessageContent::Text(t) => {
                        t.clone()
                    }
                    async_openai::types::chat::ChatCompletionRequestUserMessageContent::Array(parts) => {
                        parts.iter()
                            .filter_map(|p| match p {
                                async_openai::types::chat::ChatCompletionRequestUserMessageContentPart::Text(t) => Some(t.text.as_str()),
                                _ => None,
                            })
                            .collect::<Vec<_>>()
                            .join(" ")
                    }
                };
                let truncated = truncate_str(&text, 200);
                transcript.push_str(&format!("User: {}\n", truncated));
            }
            ChatCompletionRequestMessage::Assistant(assistant_msg) => {
                let content_str = assistant_msg.content.as_ref().map(|c| {
                    match serde_json::to_string(c) {
                        Ok(json) => json.trim_matches('"').to_string(),
                        Err(_) => format!("{:?}", c),
                    }
                }).unwrap_or_default();
                let truncated = truncate_str(&content_str, 300);
                transcript.push_str(&format!("Assistant: {}\n", truncated));
                if let Some(tool_calls) = &assistant_msg.tool_calls {
                    for tc in tool_calls {
                        if let async_openai::types::chat::ChatCompletionMessageToolCalls::Function(f) = tc {
                            transcript.push_str(&format!(
                                "  [Tool: {}({})]\n",
                                f.function.name,
                                truncate_str(&f.function.arguments, 100)
                            ));
                        }
                    }
                }
            }
            ChatCompletionRequestMessage::Tool(tool_msg) => {
                let content_str = match serde_json::to_string(&tool_msg.content) {
                    Ok(json) => json.trim_matches('"').to_string(),
                    Err(_) => format!("{:?}", tool_msg.content),
                };
                let truncated = truncate_str(&content_str, 150);
                transcript.push_str(&format!("  [Result: {}]\n", truncated));
            }
            _ => {}
        }
    }

    if transcript.is_empty() {
        transcript.push_str("(no conversation content)");
    }

    transcript
}

/// Truncate a string to at most `max_chars` characters, adding "..." if truncated.
/// Respects UTF-8 character boundaries.
fn truncate_str(s: &str, max_chars: usize) -> String {
    if s.len() <= max_chars {
        s.to_string()
    } else {
        let mut end = max_chars;
        while end > 0 && !s.is_char_boundary(end) {
            end -= 1;
        }
        format!("{}...", &s[..end])
    }
}

// ============================================================================
// Tests
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;
    use async_openai::types::chat::ChatCompletionRequestUserMessage;

    fn make_user_message(content: &str) -> ChatCompletionRequestMessage {
        ChatCompletionRequestMessage::User(
            ChatCompletionRequestUserMessage {
                content: content.into(),
                name: None,
            }
            .into(),
        )
    }

    fn make_system_message(content: &str) -> ChatCompletionRequestMessage {
        ChatCompletionRequestMessage::System(
            async_openai::types::chat::ChatCompletionRequestSystemMessage {
                content: content.into(),
                name: None,
            }
            .into(),
        )
    }

    fn make_test_config() -> ContextConfig {
        ContextConfig {
            max_output_lines: Some(500),
            max_output_bytes: Some(51200),
            reserve_ratio: Some(0.2),
            truncation_ratio: Some(0.7),
            min_keep_rounds: Some(3),
            token_safety_margin: Some(1.3),
            compression_token_threshold: Some(5000),
            compression_enabled: Some(true),
            max_tool_calls_per_turn: Some(30),
            progressive_compression: Some(true),
            rounds_per_summary: Some(3),
            max_summary_segments: Some(5),
            merge_count: Some(2),
            max_merges_per_segment: Some(2),
            max_image_dimension: Some(1024),
        }
    }

    fn make_user_message_named(content: &str, name: &str) -> ChatCompletionRequestMessage {
        ChatCompletionRequestMessage::User(
            ChatCompletionRequestUserMessage {
                content: content.into(),
                name: Some(name.to_string()),
            }
            .into(),
        )
    }

    fn make_summary_segment(content: &str, merge_level: usize) -> ChatCompletionRequestMessage {
        let name = if merge_level == 0 {
            "summary_segment".to_string()
        } else {
            format!("summary_segment_m{}", merge_level)
        };
        make_user_message_named(content, &name)
    }

    fn make_legacy_notice(content: &str) -> ChatCompletionRequestMessage {
        make_user_message_named(content, "system_notice")
    }

