cc_token_usage/analysis/

validate.rs

use anyhow::{Context, Result};
use std::collections::{HashMap, HashSet};
use std::fmt;
use std::fs::File;
use std::io::{BufRead, BufReader};
use std::path::Path;

use crate::data::models::{GlobalDataQuality, SessionData, SessionFile};
use crate::data::scanner::{resolve_agent_parents, scan_claude_home};
use crate::pricing::calculator::PricingCalculator;

// ─── Result Types ──────────────────────────────────────────────────────────

#[derive(Debug)]
pub struct ValidationReport {
    pub session_results: Vec<SessionValidation>,
    pub structure_checks: Vec<Check>,
    pub summary: ValidationSummary,
}

#[derive(Debug)]
pub struct SessionValidation {
    pub session_id: String,
    pub project: String,
    pub token_checks: Vec<Check>,
    pub agent_checks: Vec<Check>,
}

#[derive(Debug)]
pub struct Check {
    pub name: String,
    pub expected: String,
    pub actual: String,
    pub passed: bool,
}

impl Check {
    fn pass(name: impl Into<String>, value: impl fmt::Display) -> Self {
        let v = value.to_string();
        Self {
            name: name.into(),
            expected: v.clone(),
            actual: v,
            passed: true,
        }
    }

    fn compare(
        name: impl Into<String>,
        expected: impl fmt::Display,
        actual: impl fmt::Display,
    ) -> Self {
        let e = expected.to_string();
        let a = actual.to_string();
        let passed = e == a;
        Self {
            name: name.into(),
            expected: e,
            actual: a,
            passed,
        }
    }

    #[allow(dead_code)]
    fn compare_f64(name: impl Into<String>, expected: f64, actual: f64, tolerance: f64) -> Self {
        let passed = (expected - actual).abs() < tolerance;
        Self {
            name: name.into(),
            expected: format!("{:.2}", expected),
            actual: format!("{:.2}", actual),
            passed,
        }
    }
}

#[derive(Debug, Default)]
pub struct ValidationSummary {
    pub total_checks: usize,
    pub passed: usize,
    pub failed: usize,
    pub sessions_validated: usize,
    pub sessions_passed: usize,
}

// ─── Raw Token Counter (independent from main pipeline) ────────────────────

/// Token totals computed independently from the main parsing pipeline.
/// Uses serde_json::Value to ensure complete code path independence.
#[derive(Debug, Default)]
struct RawTokenCount {
    input_tokens: u64,
    output_tokens: u64,
    cache_creation_tokens: u64,
    cache_read_tokens: u64,
    turn_count: usize,
}

/// Check if a raw JSON entry passes the same validation as the main pipeline:
/// - type == "assistant"
/// - sidechain filter (if applicable)
/// - not synthetic, model exists
/// - usage exists with non-zero tokens
/// - valid timestamp <= now
fn is_valid_assistant(
    val: &serde_json::Value,
    skip_sidechain: bool,
    now: &chrono::DateTime<chrono::Utc>,
) -> bool {
    if val.get("type").and_then(|t| t.as_str()) != Some("assistant") {
        return false;
    }
    if skip_sidechain && val.get("isSidechain").and_then(|v| v.as_bool()) == Some(true) {
        return false;
    }
    let model = val.pointer("/message/model").and_then(|m| m.as_str());
    if model == Some("<synthetic>") || model.is_none() {
        return false;
    }
    // Check usage exists
    if val.pointer("/message/usage").is_none() {
        return false;
    }
    let input = val
        .pointer("/message/usage/input_tokens")
        .and_then(|v| v.as_u64())
        .unwrap_or(0);
    let output = val
        .pointer("/message/usage/output_tokens")
        .and_then(|v| v.as_u64())
        .unwrap_or(0);
    let cache_creation = val
        .pointer("/message/usage/cache_creation_input_tokens")
        .and_then(|v| v.as_u64())
        .unwrap_or(0);
    let cache_read = val
        .pointer("/message/usage/cache_read_input_tokens")
        .and_then(|v| v.as_u64())
        .unwrap_or(0);
    if input + output + cache_creation + cache_read == 0 {
        return false;
    }
    // Validate timestamp
    if let Some(ts_str) = val.get("timestamp").and_then(|t| t.as_str()) {
        if let Ok(ts) = ts_str.parse::<chrono::DateTime<chrono::Utc>>() {
            if ts > *now {
                return false;
            }
        } else {
            return false;
        }
    } else {
        return false;
    }
    true
}

