cc_token_usage/analysis/

session.rs

use std::collections::HashMap;

use crate::data::models::SessionData;
use crate::pricing::calculator::PricingCalculator;

use super::{
    AgentDetail, AgentSummary, AggregatedTokens, SessionResult, SubagentSummary, TurnCostBreakdown,
    TurnDetail,
};

/// Agent metadata loaded from .meta.json files.
#[derive(Debug, Clone, Default)]
pub struct AgentMeta {
    pub agent_type: String,
    pub description: String,
}

pub fn analyze_session(
    session: &SessionData,
    calc: &PricingCalculator,
    agent_meta: &std::collections::HashMap<String, AgentMeta>,
) -> SessionResult {
    let all_turns = session.all_responses();

    let mut turn_details = Vec::new();
    let mut total_tokens = AggregatedTokens::default();
    let mut total_cost = 0.0;
    let mut stop_reason_counts: HashMap<String, usize> = HashMap::new();
    let mut agent_summary = AgentSummary::default();
    let mut model_counts: HashMap<&str, usize> = HashMap::new();
    let mut max_context: u64 = 0;
    let mut prev_context_size: Option<u64> = None;
    let mut agent_acc: HashMap<String, (usize, u64, f64)> = HashMap::new();

    // Phase 1: new accumulators
    let mut tool_error_count: usize = 0;
    let mut truncated_count: usize = 0;
    let mut service_tiers: HashMap<String, usize> = HashMap::new();
    let mut speeds: HashMap<String, usize> = HashMap::new();
    let mut inference_geos: HashMap<String, usize> = HashMap::new();
    let mut git_branches: HashMap<String, usize> = HashMap::new();

    for (i, turn) in all_turns.iter().enumerate() {
        let input = turn.usage.input_tokens.unwrap_or(0);
        let output = turn.usage.output_tokens.unwrap_or(0);
        let cache_create = turn.usage.cache_creation_input_tokens.unwrap_or(0);
        let cache_read = turn.usage.cache_read_input_tokens.unwrap_or(0);

        // Extract 5m/1h TTL breakdown
        let (cache_write_5m, cache_write_1h) = if let Some(ref detail) = turn.usage.cache_creation {
            (
                detail.ephemeral_5m_input_tokens.unwrap_or(0),
                detail.ephemeral_1h_input_tokens.unwrap_or(0),
            )
        } else {
            (0, 0)
        };

        let context_size = input + cache_create + cache_read;
        let cache_hit_rate = if context_size > 0 {
            (cache_read as f64 / context_size as f64) * 100.0
        } else {
            0.0
        };

        // Track max context
        if context_size > max_context {
            max_context = context_size;
        }

        // Compaction detection and context delta
        let is_compaction = match prev_context_size {
            Some(prev) => prev > 0 && (context_size as f64) < (prev as f64 * 0.9),
            None => false,
        };
        let context_delta = match prev_context_size {
            Some(prev) => context_size as i64 - prev as i64,
            None => 0,
        };
        prev_context_size = Some(context_size);

        let pricing_cost = calc.calculate_turn_cost(&turn.model, &turn.usage);

        let cost_breakdown = TurnCostBreakdown {
            input_cost: pricing_cost.input_cost,
            output_cost: pricing_cost.output_cost,
            cache_write_5m_cost: pricing_cost.cache_write_5m_cost,
            cache_write_1h_cost: pricing_cost.cache_write_1h_cost,
            cache_read_cost: pricing_cost.cache_read_cost,
            total: pricing_cost.total,
        };

        // Track stop reasons
        if let Some(ref reason) = turn.stop_reason {
            *stop_reason_counts.entry(reason.clone()).or_insert(0) += 1;
            if reason == "max_tokens" {
                truncated_count += 1;
            }
        }

        // Phase 1: accumulate new metrics
        tool_error_count += turn.tool_error_count;
        if let Some(ref tier) = turn.service_tier {
            if !tier.is_empty() {
                *service_tiers.entry(tier.clone()).or_insert(0) += 1;
            }
        }
        if let Some(ref spd) = turn.speed {
            if !spd.is_empty() {
                *speeds.entry(spd.clone()).or_insert(0) += 1;
            }
        }
        if let Some(ref geo) = turn.inference_geo {
            if !geo.is_empty() {
                *inference_geos.entry(geo.clone()).or_insert(0) += 1;
            }
        }
        if let Some(ref branch) = turn.git_branch {
            if !branch.is_empty() {
                *git_branches.entry(branch.clone()).or_insert(0) += 1;
            }
        }

        // Aggregate tokens
        total_tokens.add_usage(&turn.usage);
        total_cost += pricing_cost.total;

        // Model frequency
        *model_counts.entry(&turn.model).or_insert(0) += 1;

        // Agent summary
        let is_agent = turn.is_agent;
        if is_agent {
            agent_summary.total_agent_turns += 1;
            agent_summary.agent_output_tokens += output;
            agent_summary.agent_cost += pricing_cost.total;

            // Per-agent accumulation
            let aid = turn.agent_id.clone().unwrap_or_default();
            if !aid.is_empty() {
                let entry = agent_acc.entry(aid).or_insert((0usize, 0u64, 0.0f64));
                entry.0 += 1;
                entry.1 += output;
                entry.2 += pricing_cost.total;
            }
        }

        turn_details.push(TurnDetail {
            turn_number: i + 1,
            timestamp: turn.timestamp,
            model: turn.model.clone(),
            input_tokens: input,
            output_tokens: output,
            cache_write_5m_tokens: cache_write_5m,
            cache_write_1h_tokens: cache_write_1h,
            cache_read_tokens: cache_read,
            context_size,
            cache_hit_rate,
            cost: pricing_cost.total,
            cost_breakdown,
            stop_reason: turn.stop_reason.clone(),
            is_agent,
            is_compaction,
            context_delta,
            user_text: turn.user_text.clone(),
            assistant_text: turn.assistant_text.clone(),
            tool_names: turn.tool_names.clone(),
        });
    }