    // fn make_discard_notice_msg() -> ChatCompletionRequestMessage {
    //     make_user_message_named(
    //         "[Note: Earlier conversation history beyond the earliest summary has been discarded.]",
    //         "discard_notice",
    //     )
    // }

    // ==========================================================================
    // estimate_tokens tests
    // ==========================================================================

    #[test]
    fn test_estimate_tokens_english() {
        let text = "Hello world, this is a test of the token estimation system.";
        let tokens = estimate_tokens(text);
        assert!(tokens >= 10, "Expected at least 10 tokens, got {}", tokens);
        assert!(tokens <= 30, "Expected at most 30 tokens, got {}", tokens);
    }

    #[test]
    fn test_estimate_tokens_chinese() {
        let chinese = "你好世界，这是一个测试。";
        let tokens = estimate_tokens(chinese);
        assert!(tokens >= 5, "Expected at least 5 tokens, got {}", tokens);
        assert!(tokens <= 15, "Expected at most 15 tokens, got {}", tokens);
    }

    #[test]
    fn test_estimate_tokens_code() {
        let code = "fn main() {\n    println!(\"Hello\");\n}";
        let tokens = estimate_tokens(code);
        assert!(tokens >= 10, "Expected at least 10 tokens, got {}", tokens);
        assert!(tokens <= 40, "Expected at most 40 tokens, got {}", tokens);
    }

    #[test]
    fn test_estimate_tokens_empty() {
        assert_eq!(estimate_tokens(""), 0);
    }

    #[test]
    fn test_estimate_tokens_mixed() {
        let mixed = "Hello 你好 world 世界！fn test() {}";
        let tokens = estimate_tokens(mixed);
        assert!(tokens > 0);
        assert!(tokens <= 40, "Expected at most 40 tokens, got {}", tokens);
    }

    #[test]
    fn test_estimate_messages_tokens_with_margin() {
        let messages = vec![
            make_system_message("You are a helpful assistant"),
            make_user_message("Hello world"),
        ];
        let raw = estimate_messages_tokens(&messages);
        let with_margin = estimate_messages_tokens_with_margin(&messages, 1.3);
        assert!(with_margin > raw);
        // 1.3x margin should be ~30% higher
        let expected = (raw as f32 * 1.3).ceil() as usize;
        assert_eq!(with_margin, expected);
    }

    // ==========================================================================
    // ContextManager tests
    // ==========================================================================

    #[test]
    fn test_truncation_threshold() {
        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));
        // 65536 * 0.7 = 45875
        assert_eq!(manager.truncation_threshold(), 45875);
    }

    #[test]
    fn test_truncation_result_no_truncation() {
        let mut messages = vec![
            make_system_message("You are a helpful assistant"),
            make_user_message("Hello"),
        ];

        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        assert_eq!(result.rounds_removed, 0);
        assert!(!result.needs_compression);
    }

    #[test]
    fn test_truncation_respects_min_keep_rounds() {
        let mut messages = vec![
            make_system_message("You are a helpful assistant"),
        ];

        // Add 10 rounds of large messages
        for i in 0..10 {
            let content = format!("User message {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.min_keep_rounds = Some(3); // Must keep at least 3 rounds

        // Use small context window to force aggressive truncation
        let manager = ContextManager::new(Some(5000), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        // Should have removed some rounds...
        assert!(
            result.rounds_removed > 0,
            "Should have removed some rounds"
        );
        // ...but should still have at least 3 user rounds + notice
        let user_count = messages
            .iter()
            .filter(|m| matches!(m, ChatCompletionRequestMessage::User(_)))
            .count();
        assert!(
            user_count >= 4,
            "Should have at least 3 user rounds + notice, got {}",
            user_count
        );
    }

    #[test]
    fn test_truncation_early_trigger() {
        let mut messages = vec![
            make_system_message("You are a helpful assistant"),
        ];

        // Add 8 rounds of messages, each ~2000 chars
        for i in 0..8 {
            let content = format!("User message {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.truncation_ratio = Some(0.7);
        config.min_keep_rounds = Some(2);
        config.token_safety_margin = Some(1.3);

        // With 65536 context, truncation threshold = 45875
        // 8 rounds * ~2000 chars each ≈ much less than 45875, so no truncation
        let manager = ContextManager::new(Some(65536), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);
        assert_eq!(
            result.rounds_removed, 0,
            "Should not truncate small messages in large window"
        );

