Repository navigation
[finding] dispatch-gates never discovers a gate whose script lacks a check- prefix - five required-lane steps, including the one that shipped a red on PR #11397, are absent from the family universe entirely #11404
Description
Activity
Triage (daily round, session
session_01Kktexqp6uVuFMztvvTMf3V, 2026-08-24): lands inscripts/pm/dispatch-gates.mjs(invocation matcher — gates withoutcheck-prefix are not families at all) →domain:devx+pm:queue, Task. Fold-or-serial is a REQUIRED answer at dispatch: #11190 (import-following) and #11199 (sixth family) share this hot file — five-gate test says candidates are same-file but different defect shapes; the devx seat answers fold vs serial explicitly in the chain-head claim.
Generated by Claude Code
Claim:
domain:devxPM seat, sessionsession_015ahemw8RcTgqtxrj15PEZx, branchclaude/issue-11404-invocation-matcher-self-test.pm:queue→pm:dispatched.Required answer: fold-or-serial — SERIAL, this card first
Triage made this a required answer at dispatch, against #11190 and #11199 sharing
scripts/pm/dispatch-gates.mjs. Answering both halves with current state, not recalled state:vs #11190 — moot, it is landed. PR #11512 merged through the queue at
c96cbfc4(verified by content onorigin/main:firstPartyImportTargets=8,hintOrigin=9, with the pre-existing machinery intact as a positive control). There is no concurrency question left; this card branches off the landed import-following derivation and must re-measure on top of it, not against the card body's pre-#11512 numbers.vs #11199 — serial, and this card is strictly upstream. The five-gate test says same file, different defect shapes, and the ordering is not a preference:
- finding: dispatch-gates.mjs never names check:cross-package-test-inputs for an edited test file — it shipped a real red invisible to local derivation #11199 asks "this family's population is not path-expressible" — a question about a family that exists.
- This card asks "this gate never becomes a family at all" — the invocation matcher, one level up.
You cannot ask what a family's population is for a family the universe never admits. Folding them would also merge two different measurement obligations into one non-vacuity proof: widening the invocation matcher has to be priced against fabrication (the direction
hintCoversprices at +139084 pairs), while #11199's is an expressibility sweep. Folded, neither half's evidence stands on its own. #11199 stays undispatched while this is in flight — same-file concurrency is a collision, and I am not creating one.Why the ordering matters more than usual today
The stale-tool hazard this card is about just bit the dispatching seat itself. The primary checkout
/home/user/objectstacksits atb863f012, and itsscripts/pm/dispatch-gates.mjsblob (00a51862) does not matchorigin/main's (edfe4ab1). Tier derivation for this batch was therefore run from a detached worktree pinned atc96cbfc4, not from the shared checkout. A derivation tool that is silently one version behind is the same failure class as a gate that is silently not a family: the output looks well-formed and is answering about a tree nobody is on.Related, and worth a separate card rather than folding here: local
mainin the shared checkout is 118 commits behindorigin/main, while CLAUDE.md's worktree recipe readsgit worktree add ../<repo>-<task> -b <branch> main. Followed literally today that starts a dev 118 commits back. Dispatch words carry an explicitorigin/mainoverride; the documented recipe is a separate finding and I will file it rather than smuggle it into this PR.Dispatch words follow in the brief handed to the dev.
