test(#3466): replace 8 source-grep assertions with behavioral tests (#3500)

Phase 2 of #3464, following #3465. Rewrites every assertion that read a
shipped .cjs/.js file and text-searched it, so the file no longer needs an
allow-test-rule exemption. 6 of the 8 files are now marker-free; the ceiling
drops 285 -> 278 against a measured 277.

A source-grep passes when a STRING is present, not when the code WORKS. It
survives a refactor that keeps the string but breaks the behavior, and breaks
on a refactor that keeps the behavior but renames the string. Both failure
modes are silent about the thing the test claims to protect. That is the
anti-pattern ADR-456 and local/no-source-grep exist to prevent.

Rewrites, each against the real exported seam:

- discuss-mode: calls cmdInitPlanPhase() against a fixture whose config sets
  workflow.text_mode, asserts the value propagates to its emitted JSON.
- effort-surface-axis: runs review-lane invoke against a real project with a
  fake `claude` PATH shim, asserts the resolved --effort actually lands in the
  shim's captured argv.
- install-minimal-hooks: calls applySettingsJsonHooks() with a hook source
  missing, asserts it is neither registered nor silently registers wholesale,
  with sibling present hooks as the positive control.
- install: calls the exported inferPreferredRuntime({fs, env,
  preferredConfigDir}) via its injected fs seam, asserting 'kilo' from both
  the config-marker and env-var paths.
- opencode-permissions: spawns the real installer with a custom config dir,
  asserts the written opencode.json permission paths are anchored on it.
- repo-layout: spawns the installer for copilot local vs global, asserting
  AGENTS.md is written only in the local case.
- runtime-config-adapter-registry: stubs resolveInstallPlan and force-reloads
  install.js, asserting the runtime's artifact stops being written -- proving
  install.js genuinely routes through the registry.
- runtime-homes-descriptor-drive: calls buildAgentSkillsBlock() for cursor and
  claude against real fixture SKILL.md files, asserting each runtime's refs
  land under its own config dir and never the other's.

Every rewrite was mutation-checked before its marker was dropped. Because
node --test is not runnable locally in this repo, each assertion was
replicated in a standalone probe that requires the same module: run green
against the real file, then red against a deliberately broken one (text_mode
propagation deleted, existsSync guard removed, config dir hardcoded, !isGlobal
guard dropped, kilo branch removed, effort resolution bypassed, skills base
hardcoded back to .claude), then the production file restored and confirmed
byte-identical. An assertion that could not be made to fail would not have
shipped -- a behavioral test that passes regardless of correctness is strictly
worse than the source-grep it replaces, because it looks rigorous while
asserting nothing.

Three claims were checked rather than trusted. All three were wrong:

- repo-layout's own marker cited #1188 asserting the `!isGlobal` lexical scope
  was "unprovable at runtime". It is provable: the guard decides whether
  AGENTS.md is written, which is directly observable. Both directions verified.
- The triage for runtime-config-adapter-registry claimed its source-grep was
  redundant with the file's EXPECTED_TABLE tests, so deletion would be safe.
  Those tests only exercise resolveInstallPlan() directly and never load
  bin/install.js, so they do not cover it. A real behavioral assertion was
  written instead of deleting coverage.
- An earlier revision of this change dropped runtime-config-adapter-registry's
  two markers on the grounds that ESLint stayed silent without them. An
  adversarial review caught that this was wrong, and it is restored here. The
  file still genuinely source-greps bin/install.js at two sites; ESLint is
  silent only because no-source-grep's TEXT_METHODS omits matchAll. Dropping a
  marker because the linter cannot see the violation is exploiting the blind
  spot, not resolving it -- and it would go red the moment the rule is
  widened. Those two assertions are also irreducible: they assert that EVERY
  inline `runtime === '...'` branch in install.js names a registry-known
  runtime, and a branch naming an unregistered runtime would simply never
  execute, so no runtime observation can prove its absence. The markers now
  say so explicitly.

Two coverage gaps in no-source-grep surfaced while doing this, recorded on
#3464 rather than fixed here, since widening the rule is its own change with
its own blast radius:

- TEXT_METHODS omits matchAll, so a matchAll source-grep never trips the rule
  (the case above).
- The path test requires a literal quoted bin/lib/gsd-core/src segment and
  tracks the binding one hop, so a read through a dynamic path or an
  intermediate variable evades it. install-minimal-hooks' genuinely
  load-bearing read at line 975 is itself unmarked for a related reason.

install-minimal-hooks therefore keeps its markers too: its remaining real
source-grep of bin/install.js (the Codex legacy gsd-update-check migration
check, line 975) is outside this issue's 8 sites. It is the last blocker for
that file and is a clean follow-up.