        // With 8000 context, truncation threshold = 5600
        let manager2 = ContextManager::new(Some(8000), Some(&config));
        let mut messages2 = messages.clone();
        let result2 = manager2.maybe_truncate(&mut messages2, None, 0);
        assert!(
            result2.rounds_removed > 0,
            "Should truncate when exceeding small window"
        );
    }

    #[test]
    fn test_token_safety_margin_effect() {
        let mut messages = vec![
            make_system_message("You are a helpful assistant"),
        ];

        for i in 0..10 {
            let content = format!("User message {}: {}", i, "x".repeat(500));
            messages.push(make_user_message(&content));
        }

        // With margin 1.0 (no safety), truncation may not trigger
        let mut config_low = make_test_config();
        config_low.token_safety_margin = Some(1.0);
        config_low.truncation_ratio = Some(0.7);
        config_low.min_keep_rounds = Some(1);

        let mut msgs_low = messages.clone();
        let manager_low = ContextManager::new(Some(8000), Some(&config_low));
        let result_low = manager_low.maybe_truncate(&mut msgs_low, None, 0);

        // With margin 2.0 (very conservative), truncation more likely triggers
        let mut config_high = make_test_config();
        config_high.token_safety_margin = Some(2.0);
        config_high.truncation_ratio = Some(0.7);
        config_high.min_keep_rounds = Some(1);

        let mut msgs_high = messages.clone();
        let manager_high = ContextManager::new(Some(8000), Some(&config_high));
        let result_high = manager_high.maybe_truncate(&mut msgs_high, None, 0);

        // Higher margin should result in >= rounds removed
        assert!(
            result_high.rounds_removed >= result_low.rounds_removed,
            "Higher safety margin should trigger at least as much truncation: high={}, low={}",
            result_high.rounds_removed,
            result_low.rounds_removed
        );
    }

    #[test]
    fn test_compression_flag_in_old_tests() {
        let mut messages = vec![
            make_system_message("You are a helpful assistant"),
        ];

        // Add 20 rounds of large messages
        for i in 0..20 {
            let content = format!("User message {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.compression_enabled = Some(false);
        config.compression_token_threshold = Some(1000);
        config.min_keep_rounds = Some(1);

        let manager = ContextManager::new(Some(5000), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        assert!(result.rounds_removed > 0);
        assert!(
            !result.needs_compression,
            "Should be false when compression disabled"
        );
    }

    #[test]
    fn test_truncate_output() {
        let content = "line1\nline2\nline3\nline4\nline5";
        let truncated = truncate_output(content, 3, 100);
        assert!(truncated.contains("line1"));
        assert!(truncated.contains("line2"));
        assert!(truncated.contains("line3"));
        assert!(!truncated.contains("line4"));
        assert!(truncated.contains("Output truncated"));
    }

    // ==========================================================================
    // Transcript formatting tests
    // ==========================================================================

    #[test]
    fn test_truncate_str_no_truncation() {
        let result = truncate_str("hello", 10);
        assert_eq!(result, "hello");
    }

    #[test]
    fn test_truncate_str_with_truncation() {
        let result = truncate_str("hello world this is long", 10);
        assert_eq!(result, "hello worl...");
    }

    #[test]
    fn test_format_transcript_user_and_assistant() {
        let messages = vec![
            make_user_message("Fix the bug in auth.rs"),
            make_system_message("System message should be skipped"),
        ];

        let transcript = format_removed_messages_as_transcript(&messages);
        assert!(transcript.contains("User: Fix the bug in auth.rs"));
        assert!(!transcript.contains("System message"), "System messages should be skipped");
    }

    #[test]
    fn test_format_transcript_empty() {
        let messages: Vec<ChatCompletionRequestMessage> = vec![];
        let transcript = format_removed_messages_as_transcript(&messages);
        assert!(transcript.contains("no conversation content"));
    }

    #[test]
    fn test_format_transcript_truncates_long_messages() {
        let long_text = "x".repeat(500);
        let messages = vec![
            make_user_message(&long_text),
        ];

        let transcript = format_removed_messages_as_transcript(&messages);
        assert!(transcript.contains("..."));
        // Should not contain the full 500 chars
        assert!(transcript.len() < long_text.len() + 50);
    }

    // ==========================================================================
    // Progressive compression tests
    // ==========================================================================

    #[test]
    fn test_is_summary_segment_recognizes_all_levels() {
        let m0 = make_summary_segment("summary 0", 0);
        let m1 = make_summary_segment("summary 1", 1);
        let m2 = make_summary_segment("summary 2", 2);
        let legacy = make_legacy_notice("old notice");
        let normal = make_user_message("hello");

        assert!(is_summary_segment(&m0));
        assert!(is_summary_segment(&m1));
        assert!(is_summary_segment(&m2));
        assert!(is_summary_segment(&legacy));
        assert!(!is_summary_segment(&normal));
    }

