Description 709 -> 161 chars, body 683 -> 353 words, Gotchas 8 entries/56% of
body -> 2/24.9%. Clears the description FAIL and both Vale CompositionNote
errors.
Fixes the three stale claims recorded on issue #99, all re-verified against the
deployed gitea-mcp schema during review:
- The description no longer advertises 'reviewers' as an update capability.
editPullRequestFn never reads reviewers or team_reviewers; only
add_reviewers/remove_reviewers do.
- milestone is now marked honoured on "update" only, in the Gotcha, the body
and the dispatch table's create row. On "create" the server discards it and
omits the key from the response, so the drop is indistinguishable from never
passing it -- and labels DOES apply on create, so labels landing is no
evidence the milestone did. The old text told callers to resolve a milestone
before any write, wasting the lookup on create.
- The superseded un-draft workaround is gone. "update" with draft:false and no
title makes the server strip the prefix itself, including [WIP],
case-insensitively -- carried by references/pull-requests.md, corrected in
PR #106.
The 22-row tool/method table becomes a 5-row dispatch table; all 20 operations
it named remain reachable, including update_branch and the reviewer methods.
Refs #99
Description 787 -> 263 chars, body 922 -> 302 words, Gotchas 9 entries/69% of
body -> 3/24.8%. Clears both size FAILs and both Gotchas suggestions.
Deleted the second trigger register outright -- ~300 chars re-quoting the same
six verbs as user phrasings, which ADR-0020 names this skill for specifically.
Split the body on the read/write boundary: one invocation cannot both read and
write, so a dispatch table is mandatory. Six operations collapse into two flow
files rather than six -- create/update/delete share one tool pair and one
SHA-first lifecycle whose preamble would otherwise be triplicated, and the three
read tools share ref selection plus a 'neither listing is a SHA source'
comparison that only exists between them.
references/examples.md removed; all eight of its content blocks and all nine
named parameters carry into references/writing.md, verified against git HEAD.
Two deltas are corrections: the repo tree is now ruled out as a SHA source, and
reusing a SHA captured earlier in the conversation is now forbidden.
Four Gotchas relocated to the flow file that needs them; two promoted to gates
(SHA-as-concurrency-token opens writing.md; owner/repo became ## Inputs).
Refs #99
Description 688 -> 282 chars, body 435 -> 293 words, Gotchas 5 entries/54% of
body -> 3/23.5%. Clears the description FAIL and the Gotchas suggestion.
Cut the second trigger register (six re-quoted user phrasings) and the capability
enumeration; both moved to a new Boundaries section in the skill's own README.
Kept all three boundary clauses -- git-branches, git-history and gitea-prs each
defend a real activation steal, and gitea-prs is now the only guard on the
branch/PR collision in either direction since gitea-prs's own retrofit narrowed
its boundary to issues. Written as one arrow per target: the resolver extracts
only the first name per arrow clause, so conjoined targets go unchecked.
Two Gotchas deleted -- one paraphrased the step below it (its non-obvious half,
the get_me/list_my_repos token-scope block, was folded into that step), the other
is carried in full by references/branches.md:40-52.
Repoint two reference pointers the rename broke: branches.md and commits.md named
Gotchas by titles this retrofit changed. Now named by stable descriptors.
Refs #99
Description 500 -> 211 chars, body 676 -> 462 words, Gotchas 5 entries/41% of
body -> 3/21%. Clears the description FAIL and both Gotchas suggestions.
Cut the second trigger register (the re-quoted user phrasings), the doubled
capability enumeration across releases and tags, and the gitea-issues/gitea-prs
boundary, which defended against nothing -- neither was going to win a release
request. Kept the indirect trigger and the one real near-miss, gitea-branches.
Releases and tags are one flow, not two: 'delete the release and its tag' is a
single invocation that takes both branches, which disqualifies mutual
exclusivity, so no dispatch split. Two Gotchas moved to references/ behind
explicit triggers; one deleted as a paraphrase of the step below it.
Refs #99
The Hand-invoked skill entry covered listing visibility and preload tax but not
invocability. Measured: disable-model-invocation: true hard-blocks the Skill
tool, not just the model-visible listing, so a `Call `x`` step in another
skill's body stops working the moment x takes the flag.
ADR-0020's 'Invocation as a design axis' verified only the visibility and slash
directions. Without this half, marking grill-with-docs hand-invoked -- one of
issue #99's four named candidates -- would silently break forge Step 1, which
calls it, plus three other inbound routes.
Refs #99
The SessionStart hook found six packages behind the remote default branch and
ran apm update. Committing the resulting lock separately so it does not sit in
the ADR-0020 retrofit diff.
Follow-up to the review of #106, which found four parameter claims the
first pass left wrong or missing. All four verified against gitea-mcp
v1.6.0 source before changing anything.
`remove_deadline` is dropped on "create" like the others, but the new
Gotcha enumerated a closed list that omitted it — so the paragraph
contradicted its own opening sentence. The list is now open ("including")
and the parameter carries the same "update" only marker as its neighbours.
`base` is settable on "update": editPullRequestFn reads it and retargets
the PR onto a different base branch. The file documented it as required
for "create" and said nothing else, hiding a real capability. This one is
under-claiming rather than over-claiming, but a file whose subject is
which parameter applies to which method is the place to fix it.
applyDraftPrefix strips two prefixes, `WIP:` and `[WIP]`, matched with
strings.EqualFold, while only "WIP: " is ever added. The un-draft
correction said "the prefix" singular, which understates what a caller
can hand it.
README.md bundled `reviewers` and `milestone` into "creating and
updating". Capability prose rather than a parameter contract, so not
strictly false, but it is the same conflation that produced the original
bug and it is not behind any gate.
Not addressed here: SKILL.md's description still advertises updating
reviewers, its draft guidance still prescribes the manual workaround, and
its milestone-resolution instruction carries no method qualifier — so an
agent working from SKILL.md alone can still reach the failure mode. Those
edits trip skill-size-check and the Vale prefilter, pulling in the
ADR-0020 retrofit; tracked on #99 rather than done silently here.
Refs: #104
Refs: #99
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ETxbGFetkbJQBHsx442Brt
The gitea-prs reference documented `milestone` as "settable on both
"create" and "update"". It is not: `pull_request_write method: "create"`
accepts the parameter, returns no error or warning, and discards it.
Verified against the gitea-mcp v1.6.0 source rather than by observation
alone. `createPullRequestFn` builds its `CreatePullRequestOption` from
owner, repo, title, body, head, base, draft, labels and deadline only —
so the drop is not limited to `milestone` as issue #104 supposed.
`assignee`, `assignees`, `reviewers` and `team_reviewers` are discarded
on create too, and `reviewers`/`team_reviewers` are discarded on
"update" as well; they are only ever read by "add_reviewers" and
"remove_reviewers".
Two properties made the original error easy to make and hard to catch,
so both are now written down next to the correction: `labels` sits
beside `milestone`, reads identically, and does apply on create; and
`issue_write method: "create"` honours `assignees` and `milestone`, so
the asymmetry is specific to pull requests.
While in the file, corrected the adjacent draft gotcha. It prescribed
reconstructing the un-prefixed title by hand to un-draft a PR;
`applyDraftPrefix` shows "update" with `draft: false` and no `title`
fetches the stored title and strips the prefix server-side.
Impact: documentation only, no behaviour change. Callers following the
old text silently created PRs with no milestone, assignee or reviewer.
Confined to references/pull-requests.md, so the ADR-0020 skill gates do
not apply and gitea-prs needs no #99 retrofit first.
Fixes: #104
Refs: #99
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ETxbGFetkbJQBHsx442Brt
Why: two blind verifiers re-ran the five preceding commits and found four
defects of the same class this branch exists to close -- a confidently stated
measured claim that does not survive re-measurement -- this time inside the
fixes themselves.
- AGENTS.md:41 still carried both phrasings c68e864 reports having corrected.
`grep -rn repo-defined` returned exactly one hit repo-wide: that line, in the
file every session preloads. 4d336bb edited the line directly above it.
- ADR-0021 asserted twice, in the section justifying that no gate is added,
that the ADR-0020 validators "never open an apm.yml". All three open and
yaml.safe_load it (skill-size-check.sh:342, both validate.sh). The conclusion
survives -- none reads the description: key, and their globs are SKILL.md and
*.agent.md only -- but the stated mechanism is falsified by one grep.
- architecture.md said the ADR directory holds 20 numbered ADRs; c7ba3d2 made
it 21, and c68e864 audited that file for exactly this class of stale count.
