fix(worker): retry the GitLab repository catalog step on transient timeouts - #302

Problem

A single upstream GitLab hiccup fails a whole run. Observed failure:

Task UP-4859 failed: The block could not be completed.
(Step "step//./src/workflows/agent//listFreshRepositoryCatalogStep" failed after 0 retries:
 GitLab projects list timed out after 15000ms)

listFreshRepositoryCatalogStep (agent.ts) has maxRetries = 0 at the WDK level, so once its internal listing gives up, the run dies and a human has to re-move the ticket back into the AI column. The internal listing already retried (2 attempts, fixed 500ms backoff, 15s per-page timeout), but two consecutive 15s timeouts still exhausted it.

Change

All GitLab/GitHub listing goes through listRepositoriesWithRetry in apps/worker/src/adapters/vcs/repository-directory.ts. Hardened there, surgically:

• Attempts: 2 -> 3 (LISTING_MAX_ATTEMPTS) - one more chance to ride out a transient blip. Still bounded, no infinite retry.
• Fixed 500ms backoff -> jittered exponential backoff (listingRetryDelayMs): full jitter over [0, 500ms] then [0, 1000ms], capped at 4s, so correlated upstream blips don't all retry in lockstep.
• Per-page timeout: 15s -> 18s (GITLAB_PROJECTS_TIMEOUT_MS) - the failure showed GitLab responding slower than 15s, so a modest raise survives genuine slowness.

Worst case is 3 * 18s + <=1.5s backoff ~= 55.5s, which stays inside the ~60s pre-sandbox listing budget the surrounding comments reason about (also honored by the dashboard catalog endpoint that shares this path). The listing is already paginated (per_page=100, follows x-next-page), so unbounded fetching was not the cause.

Permanent failures still surface unchanged: after the attempts are exhausted the original error (with its clear "GitLab projects list timed out after 18000ms" / status reason) is rethrown, not swallowed.

Tests

apps/worker/src/adapters/vcs/repository-directory.test.ts:

• transient GitLab timeout twice then success on the third, bounded attempt -> resolves with the recovered listing (3 fetches)
• a timeout that never clears -> fails after exactly 3 attempts with the timeout reason (bounded, no infinite retry)

Touched-file run: 19 passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_015kfeohXE66xx7RJPxZ2pvH