docs(decisions): 0255 — skill namespaces keep v1 and fabrika apart (#4829)

[usirin]

The repo carries two skill sets right now — the old pipeline (v1) and its fabrika rebuild — and a
worry had been recorded that a rebuilt skill and its v1 original were fighting over the same name.
They are not. I measured it: Claude Code names a plugin's skills plugin:skill, and a plain skill's
name is just its directory name — so the plain report is always v1's and the rebuild is only ever
fabrika:report. Both show up in a real session, side by side, under different names.

That changes the answer. There is no name fight to fix, so this PR keeps the current setup, writes
down the rule (plain name means v1, fabrika: means the rebuild), and names the single edit that
retires the whole old set later. What does genuinely overlap is the descriptions — which is what
the model actually picks a skill by — and that was already the recorded mitigation.

Fixes #4829

What this changes

• New ADR 0255 (.decisions/0255-skill-namespaces-keep-v1-and-fabrika-apart.md) — takes the
  issue's branch 2 (keep the loading path, correct the record), states the namespace rule, explains
  why the symlink is not cut now, and sequences the cutover.
• claude-plugins/fabrika/skills/report/contract.md — the "dormant only by configuration"
  paragraph is corrected in place and points at the ADR.
• claude-plugins/fabrika/skills/triage/contract.md — its already-correct correction is
  sharpened to say the overlap is in the description, not the name, and cites the ADR instead of the
  open issue.

Acceptance criteria

• An ADR settles which branch is taken, naming the cutover sequencing. ADR 0255 §2 (why the
  path is not cut now) and §3 (the three-step sequence, ending in one edit that deletes both
  symlinks and removes the ADR 0077 suppression together).
• The decision covers .claude/agents and states whether .claude/.pipeline is affected.
  §4 covers .claude/agents — same mechanism, same lever, and no collision today because fabrika
  ships no agent definitions. §5 confirms triage's read: .claude/.pipeline is a path alias, not
  a roster-load directory, so it is unaffected. §5 adds the qualifier the write-code skill scripts are unreachable in an agent worktree — .claude/.pipeline symlink is not provisioned (exit 127 on the Step-4 preflight) #4818 precedent needed —
  tracking a symlink is safe for a path alias and is a roster decision for a load path.
• The report contract's paragraph is reconciled. Corrected in place with a pointer to the
  ADR.
• The bare-name resolution order is established by observation. Two independent observations,
  both against Claude Code 2.1.226 and both recorded in the ADR's Context: the Skill tool's own
  instruction text ("Plugin skills use plugin:skill.") plus the skill-name validator
  ("Skill names match the skill's directory name (or 'plugin:skill' for plugin-qualified
  skills)"), and a live phoenix session's roster carrying bare adr/report/triage
  alongside fabrika:adr/fabrika:report/fabrika:triage.
• Mechanical follow-ups are filed as separate issues. Retire the v1 skill roster in one change: both .claude symlinks plus the ADR 0077 suppression #5276 (the one-change retirement) and
  DEVELOPMENT.md pins 13 skill references to .claude/skills paths, so they resolve to v1 only #5275 (DEVELOPMENT.md's 13 path pins, which neither CLAUDE.md pins skill routing to a filesystem path, so a fabrika skill can never inherit routing #4761 nor The crew reaches its skills by hardcoded path, so the fabrika cutover is a 22-site path rewrite with no ticket #4762 scopes). Both named in
  the ADR's Records.

Notes for review

• The four names live in both sets today: adr, plan-epic, report, triage.
• ADR 0077 is confirmed, not superseded — its status and decision text are untouched. The ADR
  only adds the measured detail that the plugin toggle's scope was always the plugin copy, never the
  symlink, which is what 0077 intended.
• No claude-plugins/kampus-pipeline/ file is touched; v1 stays the frozen baseline (ADR 0238).
• Docs-only diff. pnpm lint:worktree is a clean skip (no biome-handled files changed);
  pnpm typecheck passes across all 31 workspace tasks.
• Relates to CLAUDE.md pins skill routing to a filesystem path, so a fabrika skill can never inherit routing #4761 and The crew reaches its skills by hardcoded path, so the fabrika cutover is a 22-site path rewrite with no ticket #4762 (the routing-pin half), and to write-code skill scripts are unreachable in an agent worktree — .claude/.pipeline symlink is not provisioned (exit 127 on the Step-4 preflight) #4818 / fix(pipeline): track .claude/.pipeline in git, and leave a breadcrumb when the hook fails (#4818) #4825 (the tracked-symlink
  precedent).
• Numbered 0255, not 0253. Three in-flight lanes each derived 0253 off the same
  origin/main max (0252). Ordered by PR creation time, docs(decisions): 0253 — the eval record is an eval-namespaced PR comment (#4769) #5271 keeps 0253, docs(decisions): 0254 — the ready-for: gap closes lazily at triage time, never by backfill (#5041) #5273 takes 0254,
  and this PR takes 0255. pipeline-cli decisions-index validate passes.

