fix(hooks): passThrough() two-branch contract — never emit permissionDecision:'defer' (#2035, #2006)

Every AskUserQuestion died with "Tool result missing due to internal error"
on current Claude Code builds (Desktop 1.14271.0, CC 2.1.177). Root cause:
the question-preference-hook emitted permissionDecision:'defer' on every
pass-through path. 'defer' is a real PreToolUse value, but since CC v2.1.89
its semantics are "pause this tool call for external resumption" (headless
resume) — never "abstain". Interactive sessions have nothing to resume the
paused call, so the tool orphaned. Pre-2.1.89 builds ignored the unknown
value, which is why the hook worked when it shipped and broke later.

The fix is the two-branch pass-through contract:
- no context -> exit 0 with EXACTLY empty stdout
- memory nuggets present -> hookSpecificOutput with hookEventName +
  additionalContext ONLY (the documented shape; plan-tune Layer 8 memory
  injection ships through this branch and keeps working)

defer() is renamed passThrough() so the function says what it does, and
docs/spikes/claude-code-hook-mutation.md's protocol contract (cited by the
hook header) is corrected in the same commit — it taught '"defer" — let
permission flow continue' and was the reintroduction vector.

Test contract rewritten in the same commit (13 assertions across 3 files,
verified fail-first against the unfixed hook): pass-through paths assert
exact-empty stdout (a garbage/partial write cannot slip past an
optional-chained parse), the nugget path asserts permissionDecision is
ABSENT while additionalContext survives, and a new tripwire asserts no
non-deny path ever puts the string "permissionDecision" on stdout. The
deny (auto-decide) and Conductor prose-redirect paths are unchanged.

Deployment: no migration needed — settings.json points at the absolute
bash shim which execs the .ts live; /gstack-upgrade delivers the fix.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Garry Tan 2026-07-09 19:08:49 -07:00
parent c4efc2a4f7
commit 3d97863b14
No known key found for this signature in database
GPG Key ID: C1F69E85C74EFE1D
5 changed files with 137 additions and 61 deletions

View File

@ -51,7 +51,13 @@ Optional in subagent context: `agent_id`, `agent_type`.
- `"deny"` — block (feedback to Claude, NOT a synthetic answer per Codex
correction in D-prefixed decisions)
- `"ask"` — escalate to user
- `"defer"` — let permission flow continue
- `"defer"` — pause the tool call for EXTERNAL resumption (Claude Code
v2.1.89+, a headless feature: resume with `-p --resume` to re-evaluate).
NEVER emit this to mean "no opinion" — in an interactive session nothing
resumes the paused call and the tool dies with "Tool result missing due to
internal error" (#2035, #2006). To abstain, exit 0 with EMPTY stdout
(optionally `hookSpecificOutput` with `additionalContext` only, no
`permissionDecision`).
**`updatedInput` semantics:** shallow merge of fields present in the returned
object onto the original `tool_input`. Only valid with
@ -106,15 +112,20 @@ required for our hook to fire there.
}
```
**Pass-through (no preference, or one-way safety override):**
**Pass-through (no preference, or one-way safety override):** exit 0 with
EMPTY stdout. When there is context to inject (plan-tune memory nuggets),
emit `additionalContext` WITHOUT a `permissionDecision`:
```json
{
"hookSpecificOutput": {
"hookEventName": "PreToolUse",
"permissionDecision": "defer"
"additionalContext": "[plan-tune memory] Past answers suggest: ..."
}
}
```
(Historical note: this example originally emitted `permissionDecision:
"defer"`, which broke every AskUserQuestion once CC v2.1.89 gave 'defer'
pause-for-resume semantics — #2035.)
**PostToolUse capture (always):**
```json