    #[test]
    fn test_get_merge_level() {
        assert_eq!(get_merge_level(&make_summary_segment("a", 0)), 0);
        assert_eq!(get_merge_level(&make_summary_segment("b", 1)), 1);
        assert_eq!(get_merge_level(&make_summary_segment("c", 2)), 2);
        assert_eq!(get_merge_level(&make_legacy_notice("d")), 0);
        assert_eq!(get_merge_level(&make_user_message("e")), 0);
    }

    #[test]
    fn test_progressive_no_truncation_needed() {
        let mut messages = vec![
            make_system_message("sys"),
            make_user_message("hi"),
        ];
        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        assert_eq!(result.rounds_removed, 0);
        assert!(!result.needs_compression);
        assert_eq!(result.action, TruncationAction::TruncateOnly);
    }

    #[test]
    fn test_progressive_new_segment() {
        let mut messages = vec![make_system_message("sys")];
        // 10 rounds of large content
        for i in 0..10 {
            let content = format!("Round {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.min_keep_rounds = Some(3);
        config.rounds_per_summary = Some(3);
        config.compression_token_threshold = Some(100); // low threshold

        let manager = ContextManager::new(Some(8000), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        assert_eq!(result.action, TruncationAction::NewSegment);
        assert!(result.needs_compression);
        assert_eq!(result.rounds_removed, 3);
        assert!(result.removed_messages.len() > 0);

        // Verify the placeholder was inserted
        let has_seg = messages.iter().any(|m| is_summary_segment(m));
        assert!(has_seg, "Should have a summary segment placeholder");
    }

    #[test]
    fn test_progressive_disabled_falls_back_to_legacy() {
        let mut messages = vec![make_system_message("sys")];
        // 20 rounds of large content — definitely over threshold
        for i in 0..20 {
            let content = format!("Round {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.progressive_compression = Some(false);
        config.min_keep_rounds = Some(3);
        config.compression_token_threshold = Some(100);

        let manager = ContextManager::new(Some(8000), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        // Legacy mode removes as many rounds as needed to go below threshold,
        // which for 20 rounds of 2000 chars in an 8000-token window is more than 3.
        assert!(result.rounds_removed > 3,
            "Legacy should remove more than rounds_per_summary (3) rounds, removed {}",
            result.rounds_removed);
        // And the action type should be NewSegment (legacy single summary)
        assert!(matches!(result.action, TruncationAction::NewSegment));
    }

    #[test]
    fn test_progressive_merge_segments() {
        // Build history where full rounds are within budget (fewer than min_keep + rounds_per_summary)
        // but we have too many summary segments, forcing a merge.
        let mut messages = vec![make_system_message("sys")];
        // 6 summary segments at level 0 — exceeds max of 5, each large enough to matter
        for i in 0..6 {
            let content = format!("Summary {}: {}", i, "x".repeat(500));
            messages.push(make_summary_segment(&content, 0));
        }
        // Only 2 full user messages — less than min_keep, so NewSegment won't trigger
        for i in 0..2 {
            let content = format!("User {}: {}", i, "x".repeat(300));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.max_summary_segments = Some(5);
        config.merge_count = Some(2);
        config.min_keep_rounds = Some(3);
        config.rounds_per_summary = Some(3);
        config.compression_token_threshold = Some(10);

        // Tiny context window to force over-threshold
        let manager = ContextManager::new(Some(2000), Some(&config));

        // First check: confirm we're over threshold
        let estimated = estimate_messages_tokens_with_margin(&messages, 1.3);
        assert!(estimated > manager.truncation_threshold(),
            "Test setup error: should be over threshold, est={}, threshold={}",
            estimated, manager.truncation_threshold());

        let result = manager.maybe_truncate(&mut messages, None, 0);

        // With full rounds < min_keep + rounds_per_summary, and segments > max,
        // try_new_segment returns None, try_merge_segments should run
        match &result.action {
            TruncationAction::MergeSegments { summaries, start_position, count } => {
                assert_eq!(*count, 2, "Should merge 2 segments");
                assert_eq!(summaries.len(), 2);
                assert!(*start_position >= 1, "Start after system message");
            }
            other => {
                panic!("Expected MergeSegments, got {:?}", other);
            }
        }

