Skip to content

fix(core): loosen auto-mode classifier timeouts, disable stage-2 thinking #4680

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
pomelo-nwu merged 2 commits into from
Jun 2, 2026
Merged

fix(core): loosen auto-mode classifier timeouts, disable stage-2 thinking #4680

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4a12676
fix(core): loosen auto-mode classifier timeouts, disable stage-2 thin…
qqqys Jun 1, 2026
9c14e6a
docs(core): trim verbose comments in auto-mode classifier
qqqys Jun 2, 2026
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
5 changes: 3 additions & 2 deletions packages/core/src/permissions/classifier.test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,7 +189,7 @@ describe('classifier configuration', () => {
expect(opts.config?.thinkingConfig?.includeThoughts).toBe(false);
});

it('uses max_output_tokens=4096 with thinking enabled for stage 2', async () => {
it('uses max_output_tokens=4096 with thinking disabled for stage 2', async () => {
runSideQueryMock
.mockResolvedValueOnce({ shouldBlock: true })
.mockResolvedValueOnce({ thinking: 't', shouldBlock: false, reason: '' });
Expand All @@ -201,7 +201,8 @@ describe('classifier configuration', () => {
};
};
expect(opts.config?.maxOutputTokens).toBe(4096);
expect(opts.config?.thinkingConfig?.includeThoughts).toBe(true);
// Thinking is disabled in every stage (latency-sensitive permission gate).
expect(opts.config?.thinkingConfig?.includeThoughts).toBe(false);
});

it('does not pin a model — defaults to the fast model via sideQuery', async () => {
Expand Down
23 changes: 16 additions & 7 deletions packages/core/src/permissions/classifier.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,8 +9,9 @@
* Stage 1 (fast): shouldBlock-only output, max_tokens=32, thinking off.
* Allow path returns immediately (~300ms).
* Stage 2 (review): full output { thinking, shouldBlock, reason },
* max_tokens=4096, thinking on. Reviews stage-1 blocks
* to reduce false positives.
* max_tokens=4096, thinking off. Reviews stage-1 blocks
* to reduce false positives. (`thinking` is a plain output
* field, not an allocated reasoning budget.)
*
* Fail-closed: any non-abort failure (API error, timeout, schema failure,
* context overflow) returns shouldBlock=true with unavailable=true.
Expand All @@ -33,10 +34,14 @@ import { buildClassifierContents } from './classifier-transcript.js';
// the underlying API / timeout / context-overflow error.
const debugLogger = createDebugLogger('CLASSIFIER');

/** Stage-1 timeout: fast model p99 is ~1.5s; 3s catches stuck cases. */
export const STAGE1_TIMEOUT_MS = 3_000;
/** Stage-2 timeout: thinking takes longer; 10s caps infrastructure failure. */
export const STAGE2_TIMEOUT_MS = 10_000;
// A timeout is fail-closed (action BLOCKED as "unavailable"), so too tight a
// budget turns transient slowness into spurious blocks. The fast model's p99
// is ~1.5s but the tail is long under load, so budgets are kept generous —
// better to wait than fail closed on a healthy call.
/** Stage-1 timeout: generous headroom over the fast model's p99 (~1.5s). */
export const STAGE1_TIMEOUT_MS = 10_000;
/** Stage-2 timeout: review stage runs a larger prompt; cap infra failure. */
export const STAGE2_TIMEOUT_MS = 30_000;

/** Token usage attributed to a single classifier call. */
export interface ClassifierUsage {
Expand Down Expand Up @@ -219,7 +224,11 @@ export async function classifyAction(
config: {
temperature: 0,
maxOutputTokens: 4096,
thinkingConfig: { includeThoughts: true },
// Thinking off: this gate is latency-sensitive (the user is waiting),
// and a reasoning budget would slow the review path and worsen the
// fail-closed timeout above. The `thinking` output field still carries
// the model's reasoning.
thinkingConfig: { includeThoughts: false },
},
})) as Stage2Response;
} catch (err) {
Expand Down
Loading