When gh pr view / gh pr checks / gh run view / gh api returns GraphQL rate-limited, 401 Bad credentials, or 503 No-server-available, this is the diagnostic-and-recovery playbook. Encodes the resolution matrix that distinguishes the three failure modes by their first-word signature and gives the working REST-endpoint recipe that bypasses GraphQL rate-limits.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
When gh pr view / gh pr checks / gh run view / gh api returns GraphQL rate-limited, 401 Bad credentials, or 503 No-server-available, this is the diagnostic-and-recovery playbook. Encodes the resolution matrix that distinguishes the three failure modes by their first-word signature and gives the working REST-endpoint recipe that bypasses GraphQL rate-limits.
version
1.4.0
author
hermes (learned during PR
changelog
["1.4.0 (2026-07-24) Added cron-defaults section (REST over GraphQL) plus gh-pr-view-json field-name footgun. Verified during the 20m inbox-triage follow-up cron — a 4-PR state sweep with gh pr view --json hit GraphQL rate-limit on the first call. Field-name footgun also surfaced: changed_files does NOT exist in gh pr view --json output; the field is changedFiles. CLI is CamelCase, REST is snake_case. New dedicated section plus a field-name footgun subsection. Cron sessions are single-shot with no interactive budget recovery, so defaulting to gh api REST preserves the operator GraphQL budget for interactive PR-debug sessions.","1.3.0 (2026-07-23) Added gh-pr-create GraphQL-rate-limited section. The existing recipe covered gh pr view / gh pr checks / gh run view / gh api, but NOT gh pr create — which also uses a GraphQL mutation (createPullRequest) and 422s when the bucket is empty. The fix is POST /repos/{owner}/{repo}/pulls via REST with --input <payload.json> (handles the maintainer_can_modify boolean type-coercion that bare -f flags get wrong). Verified 2026-07-23 on $GITHUB_REPOSITORY PR #8548 where the push was already durable on origin/fix/companion-quest-cadence-mirror-8526-clean and the only missing piece was the PR-creation metadata. New matrix row + dedicated section + the working REST recipe + anti-pattern list (don't loop on gh pr create, don't use -f flags, don't try updatePullRequest to rename). Distinct from the existing 1.2.0 gh-safe-publish-wrapper-broken bypass — that one was about the wrapper script itself failing, this one is about GraphQL budget exhaustion.","1.2.0 (2026-07-21) Added gh-safe-publish wrapper broken bypass via REST section. The ~/.hermes/scripts/gh-safe-publish wrapper has a bash parse bug at line 14 (case ... in ... ;; esac mismatched bracket — verified 2026-07-21, hit while filing issue #8501). Symptom is a Python SyntaxError: closing parenthesis does not match opening parenthesis. When the wrapper is broken, the secret-redaction gate can't run via the wrapper; bypass by POSTing directly to REST with urllib.request while still calling outbound_secret_gate.py check manually. New matrix row + dedicated section + the working urllib.request POST recipe.","1.1.0 (2026-07-20) Initial recipe set."]
triggers
["gh is rate limited","GraphQL API rate limit exceeded","gh returns 401","GitHub 503","no server is currently available","curl returns Bad credentials from gh auth token","read CI logs when gh is broken","how to inspect a failed check when pr checks stalls","Bad credentials on api.github.com"]
GitHub CLI / API — Rate-Limit, 401, and 503 Failure Playbook
When gh breaks under load, you still need to debug failed CI checks. This skill gives you the resolution matrix and the working REST recipe.
Why this skill exists
Three failures look identical on the surface — gh returns an error, no data — but each has a different fix. Burning 5-10 minutes on the wrong branch of the decision tree is the failure mode. The matrix below is verified against PR #8467 ($GITHUB_REPOSITORY, 2026-07-20) where all three fired in sequence within 90 seconds.
Symptom → fix matrix (FIRST WORD test)
First word(s) of error
Failure class
Fix
GraphQL: API rate limit already exceeded
GraphQL rate limit (5000/h, primary budget)
Switch to REST endpoints via gh api
No server is currently available to service your request (503)
Transient 503 (load shedding)
Retry with 30-60s backoff; usually clears in <2 min
Bad credentials (401) from raw curl with gh auth token
Token format mismatch
Use gh api wrapper, NOT raw Authorization: Bearer ...
