Ops Integrate

ops-integrate

The service that answers one question: may this PR land, and if so, land it.

Visibility: service. A framework loop may command it. It is the only thing ops-merge-loop commands.

Why it exists. The gates used to live in the merge loop, alongside the sweep, the poll cadence and the caps. That put merge policy in the same place as merge scheduling, and made the loop resolve base branches and merge strategies itself — the leak ops-branching's privacy rule exists to close. Split the two: policy here, scheduling in the loop.

It is PR-generic. Nothing below asks who wrote the PR or which issue produced it. Every gate is a machine-verifiable property of the PR itself, which is why dependabot PRs and human-authored PRs land through the same path as loop-authored ones.

Invocation

ops-integrate land '<context-json>'

An absent context is {}. Reject any other action — do not guess.

The outcome vocabulary

outcome is exactly one of:

detail is one plain-language line naming the specific blocker — "check Build (Release) failing", "conflicts with its base" — so the caller can comment it verbatim.

Standing: a convention, not a contract. The conformance spec requires no error vocabulary and makes the result shape a SHOULD verified by evals. This list is a fixed set the evals assert on, not a validated envelope, and it travels inside the recommended { "ok": … } shape. It becomes a MUST only if a non-model consumer ever parses it — a shell step, not a loop. Today every caller is a model deciding what to say.

Action: land

Step 1 — resolve what you need

• The landing label: ops-repo-meta · identity → labels.land. Ask for the purpose, not the name — a repo may have renamed it.
• Nothing else. In particular do not resolve a base branch, a release base or a merge strategy. You cannot, and that is deliberate: ops-branching holds them privately.

Step 2 — the gates

Three gates you can check, then a fourth that is the merge attempt itself. Gates 1 to 3 are reads: check them in order and on the first failure stop and return that outcome, having changed nothing. The base check cannot work that way, and pretending otherwise is how this section used to read: ops-branching is command-only, so there is no read that answers "is this base a merge target" — the only way to ask is to tell it to merge and see whether it refuses. That is safe, because a refusal changes nothing, and it is why the base is checked last, after everything cheap has already passed.

1. The landing label is present. Read the PR (github-ops → Get a PR) and confirm it still carries labels.land. A caller that swept for the label a minute ago is not evidence: a human may have pulled it since. → blocked:no-label.

   The label is the human approval. A GitHub review "approve" is not used as the signal, because a maintainer cannot approve a PR they authored and these PRs are commonly theirs. As a second safeguard, an unresolved "changes requested" review is a human veto that outranks the label → blocked:changes-requested.

2. CI is genuinely green. ops-ci · status. Require state == "green", which means every check passed — not "nothing failed yet". pending → blocked:ci, same as red; waiting is the loop's business, not this service's.

   Re-check it even though the loop already polled. That duplication is deliberate: neither current consumer has branch protection and native auto-merge is disabled, so this check is the only gate that actually holds. A service that trusts its caller has no gate at all.

3. Mergeable, no conflicts. From the same PR read. If it is merely behind its base, update the branch and go back to gate 2 — updating restarts CI, and the green you saw was for different code. If it genuinely conflicts → blocked:conflict.

Step 3 — the base gate and the merge, in one call

Once gates 1 to 3 hold, hand the PR to ops-branching · merge. That single call is both the last gate and the landing: the owner of the branch model decides whether the base is a merge target, and merges if it is.

Call it once. An earlier version of this skill listed the base as a fourth gate and then listed the merge as a separate step, which reads as two calls to the same command. Idempotency made that survive, which is worse than it failing: nothing showed the second call was doing nothing.

Map what comes back:

Read refused, never detail. detail is a sentence for a human comment; the classification is the field. A skill that string-matches prose to tell a release PR from a mistargeted one will get it wrong the first time the wording changes.

A release-base PR is not an error. It is the release path doing its job, and ops-release owns it. Return ok: true and let the caller stay quiet.

Pass line straight through on a merged outcome. You do not derive it and you do not interpret it: the merge reported which live line the PR landed on, and the caller needs it to close the right issue and to start a port. Dropping it here forces the loop to read a line out of a branch name, which is the leak everything above is arranged to prevent.

Idempotency

An already-merged PR MUST return merged with the existing commit, never a second merge attempt. This is not theoretical: the landing label stays on the PR after it lands, so a sweeping caller hands you the same PR on its next run. ops-branching · merge is idempotent for the same reason; check the PR's merged state at gate 1 and short-circuit.

On failure

Leave the system safe. Never a partial merge, never a force, never a branch deleted without a merge. If the merge call itself fails, return { ok: false } with the reason and change nothing else — a failed landing must be safe to retry on the next sweep.

Rules

• Never merge with an unmet gate. No exceptions, no "it's probably fine", no override context. The gates are the whole reason this service exists.
• Never call ops-branching · merge twice in one land. It is the base gate and the merge. Calling it to "check" and again to "do" is one call too many, hidden by idempotency.
• Never use GitHub's native auto-merge, and never treat a mergeable state as evidence of green — with no branch protection, GitHub will land a red PR.
• Never force-merge, never force-push.
• Never resolve or return a branch name or a merge strategy. Ask for an outcome.
• Never label, comment or notify. The caller reports; this service decides. Two things commenting on one PR is how a loop starts arguing with itself.
• Never touch the release path. A release-base PR is skipped, not landed, not flagged.
• One PR per call. Sweeping is the loop's job.