Closes #3466

Co-authored-by: sim <sim@local>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Tom Boucher
2026-08-14 15:57:25 -04:00
committed by GitHub
parent 895d9df96d
commit 8a6d87538f
10 changed files with 588 additions and 311 deletions

View File

@@ -35,6 +35,7 @@ const { resolveRuntimeArtifactLayout } = require(
path.join(ROOT, 'gsd-core', 'bin', 'lib', 'runtime-artifact-layout.cjs'),
);
const { getGlobalConfigDir } = require(path.join(ROOT, 'gsd-core', 'bin', 'lib', 'runtime-homes.cjs'));
const { createTempDir, cleanup } = require('./helpers.cjs');
const sorted = (iterable) => [...iterable].sort();
@@ -492,7 +493,8 @@ describe('issue-57 AC2 — config-mutation dispatch is closed over the explicit
// structural guard over bin/install.js source. Behavioral assertions
// cannot observe inline `runtime === '...'` config branching, so this enforces that
// every inline per-runtime branch references a runtime the adapter registry knows
// about — a NEW branch against an unregistered runtime name fails here.
// about — a NEW branch against an unregistered runtime name fails here. ESLint cannot
// currently see this grep because no-source-grep's TEXT_METHODS omits matchAll (#3464).
test('every inline `runtime === "..."` branch references a registry-known runtime', () => {
const src = fs.readFileSync(path.join(ROOT, 'bin', 'install.js'), 'utf8');
const literals = new Set(
@@ -514,7 +516,8 @@ describe('issue-57 AC2 — config-mutation dispatch is closed over the explicit
// allow-test-rule: structural-regression-guard (#2103)
// VS Code is a registry runtime but is NEVER CLI-installed (Marketplace/VSIX
// extension); it must stay fully descriptor-driven — bin/install.js must
// never special-case it by name.
// never special-case it by name. ESLint cannot currently see this grep because
// no-source-grep's TEXT_METHODS omits matchAll (#3464).
test('#2103: bin/install.js has ZERO runtime === "vscode" / isVscode branches (vscode stays fully descriptor-driven)', () => {
const src = fs.readFileSync(path.join(ROOT, 'bin', 'install.js'), 'utf8');
const runtimeComparisons = [...src.matchAll(/runtime === (?:'vscode'|"vscode")/g)];
@@ -534,19 +537,63 @@ describe('issue-57 AC2 — config-mutation dispatch is closed over the explicit
);
});
// allow-test-rule: structural-regression-guard (#3336)
// delegation-presence guard. Catches wholesale removal of the registry
// dispatch (a regression to scattered per-runtime config branching).
test('bin/install.js requires the config adapter registry and dispatches through it', () => {
const src = fs.readFileSync(path.join(ROOT, 'bin', 'install.js'), 'utf8');
assert.ok(
src.includes('runtime-config-adapter-registry'),
'bin/install.js no longer requires the runtime config adapter registry',
);
assert.ok(
src.includes('resolveInstallPlan('),
'bin/install.js no longer dispatches config through resolveInstallPlan',
);
// Behavioral replacement for the delegation-presence source grep (#3466).
//
// The EXPECTED_TABLE / resolveInstallPlan projection-contract tests above
// (`resolveInstallPlan — descriptor-projection contract`) prove resolveInstallPlan
// ITSELF maps every registry descriptor correctly — but they call the registry
// function directly and never touch bin/install.js, so they cannot by themselves
// prove install.js actually CONSULTS it rather than reimplementing an equivalent
// per-runtime branch. This test closes that gap: it stubs the registry's
// resolveInstallPlan (the SAME module object bin/install.js requires and
// destructures) so it reports a different installSurface for every runtime,
// re-requires a fresh bin/install.js so its destructured reference picks up the
// stub, and asserts a REAL install(false, 'copilot') run STOPS producing the
// copilot-instructions.md artifact that installSurface === 'copilot-instructions'
// gates directly inside install() (bin/install.js:~12164 — no finishInstall/CLI
// orchestration layer involved, so this is reachable from install() alone). If
// install.js ever reverts to a `runtime === 'copilot'` inline branch instead of
// reading resolveInstallPlan(runtime).installSurface, the stub has no effect on
// that branch and the artifact keeps getting written — which is exactly the
// divergence this test would then catch (verified by mutating install.js to that
// exact inline form during authoring: the assertion below goes red).
test('bin/install.js dispatches config through the REAL resolveInstallPlan (stubbing it changes install() output)', () => {
const installPath = require.resolve('../bin/install.js');
const registryPath = require.resolve('../gsd-core/bin/lib/runtime-config-adapter-registry.cjs');
const registryModule = require(registryPath);
const originalResolveInstallPlan = registryModule.resolveInstallPlan;
registryModule.resolveInstallPlan = (runtime) => ({
...originalResolveInstallPlan(runtime),
// Force every gate bin/install.js checks against
// resolveInstallPlan(runtime).installSurface to read as "nothing special for
// this runtime" — including copilot's own surface, which the real descriptor
// sets to 'copilot-instructions'.
installSurface: 'settings-json',
});
delete require.cache[installPath];
const stubbedInstaller = require(installPath);
const tmpDir = createTempDir('gsd-3466-delegation-guard-');
const previousCwd = process.cwd();
process.chdir(tmpDir);
try {
const result = stubbedInstaller.install(false, 'copilot');
const instructionsPath = path.join(result.configDir, 'copilot-instructions.md');
assert.equal(
fs.existsSync(instructionsPath), false,
'with resolveInstallPlan stubbed to report installSurface: "settings-json" for every '
+ 'runtime, install(\'copilot\') must NOT write copilot-instructions.md — if it still '
+ 'does, bin/install.js is not actually gating on resolveInstallPlan(runtime) for this '
+ 'decision (a regression to scattered per-runtime branching)',
);
} finally {
process.chdir(previousCwd);
cleanup(tmpDir);
registryModule.resolveInstallPlan = originalResolveInstallPlan;
delete require.cache[installPath];
}
});
});
});