Skip to content

fix(review): reject forbidden Claude tool use #165

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
BaseInfinity merged 1 commit into from
Aug 22, 2026
Merged

fix(review): reject forbidden Claude tool use #165

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
56 changes: 47 additions & 9 deletions .codex/hooks/dual-review.cjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -243,24 +243,35 @@ function runSupervisedProcess(command, args, options = {}) {
let cancelled = false;
let timeoutReason = null;
let terminationRequested = false;
let rejectedEvent = null;
let requestTermination = () => {};
let lineBuffer = "";
child.stdout.setEncoding("utf8");
child.stderr.setEncoding("utf8");
const recordChunk = (stream, chunk) => {
lastActivityAt = Date.now();
if (logPath !== "") fs.appendFileSync(logPath, `[${stream}] ${chunk}`);
};
const recordEvents = (chunk) => {
const recordEvents = (chunk, flush = false) => {
lineBuffer += chunk;
const lines = lineBuffer.split(/\r?\n/);
lineBuffer = lines.pop() || "";
if (flush) lineBuffer = "";
else lineBuffer = lines.pop() || "";
for (const line of lines) {
if (line.trim() === "") continue;
try {
const event = JSON.parse(line);
lastEventAt = Date.now();
lastEventType = String(event.type || event.event || event.subtype || "json");
options.onEvent?.(event);
const rejection = options.rejectEvent?.(event);
if (rejectedEvent === null && rejection) {
rejectedEvent = {
type: lastEventType,
reason: String(rejection),
};
requestTermination();
}
} catch {
// Human-readable diagnostics remain in the local log and terminal result.
}
Expand All @@ -271,6 +282,7 @@ function runSupervisedProcess(command, args, options = {}) {
recordChunk("stdout", chunk);
recordEvents(chunk);
});
child.stdout.on("end", () => recordEvents("", true));
child.stderr.on("data", (chunk) => {
stderr += chunk;
recordChunk("stderr", chunk);
Expand All @@ -294,9 +306,12 @@ function runSupervisedProcess(command, args, options = {}) {
timedOut,
cancelled,
timeout_reason: timeoutReason,
terminal_state: timedOut
? REVIEW_STATES.TIMED_OUT
: cancelled
rejected_event: rejectedEvent,
terminal_state: rejectedEvent !== null
? REVIEW_STATES.INVALID_VERDICT
: timedOut
? REVIEW_STATES.TIMED_OUT
: cancelled
? REVIEW_STATES.CANCELLED
: result.error
? REVIEW_STATES.INTERNAL_RUNNER_FAILURE
Expand All @@ -311,7 +326,7 @@ function runSupervisedProcess(command, args, options = {}) {
log_path: logPath || null,
});
};
const requestTermination = () => {
requestTermination = () => {
if (terminationRequested) return;
terminationRequested = true;
terminateProcessTree(child, "SIGTERM");
Expand Down Expand Up @@ -392,6 +407,7 @@ function runtimeAttemptMetadata(attempt) {
timeout_reason: attempt.timeout_reason || null,
last_event_at: attempt.last_event_at || null,
last_event_type: attempt.last_event_type || "none",
rejected_event: attempt.rejected_event || null,
log_path: attempt.log_path || null,
retry_reason: attempt.retry_reason || null,
};
Expand Down Expand Up @@ -1127,11 +1143,23 @@ function actualClaudeModel(parsed, envelope) {
return candidates.find((candidate) => typeof candidate === "string" && candidate !== "") || "";
}

function rejectClaudeReviewerToolUse(event) {
const events = Array.isArray(event) ? event : [event];
const blocks = events.flatMap((entry) => [
entry,
entry?.content_block,
...(Array.isArray(entry?.message?.content) ? entry.message.content : []),
]).filter((block) => block && typeof block === "object");
const toolBlock = blocks.find((block) => /(?:^|_)tool_use$/.test(String(block.type || "")));
if (!toolBlock) return "";
return `forbidden reviewer tool use: ${toolBlock.name || "unknown"}`;
}