Deviations

• Out-of-scope change — Said: The .claude/skills symlink loads v1 skills regardless of the plugin toggle, so v1 and fabrika skills are model-invoked under one name today #4829's acceptance criteria name exactly one contract to
  reconcile, claude-plugins/fabrika/skills/report/contract.md. Did: also edited
  claude-plugins/fabrika/skills/triage/contract.md (two sentences). Why: that file had already
  independently falsified the report contract's premise and cited this issue as the open question;
  leaving it citing an open issue after the issue is settled, and leaving its "the stated mitigation
  is inert" line standing when the ADR concludes the mitigation is the right one, would put the
  corpus at odds with the ADR the same PR lands. Disposition: for the reviewer to judge — the
  edit is confined to the paragraph that names The .claude/skills symlink loads v1 skills regardless of the plugin toggle, so v1 and fabrika skills are model-invoked under one name today #4829 and adds no new claim beyond the ADR's.
• (repair round 1) Head moved on a PASS, not a FAIL — Said: repair mode acts on a
  current-head gate FAIL. Did: pushed a round while both namespaces stood at PASS/advisory,
  which staleness-invalidates the markers bound to 951a1fb. Why: the review-doc verdict
  recorded a non-blocking advisory that measurement 1 attributed the message "names must not
  contain ':' (reserved for plugin namespacing)" to the skill frontmatter validator, when both
  of its two occurrences in the shipped CLI are on the agent path (parseAgentFromMarkdown and
  the agent parse-error helper). An ADR that lands carrying a wrong attribution is far more
  expensive to correct later than now, and the PR is §CP and not yet approved. Disposition: the
  new head needs a fresh review-doc + review-skill pass before approval.
• (repair round 1) Narrower than the finding, plus one accepted optional note — Said: the
  advisory suggested a one-sentence swap of the cited mechanism. Did: that swap, re-grounded on
  two skill-side strings re-verified first-party against the same binary, plus the reviewer's
  second, optional note — softened "no bare name is ever contested" by one clause naming the
  shipped opt-in fallback frontmatter field. Why: the fallback qualifier cost one clause and
  no skill in this repo declares it, so the conclusion is untouched. Disposition: the ADR's
  conclusion, §1 naming rule, branch-2 decision and Banned list are unchanged; the branch was also
  rebased onto latest origin/main (clean).

[github-actions]

No preview deploy

• No preview deploy for this PR — its diff touches no deploy-relevant path, so no preview stack was minted and e2e is not applicable. _(c6fb167)

• web — Stage pr-5277 torn down.

[usirin]

review-doc: advisory — blocking-set PR (§CP — approval-gated)

Verified fresh against issue #4829's acceptance criteria, read-only from the PR head via a per-run ref.

Reviewed-head: @ 951a1fb

§CP classification: control-plane, decided on the content axis, not the path axis.
cp-classify classify returned content-undetermined (no CODEOWNERS path matched — CODEOWNERS,
read live from main, carries no /claude-plugins/fabrika/ row and no .decisions/ row).
guard-content-probe classify then returned guard-touching (§CP, ADR 0164)
[guard-vocabulary-match] on .decisions/0255-…. So the advisory form above, per §ADVISORY.

Class fan: class-probe classify --namespaces at this head names 2 namespaces —
review-doc, review-skill. Both are dispatched this pass, one marker per comment. No
UI-affecting path, so review-design is correctly not required.

Acceptance criteria

• [PASS] An ADR is recorded settling which branch is taken, naming the cutover sequencing rather
  than only the end state. .decisions/0255-skill-namespaces-keep-v1-and-fabrika-apart.md takes
  branch 2 explicitly ("This is the issue's branch 2, taken because branch 1 solves a problem that
  was not there"). §3 "Cutover sequencing — one lever, pulled once" gives the running-v1-routing
  sequence in three ordered steps (now / while the wave runs / at retirement), and bans the
  per-skill partial cut. Sequencing is named, not just the end state.

