fix(hooks): spawned-session escape in Conductor AUQ deny; override coverage in AUQ-error fallback (#2733)

Hooks inherit the harness env, so a per-command GSTACK_SESSION_KIND
prefix inside a subagent's bash can never reach them. Levers added:
a deterministic [conductor][spawned] auto-choose deny for env-level
spawned sessions (OPENCLAW_SESSION or session-wide GSTACK_SESSION_KIND),
and a spawned escape sentence appended to both hooks' prose directives
so a marked subagent that slips and calls AUQ resolves to auto-choose
instead of prose-STOP. The sentence lives in one shared constant
(hosts/claude/hooks/spawned-directive.ts) so the two paths can never
drift; destructive semantics are unified to conservative-continue.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Garry Tan
2026-08-30 23:42:38 +00:00
co-authored by Claude Fable 5
parent 5486cd6620
commit 9c7b0a2c95
5 changed files with 158 additions and 2 deletions
@@ -32,6 +32,7 @@ import * as fs from 'fs';
import * as path from 'path';
import * as os from 'os';
import { runBin } from './spawn-bin';
import { SPAWNED_ESCAPE_SENTENCE } from './spawned-directive';
interface HookStdin {
tool_name?: string;
@@ -158,12 +159,17 @@ export function directiveFor(kind: 'spawned' | 'headless' | 'interactive'): stri
);
case 'interactive':
default:
// #2733: the shell-out above runs in the HARNESS env, so a subagent
// marked spawned via a per-command env prefix still classifies as
// interactive here — the escape sentence is the only lever for that
// topology (see spawned-directive.ts).
return (
lead +
'SESSION_KIND=interactive — render the decision as a PROSE message now: a clear ELI10 of the issue, ' +
'then a Recommendation line, then ONE paragraph per choice carrying its `(recommended)` marker, its ' +
'`Completeness: X/10`, and 2-4 sentences of reasoning. Tell the user to reply with a letter, then STOP. ' +
'(Retry the call once first only if no answer could have surfaced.)'
'(Retry the call once first only if no answer could have surfaced.) ' +
SPAWNED_ESCAPE_SENTENCE
);
}
}
+14 -1
View File
@@ -46,6 +46,7 @@ import * as os from 'os';
import { runBin, repoRoot } from './spawn-bin';
import { isConductor } from '../../../lib/is-conductor';
import { classifyQuestion } from '../../../scripts/one-way-doors';
import { SPAWNED_ESCAPE_SENTENCE, CONDUCTOR_SPAWNED_DENY_REASON, spawnedByEnv } from './spawned-directive';
interface HookStdin {
session_id?: string;
@@ -484,6 +485,17 @@ async function main(): Promise<void> {
// preference, or door type — including one-way doors, which must reach the
// human via prose rather than the unreliable tool.
if (isConductor()) {
// #2733: env-level spawned sessions (OpenClaw inside a Conductor
// workspace, or a harness launched with GSTACK_SESSION_KIND=spawned in
// its env) get an auto-choose deny — a prose brief has no reader there.
// LIMITATION: a per-command GSTACK_SESSION_KIND prefix inside a
// subagent's bash never reaches this hook (hooks inherit the harness
// env); that case is covered by the escape sentence below plus the
// dispatching skill's prompt.
if (spawnedByEnv()) {
deny(CONDUCTOR_SPAWNED_DENY_REASON + (memoryContext ? `\n${memoryContext}` : ''));
return;
}
const conductorReason =
'[conductor] AskUserQuestion is unreliable in Conductor (native disabled, MCP variant flaky). ' +
'Do NOT call AskUserQuestion (native or any mcp__*__AskUserQuestion). Render this decision as a ' +
@@ -491,7 +503,8 @@ async function main(): Promise<void> {
'paragraph per choice carrying its `(recommended)` marker and `Completeness: X/10`; tell the user ' +
'to reply with a letter, then STOP. For a one-way/destructive confirmation, require an explicit ' +
'typed confirmation and do NOT proceed on a vague reply. Capture the decision with gstack-question-log ' +
'(PostToolUse will not fire on a prose path).' +
'(PostToolUse will not fire on a prose path). ' +
SPAWNED_ESCAPE_SENTENCE +
(memoryContext ? `\n${memoryContext}` : '');
deny(conductorReason);
return;
+47
View File
@@ -0,0 +1,47 @@
/**
* Shared spawned-session directive text for the AUQ hooks (#2733).
*
* Hook processes inherit the HARNESS env, so a per-command
* `GSTACK_SESSION_KIND=spawned` prefix inside a subagent's bash call can
* never reach a hook — the only levers a hook has for the subagent case are
* (a) env-level markers that ARE session-wide (OPENCLAW_SESSION, or a harness
* launched with GSTACK_SESSION_KIND in its env) and (b) directive TEXT the
* model reads. Both AUQ hooks (question-preference PreToolUse deny,
* auq-error-fallback PostToolUse directive) carry the same escape sentence;
* it lives here as one constant so the two paths can never drift into
* contradictory instructions.
*
* Destructive semantics are unified across every spawned surface (dispatch
* prompt, spawned-session block, AUQ prose rule, both hooks):
* conservative-continue, never prose-STOP — a prose brief with no reader is
* always wrong in a spawned session, and the conservative choice guarantees
* nothing irreversible happens.
*/
/** Appended to prose-directing hook texts so a marked subagent that slips
* and calls AUQ still resolves to auto-choose instead of prose-STOP. */
export const SPAWNED_ESCAPE_SENTENCE =
'If this session was spawned by an orchestrator or a parent agent and no human reads its ' +
'output mid-run (e.g. your dispatch prompt says you are a spawned subagent), do not render ' +
'the prose brief either — auto-choose the recommended option and continue; at a destructive ' +
'or irreversible gate, do not execute the destructive action: take the conservative ' +
'non-destructive choice (skip/defer), record it, and continue.';
/** Deterministic deny reason for env-detected spawned sessions inside Conductor. */
export const CONDUCTOR_SPAWNED_DENY_REASON =
'[conductor][spawned] AskUserQuestion is unreliable in Conductor and this session is ' +
'orchestrator-spawned — no human reads its output. Do NOT retry the tool and do NOT render ' +
'a prose decision brief: auto-choose the recommended option for each question above, note ' +
'the choice, and continue the workflow. Exception: never auto-approve a destructive or ' +
'irreversible option — take the conservative non-destructive choice (skip/defer), note it, ' +
'and continue.';
/**
* Env-level spawned detection (direct env read — PreToolUse hot path, no
* shell-out). Mirrors bin/gstack-session-kind steps 0-1. True only for
* session-wide markers; a per-command prefix in subagent bash is invisible
* here by construction.
*/
export function spawnedByEnv(env: NodeJS.ProcessEnv = process.env): boolean {
return !!env.OPENCLAW_SESSION || env.GSTACK_SESSION_KIND === 'spawned';
}