claude_session_parser/

usage.rs

use chrono::DateTime;
use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use std::fs;
use std::io::{BufRead, BufReader};
use std::path::{Path, PathBuf};
use std::time::SystemTime;

use super::activity::{classify, normalize_user_text, Activity};

/// Pricing per 1M tokens (USD) by model family.
pub struct ModelPricing {
    pub input: f64,
    pub output: f64,
    pub cache_read: f64,
    pub cache_write: f64,
}

/// Hardcoded as of 2025-02. Requires a code change + release when Anthropic updates pricing.
pub fn pricing_for_model(model: &str) -> ModelPricing {
    if model.contains("opus") {
        ModelPricing {
            input: 5.0,
            output: 25.0,
            cache_read: 0.5,
            cache_write: 6.25,
        }
    } else if model.contains("haiku") {
        ModelPricing {
            input: 1.0,
            output: 5.0,
            cache_read: 0.1,
            cache_write: 1.25,
        }
    } else {
        // Sonnet or unknown
        ModelPricing {
            input: 3.0,
            output: 15.0,
            cache_read: 0.3,
            cache_write: 3.75,
        }
    }
}

pub fn context_window_for_model(model: &str) -> u64 {
    if model.contains("[1m]") {
        1_000_000
    } else {
        200_000
    }
}

/// Flat row representing a single completed assistant turn.
///
/// One entry per unique `message.id`. Claude Code writes multiple JSONL
/// entries per assistant message as the stream progresses (typically 2-3),
/// all sharing the same `id` and final usage numbers — dedup keeps only the
/// first occurrence.
///
/// Token field semantics (inherited from Anthropic's API):
/// - `input_tokens`: *fresh* tokens in this turn's prompt that were NOT served
///   from cache. Small when conversation is reused, large on cache misses.
/// - `cache_read_tokens`: tokens served from the prompt cache (cheap).
/// - `cache_write_tokens`: tokens written into the prompt cache (~25% premium).
/// - `output_tokens`: tokens generated by the model.
///
/// "Total context sent" for a turn = input + cache_read + cache_write.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct UsageEntry {
    pub session_id: String,
    /// Unix epoch ms from the JSONL `timestamp` field. 0 if unparseable.
    pub timestamp_ms: u64,
    pub model: String,
    pub input_tokens: u64,
    pub output_tokens: u64,
    pub cache_read_tokens: u64,
    pub cache_write_tokens: u64,
    /// Names of tools this turn invoked (from `tool_use` content blocks).
    /// Empty for pure-text turns. Duplicates preserved when the model called
    /// the same tool multiple times in one turn.
    pub tool_uses: Vec<String>,
    /// Sum of `text` content block character counts for this turn. Zero means
    /// a silent tool-only turn; non-zero means the model said something.
    pub text_chars: u32,
    /// Classified activity category (Coding, Debugging, Exploration, …).
    /// See the [`crate::activity`] module for the decision tree.
    pub activity: Activity,
    /// JSONL `entrypoint` field captured from the first row of each message id.
    /// `"cli"` = interactive Claude Code session. `"sdk-cli"` = headless
    /// `claude -p` spawn (any Agent SDK consumer lands here). `None` for legacy
    /// rows that predate the field. Useful for distinguishing interactive vs.
    /// headless usage; also load-bearing once Anthropic splits Agent SDK
    /// billing from interactive plans (announced June 2026).
    pub entrypoint: Option<String>,
}

/// A single hook execution event from a `type: "attachment"` JSONL record
/// where `attachment.type == "hook_success"`.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct HookExecution {
    pub session_id: String,
    pub timestamp_ms: u64,
    /// Lifecycle phase: `PreToolUse`, `PostToolUse`, `SessionStart`, `Stop`, etc.
    pub hook_event: String,
    /// Full hook label including matcher: `PreToolUse:Bash`, `Stop`, etc.
    pub hook_name: String,
    /// The shell command that was executed.
    pub command: String,
    pub exit_code: i32,
    pub duration_ms: u64,
    /// The tool-use ID that triggered this hook (groups multiple hooks per tool call).
    pub tool_use_id: Option<String>,
}

/// Combined result from parsing a single JSONL file.
#[derive(Default)]
pub struct ParsedJsonl {
    pub entries: Vec<UsageEntry>,
    pub hooks: Vec<HookExecution>,
}

/// Usage summary for a single session, returned to the frontend status bar.
#[derive(Debug, Clone, Serialize, Default)]
#[serde(rename_all = "camelCase")]
pub struct SessionUsage {
    pub model: String,
    pub input_tokens: u64,
    pub output_tokens: u64,
    pub cache_read_tokens: u64,
    pub cache_write_tokens: u64,
    pub total_tokens: u64,
    pub estimated_cost_usd: f64,
    pub context_window_size: u64,
    pub context_window_used_pct: f64,
    /// Last message's total input context (current context window usage).
    pub current_context_tokens: u64,
    /// Cumulative context sent (input + cache_read + cache_write).
    pub total_context_sent: u64,
    /// cache_read / (input + cache_read), in percent. 0 when no cache activity.
    pub cache_hit_pct: f64,
    pub message_count: u64,
    pub session_id: String,
}

/// Minimal JSONL entry structure.
#[derive(Deserialize)]
#[serde(rename_all = "camelCase")]
struct JsonlRaw {
    #[serde(rename = "type")]
    entry_type: Option<String>,
    timestamp: Option<String>,
    message: Option<MessageData>,
    attachment: Option<AttachmentData>,
    // Fields for `type: "system"` hook summary records (v2.1.119+).
    subtype: Option<String>,
    #[allow(dead_code)]
    hook_count: Option<usize>,
    hook_infos: Option<Vec<HookInfoEntry>>,
    hook_errors: Option<Vec<serde_json::Value>>,
    prevented_continuation: Option<bool>,
    #[serde(rename = "toolUseID")]
    tool_use_id: Option<String>,
    /// Origin of the session: `"cli"` (interactive Claude Code) or `"sdk-cli"`
    /// (headless `claude -p` / Agent SDK). Present on every assistant + user row.
    entrypoint: Option<String>,
}

#[derive(Deserialize)]
struct MessageData {
    id: Option<String>,
    model: Option<String>,
    usage: Option<UsageData>,
    /// Raw content — assistant messages use `Vec<ContentBlock>`, user messages
    /// use either a string or an array of blocks (including tool_results).
    /// Kept as a raw Value so we can handle both shapes without maintaining
    /// two parallel structs.
    #[serde(default)]
    content: Option<serde_json::Value>,
}

#[derive(Deserialize, Default)]
struct UsageData {
    input_tokens: Option<u64>,
    output_tokens: Option<u64>,
    cache_read_input_tokens: Option<u64>,
    cache_creation_input_tokens: Option<u64>,
}

/// Single hook execution in a `hookInfos` array (v2.1.119+ system records).
#[derive(Deserialize)]
#[serde(rename_all = "camelCase")]
struct HookInfoEntry {
    command: Option<String>,
    duration_ms: Option<u64>,
}

/// Minimal deserialization of the `attachment` field on `type: "attachment"` rows.
#[derive(Deserialize)]
#[serde(rename_all = "camelCase")]
struct AttachmentData {
    #[serde(rename = "type")]
    attachment_type: Option<String>,
    hook_name: Option<String>,
    hook_event: Option<String>,
    command: Option<String>,
    exit_code: Option<i32>,
    duration_ms: Option<u64>,
    #[serde(rename = "toolUseID")]
    tool_use_id: Option<String>,
}

/// Claude Code active session file in `{config_dir}/sessions/{pid}.json`.
#[derive(Deserialize)]
#[serde(rename_all = "camelCase")]
struct ActiveSession {
    session_id: String,
    cwd: String,
    started_at: u64,
}

/// Hard cap to avoid pathological parsing of a corrupted file.
///
/// Parsing is streaming (`BufReader::lines()`) so memory is O(1) per line —
/// the cap only exists to keep a runaway file from blocking the UI. Real
/// agent sessions (Task-tool sub-agents) frequently exceed 30MB, so the cap
/// needs generous headroom.
use super::MAX_JSONL_BYTES;

