Files
gstack/test/skill-md-size-guard.test.ts
82530ac143 v1.91.117.0 fix: pregate bats lanes, WTREE_UNAVAILABLE, doctor rows, SKILL.md size guard, planMode hosts (fix wave PR 5) (#3236)
* fix(pregate): recognise bats test files so lanes does not fail in bats repos (#3161)

gstack-pregate never treated a .bats file as a test, so in a repository whose
tests are bats files the lanes check failed every run with PREGATE_NO_LANE: a
.gstack/pregate.json dependency pointing at tests/foo.bats selected nothing, a
selection manifest printing .bats paths yielded 0 files, and a bats run that
executed no tests could not trip the zero-run rule.

Recognise .bats in the test-file pattern and in the manifest-output parser, and
read the bats TAP plan and pretty summary (as a last resort, after every
existing runner's summary) so a run that reports zero tests is a failure.

Fixes #3161.

Cherry-picked from #3177 (d355c116b). Conflicts resolved as in the author's own
merges (b09476beb, 444f4e768): lane-selection.ts keeps main's isTestFile-based
manifest parser, and lanes.ts carries the bats summary on main's supervised runner.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(pregate): a nested .bats file is selected and run end to end (#3161)

Reviewer test for #3177: touched itself with no declaration, and through a
declared dependency; a failing nested bats run reports the bats counts.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(pregate): a wtree failure is WTREE_UNAVAILABLE, not an unresolvable base (#3197)

The pregate consumers of the working-tree fingerprint (buildContext,
withFingerprint, verifyPregate) move to lib/script-invocation.ts's
{tree}|{error} form, and lib/regen.ts drops its string-returning wrapper.

- A fingerprint failure was reported as "policy source unavailable:
  origin/<base> is not a resolvable commit" with git fetch as the fix. It is
  now WTREE_UNAVAILABLE: a launch failure lists each Windows bash candidate
  with its rejection reason and names GSTACK_GIT_BASH; a gstack-wtree exit
  keeps the repair it printed (set TMPDIR to a writable directory).
- A fingerprint failing after a check is incomplete with that cause, not
  "working tree changed"; verify says the tree cannot be fingerprinted.
- #3190's gstack-config read (test_backend) goes through runScriptSync, so
  Windows reaches it through Git Bash. The tripwire learns the bin-constant
  shape that let it through.
- Under --json every context refusal is one envelope on stdout
  ({ok:false,status:'refused',code,detail,fix,docs}, exit 3); the run object
  carries ok and status too, and GSTACK_RESULT goes to stderr.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* ci(pregate): gstack-pregate joins the Windows pregate job; pregate tests return to the Windows lane (#3197)

- windows-pregate-e2e runs gstack-pregate run (preflight: regen, secrets
  through the gstack-redact bun bin, literals, guards) in both journeys, and
  asserts a missing GSTACK_GIT_BASH refuses with one WTREE_UNAVAILABLE --json
  envelope.
- test/helpers/pregate-fixture.ts and restamp-fixture.ts runBin launch bins
  through lib/script-invocation.ts (bun bins as Bun argv, bash bins through
  Git Bash on Windows), which removes the "shebang via CreateProcess" reason
  for the pregate tests; they move from KNOWN_WINDOWS_INCOMPATIBLE to
  KNOWN_WINDOWS_SAFE.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* doctor: list rescue refs and warn about CLAUDE.md skill lists

gstack-doctor's plain report gains two rows. 'rescue refs' lists each
refs/gstack/rescue/* commit in the install checkout, newest first, with its
date, a needs-manual-resolution label when the rescue held unmerged files, and
the shell-quoted inspect and delete commands bin/gstack-ff-update's banner
prints. 'CLAUDE.md' warns when a gstack section in ~/.claude/CLAUDE.md or the
project's CLAUDE.md enumerates skills (#3169) and prints the short block to
paste instead; docs/troubleshooting.md#claude-md-skill-list explains it.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* doctor: name the Windows bash gstack resolves and what chose it

