neo PR #20321 changed learn/blog/persistent-memory-for-ai-coding-agents.md and apps/portal/resources/data/blog.json. It ran the engine e2e tier (4 shards) and the Brain-gated tier (10 shards) in full, and its Brain shards queued 7–20 minutes on the organization's 20-job cap. The operator, 2026-11-05: "you changed a markdown blog post file, and that triggered ALL unit and E2E pipelines to run. this is horribly wasteful." Baseline and method: #20328.
The Problem
classify-test-scope.yml admits a tier whenever a changed path falls inside its boundary. Four admissions reach nothing the tier runs. Counts are over the 346 PRs merged into dev from 2026-10-13 to 2026-11-05:
Admitted by
Why the tier cannot reach it
PRs
apps/portal/resources/data/**, through apps/
No e2e spec, fixture or config reads portal data; unit's content guards own it
6 (every blog post)
buildScripts/{create,docs,helpers,release}/**, through buildScripts/
Outside the e2e launch chain (below)
8
A neo-agent-skills-only bump (package.json + package-lock.json)
Agent instructions, which webpack and Playwright never load. Unit's prepare.spec.mjs and baseline / Skills materialized cover its install path
12
Any change outside the E2E boundary, docs included
The Brain tier still runs a one-shard smoke. It tests the Brain's dev tip, not the change, and the dev push runs the full tier anyway
67 today, 93 after the rows above
The Architectural Reality
isE2eRelevantPath admits all of apps/ and buildScripts/ on purpose. The comment on that predicate in classify-test-scope.yml reads: "this gate's failure mode is SILENCE. A false positive costs one ~3-minute job". Since #19705, a false positive costs 14 jobs.
The e2e launch chain draws buildScripts/ code from:
webpack.server.config.mjs (server-start);
the bundle-* and build-highlightjs builds;
e2eCiSelection.mjs;
the Playwright config, fixtures and specs.
Their static imports stay inside build/, webpack/ and util/.
packageJsonRelevance() already reads package.json at both refs. package-lock.json admits components and e2e unconditionally.
A skills bump changes three entries in the lock: the root specifier, plus version, resolved and integrity on node_modules/neo-agent-skills.
test-e2e-brain.yml builds its matrix from e2e_brain_tier: one shard for smoke, ten for full. WhiteboxE2E.md:44-46 says the tier "never skips", and "Any other pull request runs the smoke subset".
The dev ruleset requires unit, components (1/3) to (3/3), e2e-engine, build-all and the baselines. It requires no e2e-brain context.
Design authority: the classifier comment quoted above prices a false positive at one ~3-minute job, which no longer holds. The Brain tier's "never skips" is my own AC on #19657, under #19655 (steward @neo-opus-ada). This change keeps that ticket's "always reports".
It drops buildScripts/{create,docs,helpers,release}/. Because this is a deny-list, a new buildScripts/ directory still admits the tier.
Packages that never reach a browser tier (neo-agent-skills) leave the dependency-kind comparison. package-lock.json gets the same both-refs read: once those packages' entries are removed, an unchanged lock admits nothing. An unreadable ref fails toward running.
e2e_brain_tier gains none for a pull request outside the boundary. test-e2e-brain.yml then runs one shard whose only step reports the skip and its reason, so the context still reports. A push stays full.
Each change gets a red-first arm in WorkflowScopeClassifier.spec.mjs. WhiteboxE2E.md names the none tier.
Acceptance Criteria
#20321's file set (a learn/blog/*.md + blog.json) yields run_e2e=false, run_components=false, e2e_brain_tier=none and run_unit=true.
A file under buildScripts/create|docs|helpers|release/ alone runs no e2e. A file under buildScripts/build|webpack|util/, or under a new buildScripts/ directory, runs both tiers in full.
A bump of neo-agent-skills alone runs neither components nor e2e. The same bump plus any other lock change runs both, and so does an unreadable package-lock.json.
A docs-only PR's Brain job reports none with its reason and runs no Playwright. A push still selects full.
The 346-PR corpus, re-run against the changed script, admits the full tiers for fewer than 279 PRs, and every flip falls in one of the four rows above. The PR records this.
WhiteboxE2E.md states when the tier selects none.
The Brain job's skip step prints each changed file followed by in, or by the rule that left it out, passing the report through env. This was @neo-opus-ada's condition for none: a misclassification then shows on the PR that carries it.
Out of Scope
Spec-level selection inside the boundary (the breadth leaf of #20328).
Scoping unit, unit-tz, CodeQL or build-all.
Avoided Traps
Dropping all of apps/: the engine specs navigate more than 20 apps.
An allow-list of buildScripts/ files: such lists rot, as the classifier comment records.
Removing the Brain job on a skip: a reported skip costs one short job, while proving that no consumer reads the context costs more.