/// Encode a project path the same way Claude Code does.
///
/// Windows: `S:\Loom` -> `S--Loom`, `C:\Users\Seven\foo` -> `C--Users-Seven-foo`
/// macOS/Linux: `/Users/seven/Loom` -> `-Users-seven-Loom`, `/home/alice/proj` -> `-home-alice-proj`
///
/// The transform collapses any leading path separator(s) into a single leading
/// dash, which is how Claude Code writes `~/.claude/projects/` directory names
/// on Unix. Verified against ccusage's decoder (see GitHub issue
/// anthropics/claude-code#19972 for the path-encoding discussion).
pub fn encode_project_path(project_path: &str) -> String {
    let normalized = project_path.replace('/', "\\");
    let mut result = String::new();
    let mut chars = normalized.chars().peekable();

    while let Some(ch) = chars.next() {
        if ch == ':' {
            if chars.peek() == Some(&'\\') {
                chars.next();
                result.push_str("--");
            } else {
                result.push('-');
            }
        } else if ch == '\\' {
            result.push('-');
        } else {
            result.push(ch);
        }
    }

    while result.ends_with('-') {
        result.pop();
    }

    result
}

/// Extracted signal from an assistant message's `content` array.
struct AssistantContent {
    tool_uses: Vec<String>,
    bash_cmds: Vec<String>,
    text_chars: u32,
}

/// Walk an assistant message's content blocks to pull out tool names, any
/// bash-style commands (used by the classifier for Testing/GitOps/Build
/// detection), and the total character count of `text` blocks.
fn extract_assistant_content(content: Option<&serde_json::Value>) -> AssistantContent {
    let mut out = AssistantContent {
        tool_uses: Vec::new(),
        bash_cmds: Vec::new(),
        text_chars: 0,
    };
    let arr = match content.and_then(|v| v.as_array()) {
        Some(a) => a,
        None => return out,
    };

    for block in arr {
        let block_type = match block.get("type").and_then(|v| v.as_str()) {
            Some(t) => t,
            None => continue,
        };
        match block_type {
            "tool_use" => {
                if let Some(name) = block.get("name").and_then(|v| v.as_str()) {
                    // If this is a Bash-family tool, grab the command string so
                    // the classifier can distinguish Testing / GitOps / Build.
                    if matches!(name, "Bash" | "BashTool" | "PowerShellTool") {
                        if let Some(cmd) = block
                            .get("input")
                            .and_then(|i| i.get("command"))
                            .and_then(|c| c.as_str())
                        {
                            out.bash_cmds.push(cmd.to_string());
                        }
                    }
                    out.tool_uses.push(name.to_string());
                }
            }
            "text" => {
                if let Some(text) = block.get("text").and_then(|v| v.as_str()) {
                    // Saturating because a single turn's text can exceed u32
                    // only in pathological cases (would be >4 GB of text).
                    out.text_chars = out
                        .text_chars
                        .saturating_add(text.chars().count() as u32);
                }
            }
            _ => {}
        }
    }

    out
}

/// Extract the human-typed portion of a user message's content, if any.
///
/// User rows in Claude Code JSONL carry two distinct shapes:
/// - A plain string (`"content": "hello"`) — always a real user turn.
/// - An array of content blocks, which may be:
///   - `[{"type":"text","text":"..."}]` → real user turn
///   - `[{"type":"tool_result",...}]` → injected tool output, NOT a user turn
///   - Mixed → treated as a real user turn
///
/// Sanitize the per-hook `command` for use inside a synthesized `hook_name`
/// bucket key. Takes the first line, collapses internal whitespace, and caps
/// at 64 chars (with `…` suffix). The full command is preserved separately on
/// the `HookExecution.command` field — this only affects the display label /
/// aggregation key used by the Usage page and Refinery hook stats.
///
/// Without this, `/goal`-style hooks whose "command" is a multi-KB goal
/// statement explode the hook-table layout AND defeat aggregation (one bucket
/// per invocation).
fn sanitize_hook_command_label(cmd: &str) -> String {
    const MAX_LEN: usize = 64;
    let first_line = cmd.lines().next().unwrap_or("").trim();
    let collapsed: String = first_line.split_whitespace().collect::<Vec<_>>().join(" ");
    if collapsed.chars().count() <= MAX_LEN {
        return collapsed;
    }
    let truncated: String = collapsed.chars().take(MAX_LEN).collect();
    format!("{truncated}…")
}

/// Convert `snake_case` to `PascalCase`: `"stop"` → `"Stop"`, `"pre_tool_use"` → `"PreToolUse"`.
fn snake_to_pascal(s: &str) -> String {
    s.split('_')
        .filter(|w| !w.is_empty())
        .map(|w| {
            let mut c = w.chars();
            match c.next() {
                Some(first) => {
                    let mut out = first.to_uppercase().to_string();
                    out.extend(c);
                    out
                }
                None => String::new(),
            }
        })
        .collect()
}

/// Returns `None` for tool-result-only rows so the rolling `last_user_text`
/// stays pinned to the most recent actual user message.
fn extract_user_text(content: &serde_json::Value) -> Option<String> {
    match content {
        serde_json::Value::String(s) => {
            if s.is_empty() {
                None
            } else {
                Some(s.clone())
            }
        }
        serde_json::Value::Array(arr) => {
            let mut parts: Vec<&str> = Vec::new();
            for item in arr {
                let t = item.get("type").and_then(|v| v.as_str()).unwrap_or("");
                if t == "text" {
                    if let Some(text) = item.get("text").and_then(|v| v.as_str()) {
                        if !text.is_empty() {
                            parts.push(text);
                        }
                    }
                }
            }
            if parts.is_empty() {
                None
            } else {
                Some(parts.join(" "))
            }
        }
        _ => None,
    }
}

/// Per-message-id accumulator for content blocks that arrive split across
/// multiple JSONL rows. See `parse_jsonl_entries` for the why.
struct EntryAccumulator {
    out_idx: usize,
    bash_cmds: Vec<String>,
    user_text_at_first: Option<String>,
}

/// Parse a single JSONL file into one `UsageEntry` per unique `message.id`.
///
/// Pure function: depends only on the file contents. Non-assistant entries,
/// malformed lines, and missing IDs are skipped silently.
///
/// **Dedup is split by field semantics**, because Claude Code emits one JSONL
/// row per assistant content block (thinking / text / each tool_use) all
/// sharing the same `message.id`:
///
/// - Billing fields (token counts, model, timestamp, session_id) come from the
///   **first** row seen for each id — they repeat identically across rows.
/// - Content fields (`tool_uses`, `text_chars`, and the bash commands fed to
///   the activity classifier) are **unioned across every row** with that id,
///   then `activity` is classified once at the end using the merged content.
///
/// Folding only by `message.id` (as this parser did before 2026-05-13) drops
/// every tool_use after the first row, which made the Toolkit and Usage tool
/// counters undercount by an order of magnitude. The chat parser already
/// uses `(message_id + content)` for the same reason — see CLAUDE.md.
///
/// Runs a rolling `last_user_text` cursor during the stream: each genuine
/// `type: "user"` row updates it; the snapshot taken at the **first** row of
/// each assistant message is used for activity classification.
pub fn parse_jsonl_entries(path: &Path) -> std::io::Result<ParsedJsonl> {
    let meta = fs::metadata(path)?;
    if meta.len() > MAX_JSONL_BYTES {
        tracing::warn!(
            path = %path.display(),
            size = meta.len(),
            "skipping JSONL larger than {}MB",
            MAX_JSONL_BYTES / 1024 / 1024
        );
        return Ok(ParsedJsonl::default());
    }

    let file = fs::File::open(path)?;
    let reader = BufReader::new(file);

    let session_id = path
        .file_stem()
        .and_then(|s| s.to_str())
        .unwrap_or("")
        .to_string();

    let mut accums: HashMap<String, EntryAccumulator> = HashMap::new();
    let mut out: Vec<UsageEntry> = Vec::new();
    let mut hooks: Vec<HookExecution> = Vec::new();
    let mut last_user_text: Option<String> = None;

    for line in reader.lines() {
        let line = match line {
            Ok(l) => l,
            Err(_) => continue,
        };
        let line = line.trim();
        if line.is_empty() {
            continue;
        }

        let raw: JsonlRaw = match serde_json::from_str(line) {
            Ok(r) => r,
            Err(_) => continue,
        };

        let entry_type = raw.entry_type.as_deref();

