fix(#4776): resolve the artifact-exists prompt under --auto (#4832)

* fix(#4776): resolve the artifact-exists prompt under --auto

/gsd-ui-phase <phase> --auto stopped at 'UI-SPEC.md already exists for
Phase {N}. What would you like to do?' whenever the file was on disk —
which is most often after an earlier run wrote the contract as a draft
and ended before its checker ran, exactly the state a re-run exists to
verify. Step 4 had no --auto arm; step 9.5 in the same file has had one
since it was added, which is how the drift went unnoticed.

Step 4 now auto-selects Skip: the existing UI-SPEC is left untouched and
the run proceeds to the checker. Skip is the only non-destructive choice
— Update re-runs the researcher, which rewrites the whole contract and
drops answers a person already recorded in it, and View exits without
verifying anything.

spec-phase.md's artifact-exists arm auto-selected 'Update it', which is
the same defect with the opposite sign: an unattended run regenerating a
spec nobody is watching. Per the decision recorded on #4776 — an
unattended run reuses an existing artifact rather than regenerating it —
it now auto-selects Skip and leaves the spec unchanged.

The max-revision-iterations escalation (Force approve / Edit manually /
Abandon) is deliberately untouched and pinned by a test: accepting
blocking findings is a decision a person makes.

Closes #4776

* chore(#4776): add changeset fragment

Emitted-Drift-Ack-Growth: ui-phase.md — --auto arm added to the existing-UI-SPEC branch (#4776)
Emitted-Drift-Ack-Growth: spec-phase.md — --auto arm reworded to reuse the existing SPEC (#4776)

* fix(#4776): extend the reuse-as-is --auto fix to the 3 sibling files

The PR's original scope claim -- that ai-integration-phase.md,
eval-review.md and ui-review.md were "scoped by triage to their own
issues" -- was false; no such issues existed, and it contradicted #4776's
own most recent (2026-09-16) triage comment, which explicitly widened the
fix to require all 5 files under one recommended fix.

Applies the same reuse-as-is pattern: ai-integration-phase.md mirrors
ui-phase.md's 3-way Update/View/Skip shape (auto-selects Skip);
eval-review.md and ui-review.md have only Re-audit/View (auto-selects
View, the only non-regenerating choice). None of the three had any prior
--auto handling at all -- each has exactly one AskUserQuestion call site
total, and it is the one this fix resolves, so an --auto run through any
of them no longer stalls anywhere.

Emitted-Drift-Ack-Growth: ai-integration-phase.md — the --auto arm reusing an existing AI-SPEC is the deliverable (#4776)
Emitted-Drift-Ack-Growth: eval-review.md — the --auto arm reusing an existing EVAL-REVIEW is the deliverable (#4776)
Emitted-Drift-Ack-Growth: ui-review.md — the --auto arm reusing an existing UI-REVIEW is the deliverable (#4776)

---------

Co-authored-by: Tom Boucher <trekkie@nomorestars.com>
This commit is contained in:
Michel Moreira
2026-09-23 12:13:34 -03:00
committed by GitHub
parent b90eef28e8
commit af822a8024
8 changed files with 153 additions and 6 deletions