async function runClaudeReviewer(prompt, temporaryDirectory, signal, reviewer, route, fallbackReason = null, logDirectory, phase = "initial") {
const launch = claudeLaunch();
const args = [
"-p", "--model", reviewer.model, "--effort", reviewer.effort, "--safe-mode", "--max-turns", "2",
"--setting-sources", "user", "--tools", "", "--disable-slash-commands",
"--setting-sources", "user", "--tools", "", "--disallowedTools", "advisor", "--disable-slash-commands",
"--no-session-persistence", "--mcp-config", '{"mcpServers":{}}', "--strict-mcp-config",
"--json-schema", JSON.stringify(REVIEW_SCHEMA), "--output-format", "stream-json", "--verbose",
];
Expand All @@ -1146,11 +1174,19 @@ async function runClaudeReviewer(prompt, temporaryDirectory, signal, reviewer, r
heartbeatInterval: configuredDuration("CODEX_SDLC_REVIEW_HEARTBEAT_MS", 30 * 1000),
killGrace: configuredDuration("CODEX_SDLC_REVIEW_KILL_GRACE_MS", 2000),
structuredEventsRequired: true,
rejectEvent: rejectClaudeReviewerToolUse,
signal,
label: `${reviewer.label} ${phase}`,
logPath: path.join(logDirectory, `cross-model-${phase}-attempt-${attempt}.log`),
windowsVerbatimArguments: prepared.windowsVerbatimArguments,
});
if (result.rejected_event) {
return {
...result,
retryable: false,
terminal_state: REVIEW_STATES.INVALID_VERDICT,
};
}
const availabilityReason = claudeAvailabilityReason(result);
if (availabilityReason !== "") {
return {
Expand Down Expand Up @@ -1208,7 +1244,7 @@ async function confirmClaudeReviewerUnavailable(temporaryDirectory, signal, revi
const launch = claudeLaunch();
const args = [
"-p", "--model", reviewer.model, "--effort", reviewer.effort, "--safe-mode", "--max-turns", "1",
"--setting-sources", "user", "--tools", "", "--disable-slash-commands",
"--setting-sources", "user", "--tools", "", "--disallowedTools", "advisor", "--disable-slash-commands",
"--no-session-persistence", "--mcp-config", '{"mcpServers":{}}', "--strict-mcp-config",
"--output-format", "stream-json", "--verbose",
];
Expand All @@ -1222,12 +1258,13 @@ async function confirmClaudeReviewerUnavailable(temporaryDirectory, signal, revi
heartbeatInterval: configuredDuration("CODEX_SDLC_REVIEW_HEARTBEAT_MS", 30 * 1000),
killGrace: configuredDuration("CODEX_SDLC_REVIEW_KILL_GRACE_MS", 2000),
structuredEventsRequired: true,
rejectEvent: rejectClaudeReviewerToolUse,
signal,
label: `${reviewer.label} availability probe`,
logPath: path.join(logDirectory, "cross-model-availability-probe.log"),
windowsVerbatimArguments: prepared.windowsVerbatimArguments,
});
if (result.error || result.timedOut || result.cancelled) return false;
if (result.rejected_event || result.error || result.timedOut || result.cancelled) return false;
return claudeAvailabilityReason(result) === suspectedReason;
}

Expand Down Expand Up @@ -1471,6 +1508,7 @@ module.exports = {
REVIEW_STATES,
buildWindowsCommandLine,
classifyReviewFailure,
rejectClaudeReviewerToolUse,
runSupervisedProcess,
runWithInfrastructureRetry,
};
Expand Down
50 changes: 49 additions & 1 deletion tests/test-adapter.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5394,6 +5394,13 @@ if (args[0] === "auth" && args[1] === "status") {
const promptBuffer = fs.readFileSync(0);
const prompt = promptBuffer.toString("utf8");
const mode = process.env.DUAL_TEST_MODE || "clean";
if (mode === "forbidden-tool-then-clean") {
const forbidden = { type: "assistant", message: { content: [{ type: "server_tool_use", name: "advisor" }] } };
const clean = { type: "result", model: "fable", structured_output: { findings: [], verdict: "CERTIFIED", confidence: 99 } };
fs.appendFileSync(process.env.FABLE_MARKER, `${JSON.stringify({ args, prompt, promptBase64: promptBuffer.toString("base64") })}\n`);
process.stdout.write(JSON.stringify([forbidden, clean]));
return;
}
if (mode === "sol-fail-fable-hang") {
process.on("SIGTERM", () => {
fs.writeFileSync(process.env.CANCEL_MARKER, "cancelled\n");
Expand Down Expand Up @@ -5444,6 +5451,8 @@ if (!solCalls[0].args.includes("gpt-5.6-sol") || !solCalls[0].args.some((arg) =>
if (!solCalls[0].args.includes("--json")) process.exit(1);
if (!fableCalls[0].args.includes("fable") || !fableCalls[0].args.includes("high")) process.exit(1);
if (!fableCalls[0].args.includes("stream-json") || !fableCalls[0].args.includes("--verbose")) process.exit(1);
const denyIndex = fableCalls[0].args.indexOf("--disallowedTools");
if (denyIndex === -1 || fableCalls[0].args[denyIndex + 1] !== "advisor") process.exit(1);
const expectedPatch = require("node:child_process").spawnSync(
"git", ["-C", process.env.REPO_PATH, "diff", "--cached", "--binary", "HEAD"],
).stdout;
Expand All @@ -5453,6 +5462,19 @@ const fablePrompt = Buffer.from(fableCalls[0].promptBase64, "base64");
if (!fablePrompt.includes(expectedPatch)) process.exit(1);
NODE