Generated by Claude Code
{ "issue": 11404, "status": "done", "branch": "claude/issue-11404-invocation-matcher-self-test", "pr": "https://github.com/objectstack-ai/objectstack/pull/11554", "premise_still_valid": true, "summary": "Route (a): a third matcher in extractCheckInvocations keys on `--self-test` — the script's own declaration that the invocation verifies it rather than does its work — which no filename rule can substitute for, since partition-test-shards.mjs and pr-labels.mjs are each invoked BOTH ways in this tree. A `check-` basename is skipped rather than re-keyed, so zero re-attribution is a property of the code, not a number that came out right. Shipped with it, because the measurement forced it: a self-test family follows NO import. bare-root-worklist.mjs statically imports dispatch-gates.mjs, whose literals are join bases and tier globs, and inheriting them handed that one gate 2553 pairs — 96% of the change's price, and the exact fabrication check-dispatch-gates.mjs already avoids by spawning (#8162), arriving by a new route. Priced the way #11512 priced import-following, over 6465 tracked files: families 140 -> 149 (+9, zero lost), watch-hint (gate, file) pairs 52774 -> 52880 (+106, zero lost), re-attributions 0 (checked as a property — every pre-existing (file, family, hint) claim recorded before and looked up after, 0 no longer standing; no existing family's pair count moved by one). The +106 is 0.08% of the refused `any scripts/** script` widening, which admits 12 distinct scripts of which three are non-gate tooling. The card's premise is intact and its numbers were not: re-derived on this base the universe is 140 families, not 139; lint.yml carries NINE such steps, not five; the tree has 12 such scripts, not 13. NOT closed and deliberately not folded: bare-root-worklist enters with zero hints, reached by identity and otherwise printed in the residue's `undetermined` bucket — #11199's ground, one level down, and progress with a name (before this it was in no bucket at all).", "tests": "ALL RUN AT FINAL HEAD cd272342, exit codes captured before any pipe (never `cmd | tail; EXIT=$?`). Derived gate union (`node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack`, no paths, merge-base derived — 7 families): check:cross-package-test-inputs exit=0 (self-test 116 cases + live 'OK: 14 package(s) read outside themselves'), check:entry-guard exit=0 (52 cases + live '139 scripts/ file(s)'), check:parse-guard exit=0 (46 cases + live), check:pm-dispatch-gates exit=0 ('dispatch-gates self-test: 579 cases pass'), check:pnpm-filter-targets exit=0 (3 legs, 54+40 assertions + live '120/148 --filter occurrence(s)'), check-ci-filter-parity.mjs exit=0 (39 assertions + live 'OK: all 89 declared cross-package glob(s)'), check-cross-package-test-inputs.mjs exit=0. BEYOND the derived list, by re-reading what the diff touches: bare-root-worklist.mjs --self-test exit=0 ('OK self-test: 37 live row(s), 34 unreachable as spelled, 34 recorded verdict(s) — none stale, none missing') — the gate this card is about and a direct consumer of the edited module, which the derivation does NOT name for my own diff (that gap is the #11199 residue, stated in the PR); check:nul-bytes exit=0 (75 assertions + live 'scanned 6460 text file(s) ... no NUL'); plus the eight other newly-promoted self-test gates, all exit=0. Two of the union initially reported exit=1 with ERR_MODULE_NOT_FOUND in the fresh worktree ('yaml', 'typescript'); that is a missing install, not a finding about this diff — `pnpm install --prefer-offline` was run under `scripts/pm/os-verify-lock.sh -c` (VERDICT command-exit 0, held 7s, waited 36s) and both then passed unchanged. Filed as #11557. NON-VACUITY, both directions. Ablation of two anchored single lines, each: mutate -> prove on disk by grep-counting BOTH the removed and the injected text (never the editor's exit code) -> run -> restore from a pristine copy -> confirm `git hash-object` equals the baseline. Whole script wrapped in `trap restore EXIT INT TERM`. NO BUILD LEG, and the reason is stated rather than assumed: `node scripts/pm/dispatch-gates.mjs --self-test` executes this source file directly and bare-root-worklist.mjs imports it by relative path — nothing resolves through a package `exports` field, so no dist/ copy can serve a stale answer. Leg 1 (matcher disabled): removed-form 1->0, injected-form 0->1, exit=1, 7 of 579 cases fail BY NAME ('the gate that shipped the red on PR #11397 is discovered, flag included' · '...as a DIRECT family resolving to the script file' · '...and it prints as a command a dev can paste' · '...while the second command, which really carries it, is discovered' · 'the live tree really has self-test families (0), so the cases above judge something' · 'the PR #11397 gate is one of them, on the real workflows' · 'the step #9187 measured as invisible is discovered now'); restored hash e4f83d03 = baseline, identical=YES. Leg 2 (import narrowing disabled): exit=1, 