• [PASS] The decision explicitly covers .claude/agents, and states whether .claude/.pipeline
  from write-code skill scripts are unreachable in an agent worktree — .claude/.pipeline symlink is not provisioned (exit 127 on the Step-4 preflight) #4818 / PR fix(pipeline): track .claude/.pipeline in git, and leave a breadcrumb when the hook fails (#4818) #4825 is affected. §4 covers .claude/agents on its own terms: same symlink
  mechanism, retires on the same lever, and no agent-name collision today. I confirmed that last
  point directly — the fabrika plugin root carries skills/, docs/, hooks.json, README.md
  and .claude-plugin/plugin.json, and no agents/ directory. §5 confirms triage's read
  that .claude/.pipeline is unaffected, and adds the qualifier that makes the precedent safe to
  reuse (tracking a symlink is safe for a path alias, and is a roster decision for a load-path
  directory).

• [PASS] report/contract.md's "dormant only by configuration" paragraph is reconciled.
  Corrected in place and pointed at the ADR. The replacement states the toggle never reaches the
  symlink, that both sets are live, that the names never collide, and that the real overlap is the
  description. Relative link ../../../../.decisions/0255-… resolves to the repo root correctly.

• [PASS] Bare-name resolution order established by observation and recorded. §"What was
  measured, and how" records two observations and §1 records the resulting rule as a table. I
  re-verified the conclusion independently against the shipped CLI (2.1.226 — the version the ADR
  names) rather than taking it from the PR body; see the grounding note below. The conclusion holds.

• [PASS] Mechanical follow-ups filed as separate issues; this issue closes on the recorded
  decision, not the implementation. DEVELOPMENT.md pins 13 skill references to .claude/skills paths, so they resolve to v1 only #5275 (DEVELOPMENT.md path pins) and Retire the v1 skill roster in one change: both .claude symlinks plus the ADR 0077 suppression #5276 (one-change v1
  retirement) both exist and are open. §"Records" names both, plus the untouched CLAUDE.md pins skill routing to a filesystem path, so a fabrika skill can never inherit routing #4761/The crew reaches its skills by hardcoded path, so the fabrika cutover is a 22-site path rewrite with no ticket #4762
  routing-pin half.

Grounding check — the premise-is-false claim, verified against the shipped CLI

The ADR's whole branch-2 conclusion rests on "the premise is false: v1 and fabrika are never invoked
under one name." I verified this against Claude Code 2.1.226 directly, not from the PR body. The
conclusion is correct, on three independent pieces of shipped evidence:

1. The Skill tool's own instruction text states verbatim: "Plugin skills use plugin:skill."
2. The skill-name validator states verbatim that "Skill names match the skill's directory name (or
   'plugin:skill' for plugin-qualified skills)" — a filesystem skill's name is directory-derived,
   and the colon form is reserved for the plugin-qualified spelling.
3. The plugin-skill dedupe is scoped exactly as the ADR says. The shipped log line reads "Skipping
   plugin skill '…' — … is a user-level skill already surfaced by the skills directory loader",
   confirming the check is user-level-scoped and does not reach a project's .claude/skills.

The live setup also corroborates the ADR's framing independently of the PR: .claude/settings.json
carries {"fabrika@kampus": true, "kampus-pipeline@kampus": false}, .claude/skills and
.claude/agents are symlinks into the v1 tree, and the both-sets overlap is exactly the four names
the ADR claims — adr, plan-epic, report, triage — which I confirmed by intersecting the two
skill directories.

Advisory (non-blocking) — one citation in §"What was measured" is mis-scoped

Measurement 1 says "the frontmatter validator rejects a non-plugin skill whose name contains
a colon — the message is verbatim 'names must not contain ':' (reserved for plugin namespacing)'."
That message is real and verbatim in the shipped binary, but it occurs at exactly two code sites and
both are on the agent path, not the skill path: the agent-frontmatter parse-error helper
(Invalid "name": …) and parseAgentFromMarkdown (Agent file … has invalid name …). I found no
skill-side occurrence.

This does not falsify the finding — the skill-side guarantee is real, it is just a different and
arguably cleaner mechanism (evidence 1–2 above: directory-derived names, colon reserved for the
plugin-qualified form). So the conclusion, the naming rule in §1, and the branch-2 decision all
stand, and no acceptance criterion is unmet. But the sentence as written sends a future auditor
looking for a skill validator that does not carry that rule, which is the exact "confident wrong
answer" failure §1 of this ADR warns about. Worth a one-sentence correction swapping the cited
mechanism for evidence 1–2. Recorded here rather than blocking.

