Author Agent

---
name: author-agent
description: >
  Runs the GATED AUTHORING loop for load-bearing instruction text in this
  repo: creating a new agent from the template, or overhauling existing
  agents, skills, and reference contracts. Use when the text is reused and
  load-bearing, its defects are silent in production, and the author is an
  interested party — the loop forces an independent gate between authoring
  and applying. Do NOT use for one-off or low-stakes text — edit that
  directly. Do NOT use when nothing is being authored — reviewing existing
  text alone is review-agent.
metadata:
  internal: true
---

0. COST/VALUE LINE. Before any authoring, one sentence to the user: what
   this thread will cost — gate rounds × dispatches, roughly, plus an
   exact count of rework rounds covered (every request for a gate
   ruling after the first consumes one covered round, whatever comes
   back — a refusal included) — and what it buys; name the
   step-8 closing report as the thread's deliverable.
   Proceed on their green light. A new thread opened mid-loop is
   prohibited until the user green-lights its own line, and a rework
   round beyond the covered count is prohibited the same way — scope
   ratchets by decision, never by drift.

1. MODE.
   SCALED: fire when the work alters the MODEL — a new abstraction, a
   doctrine change, an invariant crossing files; skip straight to the
   branches below when the work edits text within the standing model. The
   model flow ADDS to findings-first, it never replaces it: the incumbent
   text still gets its sweep, and the packet may drop only indicted
   lines. The move: author the model as ONE canonical lean markdown file
   — derived from evidence and verified canon; usage is evidence of
   failure modes, never the spec. The published rendered artifact (HTML,
   diagrams) is a VIEW of that file for the user only: render it FROM the
   .md, never the reverse, and never hand it to a dispatched agent — the
   judge and every apply hand read the canonical .md, which packets cite.
   Discuss the inferences with the user and lock their decisions on the
   model. The law: the model is advisor-authored design and is never
   self-judged — it goes to the judge before any packet is authored
   against it; the packet is then authored against the locked model.
   A fired trigger re-opens step 0: building the model artifact is
   prohibited until the user green-lights the model phase's own cost line —
   scope ratchets by decision, never by drift. (Two model gates, 35
   findings total — 26 + 9, one a silent safety defect — all before a
   line of skill text existed.)
   - CREATE (agent): classify the dimension FIRST from the table in
     `.claude/skills/review-agent/references/agent-template.md` — an agent
     mixing dimensions is stillborn, not fixable later. Author the
     description against the checklist's Dimension 2 slots (the checklist
     is the template's sibling, `references/checklist.md`) and the body
     against the matching overlay — for classify-first EXECUTION, laws per
     rules 1-5, examples per 6-8 and 10, artifact tooth per 9; for a
     schema-rendered body, the locked text is copied verbatim from the schema
     and only the rendering is authored; for PERSONA, the
     overlay's element order (charter, ARTIFACT CONTRACT, WRITE BOUNDARY,
     pipeline-position slot). Before gating, run the routing partition
     against every existing description yourself: the newcomer must not
     double-match a neighbor's shape or monopolize a keyword. A new agent
     is unreachable until named: the by-name references that must learn
     it — or that forward to it — (routing tables, plugin manifest version
     bump when the shipped set changes) ride in the packet as their own
     items, ordered so a forwarding reference lands only after the file it
     names exists.
   - CREATE (skill or reference): `review-agent/references/description-standard.md`
     governs any frontmatter description; no template contract exists for
     other non-agent text, and the checklist's own Scope excludes references
     that are not force-loaded — name in the packet exactly which
     dimensions you authored against and which do not apply, so the gate
     rules against a stated contract rather than inferring one. Where
     the efficient-md skill is installed, its residency classes are the
     contract for non-agent text — name in the packet the class each
     item was authored against.
   - OVERHAUL: findings first. Run the review-agent gate (or a scoped
     sweep) on the existing text; author fixes only against accepted
     findings — never freestyle rewrites of text nobody indicted.

2. PACKET. All verbatim texts in one scratchpad file with item IDs
   (P1..Pn). The packet opens with the step-0 line, the user's green
   light, and the running rework-round count — the gate refuses a
   packet missing any of the three or whose step-0 line lacks the
   exact covered count. Each item names its
   target file and either its anchor plus the replacement text (OVERHAUL)
   or "new file — full contents follow" (CREATE), plus
   the finding, template rule, or locked-model section it
   answers. Item headers and markers are packet scaffolding,
   never file content. The packet is the single authority —
   apply dispatches read it, never chat memory.

3. GATE. Dispatch ONE top-tier judge per the review-agent procedure:
   evidence set = target files (existing ones; a CREATE item's evidence is
   the packet contents plus the family it joins), the full agent family,
   the checklist, the agent template, the packet, and any binding
   artifacts. Verdict per item: APPLY / APPLY-WITH-CUTS / REWORK, findings
   with failure scenarios, free token cuts named verbatim. The judge
   rules; it never authors replacement text.

4. REWORK LOOP. Fold the named cuts exactly — no more, no less. Rework
   each REWORK per its finding. Re-gate ONLY the changed items. Use the
   SAME judge instance while it is fresh — back-to-back rounds, transcript
   still hot; once it has idled past cache life or its transcript has
   grown heavy, resuming re-pays the whole transcript uncached, and a
   fresh judge supplied every prior ruling verbatim (the filed rulings
   are compact) is the cheaper equivalent and counts as the same re-gate.
   Either way the earlier precedents bind, the re-gate dispatch restates
   them verbatim, and the judge audits the cut-folding. A judge with no
   precedents supplied is never a re-gate. Repeat until every item clears.
   The author never self-clears an item, whatever the round count.

5. USER APPROVAL. Final verbatim texts on screen, referenced by item ID,
   each item citing the gate verdict that cleared it — an item with no
   verdict behind it is visibly unapproved. Instruction-file targets
   (CLAUDE.md and kin) always show exact before/after — approval of a
   category is not approval of a text.

6. APPLY. Fast-tier dispatches, exact texts from the packet, no two
   concurrent dispatches writing one path, and packet-stated apply
   ordering honored. Every dispatch states verbatim: the working tree's
   uncommitted changes are the deliverable; git checkout, restore, stash,
   and clean are forbidden; DONE-WHENs describe the dispatch's own delta
   ("your edits touch only X"), never tree state.

7. VERIFY.
   (a) placement/fidelity/standard — every OVERHAUL item at its anchor,
       every CREATE item's file existing in full, whitespace-normalized
       text match against the packet, indentation and wrap per the file's
       own (or its family's) convention;
   (b) behavior probes — for any description added or changed,
       fresh-session routing probes (`claude -p`; the live session's
       registry is a stale snapshot): the happy path plus each moved
       boundary in both directions — on a CREATE the moved boundaries are
       the neighbors'. A CREATE also probes the new agent on one
       representative task. If a probe cannot run in this environment,
       the item is recorded UNVERIFIED — never silently skipped;
   (c) fixtures — this repo has no fixture home (evals/ was removed
       2026-09-15); record each keeper finding as named fixture-debt
       instead of pretending.

8. CLOSING REPORT. Per item: the verdict ID that cleared it, its
   placement result, its probe output or UNVERIFIED marker, its fixture
   or fixture-debt entry. An item missing any of the four is not closed,
   and the report says so.