Skip to content

chore: isolate COPILOT_REQUESTS_PAT from the owner's REST rate limit #365

Open

Description

@DevSecNinja

What happened

News run 44092284114 (2026-11-17 13:24Z) failed before curation:

Failed to fetch PAT user login (403): API rate limit exceeded for user ID 14926452

Copilot CLI calls GET /user with COPILOT_GITHUB_TOKEN (= COPILOT_REQUESTS_PAT) at startup. That PAT belongs to DevSecNinja. GitHub counts REST requests from all of a user's PATs and OAuth tokens (classic and fine-grained) against one 5,000/hour quota per user. The owner's fleet automation used about 5,100 requests that hour, so Copilot CLI could not authenticate and the briefing failed.

A new fine-grained PAT on the same account does not help. It draws from the same per-user quota.

Options

# Option Isolates quota? Cost / effort Owner actions
1 Move fleet automation to a GitHub App (installation tokens) Yes, and it removes the cause. Installation tokens have their own 5,000–12,500/h per installation, separate from the user. Medium: change the fleet's auth. No licence cost. Create or reuse a GitHub App, install it on the fleet's repos/org, store APP_ID + private key as fleet secrets, and switch fleet scripts to actions/create-github-app-token (or the equivalent).
2 Dedicated machine user for Copilot requests Yes. The machine user has its own 5,000/h. Needs a Copilot seat (Pro, or Business via org) for the machine user. Premium requests bill to that seat. GitHub ToS allows one free machine account per person. Create the account, give it Copilot, create a fine-grained PAT with Copilot Requests permission only (no repo access needed), and replace the COPILOT_REQUESTS_PAT secret.
3 Rate-limit pre-flight + wait in Wazzup No, but it makes collisions survivable. Small code change. Can lengthen a run by up to the wait cap. None (code-only PR).
4 Reduce the fleet's REST usage (ETag/conditional requests, which don't count when they return 304; GraphQL batching; caching) Partly. Depends on the fleet code. Review the fleet's hottest endpoints.

Option 3 in more detail (Wazzup-side)

Before starting Copilot CLI, call GET /rate_limit with the PAT. This call does not count against the quota. If core.remaining is below a small floor (for example 50):

  • If core.reset is within about 10 minutes, sleep until the reset, then continue. The job timeout is 25 minutes.
  • Otherwise, end the run as skipped (not failed) with a clear summary line, so the freshness gate retries on the next 20-minute tick after the quota resets.

PR #364 already adds a 55-minute backoff after a failed run and a freshness gate that retries automatically. Today's collision now costs about one cadence slot instead of a stale site.

Recommendation

  1. Now: nothing blocking. fix(news): gate scheduled generation on persisted-state freshness #364 plus fix(state): retry news-state persist on transient release lookup errors #362 make a rate-limited hour self-heal on the next tick.
  2. Root cause: option 1, moving the fleet to a GitHub App. It frees the owner's quota for the Copilot PAT and for interactive gh use, with no licence cost. If the fleet can't move soon, use option 2 for Wazzup alone.
  3. Optional hardening: option 3, if collisions keep happening after 1 or 2.

No tokens or secrets were created.

Activity

  1. added
    area/backendPython ingestion, scoring, AI provider, and publisher code
    area/frontendStatic PWA assets, styling, service worker, and manifest
    and removed on Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area/backendPython ingestion, scoring, AI provider, and publisher codearea/frontendStatic PWA assets, styling, service worker, and manifestarea/githubGitHub Actions, Renovate, labels, and repository automation

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions