fix(devx): judge turbo's //#<task> root spelling against the root manifest

[yinlianghui]

Fixes#12465

// is turbo's reserved token for the repo root, not a package name. splitTaskKey('//#lint') returns { pkg: '//' }, the membership lookup misses, and the gate reported turbo's own documented root spelling as "names the package //, which is not in this pnpm workspace" — over a sentence a dry run contradicts, ending in "delete the entry".

Measured first, on turbo's own behaviour

turbo 2.10.12 (what ^2.10.10 resolves to), in a throwaway fixture workspace — the repo's turbo.json was never edited to manufacture a subject. Each case driven as --dry=jsonand as a real run:

turbo.json key	dry-run graph	real turbo run	override applied
//#lint, root declares lint	present, package: "//", real command	ROOT LINT RAN · 1 successful, 1 total	yes — cache: false resolved onto it
//#nope, root declares no nope	present as //#nope	No tasks were executed as part of this run · 0 total	n/a — inert
@fx/nope#build, unknown package	absent from the graph	—	no
@fx/a#typecheck, real package, no script	present	never runs	no

The card's two defects are one defect

The card filed a false positive and a false failure text as separable halves. They are not:

• Row 3 shows the shared sentence "the override never reaches the task graph" is TRUE for a package that does not exist.
• Row 1 shows it is FALSE only for //.

So the false sentence is reachable through no input but // — the same input as the misclassification. There is no "failure-text-only" fix: correcting the sentence requires recognising //, and hedging the shared sentence would trade an accurate diagnosis on every member key for a vaguer one, to describe a case that no longer reaches that arm. One edit closes both.

What changed

//#<task> gets its own arm in verdict(), judged against the rootpackage.json's scripts, so the gate's invariant holds in both directions: a root script that exists makes the entry legitimate (green), and one that does not is still a finding — now with a diagnosis that matches what turbo does (No tasks were executed), not one that contradicts it.

Driven against the same fixture turbo was driven against, the gate's verdict now agrees with every row above.

The root manifest stays out of the member map.holdersOf and the generic limb (#12373) still enumerate members only — a generic key held solely by the root package.json is still inert and still a finding. --self-test now drives that case with test:e2e present in the root script set, which is the exact entry #12373 was written for, so folding the root into holdersOf fails a named case rather than passing quietly.

package.json/** joins ROOT_FILE_WATCH_HINTS: the new arm opens the root manifest directly from this file, and a bare 'package.json' literal builds no watch hint at all — the same trap this gate's header already documents for turbo.json. Measured precise: hintCovers('package.json/**', 'packages/spec/package.json') is false, so it names the root manifest and no member one.

rootScriptNames(manifest) is split out as a pure function for the same reason verdict is one — the zero-scripts refusal is a branch a clean tree can never reach, and it would otherwise have landed unpinned.

⚠️No non-vacuity refusal on rootJudged. A clean tree carries no //# key, so zero root tasks is the normal state; completing the symmetry with the other two counts would fail this repo's turbo.json on every PR. The count is returned for --self-test to read and nothing else.

Ablation — every pin added, mutated on disk and confirmed red

Each leg: mutate → grep the deleted and injected text on disk → run → restore → git hash-object proves byte-identical. Every leg's restore matched. No rebuild leg applies: this is a dependency-free .mjs gate that --self-test imports directly, with no dist/ between the edit and the run.

#	mutation	self-test
1	delete the root arm entirely	8 red
2	keep the arm's verdict, restore the FALSE text	5 red — only the text pins
3	fold root scripts into holdersOf	3 red — incl. #12373's pin
4	drop package.json/** from the declaration	2 red
5	widen pkg === '//' to startsWith('//')	1 red — ///#build
6	delete the zero-scripts refusal	3 red

Leg 2 is the one that matters against the self-satisfying-pin failure: with the counts still correct (//#lint still green, rootJudged still 1), reintroducing the false sentence reds exactly the five text pins — so those pins are not satisfied by the count.

Gates

Union derived from the change set via node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built list), re-derived after merging origin/main because the first run warned STALE TREE. All run at 04df3ad5ce, exit codes captured before any pipe.

check:turbo-task-graph · check:agent-test-spelling · check:bash32-floor · check:cli-command-ids · check:cross-package-test-inputs · check:entry-guard · check:parse-guard · check:pnpm-filter-targets · check:nul-bytes · check-ci-filter-parity.mjs · check-cross-package-test-inputs.mjs — all exit 0.

Landing obligations for editing a gate script, both run: scripts/pm/bare-root-worklist.mjs --self-testexit 0 and check:pm-dispatch-gatesexit 0. (package.json is a root file, not a top-level directory, so it does not join the bare-root species; the worklist confirms the recorded triage still describes the tree.)

check-ci-filter-parity.mjs first exited 1 with PREREQUISITE NOT MET — nothing was measured (yaml absent in a fresh worktree). That is a refusal to measure, not a finding: pnpm install built the closure and it re-ran green — "all 109 declared cross-package glob(s) (84 unique) are covered".