The number is dropped rather than corrected: `ls docs/adr/` is already the
index, so a count in prose is a second thing to maintain.
- gates.md's new three-verdict table said `-> name` promotes an unresolved
target to ERROR. Reproduced with fixtures: NAME_HYPH (skill-size-check.sh:543)
requires a hyphen, so `-> gitea-prs` is checked and `-> triage` is not
extracted at all, and the unicode arrow is never recognised. The SUGGESTION
text advises that spelling, so taking its advice can silence the finding. The
gap is now documented as a defect; nothing covers it, since the one arrow case
in test-adr0020-targets.sh happens to use a hyphenated target.
Implementation notes:
- AGENTS.md:48's coverage claim is shrunk rather than chased. Restoring six
glossary entries did not make it true: 12 more sampled terms are undefined,
three of them (trigger/capability/boundary clause) used inside CONTEXT.md
itself. It now says CONTEXT.md is the glossary and is not exhaustive.
- CONTEXT.md's output profile and near-miss entries are corrected against their
sources. The first stated a false exclusion -- .github/plugin/plugin.json IS
apm-generated; only the marketplace mirror has no profile. The second
inverted its source's referent: description-quality.md defines a near-miss as
a query, not a sibling skill.
- The strict-mode message named jq, which no suite guards on (`command -v jq`
appears nowhere in tests/), while omitting python3/PyYAML, which three do.
- README's git and gitea bullets now name git-workflow and gitea-workflow.
ADR-0021 leaves README the only inventory and architecture.md now points at
it, so the two bullets that were short had to be completed.
- ADR-0018's 2026-08-14 correction is marked superseded in place. It asserted
machine state in the present tense that its own 2026-08-17 note retracts.
- ADR-0021's remaining errors: six files -> four (measured from de84d1b), the
wiki description's length 114 -> 96 chars, the codex self-contradiction, the
cost argument overstating bumps already owed for any skill addition, and two
claims about files this branch went on to edit.
- The "15 of 17 suites" figure is restored where I had removed it: it is a dated
record of one incident, not a live count, and four sites now describe it the
same way.
Impact: 16/16 pre-push hooks pass, suite 24 passed 0 skipped 0 failed. No
behaviour change; every edit is prose or a comment.
Refs: #105
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TmFqzpuExLJv3m114XVE9w
Why: the relocation into gates.md and architecture.md moved text faster than it
verified it, and a review found eight assertions that do not survive contact
with the repo.
- gates.md cited docs/research/examples/, which does not exist; the tree is at
plugins/kyberforge/docs/research/examples/. The same wrong string survives at
ADR-0013:98, where it originated, and is tracked separately.
- gates.md's description-budget arithmetic was built on 23,427, a figure
ADR-0020 measured at f9b919d and pins there, and compared a name-inclusive
total against a description-only ceiling. Dropped rather than re-measured,
following the precedent this branch set for CONTEXT.md's token figure: a
second live copy is a second thing to go stale.
- gates.md listed "every boundary-clause routing target must resolve" as a hard
FAIL. skill-size-check.sh emits unresolved targets as SUGGESTION by default,
promotes to ERROR only when terminal and either route-notated or corroborated
by a resolving target in the same sentence, and reports INFO DID NOT RUN when
no universe resolves. Verified with five fixtures; replaced with a
three-verdict table.
- gates.md and README called `pre-commit run --hook-stage pre-push --all-files`
the whole push gate. check-release-needed.sh exits 0 unless
PRE_COMMIT_REMOTE_BRANCH is refs/heads/main, which only the real git hook
exports -- its own header comment titles this a "Known gap".
- "repo-defined" was loose: the two meta hooks are also declared here, and what
separates them is repo: meta from repo: local.
- architecture.md claimed `when:` appears in two of 39 skill sources. Exactly
one SKILL.md carries it; the second hit is a META.md. The likely origin is
skill-implementation-workflow.md:233, which claims META.md sits alongside
every skill when one exists in the whole tree.
- architecture.md said core/AGENTS.md is "deployed beside" CLAUDE.md.
deploy-manifest.sh sends them to ~/.claude/ and ~/.agents/ respectively; the
relation is import, not adjacency.
- gates.md's see-also cited ADRs as bare stems where its neighbours use paths.
Implementation notes: architecture.md's closing pointer sent readers to the
published description "for what a consumer actually gets", which was true
against an enumeration and is not after ADR-0021. It now names the boundary
relationship and points at the skills tree and README for an inventory. README's
bin bullet gains caveman and zoom-out, since ADR-0021 leaves it the only place
an inventory lives.
Impact: 16/16 pre-push hooks pass. Deliberately not fixed here and tracked
separately: the 12 pre-commit-stage hooks gates.md does not document, its own
SentenceOpenerThereIs violation, and README's missing scripts/install.sh step.
Refs: #105
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TmFqzpuExLJv3m114XVE9w
Why: enumerating a plugin's skills in its published description has now failed
three times in four days, the third time inside the correction for the second.
bb9158d (Aug 14) fixed core's description, which described bin's skills. That
failure was wrong content, not an incomplete list -- a syntactically perfect
four-item enumeration that belonged to a different plugin -- so enumerating
harder could not have caught it, and bb9158d's own fix replaced the enumeration
with a domain boundary. 65bac15 and de84d1b then cited that precedent while
doing the opposite, and the re-enumeration is itself incomplete: caveman and
zoom-out are unnamed in bin, git-workflow is unnamed in git despite 65bac15's
message claiming it was added, and gitea-workflow is unnamed in gitea. 23 of 27
skills named at the third attempt.
Nothing checks any of this. check-manifests.sh does not contain the string
"description"; apm audit --ci, apm pack --check-clean and
sync-plugin-content.sh --check all compare compiled output against apm.yml, so
their job is to propagate whatever it says byte-for-byte. The wiki claim passed
all fourteen pre-push hooks every day it was published. Enumeration also makes a
marketplace release the consequence of a directory appearing.
Implementation notes:
- ADR-0021 records the decision, the three failures, and why no gate is added:
the check is one-directional, cannot see an invented capability, and a
correspondence map would relocate the same per-skill maintenance one file over.
- bin 1.1.4 -> 1.1.5, git 1.3.4 -> 1.3.5, gitea 1.3.5 -> 1.3.6,
marketplace.version 0.4.4 -> 0.4.5, all PATCH under per_package.
- apm.yml's top-level version: is restored to lockstep at 0.4.5. It has moved
with marketplace.version in all eleven commits that ever touched the file
until this branch left it at 0.4.2, and apm.yml's own comment records it as
live and consumed by apm audit.
- git's boundary names the pre-commit hooks explicitly: pc-author and pc-run are
not git operations, and a boundary drawn narrower than the contents is the
same defect one level up.
- Regenerated with apm pack for the marketplace manifests and
sync-plugin-content.sh --all for the per-plugin ones; both mirrors verified
byte-identical.
Impact: consumer-facing. 16/16 pre-push hooks pass, suite green, generated diff
is 26 insertions and 26 deletions across 8 files with no incidental churn.
Refs: #105
ADR: 0021
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TmFqzpuExLJv3m114XVE9w
Why: four defects in the files every session pays for, all introduced or left
behind by the trim.
AGENTS.md told agents the `<plugin>:` form still resolves "because user-scope
native installs were left enabled on purpose", and that a working namespaced
call "is not something to fix". That premise is false on this machine:
installed_plugins.json is empty, no enabledPlugins key exists in ~/.claude.json,
and ~/.apm/marketplaces.json is empty. ADR-0018 already reversed itself once on
this exact claim (Correction 2026-08-14) using that same enablement as its
evidence, so flipping the assertion again would be the third revision in three.
Both files now assert nothing about install state at all, which removes the
flip-flop surface instead of re-aiming it.
The other three are guard-rails whose instruction survived the trim while the
caveat that made it safe did not:
- The run-tests.sh line omitted --strict, so it named the one invocation that
reports SKIPPED rather than failed when a dependency is missing. gates.md
records this gate going green having verified 15 of 17 suites on a vale-less
PATH. .pre-commit-config.yaml:70 already uses --strict for that reason.
- The .claude/settings.json prohibition lost its ADR-0019 exception, so an agent
applying it literally would strip apm's own merged SessionStart entry and
create the drift the rule exists to prevent.
- LESSONS.md still routed graduated rules to CONTEXT.md's Principles section,
which this branch deleted.