/// Count tokens from a JSONL file using raw JSON parsing (no JournalEntry types).
/// Applies the same validation filters as the main pipeline for apples-to-apples comparison.
fn count_raw_tokens(path: &Path, skip_sidechain: bool) -> Result<RawTokenCount> {
    let file = File::open(path)
        .with_context(|| format!("raw counter: failed to open {}", path.display()))?;
    let reader = BufReader::new(file);
    let now = chrono::Utc::now();

    // requestId -> token counts (keep last for streaming dedup)
    let mut by_request: HashMap<String, (u64, u64, u64, u64)> = HashMap::new();
    let mut no_request_id_count = RawTokenCount::default();

    for line in reader.lines() {
        let line = line?;
        let val: serde_json::Value = match serde_json::from_str(&line) {
            Ok(v) => v,
            Err(_) => continue,
        };

        if !is_valid_assistant(&val, skip_sidechain, &now) {
            continue;
        }

        let input = val
            .pointer("/message/usage/input_tokens")
            .and_then(|v| v.as_u64())
            .unwrap_or(0);
        let output = val
            .pointer("/message/usage/output_tokens")
            .and_then(|v| v.as_u64())
            .unwrap_or(0);
        let cache_creation = val
            .pointer("/message/usage/cache_creation_input_tokens")
            .and_then(|v| v.as_u64())
            .unwrap_or(0);
        let cache_read = val
            .pointer("/message/usage/cache_read_input_tokens")
            .and_then(|v| v.as_u64())
            .unwrap_or(0);

        let request_id = val.get("requestId").and_then(|r| r.as_str());

        match request_id {
            Some(rid) if !rid.is_empty() => {
                by_request.insert(rid.to_string(), (input, output, cache_creation, cache_read));
            }
            _ => {
                no_request_id_count.input_tokens += input;
                no_request_id_count.output_tokens += output;
                no_request_id_count.cache_creation_tokens += cache_creation;
                no_request_id_count.cache_read_tokens += cache_read;
                no_request_id_count.turn_count += 1;
            }
        }
    }

    let mut result = no_request_id_count;
    for (input, output, cc, cr) in by_request.values() {
        result.input_tokens += input;
        result.output_tokens += output;
        result.cache_creation_tokens += cc;
        result.cache_read_tokens += cr;
        result.turn_count += 1;
    }

    Ok(result)
}

/// Count output tokens per requestId (for precise cross-file dedup token verification).
/// Returns (HashMap<requestId, output_tokens>, no_rid_output_total).
/// Streaming dedup: last entry per requestId wins.
fn count_tokens_by_request_id(
    path: &Path,
    skip_sidechain: bool,
) -> Result<(HashMap<String, u64>, u64)> {
    let file = File::open(path)?;
    let reader = BufReader::new(file);
    let now = chrono::Utc::now();
    let mut by_rid: HashMap<String, u64> = HashMap::new();
    let mut no_rid_output: u64 = 0;

    for line in reader.lines() {
        let line = line?;
        let val: serde_json::Value = match serde_json::from_str(&line) {
            Ok(v) => v,
            Err(_) => continue,
        };
        if !is_valid_assistant(&val, skip_sidechain, &now) {
            continue;
        }
        let output = val
            .pointer("/message/usage/output_tokens")
            .and_then(|v| v.as_u64())
            .unwrap_or(0);
        match val.get("requestId").and_then(|r| r.as_str()) {
            Some(rid) if !rid.is_empty() => {
                by_rid.insert(rid.to_string(), output);
            }
            _ => {
                no_rid_output += output;
            }
        }
    }
    Ok((by_rid, no_rid_output))
}