: > "$fable_marker"
set +e
output=$(cd "$ws" && CODEX_SDLC_TEST_MODE=1 CODEX_SDLC_CODEX_PATH="$fake_codex" \
CODEX_SDLC_CLAUDE_PATH="$fake_claude" SOL_MARKER="$sol_marker" \
FABLE_MARKER="$fable_marker" DUAL_TEST_MODE=forbidden-tool-then-clean \
node .codex/hooks/dual-review.cjs --base HEAD --consent-subscription-quota 2>&1)
status=$?
set -e
[ "$status" -eq 2 ] || valid=false
[ ! -f "$receipt" ] || valid=false
[ "$(wc -l < "$fable_marker" | tr -d ' ')" -eq 1 ] || valid=false
[[ "$output" == *"review INVALID_VERDICT"* ]] || valid=false

: > "$sol_marker"
: > "$fable_marker"
set +e
Expand Down Expand Up @@ -5626,6 +5648,10 @@ if (model === "fable" && ["fallback", "both-unavailable"].includes(mode)) {
process.stdout.write("You're out of usage credits. Run /usage-credits to keep using Fable 5.\n");
process.exit(1);
}
if (model === "fable" && mode === "probe-tool-use" && !probe) {
process.stdout.write("You're out of usage credits. Run /usage-credits to keep using Fable 5.\n");
process.exit(1);
}
if (model === "fable" && mode === "zero-exit-quota") {
process.stdout.write(`${JSON.stringify({ type: "rate_limit_event", rate_limit_info: { status: "rejected" } })}\n`);
process.stdout.write(`${JSON.stringify({
Expand All @@ -5652,6 +5678,13 @@ if (model === "fable" && mode === "zero-exit-malformed" && !probe) {
}));
process.exit(0);
}
if (probe && mode === "probe-tool-use") {
process.stdout.write(JSON.stringify([
{ type: "assistant", message: { content: [{ type: "server_tool_use", name: "advisor" }] } },
{ type: "rate_limit_event", rate_limit_info: { status: "rejected" } },
]));
process.exit(0);
}
if (probe) {
process.stdout.write("available\n");
process.exit(0);
Expand Down Expand Up @@ -5759,9 +5792,24 @@ if (calls.length !== 3 || !calls[1].probe) process.exit(1);
if (calls[0].model !== "fable" || calls[1].model !== "fable" || calls[2].model !== "claude-opus-4-8") process.exit(1);
if (receipt.reviewers.cross_model.route !== "fallback") process.exit(1);
if (receipt.reviewers.cross_model.fallback_reason !== "quota_exhausted") process.exit(1);
if (receipt.reviewers.cross_model.model !== "claude-opus-4-8-20260801") process.exit(1);
if (receipt.reviewers.cross_model.model !== "claude-opus-4-8-20260801") process.exit(1);
NODE

rm -f "$receipt"
: > "$calls"
set +e
output=$(cd "$ws" && CODEX_SDLC_TEST_MODE=1 CODEX_SDLC_CODEX_PATH="$fake_codex" \
CODEX_SDLC_CLAUDE_PATH="$fake_claude" CLAUDE_CALLS="$calls" \
FALLBACK_TEST_MODE=probe-tool-use node .codex/hooks/dual-review.cjs --base HEAD \
--consent-subscription-quota 2>&1)
status=$?
set -e
[ "$status" -eq 2 ] || valid=false
[ "$(wc -l < "$calls" | tr -d ' ')" -eq 2 ] || valid=false
[ "$(grep -c '"model":"fable"' "$calls")" -eq 2 ] || valid=false
grep -q 'claude-opus-4-8' "$calls" && valid=false
[ ! -f "$receipt" ] || valid=false

rm -f "$receipt"
: > "$calls"
set +e
Expand Down
99 changes: 99 additions & 0 deletions tests/test-review-supervision.cjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@ const path = require("node:path");
const {
REVIEW_STATES,
classifyReviewFailure,
rejectClaudeReviewerToolUse,
runSupervisedProcess,
runWithInfrastructureRetry,
} = require("../.codex/hooks/dual-review.cjs");
Expand Down Expand Up @@ -98,6 +99,99 @@ async function testUnstructuredChatterDoesNotFakeProgress() {
assert.equal(result.terminal_state, REVIEW_STATES.TIMED_OUT);
}

