CLAUDE_PLUGIN_DATA is empty in the Bash tool's process env, and the Phase 5
bash snippet is the cap's only caller. resolveLedgerPath() returned null there
and allowTurn() failed closed, so the budget gate denied turn 1 of every real
run: the loop this delivery exists to bound could never spend a turn, and the
pre-registered measurement could not be run at all.
resolveDataRoot() is now the single root for everything the loop writes --
CLAUDE_PLUGIN_DATA when the harness sets it, ~/.claude/voyage when it does
not. Three consumers resolve through it, which is the point: the cap ledger,
the PreToolUse hook's scope-marker lookup, and the command's bash snippets.
A writer and a reader that resolved the root separately are what made the
enforcement hook allow unconditionally in every real run while CLAUDE.md and
docs/architecture.md called it enforcing.
Same root cause, same commit:
- Marker write and remove now share ONE absolute-path guard and one root; the
write requires a non-empty CLAUDE_CODE_SESSION_ID before composing the path
(unset, the marker was named `.json`, which no lookup matches and no TTL
sweep cleans up).
- The per-turn gates resolve VOYAGE_ROOT with a plugin-cache fallback and
reserve exit 2 for "gate could not run". Interpolating an empty
${CLAUDE_PLUGIN_ROOT} ran `node /lib/...` -> exit 1, which the contract read
as "privacy gate says no" -- an unsatisfiable rewrite loop no query could
clear.
Two now-unreachable deny branches are removed rather than left as dead safety
claims (allowTurn's no_plugin_data_dir; the hook's uncountable-ledger deny).
The fail-closed stance stays where it is still real: a ledger that cannot be
WRITTEN denies the turn.
Verified end-to-end through the real bash snippets and the real hook with both
variables stripped and HOME sandboxed: marker written under the fallback root,
8 turns spent, 9th denied, hook exits 2, and exits 0 again after removal.
Note: the fallback exit-2 branch fires against the installed v5.9.1 cache,
which predates lib/util/research-loop-cap.mjs -- correct behaviour, and it
clears when the plugin is reinstalled.
Review findings 2670c10a, fbd6d534, 93550dfb.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011vPSXe88qp5aqWUqbDNWoF
145 lines
5.9 KiB
JavaScript
145 lines
5.9 KiB
JavaScript
#!/usr/bin/env node
|
||
// Hook: pre-agent-cap.mjs
|
||
// Event: PreToolUse (WebSearch | WebFetch | Task)
|
||
// Purpose: Enforce the /trekresearch Phase 5 loop bound at the harness level,
|
||
// so the cap is a reader that fells rather than prose the model obeys.
|
||
//
|
||
// Why this exists: the Phase 5 budget gate (lib/util/research-loop-cap.mjs) is
|
||
// invoked BY the loop. A gate the caller chooses to consult is advice. The
|
||
// spike in docs/spike-pretooluse-subagent-reach.md (RESULT: FIRES) established
|
||
// that a plugin PreToolUse hook does observe tool calls made INSIDE sub-agents
|
||
// on CC 2.1.226, which is what makes a second, non-optional gate possible.
|
||
//
|
||
// Two limits carried over from that spike, neither of which changes the design:
|
||
// - Reach was measured on one CC version and regressed once before (#34692),
|
||
// so this hook is defence in depth, never the only gate. research-loop-cap
|
||
// must stay correct if this hook silently stops firing.
|
||
// - Whether a blocking (exit 2) decision from inside a sub-agent propagates
|
||
// usefully was NOT measured — the probe always exited 0 by design.
|
||
//
|
||
// Scope key — the property that makes this safe to wire globally:
|
||
// session_id + a scope marker file that only the Phase 5 loop writes, at
|
||
// <data root>/trekresearch-loop-scope/<session_id>.json, where the data root
|
||
// comes from research-loop-cap.mjs's resolveDataRoot() — the same function
|
||
// the writer resolves through, because a writer and a reader that resolve
|
||
// the root separately are a hook that enforces nothing while reporting that
|
||
// it does:
|
||
// { "runId": "<run id>", "startedAt": "<ISO-8601>" }
|
||
// No marker for this session => out of scope => allow, unconditionally. An
|
||
// unrelated session must never be denied because some other run spent its
|
||
// budget; a PreToolUse hook that over-blocks breaks every session on the box.
|
||
//
|
||
// Fail-open vs fail-closed, deliberately split:
|
||
// - Out of scope (no marker, no session_id, unparsable stdin, stale marker,
|
||
// kill switch, STORM off) => exit 0. Fail OPEN.
|
||
// - In scope and over budget => exit 2. Fail CLOSED, mirroring
|
||
// research-loop-cap.mjs's own stance: a budget control that cannot count
|
||
// must not grant. (The former "CLAUDE_PLUGIN_DATA absent" deny is gone —
|
||
// the root now always resolves, so that branch could no longer fire.)
|
||
//
|
||
// Counting is read-only. The ledger is append-only and written solely by
|
||
// research-loop-cap.mjs's allowTurn(); if this hook appended, the cap would
|
||
// count its own enforcement.
