08-tooling/brigades/skill-agent-brigade

menu

Menu — skill-agent-brigade

What this brigade can do for you, and what your ticket must carry so the kitchen can cook from it. Envelope shape (identity, status, context-manifest, Gate A) is universal — see TICKET-CONTRACT.md; this menu covers only what is specific to ordering from this brigade.

artifact: skill — build a Claude skill

Status: live

What you get: a trigger-tuned, procedure-first SKILL.md (+ progressive-disclosure references), passed through 5 LLM critic axes + deterministic skill-lint, with an optional execution-eval (measured lift over the base model, per-fixture, per-tier).

What the ticket must carry, by type_hint:

type_hint mandatory payload sources why
computational / corpus worked examples with known answers the test station's oracle source — fixtures are graded deterministically against them
generative / advisory exemplars of acceptable output, provenance cited the only honest quality anchor for subjective artifacts
all types the core competency/knowledge source (when: "always…") the depth the spec station translates from knowsdoes

What the Order should specify: the ONE skill (scope-fenced against neighbors — "variance analysis, NOT budgeting"), who it's for, and what done looks like. Optional but sharpening: menu: pairing ref, known near-miss asks the trigger must NOT fire on.

artifact: brigade — assemble a domain brigade

Status: live

What you get: a roster of stations (each an ordered artifact: skill build), a pass policy (expo decision config: exit criteria, round budgets), and a rail binding — a runnable brigade. (Exercised live twice on 2026-07-02: the assessment-agent-brigade and sales-collateral-brigade roster tickets both rode spec→tests→expo convergence to advance.)

What the ticket must carry: the domain's station roster intent in the Order (which phases, which handoffs), one context source per station-to-be (same per-type payload rules as above, applied per station), and the pass-policy constraints (budgets, gates) if they differ from house defaults.

artifact: add-station — add a station to an existing brigade

Status: live

What you get: a new station built to the full skill rigor (spec → tests → author → critic, incl. deterministic lint) AND wired in: the target brigade's roster updated, its menu re-published with a version bump, its station registry mapping the new artifact kind. A station IS a skill (the russian-doll definition) — this type exists so the WIRING is part of the order, not an afterthought.

What the ticket must carry: the target brigade (its menu ref), the station's competency/source material (same per-type payload rules as artifact: skill), and the wiring intent in the Order (which artifact kind(s) it serves, where outputs land). Precedent: the sec-filings station (2026-07-02) ran this shape manually — built, then fresh-eyes critic gates retroactively (verdict ITERATE, 5 hardenings applied before acceptance); this artifact type front-loads those same gates.

artifact: iterate-skill — refine an existing skill

Status: live

What you get: the skill re-authored against the refinement Order, re-gated by the critics, and eval-gated against the existing baseline: the execution-eval station re-runs the oracle fixtures two-arm (current skill vs refined skill, per-fixture, per-tier) — the refinement only advances if it improves the targeted axis WITHOUT regressing any other fixture. No eval pass, no ship. (Machinery exists and is demonstrated: the 2026-06-29 variance-analysis and generate-tests eval runs, per-fixture expo.)

What the ticket must carry: the existing skill (cellar/plugin ref), the refinement intent (what's underperforming, with evidence), and any NEW oracle fixtures the refinement targets (a refinement without a measurable target is a smell — the steward should push back).

artifact: iterate-brigade — refine an existing brigade

Status: planned

What you get (when live): a targeted refinement of a brigade — a single station's behavior, or the expo's routing/decision policy between stations. Policy changes are the highest-authority edits a brigade can receive, so the acceptance bar is REPLAY: the brigade's closed tickets are its build records, and a routing/policy change must be replayed against recorded tickets to show the new policy reaches equal-or-better exits before it lands.

Why planned, honestly: the replay-based eval for routing changes is designed-in-principle (ticket-as-build-record makes it possible) but not built — and per the house rule that iterations MUST eval against baseline, this type stays off the live menu until that eval exists. Station-scoped iterations that reduce to iterate-skill + re-wire can ride the live types today.

artifact: menu — publish this menu

Status: live

What you get: this document, regenerated — the expo introspects stations, critic axes, and eval config, bumps version, publishes to <cellar>/brigades/skill-agent-brigade/menu.md.

What the ticket must carry: one pointer at this brigade's home (plugin dir). Order = "what can your brigade do?"

House quality gates (apply to everything ordered)