fix: Cap fix retries and classify failures instead of retrying forever

[Senna46]

Problem

A bug whose fix could not be pushed was re-marked FAILED every cycle and retried indefinitely. Each retry re-ran a full claude -p fix generation (2–6 minutes) and threw the result away at the push step.

This is not hypothetical — on 2026-07-22 a PR touching .github/workflows/ failed to push because the PAT lacked the workflow scope, and Fixooly burned six full generations at ~2.5 minute intervals before the token was fixed by hand. Nothing in the code would have stopped it. With effortLevel: max each wasted generation now costs more.

A secondary effect: computeSinceForRepo dropped the since filter entirely whenever a repo had any FAILED bug, paginating every review comment in the repo each cycle (measured: 6 pages/cycle on cauchye/ai-talent-mapping, which has 572 review comments) for as long as the failure persisted.

Changes

Failure classification (src/failureClassifier.ts) — three classes, each with a different retry policy:

Class	Examples	Behavior
permanent	branch protection / GH006, PAT missing workflow scope, permission denied	give up immediately
transient	Claude usage limit, expired or invalid auth, network errors	retry, without spending an attempt
retryable	generation errored or timed out	retry, up to MAX_FIX_ATTEMPTS (3)

Transient outages deliberately do not consume the retry budget — a multi-hour usage limit must not permanently condemn a perfectly fixable bug.

Backoff — retries wait 5 → 15 → 30 minutes based on attempts so far, replacing the every-cycle hammering. Checked locally in bugbotMonitor before any API call, so a backed-off bug costs nothing.

Give-up reporting — when a bug is abandoned it is recorded as FAILED_PERMANENT (terminal) and a comment is posted on the PR with the reason and the error detail, deduplicated by a per-bug-set marker. The error text goes through sanitizeGitError so tokens cannot leak into a public comment.

Bounded since bypass — the widened window is now 30 days instead of unlimited.

Schema — adds processed_bugs.attempts with an in-place ALTER TABLE migration for existing databases, and switches writes from INSERT OR REPLACE to UPSERT so the counter is not silently reset on every status update.

Verification

npm run typecheck and npm run build pass. The state machine and classifier were exercised against a throwaway SQLite database — 25 checks, all passing:

• transient failures ×5 keep attempts at 0 and never go permanent
• counted failures increment 1 → 2 → 3 and stop exactly at the cap
• FAILED_PERMANENT is terminal and clears the repo's retryable flag
• attempts survives a later success (UPSERT, not REPLACE)
• a legacy database without the column migrates in place with rows preserved
• the classifier returns the expected class for the exact error strings found in the production logs (workflow-scope rejection, GH006, You've hit your limit, 401 OAuth access token has expired, EADDRNOTAVAIL, timeout, generic exit 1)

Running on the Mac mini daemon from this branch before merge.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes core daemon failure handling and SQLite state semantics (migration + new terminal statuses); misclassification could give up too early or retry too aggressively, but scope is localized to autofix retry logic.

Overview
Stops failed bugs from triggering a full claude -p run every poll cycle by classifying errors, capping counted retries, and backing off between attempts.

Adds failureClassifier.ts with permanent (give up + PR comment), transient (retry without spending an attempt; abort the rest of the cycle), and retryable (up to 3 attempts). SQLite state gains attempts, recordFailedBugs, backoff checks in bugbotMonitor, expireStaleFailures, and FAILED_PERMANENT; writes use upsert so attempt counts are not reset. Repo comment scan widens to 30 days when retryable failures exist instead of dropping since entirely. fixGenerator embeds CLI output in errors for classification; main posts deduplicated give-up comments with sanitized error detail. README documents the retry policy.

^{Reviewed by Cursor Bugbot for commit 17b3cdd. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d6b93fa. Configure here.}