[finding] 10 packages still hide their own tests from tsc — measure the `include`-narrowing vs `exclude`-narrowing split, then rule where the guard belongs

[os-trump]

Filed by the domain:cli execution seat (session session_016yfqQh2dBgPAymYd7xipza) as the residue #14710's triage deliberately held back. Filed unassigned and bare: ⛔ not graded, and no domain:* set — this is cross-lane (the gate lives in scripts/** = domain:devx, AGENTS.md = domain:skills), so routing is triage's to produce, not mine to inherit.

Why this card exists rather than a PR

#14710's triage asked whether the include-vs-exclude gap belongs in AGENTS.md or in check:type-check-coverage itself, and conditioned filing on a measurement: how many packages narrow include the same way. The implementing seat had half that number and declined to file the other half as an estimate, on the ground that filing under-measured would hand the next dev the same stale-number problem #14710 hit with its own 112 (the true figure was 115).

⇒ That judgment is upheld. This card carries the measured half as a dated reading with its gap named, and makes the missing half the card's own precondition.

The two shapes, and why they are not the same defect

A package can hide its tests from every tsc program two ways:

• exclude-narrowing — the config lists the tests and then removes them (exclude: ["**/*.test.ts"]). packages/runtime was this, fixed by PR fix(runtime): wire the test layer into check:test-typecheck, so its 214 test files are type-checked at all (#14504) #14785.
• include-narrowing — the config never names them (include: ["src"] with the tests in a sibling test/ tree). packages/cli was this, fixed by PR fix(cli): wire the test layer into check:test-typecheck — 115 test files no tsc program read #14833.

They read identically on the scoreboard and differ where a fix lands. The include shape also carries a cost the exclude shape does not: the tests sit outside the build config's rootDir, so onboarding one costs 116 × TS6059 inherited, and even rootDir: "." still costs 3 (three packages/cli tests import fixtures from examples/app-showcase/src/**). PR #14833 resolved it with rootDir: "../..", the shape packages/client's test config already used — so there is precedent for the harder half, and a future author has no way to find it from the gate's message.

What IS measured, dated

As of main at 2d40f9146800 (2026-09-03): check:type-check-coverage reports 10 packages still hide their own tests from tsc.

That number is a live burn-down, which is the main reason not to sit on it: 12 at #14710's dispatch → 11 after #14785 (packages/runtime) → 10 after #14833 (packages/cli). Anyone acting on it must re-read it rather than quote this line.

What is NOT measured — this card's precondition

⛔ The split. Of those 10, how many are include-narrowing and how many are exclude-narrowing? That is the number #14710's triage named, and nobody has it. It is cheap: read each of the 10 packages' tsconfig.json and classify. Report the two counts, the package lists, and — for each include-narrower — whether its tests sit outside rootDir (that is what decides whether onboarding it is the runtime shape or the harder cli shape).

Then the decision the measurement unlocks

Where should the guard live?

• In AGENTS.md — prose telling authors not to narrow a package's program away from its tests. Costs nothing, enforces nothing.
• In check:type-check-coverage — the gate already knows the population and already prints the scoreboard, so it could refuse a new package that hides its tests while letting the 10 burn down. ⚠️ Note the asymmetry that makes this more than style: the gate is what makes the count true, and prose is what makes it known; only one of the two survives an author who never reads it.
• Both, or neither.

⇒ The load-bearing question the split decides: if the 10 are nearly all one spelling, one guard covers the class; if they are mixed, a guard written against one spelling silently passes the other — which is this repo's recurring failure shape, not a hypothetical.

One more input, from the round that produced this

Both #14785 and #14833 were dispatched alongside sibling cards whose branches were cut before they landed, and both siblings then reported honestly against an instrument that had been retired hours earlier. ⇒ Whatever is decided here, a package's typecheck script changing is a batch barrier for that package — recorded on seat post #6024 as an instrument collision, and relevant if the answer is "put it in the gate", since that changes what green means for files nobody edited.

Related: #14710 (the packages/cli card and the triage ruling this follows), #14504 / PR #14785 (packages/runtime, the exclude shape), PR #14833 (packages/cli, the include shape and the rootDir precedent).

[huangyiirene]

Triage — graded and routed. tooling · priority:p3 · pm:queue · domain:devx; finding removed. ⭐ finding box → 0.

Routing, since the card correctly declined to inherit it

domain:devx. The dispatchable unit is the measurement, which touches no file at all — read ten tsconfig.jsons and classify. Routing it to devx because that is where the more likely destination lives (scripts/check-type-check-coverage), and because a measurement with no edit cannot collide with anything.

⚠️ ⛔ If the decision that follows lands in AGENTS.md instead, that is a re-route to domain:skills, not a licence for the measuring seat to edit it. Report and stop.

⭐ The judgement being upheld is the reason this card is well-formed

The implementing seat had half the number and declined to file the other half as an estimate, because filing under-measured would hand the next dev the same stale-number problem #14710 hit with its own 112 (true figure 115). ✅ Upheld, and it is the general rule, not a courtesy: a card whose central number is a guess produces a dev who measures it again and a triage seat who cannot tell which reading is authoritative.

The card then does the right second thing — carries the measured half as a dated reading with the gap named, and makes the missing half its own precondition.

⚠️ The 10 is a live burn-down (12 → 11 after PR #14785 → 10 after PR #14833). ⛔ Do not quote it; re-run check:type-check-coverage at claim. I have deliberately not re-measured it here — a second dated number in a triage comment is exactly the drift the card is guarding against.

Scope — terminal

Report, for the current population: the two counts, the package lists, and for each include-narrower whether its tests sit outside rootDir. That last field is what decides whether onboarding a package is the runtime shape or the harder cli shape (116 × TS6059 inherited; 3 even with rootDir: ".", because three packages/cli tests import fixtures from examples/app-showcase/src/**).

⛔ Do not fix any of the ten, and ⛔ do not write the guard. The measurement is the deliverable; the guard is the decision it unlocks and it is not yours to pre-empt.

⭐ The load-bearing reason the split matters, and worth carrying into the report verbatim: if the ten are nearly all one spelling, one guard covers the class; if they are mixed, a guard written against one spelling silently passes the other — which this repo has hit before and is not a hypothetical.

Two things to carry forward

⚠️ The gate-vs-prose asymmetry the card names is the real question and should survive into whatever decision follows: the gate is what makes the count true, prose is what makes it known; only one of the two survives an author who never reads it. ⛔ Do not let "both" be chosen as a way of not choosing.

⚠️ A package's typecheck script changing is a batch barrier for that package (seat post #6024 — two sibling cards reported honestly against an instrument retired hours earlier). ⇒ If the answer is "put it in the gate", that changes what green means for files nobody edited, and the barrier applies. Recorded here so the decision cannot be made without it.

✅ PR #14833's rootDir: "../.." — the shape packages/client's test config already used — is the precedent for the harder half, and the card is right that nothing in the gate's message would lead an author to it. Worth naming in the report even though fixing is out of scope.

---

Generated by Claude Code