Adds a scheduled GitHub Actions workflow that runs `core demo learning-arc --json`, writes the report to contemplation/runs/<stamp>.json, and opens a PR against main. Operator review on the PR is the ratification gate — preserves the HITL invariant from ADR-0150/0152. Workflow stays disabled until repo variable CONTEMPLATION_ENABLED is set to "true" (soft kill switch in repo settings). Default cadence is nightly; ADR includes a budget table for the 3000 Linux minutes/month available on GitHub Pro. CI never: - commits to main directly - mutates corpora/ or packs/ - ratifies proposals - registers recognizers CI only writes a report file under contemplation/runs/ and proposes the diff via PR. Determinism check (first-run verification): local + CI runs at same SHA must byte-match on proposal_id / trace_hash. Out of scope (noted in ADR): persisted engine_state across CI runs, auto-merge, cross-runner determinism, recognizer growth from CI synthetic traffic. To enable: 1. Repo Settings → Variables → CONTEMPLATION_ENABLED=true 2. Actions → contemplation → Run workflow 3. Review the resulting PR before merging
4.9 KiB
ADR-0155 — CI contemplation runner (W-021)
Status: scoping Date: 2026-05-25
Context
ADR-0150 (W-018) made contemplation autonomous at checkpoint. ADR-0151 (W-017) auto-proposes from enriched candidates at load. ADR-0152 (W-019) closes the engine-authored proposal loop. Operator (Shay) currently runs sessions from a single workstation with intermittent connectivity.
A GitHub Actions runner is deterministic Linux compute that the operator can trigger from anywhere. Running contemplation cycles on that compute amortizes wall-clock contemplation cost without sacrificing CORE's HITL doctrine — provided the output is gated through pull-request review before any corpus mutation.
Budget on GitHub Pro (Student): 3,000 Actions minutes/month on Linux runners (1× multiplier). A 10-min contemplation run every 4 hours costs ~1,800 min/mo (60% of budget); nightly costs ~900 min.
Decision
Add .github/workflows/contemplation.yml:
- Triggers:
schedule:(nightly at 09:00 UTC = 01:00 PST) andworkflow_dispatch:(manual). - Soft kill switch: skips when repo variable
CONTEMPLATION_ENABLEDis not"true". Operator toggles in repo settings without editing the workflow. - Runs
core demo learning-arc --json, writes the report tocontemplation/runs/YYYY-MM-DD-HHMMSS.json. - Opens a PR against
mainwith the new run viapeter-evans/create-pull-request@v7. Operator review on the PR is the ratification gate. - Concurrency group prevents overlapping runs.
The CI runner never commits directly to main, never
mutates corpora/, never registers recognizers, never
ratifies proposals. It only writes a report file under
contemplation/runs/ and proposes the diff via PR.
Invariants preserved
- ADR-0150 HITL gate: every proposal still passes through operator review. CI just stages the candidate.
- Determinism (CLAUDE.md):
ubuntu-latestis consistent enough for trace_hash equality. First-run verification: compare a localcore demo learning-arc --jsonagainst the CI output for the same commit SHA; they must byte-match on theproposal_id/trace_hashfields. If they diverge, the underlying determinism gap is a substrate bug to fix, not a reason to relax the invariant. - No new trust boundary on disk: CI writes only to
contemplation/runs/(a new directory dedicated to CI output); existing trust boundaries are unchanged. - Acceptable Use Policy: output is a project artifact (proposals about CORE's corpus), so contemplation runs are defensibly "production, testing, deployment, or publication of the software project" per GitHub AUP §5. Idle / unbounded compute is not scheduled.
Trust boundary
The CI workflow has contents: write and pull-requests: write
on a branch named contemplation/<date>. It cannot push to main
(protected branch). The HITL surface is the PR review UI —
identical to existing operator workflow for human-authored
proposals.
Out of scope
- Persisted engine state across CI runs. Each run starts from
the committed corpus and produces a one-shot report. A future
ADR may track engine-state evolution across runs by committing
engine_state/under a CI-only branch, but only after operator review on each step. - Auto-merge. Never. Every CI proposal stays open until the operator merges or closes.
- Cross-runner determinism. Pinning to
ubuntu-latestis acceptable; switching runner classes invalidates the trace_hash equality check. - Recognizer growth. ADR-0154 enables the registry to grow from live traffic, but CI runs do not produce traffic the registry should learn from (they are synthetic exercises). The CI runtime sets the producer queue but does not persist derived recognizers to the committed engine_state.
Validation
- First run (manual
workflow_dispatch) produces a PR with acontemplation/runs/<date>.jsonfile. - Local + CI runs at the same SHA produce byte-identical
proposal_idandtrace_hashfields (manual check on first enable). - Workflow exits 0 with no PR created when no proposal is produced (idempotent runs).
- Soft kill: setting
vars.CONTEMPLATION_ENABLED=""causes the job to skip on the next scheduled tick.
Operator runbook
- Repo Settings → Secrets and variables → Actions → Variables:
set
CONTEMPLATION_ENABLED=true. - Actions tab → "contemplation" workflow → "Run workflow" once to verify.
- Watch the resulting PR; review the proposal as you would any
core teaching proposalsentry. - Merge to accept (proposal becomes part of the audit trail) or close to reject.
- Disable: set
CONTEMPLATION_ENABLED=""(empty); the workflow exits early without consuming meaningful minutes.
Closure
After this ADR, CORE has a remote compute path for contemplation that preserves the operator-as-gate invariant. The operator gains asynchronous contemplation cycles tied to the project's own audit trail, with no new infrastructure to maintain beyond a single workflow file.