        // Update the rolling user-message cursor on genuine user turns.
        // Tool-result-only rows (also type="user") leave it untouched.
        if entry_type == Some("user") {
            if let Some(msg) = raw.message.as_ref() {
                if let Some(content) = msg.content.as_ref() {
                    if let Some(text) = extract_user_text(content) {
                        last_user_text = Some(normalize_user_text(&text));
                    }
                }
            }
            continue;
        }

        // Extract hook execution events from attachment records (v2.1.101 format).
        if entry_type == Some("attachment") {
            if let Some(att) = raw.attachment {
                if att.attachment_type.as_deref() == Some("hook_success") {
                    let timestamp_ms = raw
                        .timestamp
                        .as_deref()
                        .and_then(|s| DateTime::parse_from_rfc3339(s).ok())
                        .map(|dt| dt.timestamp_millis().max(0) as u64)
                        .unwrap_or(0);
                    hooks.push(HookExecution {
                        session_id: session_id.clone(),
                        timestamp_ms,
                        hook_event: att.hook_event.unwrap_or_default(),
                        hook_name: att.hook_name.unwrap_or_default(),
                        command: att.command.unwrap_or_default(),
                        exit_code: att.exit_code.unwrap_or(0),
                        duration_ms: att.duration_ms.unwrap_or(0),
                        tool_use_id: att.tool_use_id,
                    });
                }
            }
            continue;
        }

        // Extract hook summaries from system records (v2.1.119+ format).
        // Subtype pattern: "<event>_hook_summary" → hook_event = "<Event>".
        if entry_type == Some("system") {
            if let Some(subtype) = raw.subtype.as_deref() {
                if let Some(event_snake) = subtype.strip_suffix("_hook_summary") {
                    let hook_event = snake_to_pascal(event_snake);
                    let timestamp_ms = raw
                        .timestamp
                        .as_deref()
                        .and_then(|s| DateTime::parse_from_rfc3339(s).ok())
                        .map(|dt| dt.timestamp_millis().max(0) as u64)
                        .unwrap_or(0);
                    let had_errors = raw
                        .hook_errors
                        .as_ref()
                        .is_some_and(|e| !e.is_empty());
                    let prevented = raw.prevented_continuation.unwrap_or(false);
                    if let Some(infos) = raw.hook_infos {
                        for info in infos {
                            let cmd = info.command.unwrap_or_default();
                            let hook_name =
                                format!("{}:{}", hook_event, sanitize_hook_command_label(&cmd));
                            hooks.push(HookExecution {
                                session_id: session_id.clone(),
                                timestamp_ms,
                                hook_event: hook_event.clone(),
                                hook_name,
                                command: cmd,
                                exit_code: if had_errors || prevented { 1 } else { 0 },
                                duration_ms: info.duration_ms.unwrap_or(0),
                                tool_use_id: raw.tool_use_id.clone(),
                            });
                        }
                    }
                }
            }
            continue;
        }

        if entry_type != Some("assistant") {
            continue;
        }

        let msg = match raw.message {
            Some(m) => m,
            None => continue,
        };
        let mid = match msg.id {
            Some(id) => id,
            None => continue,
        };

        let mut content_parsed = extract_assistant_content(msg.content.as_ref());

        // Subsequent row for an already-seen message: union its content into
        // the existing entry. Billing/model/timestamp on the entry stay as-is.
        if let Some(acc) = accums.get_mut(&mid) {
            let entry = &mut out[acc.out_idx];
            entry
                .tool_uses
                .append(&mut content_parsed.tool_uses);
            entry.text_chars = entry
                .text_chars
                .saturating_add(content_parsed.text_chars);
            acc.bash_cmds.append(&mut content_parsed.bash_cmds);
            continue;
        }

        // Skip synthetic entries. Claude Code / the Agent SDK writes
        // `model: "<synthetic>"` rows for internal routing (skill listings,
        // queue ops, tool dispatch, etc.) — these never hit the LLM and have
        // all-zero token counts. Counting them would inflate turn counts
        // without representing real model activity.
        let model = msg.model.unwrap_or_default();
        if model == "<synthetic>" {
            continue;
        }

        let usage = msg.usage.unwrap_or_default();
        let timestamp_ms = raw
            .timestamp
            .as_deref()
            .and_then(|s| DateTime::parse_from_rfc3339(s).ok())
            .map(|dt| dt.timestamp_millis().max(0) as u64)
            .unwrap_or(0);

        // First row for this message: push placeholder entry with the content
        // we've seen so far. Activity is classified at the end of the file
        // against the merged content from all rows sharing this `message.id`.
        accums.insert(
            mid.clone(),
            EntryAccumulator {
                out_idx: out.len(),
                bash_cmds: std::mem::take(&mut content_parsed.bash_cmds),
                user_text_at_first: last_user_text.clone(),
            },
        );
        out.push(UsageEntry {
            session_id: session_id.clone(),
            timestamp_ms,
            model,
            input_tokens: usage.input_tokens.unwrap_or(0),
            output_tokens: usage.output_tokens.unwrap_or(0),
            cache_read_tokens: usage.cache_read_input_tokens.unwrap_or(0),
            cache_write_tokens: usage.cache_creation_input_tokens.unwrap_or(0),
            tool_uses: content_parsed.tool_uses,
            text_chars: content_parsed.text_chars,
            activity: Activity::General,
            entrypoint: raw.entrypoint.clone(),
        });
    }

    // Classify each accumulated entry using the merged content from all rows
    // that shared its `message.id`. Order doesn't matter — we mutate in place.
    for acc in accums.values() {
        let entry = &mut out[acc.out_idx];
        let tool_refs: Vec<&str> = entry.tool_uses.iter().map(|s| s.as_str()).collect();
        let bash_refs: Vec<&str> = acc.bash_cmds.iter().map(|s| s.as_str()).collect();
        entry.activity = classify(
            &tool_refs,
            &bash_refs,
            acc.user_text_at_first.as_deref(),
            entry.text_chars,
        );
    }

    Ok(ParsedJsonl {
        entries: out,
        hooks,
    })
}

/// Find the JSONL file for the current session by looking up Claude Code's
/// active sessions directory, matching on project path.
/// Falls back to finding the newest JSONL created after `spawned_at`.
pub fn find_current_session_jsonl(
    config_dir: &str,
    project_path: &str,
    spawned_at: SystemTime,
) -> Option<PathBuf> {
    let config = Path::new(config_dir);
    let encoded = encode_project_path(project_path);
    let projects_dir = config.join("projects").join(&encoded);

    if !projects_dir.exists() {
        return None;
    }

    // Strategy 1: Look up Claude Code's sessions/ directory for an active
    // session matching our project path. This gives us the exact session ID.
    let sessions_dir = config.join("sessions");
    if sessions_dir.exists() {
        if let Some(session_id) = find_active_session_id(&sessions_dir, project_path, spawned_at) {
            let jsonl_path = projects_dir.join(format!("{}.jsonl", session_id));
            if jsonl_path.exists() {
                return Some(jsonl_path);
            }
        }
    }

    // Strategy 2: Fall back to finding the earliest-created JSONL after spawn
    // time. "Earliest" is critical: with two concurrent sessions for the same
    // project, the older session must claim the first JSONL created after its
    // spawn, not the latest one (which belongs to the newer session).
    let mut candidates: Vec<(PathBuf, SystemTime)> = Vec::new();

    if let Ok(entries) = fs::read_dir(&projects_dir) {
        for entry in entries.flatten() {
            let path = entry.path();
            if path.extension().and_then(|e| e.to_str()) != Some("jsonl") {
                continue;
            }
            if let Ok(meta) = path.metadata() {
                let created = meta.created().unwrap_or(SystemTime::UNIX_EPOCH);
                if created >= spawned_at {
                    candidates.push((path, created));
                }
            }
        }
    }

