deployments: specs

Constants at the top of deployments.py: DAYS=30, MAX_RUNS=500, MAX_FAILED_DETAIL=20, CI_DEFAULT="github_actions", RUN_FIELDS, FAILED, PASSED.

D-repo: which repos

--repo values, else config impact.repos, else the current dir. A value that is an existing directory is a local clone: gh runs with it as cwd (gh resolves the GitHub repo from its remote) and repo_label = basename(realpath), which is how impact labels repos. Anything else is passed as gh -R OWNER/NAME with repo_label = NAME. A repo whose gh call fails (not authenticated, no GitHub remote, no access) is skipped with a stderr line; the others still run.

D-scan: run metadata

gh run list --limit MAX_RUNS --created >=<since date> --json RUN_FIELDS, then runs with createdAt < since dropped (since = now − --days). Kept per run: run_id (databaseId as string), workflow, event, head_sha (lowercase), head_branch, status, conclusion (empty → null), created (UTC), duration_s = updatedAt − startedAt for completed runs, else null. gh 2.45 exposes no actor or attempt in run list; neither is stored. Commit subject (displayTitle) is not read.

D-failed-step: why a run failed

For runs whose conclusion is in FAILED, up to MAX_FAILED_DETAIL per repo (oldest first), gh run view <id> --json jobs: failed_step = "<job> / <step>" of the first step in FAILED of the first job in FAILED; the job name alone if no step failed; null otherwise or beyond the cap.

D-join: run to session

Index = (repo_label, commit sha) over impact.jsonl records with non-null link_p (trailer links, and window links decide confirmed). A run joins the record holding its head_sha in the same repo_label; exact match first, else either sha a prefix of the other. The record copies session_id, harness, link, link_p from that impact record; all four null when unmatched. Only the head commit counts: CI tests the tip of a push or PR.

D-rate: failure rate

Over runs with conclusion in FAILED ∪ PASSED (cancelled, skipped, neutral and unfinished runs excluded): rate = failed / n, separately for runs with a session (agent) and without (baseline). Pitfall from ../deployments.md: flaky tests make the rate noisy; read agent against baseline, not alone.

D-ex: worked example (04-auto-analysis/tests/test_deployments.py, fixtures/deployments/)

since 2026-10-01. Runs: 99 (Sep 30, dropped); 101 push success at sha a (s-a, trailer, p 1.0), 150 s; 102 PR failure at sha b, impact holds b[:12] (s-b, window, p 0.8 by prefix), failed_step "test / unittest"; 103 failure at sha c, whose only linked record is in repo other and whose app record is pending (no session), failed_step "test"; 104 success at c (no session); 105 in progress (conclusion and duration null). Rates: agent 0.5 (n 2), baseline 0.5 (n 2).