/// Collect requestIds from valid assistant entries in a JSONL file.
/// Applies the same validation filters as the pipeline for accurate cross-file dedup checking.
fn collect_valid_request_ids(path: &Path, skip_sidechain: bool) -> Result<HashSet<String>> {
    let file = File::open(path)?;
    let reader = BufReader::new(file);
    let now = chrono::Utc::now();
    let mut ids = HashSet::new();

    for line in reader.lines() {
        let line = line?;
        let val: serde_json::Value = match serde_json::from_str(&line) {
            Ok(v) => v,
            Err(_) => continue,
        };
        if !is_valid_assistant(&val, skip_sidechain, &now) {
            continue;
        }
        if let Some(rid) = val.get("requestId").and_then(|r| r.as_str()) {
            if !rid.is_empty() {
                ids.insert(rid.to_string());
            }
        }
    }
    Ok(ids)
}

// ─── Validation Engine ─────────────────────────────────────────────────────

/// Run full validation across all sessions.
pub fn validate_all(
    sessions: &[&SessionData],
    quality: &GlobalDataQuality,
    claude_home: &Path,
    calc: &PricingCalculator,
) -> Result<ValidationReport> {
    // Re-scan files independently for structure validation
    let mut files = scan_claude_home(claude_home)?;
    resolve_agent_parents(&mut files)?;

    let (main_files, agent_files): (Vec<&SessionFile>, Vec<&SessionFile>) =
        files.iter().partition(|f| !f.is_agent);

    let mut structure_checks = Vec::new();
    let mut session_results = Vec::new();

    // ── Structure Checks ────────────────────────────────────────────────

    // Check 1: Session count = main file count
    structure_checks.push(Check::compare(
        "session_count == main_file_count",
        main_files.len(),
        quality.total_session_files,
    ));

    // Check 2: Agent file count
    structure_checks.push(Check::compare(
        "agent_file_count",
        agent_files.len(),
        quality.total_agent_files,
    ));

    // Check 3: Orphan agent count (agents without a matching main session file)
    let main_session_ids: HashSet<&str> =
        main_files.iter().map(|f| f.session_id.as_str()).collect();
    let orphan_count = agent_files
        .iter()
        .filter(|f| {
            let parent = f.parent_session_id.as_deref().unwrap_or(&f.session_id);
            !main_session_ids.contains(parent)
        })
        .count();
    structure_checks.push(Check::pass(
        format!("orphan_agents (no main session file): {}", orphan_count),
        orphan_count,
    ));

    // Check 4: Report duplicate session IDs in main files (pipeline deduplicates by HashMap)
    let unique_main_ids: HashSet<&str> = main_files.iter().map(|f| f.session_id.as_str()).collect();
    let dup_count = main_files.len() - unique_main_ids.len();
    structure_checks.push(Check::pass(
        format!(
            "main_session_files: {} files, {} unique IDs ({} duplicates)",
            main_files.len(),
            unique_main_ids.len(),
            dup_count
        ),
        main_files.len(),
    ));

    // Check 5: Cross-file dedup — agent turns in main session should not be double-counted
    let mut cross_file_overlap = 0usize;
    for agent in &agent_files {
        let parent_id = agent
            .parent_session_id
            .as_deref()
            .unwrap_or(&agent.session_id);
        let parent_file = main_files.iter().find(|f| f.session_id == parent_id);
        if let Some(pf) = parent_file {
            let parent_rids = collect_valid_request_ids(&pf.path, true).unwrap_or_default();
            let agent_rids = collect_valid_request_ids(&agent.path, false).unwrap_or_default();
            cross_file_overlap += parent_rids.intersection(&agent_rids).count();
        }
    }
    structure_checks.push(Check::pass(
        format!(
            "cross_file_overlapping_request_ids (deduped: {})",
            cross_file_overlap
        ),
        cross_file_overlap,
    ));

    // ── Per-Session Validation ──────────────────────────────────────────

    // Build lookup: session_id -> Vec<agent SessionFile>
    let mut agents_by_parent: HashMap<&str, Vec<&SessionFile>> = HashMap::new();
    for af in &agent_files {
        let parent_id = af.parent_session_id.as_deref().unwrap_or(&af.session_id);
        agents_by_parent.entry(parent_id).or_default().push(af);
    }

    // Build lookup: session_id -> main SessionFile
    let main_file_map: HashMap<&str, &SessionFile> = main_files
        .iter()
        .map(|f| (f.session_id.as_str(), *f))
        .collect();

    for session in sessions {
        let mut token_checks = Vec::new();
        let mut agent_checks = Vec::new();