        // After merge: 6 - 2 + 1 = 5 segments (including the placeholder)
        let seg_count = messages.iter().filter(|m| is_summary_segment(m)).count();
        assert_eq!(seg_count, 5, "Should have 5 segments after merge");
    }

    #[test]
    fn test_progressive_discard_after_merge_limit() {
        // 6 summary segments at merge level 2 (at the max), plus some full rounds
        let mut messages = vec![make_system_message("sys")];
        for i in 0..6 {
            messages.push(make_summary_segment(&format!("Old summary {}", i), 2));
        }
        for i in 0..5 {
            let content = format!("User {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.max_summary_segments = Some(5);
        config.max_merges_per_segment = Some(2);
        config.merge_count = Some(2);

        let manager = ContextManager::new(Some(8000), Some(&config));

        // First call may do NewSegment, so call a few times to reach discard
        let mut did_discard = false;
        for _ in 0..5 {
            let result = manager.maybe_truncate(&mut messages, None, 0);
            if result.rounds_removed == 0 && result.messages_removed > 0 && !result.needs_compression {
                // Likely a discard
                if find_discard_notice_pos(&messages).is_some() {
                    did_discard = true;
                    break;
                }
            }
        }

        // Verify segment count went down or discard notice appeared
        let seg_count = messages.iter().filter(|m| is_summary_segment(m)).count();
        assert!(
            seg_count <= 6,
            "Segment count should decrease or stay same, got {}",
            seg_count
        );

        // Just verify no panics and something happened
        let _ = did_discard;
    }

    #[test]
    fn test_legacy_notice_recognized_as_summary_segment() {
        let mut messages = vec![
            make_system_message("sys"),
            make_legacy_notice("[Old compressed context notice]"),
        ];
        // Add several full rounds to push over threshold
        for i in 0..8 {
            let content = format!("User {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.min_keep_rounds = Some(3);
        config.max_summary_segments = Some(5);
        config.compression_token_threshold = Some(100);

        let manager = ContextManager::new(Some(8000), Some(&config));
        let result = manager.maybe_truncate(&mut messages, None, 0);

        // Should not crash; legacy notice is treated as a summary segment
        assert!(result.messages_removed > 0 || result.rounds_removed > 0);
    }

    #[test]
    fn test_discard_notice_inserted_once() {
        let mut messages = vec![
            make_system_message("sys"),
            make_summary_segment("old seg 1", 2),
            make_summary_segment("old seg 2", 2),
            make_summary_segment("old seg 3", 2),
            make_summary_segment("old seg 4", 2),
            make_summary_segment("old seg 5", 2),
            make_summary_segment("old seg 6", 2),
        ];
        for i in 0..4 {
            let content = format!("User {}: {}", i, "x".repeat(2000));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.max_summary_segments = Some(5);
        config.max_merges_per_segment = Some(2);

        let manager = ContextManager::new(Some(6000), Some(&config));

        // Trigger a few discards
        for _ in 0..3 {
            let _ = manager.maybe_truncate(&mut messages, None, 0);
        }

        // Count discard notices — should be at most 1
        let discard_count = messages.iter().filter(|m| is_discard_notice(m)).count();
        assert!(
            discard_count <= 1,
            "Should have at most 1 discard notice, found {}",
            discard_count
        );
    }

    #[test]
    fn test_multiple_progressive_rounds_gradual() {
        // Build a long history and verify compression happens gradually
        let mut messages = vec![make_system_message("sys")];
        for i in 0..20 {
            let content = format!("Round {} user message: {}", i, "x".repeat(1500));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.min_keep_rounds = Some(3);
        config.rounds_per_summary = Some(3);
        config.max_summary_segments = Some(4);
        config.compression_token_threshold = Some(100);

        let manager = ContextManager::new(Some(10000), Some(&config));

        let mut seg_count_before = 0;
        let mut did_new_segment = false;
        let mut did_merge = false;

        for round in 0..10 {
            let result = manager.maybe_truncate(&mut messages, None, 0);

            let seg_count = messages.iter().filter(|m| is_summary_segment(m)).count();

            match &result.action {
                TruncationAction::NewSegment => {
                    did_new_segment = true;
                    assert_eq!(result.rounds_removed, 3);
                    assert!(seg_count > seg_count_before || seg_count_before == 0);
                }
                TruncationAction::MergeSegments { .. } => {
                    did_merge = true;
                    assert!(seg_count <= seg_count_before);
                }
                TruncationAction::TruncateOnly => {
                    // Could be discard or nothing
                }
            }

            seg_count_before = seg_count;

            let estimated = estimate_messages_tokens_with_margin(&messages, 1.3);
            if estimated <= manager.truncation_threshold() {
                break;
            }

            tracing::debug!("Round {}: segments={}, estimated={}", round, seg_count, estimated);
        }