On Windows (OSTYPE msys/cygwin) gstack-doctor adds a 'windows bash' row from
lib/script-invocation.ts resolveWindowsBash, run through the new thin
bin/gstack-windows-bash.ts with the bun on PATH: the resolved bash and the
override or discovery step that chose it, a warn when GSTACK_GIT_BASH and
GSTACK_BASH differ, and a fail carrying the BASH_UNAVAILABLE line (every
candidate with its rejection reason) when none qualifies. Tests inject the
platform (GSTACK_DOCTOR_PLATFORM) and the discovery inputs
(GSTACK_DOCTOR_BASH_PROBE) so the row runs on every OS.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* doctor --check: name the held-install recovery on Windows, not bin/gstack-upgrade

bin/gstack-upgrade compares a pwd -P path (/c/...) with git's toplevel
(C:/...) and refuses on every Git Bash install, so on Windows the revision
row's upgrade fix (REVISION_BELOW_MINIMUM, the Capy-wrapper variant and
REVISION_UNMET) names the held-install recovery forms from
docs/troubleshooting.md#auto-update-held-install-recovery instead: the Git
Bash form and the Windows PowerShell 5.1 form, each labelled by shell with the
install path quoted for that shell. Other platforms keep bin/gstack-upgrade.
A test pins both printed forms to the documented ones.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(skills): omit the plan-mode section on hosts without plan mode (#3168)

generatePlanModeInfo now honours capabilities.planMode. Hosts that declare
planMode:false (codex, agy, hermes, openclaw, and define-host's default:
factory, kiro, slate, cursor, gbrain) no longer render the
'## Plan Mode Safe Operations' and '## Skill Invocation During Plan Mode'
preamble sections. Claude, copilot and opencode are unchanged.

The codex, factory and agy ship goldens are regenerated. The clarity test that
pinned the plan-mode entry-order prose on every host now pins it on planMode
hosts and asserts its absence elsewhere (intentional contract change).

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(skills): guard generated SKILL.md size at 100,000 chars on every host (#3199)

A free test renders every host and fails when a generated SKILL.md exceeds
100,000 characters, naming host, skill, actual size, limit and the fix. The 89
current offenders are pinned at their exact size in
test/fixtures/skill-md-size-allowlist.json: growth past a pin fails, and a
shrink fails until the pin is lowered. The shared-preamble hoist is deferred
(owner Garry, revisit 2026-11-15).

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test(pregate): keep #3177's bats tests beside the nested-directory test

Moves #3177's bats describe block verbatim from test/pregate-selection.test.ts
into test/pregate-bats-nested.test.ts, so pregate-selection.test.ts is
unchanged on this branch and the in-flight field-list-2 pregate lane
(capy/fl2-L6-pregate, which edits that file's imports and adds cases) merges
without a conflict. No assertion changes.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* doctor --check: keep bin/gstack-upgrade as the Windows fix (path spelling fixed on main)

Reverts 37bac07f8. Plan item D named the held-install recovery forms on
Windows only until bin/gstack-upgrade stopped refusing Git Bash installs.
bin/gstack-upgrade:63-67 now compares with gstack_path_same from
bin/gstack-path-spelling.sh, pinned by test/gstack-upgrade-path-spelling.test.ts,
so bin/gstack-upgrade is the right fix on every platform again.

* test(skills): measure the size-guard pins on v1.91.109.0

#3225 (merged into this branch) grew 53 pinned skills: 3 characters each from
the Aside driver resolver, and about 340 more in plan-design-review and
plan-devex-review from the #3219 summary sections. The pins are this PR's
starting measurement, so they are taken on the merged base; no skill this PR
renders grew.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* ci(windows): only the pregate tests that pass on windows-latest join the lane

windows-free-tests run 38984824731 (this branch at 8ed6da402) ran the moved
files. gstack-regen and the new pregate-wtree-refusal passed. pregate-guards,
-remote, -regate, -invalidation, -selection and -bats-nested launched
gstack-pregate fine but fail on Linux-host expectations: their fixture lanes are
Linux workflow jobs, which a Windows host records as requires-remote
obligations. They go back to KNOWN_WINDOWS_INCOMPATIBLE with that reason
instead of "shebang via CreateProcess".