Implementation notes: the six terms the trim dropped while AGENTS.md still
claimed CONTEXT.md glosses everything -- authoring root, content mirror, apm
package, output profile, near-miss, vacuous green -- are restored as one-line
entries per CONTEXT-FORMAT.md, sourced from architecture.md, gates.md and
skill-audit's description-quality.md rather than reworded. ADR-0018 gets a third
dated note recording the observation and the fact that the state has now been
described two ways, and its stale user-scope inventory is replaced by a pointer
to it; the decision it records is untouched. LESSONS.md:3 carried the identical
stale claim as :5 and is fixed with it.
Impact: preloaded context is now free of assertions about machine state.
Refs: #105
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TmFqzpuExLJv3m114XVE9w
Why: this branch moved the prerequisites list out of AGENTS.md into README.md
but left three references behind. The worst is run-tests.sh's --strict failure
message, which a developer is handed at the exact moment a push gate fails and
they need the dependency list: it named AGENTS.md, which no longer documents
vale, apm or jq anywhere. 56cc173 re-anchored this same class of citation in
.pre-commit-config.yaml, scripts/check-vale-style-sync.sh,
scripts/sync-marketplace-mirror.sh and tests/test-check-vale-style-sync.sh --
tests/run-tests.sh was missed, and it is the only one a human actually reads.
Implementation notes:
- run-tests.sh:262 (user-visible), run-tests.sh:16 (header rationale) and
test-run-tests.sh:395 all now cite README.md's Prerequisites table, verified
to name vale, apm, jq, python3/PyYAML and the claude CLI.
- Every assertion on this message greps for "a skip is a SETUP ERROR", which
sits before the changed clause, so no test needed updating.
- test-vale-wrap.sh's bash 3.2 citation is corrected in the same pass: it
credited LESSONS.md and the script headers, but LESSONS.md's only mention of
3.2 is an aside inside an unrelated lesson, not a target declaration. It now
cites scripts/lib/batch-run.sh and providers/claude-code/statusline-command.sh,
which do declare it. The citation 56cc173 replaced was already false, so this
finishes a half-fix rather than correcting a regression.
Impact: forced the strict-failure path with vale removed from PATH and confirmed
the corrected message printing -- 21 passed, 3 skipped, exit 1. Full suite with
vale present: 24 passed, 0 skipped, 0 failed.
Refs: #105
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TmFqzpuExLJv3m114XVE9w
Why: bd2bf66 restored the `apm pack` guard-rail by appending it to the `Fix:`
command after `--`, which made the printed line stop being runnable. Pasting it
ran the script with ~24 stray argv entries: `${1:-}` became `--`, so CHECK
stayed 0, no shift occurred, and `[[ $# -eq 0 ]] || usage` printed usage and
exited 1. The user got a usage error from the tool meant to fix their problem,
and the mirror stayed stale.
The unquoted backticks around `apm pack` were a second hazard in the same line:
the paste command-substituted a real `apm pack` run before this script was ever
reached, so the first error a user saw came from apm, not from here.
Implementation notes:
- The runnable command now stands alone on its own line, and the rationale
follows as a separate `Note:` echo.
- Backticks downgraded to single quotes; a line printed next to a
copy-pasteable command must not contain shell metacharacters.
- The guard-rail text is otherwise preserved verbatim. It exists because apm
ships no output profile targeting this path, so `apm pack` does not refresh
it, and expecting it to is the drift this hook prevents.
Impact: reproduced the break on a scratch copy, then verified the fix by pasting
the printed command verbatim — exit 0, mirror synced, re-check clean.
tests/test-sync-marketplace-mirror.sh asserts only exit codes and file contents,
so nothing pins this message and it could regress silently; tracked separately.
Refs: #105
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TmFqzpuExLJv3m114XVE9w
Why: bin's marketplace description read "A place for things to be binned" — a
routing boundary sitting in the enumeration slot, phrased as a joke. It is what
a consumer sees in the marketplace listing for a plugin shipping eleven
substantive skills: research, write-docs, tdd, prototype, triage, diagnose,
improve-codebase-architecture, grill-me, grill-with-docs, zoom-out and caveman.
This is the same category error b0ef503 fixed for git and gitea, and the same
text bb9158d moved out of core's description in August — "triage, diagnosis,
architecture review" left core precisely because those skills live in bin, so
bin is where it should have landed.
Implementation notes: bin 1.1.3 -> 1.1.4, marketplace 0.4.3 -> 0.4.4, PATCH for
metadata under the per_package strategy. README's bin line mirrors the new
description. architecture.md's routing boundary is deliberately left as
"unsorted skills that have not earned a home yet" — that is still true, and it
answers a different question than the listing does.
Impact: consumer-facing. apm-pack-check-clean, check-plugin-content-sync,
check-marketplace-mirror-sync, validate-marketplace, validate-plugins and
apm-audit-ci all pass.
Why: both plugins' published descriptions named capabilities that do not exist
and omitted ones that do. `gitea` advertised wikis — no wiki skill exists and
`grep -ri wiki plugins/gitea/.apm/` is empty — while omitting gitea-branches and
gitea-files. `git` advertised pull requests, which are gitea's, while omitting
pc-author/pc-run, git-submodules and git-workflow. These are compiled into
plugin.json and both marketplace.json files, so they are published metadata, not
internal notes.
This is the same defect bb9158d fixed on `core` in August, so the fix follows
that precedent exactly: correct the package apm.yml and the root marketplace
entry together — fixing only the package manifest leaves it half-propagated,
since the root entry is what reaches the compiled marketplace.
Implementation notes:
- git 1.3.3 -> 1.3.4, gitea 1.3.4 -> 1.3.5, marketplace 0.4.2 -> 0.4.3, all
PATCH for metadata under the per_package strategy at apm.yml:76.
- Regenerated with `apm pack` for the two marketplace manifests and
`scripts/sync-plugin-content.sh --all` for the four per-plugin plugin.json
files; `apm pack` alone does not touch those. `scripts/sync-marketplace-mirror.sh`
refreshed the Copilot legacy mirror, which no apm profile targets.
- README's gitea line mirrors the new published description. The routing
boundaries in architecture.md deliberately do not, since they answer a
different question.
Impact: consumer-facing. Anyone reading the marketplace now sees what these
plugins ship. apm-pack-check-clean, check-plugin-content-sync,
check-marketplace-mirror-sync, validate-marketplace and apm-audit-ci all pass.
Why: the review found architecture.md's plugin table and README's plugin list
were both written by enumerating what happened to be in each plugin, so both
went stale immediately — README credited `git` with pull requests it has no
skill for, and both omitted `pc-author`/`pc-run`. bb9158d fixed this same class
on `core` in August and it recurred here, so the fix is to stop enumerating in
the place whose job is routing.
Implementation notes:
- architecture.md's table now states routing boundaries only, with a note saying
so and pointing at each plugin's published apm.yml description for the actual
inventory. Two boundaries are named explicitly: `core` vs `kyberforge` (kept),
and `git` vs `gitea` — wire protocol against a local clone versus the forge's
HTTP API, which is why git-branches and gitea-branches are not duplicates.
- The `git` row is widened to cover git hook tooling rather than moving
pc-author/pc-run elsewhere. pre-commit manages .git/hooks/, so the placement
was always right and the row was simply under-described.
- README's setup block had a comment where step 2's command should be. It now
carries `pre-commit install -t pre-commit -t commit-msg -t pre-push`, verified
against the three stages in .pre-commit-config.yaml and the three hooks in
.git/hooks/. pc-run is still named, as the option rather than the only path.
- README:11 said each plugin carries agents, hooks and MCP servers. Only
kyberforge ships hooks and only bin ships an MCP server.
- gates.md:30 claimed its hook list was in config order; it is grouped by
concern, which is the better layout, so the claim is corrected to match.
Impact: docs-only. gitea's published description is corrected separately, since
that regenerates consumer-facing manifests.
Why: 1929ffd and b9c7762 preserved every constant, hook ID and error string,
but two don't-re-litigate notes were lost. Both were guard-rails — they fired
from preloaded context, before the mistake. Restoring the sentence to an
on-demand spec doc restores the text without restoring the function, since an
on-demand file is opened after a plan is formed, not before.
Implementation notes: placed by when each note must fire, not by what it is
about.
- "Do not expect `apm pack` to refresh the .github/plugin/marketplace.json
mirror" now lives in the DRIFT failure message of the script and hook that
already catch that exact mistake. It costs no preloaded tokens, fires at the
moment of the error, and cannot be missed. No test asserts on that message;
--check and shellcheck both pass.
- Plugin self-containment has no gate that fires on it, so it goes in
architecture.md's Plugin model beside "self-contained plugin units", stated as
the constraint it is rather than a description, with the two consequences that
make it load-bearing: the duplicated Vale styles (ADR-0014) and ADR-0020's
constants copied into three validators.