View File

@ -12,11 +12,15 @@
* 2. Look up door_type from scripts/question-registry.ts (default two-way).
* 3. Read preferences with precedence: project-local > global (D8).
* 4. Apply:
* never-ask + one-way defer (safety override; one-way always asks).
* never-ask + one-way pass through (safety override; one-way always asks).
* never-ask + two-way + marker deny with auto-decided recommendation
* in reason. Mark tool_use_id so PostToolUse logs as 'auto-decided'.
* ask-only-for-one-way + two-way + marker same as never-ask.
* always-ask, or no preference defer.
* always-ask, or no preference pass through.
*
* Pass-through = exit 0 with empty stdout (or additionalContext-only output
* when memory nuggets exist) NEVER permissionDecision:'defer', whose
* CC v2.1.89+ semantics are pause-for-external-resumption (#2035, #2006).
*
* Why deny+reason instead of allow+updatedInput:
* AskUserQuestion's `updatedInput` shape for "pre-resolve this question"
@ -31,7 +35,7 @@
* - First: (recommended) label suffix on an option.
* - Fall back: "Recommendation: X" prose match against option labels.
* - Refuse to auto-decide if ambiguous (multiple labels OR no parseable
* recommendation): defer instead of silent-wrong.
* recommendation): pass through instead of silent-wrong.
*
* Always exits 0. Hook errors land in ~/.gstack/hook-errors.log.
* See docs/spikes/claude-code-hook-mutation.md for the protocol contract.
@ -92,13 +96,25 @@ function readStdin(): Promise<string> {
});
}
function defer(additionalContext?: string): void {
const out: Record<string, unknown> = {
hookEventName: 'PreToolUse',
permissionDecision: 'defer',
};
if (additionalContext) out.additionalContext = additionalContext;
process.stdout.write(JSON.stringify({ hookSpecificOutput: out }));
function passThrough(additionalContext?: string): void {
// Abstain = exit 0 with EMPTY stdout (#2035, #2006). Never emit a
// permissionDecision here: 'defer' is a real PreToolUse value, but since
// Claude Code v2.1.89 its semantics are "pause this tool call for external
// resumption" (a headless-resume feature) — NOT "no opinion". In an
// interactive session nothing resumes the paused call, so every
// AskUserQuestion died with "Tool result missing due to internal error".
// additionalContext-only hookSpecificOutput is the documented shape for
// injecting context (plan-tune memory nuggets) without a decision.
if (additionalContext) {
process.stdout.write(
JSON.stringify({
hookSpecificOutput: {
hookEventName: 'PreToolUse',
additionalContext,
},
}),
);
}
process.exit(0);
}
@ -347,7 +363,7 @@ function logAutoDecided(
async function main(): Promise<void> {
const raw = await readStdin();
if (!raw.trim()) {
defer();
passThrough();
return;
}
let stdin: HookStdin;
@ -355,7 +371,7 @@ async function main(): Promise<void> {
stdin = JSON.parse(raw);
} catch (e) {
logHookError(`stdin parse failed: ${(e as Error).message}`);
defer();
passThrough();
return;
}
@ -364,26 +380,26 @@ async function main(): Promise<void> {
toolName !== 'AskUserQuestion' &&
!toolName.match(/^mcp__.+__AskUserQuestion$/)
) {
defer();
passThrough();
return;
}
const questions = stdin.tool_input?.questions || [];
if (questions.length === 0) {
defer();
passThrough();
return;
}
// For multi-question AUQ, enforcement is all-or-nothing per call:
// we deny only if ALL questions have marker + never-ask + safe door type.
// Mixed cases pass through (defer) so the user still gets to answer.
// Mixed cases pass through so the user still gets to answer.
const registry = loadRegistry();
const slug = slugFromCwd(stdin.cwd);
const memoryNuggets = loadMemoryNuggets(stdin.session_id);
// Compute Layer 8 memory context inline: any nuggets matching the
// signal_keys of the questions in this AUQ get surfaced as additionalContext.
// This applies whether we defer OR deny — gives the agent + user the
// This applies whether we pass through OR deny — gives the agent + user the
// relevant prior context either way.
const contextNuggets: string[] = [];
for (const q of questions) {
@ -402,7 +418,7 @@ async function main(): Promise<void> {
: undefined;
// Determine whether EVERY question is eligible for never-ask auto-decide.
// We deliberately do NOT early-return defer on the first ineligible question:
// We deliberately do NOT early-return pass-through on the first ineligible question:
// a Conductor session still needs the [conductor] prose deny as a fallback,
// so we compute eligibility, then branch. memoryContext is preserved on every
// non-enforcing exit. (All-or-nothing per-call semantics are unchanged: any
@ -471,10 +487,10 @@ async function main(): Promise<void> {
return;
}
defer(memoryContext);
passThrough(memoryContext);
}
main().catch((e) => {
logHookError(`main crash: ${(e as Error).message}`);
defer();
passThrough();
});