Second, smaller note: the shipped skill frontmatter carries an @internal fallback field whose
own description says a skill with it set "yields to a same-suffix plugin or MCP skill
(<plugin>:<name> / <server>:<name>) when one is loaded" — i.e. there is a shipped mechanism
by which a bare name can be shadowed by a same-suffix plugin skill. It is opt-in and off by default,
and no skill in this repo declares it (checked), so the ADR's conclusion is unaffected in practice.
It is a qualifier the absolute phrasing "no bare name is ever contested" does not carry.

ADR numbering — renumber verified clean at this head

• 0255 is uncontested. Resolved each open PR's claim by its actual changed .decisions/**
  filename, not by title (titles are stale): docs(decisions): 0253 — the eval record is an eval-namespaced PR comment (#4769) #5271 → 0253, docs(decisions): 0254 — the ready-for: gap closes lazily at triage time, never by backfill (#5041) #5273 → 0253 (still pending its
  renumber to 0254), docs(decisions): 0255 — skill namespaces keep v1 and fabrika apart (#4829) #5277 → 0255, docs(decisions): 0256 — the kill audit keys on the not-planned close, not the label (#4921) #5281 → 0256, docs(decisions): 0257 — a platform incident is eval material only through an artifact, at the deterministic tier (#4824) #5282 → 0257. Main's highest is 0252.
  No other open PR claims 0255.
• No dangling 0253 reference survives. git grep '\b0253\b' over the entire tree at this
  head returns zero hits — not just in .decisions/ and claude-plugins/fabrika/. Every ADR
  reference in the two contract files points at 0255 (plus pre-existing 0251 links), and all
  four ADR links inside the new file (0077, 0238, 0246, 0249) resolve to files that exist at
  this head.
• CI's own validate ADR files — no duplicate/mismatched ADR number check is green at this head,
  independently.
• Note, non-blocking and not a repo artifact: the PR title and branch name still say 0253
  (docs(decisions): 0253 — v1 and fabrika skills coexist…, branch
  usirin/adr-0253-…). Nothing in the tree is stale, so this does not gate the merge, but the
  squash-merge commit subject will carry the wrong number unless the title is edited before merge.
  Recommend retitling to 0255 first.

Hygiene

• No home / absolute / sibling-repo paths in the committed diff, the PR body, or the comments on
  either docs(decisions): 0255 — skill namespaces keep v1 and fabrika apart (#4829) #5277 or The .claude/skills symlink loads v1 skills regardless of the plugin toggle, so v1 and fabrika skills are model-invoked under one name today #4829 — checked all four surfaces specifically.
• ADR frontmatter (id/title/status/date/tags) is well-formed and id: 0255 matches the
  filename.
• All CI checks green at this head (0 failures).

Verified read-only from a per-run ref at 951a1fb6d38f99c2b59f0fccf8c2ec7d1fc511cc; no working
tree was moved.

[usirin]

review-skill: advisory — blocking-set PR (§CP — approval-gated)

Behavioral-artifact gate over the two changed fabrika contracts. Verified read-only from the PR head
via a per-run ref.

Reviewed-head: @ 951a1fb

§CP classification: control-plane by content, not path. CODEOWNERS (read live from main)
has no /claude-plugins/fabrika/ row, so the two contract files are not path-owned;
guard-content-probe classified the sibling .decisions/0255-… file as guard-touching (ADR 0164),
which puts the PR in the blocking set. Hence the advisory form.

Why this gate is required: class-probe classify --namespaces at this head names review-doc
and review-skill. The skills class fires on HAS_SKILLS_RE matching
claude-plugins/fabrika/skills/**. Both markers are posted this pass, one per comment.

Scope of this gate

Two files, both contract.md — authoring contracts, not SKILL.md. No skill frontmatter, name,
description, or trigger surface changed in this PR, so nothing here can alter what a skill is
invoked for. That keeps the skill-hygiene surface narrow: what is under review is whether the prose
these contracts commit fabrika's authors to is now true.