|
||
//
|
||
// Kill switch: VOYAGE_DISABLE_CAP_HOOK=1 disables enforcement entirely.
|
||
|
||
import { readFileSync, existsSync, rmSync } from 'node:fs';
|
||
import { join, dirname } from 'node:path';
|
||
import { fileURLToPath } from 'node:url';
|
||
|
||
const HERE = dirname(fileURLToPath(import.meta.url));
|
||
const { resolveLedgerPath, resolveDataRoot, resolveMaxConvTurns, isStormEnabled, MAX_TOTAL_DIMENSIONS } =
|
||
await import(join(HERE, '..', '..', 'lib', 'util', 'research-loop-cap.mjs'));
|
||
|
||
const SCOPE_DIRNAME = 'trekresearch-loop-scope';
|
||
const DEFAULT_TTL_MS = 6 * 60 * 60 * 1000; // 6h — longer than any real research run
|
||
|
||
const env = process.env;
|
||
|
||
function allow() {
|
||
process.exit(0);
|
||
}
|
||
|
||
function deny(message) {
|
||
process.stderr.write(`[voyage] BLOCKED: trekresearch loop cap\n${message}\n`);
|
||
process.exit(2);
|
||
}
|
||
|
||
// 1. Kill switch.
|
||
if (env.VOYAGE_DISABLE_CAP_HOOK === '1') allow();
|
||
|
||
// 2. Default-off: no loop runs unless STORM is enabled, so nothing to enforce.
|
||
if (!isStormEnabled(env)) allow();
|
||
|
||
// 3. Parse stdin. Unparsable input is not evidence of a loop turn.
|
||
let input;
|
||
try {
|
||
input = JSON.parse(readFileSync(0, 'utf-8'));
|
||
} catch {
|
||
allow();
|
||
}
|
||
|
||
const sessionId = input?.session_id;
|
||
if (!sessionId || typeof sessionId !== 'string') allow();
|
||
|
||
// 4. Resolve the scope marker through the writer's own root resolution.
|
||
// VOYAGE_CAP_SCOPE_DIR stays as a test/override seam; unset, this lands on
|
||
// exactly the directory the Phase 5 snippet writes into.
|
||
const scopeDir = env.VOYAGE_CAP_SCOPE_DIR || resolveDataRoot(env);
|
||
|
||
const markerPath = join(scopeDir, SCOPE_DIRNAME, `${sessionId}.json`);
|
||
if (!existsSync(markerPath)) allow();
|
||
|
||
let marker;
|
||
try {
|
||
marker = JSON.parse(readFileSync(markerPath, 'utf-8'));
|
||
} catch {
|
||
allow(); // A marker we cannot read cannot tell us which run we are in.
|
||
}
|
||
|
||
if (!marker?.runId) allow();
|
||
|
||
// 5. TTL / auto-reset. A marker left behind by a crashed run must not deny
|
||
// tool calls for the rest of the machine's life.
|
||
const ttlRaw = Number(env.VOYAGE_CAP_SCOPE_TTL_MS);
|
||
const ttlMs = Number.isFinite(ttlRaw) && ttlRaw > 0 ? ttlRaw : DEFAULT_TTL_MS;
|
||
const startedAt = Date.parse(marker.startedAt ?? '');
|
||
if (!Number.isFinite(startedAt) || Date.now() - startedAt > ttlMs) {
|
||
try { rmSync(markerPath, { force: true }); } catch { /* best effort */ }
|
||
allow();
|
||
}
|
||
|
||
// --- In scope from here on. ---
|
||
|
||
// 6. The ledger is the only source of truth for turns spent.
|
||
const ledgerPath = resolveLedgerPath(env);
|
||
|
||
function countTurns(path, runId) {
|
||
if (!existsSync(path)) return 0;
|
||
let text;
|
||
try { text = readFileSync(path, 'utf-8'); } catch { return 0; }
|
||
let count = 0;
|
||
for (const line of text.split('\n')) {
|
||
if (!line) continue;
|
||
try {
|
||
if (JSON.parse(line).runId === runId) count++;
|
||
} catch { /* skip malformed lines */ }
|
||
}
|
||
return count;
|
||
}
|
||
|
||
// 7. Same bound the primitive uses: turns-per-dimension × the whole dimension
|
||
// list under settings.json:16's maxDimensions ceiling.
|
||
const budget = resolveMaxConvTurns(env) * MAX_TOTAL_DIMENSIONS;
|
||
const used = countTurns(ledgerPath, marker.runId);
|
||
|
||
if (used >= budget) {
|
||
deny(
|
||
` Run ${marker.runId} has spent ${used}/${budget} loop turns.\n` +
|
||
` Tool: ${input?.tool_name ?? 'unknown'}${input?.agent_type ? ` (agent: ${input.agent_type})` : ''}\n` +
|
||
` Remaining gaps belong in the brief as open questions, not in another turn.\n` +
|
||
` Raise TREKRESEARCH_MAX_CONV_TURNS deliberately, or set VOYAGE_DISABLE_CAP_HOOK=1.`,
|
||
);
|
||
}
|
||
|
||
allow();
|