Agent Skills: Quiesce

Quiesce

Thin session-quiescence preset. quiesce does not re-implement orchestration, delegation, convergence, or anomaly detection — it composes the native /goal condition-loop with a pluggable inner driver. Every behavior below lands on a primitive that already exists in this repo (or a host built-in).

Purpose

Drive the current work session to a QUIESCENT steady state — no open ticket/gap/fix/failure/PR, every PR green and answered, agentic convergence reached — by wrapping the native /goal condition-loop around a pluggable inner work driver, and codifying the operator's recurring "converge the whole session" invocation as one reusable, override-friendly, token-economic command.

When to use

• The operator wants the session driven to a clean/green/converged steady state.
• One or more open PRs of this session need PDCA convergence before merge.
• Out-of-radar items (gaps/failures/warnings) should be captured as tickets, not lost.
• A long unattended run where per-step approval would dominate wall-clock.

When not to use

• Single-shot edit / single-file fix / typo — disproportionate ceremony.
• Read-only Q&A — answer directly.
• ONE specific goal needs decompose+delegate across agents -> use auto-pilot.
• Destructive ops (force-push protected, drop prod) — always HITL.
• Goal not yet stable / operator still exploring — wait for stability.

Trigger Phrases

Canonical invocations that should activate this skill:

• "quiesce" / "/quiesce"
• "session quiescence" / "drive this session to green" / "sessao limpa/convergente"
• "converge all open PRs" / "converge open PRs"
• "zero-open loop" / "leave nothing pending" / "clean session loop"

Quiescence predicate (default --condition)

Override preservation: a custom --condition REPLACES the default predicate, but the three rubric conjuncts below (C1 loose-end sweep · C2 check classification · C3 stale-verdict re-measurement) are invariants that survive any override — if the custom condition cannot honor them, the run must report them as skipped-by-override with rationale (never silently dropped). Silent weakening via override is the failure this clause exists to prevent.

QUIESCENCE := NOT(unaddressed in-scope TICKET or GAP or pending FIX or FAILURE or open PR)
              AND every PR green (all required checks pass)
              AND every PR comment answered
              AND agentic convergence reached
              AND loose-end sweep across ALL worktrees (rubric C1)
              AND check-failure classification recorded (rubric C2)
              AND stale-verdict re-measurement (rubric C3)

Rubric conjuncts (pilot-eval FAILs 2026-08-16, PR #354): C1 — before quiescence, list unpushed commits in EVERY worktree (others' = reported, never reaped; own = blocker). C2 — every failed check classified [externo-quota | gate-real] with the error message as evidence BEFORE any merge. C3 — reviewDecision pinned to an old commit ⇒ re-measure the findings against current headRefOid; never read the verdict, re-read the code.

Agentic convergence = all bot reviewers + CI (e.g. CodeRabbit, Amazon Q, Qodo, gitleaks) GREEN or resolved — see CONTRIBUTING.md (Bot review convergence) and .claude/rules/pr-reviewer-communication.md.

Non-blocking tickets: a tracking ticket filed for an out-of-radar / out-of-scope / deferred item does NOT block quiescence — capturing-and-deferring IS the resolution (Boy-Scout "don't lose it"). Only unaddressed in-scope obligations keep the loop open; otherwise filing tickets would make QUIESCENCE unreachable.

How it works

operator invokes /quiesce
        |
        v
  resolve flags -> defaults unless overridden
        |
        v
/goal --goal-aware --scope=<scope> --condition='<predicate>'   (OUTER loop — host built-in)
        |  each iteration v
        v
<inner driver: default skills/auto-pilot>                      (INNER work driver)
        |-- PDCA-loop each OPEN PR of <scope> (steelman -> critique -> fix -> re-review)
        |-- PR red/blocked on a bot-reviewer finding -> route EACH finding to
        |     skills/bot-finding-arbiter (*Praetor*) — the DEFAULT per-finding handler
        |     (OODA -> 7-way disposition + teach-the-bot edict), instead of ad-hoc PDCA
        |-- out-of-radar item -> file a tracking ticket (per CONTRIBUTING tracker) + cross-link
        \-- emit exactly ONE STOP marker as the last line of the turn
        |
        v
  /goal Stop-hook evaluator reads the marker -> CONTINUE or terminate at QUIESCENCE

Composition (the wiring it emits)

The resolved control flags are applied per-driver (translated for auto-pilot, native for auto-orchestrator) — never silently dropped. auto-pilot's own surface is only --mode/--band/--max-depth, so its controls are translated (see table):

/goal --goal-aware --scope=<scope> --condition='<condition>' --action:{
  <driver-invocation>      # resolved per the table below (flags translated, not dropped)
    "<instructions>; honor: auto-merge=<auto-merge> (reason "<reason>"),
     auto-fix=<auto-fix>, self-fix=<self-fix>, max-pdca=<n>;
     PDCA-loop each OPEN PR of <scope> as a converge-prompt;
     for any out-of-radar item (gap/pending/failure/warning/error) file a
     tracking ticket (per CONTRIBUTING.md) + cross-link; keep created tickets updated."
}

<driver> resolves from --driver:

| --driver | Resolves to | How resolved flags are applied | |---|---|---| | auto-pilot (default) | /auto-pilot "<goal+controls>" --band=<L1\|L2\|L3> | auto-pilot's surface is --mode/--band/--max-depth only, so --autonomy-threshold -> --band (>=0.85 L3, >=0.65 L2, else L1) and --auto-merge/--auto-fix/--self-fix/--max-pdca are carried in the goal text (honored by its PDCA + merge gate). In-repo (portable). | | auto-orchestrator | /auto-orchestrator --goal-aware --scope=... --auto-merge=... --auto-fix=... --self-fix=... --autonomy-threshold=... | controls passed as native flags; only if installed | | <custom> | any host orchestrator the operator names | operator maps the flags |

The operator's original /goal ... /auto-orchestrator ... pattern is reproduced exactly via --driver=auto-orchestrator. The default stays in-repo for portability.

