Add bounded retries for transient Codex server errors

[tsubasakong]

Summary

• add bounded retries for transient OpenAI/Codex server_error responses in the legacy runtime path
• add the same retry semantics to pi_session_bridge so both backends recover consistently from transient upstream failures
• extend runtime seam coverage for retryable and non-retryable Codex errors on both backends

Why

Transient upstream OpenAI/Codex server errors were terminating phone tasks immediately, which broke long-running automations even when a simple retry would have recovered. The legacy and bridge backends also behaved differently, so the runtime needed a consistent recovery path.

Changes

• add shared retry classification helpers and reuse the formatted model error logger for retry handling
• retry bounded server_error/5xx-style model failures with exponential backoff (1s, 2s) before surfacing a failure
• keep non-retryable model errors and aborted runs on the existing failure path
• update continuation handling so follow-up work spawned by retries and textual fallback execution is fully drained before finalizing the run
• add seam tests covering retry success and non-retryable failure for both legacy_agent_core and pi_session_bridge

Testing

• node --test test/runtime-seams.test.mjs
• npm test

Checklist

• I ran relevant tests, or the Testing section explains why I did not.
• I updated docs, or confirmed no doc changes were needed.
• I confirmed the PR does not include secrets, credentials, or private data.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openpocket	Ready	Preview, Comment	Mar 13, 2026 6:42pm