servo

Run a headless agent loop against a scaffolded target. Subprocesses `claude -p --agent <runner|judge>` against the target N iterations under hard guardrails (iteration cap, cumulative cost ceiling, context-fill refusal, oracle-score plateau detection, SIGINT/SIGTERM cleanup), uses `gate.py` as the truth-source after each iteration, and checkpoints state to `<target>/.servo/runs/<run-id>/state.json` so an interrupted or capped run can be resumed.

Decide whether a goal's SCOPE and INITIAL PROMPT are precise and bounded enough to hand to an unattended, long-horizon agent loop — *before* any budget is burned. Emits the ADR-0029 **readiness verdict**: a closed three-state gate (`ready` / `needs_tightening` / `unsafe_for_autonomy`), fail-closed, written to `<target>/.servo/readiness/<goal-id>.json`. Human-owned — the artifact starts `approval_status: proposed` and is never auto-approved; a human reviews the scorecard and flips it to `approved`. Readiness sits **upstream of** `/servo:edd-suitability`: suitability asks "can this spec be evaluated?", this asks the prior question "is the premise safe to run unattended at all?".

Author a frozen text content-fidelity eval component for a project — the non-visual sibling of servo's design-eval. Use when generated text (a reply, a doc, a templated message) must match an intended rubric or spec ("does this text match the intended tone/content/style?") and that fidelity should drive or gate the loop. Gathers the generated text under test (file read or generator command), judges fidelity with a pinned text-capable model (n-sampled, lower-bound), freezes the definition, and installs a `score_content_fidelity` component into the project's `oracle.sh`. Do not use for deterministic checks (use scaffold-init / spec-oracle), for visual/UI fidelity (use design-eval), or for the per-iteration judge agent.

Author a frozen UI design-fidelity eval component for a project — the non-deterministic sibling of servo's deterministic component templates. Use when a project must be built to match a design mockup ("does the UI match the design?") and that fidelity should drive or gate the loop. Captures app vs reference screenshots, judges fidelity with a pinned vision model (n-sampled, lower-bound), freezes the definition, and installs a `score_design_fidelity` component into the project's `oracle.sh`. Do not use for deterministic checks (use scaffold-init / spec-oracle) or for the per-iteration judge agent.

Decide whether an engineering spec is suitable for Evaluation-Driven Development before any unattended loop runs against it. Emits the ADR-0015 **suitability verdict** — a closed three-state gate (`suitable` / `needs_evidence` / `unsuitable`), fail-closed, with an actionable `missing_evidence` checklist — written to `<target>/.servo/suitability/<spec-id>.json`. A gate, not a score: the loop never optimizes toward a meaningless green oracle on work that cannot be evaluated.

Author a frozen text-judged eval from a spec-006 `residual_judgment` AC or a free-form goal. Guides triage, rubric shaping, reference-set collection, frozen n/δ/threshold/judge parameters, ADR-0024 freeze/install, and advisory judge audit. Use to author an eval, turn a goal into acceptance criteria, shape a rubric or reference set, freeze/install an eval, or audit judge trust. Do NOT use to run evals (`quality-gate`/`agent-loop`), classify deterministic checks (`spec-oracle`), decide EDD suitability (`edd-suitability`), or build design/content fidelity evals (use the `design-eval` / `content-fidelity` presets). This is the generic surface for other text-judged criteria.

Assemble the [ADR-0016](../../docs/decisions/adr-0016-execution-plan-artifact.md) **execution plan** — the durable Compile→Run handoff artifact — from the Compile inputs already produced upstream (the suitability verdict, the oracle, the spec-oracle overlay) and write it to `<target>/.servo/plans/<spec-id>/plan.json`. This is the **last step of Servo Compile**: it decides nothing new, classifies nothing, and adds no gate — it only references what Compile already produced.

Run heartbeat or set up a scheduled heartbeat for a servo-scaffolded target. The heartbeat is servo's Routine-triggered front-end: discover project signals into a triage inbox, read the heartbeat inbox, dispatch heartbeat findings, or run heartbeat end to end under one whole-heartbeat cost ceiling. Fire this skill for requests mentioning a Routine, cron, GitHub Actions schedule, or a scheduled agent that should invoke servo on an interval.

Install, uninstall, or report the status of servo's meta-judge `Stop` hook in a target project. The installed hook scores every assistant turn against the scaffolded oracle (via `gate.py`) and blocks the stop with a structured retry hint when the oracle is below threshold — the interactive cousin of `/servo:agent-loop`. Installing it mutates `<target>/.claude/settings.json`, so it is a **Tier-2** surface: offered explicitly, never auto-installed.

Run the scaffolded `oracle.sh` against a target project and surface a closed-contract result (0 = pass / 1 = below threshold / 2 = env error). Also exposes an `audit` subcommand that prints the servo install manifest without invoking the oracle.

Compile a target project's signals into an executable oracle — the first Servo Compile step of servo's Evaluation-Driven Development engine. Probes the target for signals (tests, lint, CI, language) and drops a tailored `oracle.sh`, `.servo/install.json` manifest, and `.servo/refinement-todo.md` deferred-decisions list.

Compile a spec or slice into a reviewable, project-owned **evidence overlay**: map each acceptance criterion to a deterministic check family, write a `plan.md` + `checks.json` under the spec's own `oracle/<id>/` directory, compile the checks into an installable `oracle.sh` component, and freeze it behind an explicit approval so an unattended loop can optimize against the spec's promises instead of a hand-waved single judge.