View File

@ -2,7 +2,7 @@
* Layer 8 memory cache + injection (plan-tune cathedral T12).
*
* Verifies the PreToolUse hook reads ~/.gstack/free-text-memory.json and
* surfaces matching nuggets via additionalContext on the hook response.
* surfaces matching nuggets via additionalContext-only output (#2035: never a permissionDecision).
* Cache: per-session memory-cache.json populated on first read, sub-1ms
* thereafter (D13 perf).
*/
@ -43,9 +43,9 @@ function runHook(stdin: object): { stdout: string; stderr: string; status: numbe
env.GSTACK_STATE_ROOT = stateRoot;
env.GSTACK_QUESTION_LOG_NO_DERIVE = '1';
delete env.GSTACK_HOME;
// These cases assert the defer-path memoryContext injection. Strip ambient
// These cases assert the pass-through memoryContext injection. Strip ambient
// Conductor markers so running inside Conductor (CONDUCTOR_WORKSPACE_PATH/PORT
// set) doesn't flip the hook into the [conductor] prose deny instead of defer.
// set) doesn't flip the hook into the [conductor] prose deny instead of pass-through.
delete env.CONDUCTOR_WORKSPACE_PATH;
delete env.CONDUCTOR_PORT;
const res = spawnSync(HOOK, [], {
@ -69,7 +69,7 @@ function runHook(stdin: object): { stdout: string; stderr: string; status: numbe
// ----------------------------------------------------------------------
describe('memory injection', () => {
test('injects matching nugget into additionalContext on defer', () => {
test('injects matching nugget into additionalContext on pass-through', () => {
writeMemory([
{
nugget: 'User prefers verbose explanations with tradeoffs',
@ -91,7 +91,10 @@ describe('memory injection', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
// #2035: nugget delivery is additionalContext-ONLY — a permissionDecision
// here (any value) would orphan the tool call on CC >= 2.1.89.
expect('permissionDecision' in (r.parsed?.hookSpecificOutput ?? {})).toBe(false);
expect(r.parsed?.hookSpecificOutput?.hookEventName).toBe('PreToolUse');
expect(r.parsed?.hookSpecificOutput?.additionalContext).toContain('verbose explanations');
});
@ -115,8 +118,9 @@ describe('memory injection', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expect(r.parsed?.hookSpecificOutput?.additionalContext).toBeUndefined();
// No nugget → pure pass-through: exit 0 with EXACTLY empty stdout.
expect(r.status).toBe(0);
expect(r.stdout).toBe('');
});
test('caps to 3 most-recent nuggets when many match', () => {
@ -219,7 +223,8 @@ describe('per-session memory cache', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expect(r.parsed?.hookSpecificOutput?.additionalContext).toBeUndefined();
// No nugget → pure pass-through: exit 0 with EXACTLY empty stdout.
expect(r.status).toBe(0);
expect(r.stdout).toBe('');
});
});

