Files
msd-core/tests/bug-936-no-nested-spawner-wrap.test.cjs
Tom Boucher e4e8a9fcb8 fix(#936): run plan-phase inline in convergence; guard against nested spawner wraps (#939)
Both sites in plan-review-convergence.md that wrapped gsd-plan-phase in
Agent() (initial planning + replan loop) are now bare Skill() calls at depth 0.
On Claude Code, a depth-1 Agent has no Agent tool so wrapped plan-phase could
never spawn gsd-planner/gsd-plan-checker — the replan loop silently produced no
revised plan when HIGHs were found. Running plan-phase inline from the depth-0
orchestrator (which retains the Agent tool) restores the full sub-agent chain.

A full audit of all workflow files confirmed these two sites were the only
instances of the anti-pattern (no other workflow wraps a spawner orchestrator
in Agent() without a RUNTIME carve-out).

Added structural guard test bug-936-no-nested-spawner-wrap.test.cjs that
dynamically derives the spawner set (workflows containing subagent_type=) and
asserts no workflow wraps a spawner inside Agent() without a RUNTIME != claude
carve-out — prevents silent regression. Test passes on fixed code, would fail
on pre-fix code at the two de-wrapped sites.

Also applied two low-severity prose nits flagged in review:
- commands/gsd/plan-review-convergence.md: orchestrator role updated to
  describe inline plan-phase + Agent for review (was generic "spawn Agents")
- gsd-core/workflows/plan-review-convergence.md success_criteria: narrowed
  "Each Agent fully completes" to the review Agent (plan-phase is inline now)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-09 12:37:32 -04:00

