Current state, 2026-09-15. This body is the record; the thread below it is history.
The defect
Both archaeology guards read an all-numeric hex colour in a durable comment as a ticket reference. /#\d{4,}\b/ matches #000000, #111111 and #123456; the trailing \b only spares hex that contains a letter, such as #1234ff.
the pre-commit hook on staged files, and ticket-archaeology-lint.yml with --base origin/dev
fires
published, neo-agent-skills@0.1.4
pr-baseline.yml's archaeology job, over every changed file in full
fires
What changed since filing
The escape no longer lies. The original harm was that the only way past was ticket-ref-ok, which asserts a deliberate ticket reference where there is only a colour. Both guards now accept a typed escape, [not-ticket-ref: css-color], scoped to the colour token (#18433). The two affected lines use it: src/component/Gallery.mjs:33 and src/component/Helix.mjs:29.
The annotation itself remains: every all-numeric colour in a durable comment still needs one.
Why the obvious fixes were not taken
attempt
outcome
exempt a match in value position (quoted, or after =), PR #16968
closed: a quoted real ref such as 'see #16538' goes silent, and that false negative is worse than the false positive
exempt quoted matches of CSS hex length (3, 4, 6 or 8 digits)
still leaks: #9473 is a real 4-digit ticket, and 6-digit ticket numbers are coming
dropped: it would fork the policy the published guard owns
drop the digit bound to \d+ so short cross-repo refs are caught
the audit went from 9 findings to 27, nearly all ordinals like Refresh #2; short refs are only decidable when repo-qualified
The fix, in order
The published guard: neomjs/neo-agent-skills#69, carried by neomjs/neo-agent-skills#68. A 3-, 4-, 6- or 8-digit number directly after colour syntax (color:, backgroundColor_=, fillStyle=, CSS color) is a colour, and a number with a leading zero is never a ticket. Every other bare number still reports, so both counterexamples above still fail: 'see #16538' and #9473. The typed escape stays valid on a colour, so nothing reds while it is in place.
A release, then a bump of neo-agent-skills here.
The engine guard takes the same colour-context rule and the entity rule already written on agent/16553-engine-entity-guard, and the two escapes come out, in the PR that resolves this ticket. Removing an escape before the bump reds the baseline job.
AC-1's controls are read in colour syntax. A bare #123456 in prose stays a ref, because nothing distinguishes it from a six-digit ticket (the second row above).
A second false positive of the same family, found 2026-09-15 by @neo-opus-vega on #18729: a numeric HTML entity. A comment describing rendered markup carries ' or —, and the digits are a codepoint. It needs no release ordering, because the discriminator is syntactic — &#\d+;, with the & immediately before the # — so no typed marker and no context inference. Measured, the two guards trip on different entities, which is why both copies carry it: the published #\d+ sees ' while the engine's #\d{4,} cannot, but the engine sees — and ’ — the em dash and right quote such a comment carries most often. Published: neomjs/neo-agent-skills#68 at b1fd043. Engine: written and green on agent/16553-engine-entity-guard at 68e8c4879b — withHtmlEntitiesRemoved blanking &#\d+; before the patterns run, five arms, mutation red — and held rather than shipped separately. @tobiu's call (2026-09-15): a PR must carry Resolves #N, this ticket cannot close until its colour half lands, and the entity half does not earn a leaf of its own. It ships inside step 3 below. Until then the hook still reads — and ’ as tickets: rephrase the comment rather than embedding the entity.
Found on the same guards
The engine guard's audit mode fails in a clean checkout.DEFAULT_SCAN_PATHS still lists ai, which moved to the Brain repo, so find stops with ai: No such file or directory. Staged-file and --base runs are unaffected.
The published guard rejects the legacy marker. It classifies ticket-ref-ok as invalid-escape, while the engine guard still honours it as the escape for a deliberate, load-bearing ref. dev carries 134 such markers in 73 files, and touching any of those files reds the baseline job. Whether a deliberate ticket ref may stay in a durable comment is a policy question for both repositories, not a regex fix.
Delegating the engine guard to the published one is blocked upstream: neo-agent-skills exports only ./manifest and ./package.json.
Acceptance Criteria
An all-numeric hex colour in a durable comment passes both guards without an annotation, and a real ref still fails in prose and inside a quoted string. Controls, in colour syntax: backgroundColor_='#123456' and borderColor="#111111" pass — six digits with no leading zero and no letter, so syntax is the only thing that tells them from a ticket. Bare: #000000 and #1234ff pass, on the leading zero and the letter. Failing: #16538, see #16538, 'see #16538' and #9473.
The fix lands in the published guard first and the engine guard follows, so the two cannot disagree on the same line.
The two [not-ticket-ref: css-color] escapes are removed once they are no longer needed.