    // Pick the earliest-created file after our spawn time
    candidates.sort_by(|a, b| a.1.cmp(&b.1));
    candidates.into_iter().next().map(|(p, _)| p)
}

/// Search Claude Code's sessions/ directory for an active session matching
/// the project path and started after our spawn time.
fn find_active_session_id(
    sessions_dir: &Path,
    project_path: &str,
    spawned_at: SystemTime,
) -> Option<String> {
    let spawned_ms = spawned_at
        .duration_since(SystemTime::UNIX_EPOCH)
        .ok()?
        .as_millis() as u64;

    // Allow 10s grace period (Claude Code may start slightly after PTY spawn)
    let threshold_ms = spawned_ms.saturating_sub(10_000);

    let normalized_project = normalize_path(project_path);

    let mut best: Option<(String, u64)> = None;

    if let Ok(entries) = fs::read_dir(sessions_dir) {
        for entry in entries.flatten() {
            let path = entry.path();
            if path.extension().and_then(|e| e.to_str()) != Some("json") {
                continue;
            }
            if let Ok(content) = fs::read_to_string(&path) {
                if let Ok(session) = serde_json::from_str::<ActiveSession>(&content) {
                    let session_cwd = normalize_path(&session.cwd);
                    if session_cwd == normalized_project && session.started_at >= threshold_ms {
                        // Pick the earliest session started after our spawn
                        // threshold. With multiple concurrent sessions for the
                        // same project, each has a different spawned_at: the
                        // older session's threshold is lower, so it must claim
                        // the first matching Claude session, not the latest one.
                        if best.as_ref().is_none_or(|(_, t)| session.started_at < *t) {
                            best = Some((session.session_id, session.started_at));
                        }
                    }
                }
            }
        }
    }

    best.map(|(id, _)| id)
}

/// Normalize a path for comparison: lowercase, forward slashes, no trailing slash.
fn normalize_path(p: &str) -> String {
    p.replace('\\', "/")
        .trim_end_matches('/')
        .to_lowercase()
}

/// Reduce a list of `UsageEntry` for a single session into a `SessionUsage`.
///
/// Preserves existing semantics:
/// - Primary model = the model used by the most messages (mode).
/// - Context window % uses the LAST entry's total input (API resends the full
///   conversation each turn; the last turn's input is the current usage).
/// - `launched_model` (from launch options) is used for context window sizing
///   because JSONL strips the `[1m]` suffix; fall back to mode when empty.
pub fn reduce_session(
    entries: &[UsageEntry],
    launched_model: &str,
    session_id_hint: &str,
) -> Option<SessionUsage> {
    if entries.is_empty() {
        return None;
    }

    let mut input_tokens = 0u64;
    let mut output_tokens = 0u64;
    let mut cache_read = 0u64;
    let mut cache_write = 0u64;
    let mut model_counts: HashMap<&str, u64> = HashMap::new();

    for e in entries {
        input_tokens += e.input_tokens;
        output_tokens += e.output_tokens;
        cache_read += e.cache_read_tokens;
        cache_write += e.cache_write_tokens;
        if !e.model.is_empty() {
            *model_counts.entry(e.model.as_str()).or_insert(0) += 1;
        }
    }

    let primary_model = model_counts
        .into_iter()
        .max_by_key(|(_, c)| *c)
        .map(|(m, _)| m.to_string())
        .unwrap_or_else(|| {
            entries
                .last()
                .map(|e| e.model.clone())
                .unwrap_or_default()
        });

    let pricing = pricing_for_model(&primary_model);
    let estimated_cost = (input_tokens as f64 * pricing.input
        + output_tokens as f64 * pricing.output
        + cache_read as f64 * pricing.cache_read
        + cache_write as f64 * pricing.cache_write)
        / 1_000_000.0;

    let context_model = if launched_model.is_empty() {
        &primary_model
    } else {
        launched_model
    };
    let context_window = context_window_for_model(context_model);

    let last = entries.last().unwrap();
    let last_context_used =
        last.input_tokens + last.cache_read_tokens + last.cache_write_tokens;
    let context_pct = if context_window > 0 {
        (last_context_used as f64 / context_window as f64 * 100.0).min(100.0)
    } else {
        0.0
    };

    let total = input_tokens + output_tokens + cache_read + cache_write;
    let total_context_sent = input_tokens + cache_read + cache_write;

    // Cache hit % reflects "what fraction of billable input-side tokens came
    // from cache". Denominator includes cache_write because writes are fresh
    // input being laid down for later reuse (billed at ~125% of input rate) —
    // they are NOT a hit. Only cache_read counts as reusing prior work.
    let cache_denom = input_tokens + cache_read + cache_write;
    let cache_hit_pct = if cache_denom > 0 {
        (cache_read as f64 / cache_denom as f64) * 100.0
    } else {
        0.0
    };

    let session_id = entries
        .first()
        .map(|e| e.session_id.clone())
        .filter(|s| !s.is_empty())
        .unwrap_or_else(|| session_id_hint.to_string());

    Some(SessionUsage {
        model: primary_model,
        input_tokens,
        output_tokens,
        cache_read_tokens: cache_read,
        cache_write_tokens: cache_write,
        total_tokens: total,
        estimated_cost_usd: estimated_cost,
        context_window_size: context_window,
        context_window_used_pct: context_pct,
        current_context_tokens: last_context_used,
        total_context_sent,
        cache_hit_pct,
        message_count: entries.len() as u64,
        session_id,
    })
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::io::Write;
    use tempfile::TempDir;

    #[test]
    fn test_encode_project_path_drive_root() {
        assert_eq!(encode_project_path("S:\\Loom"), "S--Loom");
    }

    #[test]
    fn test_encode_project_path_deep() {
        assert_eq!(
            encode_project_path("C:\\Users\\Seven\\Desktop\\SF\\project"),
            "C--Users-Seven-Desktop-SF-project"
        );
    }

    #[test]
    fn test_encode_project_path_forward_slashes() {
        assert_eq!(encode_project_path("S:/Loom"), "S--Loom");
    }

    #[test]
    fn test_encode_project_path_trailing_slash() {
        assert_eq!(encode_project_path("S:\\Loom\\"), "S--Loom");
    }

    #[test]
    fn test_encode_project_path_macos() {
        assert_eq!(
            encode_project_path("/Users/seven/Loom"),
            "-Users-seven-Loom"
        );
    }

    #[test]
    fn test_encode_project_path_linux() {
        assert_eq!(
            encode_project_path("/home/alice/projects/foo"),
            "-home-alice-projects-foo"
        );
    }

    #[test]
    fn test_encode_project_path_unix_trailing_slash() {
        assert_eq!(encode_project_path("/Users/seven/Loom/"), "-Users-seven-Loom");
    }

    #[test]
    fn test_normalize_path() {
        assert_eq!(normalize_path("S:\\Loom\\"), "s:/loom");
        assert_eq!(normalize_path("S:/Loom"), "s:/loom");
        assert_eq!(normalize_path("C:\\Users\\Seven\\"), "c:/users/seven");
    }

    fn write_jsonl(lines: &[&str]) -> (TempDir, PathBuf) {
        let tmp = TempDir::new().unwrap();
        let path = tmp.path().join("abc123.jsonl");
        let mut f = fs::File::create(&path).unwrap();
        for l in lines {
            writeln!(f, "{}", l).unwrap();
        }
        (tmp, path)
    }

    fn assistant_line(id: &str, model: &str, ts: &str, inp: u64, out: u64, cr: u64, cw: u64) -> String {
        format!(
            r#"{{"type":"assistant","timestamp":"{ts}","message":{{"id":"{id}","model":"{model}","usage":{{"input_tokens":{inp},"output_tokens":{out},"cache_read_input_tokens":{cr},"cache_creation_input_tokens":{cw}}}}}}}"#
        )
    }

