Files

Claude Code AI - Gitea MCP 598a7c326a refactor(skills): retrofit the corpus to the ADR-0020 context contract (#129 )

Retrofits all 39 skills to ADR-0020's description/body context contract, then fixes what six rounds of independent review found in that retrofit — including four ways the hot gate itself failed open.

Closes #99, #107, #108, #110, #111, #114, #115, #120.

## The retrofit (waves 1-5)

| | Start | Now |
|---|---|---|
| Description FAILs (>400 chars) | 26 | **0** |
| Body FAILs (>900 words, body-only) | 9 | **0** |
| Dangling routing targets | 2 | **0** |
| `Kyberforge.CompositionNote` | 10 | **0** |
| Preload tax | 21,005 chars | **~10,500** |

Under the 12,000-char success criterion. Per-wave detail is on #99.

## The review fixes

**The gate failed open four ways, three of them found after the retrofit shipped.** An unrecognised follower token made a dangling target vanish. A skill directory with no `SKILL.md` resolved as a valid target, so a commit could be green locally and red in a fresh clone — three existing fixtures were relying on that, one of which made the install-leak A/B pass vacuously. Then the free-standing `/name` sweep turned out to be gated on the sentence carrying a boundary marker, so route notation in any other sentence was invisible — not an ERROR, not a SUGGESTION, not an INFO — which left the documented "`/name` always blocks" promise false from a second direction. All four fixed and pinned.

**Two checks were silently not running.** `validate-provenance.sh` checks 7-8 were dead across nine skills. Waking them exposed a deeper problem: they assume `Research doc:` names a source index, but 30 of 121 entries point at topic content documents, so every new check-7 INFO was a false positive and check 8 was saved from a false-FAIL flood only by an *unannounced* skip. Checks 7/8 are now scoped to source indexes and every skip announces itself (#121).

**The retrofit's own anti-goal, four times.** ADR-0020 warns that a blunt gate gets satisfied by deleting content rather than relocating it. `diagnose` and `skill-audit` relocated prose and then read it unconditionally; `prototype` and `vale-config` deleted rules outright that survived nowhere. All four addressed.

## Verification

- `bash tests/run-tests.sh --strict` — 24 suites, 0 skipped, 0 failed
- `bash tests/run-bats.sh` — 325 tests, 0 failures
- `pre-commit run --all-files` — 17/17
- `pre-commit run --hook-stage pre-push --all-files` — 16/16, with `apm marketplace check` and `apm pack --check-clean` run against the remote, not skipped
- `scripts/skill-size-check.sh` over all 39 skills — rc 0, 0 ERROR/FAIL, SUGGESTION-only
- Preload tax measured at **10,498 chars**, max description 390 — both inside budget
- Every new test proven non-vacuous by a deliberate mutation of the behaviour it covers

**Per-commit sync, stated accurately:** the ten commits from the latest review round each pass `check-plugin-content-sync` in isolation, verified by checking each out in a detached worktree with a clean between. The earlier gitea window (`dfacf05..bedbd1d`, nine commits) does **not** — its mirror was regenerated in one batch at `bbc7300`. An earlier revision of this description claimed the property held for every commit; it does not, and a bisect through that window lands on a red commit. **Squash-merge** to collapse it, or accept that this range is not bisectable.

## Version bump

Six plugins and the catalog take a **patch**, not a minor. The branch is **89 commits — 40 `fix` / 30 `refactor` / 12 `docs` / 5 `chore` / 2 `test` — zero `feat`, zero `!`, zero `BREAKING CHANGE`** — and adds no skill, agent, command or hook. (Two earlier revisions of this section cited a stale histogram, most recently 78 commits; the figures above are measured at HEAD.) Both rules this repo ships (`forge/references/version-bump.md`, landing in this PR, and `git-commits/references/conventional-commits-spec.md`) make that a patch, and the catalog set is unchanged at 7 entries.

Not settled by that: four published files were removed from the installed tree, three moved, and `caveman` gained `disable-model-invocation`, retiring its old triggers. Under a strict reading those are major-class and currently ship under `refactor:` with no marker. Whether the deployed skill surface is a public contract is written down nowhere — worth deciding, but it outlives this PR.

## Deliberately not in scope

#112 (cherry-pick ownership, now resolved in favour of `git-commits`), #113 (`rtk git` normalisation), #116 (research fan-out), #101 (audit-skill merge), #122 (non-spec skill-root files), #123 (no PRD producer) stay open. #117 is the one worth reading: the contract's remedy is to move prose into `references/`, which is exactly where neither the size gate nor Vale looks — and the blind spot is wider than #117 currently records, since there is no root `.vale.ini` at all, so every ADR, `CONTEXT.md` and `README.md` is unlinted too.