• [PASS] report/contract.md — the falsified premise is corrected, not just re-pointed. The old
  paragraph asserted the collision was "dormant only by configuration" because
  "kampus-pipeline@kampus": false. The replacement states the mechanism correctly: the symlink
  loads v1 project-level so the toggle never reaches it, both sets are live, the names never
  collide, and the real overlap is the description — which is what a model actually invokes off.
  This matches what I verified against the shipped CLI (see the review-doc verdict for the
  grounding detail). It correctly hands the "does the differentiation actually fire" question to the
  eval set per ADR 0249 rather than asserting it, and preserves the frozen-v1-baseline stance of
  ADR 0238. Relative ADR link resolves.

• [PASS] triage/contract.md — corrected on both counts, and it was already the more accurate of
  the two. Note in the ADR's favour: this file's pre-existing text on main already stated that
  a live roster carries both adr/fabrika:adr and report/fabrika:report. That is a committed
  observation predating this PR, so the ADR's roster claim has corroboration independent of the PR
  body. The edit makes the two changes the author disclosed: #4829 is no longer described as an
  open question (now "Settled by ADR 0255"), and the mitigation is no longer called "inert" (now
  "the right one, and it has to be load-bearing here"). Both were flatly contradicted by the ADR
  landing in the same commit.

• [PASS] No behavioral drift. Neither edit changes a verb, an invariant, a gate, or a
  fail-closed rule. The triage edit keeps the "model-invoked deliberately" requirement and the
  two non-optional consequences that follow it intact — it only rewrites the justification
  paragraph above them. Diff sizes (11+/7− and 8+/4−) are consistent with a prose-only reconcile.

Ruling on the disclosed class-7 deviation — the triage/contract.md edit is IN SCOPE

The author flagged that #4829's acceptance criteria name only report/contract.md, and that they
edited triage/contract.md too. Accepted as in-scope. Three reasons:

1. It is the same obligation the criteria already impose, applied to the one other file carrying
   the same falsified claim. AC3 exists because report/contract.md recorded a premise the ADR
   disproves. triage/contract.md recorded the same disproved premise plus a second claim the ADR
   contradicts ("its stated mitigation is inert" vs. the ADR's "the premise was wrong; the
   mitigation was right"). Reconciling one and not the other satisfies the letter of AC3 while
   leaving the repo self-contradicting.
2. The ADR's own Banned list forbids "Recording anywhere that the plugin toggle gates what the
   symlink loads, or that the bare name is contested." Leaving the stale triage paragraph would
   land the ADR and a violation of its own banned list in the same commit.
3. The edit is minimal and bounded — one paragraph, no new claims, no verb or invariant touched.

This is the correct handling of a class-7 deviation: disclosed rather than silent, and narrowly
scoped to what the primary change makes false. No further scope was taken; the routing-pin half
(#4761/#4762) and the retirement work (#5276) were correctly left filed rather than done here.

Hygiene

• No home / absolute / sibling-repo paths in either contract file, the diff, the PR body, or the
  comments on docs(decisions): 0255 — skill namespaces keep v1 and fabrika apart (#4829) #5277 or The .claude/skills symlink loads v1 skills regardless of the plugin toggle, so v1 and fabrika skills are model-invoked under one name today #4829.
• Both ../../../../.decisions/0255-… links resolve to the repo root from
  claude-plugins/fabrika/skills/<skill>/, and the target file exists at this head.
• No dangling 0253 reference anywhere in the tree at this head.
• validate skill frontmatter and the skill-corpus lint checks are green at this head.

Verified read-only from a per-run ref at 951a1fb6d38f99c2b59f0fccf8c2ec7d1fc511cc; no working
tree was moved.

[usirin]

review-doc: advisory — blocking-set PR (§CP — approval-gated)

Re-review at the new head after repair round 1. The prior verdicts (bound to 951a1fb6) are
staleness-invalidated per ADR 0058 and are records, not current. Everything below is re-derived
first-party at this head, read-only from a per-run ref.

Reviewed-head: @ c6fb167

§CP classification — re-derived on BOTH axes at this head. Control plane, decided on the
content axis; the path axis is clear.
cp-classify classify → content-undetermined [content-source-present] (no path matched
CONTROL_PLANE_RE, re-resolved live from main; the two changed files under
claude-plugins/fabrika/skills/ are not kampus-pipeline, and .decisions/** is not path-§CP).
guard-content-probe classify on .decisions/0255-… at this head → guard-touching (§CP, ADR
0164) [guard-vocabulary-match]. Hence the advisory form (§ADVISORY): this verdict does not
authorize a merge — a @kamp-us/control-plane approval at head does, and ship-it then enqueues
(ADR 0135/0048).