Impact: the every-session set is unchanged — neither note returns to AGENTS.md.
Why: b9c7762 gave dated amendments to ADR-0013, ADR-0015 and ADR-0017, but two
pointers were missed. ADR-0007 cited the "Provider-agnostic issue tracker"
glossary entry, which the trim renamed to "Issue". ADR-0015 cited the "lint
plugin" entry, which the trim deleted outright — and ADR-0015 was amended for
its other CONTEXT.md pointer in the same commit, so this one was missed inside
an already-edited file.
Implementation notes: ADR-0007 gets a dated parenthetical matching the pattern
the other three carry, confirming the entry survived under a new name rather
than being dropped. ADR-0015's Considered-options parenthetical is re-pointed
at docs/spec/architecture.md's plugin scope table, which now carries the
repo-agnostic-versus-marketplace-specific argument the deleted entry held, and
its existing 2026-08-17 supersession note is extended to record that second
redirect in one place.
Impact: all four ADRs that referenced moved CONTEXT.md content now say where the
content went. No decision is altered.
Why: eight comments and one status note cited CONTEXT.md or AGENTS.md text that
b9c7762 and 1929ffd moved or deleted. All are inert at runtime, but they are the
rationale comments that tell the next maintainer why an assertion exists, and
they now name a file that no longer explains it.
Implementation notes: re-anchored by what the citation is for, not uniformly.
- Four sites quoted facts ADR-0013 owns — every rule is `level: error` with no
ignorable tier (ADR-0013:59-70), and KyberforgeCopilot's `.agent.md`-only
scope (ADR-0013:43-46). These now cite ADR-0013. ADRs are append-only here;
the spec docs are refactored, which is what caused this rot.
- Two sites quoted the glob location-independence property, which no ADR owns.
The quote was already inline and carried the full rationale, so the citation
added a rot surface and no information — dropped, statement kept.
- sync-marketplace-mirror.sh's header attributed the mirror-not-a-profile fact
to CONTEXT.md; the parenthetical beside it already carries the evidence, so
the attribution is dropped rather than re-pointed.
- .pre-commit-config.yaml cited an AGENTS.md instruction that no longer exists;
generalised to "the documented instruction".
- LESSONS.md:29 misquoted AGENTS.md's current session-start line.
Also corrects a pre-existing misattribution at tests/test-vale-wrap.sh:454:
AGENTS.md has never named bash 3.2 as a repo target (`git log -S'3.2'` on it is
empty). LESSONS.md and the script headers do.
Impact: no behaviour change. test-check-vale-style-sync.sh and test-vale-wrap.sh
both pass (42 passed, 0 failed).
Why: CONTEXT.md is an instructed every-session read, so a wrong assertion in it
is cited downstream without being checked. The b9c7762 rewrite introduced one
broken pointer and carried over two claims that were already false.
Implementation notes:
- The Preload tax entry pointed at docs/spec/gates.md for the measurement
method. gates.md has no such method — it uses the 23,427-char figure once, in
passing, without deriving it. The method and the current figure both live in
ADR-0020, which is where a dated measurement belongs: it is superseded there
deliberately, and issue #99 tracks its drift.
- Dropped the "~5,900 tokens across 39 skills" measurement rather than
re-measuring it. It was already wrong three days after being written (~5,340
today) and it was the only live number in the file. Every sibling entry is one
definitional sentence plus a pointer fragment; this one now matches, and
CONTEXT-FORMAT.md:43 asks for exactly that.
- Dropped "Written by the session-handoff skill or directly by the human" from
the LESSONS.md entry. No session-handoff skill exists — LESSONS.md:7 records
it as planned work. Who writes a file is not its definition.
Impact: removes the only rot surface in the file and the only assertion in it
that names an artifact the repo does not have.
Refs: #99
Why: CONTEXT.md is an instructed every-session read at 33,799 bytes (~8,450
tokens), and CONTEXT.md:17 layered docs/VISION.md on top. After the AGENTS.md
trim, that left ~9,650 tokens of instructed reads sitting behind a file whose
own format contract — grill-with-docs/CONTEXT-FORMAT.md:43, "Keep definitions
tight. One sentence max." — it was not following. 22 glossary entries averaged
190 words; the top four were 55% of the file, and one ("Vale audit prefilter")
was 1,444 words of pre-commit forensics filed as a definition.
Implementation notes:
- CONTEXT.md rewritten to CONTEXT-FORMAT.md's prescribed template: ## Language
(five clusters, 22 terms with _Avoid_ aliases), ## Relationships,
## Example dialogue, ## Flagged ambiguities. 33,799 -> 9,762 bytes.
- Vale prefilter forensics moved to docs/spec/gates.md under nine subsections:
the two-copy style layout, vale-wrap.sh and the Vale 3.15.2
text.frontmatter.description limitation, the --config argv defect (E100),
glob scoping, the rule inventory, and 0-files-means-NOT-RUN.
- Instruction-file format and the bidirectional-reference principle moved to
docs/spec/architecture.md, the latter flagged as aspirational: no instruction
file carries a `when:` field, only 2 of 39 skills do, and the reference
scanner script it names does not exist.
- Added a per-plugin scope table to architecture.md so "which plugin owns this
skill" survives the trim, and corrected its claim that the content index
lives in providers/claude-code/CLAUDE.md — that file is now two import lines
and the index sits in core/AGENTS.md.
- "Repo and Gitea are the only source of truth" moved from CONTEXT.md's
Principles into AGENTS.md session rules; it was operative guidance in a
glossary. docs/VISION.md and LESSONS.md added to the on-demand list.
- The apm-consumed install entry (432 words) was deleted outright: all four
consequences are already carried by README.md, ADR-0018, gates.md and
AGENTS.md.
- ADR-0013, ADR-0015 and ADR-0017 each pointed at CONTEXT.md sections that
moved; amended with dated supersession notes rather than left dangling.
Impact: the every-session set drops from ~15,700 to ~4,140 tokens across both
passes (-74%). The deleted "lint plugin" entry was also stale on a safety
claim — it said lint-runner inherits Edit, but ADR-0016 restored
`disallowedTools: Edit, Write, NotebookEdit` and CONTEXT.md was never updated.
Refs: #99
Why: AGENTS.md is preloaded into every session via CLAUDE.md's @AGENTS.md
import, so every word is a tax paid whether or not it is relevant. It had
grown to 3,371 words (~4,550 tokens), 75% of it a single "Setup and testing"
section that was mostly neither: roughly 1,600 words were gate forensics —
why a hook is shaped the way it is, what was tried and rejected, verification
numbers — and only ~390 words were genuine one-time setup.
Implementation notes: split three ways by audience rather than by heading.
One-time setup, prerequisites and test commands move to a new root README.md,
which did not previously exist. The gate forensics move to a new
docs/spec/gates.md, preserving every constant, hook ID, error string and
tried-and-rejected note (policy.fetch_failure_default, the removed fallback
frontmatter reader, extending skill-size-check's files: pattern to agents,
tidying .claude/settings.json out of the pretty-format-json exclude list) —
these exist to stop settled decisions being re-litigated, so none were
dropped. AGENTS.md keeps only what applies every session, at one clause of
rationale per rule. The .apm/-mirror section collapses to a pointer because
docs/spec/architecture.md already covers it in full. The hot ADR-0020 gate
counts stay in AGENTS.md deliberately: they block any skill edit today.
Also corrects docs/spec/architecture.md's stale claim of 19 ADRs; ADR-0020
landed without updating the count.
Impact: per-session preload drops from ~4,550 to ~1,116 tokens, a 75% cut.
agentsmd-audit passes: secrets clean, drift clean, 4 informational findings
for sections deliberately relocated to the README.
Why: kyberforge's SessionStart hook found six packages behind the holocron
default branch and ran `apm update --yes`, which re-resolved every unpinned
dependency ref and rewrote the lock. The dependencies are declared unpinned
against the default branch, so this churn is expected rather than a defect.
Impact: pins the deployed skill and agent set to the refs resolved this
session. No manifest or plugin content changed.
Closes the second open design decision on PR #103. The `/compact` finding was
recorded as a false positive needing an allowlist or a suppression mechanism.
It is neither: the routing universe is the apm marketplace, so a target either
resolves to a skill or an agent or it does not resolve, and `/compact`,
`/clear` and `/init` are Claude Code slash commands with no counterpart in
Copilot CLI or Codex. `.apm/` source compiles for all three, so a
vendor-neutral description routing to one is a portability defect and the hard
FAIL is a true positive.