View File

@ -3,15 +3,18 @@
*
* Covers:
* - never-ask + marker + two-way + clean recommendation deny+reason
* - never-ask + no marker defer (D18 marker gate)
* - never-ask + one-way defer (safety override)
* - never-ask + ambiguous recommendation defer (D2 refuse-on-ambiguous)
* - always-ask defer
* - no preference defer
* - never-ask + no marker pass-through (D18 marker gate)
* - never-ask + one-way pass-through (safety override)
* - never-ask + ambiguous recommendation pass-through (D2 refuse-on-ambiguous)
* - always-ask pass-through
* - no preference pass-through
* - project preference wins over global (D8 precedence)
* - global preference applies when no project preference set
* - mcp__*__AskUserQuestion matcher accepted
* - empty stdin defer (crash safety)
* - empty stdin pass-through (crash safety)
*
* Pass-through contract (#2035/#2006): exit 0 + EXACTLY empty stdout, or
* additionalContext-only hookSpecificOutput never a permissionDecision.
* - auto-decided event logged via gstack-question-log (PostToolUse won't fire)
* - auto-decided marker written to ~/.gstack/sessions/<id>/.auto-decided-<tool_use_id>
*/
@ -97,6 +100,20 @@ function runHook(stdin: object, cwd?: string, extraEnv?: Record<string, string>)
};
}
/**
* #2035/#2006 contract: pass-through (abstain) is exit 0 with EXACTLY empty
* stdout never a permissionDecision. 'defer' is a real PreToolUse value,
* but its semantics are pause-for-external-resumption (CC v2.1.89+), so
* emitting it orphans the tool call in interactive sessions. Exact-empty
* (not trim) is deliberate: whitespace on stdout is still hook output, and a
* garbage/partial write must fail this assertion rather than slip past an
* optional-chained parse.
*/
function expectPassThrough(r: { status: number; stdout: string }): void {
expect(r.status).toBe(0);
expect(r.stdout).toBe('');
}
function autoDecidedEvents(): Array<Record<string, unknown>> {
const f = path.join(stateRoot, 'projects', cwdSlug, 'question-log.jsonl');
if (!fs.existsSync(f)) return [];
@ -113,8 +130,8 @@ function autoDecidedEvents(): Array<Record<string, unknown>> {
// Defer paths
// ----------------------------------------------------------------------
describe('defers (no enforcement)', () => {
test('no preference set → defer', () => {
describe('passes through (no enforcement)', () => {
test('no preference set → pass-through (empty stdout, no permissionDecision)', () => {
const r = runHook({
session_id: 's1',
tool_name: 'AskUserQuestion',
@ -125,11 +142,10 @@ describe('defers (no enforcement)', () => {
],
},
});
expect(r.status).toBe(0);
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
test('marker missing → defer (D18)', () => {
test('marker missing → pass-through (D18)', () => {
writeProjectPref('test-q', 'never-ask');
const r = runHook({
session_id: 's2',
@ -141,10 +157,10 @@ describe('defers (no enforcement)', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
test('always-ask preference → defer', () => {
test('always-ask preference → pass-through', () => {
writeProjectPref('test-q', 'always-ask');
const r = runHook({
session_id: 's3',
@ -156,10 +172,10 @@ describe('defers (no enforcement)', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
test('empty stdin → defer (crash safety)', () => {
test('empty stdin → pass-through (crash safety)', () => {
const env: Record<string, string> = {};
for (const [k, v] of Object.entries(process.env)) {
if (v !== undefined) env[k] = v;
@ -167,14 +183,39 @@ describe('defers (no enforcement)', () => {
env.GSTACK_STATE_ROOT = stateRoot;
const res = spawnSync(HOOK, [], { env, input: '', encoding: 'utf-8' });
expect(res.status).toBe(0);
const parsed = JSON.parse(res.stdout || '{}');
expect(parsed.hookSpecificOutput?.permissionDecision).toBe('defer');
expect(res.stdout).toBe('');
});
test('non-AUQ tool_name → defer (defensive)', () => {
test('non-AUQ tool_name → pass-through (defensive)', () => {
writeProjectPref('test-q', 'never-ask');
const r = runHook({ session_id: 's4', tool_name: 'Bash', tool_use_id: 'tu-4', tool_input: {} });
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