Class fan — re-run at this head, not reused. class-probe classify --namespaces over the
3-file changed set names 2 namespaces: review-doc, review-skill. Both are dispatched this
pass, one marker per comment. No UI-affecting path, so review-design is correctly not named.

Acceptance criteria (#4829)

• [PASS] An ADR settles which branch is taken, naming the cutover sequencing. ADR 0255 takes
  branch 2 explicitly; §2 gives why the path is not cut now (three ordered reasons) and §3 gives the
  three-step sequence ending in one edit, plus the ban on a per-skill partial cut. Unchanged by the
  repair round.

• [PASS] Covers .claude/agents; states whether .claude/.pipeline is affected. §4 covers
  .claude/agents (same mechanism, same lever, no agent-name collision today — re-confirmed: the
  fabrika plugin root carries no agents/ directory). §5 confirms triage's read that
  .claude/.pipeline is unaffected and adds the path-alias-vs-load-path qualifier. Unchanged.

• [PASS] report/contract.md's "dormant only by configuration" paragraph is reconciled.
  Corrected in place and pointed at the ADR; the relative link ../../../../.decisions/0255-…
  resolves to repo root from claude-plugins/fabrika/skills/report/ and the target exists at this
  head. Unchanged by the repair round.

• [PASS] Bare-name resolution order established by observation, and now correctly attributed.
  This is the criterion the repair touched, so I re-verified it against the shipped CLI myself
  rather than reading the diff. See the grounding section below — all three claims are true at
  source.

• [PASS] Mechanical follow-ups filed as separate issues. DEVELOPMENT.md pins 13 skill references to .claude/skills paths, so they resolve to v1 only #5275 and Retire the v1 skill roster in one change: both .claude symlinks plus the ADR 0077 suppression #5276 both exist, are open,
  and are named in §Records. Unchanged.

Grounding — is the CORRECTED attribution actually TRUE at source?

The whole point of the repair was factual accuracy, so I re-grounded every claim in the rewritten
paragraph against the shipped Claude Code 2.1.226 binary (the version the ADR names), matching
byte offsets in its JS source region and reading the surrounding function bodies. Verdict:
factually true.

1. "the Skill tool's own instruction text says verbatim 'Plugin skills use plugin:skill.'" —
   TRUE. One occurrence, inside the Skill tool's description of its skill parameter (the block
   that also documents "exact name from the listing, no leading slash"). The stored form carries JS
   template-literal escaping on the backticks; the rendered text is exactly as quoted.

2. "the skill-name validator says verbatim 'Skill names match the skill's directory name (or
   'plugin:skill' for plugin-qualified skills)'" — TRUE. It sits in the skill-name validation
   function (the one that also rejects unpaired surrogates, surrounding whitespace, *,
   wildcard-suffix names and a leading /), as the explanatory clause of its
   parentheses/commas/control-characters rejection. Correctly attributed to a skill-side
   validator.

3. "that message occurs at exactly two sites in the shipped binary and both are on the agent path
   (parseAgentFromMarkdown and the agent parse-error helper)" — TRUE. Four raw matches of the
   phrase, of which two are the compiled-snapshot copies of the same two source strings; the two
   source sites are (a) the frontmatter parse-error helper, which returns the
   Invalid "name": … string and whose only call site is the agent loader (it logs
   Failed to parse agent from … and fires the tengu_agent_parse_error metric), and (b) the
   agent-from-markdown parser, which emits Agent file … has invalid name '…'. No skill-side
   occurrence exists. The redirect the ADR now gives the reader is correct.

Also re-verified, for the softening clause: the fallback frontmatter field is real (it is in the
shipped known-frontmatter-key list and in the parsed-frontmatter default object), its own shipped
description says a skill carrying it "yields to a same-suffix plugin or MCP skill … when one is
loaded", and git grep '^fallback:' over the whole tree at this head returns zero hits. So
"which no skill in this repo declares" holds, and the hedge is accurate rather than decorative.

• [PASS] Advisory from the prior round is discharged. The prior head's non-blocking advisory
  asked for exactly this swap. It was made, and made correctly.

