Harden GitHub connector: distributed token cache, error classification, write operations

[haasonsaas]

Summary

The GitHub connector works for basic read-only proxy flows but has several gaps for production use.

Current state

• internal/connectors/github/connector.go — 4 read-only operations allowlisted
• internal/connectors/github/executor.go — in-memory token cache, basic error handling
• internal/connectors/github/app_token_source.go — in-memory installation→token cache with no distributed backing

Required work

• Distributed token cache: move GitHub App installation-token cache from in-memory to Redis so multiple API replicas share cached tokens and don't each hit GitHub's token creation endpoint
• Error classification: distinguish GitHub API rate limits (429), permission errors (403), transient failures (5xx), and not-found (404) — return appropriate broker error codes for each
• Token pre-refresh: if a cached token has < 5 minutes remaining, refresh proactively before the next request uses it
• Write operations (when needed): add allowlisted operations for create_issue, create_pull_request_comment, create_check_run — each with appropriate permission scoping
• Per-repository permission scoping: ensure installation tokens are minted with only the permissions needed for the specific operation, not the full app permission set
• Fallback logic: if GitHub App token exchange fails (misconfigured app, installation removed), fall back to static token if available, or return a clear error

Files

• internal/connectors/github/connector.go — add operations
• internal/connectors/github/executor.go — error classification
• internal/connectors/github/app_token_source.go — Redis-backed cache, pre-refresh

Priority

Medium.

🤖 Generated with Claude Code

[haasonsaas]

Working remaining #12 slice now: GitHub write-operation allowlisting (create_issue, create_pull_request_comment, create_check_run) with per-operation GitHub App permission scoping and scoped token-cache keys. Full go test ./... and repo-wide lint are green locally; PR incoming next.