        // --- Token validation: raw counter vs pipeline ---
        if let Some(mf) = main_file_map.get(session.session_id.as_str()) {
            let raw_main = count_raw_tokens(&mf.path, true).unwrap_or_default();

            // Pipeline's main turns
            let pipeline_main_input: u64 = session
                .turns
                .iter()
                .map(|t| t.usage.input_tokens.unwrap_or(0))
                .sum();
            let pipeline_main_output: u64 = session
                .turns
                .iter()
                .map(|t| t.usage.output_tokens.unwrap_or(0))
                .sum();
            let pipeline_main_cache_creation: u64 = session
                .turns
                .iter()
                .map(|t| t.usage.cache_creation_input_tokens.unwrap_or(0))
                .sum();
            let pipeline_main_cache_read: u64 = session
                .turns
                .iter()
                .map(|t| t.usage.cache_read_input_tokens.unwrap_or(0))
                .sum();
            let pipeline_main_turns = session.turns.len();

            token_checks.push(Check::compare(
                "main_turn_count",
                raw_main.turn_count,
                pipeline_main_turns,
            ));
            token_checks.push(Check::compare(
                "main_input_tokens",
                raw_main.input_tokens,
                pipeline_main_input,
            ));
            token_checks.push(Check::compare(
                "main_output_tokens",
                raw_main.output_tokens,
                pipeline_main_output,
            ));
            token_checks.push(Check::compare(
                "main_cache_creation_tokens",
                raw_main.cache_creation_tokens,
                pipeline_main_cache_creation,
            ));
            token_checks.push(Check::compare(
                "main_cache_read_tokens",
                raw_main.cache_read_tokens,
                pipeline_main_cache_read,
            ));
        }

        // --- Agent validation ---
        let agent_session_files = agents_by_parent.get(session.session_id.as_str());
        let expected_agent_files = agent_session_files.map_or(0, |v| v.len());
        let actual_agent_file_count = if expected_agent_files > 0 {
            expected_agent_files
        } else {
            0
        };

        agent_checks.push(Check::compare(
            "agent_file_count (from scanner)",
            actual_agent_file_count,
            expected_agent_files,
        ));

        // Verify agent turn association (if agent files exist)
        if expected_agent_files > 0 {
            if let Some(afs) = agent_session_files {
                // Get main session's valid requestIds for cross-file dedup
                let main_file = main_file_map.get(session.session_id.as_str());
                let main_rids = main_file
                    .map(|mf| collect_valid_request_ids(&mf.path, true).unwrap_or_default())
                    .unwrap_or_default();

                // Calculate expected per-file (matching pipeline's per-file merge logic)
                let mut expected_unique_agent_turns = 0usize;
                let mut raw_agent_output: u64 = 0;

                for af in afs {
                    let raw = count_raw_tokens(&af.path, false).unwrap_or_default();
                    let file_rids = collect_valid_request_ids(&af.path, false).unwrap_or_default();
                    let file_overlap = file_rids.intersection(&main_rids).count();
                    let unique_turns = raw.turn_count.saturating_sub(file_overlap);
                    expected_unique_agent_turns += unique_turns;

                    // Precise: count output tokens only for non-overlapping requestIds
                    let (per_rid, no_rid_output) =
                        count_tokens_by_request_id(&af.path, false).unwrap_or_default();
                    for (rid, output) in &per_rid {
                        if !main_rids.contains(rid) {
                            raw_agent_output += output;
                        }
                    }
                    raw_agent_output += no_rid_output;
                }

                // Turn count check
                let pipeline_subagent_turn_count = session.agent_turn_count();
                agent_checks.push(Check::compare(
                    "agent_turn_count (after cross-file dedup)",
                    expected_unique_agent_turns,
                    pipeline_subagent_turn_count,
                ));

                // If expected > 0 but pipeline has 0, that's a real issue
                if expected_unique_agent_turns > 0 {
                    agent_checks.push(Check::compare(
                        "has_agent_turns (non-overlapping exist)",
                        "true",
                        (pipeline_subagent_turn_count > 0).to_string(),
                    ));
                }