    // Duration
    let duration_minutes = match (session.first_timestamp, session.last_timestamp) {
        (Some(first), Some(last)) => (last - first).num_seconds() as f64 / 60.0,
        _ => 0.0,
    };

    // Compaction count
    let compaction_count = turn_details.iter().filter(|t| t.is_compaction).count();

    // Cache write percentages from total tokens
    let total_5m = total_tokens.cache_write_5m_tokens;
    let total_1h = total_tokens.cache_write_1h_tokens;
    let total_cache_write = total_5m + total_1h;
    let cache_write_5m_pct = if total_cache_write > 0 {
        (total_5m as f64 / total_cache_write as f64) * 100.0
    } else {
        0.0
    };
    let cache_write_1h_pct = if total_cache_write > 0 {
        (total_1h as f64 / total_cache_write as f64) * 100.0
    } else {
        0.0
    };

    // Per-agent details
    let mut agents: Vec<AgentDetail> = agent_acc
        .into_iter()
        .map(|(aid, (turns, output, cost))| {
            let meta = agent_meta.get(&aid);
            AgentDetail {
                agent_id: aid,
                agent_type: meta.map_or_else(|| "unknown".into(), |m| m.agent_type.clone()),
                description: meta.map_or_else(|| "".into(), |m| m.description.clone()),
                turns,
                output_tokens: output,
                cost,
            }
        })
        .collect();
    agents.sort_by(|a, b| {
        b.cost
            .partial_cmp(&a.cost)
            .unwrap_or(std::cmp::Ordering::Equal)
    });
    agent_summary.agents = agents;

    // Primary model
    let model = model_counts
        .into_iter()
        .max_by_key(|(_, count)| *count)
        .map(|(m, _)| m.to_string())
        .unwrap_or_default();

    // Autonomy ratio
    let total_turn_count = session.total_turn_count();
    let user_prompt_count = session.metadata.user_prompt_count;
    let autonomy_ratio = if user_prompt_count > 0 {
        total_turn_count as f64 / user_prompt_count as f64
    } else {
        0.0
    };

    SessionResult {
        session_id: session.session_id.clone(),
        project: session
            .project
            .clone()
            .unwrap_or_else(|| "(unknown)".to_string()),
        turn_details,
        agent_summary,
        total_tokens,
        total_cost,
        stop_reason_counts,
        duration_minutes,
        max_context,
        compaction_count,
        cache_write_5m_pct,
        cache_write_1h_pct,
        model,
        // Phase 1: metadata
        title: session.metadata.title.clone(),
        tags: session.metadata.tags.clone(),
        mode: session.metadata.mode.clone(),
        pr_links: session.metadata.pr_links.clone(),
        // Autonomy
        user_prompt_count,
        autonomy_ratio,
        // Errors
        api_error_count: session.metadata.api_error_count,
        tool_error_count,
        truncated_count,
        // Speculation
        speculation_accepts: session.metadata.speculation_accepts,
        speculation_time_saved_ms: session.metadata.speculation_time_saved_ms,
        // Service info
        service_tiers,
        speeds,
        inference_geos,
        // Git
        git_branches,
        // Context Collapse
        collapse_count: session.metadata.collapse_commits.len(),
        collapse_summaries: session
            .metadata
            .collapse_commits
            .iter()
            .map(|c| c.summary.clone())
            .collect(),
        collapse_avg_risk: session
            .metadata
            .collapse_snapshot
            .as_ref()
            .map_or(0.0, |s| s.avg_risk),
        collapse_max_risk: session
            .metadata
            .collapse_snapshot
            .as_ref()
            .map_or(0.0, |s| s.max_risk),
        // Attribution
        attribution: session.metadata.attribution.clone(),
        // ── Phase 2: per-session capability inventory ──
        subagents: build_subagent_summaries(session, calc),
        plugins: session.plugins.clone(),
        skills: session.skills.clone(),
        hooks: session.hooks.clone(),
        subagent_types: session.subagent_type_aggregates(calc),
        is_orphan: session.is_orphan,
    }
}

/// Build one `SubagentSummary` per `Subagent`, costed via the pricing
/// calculator. Sort order: by cost descending (most expensive first), matching
/// the existing `agent_summary.agents` ordering convention.
fn build_subagent_summaries(
    session: &SessionData,
    calc: &PricingCalculator,
) -> Vec<SubagentSummary> {
    let mut out: Vec<SubagentSummary> = session
        .subagents
        .iter()
        .map(|sa| {
            let mut output_tokens: u64 = 0;
            let mut cost = 0.0f64;
            for t in &sa.turns {
                output_tokens += t.usage.output_tokens.unwrap_or(0);
                cost += calc.calculate_turn_cost(&t.model, &t.usage).total;
            }
            SubagentSummary {
                agent_id: sa.agent_id.clone(),
                agent_type: sa.agent_type.clone(),
                description: sa.description.clone(),
                turns: sa.turns.len(),
                output_tokens,
                cost,
            }
        })
        .collect();
    out.sort_by(|a, b| {
        b.cost
            .partial_cmp(&a.cost)
            .unwrap_or(std::cmp::Ordering::Equal)
    });
    out
}