An allowlist was rejected for a concrete reason, not a stylistic one: it
answers a different question ("does this exist on some host?"), it cannot
answer that portably from a single source file, and it goes stale the next
time a host ships a command — reintroducing the same-commit-two-verdicts
failure ADR-0020 already closed for deployed trees.
Nothing is blocked today: zero of the 43 descriptions name a host built-in,
and an author who needs to mention one writes it un-slashed, which is not
route notation and carries no routing claim.
Recorded in ADR-0020 and in both author-facing contract references, so the
next agent reads the decision rather than "fixing" the gate.
ADR: 0020
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A prose-form routing target blocks a commit only when its own sentence names
another target that resolves. That makes the sentence splitter part of the
ADR-0020 contract rather than an implementation detail, and the naive
"period, space, capital" rule got it wrong in both directions:
- OVER-SPLIT: `e.g. "..."` is not a sentence end, but the quote looks like a
start. The clause was cut in half and the corroborator stranded on the far
side, so a genuinely dangling target silently demoted to SUGGESTION — a
measurement taken and then discarded, the vacuous-green shape this gate
exists to prevent. Seven such splits are live in the current corpus.
- UNDER-SPLIT: a sentence opening with a code span or a lowercase skill name
was not seen as a start, so two sentences merged and a resolving target
vouched for an unresolvable one it never stood beside — a hard FAIL with no
escape hatch, which is the exact failure corroboration was added to prevent.
The splitter now excludes the five abbreviations that occur in routing prose
and admits a backtick or lowercase letter as a sentence opener. Applied
byte-identically to all three copies of the shared resolver.
Verified zero-delta against the corpus: 37 ERROR / 58 SUGGESTION / 2 dangling
before and after, findings byte-identical. The exposure this closes is to the
descriptions #99 is about to rewrite, not to the ones already measured — which
is why the deferral reason recorded on PR #103 ("can move the documented corpus
counts") does not hold and the fix lands here rather than after the retrofit.
Three regression tests, one per direction plus the backtick opener, each proven
non-vacuous by reverting the splitter alone and watching it go red.
Refs: #99
ADR: 0020
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both AGENTS.md and ADR-0020 said deployed .claude/.agents trees are consulted "only
when no authoring root exists". That stopped being true in f7cc279: the walk-up
finds a root in any git repo, so the condition is now whether that root holds
plugins, not whether one was found at all. Left alone, the two documents describe a
resolver that no longer exists — and this repo's prose is load-bearing, since the
next agent reads it instead of the code.
Both now also record why a name-count delta is not an equivalent test, because it is
the obvious simplification and it is wrong: a single-plugin monorepo re-collects its
own package, adds no new name, and would pull the deployed trees back in.
ADR: 0020
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015W3iwF9ncfRZddGBxsMCYi
Diffing each retrofitted SKILL.md against its replacement references/ files found
rules that existed on main and now existed nowhere — relocated in intent, deleted in
fact. A trim that loses a rule is not progressive disclosure, it is data loss with a
smaller word count.
Three had no survivor. The least-privilege guidance for `tools` kept its mechanics
and lost the "restrict to what the agent needs" half, so the remaining text read as
encouragement to omit the field. The improve flow lost its regression check, so
nothing compared the closing audit against the pre-edit state and a PASS quietly
becoming a SUGGESTION went unnoticed — restored on both halves of the author pair,
since agent-author had dropped its equivalent too. And agent bodies lost "would the
agent get this wrong without it?", which mattered more than it looks: ADR-0020
deliberately sets no body word gate for agents, three of the four already sit
between 933 and 1,199 words, and the delegation check only fires on procedure a
skill already owns. That heuristic was the only brake left.
Two more were reachable only from the wrong scope. agent-author tells the reader to
load only the file for the resolved scope, but the mcp__ glob syntax for
disallowedTools and the five tools no subagent ever receives had both landed in
project-user-scope.md. disallowedTools is the ONLY permitted fence at plugin/APM
scope, so the scope that needs the syntax most could not reach it, and a plugin-scope
run could write a body telling the agent to ask the user a question.
Two documents were actively wrong rather than merely thin. agent-audit told auditors
that validate.sh resolves boundary targets for skills only; it runs at both scopes,
so the auditor was hand-resolving what the script had already decided and could
contradict it. And skill-audit routed to its script-troubleshooting reference
whenever validate.sh "fails" — but it exits 1 on ordinary content FAILs, the normal
outcome for the whole #99 population, so 1,302 words loaded on nearly every audit.
A context-budget regression inside the skill that enforces the context budget.
Finally, two illustrations taught the shape the gate ERRORs on, unfenced, while an
adjacent rubric called it a hard ERROR.
LESSONS.md records the reference-chain depth rule flipping from "one level deep" to
"two hops, never three". ADR-0020 is silent on it and the reversal rode entirely on
the diff; the looser rule is what mandatory dispatch requires.
Refs: #99
ADR: 0020
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015W3iwF9ncfRZddGBxsMCYi
Every defect fixed in f7cc279 was reachable because nothing asserted against it.
The gate had 43 assertions and none of them covered a consumer repo, a non-string
description, an unclosed fence, or the two spec ceilings. Each case below fails
against the pre-fix code and passes against the current one; every one was proved
non-vacuous by mutating a scratch copy of the script and watching the test go red,
independently twice.
The two that mattered most had no fixture anywhere. A consumer repo WITH .git is
the shape the resolver exists to serve, and only the no-.git case had ever been
tested, which is exactly why the blocker was invisible. And ADR-0020 says the
walk-up runs in two passes specifically so a nested .git cannot beat a plugins/
root further up — no fixture had ever placed a .git inside a plugin.
test-adr0020-differential.sh loses _non_adr_hook_error(). It excluded MAX_LINES and
MAX_WORDS from the cross-script comparison on the untested assumption that awk and
splitlines() agree. They do not, and the divergence stayed invisible for exactly as
long as the exclusion stood. The ceilings are now compared like any other rule.
Two existing assertions were repairs, not additions. The skill-improve probe had
been fixed by this very branch, so its iteration permanently took an
assertion-free SKIP that still counted as a pass; both branches now fail loudly and
each names the other file's pin so the two stay in step. And the yaml-none fixture
emitted `---/---`, which never matched the frontmatter pattern at all — it passed on
the bare word "frontmatter", present in both messages, while never reaching the
branch it was named for. Needles throughout that file now name their branch.
Refs: #99
ADR: 0020
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015W3iwF9ncfRZddGBxsMCYi
Review of the ADR-0020 gate found four ways it could exit 0 without measuring, and
one way it hard-failed a repo it had no business failing. On a gate shipping hot
with no baseline, a silent pass is the worst outcome available and a false block is
the second worst.
Consumer resolution was the blocker. _authoring_root() fell back to the nearest
.git, so it returned truthy in ANY git repo; _collect_authoring_root() then
contributed nothing and the deployed-tree branch was dead code in precisely the
consumer case it exists for. A consumer repo routing to an installed sibling got an
unblockable ERROR, and deleting .git "fixed" it. It now keys on which of the two
walk-up passes matched. A name-count delta was tried first and is wrong: a
single-plugin monorepo re-collects its own package and adds no new name, so the
delta reads zero and drags the deployed trees — including a global ~/.claude — back
into the universe. That reintroduces the install-dependence ADR-0020 forbids, one
layer down.
The three silent passes: an indented `---` inside a block scalar truncated the
frontmatter and reclassified the rest of the description as body; a non-string
description was str()-coerced, so `description: true` measured as the four-character
"True"; and an unterminated fence blanked the rest of the body, disabling the
ERROR-tier references/ check and the gotcha counts.
Two measurement defects came with them. The awk line/word counts discarded awk's
exit status, so an unreadable file passed both spec ceilings in total silence, and
awk NR/NF disagreed with the audit script's splitlines()/split() on Unicode
whitespace — the "fix one gate, get blocked by the other" bug, on the two axes the
differential test deliberately excluded. Both counts now run in the Python block
that already reads the file. A type error also no longer reports itself as a syntax
error.
Also: glob metacharacters in the checkout path silently disabled the resolver;
re.I was applied to some extraction patterns and not others; agent-audit missed
`tools:` written as a YAML block sequence, the shape Copilot files use; and a
nonexistent agent file raised a bare FileNotFoundError instead of a diagnostic.
The shared resolver block stays byte-identical across all three scripts. Corpus
output is unchanged — 26 description FAIL, 9 body FAIL, 2 dangling, 0 missing
references, 58 SUGGESTIONs — so no documented count moves.
