impact

Links agent sessions to the git commits they produced and puts cost next to kept and reverted work. Reads git metadata only (no diffs, no file contents). Specs: METHOD.md (L-*).

Input: sessions.jsonl (analyze), decisions.jsonl (decide, optional) from the data dir; git repos on disk.
Output: impact.jsonl, impact_summary.json in the data dir (or --out).
CLI: python3 04-auto-analysis/impact/impact.py [--days N] [--harness ID] [--config PATH] [--out DIR] [--sessions FILE] [--decisions FILE]
Two passes: run 1 writes trailer links and window candidates (link_p: null); decide asks commit_link per candidate (subject per L-subject); run 2 applies the answers (L-decided).

Reconfigure

Config key impact: vcs (connector name, default local_git), repos (list of repo paths to scan instead of session.repo), read_subjects (bool, default false).

Connectors live in impact/connectors.py, dict VCS: local_git (git CLI), github_api (placeholder, raises). Add one with signature commits(repo, since, until, cfg) -> list[dict] returning the commit dicts described at the top of connectors.py, and register it in VCS.

Example: scan two repos even when sessions lack a repo path:

{"impact": {"vcs": "local_git", "repos": ["~/code/app", "~/code/lib"], "read_subjects": false}}

Adapt, trade-offs, decisions for the human

Adapt: impact.repos to the repos the team works in; a VCS connector for a hosted API instead of local git (github_api is a placeholder).
Trade-offs: trailer links are exact but need harnesses that write trailers; time-window links catch more commits but are guesses until decide confirms them (commit_link). Commit subjects are not used; read_subjects is reserved and changes nothing yet (METHOD.md).
Decisions for the human: which repos may be scanned; whether $/kept commit is a fair measure for this team.