        // With 20 rounds and a small window, we should see at least new segments
        assert!(did_new_segment, "Should have created at least one new summary segment");
        let _ = did_merge;
    }

    // ==========================================================================
    // Token calibration (estimate_context_tokens) tests
    // ==========================================================================

    #[test]
    fn test_estimate_context_tokens_fallback_no_calibration() {
        let messages = vec![
            make_system_message("You are a helpful assistant"),
            make_user_message("Hello world"),
        ];
        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));

        // No calibration data → fallback to heuristic with safety margin
        let result = manager.estimate_context_tokens(&messages, None, 0);
        let expected = estimate_messages_tokens_with_margin(&messages, 1.3);
        assert_eq!(result, expected);
    }

    #[test]
    fn test_estimate_context_tokens_calibrated_exact_match() {
        let messages = vec![
            make_system_message("sys"),
            make_user_message("hello"),
        ];
        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));

        // Calibration matches current history length exactly
        let result = manager.estimate_context_tokens(&messages, Some(100), 2);
        // No new messages → delta = 0, total = 100
        assert_eq!(result, 100);
    }

    #[test]
    fn test_estimate_context_tokens_calibrated_with_new_messages() {
        let mut messages = vec![
            make_system_message("sys"),
            make_user_message("hello"),
        ];
        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));

        // Calibration snapshot was at 2 messages with 100 tokens.
        // Now we have 3 messages (1 new).
        messages.push(make_user_message("world"));
        let result = manager.estimate_context_tokens(&messages, Some(100), 2);

        // result = 100 (baseline) + estimate(1 new message)
        let new_msg_tokens = estimate_messages_tokens(&messages[2..]);
        assert_eq!(result, 100 + new_msg_tokens);
        // Key property: calibrated uses the API baseline, NOT the heuristic.
        // The heuristic+margin may be larger or smaller depending on content.
        let heuristic = estimate_messages_tokens_with_margin(&messages, 1.3);
        // Just verify they're different (calibrated != heuristic for arbitrary baseline)
        // and that calibrated correctly adds the delta
        assert!(new_msg_tokens > 0, "New message should have non-zero token estimate");
        let _ = heuristic; // heuristic comparison is content-dependent, not asserted
    }

    #[test]
    fn test_estimate_context_tokens_stale_snapshot_falls_back() {
        let messages = vec![
            make_system_message("sys"),
            make_user_message("hello"),
        ];
        let config = make_test_config();
        let manager = ContextManager::new(Some(65536), Some(&config));

        // Snapshot was at 5 messages, but we only have 2 → history was truncated
        let result = manager.estimate_context_tokens(&messages, Some(100), 5);
        // Should fallback to heuristic because messages.len() < snapshot
        let expected = estimate_messages_tokens_with_margin(&messages, 1.3);
        assert_eq!(result, expected);
    }

    #[test]
    fn test_maybe_truncate_with_calibration_avoids_premature_truncation() {
        // Build messages that are near threshold by heuristic but clearly under by calibration
        let mut messages = vec![make_system_message("sys")];
        for i in 0..5 {
            let content = format!("User {}: {}", i, "x".repeat(500));
            messages.push(make_user_message(&content));
        }

        let mut config = make_test_config();
        config.min_keep_rounds = Some(3);

        // Use small window so heuristic+margin would trigger truncation
        let manager = ContextManager::new(Some(3000), Some(&config));

        // Without calibration: heuristic likely triggers truncation
        let mut msgs_no_cal = messages.clone();
        let result_no_cal = manager.maybe_truncate(&mut msgs_no_cal, None, 0);

        // With calibration: API says we're at exactly 500 tokens (well under 3000*0.7=2100)
        let mut msgs_cal = messages.clone();
        let result_cal = manager.maybe_truncate(&mut msgs_cal, Some(500), messages.len());

        // Calibrated should NOT truncate (500 < 2100 threshold)
        assert_eq!(result_cal.rounds_removed, 0,
            "Calibrated estimation should not trigger premature truncation");

        // Log for visibility (the no-cal case may or may not truncate depending on exact heuristic)
        tracing::debug!(
            "no_cal: rounds_removed={}, cal: rounds_removed={}",
            result_no_cal.rounds_removed, result_cal.rounds_removed
        );
    }
}