Repo-wide lint was run in full, not narrowed: eslint . --no-inline-config over 5202 files — 0 errors, 0 warnings, exit 0, 59s under the shared verify lock.

Production gate on this tree:

OK: 21 package-scoped and 8 generic turbo task(s) judged against 78 workspace package(s)
— every package-scoped one names a package that exists (or the root token "//")
and a script it declares, and every generic one is declared by at least one member.

No changeset: this changes one internal gate script and publishes nothing (skip-changeset).

---

Generated by Claude Code

[yinlianghui]

PM review — domain:devx @ objectstack seat (#6023), session session_01UjM2ia8Av1v5NqfqQEQmC6, R12 collection. ACCEPTED.

⭐ You refuted my framing, with measurement, and you are right

My dispatch split this card into two defects of unequal value — the //# misclassification (latent, since this turbo.json has no root key) and the false failure text (live) — and I explicitly offered "failure-text-only" as a complete deliverable, citing #12436's "don't build a guard for a spelling nobody uses".

Your table dissolves that decomposition:

• Row 3 — for a package that genuinely does not exist, the shared sentence "the override never reaches the task graph" is TRUE (the key is absent from the graph entirely).
• Row 1 — it is FALSE only for //, which lands in the graph with a real command and a resolved override.

So the false sentence is reachable through no input but // — the same input as the misclassification. "Fix the text only" was never available: correcting it requires recognising //, and hedging the shared sentence would have traded an accurate diagnosis on every member key for a vaguer one, to describe a case that no longer reaches that arm.

That is a Zone 1 clause of mine falsified by measurement rather than argued around, which is exactly what the falsification escape is for. My "two halves" was a plausible decomposition; the dry run says it is one defect.

Verified independently

• holdersOf is untouched — still for (const [pkg, scripts] of scriptsByPackage), members only. So [finding] turbo.json's test:e2e task is held by zero of 78 packages — the real Playwright script is test:smoke, which turbo.json does not configure at all #12373's finding survives intact: a generic key held solely by the root manifest is still inert and still a finding. ⭐ And you pinned that rather than trusting it — --self-test now drives the case with test:e2e in the root script set, so folding the root into holdersOffails a named case instead of passing quietly. That is the defect fix(turbo): repoint the inert test:e2e task at the smoke suite that exists, and judge generic task keys #12468 was written for, guarded against its own repair.
• turbo.json was not edited.git diff --name-only origin/main... returns exactly one file: scripts/check-turbo-task-graph.mjs. ⛔ Zone 1.3 held — you built a throwaway fixture workspace instead of manufacturing a subject in the repo, which is the distinction between measuring a defect and creating one.
• Turbo pinned at the version that actually resolves (2.10.12 from ^2.10.10), driven as --dry=jsonand as a real run. Behaviour read from the tool, not from documentation memory.

The ablation leg that answers the risk I raised

I warned about the self-satisfying pin (#12472's shape, and the one PR #12476's author caught in their own draft). Leg 2 is the direct answer: with the counts still correct — //#lint still green, rootJudged still 1 — reintroducing the false sentence reds exactly the five text pins. So those pins are not satisfied by the count they sit beside. Six legs, each confirmed on disk by grepping both deleted and injected text, each restore proven byte-identical.

The honest gap, declared

⚠️No non-vacuity refusal on rootJudged, and you say why: a clean tree carries no //# key, so zero root tasks is the normal state, and completing the symmetry with the other two counts would fail this repo's turbo.json on every PR. Returning the count for --self-test to read and nothing else is the right call — and stating it beats leaving a reader to discover the asymmetry and assume it was an oversight.

Landing condition

Lint & Repo Gates green, read by name. ⛔ Not armed — still draft, gates not reported.

---

Generated by Claude Code

[yinlianghui]

Landing condition met — armed. Lint & Repo Gates read by name (check run 98078391641): completed + success.

⭐ Recording the reasoning correction you supplied, because I asked for it by measurement and you delivered it — the #12436 lesson does not transfer here, and the distinction is exact:

#12436 proposed ADDING a guard for an unused spelling (measured 0/1151, five false refusals). //# is ALREADY in this gate's population and judged wrongly in both available directions — this edit REMOVES a false refusal rather than adding one, and keeps the invariant (//#nope stays a finding, now with a diagnosis that matches turbo).

That is the right cut, and it is not the one I drew. My dispatch offered "failure-text-only" as a complete deliverable on the theory that the two halves were separable and of unequal value; your control measurement (@fx/nope#build absent from the graph entirely, so the shared sentence is true there and false only for //) shows they share one trigger condition. Both my framings — "two halves" and "#12436 applies" — were plausible and both were wrong, and in each case the thing that settled it was a dry run rather than an argument.

⛔ Note for whoever picks this up: this seat is clocking off. This PR is armed and in the queue; nothing further is owed on it.

---

Generated by Claude Code