ci-routing

CI Routing

The single source of truth for acting on an orc-ci-investigator diagnosis — $orc-ci and $orc-flow's post-open CI gate both execute this protocol instead of restating it. Input: the diagnosis (verdict + evidence table + fix list), already saved verbatim to ${ORC_STATE_DIR}/<branch>/files/ci-diagnosis.md.

Single-render rule: the diagnosis is saved via Write — never echoed wholesale into conversation. It gets ONE capped rendering: the fix list + one evidence line per root cause + the file path. Every later step references the saved file; quotes are capped at 3 evidence lines.

green

Say so — one line, run ID cited. Mark the session step complete and stop.

fixable

Render the one preview (> **📋 Preview — CI fix list**: N root causes across M jobs, then the numbered <file>:<line> — <change> — clears <job> list), then the available user-question tool (or ask in chat) — one call, two questions (the landing decision is collected up front, so a green fixer report lands without a second stop):

1. Apply the fix list? — Apply (dispatch orc-code-fixer with the fix list + diagnosis path; it applies edits, runs tests, returns a diff + test summary) / Edit the list first / This needs real debugging (jump to needs-debug) / Abort.
2. After a green fixer report, land it how? (ignored unless the fixer ran and came back green) — Commit + push + re-watch (Recommended: orc-git-commit, git push, re-watch the run until it concludes) / Commit only / Show me the diff first.

Red fixer report → discard the landing answer and return the new evidence to the investigator (it sees the failed attempt) — or offer the needs-debug hand-off. Loop until green or the user explicitly advances with red CI (logged to the digest).

flake

Quote up to 3 run-history evidence lines proving intermittency, then the available user-question tool (or ask in chat): Re-run failed jobs (gh run rerun <id> --failed, then watch) / Re-run everything / Not convinced — treat as fixable or needs-debug / Skip.

infra

No code changes to make. Surface the diagnosis path + the agent's recommendation (pin the action version, report to the platform team, wait out the outage) with its capped evidence. Offer a re-run only when the agent recommended one.

needs-debug

The failure needs root-cause work beyond the diff:

> **➡️ Next**
>
> Genuine regression — run `$orc-debug` seeded with the CI diagnosis at `.orc/<branch>/files/ci-diagnosis.md`.

the available user-question tool (or ask in chat): Start $orc-debug now (diagnosis pre-seeded as the bug description — its investigator starts from the CI evidence instead of zero) / I'll run it later / Abort.

Iron rules

• No fix without the investigator's diagnosis. The fixer only ever receives an evidence-cited fix list — never "CI is red, go fix it".
• Re-runs are gated. gh run rerun only after the flake evidence is shown and the user confirms.
• Never git push --force — a red pipeline is not a license for history rewrites.