Split out of #133 so it isn't lost when that issue closes. Not load-bearing — nothing here is gated on it, and #156 already mitigated the local impact.
The inconsistency
anthropics/claude-code-action retries its OIDC token exchange 3× but does not retry its actor permission check:
GET /repos/{owner}/{repo}/collaborators/{actor}/permission
A single 503 there kills the whole run. Observed on 2026-08-17:
Attempt 1 failed: GitHub API is temporarily unavailable while verifying repository access.
Attempt 2 of 3...
App token successfully obtained ← the token exchange retried and recovered
Checking permissions for actor: smartwatermelon
GET /repos/.../collaborators/.../permission - 503
##[error]Failed to check permissions
Two calls to the same degraded API, seconds apart, with opposite retry policies. The one that retries recovers; the one that doesn't takes down the run.
Evidence
smartwatermelon/archive-resolver#25, run 31986353713 — failed twice with an identical signature ~2.5h apart, on an unrelated one-file shell change.- The endpoint was failing ~40% of probes (3 of 8) while the GitHub status page reported everything except Copilot operational.
- The step exits in ~13s; a real review takes ~1m11s.
Ask
Apply the same retry policy to the permission check that already wraps the token exchange.
Local state
v3.2.1 (#156) retries the review once when a probe confirms the endpoint is failing, which takes outage-induced failures from ~40% to ~16%. That is a mitigation, not a fix — a sustained outage still fails both attempts, and every consumer pays two runs to discover it.
Action needed here: open the issue on anthropics/claude-code-action with the above. No code change in this repo.
Refs #133, #156
Split out of #133 so it isn't lost when that issue closes. Not load-bearing — nothing here is gated on it, and #156 already mitigated the local impact.
The inconsistency
anthropics/claude-code-actionretries its OIDC token exchange 3× but does not retry its actor permission check:A single 503 there kills the whole run. Observed on 2026-08-17:
Two calls to the same degraded API, seconds apart, with opposite retry policies. The one that retries recovers; the one that doesn't takes down the run.
Evidence
smartwatermelon/archive-resolver#25, run 31986353713 — failed twice with an identical signature ~2.5h apart, on an unrelated one-file shell change.Ask
Apply the same retry policy to the permission check that already wraps the token exchange.
Local state
v3.2.1(#156) retries the review once when a probe confirms the endpoint is failing, which takes outage-induced failures from ~40% to ~16%. That is a mitigation, not a fix — a sustained outage still fails both attempts, and every consumer pays two runs to discover it.Action needed here: open the issue on
anthropics/claude-code-actionwith the above. No code change in this repo.Refs #133, #156