Repository navigation
Fleet workhorse legs failing fleet-wide: provider.auth 401 'Missing scopes: api.responses.write' (15:40 wave 9/9 roots failed) #3216
Description
Found by the OpenChamber conductor, cycle 2026-10-31 15:32 CDT (content-verified from session messages, scratch conductor-scratch-2026-10-31-1532/msgs-new.txt).
Symptom
The entire 15:40 CDT lane wave failed: 9/9 root sessions on opencode-route/workhorse (code x6, openchamber x1, opencode x1, nix x1; created 20:40:00-01Z) ended failed. The 15:00 and 15:20 waves on the same route succeeded, so this began between 15:20 and 15:40 CDT.
Error (verbatim, all 4 sampled sessions identical)
{"type":"provider.auth","status":401,"message":"You have insufficient permissions for this operation. Missing scopes: api.responses.write. Check that you have the correct role in your organization (Reader, Writer, Owner) and project (Member, Owner), and if you're using a restricted API key, that it has the necessary scopes."}
Sampled session ids: ses_e5e6c4f2b7effeHIO4alcU5YIq (code), ses_e5e6c4f2f2bffe1i6T5JblGWKv (nix), ses_e5e6c4f312bffeYDe72XVmZ14C (openchamber), ses_e5e6c4f323dffeON6AESTsjITW (opencode). Zero thought_signature hits — this is NOT the antigravity invalid-request pattern from the 12:00-13:00 waves.
Context
- The codex / openai.alternate pool (popesdrew@gmail.com) read 5h 100% (weekly 53%) in the same cycle's quota snapshot, and the second credential (drewrypope@gmail.com) weekly 100%. The failing leg appears to be an OpenAI credential whose token/key lacks the api.responses.write scope (restricted key or Reader role).
- Until this is fixed, workhorse-route lane sessions that fall through to this credential die on their first model call; queue work stalls behind claim expiry.
Acceptance
- Identify which credential/leg of the workhorse route returns the 401 and why its scope/role changed.
- Restore a working leg (scope/role fix or route the leg to a credential with api.responses.write) — plan quotas only, never PAYG.
- Verify with a fresh lane wave: roots on opencode-route/workhorse complete model calls without provider.auth errors.
Activity
dezren39 commentedon Nov 1, 2026
Intake review
Related: #3256 (the standalone-proxy endpoint defect in front of this one — distinct; both must be fixed before a GPT leaf serves through the proxy).
Notes: The failure window is pinned by the conductor's cycle evidence: 15:00 and 15:20 CDT waves on
opencode-route/workhorsesucceeded, the 15:40 wave failed 9/9 roots with an identical provider.auth 401 (Missing scopes: api.responses.write), sampled across all four repos. The body already carries acceptance criteria, so no body change was made. Addeduser/review: identifying and re-scoping/re-roling an OpenAI credential may need an org-admin decision only Drew can make — Muse review lane first, per the taxonomy's user/ flow. Plan quotas only, never PAYG, per the standing pool rules.Risks: Until a working leg is restored, workhorse-route sessions that fall through to this credential die on their first model call and queue work stalls behind claim expiry — this is a fleet-throughput blocker, not a single-app bug.
Suggested approach: Identify the credential from the failing legs' route_attempt/session records, check its key scopes and org/project role against
api.responses.write, then restore the scope/role or re-point the leg at a credential that has it, and verify with a fresh lane wave per the body's acceptance list.— fleet monitor intake, 2026-10-31 fleet-monitor run (19:12 CDT)
dezren39 commentedon Nov 1, 2026
Root cause — fixed and verified 2026-10-31 ~19:53 CDT
The failure was a credential/method mismatch in the OpenCode credential store, not quota and not the account.