Refs: #99
ADR: 0020
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015W3iwF9ncfRZddGBxsMCYi
The four preceding commits change `plugins/kyberforge/.apm/` content that reaches
the compiled artifacts — three validators, a new reference file in each of
`skill-audit` and `skill-author`, and the authoring rules across both author
skills — so per `apm-workflow`'s configure policy the package earns a bump, minor
for the new capability.
Root `apm.yml`'s `executables.allow` key moves with it, in this commit and not a
later one. apm approves a package's `hooks/` and `bin/` by an exact
`<name>#<version>` dictionary lookup with no wildcard and no version-less form, so
a `kyberforge#1.5.0` key left behind a 1.6.0 package errors nowhere: the entry
stops matching, the `SessionStart` freshness hook stops deploying, and the install
goes quietly stale. That is the failure ADR-0019 records as having actually
happened, and `check-executables-allow-sync` exists to catch it.
The catalog bump was missing from the working tree and is added here.
`apm-workflow`'s marketplace policy is explicit that an existing entry's
`version:` moving earns the catalog a **patch** — the set of packages is
unchanged, only its metadata moved — and that the root `version:` stays in step
with `marketplace.version`, since apm audit reads one and the compiled manifest
carries the other. Nothing enforces this: `apm pack --check-clean` catches a bump
made in `apm.yml` but never re-packed, while a bump never made at all fails
nothing.
Manifests regenerated with `apm pack` plus `scripts/sync-marketplace-mirror.sh`
for `.github/plugin/marketplace.json`, which no apm output profile targets.
`.agents/plugins/marketplace.json` is unchanged — the codex profile's shape
carries no version field for either the catalog or its entries.
The ADR was written against base commit `f9b919d` and then not updated as the
implementation moved, so several of its numbers were measuring one thing and being
read as another — the exact conflation the ADR exists to stop, reproduced inside
it. Corrections, all reproducible now that each figure states its method:
- The preload tax is 23,427 chars / ~5,900 tokens, not 23,612 / ~6,200.
- `MAX_WORDS=2770` is a density proxy for the agentskills.io ~5,000-token ceiling,
not "2× p90". Neither percentile reaches it: 2× the body-only p90 is 2,698 and
2× the whole-file p90 is 3,052. Reading it as a percentile pairs a whole-file
gate against a body-only distribution.
- `apm-workflow` is a 421-word body; 554 is its whole-file count. `skill-author`
and `agent-author` were 2,623 and 2,582 body words — 2,760 and 2,758 whole-file,
which is where "within twelve words of the gate" comes from. Two numbers for one
file is the point, and only one of them is what either gate measures.
- Every `file:line` citation now says it resolves against `f9b919d`, since this
change rewrites most of the cited files.
Three things the ADR asserted that no validator implemented are now filed by tier
in an exhaustive enforcement table — deterministic, prose-pattern, or auditor
judgment — because a rule filed under "Enforcement" that nothing enforces is the
failure mode this ADR is most exposed to. The Gotchas entry count moves to
SUGGESTION to match the script; the paraphrase FAIL is marked as an auditor's,
since semantic equivalence is not pattern-matchable.
Two gaps recorded rather than quietly left:
- The agent body-gate exemption lives in `agent-audit`'s validator and in the
`skill-size-check` hook's `SKILL.md`-only `files:` pattern — *not* in
`scripts/skill-size-check.sh`, which measures whatever path it is handed and
today reports 900-word body FAILs on `git-orchestrate` (933),
`gitea-orchestrate` (1,199) and `apm-orchestrate` (1,080). Agents escape by file
pattern, not because the script knows the difference, so widening that pattern
would silently enforce a gate this ADR declines to set.
- The `skill-audit`/`agent-audit` merge is deferred to #101. This change made the
split deeper, not shallower: the dispatch retrofit took them from 3 and 4
reference files to 7 and 8, and their two same-named `description-quality.md`
files now differ on 100 of ~120 lines after normalising skill/agent. The merge
reopens ADR-0008 and touches every call site in `skill-author`, `agent-author`
and `forge`, so it is its own change. #100 carries the dangling-target fixes.
AGENTS.md and CONTEXT.md take the same corrections plus the two live setup
changes: PyYAML is now a hard requirement rather than an optional accelerator (a
fallback that mis-parses an unfamiliar scalar shape reports a clean pass on a file
it never measured), and `.claude/settings.json`'s `pretty-format-json` exclusion is
documented as load-bearing rather than as a tidy-up candidate.
LESSONS.md's autofix entry is corrected on its own provenance, which it got wrong
in both directions. `git log --date=iso` puts the introducing commit at 18:47 and
the fix at 21:54 — three hours, not "weeks" — and `git branch -a --contains` puts
the introducing commit on this branch only, not on main. It was manufactured
inside the same PR that diagnosed it. The added lesson is that "pre-existing" is a
claim about history and history is queryable: a defect found while working on a
branch feels inherited, and the feeling is not evidence.
Refs: ADR-0020, #99, #100, #101
7607522 fixed the symptom in the wrong place. It made `test-run-tests.sh`'s
`run_fake()` spawn fixtures via `env -u RUN_TESTS_STRICT`, which stops that one
suite inheriting strictness — and leaves every future suite to defend itself the
same way. The variable's only job is done the moment `run-tests.sh` latches it
into the `STRICT` shell local, so it is unset there now and the leak is gone for
every child. The `env -u` stays as this suite's own defence in depth rather than
as the fix.
Two corrections to that commit's account of the bug, both overstated and both
cheap to have checked:
- The blast radius was two assertions, cases 10c and 10g, not six. Nothing else
in the repo reads `RUN_TESTS_STRICT`.
- The pre-push gate was never red. It invokes `bash tests/run-tests.sh --strict`,
and the flag sets a shell local that is never exported, so the flag spelling
never leaked at all. Only the env-var spelling did.
That asymmetry between the two documented spellings is the real finding, and
nothing asserted against it. Case 10b compared the parent's verdict, which is the
half that already matched; the halves that differed were the environments the two
spellings handed every dispatched suite. New case 10i asks a child directly —
`${VAR+set}`, so an exported empty value still counts as a leak — and asserts the
two observations equal each other rather than a hardcoded expectation, so they
cannot drift apart in a direction the case did not anticipate.
Six defects, each one a place where two files that an author reads in the same
sitting told them different things — or where the trim dropped a rule and nothing
noticed because no gate covers prose.
**"Use proactively" contradicted itself across the pair.** All three agent
templates said to add it where the runtime should delegate unprompted, while
`agent-audit`'s `KyberforgeCopilot.ProactivePhrase` rule grades it a hard FAIL in
any `*.agent.md` — which is the Copilot half of every project/user pair *and* the
vendor-neutral plugin-scope file, since that compiles to a real Copilot agent
downstream. Following the template produced a file the repo's own gate rejects.
The phrase is now permitted in exactly one place, the Claude Code `.md`, and
`references/contract.md` carries the per-file table plus the consequence authors
ask about next: a pair whose CC half has it and whose Copilot half does not is
correct, because `agent-audit` checks that both halves describe the same job, not
that they match word for word.
**The output-schema rule contradicted itself inside one file.** `contract.md`
said any content only one branch reaches moves to `references/`, and then offered
an "Output format template" body pattern with no qualification. Stated once now,
so it is not re-litigated: an output schema stays in the body only when every flow
produces it and it is roughly 50 words or less. No third option.
**Gotchas tiers disagreed with the script.** `validate.sh` emits the entry count
through `suggest()` and exits 0, while `skill-author` and `skill-audit` both
called more than five entries a FAIL. Whether a given gotcha earns its place is
judgment, so the prose moves to the script's tier rather than the reverse. The
paraphrase rule stays a FAIL and is explicitly marked as the auditor's call — no
script detects it.
**The dispatch exemplar was cited at the wrong number.** `apm-workflow`'s body is
421 words; 554 is its whole-file count. Both `contract.md` and `body-discipline.md`
cited 554 while describing a body budget, so an author calibrating against the
exemplar overshot by ~30% — the exact whole-file/body-only conflation those two
sections exist to warn against, reproduced inside the warning.
**"Error handling" came back as a required body element.** It was one of four and
is the one that gets dropped, and dropping it is not neutral: an agent handed
malformed input with no instruction invents a recovery, and a subagent's invented
recovery is invisible to its caller until the output is wrong. Restored in
`agent-audit`'s rubric as a SUGGESTION, in `agent-author`'s contract and both
scope checklists as a required element, and as an `## Errors` section in all three
templates.
**`skill-author` Step 4 gains the one check the audit misses.** An empty body
reports `PASS SKILL.md body word count 0` — a word gate cannot tell "concise"
from "absent". Step 4 now hand-checks for a non-empty section, and its commit
verification is conditioned on actually being inside a git worktree, which a skill
under `~/.claude/skills/` is not.