                // Agent output token verification (pipeline vs raw, ±5% tolerance for ratio estimate)
                let pipeline_agent_output: u64 = session
                    .subagents
                    .iter()
                    .flat_map(|s| s.turns.iter())
                    .map(|t| t.usage.output_tokens.unwrap_or(0))
                    .sum();

                let agent_output_match = {
                    if raw_agent_output == 0 && pipeline_agent_output == 0 {
                        true
                    } else {
                        let max_val = raw_agent_output.max(pipeline_agent_output) as f64;
                        if max_val == 0.0 {
                            true
                        } else {
                            (raw_agent_output as f64 - pipeline_agent_output as f64).abs() / max_val
                                < 0.05
                        }
                    }
                };

                agent_checks.push(Check {
                    name: "agent_output_tokens (±5%)".into(),
                    expected: raw_agent_output.to_string(),
                    actual: pipeline_agent_output.to_string(),
                    passed: agent_output_match,
                });

                // Verify all subagent turns have is_agent=true
                let all_marked_agent = session
                    .subagents
                    .iter()
                    .flat_map(|s| s.turns.iter())
                    .all(|t| t.is_agent);
                agent_checks.push(Check::compare(
                    "all agent_turns have is_agent=true",
                    "true",
                    all_marked_agent.to_string(),
                ));
            }
        }

        // --- Workflow validation (Claude Code 2.1.159+) ---
        // For every discovered workflow run, reconcile what the pipeline parsed
        // from the run's agent transcripts (the subagents whose workflow_run_id
        // matches) against the run snapshot.
        //
        // IMPORTANT: these are *supplementary reconciliation* checks against
        // Claude Code's own self-reported snapshot — they are NOT an independent
        // token re-count and do not by themselves prove "no loss". The genuine
        // independent no-loss guarantee comes from the pre-existing per-file
        // `agent_output` (±5%) and cross-file-dedup checks above, which re-count
        // every agent file from raw JSON via the independent `count_raw_tokens`
        // path — and those now automatically cover these workflow agent files,
        // because the scanner records them as ordinary `is_agent` files that the
        // independent re-scan picks up.
        //
        // Two checks flagged on mismatch:
        //   1. parsed tokens > 0 — the run's agents were parsed and fed into the
        //      session/overview totals (catches a wholesale drop of the run).
        //   2. parsed agent count == snapshot `agentCount` — every agent the run
        //      declared was found and parsed on disk.
        //
        // One informational check (always "passes", surfaces both numbers): the
        // snapshot's `totalTokens` figure. Empirically the snapshot's token
        // accounting does NOT equal any clean parsed sum (on observed 2.1.159
        // data it tracks the cache-write/`cache_creation` tokens, not
        // input+output+cache_read), so asserting equality would always fail. We
        // report both so a human can eyeball drift without flagging a false
        // failure.
        let workflow_runs =
            cc_session_jsonl::scanner::scan_session_workflows(&session.session_id, claude_home)
                .unwrap_or_default();
        for run in &workflow_runs {
            // Parsed totals for this run, summed across matching subagents.
            let mut parsed_input: u64 = 0;
            let mut parsed_output: u64 = 0;
            let mut parsed_cache_creation: u64 = 0;
            let mut parsed_cache_read: u64 = 0;
            let mut parsed_agent_count = 0usize;
            for sa in &session.subagents {
                if sa.workflow_run_id.as_deref() != Some(run.run_id.as_str()) {
                    continue;
                }
                parsed_agent_count += 1;
                for t in &sa.turns {
                    parsed_input += t.usage.input_tokens.unwrap_or(0);
                    parsed_output += t.usage.output_tokens.unwrap_or(0);
                    parsed_cache_creation += t.usage.cache_creation_input_tokens.unwrap_or(0);
                    parsed_cache_read += t.usage.cache_read_input_tokens.unwrap_or(0);
                }
            }
            let parsed_total_tokens =
                parsed_input + parsed_output + parsed_cache_creation + parsed_cache_read;

            // Hard check 1: the run's agents contribute tokens to the totals.
            agent_checks.push(Check::compare(
                format!("workflow[{}] parsed tokens > 0 (no loss)", run.run_id),
                "true",
                (parsed_total_tokens > 0).to_string(),
            ));

