Skip to main content

Initial rollout: dry-run

The CI reconciler initially observes only. The checked-in policy at scripts/ci/ci-reconciler-policy.json disables recovery. Missing or invalid policy fails closed; a workflow input cannot enable recovery. Mutation activation requires an exact-source Security ACK and the release owner’s freeze lift after dry-run acceptance. Source review alone does not establish scheduled or hosted execution. The workflow .github/workflows/ci-reconciler.yaml runs on standard ubuntu-latest runners, with no dependency installation or application build. It uses GitHub’s built-in job token; no provider credential is required. The fifteen-minute schedule is a desired observation cadence, not a recovery SLA. Hosted queue delays and schedule coalescing can delay execution.

Coverage and identity

The trusted operational manifest explicitly names eight core producers:
  • Test
  • TypeScript Type Check
  • Oxc Check
  • i18n Check
  • Security Checks
  • CodeQL Manual Fallback
  • Codex Plugin Check
  • Required account MFA database contract
The manifest covers main and production explicitly. It is not proof that all apps, builds, Mobile checks, or release requirements are green. App applicability and Mobile qualification remain with their owning release/check owners. New branches and gates require reviewed policy entries; historical release branches are not discovered and activated automatically. The controller reads the operational manifest and Actions history. Comparison with effective required-check rules is separate proposed work. Missing GitHub required-check configuration is a known enforcement gap, not proof of zero required coverage. External/manual gates remain outside this manifest; they are never replaced with synthetic success. Each Actions summary identifies the branch head and declared gate count:
Failed terminal checks appear separately. All terminal does not mean all green. Deterministic failures stay failed and return to their source owner. Cancelled and missing checks are recovery candidates; queued or running matches are deduplicated.

Trust and recovery boundaries

Only tutur3u/platform on the trusted main workflow ref can plan. Completion events must name an allowlisted producer in this repository on main or production; fork and pull-request events cannot authorize recovery. The reconciler excludes itself from completion triggers. Both jobs check out github.workflow_sha, never upstream branch code or artifact code. Planning has actions: read and contents: read. Recovery alone has actions: write and contents: read. Jobs have five-minute timeouts and the workflow serializes runs without cancellation. The recovery matrix contains only branches with actual admitted actions and executes one branch at a time. When recovery is enabled through reviewed source policy, the controller re-reads HEAD immediately before every mutation and stops an old plan if HEAD advances. It may rerun a cancelled allowlisted run at the exact current SHA or dispatch a supported allowlisted workflow after exact-head admission. It never cancels protected runs or retries deterministic failures into green. Dispatch identity requires API readback; an ambiguous write outcome requires reconciliation rather than blind replay. Recovery is capped at three attempts per branch/head/workflow/policy identity, with one-, five- and fifteen-minute backoff across subsequent ticks. No sleeping worker holds a runner. Bounded API history and pagination must be complete enough to establish eligibility; unavailable or truncated history blocks recovery. This workflow never starts releases, deployments, staging/production migrations, or provider publication. The disposable account-MFA CI contract is not a production migration. It does not write branch rulesets or provider settings.

Durable receipts

Planning uploads an intent before the recovery job can start: The recovery job downloads only the exact branch/SHA intent from its own workflow run. Artifact content is data for the trusted controller, never executable code. An observation artifact does not count as a recovery intent or attempt. The recovery head-age fence is 20 days, inside the 30-day receipt retention window. Recovery fails closed before an expired ledger could make an old attempt look new. Missing, expired, unavailable or incomplete durable history is a block, not an empty retry budget. Preserve receipts when investigating unknown outcomes.

Acceptance and stop procedure

Before enabling recovery, require focused controller and workflow tests for head races, dedupe, concurrent events, retry caps and backoff, unknown outcomes, unsupported workflows, invalid policy, external contexts, recursive triggers and release/migration exclusions. Require exact-source independent Security review, then dry-run scheduled and producer-completion receipts. Verify that failed checks remain visible and that no mutation occurs with recovery disabled. Stop recovery by disabling it in trusted source policy. Stop observation through the repository’s authorized workflow control if needed. Neither action grants permission to change release gates. After any source change, repeat the relevant exact-source review and hosted acceptance; old receipts do not prove a new head. A separate proposal can enforce the validated manifest through a public-repository GitHub Free ruleset. It needs a dry-run comparison of conditional checks and App identities and its own exact Security ACK before any ruleset write. This proposal is outside the initial rollout and is not performed by the reconciler. See the GitHub Actions runbook for the remaining validation and delivery evidence requirements.