Also here: absolute repo paths removed from `skill-author`'s SKILL.md and
contract.md in favour of naming the skill (`zoom-out`'s description is quoted
inline instead of pointed at), the boundary-target universe documented to match
the resolver, a two-hops-from-SKILL.md limit on reference chains, and
`new-agent.sh`'s next-steps output naming the description budget and the
deliberate absence of an agent body gate.
Refs: ADR-0020
ADR-0020 shipped its gates hot with no baseline file, so 26 of 39 descriptions
and 9 of 39 bodies are over their FAIL tier and editing any of them for any
reason requires bringing the skill into contract first. `references/improve.md`
said exactly that and stopped there — it mandated a retrofit and supplied no
procedure for one.
Four dry-run retrofits confirmed what that costs. Asked the same questions —
what to cut first, when a body is two flows rather than one, what else has to
change alongside — they invented six to ten different answers, so the same skill
retrofitted twice produced two different skills and neither run could be reviewed
against anything.
`references/retrofit.md` fixes the answers: an ordered cut list ranked by tokens
removed against behaviour lost (inverting that order is how a retrofit deletes the
instruction the skill existed to carry), the test for whether a body holds two
mutually exclusive flows, the reference-file conventions, the collateral checklist
for `README.md` and `references/sources.md`, and a worked description retrofit.
It also states the trap the dry runs kept hitting: retrofit the skill in place,
inside its package. The boundary-target universe is built by walking up from the
file being checked, so a scratch copy has no authoring root above it, the check
prints `INFO ... DID NOT RUN`, and the run still exits 0 — a line that reads as a
pass and is not one. A retrofit signed off on a copy carries an unverified
boundary target into the corpus.
Loaded from the improve flow only when a budget is actually exceeded, so a routine
improvement pays nothing for it.
Refs: ADR-0020, #99
The ADR-0020 body trim took `skill-audit` from 2,623 body words to a dispatch
shape, and two things went out with it that were not padding.
The manual structural fallback was one. Its replacement was a single sentence
telling the auditor to report an INFO when `validate.sh` cannot run — so with no
`python3` or no PyYAML, `skill-audit` reported the gap honestly and then audited
nothing structural at all. Every ADR-0020 measurement, the whole-file ceilings,
the name-to-directory match, the `references/` pointer check and the script
hygiene checks silently left the audit. A skill's whole Structure dimension
hanging on one optional interpreter is the same vacuous-pass shape the gate
scripts were just fixed for, one layer up.
The `E100 Runtime error ... does not exist` diagnostic was the other. That exit
code means an explicit relative `--config` was passed to `vale-wrap.sh` while
vale itself was installed and working; without the note, Step 1's fallback reads
exit 2 as "vale unavailable" and downgrades the description, body-discipline and
patterns dimensions to full LLM judgment for a config error it could have fixed.
That misreading is already recorded in CONTEXT.md as the reason both audit skills
stopped passing `--config` at all.
Both are restored in `references/validation-scripts.md`, loaded only when a Step 1
script fails — so the body pays nothing for them on a clean run, which is what the
dispatch pattern is for. The file also carries the by-hand boundary-target
procedure and the three ways to misread the result, including that
`INFO ... DID NOT RUN` is not a pass.
`references/file-structure.md` gains the one sanctioned spelling for a cross-skill
reference. The possessive form (``skill-audit's references/validation-scripts.md``)
is the only spelling both rules accept: a full repo path is what that section
already forbids, and a bare `references/<file>.md` is now a hard ERROR from the
ADR-0020 pointer check, which requires the file to exist in the skill's *own*
directory. Without the rule the two constraints look mutually exclusive.
Refs: ADR-0020
Three ways the gates could report green having measured nothing. All three were
invisible to a passing test suite, because pre-commit prints nothing at all for a
hook that exits 0 — a gate that declines to check and a gate that checked and
passed produce the identical signal.
- A UTF-8 BOM, a leading blank line, a trailing space after a `---` marker or
CRLF line endings defeated the `^---\n` frontmatter matcher. Every ADR-0020
check was then skipped and the file passed: measured at the time, a
550-character description with a 1,000-word body exited 0 behind a BOM.
All four shapes are now tolerated, and frontmatter that genuinely cannot be
parsed is a hard ERROR rather than a silent skip.
- An agent file with a valueless `description:` followed by another key let a
line regex capture the *next* key, which looked non-empty, so the
missing-or-empty branch never fired and every gate below it early-returned on
the empty folded value — zero output, exit 0, on a blocking gate. The one
field this contract is entirely about was the one field a gate could fail to
notice was absent. Presence is now decided on the YAML-folded value and
nowhere else, and a missing or empty description is a hard FAIL in all three
validators.
- The hand-rolled frontmatter fallback disagreed with PyYAML across the FAIL
boundary on folded scalars, so which reader happened to be available decided
the verdict. A fallback that mis-parses a scalar shape reports a vacuous pass,
which is worse than not running, so it is deleted: python3 and PyYAML are hard
requirements that fail loudly with an install pointer.
Boundary-target resolution no longer derives its universe from its own location.
A `${BASH_SOURCE}`-relative repo root leaked this repo's 39-skill universe into
every consumer repo running the hook through pre-commit, so a consumer skill
routing to `skill-audit` resolved against a plugin it had never installed. The
interim form resolved through `.claude/` and `.agents/`, which are gitignored
`apm install` output — the same commit reported 2 dangling targets on a machine
that had run the install and 6 on a fresh clone. Resolution now walks up from the
file being checked to an authoring root (nearest ancestor holding
`plugins/*/.apm/{skills,agents}`, else the nearest `.git`, in two passes so a
nested `.git` cannot outrank a real monorepo root); the universe is every skill
and agent under `<root>/plugins/*/` plus the file's own apm package and that
package's declared `dependencies.apm`. Deployed trees are consulted only when no
authoring root exists at all — the consumer case. One commit now gets one verdict,
which a gate shipping hot with no baseline file has to.
Narrowed in the same pass: a routing target inferred from the prose boundary form
and corroborated by nothing else reports at SUGGESTION instead of blocking. A
blocking check with no escape hatch is the wrong trade when the inference from
prose is the weak part of it.
New deterministic checks, all previously untested or absent: every
`references/<file>.md` a body names must exist (ERROR — a broken pointer is not a
style opinion); a description with no boundary clause at all, a Gotchas section
over five entries, and a Gotchas section over 25% of the body are SUGGESTIONs.
Where no universe can be determined the target check prints `INFO ... DID NOT
RUN` rather than passing quietly. Each prose-scanning check needed its own
false-positive fix — a fenced example of a Gotchas section was being read as the
section itself — and those fixes are pinned rather than assumed.
The resolver is one block copied verbatim into all three scripts between
BEGIN/END markers, because a cache-installed plugin's scripts cannot read outside
their own plugin directory. Nothing asserted the copies were still identical; a
one-line edit to a single copy passed every constant-agreement assertion, since
constants are not what drifts.
Tests land here rather than in a later commit. The existing suites assert the old
behaviour and go red against these scripts, so splitting them would leave a commit
whose own `run-tests` pre-push gate fails in isolation.
Refs: ADR-0020
Two pre-existing failures, both red at HEAD before ADR-0020 work began,
both invisible in an ordinary local run.
RUN_TESTS_STRICT leaked from the environment into test-run-tests.sh's
fixture children. The meta-test is itself a suite the runner discovers,
so under the gate's own invocation the variable propagated outer runner
-> batch_run -> the fixture's copy of run-tests.sh, flipping it strict.
Case 10c (a deliberate control asserting a skip is tolerated WITHOUT
strict) then failed. Six further cases were silently running strict too
and asserting against the wrong stream — case 9 was matching the stderr
strict block rather than the stdout skip list it was written to check.
run_fake now spawns via 'env -u RUN_TESTS_STRICT', so fixture strictness
is a property of the case, never of how the file was launched. No
assertion weakened; run-tests.sh itself is untouched.
pretty-format-json --autofix was re-sorting apm's output on the way into
every commit. .claude/settings.json is apm-owned (ADR-0018/0019) and its
exclude list named fifteen generated manifests but not this file, so
since 2e395a4 it has been committed in a key order apm would never write
— permanent drift on a file with an empty git diff. Content was always
byte-identical; only JSON key order differed. The exclude ships in the
same commit as the corrected file because otherwise the hook re-breaks
it during staging.
apm.lock.yaml: generated_at churn, plus lint's exec_status corrected from
'deployed' to 'gated_pending_approval' — executables.allow grants only
kyberforge#1.5.0, so lint's hooks/bin are genuinely gated.