            // Hard check 2: every declared agent was found and parsed.
            if let Some(snap_agents) = run.snapshot.as_ref().and_then(|s| s.agent_count) {
                agent_checks.push(Check::compare(
                    format!("workflow[{}] agent_count == snapshot", run.run_id),
                    snap_agents,
                    parsed_agent_count as u64,
                ));
            }

            // Informational: snapshot totalTokens vs parsed total (never fails).
            if let Some(snap_tokens) = run.snapshot.as_ref().and_then(|s| s.total_tokens) {
                agent_checks.push(Check::pass(
                    format!(
                        "workflow[{}] snapshot.totalTokens={} | parsed total={} (in={} out={} cw={} cr={})",
                        run.run_id,
                        snap_tokens,
                        parsed_total_tokens,
                        parsed_input,
                        parsed_output,
                        parsed_cache_creation,
                        parsed_cache_read,
                    ),
                    parsed_total_tokens,
                ));
            }
        }

        // --- Total (main + agent) verification ---
        let pipeline_total_output: u64 = session
            .turns
            .iter()
            .chain(session.subagents.iter().flat_map(|s| s.turns.iter()))
            .map(|t| t.usage.output_tokens.unwrap_or(0))
            .sum();
        let pipeline_total_turns = session.total_turn_count();

        // Verify total turn count matches all_responses()
        token_checks.push(Check::compare(
            "total_turn_count == turns + agent_turns",
            pipeline_total_turns,
            session.all_responses().len(),
        ));

        // Verify non-zero totals for sessions that have data
        if pipeline_total_turns > 0 {
            token_checks.push(Check::compare(
                "total_output_tokens > 0",
                "true",
                (pipeline_total_output > 0).to_string(),
            ));
        }

        // --- Cost validation ---
        let pipeline_cost: f64 = session
            .turns
            .iter()
            .chain(session.subagents.iter().flat_map(|s| s.turns.iter()))
            .map(|t| calc.calculate_turn_cost(&t.model, &t.usage).total)
            .sum();

        // Verify cost is non-negative and consistent with tokens
        let has_tokens = session
            .turns
            .iter()
            .chain(session.subagents.iter().flat_map(|s| s.turns.iter()))
            .any(|t| {
                t.usage.input_tokens.unwrap_or(0) > 0 || t.usage.output_tokens.unwrap_or(0) > 0
            });
        if has_tokens {
            token_checks.push(Check::compare(
                "cost > 0 when tokens exist",
                "true",
                (pipeline_cost > 0.0).to_string(),
            ));
        }

        // --- Project association ---
        if let Some(mf) = main_file_map.get(session.session_id.as_str()) {
            token_checks.push(Check::compare(
                "project_association",
                mf.project.as_deref().unwrap_or("(none)"),
                session.project.as_deref().unwrap_or("(none)"),
            ));
        }

        let project_name = session
            .project
            .as_deref()
            .unwrap_or("(unknown)")
            .to_string();

        session_results.push(SessionValidation {
            session_id: session.session_id.clone(),
            project: project_name,
            token_checks,
            agent_checks,
        });
    }

    // ── Compute Summary ─────────────────────────────────────────────────

    let mut summary = ValidationSummary::default();

    for check in &structure_checks {
        summary.total_checks += 1;
        if check.passed {
            summary.passed += 1;
        } else {
            summary.failed += 1;
        }
    }

    for sv in &session_results {
        summary.sessions_validated += 1;
        let mut session_pass = true;
        for check in sv.token_checks.iter().chain(sv.agent_checks.iter()) {
            summary.total_checks += 1;
            if check.passed {
                summary.passed += 1;
            } else {
                summary.failed += 1;
                session_pass = false;
            }
        }
        if session_pass {
            summary.sessions_passed += 1;
        }
    }

    Ok(ValidationReport {
        session_results,
        structure_checks,
        summary,
    })
}

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

    fn make_assistant_line(request_id: &str, input: u64, output: u64) -> String {
        format!(
            r#"{{"type":"assistant","uuid":"u-{}","timestamp":"2026-03-16T10:00:00Z","message":{{"model":"claude-opus-4-6","role":"assistant","stop_reason":"end_turn","usage":{{"input_tokens":{},"output_tokens":{},"cache_creation_input_tokens":0,"cache_read_input_tokens":0}},"content":[{{"type":"text","text":"hi"}}]}},"sessionId":"s1","cwd":"/tmp","gitBranch":"","userType":"external","isSidechain":false,"parentUuid":null,"requestId":"{}"}}"#,
            request_id, input, output, request_id
        )
    }

