feat: Adapt CodeFire-app ideas for enhanced session logging and project intelligence

[sinh-x]

Inspiration

CodeFire is an open-source (MIT) desktop companion that gives AI coding CLIs persistent memory via MCP. Several of its patterns would significantly improve ai-usage-log's session tracking and project intelligence capabilities.

What CodeFire Does Well

Feature	How It Works
Direct .jsonl parsing	Reads Claude Code session files line-by-line, extracting model, tokens (in/out/cache), tool counts, git branch, timestamps
Auto project discovery	Scans ~/.claude/projects/, decodes path-encoded folder names back to real filesystem paths
Live session monitoring	Watches active sessions — context window %, token burn rate, cost estimates
Session → task linking	Sessions reference tasks worked on; tasks reference originating sessions
Cost tracking	Per-session and aggregate cost estimates by model pricing
Codebase snapshots	Point-in-time file tree capture per session
Daily briefings	AI-generated digest summarizing activity across all projects

Proposed Adaptations

Phase 1 — Direct .jsonl Parsing (High Priority)

Currently read_claude_sessions reads JSONL files but could extract more structured data. Adapt CodeFire's parsing approach:

• Extract model ID per session (for cost calculation)
• Extract token counts (input, output, cache_creation, cache_read) per turn
• Calculate estimated cost using model pricing tables
• Extract git branch active during the session
• Extract context window utilization (% of model limit)
• Parse tool call results for success/failure status

This overlaps with #8 but focuses specifically on the .jsonl parsing layer rather than stats aggregation.

Phase 2 — Auto Project Discovery

• Scan ~/.claude/projects/ to auto-discover project folders
• Decode path-encoded folder names (e.g., -home-sinh-git-repos-foo → /home/sinh/git-repos/foo)
• Auto-link sessions to projects based on discovered paths
• Expose via MCP tool: list_discovered_projects

Phase 3 — Session ↔ Avodah Task Linking

Enable bidirectional linking between ai-usage-log sessions and Avodah tasks:

• Add optional task_ids field to session metadata
• When saving a session, accept Avodah task IDs that were worked on
• Add session_ids lookup — given a task ID, find all sessions that touched it
• Expose via MCP: link_session_task, get_task_sessions

This creates the feedback loop: AI reads tasks → works on them → session logs reference the tasks → next session can see history.

Phase 4 — Cost Tracking & Reporting

• Maintain a model pricing table (configurable, with sensible defaults)
• Calculate per-session cost from token counts × model pricing
• Aggregate cost by: day, week, project, model
• Expose via MCP: get_cost_summary(period, project?)
• Include cost in daily summary output

Phase 5 — Live Session Awareness (Stretch)

• Detect currently active Claude Code sessions (via .jsonl file locks or recent writes)
• Expose get_active_sessions MCP tool showing: project, model, duration, tokens so far, context %
• Could power the Context Guardian feature (auto-save when context runs low)

Non-Goals

• No GUI — ai-usage-log stays MCP-only; visual dashboards are Avodah Flutter app territory
• No code indexing/RAG — out of scope for session logging
• No browser automation — CodeFire feature, not relevant here

Relationship to Other Issues

• Enrich session statistics for deeper usage pattern analysis #8 (Enrich session statistics) — Phase 1 feeds richer data into Enrich session statistics for deeper usage pattern analysis #8's stats layer
• Add response token estimation to MCP tool responses #5 (Response token estimation) — Phase 1's token extraction makes Add response token estimation to MCP tool responses #5 more accurate
• Add version tracking to session logs and capability changelog #1 (Version tracking) — Orthogonal, but sessions with version info + cost would be powerful

References

• CodeFire repo — MIT license, session parsing logic in swift/ and electron/src/
• CodeFire landing page