The fork ships two OpenAI sign-in paths that wire provider
openaito different backends:chatgpt-browser/chatgpt-headless(openai.ts plugin) → Codex backendhttps://chatgpt.com/backend-api/codex, token = ChatGPT plan token from the Codex client (app_EMoamEEZ73f0CkXaXp7hrann).chatgpt-token-sharing(chatgpt.ts plugin) → platform APIhttps://api.openai.com/v1, which requires a token minted by that flow for the registered agent client, carryingresource.invoke+chatgpt.tokens.use.directscopes.For the main account (drewrypope) the store held both credentials. The token-sharing credential was the active one, but the access token stored under it is a Codex-flow token: claims show client_id
app_EMoamEEZ73f0CkXaXp7hrannand scp[openid, profile, email, offline_access, api.connectors.read, api.connectors.invoke]— byte-identical (same jti/iat) to the Muse's current token in~/.codex/auth.json. It was never minted by the token-sharing flow, so when the plugin sent it toapi.openai.com/v1/responses, the platform API rejected it: 401 "Missing scopes: api.responses.write". A refresh could not repair it — refresh re-mints from the original grant, which never included the token-sharing scopes, and the stored refresh token belongs to the Codex client, not the agent client named in the credential metadata.Timing matches this issue's window: the
chatgpt-browsercredential row was last updated at 15:40 CDT (the active-path flip), right when the waves started failing; the 15:00/15:20 waves succeeded on the Codex-backend path. Quota is excluded: after the main account's quota reset (weekly at 1% on the 19:30 hub read), attempts at 19:40–19:49 CDT still failed 401.Fix
opencode auth switch openai <chatgpt-browser credential id>— reactivated the existing browser credential for the main account (same account, same tokens, Codex-backend wiring; its metadata carries the accountID used for thechatgpt-account-idheader) and deactivated the token-sharing credential. No config edits, no token changes, no re-login required. The token-sharing row remains stored but inactive. Note: the chatgpt plugin auto-migrates browser credentials to token-sharing only when their metadata carries a clientID; this credential's metadata is accountID-only, and its method was verified unchanged after the switch and plugin reload.Verification (route_attempt ledger)
openai, modelgpt-6.1-terra, routeworkhorse, 401 Authentication.openai/gpt-6.1-terra/workhorse, from the same fleet sessions that had been 401ing. Zeroopenaifailures since the switch.Alternate account
openai.alternate(popesdrew) is methodchatgpt-headless— already on the Codex-backend path, not affected by this defect; left untouched.#3256 (standalone proxy endpoint config) is distinct and still open — it gates GPT through the proxy door (e.g. OpenClaw), not the internal routing fixed here.
(Editorial note: this comment replaces an earlier rendering whose inline-code spans were stripped by a shell-quoting accident during posting; substance unchanged.)
dezren39 commentedon Nov 1, 2026
Follow-up (2026-10-31 ~20:11–20:25 CDT) — recurrence guard ordered by Drew, both parts done:
1. Stray credential deleted. The inactive
chatgpt-token-sharingcredential for provideropenaiwas removed viaopencode auth logout openai <id>:cred_10b633957001YZ1Wclg9xbOEXM, label "OpenAI 2", methodchatgpt-token-sharing, account0227321a-a7f6-4f8c-8ce8-3c1c12dda014(drewrypope main), created 2026-10-05, active=0 at deletion. (Its token was the Codex-flow token still live in the Muse's own auth store, so no secret was lost to the machine.)openaiholds exactly one credential —cred_10b5fd438001thsQzMXLQMc9oD, methodchatgpt-browser, active.openai.alternateuntouched (headless, active).openai; all 3 failures are RateLimit/429 — zero Authentication/401. The missing-scopes failure mode is gone.2. Watchdog added.
tools/openai-cred-watchdog.pyin dezren39/hatch (commiteafdee9), run read-only against the laptop credential store. Invariant: foropenai— nochatgpt-token-sharingrow may exist, and the active method must bechatgpt-browser/chatgpt-headless; foropenai.alternate— active method must be browser/headless. Prints OK + exit 0 when healthy; VIOLATION lines naming credential id/method + exit 1 otherwise. Never prints token values.