That blind spot let this branch carry two `level: error` `Kyberforge.SentenceOpenerThereIs` violations into `references/` files it created — `provider-adapter-author/references/provider-matrix.md:31` and `agent-audit/references/finding-criteria.md:95`. Both are reworded in `afadaae`, confirmed by routing each file through the audit's own `vale-wrap.sh` (1 error each before, 0 after). Five further occurrences sit in `references/` files already on `main`; those are the pre-existing corpus and stay with #117, which is the real fix.

Also unfixed and not this PR's: `apm install` appends a duplicate `SessionStart` entry to `.claude/settings.json`, so a fresh clone cannot get pre-push green without an edit AGENTS.md warns against. Reproduces identically on `main`.

Co-authored-by: Defame1297 <gitea@rkdr.net>
Reviewed-on: https://git.dev.rkdr.net/Defame1297/holocron/pulls/129
Co-authored-by: Claude Code AI - Gitea MCP <claude@noreply.git.dev.rkdr.net>
Co-committed-by: Claude Code AI - Gitea MCP <claude@noreply.git.dev.rkdr.net>

2026-09-01 13:47:46 +00:00

4.5 KiB

Raw Blame History

source_keys

context7-websites-code-claude

claude-code-subagents-docs

context7-github-en-copilot

github-custom-agents-configuration

Agent Description Quality Reference

Upstream source: Claude Code subagent reference, GitHub Copilot custom-agents configuration. House contract: ADR-0020, the context budget. The house contract is narrower than either platform's schema rather than a reinterpretation of it: where both speak, both must be satisfied.

Why the description is the expensive part

At startup an agent loads only the name and description of every installed skill and agent. The body is never seen until the agent is invoked. The description therefore carries the entire triggering burden and is paid for in every session, whether the agent fires or not.

A second cost is less obvious and is a correctness hazard rather than a token cost: a description that summarises the workflow is a shortcut the caller takes instead of reading the body. A measured failure upstream — a description saying "code review between tasks" — produced one review where the body's flowchart specified two.

Step 0 — establish which contract applies

Read the frontmatter before judging a single word.

disable-model-invocation: true or user-invocable: false — the agent is hand-invoked. Its description is never matched against user intent, so it is not a routing string. It carries one plain human-facing sentence stating what the agent does. Audit it for that and nothing else. Reporting a missing trigger clause, a missing boundary clause or absent indirect triggers on a hand-invoked agent is a wrong finding, not a strict one. Both fields are Copilot-only and neither is on the vendor-neutral APM allowlist, so this case arises in a Copilot .agent.md at project/user scope and nowhere else. Its Claude Code counterpart has no equivalent field and stays model-invoked, so the two halves of the pair carrying differently shaped descriptions is expected there rather than a pair-consistency finding.
No such flag — the agent is model-invoked and the rest of this file applies.

The three-part shape

A model-invoked description carries exactly three things:

Trigger clause. When to invoke, phrased imperatively: Use when .... Not This agent ... — the caller is deciding whether to act, not reading a catalogue entry.
At most one capability clause. What it does, in one clause. Never an enumeration.
Boundary clause. Compressed form: Not <thing> -> <skill-name>. The target must resolve to a real skill directory or agent file in the authoring source.

Everything else belongs in the body or in the plugin's README.md.

Indirect triggers — conditional, never blanket

Add "even if the user doesn't say X" only where the user's natural phrasing genuinely omits the domain word. True for the gitea-* family: people say "create an issue", not "create a Gitea issue". False for git-commits: nobody asks for a commit without saying commit. A blanket indirect-trigger clause on an agent whose domain word is unavoidable is padding charged to every session.

Near-miss exclusions

Add a boundary clause only where a sibling skill or agent could plausibly steal the activation. Use strong near-misses — queries that share keywords but need something different — not weak ones. One boundary clause per genuine near-miss; a list of four is enumeration wearing a boundary's clothes.

Before / after

# FAIL — a noun-phrase opener rather than a trigger, capability enumeration in
# place of one capability clause, and no boundary clause at all, preloaded into
# every session forever. (The live git-orchestrate description, 254 chars.)
description: Orchestrates git workflow operations for other agents. Invoke when a
  caller needs a multi-step or destructive git operation (rebase, force-push, branch
  deletion) coordinated across domain skills with safety gates, session context, and
  structured results.

# PASS — trigger, one capability clause, boundary. The operation list and the
# safety-gate mechanics are the body's job; the router cannot act on them.
description: >
  Use when an agent caller needs a multi-step or destructive git operation
  dispatched and safety-gated. Not conversational git help -> git-workflow.

Where the criteria live

Every FAIL and SUGGESTION criterion for this dimension is in references/finding-criteria.md, which Step 3 reads on every run. This file is the reasoning behind them, loaded only when that file puts the description dimension in play.

4.5 KiB Raw Blame History