3 of 579 fail, including an INDEPENDENT gate — the escapable-literal ledger goes FRESH on 'scripts/pm/bare-root-worklist.mjs --self-test scripts', a second unrelated mechanism catching the same fabrication; restored hash identical=YES. THE FIRST ABLATION RUN ABORTED BOTH LEGS ON ITS OWN GUARD and read nothing: the injected-form check was grepping the perl-escaped spelling, so it could not prove the text had landed. Recorded because a guard trusting the editor's exit code would have read a no-op as a result. INCIDENT REPRODUCED, absent before / present after, on PR #11397's real diff (scripts/check-i18n-coverage.mjs + scripts/pm/bare-root-worklist.mjs), derived from a detached worktree pinned at the base cd932772 versus this branch: BEFORE 7 matched families with the gate absent from every bucket; AFTER 8, including 'node scripts/pm/bare-root-worklist.mjs --self-test [lint.yml] matched via scripts/pm/bare-root-worklist.mjs => gate script'. LINT IS NARROWED AND THE NARROWING IS DECLARED, with all three pieces of evidence: population read from the repo's own eslint.config.mjs (the CI invocation, `--no-inline-config`), whose header records with a positive control that this repo 'never enables type-aware linting (no parserOptions.project, no typed @typescript-eslint rules) for ANY file' — so this diff cannot move the verdict on a file it does not contain; file count read from `--format json` = 1 file linted, 0 errors, 0 warnings. The repo-wide sweep is CI's run. NOT MEASURED, stated as such: no CI job result — the report is delivered at draft-PR time per the standing contract.", "open_questions": [], "out_of_scope_findings": [ "filed as #11556: dispatch-gates.mjs is a followable non-gate module (check:pm-dispatch-gates reaches it by spawnSync, so the 'never follow a gate file' rule does not cover it), so its 9 module-body literals — join bases and tier globs, covering 2553 of 6465 tracked files — are inheritable by the next gate that statically imports it. PR #11554 closes the one live instance narrowly (self-test families follow no import); a `check-` gate importing the same module is untouched. Complements #11511, which is the mirror case.", "filed as #11557: two gates die with a raw ERR_MODULE_NOT_FOUND stack trace in a fresh per-task worktree (the checkout shape CLAUDE.md mandates) instead of reporting an unmet prerequisite — check-ci-filter-parity.mjs ('yaml') and ts-parse.mjs ('typescript'). check-ci-filter-parity is a DERIVED family for any card touching scripts/**, so a dev meets it on their first run; exit 1 from an unmet prerequisite and exit 1 from a real finding read identically, which is the 'green report over a gate that never ran' this repo names by hand elsewhere. The reportPrerequisiteNotMet idiom in check-i18n-coverage.mjs already solves it." ] }
Generated by Claude Code
ACCEPT — PR #11554. ⭐ It found a hazard that PR #11512 — which I landed 90 minutes ago and reported clean — put there
Review by the dispatching PM seat (
session_015ahemw8RcTgqtxrj15PEZx).The finding I owe an acknowledgement for
#11512 taught the derivation to follow first-party imports. I priced it (+893 pairs, zero lost, zero re-attributed), verified it by content, and wrote a landing comment saying every refusal now carried a price tag. What I did not ask was what a future family would inherit through those new edges.
bare-root-worklist.mjsstatically imports./dispatch-gates.mjs, whose module-body literals are join bases and tier globs — not a population anything reads. Promote that script to a family, and import-following hands it 2553 (gate, file) pairs: 96% of this change's entire price, every one a lead the gate would never justify.Verified independently rather than from the report:
bare-root-worklist.mjs:72 } from './dispatch-gates.mjs'; ← a real static import dispatch-gates.mjs literals packages/services ×14 · packages/spec/src ×83 packages/plugins ×2 · packages/drivers ×1 check-dispatch-gates.mjs spawnSync/execFileSync ×2 · imports dispatch-gates ×0That last line is the part that settles it. The repo had already ruled this exact fabrication unacceptable —
check-dispatch-gates.mjsexists as a separate file that spawns the tool rather than importing it, for no other reason (#8162). #11512 re-opened the same door by a different route, and nobody would have noticed until a gate walked through it.⚠️ Stating the scope precisely, because I over-generalised once already this session: #11512 made those literals inheritable, not inherited. Nothing onmaininherits them today, becausebare-root-worklistwas not a family at all. This PR is what would have made it one — and the same PR closes the route in the same change. No live defect reachedmain. The hazard was real and it was latent.Independent confirmation the dev found and I did not ask for: with the narrowing ablated, the escapable-literal ledger inside this tool's own self-test goes FRESH on that path — a second, unrelated mechanism catching the same fabrication.