function realClaudeAdvisorEvent() {
return {
type: "assistant",
message: {
content: [
{ type: "thinking", thinking: "" },
{ type: "server_tool_use", name: "advisor", input: { task: "review" } },
],
},
};
}

function testClaudeToolUseDetectorScansRealAssistantEnvelope() {
assert.equal(
rejectClaudeReviewerToolUse(realClaudeAdvisorEvent()),
"forbidden reviewer tool use: advisor",
);
}

function testClaudeToolUseDetectorScansArrayEnvelope() {
assert.equal(
rejectClaudeReviewerToolUse([
realClaudeAdvisorEvent(),
{ type: "result", structured_output: { findings: [], verdict: "CERTIFIED", confidence: 99 } },
]),
"forbidden reviewer tool use: advisor",
);
}

async function testRejectedToolUseCannotBecomeCleanResult() {
const source = [
"const forbidden={type:'assistant',message:{content:[{type:'server_tool_use',name:'advisor'}]}};",
"const clean={type:'result',model:'fable',structured_output:{findings:[],verdict:'CERTIFIED',confidence:99}};",
"process.on('SIGTERM',()=>process.exit(0));",
"process.stdout.write(JSON.stringify(forbidden)+'\\n'+JSON.stringify(clean)+'\\n');",
"setInterval(()=>{},1000);",
].join("");
const result = await runSupervisedProcess(process.execPath, ["-e", source], {
heartbeatInterval: 20,
stallTimeout: 500,
timeout: 800,
killGrace: 30,
structuredEventsRequired: true,
rejectEvent: rejectClaudeReviewerToolUse,
});

assert.equal(result.timedOut, false);
assert.equal(result.terminal_state, REVIEW_STATES.INVALID_VERDICT);
assert.equal(result.rejected_event.reason, "forbidden reviewer tool use: advisor");
assert.match(result.stdout, /\"verdict\":\"CERTIFIED\"/);
}

async function testRejectedFinalBufferCannotBecomeCleanResult() {
const source = [
"const forbidden={type:'assistant',message:{content:[{type:'server_tool_use',name:'advisor'}]}};",
"const clean={type:'result',model:'fable',structured_output:{findings:[],verdict:'CERTIFIED',confidence:99}};",
"process.stdout.write(JSON.stringify([forbidden,clean]));",
].join("");
const result = await runSupervisedProcess(process.execPath, ["-e", source], {
heartbeatInterval: 20,
stallTimeout: 500,
timeout: 800,
killGrace: 30,
structuredEventsRequired: true,
rejectEvent: rejectClaudeReviewerToolUse,
});

assert.equal(result.timedOut, false);
assert.equal(result.terminal_state, REVIEW_STATES.INVALID_VERDICT);
assert.equal(result.rejected_event.reason, "forbidden reviewer tool use: advisor");
}

async function testRejectedToolUseOutranksTimeout() {
const source = [
"process.on('SIGTERM',()=>{});",
"process.stdout.write(JSON.stringify({type:'assistant',message:{content:[{type:'server_tool_use',name:'advisor'}]}})+'\\n');",
"setInterval(()=>{},1000);",
].join("");
const result = await runSupervisedProcess(process.execPath, ["-e", source], {
heartbeatInterval: 25,
stallTimeout: 2000,
timeout: 500,
killGrace: 650,
structuredEventsRequired: true,
rejectEvent: rejectClaudeReviewerToolUse,
});

assert.equal(result.timedOut, true);
assert.equal(result.timeout_reason, "wall");
assert.equal(result.terminal_state, REVIEW_STATES.INVALID_VERDICT);
assert.equal(result.rejected_event.reason, "forbidden reviewer tool use: advisor");
}

function testAuthAndInvalidVerdictClassification() {
assert.equal(classifyReviewFailure({
status: 1,
Expand Down Expand Up @@ -139,6 +233,11 @@ async function testOneInfrastructureRetryOnly() {
await testHardTimeoutReapsDescendantTree();
await testCancellationStopsReviewerAndRetainsPrivateLog();
await testUnstructuredChatterDoesNotFakeProgress();
testClaudeToolUseDetectorScansRealAssistantEnvelope();
testClaudeToolUseDetectorScansArrayEnvelope();
await testRejectedToolUseCannotBecomeCleanResult();
await testRejectedFinalBufferCannotBecomeCleanResult();
await testRejectedToolUseOutranksTimeout();
testAuthAndInvalidVerdictClassification();
await testOneInfrastructureRetryOnly();
process.stdout.write("review supervision tests passed\n");
Expand Down