Fail loudly when tree-sitter grammars are unavailable

[DonJayamanne]

Fixes the macOS unit test failures reported in vscode-engineering#3484, without reverting #328863 — that PR is innocent.

Root cause

Build 20260804.3 (macOS, Electron unit tests) failed with 37 assertion mismatches across CommandAutoApprover (28), SessionPermissionManager (7) and AgentSideEffects (2). Every failure had the same cause: the bash and PowerShell tree-sitter grammars could not be used, so _extractSubCommands returned undefined and every command degraded to noMatch.

That degradation is deliberate and fail-closed in production, but it is silent: grammars load via Promise.allSettled and failures are only _logService.warn, which tests discard through NullLogService. One environment problem therefore surfaced as dozens of unrelated-looking assertion failures — and CI misattributed it to the unrelated commit at the head of the build.

dce00340cb3 touches none of commandAutoApprover.ts, sessionPermissions.ts, or agentSideEffects.ts.

Evidence

Deleting only the two grammar files (leaving core tree-sitter.wasm intact) reproduces CI exactly:

Condition	Result	Console noise	Matches CI?
Baseline	37 pass / 0 fail	none	—
Both grammars unusable	9 pass / 28 fail	zero	✅ exact — counts and all 9 passing test names
Same, all three suites	28 + 7 + 2	zero	✅ exact
Core tree-sitter.wasm missing	2 pass / 1 fail, aborts	loud ENOENT	❌ refuted
Bash grammar only	17–20 fail	zero	❌ no
Grammars truncated	9 pass / 28 fail	zero	✅ exact
Grammars replaced with a valid but wrong grammar	9 pass / 28 fail	zero	✅ exact

The last row matters: the signature can arise even when Language.load() succeeds, so a file-existence check is not sufficient. Missing / truncated / corrupt / wrong-ABI are all indistinguishable from the failure alone.

Changes

Make it loud (source)

• CommandAutoApprover.initialize() now resolves with { parser, bash, powershell } readiness. Readiness is determined by analyzing a probe command per shell, not by trusting Language.load() — this catches the valid-but-wrong-grammar case above.
• The test suite asserts full readiness in setup(), collapsing 28 confusing failures into 1 clear one.

Stop an incomplete dependency tree reaching the tests (build)

• When the pipeline restores its node_modules cache it skips npm ci entirely, so a truncated/partial archive is never repaired. Added verifyNodeModulesCache.ts, run after every cache extraction (7 pipelines), which clears the cache-hit variable so the install steps run and rebuild the tree.
• listNodeModules.ts swallowed statSync failures via catch { continue }, silently dropping files from the cache manifest. It now only skips dangling symlinks and throws otherwise.

Production behavior is unchanged: an unavailable grammar still fails closed and requires user confirmation rather than blocking startup.

Verification

• ✅ 225 passing across all three affected suites (was 224; +1 new readiness test)
• ✅ With grammars removed → 1 clear failure: tree-sitter must be fully available for these tests (was 28)
• ✅ Same single clear failure for the valid-but-wrong-grammar case
• ✅ typecheck-client: 7 errors before, 7 after — no regressions
• ✅ build/ typecheck clean; ESLint clean on all changed files
• ✅ verifyNodeModulesCache.ts manually exercised: passes on a good tree, detects missing and zero-length files, honors the custom variable name used by the copilot pipeline

Note on the caching hypothesis

The most plausible source is a poisoned Darwin node_modules cache (per-platform cache keys explain macOS-only). That remains plausible but unproven without the failed archive — which is precisely why this PR hardens detection at both layers rather than asserting a single physical cause. Recommend also invalidating the darwin cache entry.

Analysis methodology

The first root cause I proposed was wrong (I blamed the product.commit heuristic in appNodeModules.ts). It was falsified by experiment: that path produces loud Emscripten ENOENT errors and aborts after ~2 tests, and repo-root product.json is never stamped with a commit. A proposed isBuilt fix was also rejected — it would have broken the working test harness. Findings were independently re-derived and verified by two separate models before this PR.

[Copilot]

🟡 Not ready to approve

Some affected suites still discard readiness, and symlink stat failures can still be silently omitted.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Pull request overview

Adds explicit tree-sitter readiness reporting and hardens cached dependency restoration.

Changes:

• Probes bash and PowerShell grammars and exposes readiness.
• Adds fail-fast test assertions.
• Verifies restored caches and surfaces manifest-generation errors.

File summaries

File	Description
commandAutoApprover.test.ts	Asserts grammar readiness.
sessionPermissions.ts	Propagates readiness.
commandAutoApprover.ts	Probes parser and grammars.
agentSideEffects.ts	Adapts initialization return type.
product-build-win32-compile.yml	Verifies restored cache.
product-build-web.yml	Verifies restored cache.
product-quality-checks.yml	Verifies restored cache.
product-build-linux-compile.yml	Verifies restored cache.
product-build-darwin-compile.yml	Verifies restored cache.
copilot/setup-steps.yml	Verifies Copilot build cache.
verifyNodeModulesCache.ts	Adds required-file validation.
listNodeModules.ts	Tightens stat error handling.
product-build-alpine.yml	Verifies restored cache.

Review details

• Files reviewed: 13/13 changed files
• Comments generated: 4
• Review effort level: Balanced

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.