Try gh api repos/<o>/<r>/actions/runs/<id>/logs instead
gh-safe-publish wrapper bash syntax error (closing parenthesis ')' does not match opening parenthesis '[')
Wrapper script has a shell parse bug (verified 2026-07-21, ~/.hermes/scripts/gh-safe-publish:14)
Bypass the wrapper and POST directly to REST https://api.github.com/repos/<owner>/<repo>/issues (or /pulls) via urllib/httpx with gh auth token. See "gh-safe-publish wrapper broken" section below.
gh pr create returns GraphQL: API rate limit already exceeded for user ID <n>
GraphQL rate limit hits the createPullRequest mutation while the push is already durable on origin/<branch>
Switch to REST POST /repos/<owner}/{<repo>}/pulls with --input <payload.json>. The push is the durable state; the PR creation is a thin REST call that does NOT consume GraphQL budget. Verified 2026-07-23 on $GITHUB_REPOSITORY PR #8548 after the existing skill's recipe for gh pr checks/gh api did not cover gh pr create.
unable to determine default branch for <repo>: GraphQL: API rate limit already exceeded for user ID <n> from gh workflow run
GraphQL rate limit (5000/h, primary bucket) hits the resolver that picks the default branch before --ref is applied
Add --ref <branch> AND/OR switch to REST /actions/workflows/<file>/dispatches (which doesn't need the default branch)
Test: look at the FIRST word(s) of stderr before reading the rest. 'GraphQL' → use REST. 'No server' → wait + retry. 'Bad credentials' → wrap in gh api. Don't guess — each fix is wrong for the other two failure classes.
The working recipe — REST via gh api
gh api (without --method) hits REST endpoints and is not subject to the GraphQL rate-limit bucket. When gh pr checks returns GraphQL: API rate limit already exceeded, the same data is available via:
# PR metadata (state, mergeable, head SHA, body)
gh api \
-H 'Accept: application/vnd.github+json' \
-H 'X-GitHub-Api-Version: 2022-11-28' \
repos/<owner>/<repo>/pulls/<pr_number>
# Check runs (replaces `gh pr checks`)
gh api \
-H 'Accept: application/vnd.github+json' \
-H 'X-GitHub-Api-Version: 2022-11-28' \
repos/<owner>/<repo>/pulls/<pr_number>/check-runs
# Job logs (the load-bearing recipe for failed-CI debug)
gh api \
-H 'Accept: application/vnd.github+json' \
-H 'X-GitHub-Api-Version: 2022-11-28' \
repos/<owner>/<repo>/actions/jobs/<job_id>/logs > /tmp/job.log 2>&1
# Workflow run jobs + conclusions (replaces `gh run view`)
gh api \
-H 'Accept: application/vnd.github+json' \
-H 'X-GitHub-Api-Version: 2022-11-28' \
repos/<owner>/<repo>/actions/runs/<run_id>/jobs \
--jq '.jobs[] | {name, conclusion, html_url}'
Saved ~6 minutes during PR #8467 debug (the Green Gate Precheck GATE-6b failure was hidden behind a GraphQL rate-limit at the moment of inspection).
Why direct curl with gh auth token returns 401
gh auth token resolves to a short-lived/refreshable token whose format (ghs_... or already-rotated) rejects static Authorization: Bearer headers in curl. The gh api wrapper handles the refresh transparently — that's why gh api keeps working even when bare curl with the same token value gets 401.
Symptom: you read gh auth token and get something like ghp_tP...rzoJ*** (note the trailing *** — that's the masked portion). When you pass that string as Authorization: Bearer $TOKEN in curl, GitHub returns 401.
Fix: don't use raw curl. Use gh api. The wrapper refreshes the token on each call.
Don't try these (they waste time)
curl https://api.github.com/.../actions/runs/<id>/logs — endpoint returns 302 to Azure blob storage; urllib / curl-without-follow-redirect fails on the redirect.
httpx.post("https://api.github.com/...", headers={"Authorization": f"Bearer {gh_token}"}) — same 401.
Reading ~/.config/gh/hosts.yml for the token and exporting it — same refresh token problem.
Importing Python's urllib and following redirects manually — fragile, error-prone, and unnecessary because gh api already handles it.
gh workflow run <wf>.yml -f pr_number=N -f test_mode=real when GraphQL is rate-limited, even with --ref <branch> — the CLI hits GraphQL to resolve the default branch before applying --ref, so the GraphQL bucket still bites. Use the REST /actions/workflows/<wf>/dispatches endpoint instead (no default-branch lookup, no GraphQL). Verified 2026-07-20 on PR #8462 / mcp-smoke-tests.yml dispatch.