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* windows curation: keep CI run ids out of reason prose (pii.phone false positive)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* skill-md size pins: follow #3231's review-army line (+21 chars on five inlining hosts)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* doctor test: drop the reverted Windows upgrade-fix wording from item D's describe

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* skill-md size pins: follow #3226 (plan-devex-review +3324) and #3234 (autoplan +59)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* WTREE_UNAVAILABLE moves to lib/result-codes-release.ts (#3235 split; result-codes.ts stays at 796 of 800 lines); doctor fixture copies the release table; size pins follow #3235's review +1

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(changelog): fix wave PR 5 release notes

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* v1.91.117.0 fix: fix wave PR 5 (pregate bats + WTREE_UNAVAILABLE, doctor rows, SKILL.md size guard, planMode hosts)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* fix(qa): an invalid-input probe breaks only the rule it tests

The paid gate's qa-functional-cli-report failed again on #3236 (run
39001107574): the agent probed the digits-only amount rule with -5, which
the positivity check rejected first, so the planted parseInt defect stayed
hidden and it reported no defect. The contract map's invalid-input row now
asks for each input to break only the rule it tests, so another check
cannot reject it first.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

---------

Co-authored-by: harjothkhara <48686985+harjothkhara@users.noreply.github.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
2026-10-17 04:36:00 +00:00