    #[test]
    fn parse_deduplicates_repeated_message_ids() {
        // Mirrors real Claude Code JSONL: each assistant message is written 2-3
        // times with identical usage as the stream progresses.
        let a1 = assistant_line("msg_01", "claude-opus-4-6", "2026-04-14T08:27:38.895Z", 15, 1463, 16397, 25661);
        let a1_dup = a1.clone();
        let a2 = assistant_line("msg_02", "claude-opus-4-6", "2026-04-14T08:34:37.200Z", 10, 401, 42058, 1668);
        let a2_dup = a2.clone();
        let a2_dup2 = a2.clone();
        let (_tmp, path) = write_jsonl(&[&a1, &a1_dup, &a2, &a2_dup, &a2_dup2]);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 2, "should dedup by message.id");
        assert_eq!(entries[0].input_tokens, 15);
        assert_eq!(entries[0].output_tokens, 1463);
        assert_eq!(entries[1].input_tokens, 10);
    }

    #[test]
    fn parse_captures_entrypoint_from_assistant_row() {
        // Headless `claude -p` writes JSONL with entrypoint="sdk-cli" on every
        // assistant row. Interactive Claude Code writes "cli". A row missing
        // the field (legacy/malformed) yields None.
        let headless = r#"{"type":"assistant","timestamp":"2026-05-13T07:19:12.573Z","entrypoint":"sdk-cli","message":{"id":"msg_h1","model":"claude-opus-4-7","usage":{"input_tokens":5,"output_tokens":628,"cache_read_input_tokens":0,"cache_creation_input_tokens":14266}}}"#;
        let interactive = r#"{"type":"assistant","timestamp":"2026-05-13T08:00:00.000Z","entrypoint":"cli","message":{"id":"msg_i1","model":"claude-opus-4-6","usage":{"input_tokens":10,"output_tokens":20,"cache_read_input_tokens":0,"cache_creation_input_tokens":0}}}"#;
        let legacy = r#"{"type":"assistant","timestamp":"2026-05-13T09:00:00.000Z","message":{"id":"msg_l1","model":"claude-sonnet-4-6","usage":{"input_tokens":3,"output_tokens":4,"cache_read_input_tokens":0,"cache_creation_input_tokens":0}}}"#;
        let (_tmp, path) = write_jsonl(&[headless, interactive, legacy]);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 3);
        assert_eq!(entries[0].entrypoint.as_deref(), Some("sdk-cli"));
        assert_eq!(entries[1].entrypoint.as_deref(), Some("cli"));
        assert_eq!(entries[2].entrypoint, None);
    }

    #[test]
    fn parse_skips_non_assistant_entries() {
        let lines = vec![
            r#"{"type":"user","message":{"role":"user","content":"hi"}}"#.to_string(),
            r#"{"type":"permission-mode"}"#.to_string(),
            r#"{"type":"file-history-snapshot"}"#.to_string(),
            r#"{"type":"attachment","timestamp":"2026-04-14T08:27:26.236Z"}"#.to_string(),
            assistant_line("msg_01", "claude-sonnet-4-6", "2026-04-14T08:28:00.000Z", 100, 200, 0, 0),
        ];
        let line_refs: Vec<&str> = lines.iter().map(|s| s.as_str()).collect();
        let (_tmp, path) = write_jsonl(&line_refs);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].model, "claude-sonnet-4-6");
    }

    #[test]
    fn parse_tolerates_malformed_lines_and_empty_lines() {
        let good = assistant_line("msg_01", "claude-opus-4-6", "2026-04-14T08:27:38.895Z", 5, 10, 0, 0);
        let lines = vec![
            "",
            "not json at all",
            r#"{"type":"assistant","message":null}"#,
            r#"{"type":"assistant","message":{"id":null}}"#,
            &good,
            "   ",
            r#"{"broken":"trailing"#,
        ];
        let (_tmp, path) = write_jsonl(&lines);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
    }

    #[test]
    fn parse_extracts_session_id_from_file_stem() {
        let line = assistant_line("msg_01", "claude-opus-4-6", "2026-04-14T08:27:38.895Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].session_id, "abc123");
    }

    #[test]
    fn parse_parses_iso_timestamps() {
        let line = assistant_line("msg_01", "claude-opus-4-6", "2026-04-14T08:27:38.895Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        // 2026-04-14T08:27:38.895Z → 1776155258895 ms
        assert_eq!(entries[0].timestamp_ms, 1_776_155_258_895);
    }

    #[test]
    fn parse_skips_synthetic_model_rows() {
        // Claude Code / Agent SDK writes `model: "<synthetic>"` for internal
        // routing (skill listings, queue ops). These must not count as turns.
        let real = assistant_line("msg_real", "claude-opus-4-6", "2026-04-11T00:27:38.895Z", 15, 100, 0, 0);
        let synthetic = r#"{"type":"assistant","timestamp":"2026-04-11T00:27:39.000Z","message":{"id":"msg_synth","model":"<synthetic>","usage":{"input_tokens":0,"output_tokens":0,"cache_read_input_tokens":0,"cache_creation_input_tokens":0}}}"#;
        let (_tmp, path) = write_jsonl(&[&real, synthetic]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].model, "claude-opus-4-6");
    }

    #[test]
    fn parse_handles_missing_usage_fields() {
        let line = r#"{"type":"assistant","timestamp":"2026-04-14T08:27:38.895Z","message":{"id":"msg_01","model":"claude-opus-4-6"}}"#;
        let (_tmp, path) = write_jsonl(&[line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].input_tokens, 0);
        assert_eq!(entries[0].cache_read_tokens, 0);
    }

    #[test]
    fn reduce_computes_primary_model_by_mode() {
        let e = |m: &str| UsageEntry {
            session_id: "s".into(),
            timestamp_ms: 0,
            model: m.into(),
            input_tokens: 1,
            output_tokens: 1,
            cache_read_tokens: 0,
            cache_write_tokens: 0,
            tool_uses: Vec::new(),
            text_chars: 0,
            activity: Activity::General,
            entrypoint: None,
        };
        let entries = vec![
            e("claude-opus-4-6"),
            e("claude-opus-4-6"),
            e("claude-opus-4-6"),
            e("claude-sonnet-4-6"),
        ];
        let u = reduce_session(&entries, "", "sid").unwrap();
        assert_eq!(u.model, "claude-opus-4-6");
        assert_eq!(u.message_count, 4);
    }

    #[test]
    fn reduce_uses_launched_model_for_context_window() {
        let entries = vec![UsageEntry {
            session_id: "s".into(),
            timestamp_ms: 0,
            // JSONL strips the [1m] suffix
            model: "claude-opus-4-6".into(),
            input_tokens: 1000,
            output_tokens: 0,
            cache_read_tokens: 0,
            cache_write_tokens: 0,
            tool_uses: Vec::new(),
            text_chars: 0,
            activity: Activity::General,
            entrypoint: None,
        }];
        let with_1m = reduce_session(&entries, "claude-opus-4-6[1m]", "sid").unwrap();
        let without = reduce_session(&entries, "", "sid").unwrap();
        assert_eq!(with_1m.context_window_size, 1_000_000);
        assert_eq!(without.context_window_size, 200_000);
    }

    #[test]
    fn reduce_context_pct_uses_last_entry_only() {
        // Claude's context window reflects the LAST turn's input, not the sum.
        let mk = |input: u64, cr: u64, cw: u64| UsageEntry {
            session_id: "s".into(),
            timestamp_ms: 0,
            model: "claude-opus-4-6".into(),
            input_tokens: input,
            output_tokens: 0,
            cache_read_tokens: cr,
            cache_write_tokens: cw,
            tool_uses: Vec::new(),
            text_chars: 0,
            activity: Activity::General,
            entrypoint: None,
        };
        let entries = vec![
            mk(1000, 0, 0),
            mk(500, 0, 0),
            mk(5, 160_000, 0), // last: 160_005 / 200_000 = 80.0025%
        ];
        let u = reduce_session(&entries, "", "sid").unwrap();
        assert_eq!(u.current_context_tokens, 160_005);
        assert!((u.context_window_used_pct - 80.0025).abs() < 0.01);
    }