Advisory (non-blocking, no action required) — one residual imprecision

The rewritten sentence says the colon spelling "is reserved for the plugin-qualified form." The
same Skill-tool instruction text I verified for claim 1 continues: "Directory-scoped skills are
listed with a path prefix (apps/web:deploy)" — so a colon also appears in the directory-scoped
listing form, which is not plugin-qualified. This does not touch the conclusion: v1 loads from the
repo-root .claude/skills, which is unscoped, so its listed names are bare and disjoint from
fabrika:* exactly as the ADR says. And the sentence is a faithful reading of the validator string
it cites, which itself names only the plugin-qualified form. Recorded so a future auditor who meets
apps/web:deploy is not surprised; not worth a head move.

ADR numbering + dangling-reference check, re-run at this head

• [PASS] 0255 is uncontested. Fresh origin/main carries .decisions/0250, 0251, 0252 and
  nothing higher. Among open PRs, docs(decisions): 0253 — the eval record is an eval-namespaced PR comment (#4769) #5271 and docs(decisions): 0254 — the ready-for: gap closes lazily at triage time, never by backfill (#5041) #5273 contest 0253 between themselves, docs(decisions): 0256 — the kill audit keys on the not-planned close, not the label (#4921) #5281 claims
  0256; no other open PR claims 0255.
• [PASS] No dangling 0253 reference survives the rebase. git grep '0253' over the whole tree
  at this head returns zero hits. Every ADR cross-link inside the new file (0077, 0238,
  0246, 0249) resolves to a file present on fresh origin/main.
• [PASS] Title is correct. The PR title reads docs(decisions): 0255 — skill namespaces keep v1 and fabrika apart (#4829), so the squash subject will carry 0255. The branch name still says
  0253; that is deliberate and never reaches the squash message.

Doc hygiene

• [PASS] House format. ADR frontmatter carries id/title/status/date/tags, id: 0255
  matches the filename, and the body has ## Context / ## Decision / ## Consequences.
• [PASS] Index row. No .decisions/index.md is committed (correct — discovery is ambient, ADR
  0126). Frontmatter status: accepted is the source the ambient map renders.
• [PASS] Links resolve. All in-repo links are standard markdown relative paths; every target
  exists at this head or on fresh origin/main. No wikilinks.
• [PASS] No leaked local/home paths. The shared matcher over the diff's added lines returns
  clean.
• [PASS] Supersession. Nothing is superseded — ADR 0077 is explicitly confirmed, its status and
  decision text untouched, so no forward-pointer is owed.
• [PASS] Status sanity. accepted is in the house vocabulary and is coherent with an entry that
  settles a live fork.
• [PASS] Single Diátaxis mode. Explanation throughout; the tables and the numbered sequence serve
  the reasoning rather than turning the page into a how-to.
• [PASS] Clear, concise prose. The rewritten paragraph is direct and the added redirect earns its
  length — it is the whole point of the round. No AI-tell density.
• [PASS] deviation-disclosure (§DEV). The body carries a literal ## Deviations heading with
  three entries, two tagged (repair round 1), each with substantive Said / Did / Why / Disposition.
  The head-moved-on-a-PASS entry and the narrower-than-the-finding entry both match what the diff
  actually did, including the rebase. Nothing undisclosed that this gate could see.

CI

All checks green at this head — 23 success, 7 skipped, 0 failures.

Verified read-only from a per-run ref at c6fb167effc717dce000849f026fe2e738db4f5f; no working tree
was moved, and no prior verdict was edited or removed.

[usirin]

review-skill: advisory — blocking-set PR (§CP — approval-gated)

Re-review at the new head after repair round 1. The prior review-skill verdict (bound to
951a1fb6) is staleness-invalidated per ADR 0058. Re-derived first-party at this head, read-only
from a per-run ref.

Reviewed-head: @ c6fb167

§CP classification — re-derived on BOTH axes at this head. Control plane by content, not by
path. cp-classify classify → content-undetermined (nothing matched CONTROL_PLANE_RE,
re-resolved live from main; claude-plugins/fabrika/skills/** is a non-kampus-pipeline
plugin, which the §CP set names as non-blocking). guard-content-probe classify on the sibling
.decisions/0255-… at this head → guard-touching (§CP, ADR 0164), which puts the whole PR in the
blocking set. Hence the advisory form: this verdict does not authorize a merge.

Why this gate is required. class-probe classify --namespaces re-run at this head names
review-doc and review-skill — the skills class fires on HAS_SKILLS_RE matching
claude-plugins/fabrika/skills/**. Both markers posted this pass, one namespace per comment.

What actually changed since the reviewed head

The repair round touched only .decisions/0255-… (10 insertions, 5 deletions) plus a clean
rebase onto latest origin/main. git diff 951a1fb6..c6fb167e -- claude-plugins/ is empty — the
two contracts are byte-identical to the head that already passed this gate. I re-verified them
against fresh origin/main rather than trusting that, since a rebase can change what a diff means.