208 lines
9.3 KiB
JavaScript

'use strict';
/**
* Structural guard — bug(#936): plan-review-convergence wrapped gsd-plan-phase
* in Agent() at TWO sites (initial planning + replan). On Claude Code, a depth-1
* Agent has no Agent tool, so plan-phase cannot spawn gsd-planner / gsd-plan-checker
* → the replan loop never works when HIGHs are found.
*
* Fix: run plan-phase INLINE (bare Skill()) from the convergence orchestrator,
* which runs at depth 0 and has Agent available — exactly how autonomous.md,
* manager.md, and discuss-phase-assumptions.md already chain plan-phase.
*
* This guard dynamically derives the set of "spawner" workflows (those containing
* `subagent_type=`) and asserts that NO workflow wraps a spawner inside Agent()
* UNLESS the wrapping block includes a RUNTIME != claude carve-out (the #853
* pattern already applied to autonomous.md / manager.md).
*/
// allow-test-rule: source-text-is-the-product
// The workflow markdown IS the runtime instruction — static guards over
// workflow text are the canonical regression-test mechanism (per CONTRIBUTING
// exception matrix and tests/bug-853-bg-dispatch-runtime-gating.test.cjs).
const { describe, test } = require('node:test');
const assert = require('node:assert/strict');
const fs = require('node:fs');
const path = require('node:path');
const WORKFLOWS_DIR = path.join(__dirname, '..', 'gsd-core', 'workflows');
// ── 1. Derive spawner skill names dynamically ──────────────────────────────
// A "spawner" workflow is one that contains `subagent_type=` — it NEEDS the
// Agent tool to run and therefore cannot safely be wrapped in another Agent()
// on Claude Code (where depth-1 agents have no Agent tool).
// Recursively collect all *.md files under WORKFLOWS_DIR (covers nested fragments
// like discuss-phase/modes/*.md and execute-phase/steps/*.md).
function collectWorkflowFiles(dir) {
const entries = fs.readdirSync(dir, { withFileTypes: true });
const results = [];
for (const e of entries) {
const fullPath = path.join(dir, e.name);
if (e.isDirectory()) {
results.push(...collectWorkflowFiles(fullPath));
} else if (e.name.endsWith('.md')) {
results.push({
name: path.relative(WORKFLOWS_DIR, fullPath),
path: fullPath,
content: fs.readFileSync(fullPath, 'utf8'),
});
}
}
return results;
}
const allWorkflowFiles = collectWorkflowFiles(WORKFLOWS_DIR);
// Map: base-slug → workflow filename (e.g. "plan-phase" → "plan-phase.md")
// Skill() calls use the "gsd-<slug>" convention in all workflow files.
// We build BOTH the bare slug set and the gsd-prefixed skill-name set.
const SPAWNER_BASE_SLUGS = new Set(
allWorkflowFiles
.filter((w) => w.content.includes('subagent_type='))
.map((w) => w.name.replace(/\.md$/, ''))
);
// Skill invocations use "gsd-<slug>" (e.g. gsd-plan-phase, gsd-execute-phase).
// Build the regex from the prefixed names so it actually matches what workflows write.
const SPAWNER_GSD_NAMES = new Set([...SPAWNER_BASE_SLUGS].map((s) => `gsd-${s}`));
// Build a regex that matches Skill(skill='gsd-<spawner>') or Skill(skill="gsd-<spawner>")
const spawnerPattern = new RegExp(
`Skill\\(\\s*skill=['"](?:${[...SPAWNER_GSD_NAMES].join('|')})['"]`,
's'
);
// ── 2. Helper: extract Agent() blocks from a workflow ─────────────────────
// Each block starts at "Agent(" and ends at the balancing ")". We collect
// the text of each such block together with the surrounding context (a 400
// char window before the block) so we can check for RUNTIME carve-outs.
function extractAgentBlocks(content) {
const blocks = [];
let pos = 0;
while (pos < content.length) {
const start = content.indexOf('Agent(', pos);
if (start === -1) break;
// Walk forward to find the balancing closing paren
let depth = 0;
let i = start + 'Agent('.length - 1; // at the '('
for (; i < content.length; i++) {
if (content[i] === '(') depth++;
else if (content[i] === ')') {
depth--;
if (depth === 0) break;
}
}
const end = i + 1;
const blockText = content.slice(start, end);
// Capture context: 400 chars before the block (for RUNTIME gate detection)
const contextBefore = content.slice(Math.max(0, start - 400), start);
blocks.push({ start, end, blockText, contextBefore });
pos = end;
}
return blocks;
}
// ── 3. Helper: does a block have a RUNTIME != claude carve-out nearby? ────
// The #853 pattern looks like: "RUNTIME is `claude`" in a preceding condition
// that switches to inline Skill() instead of the Agent() block. A block is
