Summary
Enrich checkpoint metadata with tool-level event data extracted from agent session transcripts. This provides finer-grained insight into what the agent actually did during a commit — which files it read, which tools it invoked, and what edits it made.
Motivation
Current checkpoints capture the session transcript as a whole, but don't extract structured data about individual tool invocations. Tool-level events would enable:
- Better attribution: knowing exactly which files the agent touched and how
- Richer checkpoint browsing: users can see a timeline of tool calls alongside the diff
- Foundation for analytics: aggregating tool usage patterns across sessions
- More accurate "files touched" tracking for multi-file agent sessions
Implementation Notes
- Extend the JSONL parser in
internal/agent/claude/ to extract tool use events (tool name, input summary, file paths)
- Store tool events as a structured field in checkpoint metadata (e.g.,
tools_used array)
- Keep extraction best-effort: log warnings on parse failures, never block checkpoint creation
- Tool events should be lightweight summaries, not full tool input/output
Source
Inspired by entireio/cli#860
Summary
Enrich checkpoint metadata with tool-level event data extracted from agent session transcripts. This provides finer-grained insight into what the agent actually did during a commit — which files it read, which tools it invoked, and what edits it made.
Motivation
Current checkpoints capture the session transcript as a whole, but don't extract structured data about individual tool invocations. Tool-level events would enable:
Implementation Notes
internal/agent/claude/to extract tool use events (tool name, input summary, file paths)tools_usedarray)Source
Inspired by entireio/cli#860