    #[test]
    fn raw_counter_basic() {
        let mut f = NamedTempFile::new().unwrap();
        writeln!(f, "{}", make_assistant_line("r1", 100, 50)).unwrap();
        writeln!(f, "{}", make_assistant_line("r2", 200, 75)).unwrap();
        f.flush().unwrap();

        let result = count_raw_tokens(f.path(), true).unwrap();
        assert_eq!(result.turn_count, 2);
        assert_eq!(result.input_tokens, 300);
        assert_eq!(result.output_tokens, 125);
    }

    #[test]
    fn raw_counter_deduplicates_streaming() {
        let mut f = NamedTempFile::new().unwrap();
        // Same requestId, different values — last one wins
        writeln!(f, "{}", make_assistant_line("r1", 100, 50)).unwrap();
        writeln!(f, "{}", make_assistant_line("r1", 200, 75)).unwrap();
        f.flush().unwrap();

        let result = count_raw_tokens(f.path(), true).unwrap();
        assert_eq!(result.turn_count, 1);
        assert_eq!(result.input_tokens, 200);
        assert_eq!(result.output_tokens, 75);
    }

    #[test]
    fn raw_counter_skips_synthetic() {
        let mut f = NamedTempFile::new().unwrap();
        writeln!(f, r#"{{"type":"assistant","uuid":"u1","timestamp":"2026-03-16T10:00:00Z","message":{{"model":"<synthetic>","role":"assistant","stop_reason":"end_turn","usage":{{"input_tokens":100,"output_tokens":50,"cache_creation_input_tokens":0,"cache_read_input_tokens":0}},"content":[]}},"sessionId":"s1","cwd":"/tmp","gitBranch":"","userType":"external","isSidechain":false,"parentUuid":null,"requestId":"r1"}}"#).unwrap();
        writeln!(f, "{}", make_assistant_line("r2", 200, 75)).unwrap();
        f.flush().unwrap();

        let result = count_raw_tokens(f.path(), true).unwrap();
        assert_eq!(result.turn_count, 1);
        assert_eq!(result.input_tokens, 200);
    }

    #[test]
    fn raw_counter_respects_sidechain_flag() {
        let sidechain_line = r#"{"type":"assistant","uuid":"u1","timestamp":"2026-03-16T10:00:00Z","message":{"model":"claude-opus-4-6","role":"assistant","stop_reason":"end_turn","usage":{"input_tokens":100,"output_tokens":50,"cache_creation_input_tokens":0,"cache_read_input_tokens":0},"content":[]},"sessionId":"s1","cwd":"/tmp","gitBranch":"","userType":"external","isSidechain":true,"parentUuid":null,"requestId":"r1"}"#;
        let mut f = NamedTempFile::new().unwrap();
        writeln!(f, "{}", sidechain_line).unwrap();
        f.flush().unwrap();

        // Main session: skip sidechain
        let result = count_raw_tokens(f.path(), true).unwrap();
        assert_eq!(result.turn_count, 0);

        // Agent file: keep sidechain
        let result = count_raw_tokens(f.path(), false).unwrap();
        assert_eq!(result.turn_count, 1);
        assert_eq!(result.input_tokens, 100);
    }

    #[test]
    fn raw_counter_skips_non_assistant() {
        let mut f = NamedTempFile::new().unwrap();
        writeln!(f, r#"{{"type":"user","uuid":"u1","message":{{"role":"user","content":"hi"}},"timestamp":"2026-03-16T10:00:00Z","sessionId":"s1"}}"#).unwrap();
        writeln!(f, r#"{{"type":"progress","data":{{"type":"hook"}},"uuid":"u2","timestamp":"2026-03-16T10:00:00Z","sessionId":"s1"}}"#).unwrap();
        writeln!(f, "{}", make_assistant_line("r1", 100, 50)).unwrap();
        f.flush().unwrap();

        let result = count_raw_tokens(f.path(), true).unwrap();
        assert_eq!(result.turn_count, 1);
    }
}