// considered guarded when the 400-char context window before it (or the block
// body itself for block-internal guards) contains any of these markers.
function hasRuntimeCarveout(block) {
const haystack = block.contextBefore + block.blockText;
return (
/RUNTIME[^`\n]{0,30}(?:!=|≠|is not|!==)\s*[`'"]?claude/i.test(haystack) ||
/RUNTIME[^`\n]{0,30}claude[^`\n]{0,30}(?:inline|not.*Agent|do NOT)/i.test(haystack) ||
/If `RUNTIME` is `claude`/i.test(haystack) ||
/On Claude Code.*inline/is.test(haystack)
);
}
// ── 4. The guard: scan every workflow for unguarded Agent→spawner wraps ───
describe('bug-936 — no workflow wraps a spawner skill inside Agent() without a RUNTIME carve-out', () => {
test('spawner set is non-empty (self-check: subagent_type= grep must find files)', () => {
assert.ok(SPAWNER_BASE_SLUGS.size > 0, `No spawner workflows found in ${WORKFLOWS_DIR} — SPAWNER_BASE_SLUGS derivation is broken`);
// plan-phase must be a spawner (base slug)
assert.ok(SPAWNER_BASE_SLUGS.has('plan-phase'), 'plan-phase.md must be in the spawner set (contains subagent_type=)');
// gsd-plan-phase must be in the prefixed set used by the regex
assert.ok(SPAWNER_GSD_NAMES.has('gsd-plan-phase'), 'gsd-plan-phase must be in SPAWNER_GSD_NAMES — the prefixed form used in Skill() calls');
});
for (const wf of allWorkflowFiles) {
// Only scan files that have at least one Agent( call
if (!wf.content.includes('Agent(')) continue;
test(`${wf.name}: no Agent() block wraps a spawner Skill without a RUNTIME carve-out`, () => {
const blocks = extractAgentBlocks(wf.content);
const violations = blocks.filter((b) => {
const wrapsSpawner = spawnerPattern.test(b.blockText);
if (!wrapsSpawner) return false;
return !hasRuntimeCarveout(b);
});
assert.deepStrictEqual(
violations.map((v) => v.blockText.slice(0, 120).replace(/\n/g, '\\n')),
[],
`${wf.name} wraps a spawner Skill inside Agent() without a RUNTIME != claude carve-out.\n` +
`Fix: run the spawner Skill inline (bare Skill() call at depth 0) OR add a RUNTIME gate.\n` +
`See: bug #936, tests/bug-853-bg-dispatch-runtime-gating.test.cjs for the guarded pattern.`
);
});
}
});
// ── 5. Focused regression: plan-review-convergence never wraps plan-phase ─
describe('bug-936 — plan-review-convergence runs plan-phase inline, not inside Agent()', () => {
const CONVERGENCE = fs.readFileSync(
path.join(WORKFLOWS_DIR, 'plan-review-convergence.md'),
'utf8'
);
test('plan-review-convergence does NOT wrap gsd-plan-phase inside Agent()', () => {
// The anti-pattern: Agent( block whose body contains Skill(skill='gsd-plan-phase')
const blocks = extractAgentBlocks(CONVERGENCE);
const wrapping = blocks.filter((b) =>
/Skill\(\s*skill=['"]gsd-plan-phase['"]/.test(b.blockText) &&
!hasRuntimeCarveout(b)
);
assert.deepStrictEqual(
wrapping.map((v) => v.blockText.slice(0, 120).replace(/\n/g, '\\n')),
[],
'plan-review-convergence must NOT wrap gsd-plan-phase inside Agent(). ' +
'Run it inline (bare Skill() at depth 0) so it can spawn gsd-planner/gsd-plan-checker. ' +
'See: bug #936'
);
});
test('plan-review-convergence calls gsd-plan-phase inline (bare Skill call outside Agent block)', () => {
// After the fix: at least one bare Skill(skill="gsd-plan-phase") must appear
// outside any Agent( block — that is the inline call from the depth-0 orchestrator.
const blocks = extractAgentBlocks(CONVERGENCE);
// Remove all Agent block ranges from the text
let masked = CONVERGENCE;
// Work from end to start so offsets stay valid
const sorted = [...blocks].sort((a, b) => b.start - a.start);
for (const b of sorted) {
masked = masked.slice(0, b.start) + ' '.repeat(b.end - b.start) + masked.slice(b.end);
}
const hasInlineCall = /Skill\(\s*skill=["']gsd-plan-phase["']/.test(masked);
assert.ok(
hasInlineCall,
'plan-review-convergence must contain at least one bare Skill(skill="gsd-plan-phase") ' +
'outside any Agent() block — this is the inline call that lets plan-phase spawn its sub-agents. ' +
'See: bug #936'
);
});
test('plan-review-convergence still wraps gsd-review inside Agent() (leaf — isolation is correct)', () => {
// gsd-review is a leaf (shells out via Bash, no subagent_type) so the Agent wrap is fine and intentional.
const blocks = extractAgentBlocks(CONVERGENCE);
const reviewWrap = blocks.some((b) => /Skill\(\s*skill=['"]gsd-review['"]/.test(b.blockText));
assert.ok(reviewWrap, 'gsd-review must still be wrapped in Agent() — it is a Bash leaf and isolation is intentional');
});
});