Convergence Skill

Decide whether the harness loop continues based on harness gate deterministic evaluations. Uses fixed rules only — no scoring, no model confidence.

Workflow Commands

harness gate preflight
harness impact scope --format yaml
harness evidence run --type unit_test --command "pytest"
harness evidence attach --type build --file external-proof.json
harness finding resume-review FND-001

Rules

PASS

harness gate prints DECISION: CONVERGED
→ harness transition DONE
→ report to user with gate output attached

Never harness transition CONVERGED. Only harness gate may enter CONVERGED.

Continue

A printed DECISION: CONTINUE with DIRECTIVE: RESUME_TYPED_RECOVERY from harness gate grants autonomous permission to resume via typed recovery, repair within bounds, and rerun harness gate without prompting the user to continue.

harness gate prints DECISION: CONTINUE (transitioned GATING -> BLOCKED, iteration incremented)
→ run harness resume
→ repair within bounded scope
→ collect fresh evidence
→ run harness gate

Only DECISION: CONTINUE with DIRECTIVE: RESUME_TYPED_RECOVERY permits resume. If directive is absent, unknown, or conflicts with decision, stop and inspect persisted task; do not follow Next: or call harness resume. This is not a request for routine continue permission.

Blocker dispatch from harness resume:

• open finding → REPRODUCING (via reproduce-finding)
• unverified requirement / red verification → IMPLEMENTING

Escalate

DECISION: ESCALATED from harness gate, or any human/skill escalation trigger, immediately stops autonomous execution. No SPECIFYING, IMPLEMENTING, or evidence-collection resume is permitted. DIRECTIVE: NONE on CONVERGED or ESCALATED ends the loop. It is not permission to resume. An unrecognized DIRECTIVE is not permission to resume, and a free-text Next: line is not a command.

Do not retry a command that failed because a credential is unavailable, authorization was denied, or an external side effect was refused. Escalate immediately.

Precedence is strictly fixed: USER_AUTHORITY_REQUIRED > REPEATED_REGRESSION > NO_PROGRESS > MAX_ITERATIONS.

user authority required by current Gate blockers      -> detected by harness gate (USER_AUTHORITY_REQUIRED)
same finding VERIFIED then open again (regression)   -> detected by harness gate (REPEATED_REGRESSION)
identical blocker fingerprint repeated (no progress) -> detected by harness gate (NO_PROGRESS)
iteration >= max_iterations                          -> detected by harness gate (MAX_ITERATIONS)
same invariant repeatedly violated                   -> human/skill declares
test suite unstable                                  -> human/skill declares
architecture defect suspected                        -> human/skill declares
spec ambiguity blocks verification                   -> human/skill declares

Output exactly one reason code:

SPEC_AMBIGUITY | ARCHITECTURE_DEFECT | REPEATED_REGRESSION |
UNSTABLE_TEST  | REVIEW_DISAGREEMENT  | MAX_ITERATIONS     |
NO_PROGRESS    | USER_AUTHORITY_REQUIRED

Code-detected codes (USER_AUTHORITY_REQUIRED, REPEATED_REGRESSION, NO_PROGRESS, MAX_ITERATIONS) are emitted by harness gate only during GATING. Guard contract drift or unresolved decision instead emits POLICY: USER_AUTHORITY_REQUIRED / DIRECTIVE: HALT_AND_WAIT and preserves task state. Stop without calling harness gate or harness resume; report blocker to user. Text directive is guidance, not authenticated approval. When Gate emits DECISION: ESCALATED, stop and report reason + blockers. For the others YOU must declare them explicitly to the user with evidence — never silently retry past a judgment-call blocker.

Hard Boundaries

1. DONE only via CONVERGED → DONE, and CONVERGED only after DECISION: CONVERGED.
2. ESCALATED ends the autonomous loop immediately. Report reason + full status to user; never route to SPECIFYING, never retry evidence collection, and do not silently continue.
3. When DECISION: CONTINUE and DIRECTIVE: RESUME_TYPED_RECOVERY both appear on a successful Gate call, resume and rerun harness gate autonomously. Otherwise stop and inspect persisted task; never infer permission from DECISION: alone.
4. Autonomous operations must NEVER execute:
   • git commit, git tag, or git push
   • PyPI or npm publication (npm publish, twine upload, uv publish, registry release). The user publishes those registries manually. Local python -m build and npm pack --dry-run are packaging checks. Control-plane publication stays.
   • granting Harness authorizations (harness auth grant ...)
   • accepting Decision records on behalf of user
5. Never reset iteration to dodge max_iterations.