Sweeps: same as #20328, at 2026-11-05T14:21Z. They found no equivalent ticket and no prior exclusion.
Origin Session ID: 8545cff9-f4d7-4731-b93e-7dfad1347c8a
Retrieval Hint: "classifier misclassification portal data buildScripts create docs neo-agent-skills lock Brain tier none docs-only"
Context
neo PR #20321 changed
learn/blog/persistent-memory-for-ai-coding-agents.mdandapps/portal/resources/data/blog.json. It ran the engine e2e tier (4 shards) and the Brain-gated tier (10 shards) in full, and its Brain shards queued 7–20 minutes on the organization's 20-job cap. The operator, 2026-11-05: "you changed a markdown blog post file, and that triggered ALL unit and E2E pipelines to run. this is horribly wasteful." Baseline and method: #20328.The Problem
classify-test-scope.ymladmits a tier whenever a changed path falls inside its boundary. Four admissions reach nothing the tier runs. Counts are over the 346 PRs merged intodevfrom 2026-10-13 to 2026-11-05:apps/portal/resources/data/**, throughapps/buildScripts/{create,docs,helpers,release}/**, throughbuildScripts/neo-agent-skills-only bump (package.json+package-lock.json)prepare.spec.mjsandbaseline / Skills materializedcover its install pathdevtip, not the change, and thedevpush runs the full tier anywayThe Architectural Reality
isE2eRelevantPathadmits all ofapps/andbuildScripts/on purpose. The comment on that predicate inclassify-test-scope.ymlreads: "this gate's failure mode is SILENCE. A false positive costs one ~3-minute job". Since #19705, a false positive costs 14 jobs.buildScripts/code from:webpack.server.config.mjs(server-start);bundle-*andbuild-highlightjsbuilds;e2eCiSelection.mjs;build/,webpack/andutil/.packageJsonRelevance()already readspackage.jsonat both refs.package-lock.jsonadmits components and e2e unconditionally.version,resolvedandintegrityonnode_modules/neo-agent-skills.test-e2e-brain.ymlbuilds its matrix frome2e_brain_tier: one shard forsmoke, ten forfull.WhiteboxE2E.md:44-46says the tier "never skips", and "Any other pull request runs the smoke subset".devruleset requiresunit,components (1/3)to(3/3),e2e-engine,build-alland the baselines. It requires noe2e-braincontext.Design authority: the classifier comment quoted above prices a false positive at one ~3-minute job, which no longer holds. The Brain tier's "never skips" is my own AC on #19657, under #19655 (steward @neo-opus-ada). This change keeps that ticket's "always reports".
The Fix
In
classify-test-scope.yml:isE2eRelevantPathdropsapps/portal/resources/data/.buildScripts/{create,docs,helpers,release}/. Because this is a deny-list, a newbuildScripts/directory still admits the tier.neo-agent-skills) leave the dependency-kind comparison.package-lock.jsongets the same both-refs read: once those packages' entries are removed, an unchanged lock admits nothing. An unreadable ref fails toward running.e2e_brain_tiergainsnonefor a pull request outside the boundary.test-e2e-brain.ymlthen runs one shard whose only step reports the skip and its reason, so the context still reports. A push staysfull.Each change gets a red-first arm in
WorkflowScopeClassifier.spec.mjs.WhiteboxE2E.mdnames thenonetier.Acceptance Criteria
learn/blog/*.md+blog.json) yieldsrun_e2e=false,run_components=false,e2e_brain_tier=noneandrun_unit=true.buildScripts/create|docs|helpers|release/alone runs no e2e. A file underbuildScripts/build|webpack|util/, or under a newbuildScripts/directory, runs both tiers in full.neo-agent-skillsalone runs neither components nor e2e. The same bump plus any other lock change runs both, and so does an unreadablepackage-lock.json.nonewith its reason and runs no Playwright. A push still selectsfull.WhiteboxE2E.mdstates when the tier selectsnone.in, or by the rule that left it out, passing the report throughenv. This was @neo-opus-ada's condition fornone: a misclassification then shows on the PR that carries it.Out of Scope
build-all.Avoided Traps
apps/: the engine specs navigate more than 20 apps.buildScripts/files: such lists rot, as the classifier comment records.Related
#20328 (parent) · #20321 (witness) · #19655, #19657 (Brain tier) · #12705 (first docs-only scoping) · #18493 (shared classifier) · #18052 (classifier spec)
Sweeps: same as #20328, at 2026-11-05T14:21Z. They found no equivalent ticket and no prior exclusion. Origin Session ID: 8545cff9-f4d7-4731-b93e-7dfad1347c8a Retrieval Hint: "classifier misclassification portal data buildScripts create docs neo-agent-skills lock Brain tier none docs-only"