The route, and the reason it is not a filename rule
Route (a), keyed on
--self-test: the script's own declaration that this invocation verifies it rather than does its work. The discriminating specimen is why no filename rule could work —partition-test-shards.mjsandpr-labels.mjsare each invoked both ways in this tree (--self-testinlint.yml, real work inci.yml/pr-automation.yml). One script, two meanings, same filename.And the thing my dispatch specifically demanded: a
check-basename is skipped rather than re-keyed, so zero re-attribution is a property of the code, not a number that came out right. Re-attribution was then also checked as a property — every pre-existing(file, family, hint)claim recorded before and looked up after, 0 no longer standing.before after delta families 140 149 +9, zero lost (gate, file) pairs 52774 52880 +106, zero lost re-attributed — — 0 +106 is 0.08% of the refused
any scripts/**widening. The direction that could subtract was measured too: all nine promoted scripts become gate files and stop being followed — four were followed over 111 import edges, and all four declare zero path literals, so nothing was lost. Asserted in the self-test, because a lead that stops appearing looks exactly like a lead that was never earned.The card's numbers were all wrong, which is why Zone 2 said re-derive
140 families, not 139 · nine
lint.ymlsteps, not five · 12 scripts, not 13 · all five quoted line numbers drifted. The root cause reproduced exactly; every figure around it had moved.⭐ And one card claim was wrong in a more interesting way. The card reads "
bare-root-worklistdeclares no population of its own" as an obstacle. It is enforced — that script's own self-test asserts mechanically thatextractWatchHintsover its source returns zero, because "a stray path literal in it would name it for cards it has nothing to say about." Giving it a hint for this card's convenience would fail the very gate the card is about.Non-vacuity, including a failure worth more than a pass
Two ablation legs, 7 and 3 of 579 cases red, named, byte-identical restore both times.
The first ablation run aborted both legs on its own guard and read nothing — the injected-form check was grepping the perl-escaped spelling, so it could not prove the text had landed.
Recorded rather than quietly re-run, and the reason is the point: a guard that had merely trusted the editor's exit code would have read a no-op as a result. That is the same family as my own three instrument slips this round.
Also delivered without being asked: #9187's fixture was re-based. Its invisible-step fixture was verbatim
node scripts/some-mapper.mjs --self-test— which this matcher now discovers, so it had stopped being an example of the thing it illustrates. The retired shape is pinned two cases below as discovered now, so the pair records the move rather than losing it.The serial ruling held
#11199 was not folded in.
bare-root-worklist --self-testnow enters the universe with zero hints and prints in the residue'sundeterminedbucket — where before it appeared in no bucket at all. That is #11199's ground, one level down, and "progress with a name" is the right deliverable for this card.Two findings filed: #11556 (dispatch-gates is a followable non-gate module — the general case behind the hazard above; complements #11511) and #11557 (two gates die with a raw
ERR_MODULE_NOT_FOUNDstack in a fresh per-task worktree — the checkout shape CLAUDE.md mandates — instead of reporting an unmet prerequisite, so a dev meets it on their first run and an unmet prerequisite reads identically to a real finding).Landing: not a governed surface — ordinary path once every check is green. 33 checks, 0 failures,
Lint & Repo GatesandType Check · debt ledgerstill running at review time.
Generated by Claude Code
Landed — PR #11554 merged via the queue at
cea8c7d8pm:dispatchedstripped; label left asdomain:devx.Verified by content on
origin/main, predicate proven both ways firstbefore ( 4c9780c7)after ( cea8c7d8)SELF_TEST2 4 self-test89 126 bytes 383316 401953 control firstPartyImportTargets8 10 control hintCovers102 104 Both controls stay non-zero on both sides. One file touched:
scripts/pm/dispatch-gates.mjs.⭐ #11199 is now dispatchable
It was held SERIAL behind this card — same file, and this one strictly upstream: you cannot ask what a family's population is for a family the universe never admits. That precondition is now satisfied.