Override parameters

| Flag | Default | Allowed / Notes | |---|---|---| | "<instructions>" (positional) | empty | extra free-text appended to the driver action | | --scope | this.session | this.session | repo | branch | ticket:<id> | pr:<n> | | --condition | (quiescence predicate above) | override the termination predicate string — rubric conjuncts C1-C3 survive any override (skipped-by-override must be reported with rationale); MAY be a dod-as-prompt.termination_predicate (a measurable Prisma-derived DoD) when driven by ooda-loop — the recovered goal's D/T/J leaves become the stop test | | --driver | auto-pilot | auto-pilot | auto-orchestrator | <custom> | | --auto-merge | authorized | authorized | hold | off | | --auto-merge-reason | (operator invocation) | required-non-empty when authorized; invoking /quiesce --auto-merge=authorized IS the authorization | | --auto-fix | enabled | enabled | disabled | | --self-fix | enabled | enabled | disabled | | --autonomy-threshold | 0.85 (HIGH) | 0.0-1.0 — maps to auto-pilot band (>=0.85->L3, >=0.65->L2, else L1) | | --max-pdca | 6 | int — per-PR PDCA iteration cap |

STOP-marker grammar (paired with --goal-aware, not re-authored)

Emit exactly ONE terminal marker as the last line of each turn. The HTML-comment prefix is low-collision; the /goal Stop-hook evaluator reads it:

<!--ORCH-STATUS: STOP-DONE -->     quiescence reached — nothing pending
<!--ORCH-STATUS: STOP-HITL -->     HITL escalation required (also prepend above any action block)
<!--ORCH-STATUS: STOP-ERROR -->    unrecoverable error (subagent / network / rate-limit)
<!--ORCH-STATUS: CONTINUE -->      iteration done; work remains; evaluator decides next turn

Relationship to siblings

| Tool | Scope | Drives | |---|---|---| | quiesce (this) | the SESSION | termination predicate over ALL open items -> steady state | | auto-pilot | ONE operator goal | decompose -> select -> spawn -> converge | | converge | N proposals | 5-act merge (invoked inside PDCA when a PR has competing proposals) |

quiesce MAY invoke auto-pilot for sub-goals; it never re-implements delegation, convergence, or anomaly detection.

Auto-merge

--auto-merge=authorized lets the driver queue GitHub native auto-merge only when all gates pass (mergeable + green + all-comments-answered + agentic convergence

• autonomy >= --autonomy-threshold + non-empty reason) and no refusal applies (target is a protected deploy branch, CI/infra files touched, native auto-merge disabled on the repo, or operator cancel). Otherwise use hold (operator merges) or off. Always --squash --delete-branch. Fire-and-forget — no local merge queue across turns (amnesic-safe; delegates the merge contract to GitHub).

Protocol Rules (anti-loop invariants + bounds)

• --max-pdca (default 6) caps per-PR PDCA iterations; diminishing returns -> escalate.
• Worktree discipline always on (skills/worktree-policy/SKILL.md); never commit to main.
• Delegation depth <= 2; Sentinel HIGH auto-blocks (sentinel/config.json authoritative).
• 6-attempt escalation rule (different approach each attempt).
• Exactly ONE STOP marker per turn (the /goal evaluator contract).
• HUMAN_DOMAIN + non-negotiable guardrails (secrets/PII, force-push protected, prod/irreversible, cross-org) ALWAYS halt the loop -> HITL.

DNA Geracional (inherited by every spawned agent)

Transcribed to every spawned agent (liberdade-com-responsabilidade · efeito-borboleta · self-healing). Full text: governance corpus DNA-geracional (do not duplicate here).

Examples

See examples.md.

Validation

• tests/validate-plugin.sh enforces (generically, for all skills): skills/quiesce/ contains SKILL.md with valid frontmatter.
• Skill file kept < 12288 bytes — a Goldilocks guideline mirrored from auto-pilot (the script hard-codes that ceiling only for auto-pilot, not as a quiesce-specific gate).
• commands/quiesce.md carries matching name: quiesce frontmatter.
• Satisfies the 10-item checklist in skills/skill-writer/SKILL.md.

Related

• commands/quiesce.md
• skills/bot-finding-arbiter/SKILL.md
• skills/auto-pilot/SKILL.md
• skills/ooda-loop/SKILL.md
• skills/converge/SKILL.md
• skills/worktree-policy/SKILL.md
• skills/status-map/SKILL.md
• sentinel/config.json + sentinel/detection_rules.md
• CONTRIBUTING.md
• agents/orchestrator.md

Versioning

See CHANGELOG.md (kept out of SKILL.md per progressive disclosure).