TracePilot
TracePilot is built for developers who lean on coding agents and want a clearer view of what happened across their sessions: prompts, assistant turns, subagents and the messages they exchange, tool calls, todos, checkpoints, context growth, prompt-cache state, token usage, cost, search, and orchestration.
It reads the sessions Copilot CLI writes under ~/.copilot/session-state/ and Claude Code writes under ~/.claude/projects/, indexes them locally, and shows them side by side in a Tauri desktop app backed by Rust, SQLite, and Vue. Your session data stays on your machine.
Project status: TracePilot is early-stage software. It is useful today, but the UI and internal data model still change quickly.
Platform status: TracePilot is tested on Windows. Apple Silicon Macs get a preview build; Intel Macs and Linux are not yet release targets.
What you can do
Inspect sessions
Browse your session history as a searchable library, filter it by source, repository, or folder, and see which sessions are Busy or Waiting right now. Open any session for a tabbed deep-dive:
| Area | What it helps with |
|---|
| Overview | Headline numbers, how the session ended, incidents, changes, plan and checkpoints, and turn activity over time. |
| Conversation | User/assistant turns, reasoning, model switches, parallel subagents, skill invocations, tool calls, and rich tool-result renderers. |
| Events | Raw session events with filtering and pagination for debugging parser or CLI behavior. |
| Todos | The session's task list, as a list or dependency graph. |
| Metrics | Cost, token and cache usage per model or per agent, duration, code changes, and prompt-cache timing. |
| Context | How the context window grows by turn or over time (system prompt, tool definitions, conversation), where compactions happened, and which tools contribute most. |
| Explorer | Files inside the session state directory, with viewers for Markdown, JSON/JSONL, CSV, SQLite, images, and more. |
| Timeline | Swimlane, waterfall, and agent-tree views, plus a Messages view of how agents launched, messaged, and read back from each other. |
The session header shows the model, reasoning effort, and whether the prompt cache is warm, expiring, or expired, with a live countdown. Resume in Terminal reopens a session in its folder. Sessions can open in pop-out windows and auto-refresh while they run. Rich renderers cover common tool output such as diffs, patches, shell output, search and web-search results, SQL, file trees, ask_user prompts, and the agent-control tools (read_agent, write_agent, list_agents).
| Conversation | Agent timeline |
|---|
| |
| Inter-agent messages | Todos |
|---|
| |
Copilot CLI and Claude Code
Claude Code sessions sit next to Copilot CLI sessions in the library, Conversation, Timeline, Metrics, Context, Search, Analytics, and Export. First-run setup turns them on when it finds Claude Code's folder (CLAUDE_CONFIG_DIR, else ~/.claude); Settings → Claude Code changes the folder or turns them off.
- Sessions, Search, and the analytics pages switch between All, Copilot, and Claude Code. Search also takes
source:copilot and source:claude.
- Copilot usage is priced in AI Credits and Claude Code usage in API-equivalent US dollars. With both sources, Analytics and Models total them in USD.
- Launching, the SDK bridge, the config injector, and the skills and agents editors stay Copilot-only.
TracePilot only reads Claude Code's files. Claude Code deletes transcripts older than its cleanupPeriodDays setting (30 days by default), and Settings → Claude Code can raise it in one click. See Claude Code sessions for details.
Search and analyze session history
TracePilot maintains a local SQLite/FTS5 index so you can search across session metadata and conversation content without sending data to a remote service.
Analytics, Tools, and Code share a dashboard of per-day charts, split by source. They help answer questions such as:
- Which repositories, models, and tools are consuming the most time, tokens, or cost?
- How well is the prompt cache working, and how often do sessions resume after it expired?
- How often do tool calls fail, and which tools dominate a session?
- Which files and file types are touched most often?
- How do two sessions compare after normalizing by turns or duration?
Available top-level analysis pages include Search, Analytics, Tools, Code Impact, Model Comparison, and Session Comparison.
| Search | Analytics | Tool analysis |
|---|
| | |
This is the beginning of the README. Read the rest at the original. Read the full README ↗