New coverage: an ambient RUN_TESTS_STRICT must not reach a fixture that
did not ask for it, and under --strict the skip report goes to stderr
only with the stdout list suppressed. Neither was pinned.
plugins/lint/hooks/hooks.json was no longer produced from .apm/hooks/,
so check-plugin-content-sync failed on it at HEAD:
DRIFT plugins/lint/hooks/hooks.json: stale, no longer produced from .apm/hooks/
Surfaced by running sync-plugin-content.sh --all during unrelated work.
Unrelated to ADR-0020; committed separately so the contract change stays
reviewable on its own.
Per ADR-0017's 2026-08-14 amendment, a hooks.json a sync no longer
generates is deleted as stale.
Skill name+description pairs are preloaded into every session, costing
~6,200 tokens across 39 skills before any skill is invoked. The authoring
rules mandated that growth: skill-author:104 and description-quality.md:21
both required padding, while skill-author:102 (the deflating rule) had no
FAIL condition behind it.
Gates (blocking, no baseline file):
- description 250 chars SUGGESTION / 400 FAIL, measured on the folded
YAML value
- body-only 600 words SUGGESTION / 900 FAIL, independent of the unchanged
whole-file 2770-word / 500-line spec backstop
- every boundary-clause routing target must resolve to a real skill or
agent; catches skill-improve, neuledge-context and gitea-labels
- agents take the description gates but deliberately no body gate; a test
pins that absence
Vale: DescriptionOpener widened to ^This\b, new CompositionNote rule
banning architecture notes from descriptions. 10 hits, 0 false positives.
Kyberforge's own four skills retrofitted: descriptions 3,364 -> 938 chars
(-72%), bodies 8,306 -> 2,487 words (-70%), all via the apm-workflow
dispatch pattern. Fixes the skill-improve dangling route and the
agent-author misroute to manual review.
Also fixes a pre-existing false positive where any line-initial 'read '
was flagged as interactive input, which had already caused two scripts to
be rewritten around it.
Refs: ADR-0020
Every installed skill's name+description is preloaded every session -
23,612 chars (~6,200 tokens) across 39 skills. The authoring rules
optimised for triggering reliability with no counter-pressure on size:
skill-author:104 and description-quality.md:21 both mandate padding,
while skill-author:102 (the rule that would deflate it) is
judgment-only and absent from description-quality.md's FAIL conditions.
ADR-0020 sets the description shape (trigger + one capability +
boundary), two-tier size gates sitting below the unchanged
agentskills.io ceilings, a mandatory dispatch pattern for multi-branch
bodies, a Gotchas constraint, an agent-side delegation check, and
invocation as a design axis. Gates ship blocking with no baseline.
Adds five CONTEXT.md glossary entries: preload tax, skill context
contract, dispatch body, hand-invoked skill, delegation discipline.
Refs: ADR-0020
Deploys kyberforge's SessionStart apm-currency hook into
.claude/settings.json (ADR-0019), bumps bin 1.1.2->1.1.3 and
kyberforge 1.4.1->1.5.0 in apm.lock.yaml with new exec_status
fields, and removes lint's now-empty .apm/hooks/ source dir
(the generated hooks/hooks.json mirror is untouched).
AGENTS.md and CONTEXT.md asserted that the `<plugin>:<skill>` form "no
longer resolves here". It does: ~/.claude.json still enables core, git,
gitea, kyberforge and lint at user scope, which ADR-0018 left in place
deliberately. Both names are live at once, so a working `gitea:gitea-prs`
is the user-scope copy answering — not evidence that the apm install is
broken and not something to "fix". ADR-0018 contradicted itself on this,
claiming every namespaced reference went stale while its own "User scope
is untouched" consequence said otherwise; recorded as a dated correction
alongside the existing one. Bare names stay the documented default.
Five stale pre-push hook counts updated for the new
check-executables-allow-sync gate: 13 -> 14 repo-defined hooks, 15 -> 16
reported by pre-commit, eleven -> twelve passing offline. The gate reads
two local manifests and makes no network call, so the SKIP pair for
offline pushes stays exactly two. "Four pre-push hooks shell out to apm"
is unchanged and still correct — the new hook parses YAML directly.
ADR-0019 gains the timeout arithmetic, the singular/plural failure and
why mocking every apm call hid it, and a consequence recording that the
trust gate is keyed on version rather than content: an edit to a hook
script landing on main deploys and executes unattended on the next
session start, since the dependency is unpinned and the hook runs
`apm update --yes`. That is accepted, not overlooked, but it is why the
gate should not be read as a supply-chain control.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7GvKuJfy2WrdBmUttV4DT
apm prints "1 outdated dependency found" in the singular when exactly one
package is behind (apm_cli/commands/outdated.py). check-apm-current.sh
matched only "outdated dependencies found", so one stale package was
invisible: the hook exited 0 silently and no refresh ran. With six
packages merging independently, one-behind is the ordinary case, so the
freshness mechanism failed most often in the situation it exists for.
Three further defects in the same hook:
- The host timeout was below the script's own budget. hooks.json declared
320s while the script allows `timeout 60` plus `timeout 300` = 360s, so
a slow remote let the host kill the hook mid-update and leave
.claude/skills/ half-deployed with nothing emitted. Now 380. A test
asserts the invariant rather than the literal: it sums every `timeout N`
parsed out of the script and requires hooks.json to exceed it, so
changing either side alone fails.
- The lockfile guard was cwd-relative, so a session opened in a
subdirectory no-opped silently and ran both apm calls against the wrong
directory. Now anchored on CLAUDE_PROJECT_DIR, falling back to the cwd
so the hook stays inert under a host that does not set it.
- Every assertion mocked apm, so the suite was green over code that could
not detect its own most common trigger. That blind spot is what hid the
singular/plural bug, and it is the same shape as the deleted post-push
tests. The suite now stages a genuinely outdated dependency against a
local git remote — offline, via url.<path>.insteadOf, so the
pass-under-unshare property survives — runs the real `apm outdated`, and
replays its output through the real hook. Reverting the grep to
plural-only fails it.
23 -> 35 assertions. Each fix mutation-tested individually. kyberforge
stays at 1.5.0: it is untagged, so this changes what 1.5.0 ships rather
than superseding it, and executables.allow needs no edit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7GvKuJfy2WrdBmUttV4DT
plugins/bin/.mcp.json declared the obsidian server as
`npx @bitbonsai/mcpvault@latest`, so an unpinned third-party npm package
was fetched and executed at every session start. The apm-consumed install
promoted that string to committed repo-root content in .mcp.json, giving
every clone the same unpinned execution. Pinned to 0.15.0, the version
`latest` currently resolves to.
bin 1.1.2 -> 1.1.3 and marketplace 0.4.0 -> 0.4.1, following the mapping
bb9158d establishes and 3bfdf58 confirms: the marketplace takes the same
bump severity as the highest-severity package bump. kyberforge is not
bumped here, so executables.allow's `kyberforge#1.5.0` key is untouched.
The pin is not live for this working copy until this lands on the remote
and `apm update` re-resolves — apm.lock.yaml still records 1.1.2 and
`@latest`, because the six dependencies resolve from the remote rather
than from the tree beside them. Correct for a fresh clone immediately.
.gitignore gains /.claude-plugin/plugin.json: a bare `apm pack` emits a
root-package manifest there that has never been tracked on any branch.
Scoped to the file, since the sibling marketplace.json is compiled output
that is committed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7GvKuJfy2WrdBmUttV4DT
apm approves a package's hooks and bin by an exact dictionary lookup on a
composed `name#version` key (apm_cli/security/executables.py,
is_package_approved). There is no wildcard and no version-less form, so
bumping plugins/kyberforge/apm.yml without editing root apm.yml's
`kyberforge#<version>` key errors nowhere: the entry stops matching, the
SessionStart hook stops deploying, and the install goes quietly stale.
ADR-0019 already named that as a live failure mode, mitigated only by a
comment in the executables block. This repo gates generated-content
drift, marketplace mirror drift and vale style drift deterministically,
and a silent-staleness failure is worse than any of them — a comment does
not survive the release that breaks it.
check-executables-allow-sync.sh parses the version out of the plugin
manifest and asserts the matching key exists in the root manifest. It
uses PyYAML where importable and falls back to a two-shape scan
otherwise, so a missing pip package cannot become the thing that blocks
every push; the test asserts both readers agree. 23 assertions.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7GvKuJfy2WrdBmUttV4DT