To check remaining GraphQL budget without consuming it (which would worsen the rate-limit), use the REST /rate_limit endpoint:
gh api rate_limit --jq '.resources.graphql'# Returns: {"limit":5000,"used":5058,"remaining":0,"reset":1784510958}
This is REST-not-GraphQL so it stays available even when GraphQL budget is exhausted.
Anti-patterns to avoid
Don't import httpx / requests to write a "fixed" GitHub client. Use gh api. The CLI exists for this.
Don't pipe gh auth token to curl -H "Authorization: Bearer $(cat)" — passes the masked value with *** suffix → guaranteed 401.
Don't retry the GraphQL endpoint when it's rate-limited. Each retry burns the rolling-window budget further. Switch to REST.
Don't wait for the 5000/h budget reset before doing useful work. A 30-60s exponential backoff on REST clears most 503s.
Don't try --paginate or --limit 1000 on gh pr checks during a rate-limit window — that's a GraphQL endpoint, you'll exhaust the budget for nothing.
Verification
After fixing a failed check, run the working REST recipe to confirm:
# 1. PR is still open and mergeable
gh api repos/<owner>/<repo>/pulls/<pr_number> \
--jq '{state: .state, mergeable: .mergeable, head_sha: .head.sha}'# 2. Latest check-run conclusions
gh api repos/<owner>/<repo>/commits/<head_sha>/check-runs \
--jq '.check_runs[] | {name, conclusion, html_url}'# 3. (If failure persists) re-fetch the failed job's logs via the REST recipe.
If all three return data without GraphQL: API rate limit or 503 errors, your workflow is unstuck.
Reference files
references/pr-8467-recovery-transcript.md — the verified transcript showing all 5 failure modes firing in sequence within 90 seconds during PR #8467 green-gate debug (2026-07-20), with the working gh api recipe that recovered in ~30 seconds.
The REST-vs-GraphQL choice is the canonical workaround in SOUL.md gh-actions-stuck-self-hosted-runner-recovery skill (different failure class — runner queue stuck, not API rate-limit).
The 8-section PR-body template (the OTHER class of failure that hides behind gate validators) lives at wa-green-gate-pr-shape/templates/pr-body-8-section.md.
gh-safe-publish wrapper broken — bypass via REST (verified 2026-07-21)
The ~/.hermes/scripts/gh-safe-publish wrapper around gh issue create / gh pr create is meant to scan the body file with lib/outbound_secret_gate.py before publishing. But the wrapper itself can be broken (e.g. line-14 bash parse error: closing parenthesis ')' does not match opening parenthesis '['). When that fires, the wrapper is unusable for ALL publishing, not just for one call.
Symptom:
$ gh-safe-publish issue create --repo X --body-file Y
File "$HOME/.hermes/scripts/gh-safe-publish", line 14, in <module>
"issue create"|"issue comment"|"pr create"|"pr comment"|"gist create") ;;
^
SyntaxError: closing parenthesis ')' does not match opening parenthesis '[' on line 8
Bypass recipe — POST directly to the REST API while still running the gate first:
import json, subprocess, urllib.request, urllib.error
# 1. Run the gate manually (same check the wrapper would have run)
subprocess.check_call(["python3", "$HOME/.hermes/lib/outbound_secret_gate.py",
"check", "--file", body_file_path])
# 2. Read the token (NOTE: per the 401 pitfall above, this approach to token-fetching# works for `gh api`-style flows. For bare REST POST you may need a different mechanism# — see "Token-format pitfall" below.)