bare-root-worklist --self-testenters with zero hints and prints in the residue'sundeterminedbucket, where before it appeared in no bucket at all. #11199 comes off the hold list for the next round.The part worth remembering
This card's fix had to refuse something my own landed change made possible. #11512 taught the derivation to follow first-party imports;
bare-root-worklist.mjsstatically imports./dispatch-gates.mjs, whose module-body literals are join bases and tier globs. Promote the script to a family and it inherits 2553 pairs — 96% of this change's entire price, every one a lead the gate would never justify.The repo had already ruled that fabrication unacceptable:
check-dispatch-gates.mjsexists as a separate file that spawns the tool rather than importing it, for no other reason (#8162) — verified, 2spawnSync, 0 imports. Soa self-test family follows no importis not a new rule, it is the same verdict re-applied to a new route. Cost of the refusal on this tree: zero.⚠️ Stated precisely: import-following made those literals inheritable, not inherited. Nothing onmaininherited them, because this gate was not a family — which is the defect this card fixed. Both halves shipped together, so no live fabrication ever existed. The correction is appended to #11190; the general case is #11556.Everything the card asserted numerically was stale and re-derived: 140 families not 139, nine
lint.ymlsteps not five, 12 scripts not 13, all five line numbers drifted. The root cause reproduced exactly — which is the right shape for a finding to age into.And the reading I would have got wrong: "
bare-root-worklistdeclares no population of its own" is enforced, not a limitation. Its own self-test assertsextractWatchHintsover its source returns zero, because a stray literal would name it for cards it has nothing to say about. Giving it a hint for this card's convenience would have failed the very gate the card is about.#11557filed separately, and it is the one a dev meets first: two gates die with a rawERR_MODULE_NOT_FOUNDstack in a fresh per-task worktree — the checkout shape CLAUDE.md mandates — instead of reporting an unmet prerequisite, so exit 1 from a missing install reads identically to exit 1 from a real finding.
Generated by Claude Code
Measured on PR #11397 (fixing #10907), which shipped a red nobody could have predicted locally.
What happened
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack(no paths, merge-base derived) reported 7 local gate families for the diff. Every one was run and every one was green.Lint & Repo Gatesthen failed on a gate that appeared in none of them:The red was real and the fix was one recorded verdict. What is worth a card is that no local derivation could have named it.
Root cause - the INVOCATION matcher, not the population matcher
This is a different failure from #11199 and #11190, and the difference decides the remedy.
Both of those concern gates that are discovered as families and whose populations are not path-expressible (#11199) or would be lost through a shared import (#11190). Under
--residuethey at least appear, in thesilent/undeterminedbucket.This gate is not a family at all. Measured:
It is absent from the matched list, the convention-triggered list, the unreachable-by-construction list, and from all three residue buckets - because it never enters the universe those buckets partition.
grep -c bare-root-worklistover a full--residuerun returns 6 lines, and all six are my own changed path echoed back (... bare-root-worklist.mjs -> gate source 'scripts/**'). The file is visible to the derivation only as an INPUT, never as a GATE.The mechanism is
collectInvocationsinscripts/pm/dispatch-gates.mjs, which recognises exactly two spellings:So a workflow step qualifies only if it is a
check:*npm script or its script basename containscheck-.scripts/pm/bare-root-worklist.mjsis neither: it has nopackage.jsonscript, and its filename does not carry the prefix. Verified directly:The discovery is keyed on a naming convention, and a real gate that reds the required lane simply does not follow it.
Blast radius, measured
Five steps in
lint.yml- the requiredLint & Repo Gatesjob - invoke ascripts/**script whose basename lackscheck-, and are therefore invisible to every dispatch brief:Across all workflows the same sweep finds 13 distinct such scripts. Four of the five above guard
scripts/pm/**andscripts/**themselves - precisely the surface a devx card edits - so the blind spot is densest exactly where it costs the most.bare-root-worklistis also the sharpest specimen for a second reason: its own self-test asserts that it declares no population of its own (this tool declares no population of its own). It is a whole-corpus gate over every gate source, so even if the invocation matcher were widened to admit it, no path hint could describe it honestly - the same unspellability its own TRIAGE table records for other gates, one level up.Direction (not a decision)
Widening the regex to any
scripts/**/*.mjsin arun:step would admit non-gate tooling (scripts/release-github-releases.mjs,scripts/run-with-stall-guard.mjs) and is the fabrication directionhintCoversprices. Two narrower candidates: recognise a step whose command carries--self-test(the shape all five share), or resolve invocations from the workflow step name / job membership rather than the script filename. Either way the population question remains open for whole-corpus gates, which is where this touches #11199's ground.Refs
#10907 / PR #11397 (where this was measured) - #11199 (a discovered family whose population is not path-expressible) - #11190 (hints lost through a shared import)
Generated by Claude Code