119 lines
6.1 KiB
TypeScript

/**
* Generated SKILL.md size guard (#3199). Several agent runtimes and skill
* loaders cap one SKILL.md at 100,000 characters; Hermes' skill editor refuses
* every edit to a larger skill without naming the size. This free test renders
* every host and fails when any generated SKILL.md is over the limit.
*
* Skills already over the limit are pinned at their exact size in
* test/fixtures/skill-md-size-allowlist.json (a ratchet): growth past a pin
* fails, and a shrink below a pin fails until the pin is lowered. The shared
* preamble hoist that would retire most pins is deferred (owner Garry,
* revisit 2026-11-15).
*/
import { afterAll, describe, expect, test } from 'bun:test';
import * as fs from 'fs';
import * as os from 'os';
import * as path from 'path';
import { ALL_HOST_CONFIGS } from '../hosts/index';
const ROOT = path.resolve(import.meta.dir, '..');
const ALLOWLIST_FILE = 'test/fixtures/skill-md-size-allowlist.json';
interface Allowlist {
limit: number;
pins: Record<string, number>;
}
function countChars(text: string): number {
let count = 0;
for (const _ of text) count++;
return count;
}
function sizeViolations(sizes: Record<string, number>, allowlist: Allowlist): string[] {
const { limit, pins } = allowlist;
const violations: string[] = [];
for (const [key, actual] of Object.entries(sizes).sort(([a], [b]) => a.localeCompare(b))) {
const [host, skill] = key.split('/');
const pin = pins[key];
const where = `host ${host}, skill ${skill}: ${actual} chars, limit ${limit}`;
if (pin === undefined) {
if (actual > limit) violations.push(`${where}, not pinned. Shrink it to ${limit} chars or fewer.`);
} else if (actual > pin) {
violations.push(`${where}, pinned at ${pin}: grew by ${actual - pin}. Shrink it back to ${pin} chars or fewer.`);
} else if (actual < pin) {
violations.push(actual > limit
? `${where}, pinned at ${pin}: shrank by ${pin - actual}. Lower the pin "${key}" to ${actual} in ${ALLOWLIST_FILE}.`
: `${where}, pinned at ${pin}: now within the limit. Delete the pin "${key}" from ${ALLOWLIST_FILE}.`);
}
}
for (const key of Object.keys(pins).sort()) {
if (!(key in sizes)) violations.push(`pin "${key}" names no generated SKILL.md. Delete it from ${ALLOWLIST_FILE}.`);
}
return violations;
}
const SIZE_FIX = 'Fix: move skill-specific prose into sections/*.md that the skill reads on demand (on non-Claude hosts, also add the skill '
+ 'to CARVED_ON_EVERY_HOST in scripts/resolvers/sections.ts so its sections stop being inlined), or cut duplicated text; '
+ 'then run `bun run gen:skill-docs` and rerun this test. A pin may only go down: when a pinned skill shrinks, '
+ `lower its pin in ${ALLOWLIST_FILE} to the new size, and delete it once the skill is within the limit. `
+ 'Raising a pin is an owner decision. The shared-preamble hoist (#3199) is deferred: owner Garry, revisit 2026-11-15.';
function renderAllHosts(outDir: string): Record<string, number> {
const render = Bun.spawnSync(
['bun', 'run', 'scripts/gen-skill-docs.ts', '--host', 'all', '--out-dir', outDir],
{ cwd: ROOT, stdout: 'pipe', stderr: 'pipe', timeout: 180_000 },
);
if (render.exitCode !== 0) {
throw new Error(`gen-skill-docs --host all --out-dir failed (exit ${render.exitCode}):\n${render.stderr.toString()}`);
}
const sizes: Record<string, number> = {};
for (const host of ALL_HOST_CONFIGS) {
// Claude's tracked tree is what ships: an out-dir Claude render repoints
// section paths into the render, so it is not byte-identical.
const skillsRoot = host.name === 'claude' ? ROOT : path.join(outDir, host.hostSubdir, 'skills');
const skillDirs = fs.readdirSync(skillsRoot, { withFileTypes: true })
.filter(entry => entry.isDirectory() && !entry.name.startsWith('.'))
.map(entry => entry.name);
if (host.name === 'claude') skillDirs.push('.');
for (const dir of skillDirs) {
const file = path.join(skillsRoot, dir, 'SKILL.md');
if (!fs.existsSync(file)) continue;
if (host.name === 'claude' && !fs.existsSync(`${file}.tmpl`)) continue;
sizes[`${host.name}/${dir === '.' ? 'gstack' : dir}`] = countChars(fs.readFileSync(file, 'utf-8'));
}
}
return sizes;
}
describe('generated SKILL.md size guard (#3199)', () => {
const outDir = fs.mkdtempSync(path.join(os.tmpdir(), 'gstack-size-guard-'));
afterAll(() => fs.rmSync(outDir, { recursive: true, force: true }));
test('the ratchet flags over-limit, grown, shrunk, within-limit and stale entries', () => {
const allowlist = { limit: 100, pins: { 'h/grew': 150, 'h/shrank': 150, 'h/under': 150, 'h/same': 150, 'h/gone': 150 } };
const violations = sizeViolations({ 'h/new': 101, 'h/ok': 100, 'h/grew': 151, 'h/shrank': 149, 'h/under': 99, 'h/same': 150 }, allowlist);
expect(violations).toEqual([
'host h, skill grew: 151 chars, limit 100, pinned at 150: grew by 1. Shrink it back to 150 chars or fewer.',
'host h, skill new: 101 chars, limit 100, not pinned. Shrink it to 100 chars or fewer.',
`host h, skill shrank: 149 chars, limit 100, pinned at 150: shrank by 1. Lower the pin "h/shrank" to 149 in ${ALLOWLIST_FILE}.`,
`host h, skill under: 99 chars, limit 100, pinned at 150: now within the limit. Delete the pin "h/under" from ${ALLOWLIST_FILE}.`,
`pin "h/gone" names no generated SKILL.md. Delete it from ${ALLOWLIST_FILE}.`,
]);
expect(countChars('é—x')).toBe(3);
});
test('every host\'s generated SKILL.md is within 100,000 chars or at its exact pin', () => {
const allowlist: Allowlist = JSON.parse(fs.readFileSync(path.join(ROOT, ALLOWLIST_FILE), 'utf-8'));
expect(allowlist.limit).toBe(100_000);
const sizes = renderAllHosts(outDir);
for (const host of ALL_HOST_CONFIGS) {
expect(Object.keys(sizes).some(key => key.startsWith(`${host.name}/`)), `no rendered skills for ${host.name}`).toBe(true);
}
const violations = sizeViolations(sizes, allowlist);
if (violations.length > 0) {
throw new Error(`${violations.length} generated SKILL.md size violation(s):\n ${violations.join('\n ')}\n${SIZE_FIX}`);
}
});
});