tok = subprocess.check_output(["gh", "auth", "token"], text=True).strip()
# 3. POST to GitHub REST API
payload = {"title": "...", "body": open(body_file_path).read(), "labels": [...]}
req = urllib.request.Request(
f"https://api.github.com/repos/{owner}/{repo}/issues",
method="POST",
headers={
"Authorization": f"Bearer {tok}",
"Accept": "application/vnd.github+json",
"Content-Type": "application/json",
"X-GitHub-Api-Version": "2022-11-28",
},
data=json.dumps(payload).encode(),
)
with urllib.request.urlopen(req, timeout=30) as resp:
print(json.load(resp)["html_url"])
Token-format pitfall (this is the second failure mode the bare-curl approach can hit).gh auth token may return a masked token ending in *** (or a value that 401s when used raw in Authorization: Bearer). If your POST returns 401, switch to gh api:
# Let gh write the JSON body and POST it (uses gh's token-refresh wrapper)cat > /tmp/issue.json <<EOF
{"title": "...", "body": "...", "labels": ["..."]}
EOF
gh api -X POST repos/<owner>/<repo>/issues --input /tmp/issue.json
Important caveat about the gate. The gate still runs at step 1 — outbound_secret_gate.py check works even when the wrapper is broken. Never skip the gate to "save time" — that's how PATs end up in published issues (see the outbound-secret-redaction-gate security skill for the verified incident).
Fix the wrapper separately. The bash parse error is a 30-second fix once you spot it (case "..." in ... ;; esac mismatched bracket) but don't block on fixing the wrapper — bypass it via REST and file a follow-up to patch the script.
gh pr create GraphQL rate-limited — switch to REST POST /repos/{owner}/{repo}/pulls (verified 2026-07-23, PR #8548)
The skill's existing recipe covered gh pr view, gh pr checks, gh run view, and gh api. The gap that bit during PR #8548: gh pr create ALSO uses GraphQL under the hood (the createPullRequest mutation), and when the GraphQL bucket is exhausted, gh pr create returns GraphQL: API rate limit already exceeded for user ID <n>. The branch is already pushed (the push is durable state per SOUL.md push-pr-donot-stop-halfway), so the only thing missing is the PR-creation thin call.
Don't waste cycles retrying gh pr create. Each retry burns GraphQL budget. The fix is to bypass the GraphQL mutation entirely via REST.
Working recipe (verified 2026-07-23, $GITHUB_REPOSITORY PR #8548):
# 1. The push is already durable (DONE in the prior turn).
git -C <repo> rev-parse origin/<branch>
# Verify the SHA matches what you intended to push.# 2. Confirm the PR does NOT already exist via REST (NOT GraphQL — would burn budget):
gh api repos/<owner>/<repo>/pulls?head=<owner>:<branch>&state=all \
-q '.[].number' | grep -v '^$' || echo"PR does not exist yet"# 3. Build the PR payload as JSON (avoid `gh api -f` here — `maintainer_can_modify`# type-coerces badly with -f flags; --input is the clean path):cat > /tmp/pr-payload.json <<EOF
{
"title": "<pr title>",
"head": "<branch>",
"base": "main",
"body": "<full PR body — multiline JSON string>",
"maintainer_can_modify": true,
"draft": false
}
EOF# 4. POST to REST (no GraphQL consumption):
gh api -X POST \
-H "Accept: application/vnd.github+json" \
-H "Content-Type: application/json" \
repos/<owner>/<repo>/pulls \
--input /tmp/pr-payload.json
# 5. Verify the PR was created (still REST, no GraphQL):
gh api repos/<owner>/<repo>/pulls \
-q '.[] | select(.head.ref=="<branch>") | {number, html_url, state}' \
| head -5
Why this works:
POST /repos/{owner}/{repo}/pulls is a REST endpoint, NOT a GraphQL mutation. It draws from the REST rate-limit bucket (separate budget; not affected by createPullRequest GraphQL exhaustion).
The push is durable on origin/<branch> BEFORE you attempt PR creation. The PR-creation call is just metadata — nothing about the underlying work is at risk if the PR call fails.
The recipe handles the type-coercion pitfall with gh api -f: the maintainer_can_modify field must be a JSON boolean (not the string "true"). Using --input <payload.json> with explicit JSON gets the type right; -f flags string-coerce everything.
Anti-patterns:
❌ Loop on gh pr create with sleep 30 && gh pr create ... waiting for the GraphQL bucket to refresh. Burns budget and can take 8+ hours to reset (the reset window is the rolling 5000/h bucket).
❌ Try gh pr create --head <branch> with no other flags — same GraphQL path, same failure.
❌ Open the PR via the GitHub web UI from the terminal session — the agent doesn't have a browser here.