// #2035 tripwire: no non-deny/non-allow path may EVER put the string
// "permissionDecision" on stdout. Emitting one on a pass-through path (any
// value — 'defer' included) hands the platform a decision where the hook
// has none, and 'defer' specifically pauses the call for a resumption that
// never comes in interactive sessions.
test('pass-through stdout never contains "permissionDecision" (#2035)', () => {
const paths = [
runHook({
session_id: 's-trip-1',
tool_name: 'AskUserQuestion',
tool_use_id: 'tu-trip-1',
tool_input: {
questions: [
{ question: '<gstack-qid:test-q> Approve?', options: ['A) Yes (recommended)', 'B) No'] },
],
},
}),
runHook({ session_id: 's-trip-2', tool_name: 'Bash', tool_use_id: 'tu-trip-2', tool_input: {} }),
runHook({ session_id: 's-trip-3', tool_name: 'AskUserQuestion', tool_use_id: 'tu-trip-3', tool_input: { questions: [] } }),
];
for (const r of paths) {
expect(r.status).toBe(0);
expect(r.stdout).not.toContain('"permissionDecision"');
}
});
});
@ -204,7 +245,7 @@ describe('enforces never-ask preferences', () => {
expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toContain('Fix now');
});
test('one-way door → defer even with never-ask (safety override)', () => {
test('one-way door → pass-through even with never-ask (safety override)', () => {
writeProjectPref('ship-test-failure-triage', 'never-ask');
const r = runHook({
session_id: 's6',
@ -219,10 +260,10 @@ describe('enforces never-ask preferences', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
test('ambiguous recommendation (two labels) → defer (D2 refuse-on-ambiguous)', () => {
test('ambiguous recommendation (two labels) → pass-through (D2 refuse-on-ambiguous)', () => {
writeProjectPref('ship-pre-landing-review-fix', 'never-ask');
const r = runHook({
session_id: 's7',
@ -237,10 +278,10 @@ describe('enforces never-ask preferences', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
test('no recommendation marker AND no prose match → defer', () => {
test('no recommendation marker AND no prose match → pass-through', () => {
writeProjectPref('ship-pre-landing-review-fix', 'never-ask');
const r = runHook({
session_id: 's8',
@ -255,7 +296,7 @@ describe('enforces never-ask preferences', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
});
@ -301,7 +342,7 @@ describe('precedence: project wins over global (D8)', () => {
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('deny');
});
test('project always-ask + global never-ask → defer (project wins)', () => {
test('project always-ask + global never-ask → pass-through (project wins)', () => {
writeProjectPref('ship-pre-landing-review-fix', 'always-ask');
writeGlobalPref('ship-pre-landing-review-fix', 'never-ask');
const r = runHook({
@ -317,7 +358,7 @@ describe('precedence: project wins over global (D8)', () => {
],
},
});
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
});
@ -437,13 +478,13 @@ describe('Conductor prose redirect', () => {
expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).not.toContain('[conductor]');
});
test('non-AUQ tool in Conductor → still defer (no redirect on unrelated tools)', () => {
test('non-AUQ tool in Conductor → still pass-through (no redirect on unrelated tools)', () => {
const r = runHook(
{ session_id: 'c6', tool_name: 'Bash', tool_use_id: 'tu-c6', tool_input: {} },
undefined,
CONDUCTOR,
);
expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
expectPassThrough(r);
});
});

View File

@ -296,7 +296,10 @@ describeIfSelected('PlanTune cathedral E2E: annotation', ['plan-tune-annotation'
});
expect(res.status).toBe(0);
const parsed = JSON.parse(res.stdout || '{}');
expect(parsed.hookSpecificOutput?.permissionDecision).toBe('defer');
// #2035: memory-nugget delivery is additionalContext-ONLY. Emitting a
// permissionDecision here (the old 'defer') pauses the tool call for a
// resumption that never comes in interactive sessions.
expect('permissionDecision' in (parsed.hookSpecificOutput ?? {})).toBe(false);
expect(parsed.hookSpecificOutput?.additionalContext).toContain('verbose explanations');
});
});