    #[test]
    fn reduce_cache_hit_pct() {
        let e = |input: u64, cr: u64, cw: u64| UsageEntry {
            session_id: "s".into(),
            timestamp_ms: 0,
            model: "claude-opus-4-6".into(),
            input_tokens: input,
            output_tokens: 1,
            cache_read_tokens: cr,
            cache_write_tokens: cw,
            tool_uses: Vec::new(),
            text_chars: 0,
            activity: Activity::General,
            entrypoint: None,
        };
        // 3 fresh + 97 cache_read + 0 cache_write → 97 / 100 = 97%
        let u = reduce_session(&[e(3, 97, 0)], "", "sid").unwrap();
        assert!((u.cache_hit_pct - 97.0).abs() < 0.01);

        // Cache writes count as NOT a hit: 10 fresh + 80 read + 10 write → 80%
        let u = reduce_session(&[e(10, 80, 10)], "", "sid").unwrap();
        assert!((u.cache_hit_pct - 80.0).abs() < 0.01);

        // zero activity → 0 (not NaN)
        let u = reduce_session(&[e(0, 0, 0)], "", "sid").unwrap();
        assert_eq!(u.cache_hit_pct, 0.0);
    }

    #[test]
    fn reduce_total_context_sent_is_cumulative() {
        let mk = |input: u64, cr: u64, cw: u64| UsageEntry {
            session_id: "s".into(),
            timestamp_ms: 0,
            model: "claude-opus-4-6".into(),
            input_tokens: input,
            output_tokens: 0,
            cache_read_tokens: cr,
            cache_write_tokens: cw,
            tool_uses: Vec::new(),
            text_chars: 0,
            activity: Activity::General,
            entrypoint: None,
        };
        let u = reduce_session(
            &[mk(15, 16_397, 25_661), mk(10, 42_058, 1_668)],
            "",
            "sid",
        )
        .unwrap();
        assert_eq!(u.total_context_sent, 15 + 16_397 + 25_661 + 10 + 42_058 + 1_668);
    }

    #[test]
    fn reduce_cost_matches_opus_pricing() {
        // Opus: $5/$25 input/output, $0.50 read, $6.25 write per 1M tokens
        let entries = vec![UsageEntry {
            session_id: "s".into(),
            timestamp_ms: 0,
            model: "claude-opus-4-6".into(),
            input_tokens: 1_000_000,   // $5
            output_tokens: 1_000_000,  // $25
            cache_read_tokens: 1_000_000, // $0.50
            cache_write_tokens: 1_000_000, // $6.25
            tool_uses: Vec::new(),
            text_chars: 0,
            activity: Activity::General,
            entrypoint: None,
        }];
        let u = reduce_session(&entries, "", "sid").unwrap();
        assert!((u.estimated_cost_usd - 36.75).abs() < 0.001);
    }

    #[test]
    fn reduce_empty_returns_none() {
        assert!(reduce_session(&[], "", "sid").is_none());
    }

    // ---------- content extraction ----------

    fn assistant_with_content(id: &str, ts: &str, content_json: &str) -> String {
        format!(
            r#"{{"type":"assistant","timestamp":"{ts}","message":{{"id":"{id}","model":"claude-opus-4-6","usage":{{"input_tokens":1,"output_tokens":1,"cache_read_input_tokens":0,"cache_creation_input_tokens":0}},"content":{content_json}}}}}"#
        )
    }

    #[test]
    fn parse_extracts_tool_use_names() {
        let content = r#"[{"type":"tool_use","id":"tu_1","name":"Edit","input":{"file_path":"a.rs"}},{"type":"tool_use","id":"tu_2","name":"Read","input":{"file_path":"b.rs"}}]"#;
        let line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].tool_uses, vec!["Edit".to_string(), "Read".to_string()]);
    }

    #[test]
    fn parse_captures_text_char_count() {
        let content = r#"[{"type":"text","text":"Hello world"},{"type":"text","text":"!"}]"#;
        let line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].text_chars, 12); // "Hello world" + "!"
    }

    #[test]
    fn parse_handles_missing_content_gracefully() {
        // Most existing JSONL rows have no `content` field at all (older
        // Claude Code versions, or when we parse a row before content arrives).
        let line = assistant_line("msg_1", "claude-opus-4-6", "2026-04-14T08:27:38.895Z", 5, 10, 0, 0);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert!(entries[0].tool_uses.is_empty());
        assert_eq!(entries[0].text_chars, 0);
        assert_eq!(entries[0].activity, Activity::General);
    }

    #[test]
    fn parse_classifies_edit_plus_debug_as_debugging() {
        // Real-world flow: user asks to fix a bug, assistant runs Edit.
        let user_line = r#"{"type":"user","timestamp":"2026-04-14T08:27:30.000Z","message":{"role":"user","content":"Please fix the bug in login.rs"}}"#.to_string();
        let content = r#"[{"type":"tool_use","id":"tu_1","name":"Edit","input":{"file_path":"login.rs"}}]"#;
        let asst_line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&user_line, &asst_line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].activity, Activity::Debugging);
    }

    #[test]
    fn parse_classifies_bash_pytest_as_testing() {
        let user_line = r#"{"type":"user","timestamp":"2026-04-14T08:27:30.000Z","message":{"role":"user","content":"run the tests"}}"#.to_string();
        let content = r#"[{"type":"tool_use","id":"tu_1","name":"Bash","input":{"command":"pytest tests/unit"}}]"#;
        let asst_line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&user_line, &asst_line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].activity, Activity::Testing);
    }

    #[test]
    fn parse_classifies_bash_git_as_gitops() {
        let content = r#"[{"type":"tool_use","id":"tu_1","name":"Bash","input":{"command":"git push origin main"}}]"#;
        let line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].activity, Activity::GitOps);
    }

    #[test]
    fn parse_tool_result_row_does_not_overwrite_user_text() {
        // Sequence: user says "refactor the store", assistant runs a tool, the
        // injected tool_result row is also type="user", then another assistant
        // follow-up. The follow-up should still attribute to the original
        // "refactor" intent — tool_result rows must NOT overwrite last_user_text.
        let user_real = r#"{"type":"user","timestamp":"2026-04-14T08:27:30.000Z","message":{"role":"user","content":"refactor the usage store"}}"#.to_string();
        let asst1 = assistant_with_content(
            "msg_1",
            "2026-04-14T08:27:35.000Z",
            r#"[{"type":"tool_use","id":"tu_1","name":"Read","input":{"file_path":"store.ts"}}]"#,
        );
        let tool_result = r#"{"type":"user","timestamp":"2026-04-14T08:27:36.000Z","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"tu_1","content":"file contents..."}]}}"#.to_string();
        let asst2 = assistant_with_content(
            "msg_2",
            "2026-04-14T08:27:40.000Z",
            r#"[{"type":"tool_use","id":"tu_2","name":"Edit","input":{"file_path":"store.ts"}}]"#,
        );
        let (_tmp, path) = write_jsonl(&[&user_real, &asst1, &tool_result, &asst2]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 2);
        // First assistant: Read only → Exploration
        assert_eq!(entries[0].activity, Activity::Exploration);
        // Second assistant: Edit after "refactor" intent → Refactor (not Coding),
        // proving the tool_result row didn't wipe last_user_text.
        assert_eq!(entries[1].activity, Activity::Refactor);
    }

    #[test]
    fn parse_classifies_agent_turn_as_delegation() {
        let content = r#"[{"type":"tool_use","id":"tu_1","name":"Agent","input":{"description":"explore"}}]"#;
        let line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].activity, Activity::Delegation);
    }

    #[test]
    fn parse_pure_text_turn_is_conversation() {
        let user_line = r#"{"type":"user","timestamp":"2026-04-14T08:27:30.000Z","message":{"role":"user","content":"thanks"}}"#.to_string();
        let content = r#"[{"type":"text","text":"You're welcome — anything else?"}]"#;
        let asst_line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&user_line, &asst_line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].activity, Activity::Conversation);
        assert!(entries[0].text_chars > 0);
    }

