deployments

Reads CI runs (GitHub Actions through the gh CLI) and joins each run to the agent session whose commit it tested, using impact.jsonl. Reads run metadata only: workflow, trigger event, head sha and branch, status, conclusion, timestamps, and for failed runs the name of the failed job and step. No logs, no commit subjects, no actor. Specs: METHOD.md (D-*). Placeholder spec and coverage: ../deployments.md.

Input: gh run list / gh run view --json jobs per repo (uses the user's existing gh auth, read-only calls); impact.jsonl from the data dir.
Output: deployments.jsonl in the data dir (or --out), one record per run; a one-line summary on stdout (failure rate for agent-linked runs vs the repo baseline).
CLI: python3 04-auto-analysis/deployments/deployments.py [--repo PATH|OWNER/NAME ...] [--days N] [--ci NAME] [--config PATH] [--impact FILE] [--out DIR]
Pipeline stage deployments of 03-analyze/pipeline.py (after impact2, skipped without network or gh); its summary goes to section deployments of readout.json (../../03-analyze/output.md). Standalone, run it after impact (it only joins linked impact records).
Record schema: 02-measure/schemas/deployment.schema.json.

Reconfigure

Repos: --repo (repeatable; a local path gives repo_label = basename, same as impact; OWNER/NAME gives NAME), else config impact.repos, else the current dir. CI connectors: dict CI in deployments.py: github_actions (gh CLI), gitlab_ci (placeholder, raises). Add one with signature fetch(repo, since, run) -> list[dict] returning the run dicts described in METHOD.md D-scan.

Adapt, trade-offs, decisions for the human

Adapt: the repos list; MAX_RUNS for busy repos; a GitLab connector.
Trade-offs: a run links only through its head commit, so in a push of several commits only the tip's session is credited (D-join). Pending impact window candidates are not joined until decide confirms them.
Decisions for the human: whether workflow step names (which can contain the command a step runs) may be stored; which repos may be queried.