View File

@@ -199,3 +199,91 @@ describe('bug #17: AskUserQuestion options arrays respect runtime cap', () => {
});
});
}
// ────────────────────────────────────────────────────────────────────────
// #4776 — an artifact-exists question must not block an unattended run
// ────────────────────────────────────────────────────────────────────────
{
const { describe: __autoDescribe, test: __autoTest } = require('node:test');
__autoDescribe('artifact-exists prompts resolve under --auto (#4776)', () => {
const WF = path.join(__dirname, '..', 'gsd-core', 'workflows');
/**
* The decision points this issue covers. Each names the heading (or, for
* files with no dedicated heading at this point, the arm's own guard
* clause) that opens the artifact-exists branch, the option the `--auto`
* arm must take, and the sibling option that regenerates the artifact
* (the one an `--auto` arm must NOT select).
*
* The maintainer decision recorded on #4776: an unattended run REUSES an
* existing artifact rather than regenerating it. "Update"/"Re-audit"
* re-run the generator, rewriting a contract or review that may already
* carry answers or findings a person recorded, with nobody present to
* notice — so an `--auto` arm that selects one of those is the defect,
* not the fix.
*
* Widened from the original ui-phase.md/spec-phase.md pair to all 5
* files sharing the gap (#4776's 2026-09-16 triage comment: "give all 5
* files a consistent --auto branch... recommended fix").
* ai-integration-phase.md mirrors ui-phase.md's 3-way Update/View/Skip
* shape; eval-review.md and ui-review.md have only Re-audit/View, so
* their non-destructive `--auto` choice is "View", not "Skip".
*/
const DECISION_POINTS = [
{ file: 'ui-phase.md', anchor: '## 4. Check Existing UI-SPEC', reuseOption: 'Skip', regenOption: 'Update' },
{ file: 'spec-phase.md', anchor: '**Check for existing SPEC.md:**', reuseOption: 'Skip', regenOption: 'Update' },
{ file: 'ai-integration-phase.md', anchor: '## 4. Check Existing AI-SPEC', reuseOption: 'Skip', regenOption: 'Update' },
{ file: 'eval-review.md', anchor: '**If `EVAL_REVIEW_FILE` non-empty:**', reuseOption: 'View', regenOption: 'Re-audit' },
{ file: 'ui-review.md', anchor: '**If `UI_REVIEW_FILE` non-empty:**', reuseOption: 'View', regenOption: 'Re-audit' },
];
/** The text from `anchor` up to the next heading of the same or higher level. */
function section(content, anchor) {
const start = content.indexOf(anchor);
assert.notEqual(start, -1, `anchor not found: ${anchor}`);
const rest = content.slice(start + anchor.length);
const end = rest.search(/\n##? /);
return rest.slice(0, end === -1 ? rest.length : end);
}
for (const { file, anchor, reuseOption, regenOption } of DECISION_POINTS) {
__autoTest(`${file}: the --auto arm reuses the existing artifact, before any prompt`, () => {
const body = section(fs.readFileSync(path.join(WF, file), 'utf8'), anchor);
const autoIdx = body.search(/\*\*If `--auto`:\*\*/);
assert.notEqual(autoIdx, -1,
`${file}: the artifact-exists branch has no --auto arm, so an unattended run stops here`);
// `Use AskUserQuestion`, not a bare mention: ui-phase.md's TEXT_MODE
// paragraph names AskUserQuestion above this branch, and anchoring on
// that would compare the arm against prose it has nothing to do with.
const askIdx = body.indexOf('Use AskUserQuestion');
assert.notEqual(askIdx, -1, `${file}: expected a \`Use AskUserQuestion\` prompt at this decision point`);
assert.ok(autoIdx < askIdx,
`${file}: the --auto arm must resolve BEFORE the prompt, or the prompt still runs`);
// The arm's own sentence — up to the end of that line — must name the
// reuse option. Scanning the whole section would match the interactive
// option list below it and pass on a file that auto-selects the
// regenerating option instead.
const armLine = body.slice(autoIdx, body.indexOf('\n', autoIdx) === -1 ? undefined : body.indexOf('\n', autoIdx));
assert.match(armLine, new RegExp(`"${reuseOption}"`),
`${file}: the --auto arm must select "${reuseOption}" (reuse as-is), not regenerate the artifact`);
assert.doesNotMatch(armLine, new RegExp(`"${regenOption}`),
`${file}: auto-selecting "${regenOption}" regenerates an artifact nobody is watching (#4776)`);
});
}
__autoTest('the max-revision-iterations escalation stays interactive under --auto', () => {
// #4776 explicitly does NOT ask for this one: force-approving blocking
// findings is a decision a person makes. A future "resolve every prompt
// under --auto" sweep would break that, so it is pinned here.
const body = fs.readFileSync(path.join(WF, 'ui-phase.md'), 'utf8');
const idx = body.indexOf('Force approve');
assert.notEqual(idx, -1, 'ui-phase.md: expected the max-iterations escalation to still exist');
const around = body.slice(Math.max(0, idx - 600), idx + 600);
assert.doesNotMatch(around, /\*\*If `--auto`:\*\*/,
'the force-approve escalation must keep blocking for a human under --auto');
});
});
}