    #[test]
    fn parse_user_array_content_extracts_text_blocks() {
        // Some Claude Code versions write user content as an array even for
        // plain text. Ensure we pull the text out, not skip the row.
        let user_line = r#"{"type":"user","timestamp":"2026-04-14T08:27:30.000Z","message":{"role":"user","content":[{"type":"text","text":"fix the crash"}]}}"#.to_string();
        let content = r#"[{"type":"tool_use","id":"tu_1","name":"Edit","input":{}}]"#;
        let asst_line = assistant_with_content("msg_1", "2026-04-14T08:27:38.895Z", content);
        let (_tmp, path) = write_jsonl(&[&user_line, &asst_line]);
        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].activity, Activity::Debugging);
    }

    // ---------- split-row content merge (Claude Code one-block-per-row) ----------

    /// Helper: builds an assistant row that carries the same token usage
    /// numbers as the other rows for this `id` (matches what Claude Code
    /// actually writes — every row of a message repeats the final usage block).
    fn assistant_split_row(id: &str, ts: &str, content_json: &str) -> String {
        format!(
            r#"{{"type":"assistant","timestamp":"{ts}","message":{{"id":"{id}","model":"claude-opus-4-7","usage":{{"input_tokens":6,"output_tokens":746,"cache_read_input_tokens":21859,"cache_creation_input_tokens":33027}},"content":{content_json}}}}}"#
        )
    }

    #[test]
    fn parse_merges_tool_uses_across_split_content_rows() {
        // Claude Code 2.1.x emits one JSONL row per content block of an
        // assistant message: a `thinking` row, an optional `text` row, then
        // one row per `tool_use`. All rows share the same `message.id` and
        // the same `usage` block. The parser must merge tool_uses across rows
        // rather than dropping every row after the first.
        let thinking = assistant_split_row(
            "msg_01M6AvUVpAFjzSWcXeyGzUwV",
            "2026-05-13T09:09:03.500Z",
            r#"[{"type":"thinking","thinking":"plan"}]"#,
        );
        let text = assistant_split_row(
            "msg_01M6AvUVpAFjzSWcXeyGzUwV",
            "2026-05-13T09:09:03.550Z",
            r#"[{"type":"text","text":"Running tools in parallel."}]"#,
        );
        let tu1 = assistant_split_row(
            "msg_01M6AvUVpAFjzSWcXeyGzUwV",
            "2026-05-13T09:09:03.600Z",
            r#"[{"type":"tool_use","id":"tu_1","name":"mcp__example__list_configs","input":{"area":"rules"}}]"#,
        );
        let tu2 = assistant_split_row(
            "msg_01M6AvUVpAFjzSWcXeyGzUwV",
            "2026-05-13T09:09:03.700Z",
            r#"[{"type":"tool_use","id":"tu_2","name":"mcp__example__get_report","input":{}}]"#,
        );
        let tu3 = assistant_split_row(
            "msg_01M6AvUVpAFjzSWcXeyGzUwV",
            "2026-05-13T09:09:03.800Z",
            r#"[{"type":"tool_use","id":"tu_3","name":"Edit","input":{"file_path":"a.rs"}}]"#,
        );
        let (_tmp, path) = write_jsonl(&[&thinking, &text, &tu1, &tu2, &tu3]);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1, "still one entry per message.id");

        let e = &entries[0];
        assert_eq!(
            e.tool_uses,
            vec![
                "mcp__example__list_configs".to_string(),
                "mcp__example__get_report".to_string(),
                "Edit".to_string(),
            ],
            "all tool_use names merged in row order"
        );
        assert!(e.text_chars > 0, "text from the text row is preserved");

        // Billing fields come from the first row (they repeat identically, so
        // first == last in practice). Importantly, the merged content must
        // NOT cause tokens to multiply.
        assert_eq!(e.input_tokens, 6);
        assert_eq!(e.output_tokens, 746);
        assert_eq!(e.cache_read_tokens, 21859);
        assert_eq!(e.cache_write_tokens, 33027);
    }

    #[test]
    fn parse_split_row_classifies_against_merged_content() {
        // Bug pre-fix: first row is `thinking`, so `tool_uses` was empty and
        // activity classified as Conversation/General. After the fix the merged
        // content includes Edit, and a "fix the bug" user intent must route to
        // Debugging.
        let user = r#"{"type":"user","timestamp":"2026-05-13T09:09:00.000Z","message":{"role":"user","content":"please fix the bug in auth.rs"}}"#.to_string();
        let thinking = assistant_split_row(
            "msg_split",
            "2026-05-13T09:09:01.000Z",
            r#"[{"type":"thinking","thinking":"investigating"}]"#,
        );
        let edit = assistant_split_row(
            "msg_split",
            "2026-05-13T09:09:02.000Z",
            r#"[{"type":"tool_use","id":"tu_1","name":"Edit","input":{"file_path":"auth.rs"}}]"#,
        );
        let (_tmp, path) = write_jsonl(&[&user, &thinking, &edit]);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries.len(), 1);
        assert_eq!(entries[0].activity, Activity::Debugging);
        assert_eq!(entries[0].tool_uses, vec!["Edit".to_string()]);
    }

    #[test]
    fn parse_split_row_bash_commands_feed_classifier() {
        // bash_cmds accumulate across rows too — pytest in a separate row
        // from the assistant's text must still classify as Testing.
        let user = r#"{"type":"user","timestamp":"2026-05-13T09:09:00.000Z","message":{"role":"user","content":"run tests"}}"#.to_string();
        let text = assistant_split_row(
            "msg_bash",
            "2026-05-13T09:09:01.000Z",
            r#"[{"type":"text","text":"Running the unit tests."}]"#,
        );
        let bash = assistant_split_row(
            "msg_bash",
            "2026-05-13T09:09:02.000Z",
            r#"[{"type":"tool_use","id":"tu_1","name":"Bash","input":{"command":"pytest tests/"}}]"#,
        );
        let (_tmp, path) = write_jsonl(&[&user, &text, &bash]);

        let entries = parse_jsonl_entries(&path).unwrap().entries;
        assert_eq!(entries[0].activity, Activity::Testing);
    }

    // ---------- hook extraction ----------

    #[test]
    fn parse_extracts_hook_success_events() {
        let hook_line = r#"{"type":"attachment","timestamp":"2026-04-15T18:08:50.000Z","attachment":{"type":"hook_success","hookName":"PostToolUse:Edit","hookEvent":"PostToolUse","toolUseID":"tu_abc","command":"verify-edit.sh","exitCode":0,"durationMs":90}}"#;
        let hook_fail = r#"{"type":"attachment","timestamp":"2026-04-15T18:09:00.000Z","attachment":{"type":"hook_success","hookName":"PreToolUse:Bash","hookEvent":"PreToolUse","toolUseID":"tu_def","command":"lint-check.sh","exitCode":1,"durationMs":250}}"#;
        // hook_system_message should be ignored (no execution data)
        let hook_msg = r#"{"type":"attachment","timestamp":"2026-04-15T18:09:01.000Z","attachment":{"type":"hook_system_message","hookName":"Stop","hookEvent":"Stop","content":"reminder text"}}"#;
        let asst = assistant_line("msg_01", "claude-opus-4-6", "2026-04-15T18:08:51.000Z", 5, 10, 0, 0);
        let (_tmp, path) = write_jsonl(&[hook_line, hook_fail, hook_msg, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.entries.len(), 1, "assistant entries unaffected");
        assert_eq!(parsed.hooks.len(), 2, "two hook_success events");

        assert_eq!(parsed.hooks[0].hook_name, "PostToolUse:Edit");
        assert_eq!(parsed.hooks[0].hook_event, "PostToolUse");
        assert_eq!(parsed.hooks[0].command, "verify-edit.sh");
        assert_eq!(parsed.hooks[0].exit_code, 0);
        assert_eq!(parsed.hooks[0].duration_ms, 90);

        assert_eq!(parsed.hooks[0].tool_use_id.as_deref(), Some("tu_abc"));

        assert_eq!(parsed.hooks[1].hook_name, "PreToolUse:Bash");
        assert_eq!(parsed.hooks[1].exit_code, 1);
        assert_eq!(parsed.hooks[1].duration_ms, 250);
        assert_eq!(parsed.hooks[1].tool_use_id.as_deref(), Some("tu_def"));
    }