Skill-rigor checklist

• [PASS] Behavioral correctness. Both changed files are contract.md authoring contracts, not
  SKILL.md. No frontmatter, name, description, trigger surface, verb, or procedure step
  changed — the edits are confined to the premise paragraph each file recorded about the v1/fabrika
  collision. Traced as an executing agent would: nothing in either file's procedure reads
  differently after the edit.

• [PASS] report/contract.md — the falsified premise is corrected, not merely re-pointed. The
  replacement states the mechanism correctly (the symlink loads v1 project-level so the toggle never
  reaches it; both sets are live; the names never collide; the real overlap is the description,
  which is what a model invokes off), hands the does-it-actually-fire question to the eval set per
  ADR 0249 instead of asserting it, and preserves the frozen-v1-baseline stance of ADR 0238. This is
  consistent with what I re-verified against the shipped CLI at this head (grounding detail in the
  review-doc verdict).

• [PASS] triage/contract.md — corrected on both counts. #4829 is no longer described as an
  open question ("Settled by ADR 0255"), and the mitigation is no longer called "inert" (now "the
  right one, and it has to be load-bearing here"). Both were flatly contradicted by the ADR landing
  in the same commit. The "model-invoked deliberately" requirement and the two non-optional
  consequences below it are intact — only the justification paragraph above them was rewritten.

• [PASS] Trigger / description quality. Out of reach: no description line and no frontmatter
  is in the diff, so nothing here can change what either skill fires on.

• [PASS] Cross-skill conflict / shadowing. No shared contract, marker namespace, §CP row, or
  seam another skill reads is touched. Both files now agree with each other and with ADR 0255 — the
  edit removes a corpus-level contradiction rather than creating one, and it keeps the corpus off
  the ADR's own Banned list ("recording anywhere that the plugin toggle gates what the symlink
  loads, or that the bare name is contested").

• [PASS] Gate-invariant preservation. No gate invariant is in this diff's reach: nothing under
  claude-plugins/kampus-pipeline/, no ship-it refusal, no matcher, no SHA-staleness rule, no
  author-gate, no config-pin. The two contracts govern fabrika authoring, not merge authority.

• [PASS] Contract implementation ticket (ADR 0248). In scope by the file list (the diff changes
  two claude-plugins/fabrika/skills/*/contract.md), and discharged: the verb inventories are
  untouched by this PR — git diff origin/main...HEAD over both files shows no row added, removed
  or changed in either ## Verb inventory table — and each contract's verbs already carry their
  implementation ticket from its authoring lane (Implement the three /report fabrika verbs — the merged contract has no implementation ticket #4748 for the three /report verbs, Implement the nine fabrika triage verbs specified by the /triage derived contract #4831 for the
  nine /triage verbs, both closed as implemented). This PR introduces no unbuilt verb, so there is
  no new build-pool obligation to name.

• [PASS] deviation-disclosure (§DEV). The body carries a literal ## Deviations heading with
  three entries, two tagged (repair round 1), each with substantive Said / Did / Why / Disposition.
  The pre-existing class-7 entry (editing triage/contract.md beyond the AC's named file) still
  stands and is still correctly ruled in scope — same obligation, one other file carrying the same
  falsified claim, and the ADR's Banned list would otherwise be violated by the same commit that
  lands it. The two new entries accurately describe a PASS-to-PASS head move and its narrow content.
  Nothing undisclosed that this gate could see.

Specialist fan-out (skill-class dimensions)

Ran over the changed contract text at this head: no unreachable step, no contradictory
instruction, and no uncovered procedural path introduced. Nothing to route; the linked issue's
acceptance-criteria list is unchanged by this gate.

Hygiene

• [PASS] No home / absolute / sibling-repo paths in the diff's added lines (shared matcher, clean).
• [PASS] Both ../../../../.decisions/0255-… links resolve to repo root from
  claude-plugins/fabrika/skills/<skill>/, and the target file exists at this head.
• [PASS] No dangling 0253 reference anywhere in the tree at this head (git grep returns zero).
• [PASS] All CI checks green at this head — 23 success, 7 skipped, 0 failures.

Verified read-only from a per-run ref at c6fb167effc717dce000849f026fe2e738db4f5f; no working tree
was moved, and no prior verdict was edited or removed.