❌ Pre-create the PR via a placeholder title ("WIP: PR <N>") and rename later — that ALSO consumes GraphQL budget on updatePullRequest.
gh pr view --json field-name footgun. The field names returned by gh pr view --json are CamelCase and don't match REST field names one-to-one. The field that bites is changedFiles (NOT changed_files). Easy way to get the exact list without guessing:
gh pr view --help | grep -A1 '\-\-json' | head -10
# OR just call gh pr view with one bogus field to get the available-fields error:
gh pr view 123 --json bogus --repo foo/bar
# Error: Unknown JSON field: "bogus"# Available fields:# additions# assignees# author# ...# changedFiles# closed# ...
Field name → REST mapping that bit me (2026-07-24 cron): changedFiles (CLI) ↔ changed_files (REST), mergedAt (CLI) ↔ merged_at (REST). The CLI uses CamelCase, the REST API uses snake_case. When in doubt, default to gh api repos/<owner>/<repo>/pulls/<N> and --jq your way to the shape you need — the REST field names are stable and self-documenting.
Cron jobs / scheduled follow-ups: prefer REST gh api over GraphQL gh pr view by default (verified 2026-07-24)
A scheduled 20m inbox-triage follow-up fired in this session and hit GraphQL: API rate limit already exceeded for user ID 13840161 immediately on the first gh pr view --json number,title,state,mergedAt,updatedAt,headRefName,author,additions,changed_files ... call across 4 PRs. Recovery was a one-line swap to REST (gh api repos/<owner>/<repo>/pulls/<n> --jq "{number, title, state, merged, merged_at, updated_at, head: .head.ref, base: .base.ref, additions, changed_files}"), but the operator's GraphQL budget was already spent.
The lesson: cron jobs and other low-stakes scheduled state-checks should default to gh api REST from the start, not gh pr view GraphQL. Rationale:
Cron sessions have no interactive context to recover from a rate-limit. When an interactive session hits the limit, the operator can wait 30-60s and pick a different approach. A cron session is a single-shot — once the GraphQL bucket is empty, every subsequent gh pr view/gh pr list/gh pr checks in the same job will fail.
Cron state-checks are routine and well-served by REST. The fields you actually need for "did PR X change state?" — state, merged, merged_at, updated_at, head.sha, additions, changed_files — are all available via REST with stable field names. The CLI's CamelCase reshuffling adds friction, not value.
Preserve the operator's GraphQL budget for interactive sessions. A heavy 20m-tick cron that polls 4-8 PRs with gh pr view every tick burns 5000/h of budget that the operator's interactive PR-debug sessions also need. REST /repos/.../pulls/<n> does NOT consume the GraphQL bucket.
Working cron-friendly recipe (verified 2026-07-24):
# Quick "did it change?" state-check for one PR — REST, no GraphQL
gh api repos/<owner>/<repo>/pulls/<N> \
--jq '{state: .state, merged: .merged, merged_at: .merged_at, updated_at: .updated_at, head_sha: .head.sha, additions: .additions, changed_files: .changed_files}'# Multi-PR sweep (loop, not parallel GraphQL):forprin 8551 8559 8561 8564; doecho"=== PR #$pr ==="
gh api "repos/$GITHUB_REPOSITORY/pulls/$pr" \
--jq '{number, title, state, merged, merged_at, updated_at, head: .head.ref, base: .base.ref, additions, changed_files}'done
When GraphQL is still appropriate in a cron: never. If you find yourself reaching for gh pr list --search or gh pr view --json from a cron, stop and use gh api with explicit --jq filters instead. The CLI conveniences don't pay off when the budget is the binding constraint.
Edge case — the GraphQL bucket is truly exhausted for hours: this happens when an earlier session burned the bucket via heavy gh pr list --search polling. The reset window is the rolling 5000/h bucket; check via REST:
gh api rate_limit --jq '.resources.graphql'# Returns: {"limit":5000,"used":5058,"remaining":0,"reset":1784860131}# If `remaining: 0` AND the `reset` epoch is hours away, REST creation is the only option.
Edge case — gh api itself 401s: the existing skill already covers this (gh auth token returns a masked/rotated value that 401s on bare Authorization: Bearer). gh api wraps the token refresh, so it stays working. If gh api 401s, re-auth with gh auth login --scopes "repo,workflow,read:org".
Recipes for the three sibling failures
Failure
Skill to load
Self-hosted runner stuck / unhealthy
gh-actions-stuck-self-hosted-runner-recovery
Action job log shows specific error
wa-cloud-run-deploy-failure-debug (Cloud Run) or gh-actions-slow-runs (general perf)