    #[test]
    fn parse_ignores_attachment_without_hook_data() {
        // Non-hook attachments (e.g. file-history-snapshot) should be skipped.
        let non_hook = r#"{"type":"attachment","timestamp":"2026-04-14T08:27:26.236Z","attachment":{"type":"file_content","path":"foo.rs"}}"#;
        let asst = assistant_line("msg_01", "claude-opus-4-6", "2026-04-14T08:28:00.000Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[non_hook, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.entries.len(), 1);
        assert!(parsed.hooks.is_empty());
    }

    #[test]
    fn snake_to_pascal_converts_correctly() {
        assert_eq!(snake_to_pascal("stop"), "Stop");
        assert_eq!(snake_to_pascal("pre_tool_use"), "PreToolUse");
        assert_eq!(snake_to_pascal("post_tool_use"), "PostToolUse");
        assert_eq!(snake_to_pascal("session_start"), "SessionStart");
    }

    #[test]
    fn parse_extracts_system_hook_summary_v2() {
        // v2.1.119+ format: type=system, subtype=*_hook_summary
        let stop_summary = r#"{"type":"system","subtype":"stop_hook_summary","hookCount":1,"hookInfos":[{"command":"Running build checks...","durationMs":5227}],"hookErrors":[],"preventedContinuation":false,"stopReason":"","hasOutput":false,"level":"suggestion","timestamp":"2026-04-25T09:06:12.854Z","uuid":"abc","toolUseID":"tu_xyz","sessionId":"sess-1","version":"2.1.119"}"#;
        let asst = assistant_line("msg_01", "claude-opus-4-6", "2026-04-25T09:06:13.000Z", 5, 10, 0, 0);
        let (_tmp, path) = write_jsonl(&[stop_summary, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.entries.len(), 1);
        assert_eq!(parsed.hooks.len(), 1);

        let h = &parsed.hooks[0];
        assert_eq!(h.hook_event, "Stop");
        assert_eq!(h.command, "Running build checks...");
        assert_eq!(h.duration_ms, 5227);
        assert_eq!(h.exit_code, 0, "no errors and not prevented = success");
        assert_eq!(h.tool_use_id.as_deref(), Some("tu_xyz"));
    }

    #[test]
    fn parse_system_hook_summary_with_errors() {
        let summary = r#"{"type":"system","subtype":"stop_hook_summary","hookCount":1,"hookInfos":[{"command":"cargo check","durationMs":3000}],"hookErrors":["something broke"],"preventedContinuation":true,"timestamp":"2026-04-25T10:00:00.000Z","uuid":"def","toolUseID":"tu_err","sessionId":"sess-2","version":"2.1.119"}"#;
        let asst = assistant_line("msg_02", "claude-opus-4-6", "2026-04-25T10:00:01.000Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[summary, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.hooks.len(), 1);
        assert_eq!(parsed.hooks[0].exit_code, 1, "errors or prevented = failure");
    }

    #[test]
    fn parse_system_hook_summary_multiple_hooks() {
        let summary = r#"{"type":"system","subtype":"stop_hook_summary","hookCount":2,"hookInfos":[{"command":"tsc --noEmit","durationMs":1200},{"command":"cargo check","durationMs":3500}],"hookErrors":[],"preventedContinuation":false,"timestamp":"2026-04-25T11:00:00.000Z","uuid":"ghi","toolUseID":"tu_multi","sessionId":"sess-3","version":"2.1.119"}"#;
        let asst = assistant_line("msg_03", "claude-opus-4-6", "2026-04-25T11:00:01.000Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[summary, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.hooks.len(), 2);
        assert_eq!(parsed.hooks[0].command, "tsc --noEmit");
        assert_eq!(parsed.hooks[0].duration_ms, 1200);
        assert_eq!(parsed.hooks[1].command, "cargo check");
        assert_eq!(parsed.hooks[1].duration_ms, 3500);
        assert_eq!(parsed.hooks[0].tool_use_id, parsed.hooks[1].tool_use_id);
    }

    #[test]
    fn hook_name_label_sanitizes_long_multiline_commands() {
        // /goal-style stop hook: command is a multi-line goal statement.
        // The synthesized hook_name must NOT include the full body — otherwise
        // every distinct goal creates a unique bucket and breaks the Usage page.
        let long_cmd = "Goal statement: Achieve >=3x better compression ratio than zstd -19 --long=27 and xz -9e on the Lichess 2024-01 corpus.\n\n3. Validation strategy:\n- (a) Fixed + held-out corpora.";
        let escaped = serde_json::to_string(long_cmd).unwrap();
        let summary = format!(
            r#"{{"type":"system","subtype":"stop_hook_summary","hookCount":1,"hookInfos":[{{"command":{escaped},"durationMs":42}}],"hookErrors":[],"preventedContinuation":false,"timestamp":"2026-04-25T12:00:00.000Z","uuid":"goal","toolUseID":"tu_goal","sessionId":"sess-goal","version":"2.1.119"}}"#
        );
        let asst = assistant_line("msg_goal", "claude-opus-4-6", "2026-04-25T12:00:01.000Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[&summary, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.hooks.len(), 1);
        let h = &parsed.hooks[0];

        assert_eq!(h.command, long_cmd, "full command preserved on the entry");
        assert!(h.hook_name.starts_with("Stop:"), "name still scoped by event");
        // Total label length is bounded: event prefix + ":" + max 64 chars + 1-char ellipsis.
        assert!(h.hook_name.chars().count() <= "Stop:".len() + 65);
        assert!(h.hook_name.ends_with('…'), "long commands get ellipsis: {}", h.hook_name);
        assert!(!h.hook_name.contains('\n'), "no newlines in bucket key");
    }

    #[test]
    fn hook_name_label_preserves_short_commands() {
        let summary = r#"{"type":"system","subtype":"stop_hook_summary","hookCount":1,"hookInfos":[{"command":"cargo check","durationMs":3000}],"hookErrors":[],"preventedContinuation":false,"timestamp":"2026-04-25T13:00:00.000Z","uuid":"short","toolUseID":"tu_short","sessionId":"sess-short","version":"2.1.119"}"#;
        let asst = assistant_line("msg_short", "claude-opus-4-6", "2026-04-25T13:00:01.000Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[summary, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.hooks[0].hook_name, "Stop:cargo check");
    }

    #[test]
    fn parse_handles_both_formats_in_same_file() {
        // Old v2.1.101 attachment format
        let old_hook = r#"{"type":"attachment","timestamp":"2026-04-23T14:05:05.937Z","attachment":{"type":"hook_success","hookName":"PostToolUse:Edit","hookEvent":"PostToolUse","toolUseID":"tu_old","command":"lint.sh","exitCode":0,"durationMs":90}}"#;
        // New v2.1.119 system format
        let new_hook = r#"{"type":"system","subtype":"stop_hook_summary","hookCount":1,"hookInfos":[{"command":"cargo check","durationMs":5000}],"hookErrors":[],"preventedContinuation":false,"timestamp":"2026-04-25T09:06:12.854Z","uuid":"both","toolUseID":"tu_new","sessionId":"sess-both","version":"2.1.119"}"#;
        let asst = assistant_line("msg_both", "claude-opus-4-6", "2026-04-25T09:07:00.000Z", 1, 1, 0, 0);
        let (_tmp, path) = write_jsonl(&[old_hook, new_hook, &asst]);

        let parsed = parse_jsonl_entries(&path).unwrap();
        assert_eq!(parsed.hooks.len(), 2, "both formats parsed");
        assert_eq!(parsed.hooks[0].hook_event, "PostToolUse");
        assert_eq!(parsed.hooks[1].hook_event, "Stop");
    }
}