ms-continue-poetics-loop - SKILL.md Agent Skill

name: ms-continue-poetics-loop description: Run exactly one bounded poetics adaptation-recognition loop (D42/D50/D53, three iterations, two required passes) following the canonical handoff workflow. Use when continuing the poetics arc from a prior session. argument-hint: "[--targets D42,D50,D53] [--iterations 3] [--passes 2]" disable-model-invocation: true allowed-tools: Bash, Read

Run one bounded gated adaptation-recognition loop, following the literal handoff in notes/poetics/2026-05-27-handoff-adaptation-recognition-loop.md. This skill has side effects (paid OpenRouter API calls and writes to the live poetics sidecar DB), so it is user-only.

Defaults (do not broaden without explicit user instruction)

targets: D42,D50,D53
arms: routine,none,peripeteia-only
iterations: 3
required passes: 2

The handoff is explicit: do not broaden the target set until the focused loop either passes or diagnostically saturates.

Steps

1. Confirm no loop is currently running

ps -axo pid,ppid,stat,command | rg 'run-poetics-adaptation-loop|run-poetics-production-batch|generate-pedagogical-dramas|codex exec' | rg -v ' rg ' || echo "no loop running"

If a loop IS running, STOP. Tell the user and exit — do not start a parallel loop.

2. Sync the branch and run the focused-tests gate

git pull
git status --short --branch
node --test tests/generatePedagogicalDramas.test.js tests/criticPoeticsStructure.test.js tests/runPoeticsAdaptationLoop.test.js

If tests fail, STOP and report — do not run the loop on a broken contract.

3. Run exactly one bounded loop

npm run poetics:adaptation-loop -- --max-iterations 3 --required-passes 2 --skip-existing-scores --generator claude

--generator claude routes generation through the claude-code CLI (Max-plan, attended) instead of the production-batch default (codex). Override to --generator codex only if explicitly asked — the two CLIs use separate quota pools and a switch changes which one drains.

This may take a long time (multiple iterations × generation + critic-panel scoring). Stream output. The loop writes a stamp like phase2-adaptation-recognition-loop-<TIMESTAMP>Z and persists status to exports/<stamp>-loop-status.{md,json}.

4. Inspect status

ls -t exports/*adaptation-recognition-loop*loop-status.md | head -3
latest=$(ls -t exports/*adaptation-recognition-loop*loop-status.md | head -1)
sed -n '1,220p' "$latest"

Summarise to the user:

overall pass / fail
per-iteration pass count (e.g. "1/2 passes")
failure classes if any (organic_or_ambiguous_recognition, quality_warning, critic_split, insufficient_scores, scorer_error)

5. If the loop passes, package each completed iteration

For each iteration that passed:

npm run poetics:package-run -- --run-id <RUN_ID>

This archives raw artifacts to artifacts/poetics-runs/<RUN_ID>/ (gitignored) and writes a small manifest to config/poetics-calibration/runs/<RUN_ID>.manifest.json (committable).

Confirm with the user before committing — they may want to inspect first.

6. If the loop fails

Use the poetics-loop-diagnostician subagent on the failed iteration's status artifact:

Agent({ subagent_type: "poetics-loop-diagnostician", prompt: "Diagnose <RUN_ID>" })

If the same failure class has appeared for 3 consecutive iterations, STOP and recommend mechanism / prompt surgery per the handoff termination rule — do not just rerun.

Critical rules

Never --force re-run a loop that produced evidence; package and commit the manifest first.
Never broaden targets without explicit user instruction; the gated-loop discipline is the whole point.
Never claim "the mechanical adaptive tutor is complete" based on a single passing iteration. The defensible claim is in the handoff: branch-valid tutor peripeteia can produce recognitive reframe in clean low-organic anchors; one full gated iteration passed; not yet repeat-stable.
Archive payloads stay out of git; only the manifests get committed. The user must copy artifacts/poetics-runs/<RUN_ID>/ to durable external storage before deleting the worktree.