A numeric HTML entity in a durable comment passes both guards. Controls per guard, because their reach differs: published — ' quoted in markup and a bare —; engine — — and ’. Both: a real ref beside an entity still reports, the hex form ' is unaffected, and the same digits without entity syntax stay a ticket. This AC does not wait for the release; the engine copy is neo's own code.
Related
#18433 (the typed escape in the engine guard) · PR #16968, PR #17880 (closed attempts) · neomjs/neo-agent-skills#69 and neomjs/neo-agent-skills#68 (the published guard and its fix)
tobiu referenced in commit b28ac0b - "fix(buildScripts): a hex colour in value position is not a ticket ref (#16553) on Aug 21, 2026, 2:31 AM
tobiu referenced in commit 525b999 - "fix(build): the engine archaeology guard reads a numeric HTML entity as a codepoint (#16553) (#18734) on Sep 15, 2026, 3:30 PM
The defect
Both archaeology guards read an all-numeric hex colour in a durable comment as a ticket reference.
/#\d{4,}\b/matches#000000,#111111and#123456; the trailing\bonly spares hex that contains a letter, such as#1234ff.backgroundColor_='#000000'buildScripts/util/check-ticket-archaeology.mjsticket-archaeology-lint.ymlwith--base origin/devneo-agent-skills@0.1.4pr-baseline.yml's archaeology job, over every changed file in fullWhat changed since filing
ticket-ref-ok, which asserts a deliberate ticket reference where there is only a colour. Both guards now accept a typed escape,[not-ticket-ref: css-color], scoped to the colour token (#18433). The two affected lines use it:src/component/Gallery.mjs:33andsrc/component/Helix.mjs:29.Why the obvious fixes were not taken
=), PR #16968'see #16538'goes silent, and that false negative is worse than the false positive#9473is a real 4-digit ticket, and 6-digit ticket numbers are coming\d+so short cross-repo refs are caughtRefresh #2; short refs are only decidable when repo-qualifiedThe fix, in order
color:,backgroundColor_=,fillStyle=,CSS color) is a colour, and a number with a leading zero is never a ticket. Every other bare number still reports, so both counterexamples above still fail:'see #16538'and#9473. The typed escape stays valid on a colour, so nothing reds while it is in place.neo-agent-skillshere.agent/16553-engine-entity-guard, and the two escapes come out, in the PR that resolves this ticket. Removing an escape before the bump reds the baseline job.AC-1's controls are read in colour syntax. A bare
#123456in prose stays a ref, because nothing distinguishes it from a six-digit ticket (the second row above).A second false positive of the same family, found 2026-09-15 by @neo-opus-vega on #18729: a numeric HTML entity. A comment describing rendered markup carries
'or—, and the digits are a codepoint. It needs no release ordering, because the discriminator is syntactic —&#\d+;, with the&immediately before the#— so no typed marker and no context inference. Measured, the two guards trip on different entities, which is why both copies carry it: the published#\d+sees'while the engine's#\d{4,}cannot, but the engine sees—and’— the em dash and right quote such a comment carries most often. Published: neomjs/neo-agent-skills#68 atb1fd043. Engine: written and green onagent/16553-engine-entity-guardat68e8c4879b—withHtmlEntitiesRemovedblanking&#\d+;before the patterns run, five arms, mutation red — and held rather than shipped separately. @tobiu's call (2026-09-15): a PR must carryResolves #N, this ticket cannot close until its colour half lands, and the entity half does not earn a leaf of its own. It ships inside step 3 below. Until then the hook still reads—and’as tickets: rephrase the comment rather than embedding the entity.Found on the same guards
DEFAULT_SCAN_PATHSstill listsai, which moved to the Brain repo, sofindstops withai: No such file or directory. Staged-file and--baseruns are unaffected.ticket-ref-okasinvalid-escape, while the engine guard still honours it as the escape for a deliberate, load-bearing ref.devcarries 134 such markers in 73 files, and touching any of those files reds the baseline job. Whether a deliberate ticket ref may stay in a durable comment is a policy question for both repositories, not a regex fix.neo-agent-skillsexports only./manifestand./package.json.Acceptance Criteria
backgroundColor_='#123456'andborderColor="#111111"pass — six digits with no leading zero and no letter, so syntax is the only thing that tells them from a ticket. Bare:#000000and#1234ffpass, on the leading zero and the letter. Failing:#16538,see #16538,'see #16538'and#9473.[not-ticket-ref: css-color]escapes are removed once they are no longer needed.'quoted in markup and a bare—; engine ——and’. Both: a real ref beside an entity still reports, the hex form'is unaffected, and the same digits without entity syntax stay a ticket. This AC does not wait for the release; the engine copy is neo's own code.Related
#18433 (the typed escape in the engine guard) · PR #16968, PR #17880 (closed attempts) · neomjs/neo-agent-skills#69 and neomjs/neo-agent-skills#68 (the published guard and its fix)