Files
codex/codex-rs/app-server
jif 5d3fe48b08 Improve Guardian retries and review failure reporting (#44482)
## Why

Transient rate limits can end automatic approval reviews prematurely, and review failures currently report high risk even when no assessment completed.

## What changed

- Retry rate limits and recoverable exhausted-stream errors, while excluding non-transient HTTP failures.
- Preserve server retry delays after stream retries are exhausted and honor them within the review deadline. Scope retry advice to the current turn so reused sessions cannot apply stale delays.
- Keep failed reviews denied, but leave risk and authorization unset and explain that the review could not complete without declaring the action unsafe.

## Testing

Add an integration test covering rate-limit recovery through approval and tool execution, asserting that the action executes exactly once after approval. Update failure assertions to check absent assessment fields and the review-failure explanation.

GitOrigin-RevId: 1163cfde35c6b8eb23b6f24f1f24461b86ded838
2026-09-10 09:50:59 +00:00
..

User verification cancellation (experimental)

Local UI clients can cancel a native user-verification RPC by sending userVerification/cancel with {requestId} and the experimentalApi opt-in. The result is an empty acknowledgment ({}). This API does not enable desktop verification capability advertisement.

requestId is the original status, enroll, delete, or verify RPC's string or integer ID on the same connection, not the server elicitation ID. Use fresh IDs for each operation and a distinct ID for the cancel RPC. Unknown, finished, unrelated, and other-connection requests are no-ops.

The acknowledgment confirms the cancellation signal without waiting for the OS prompt to close. The original RPC completes independently, with cancelled/interrupted when cancellation prevents completion. Cancellation cannot roll back completed effects. It remains effective while a proof waits for outbound queue capacity, but cannot retract a response already enqueued.

Canceling or resolving an elicitation does not itself stop a separate userVerification/verify RPC. Clients must cancel that RPC separately and discard late proofs after the approval is canceled or resolved. Only one native worker runs per app-server; if an OS call remains active after cancellation or timeout, subsequent local operations return failed/providerError until that worker exits.

Hosted Codex Apps MCP protocol

The host-owned HTTP codex_apps server uses Legacy by default in app-server and standalone Codex. To discover the 2026-07-28 protocol, set codex_apps_mcp_2026_07_28 = true under [features], or send a true runtime override via experimentalFeature/enablement/set. Discovery falls back to Legacy when the server does not support it. Explicit config takes precedence. The dedicated setting does not apply to third-party HTTP or local codex_app stdio servers. The existing mcp_2026_07_28 flag still governs eligible other servers, regardless of whether their names or URLs resemble hosted Apps. App-server does not persist this selection.

Thread removal

thread/archive and thread/delete reject attempts to remove a live internal worker with JSON-RPC error -32600. The worker's owner controls its shutdown. For example, a Guardian reviewer remains available to its parent conversation after a client tries to archive or delete it.

After the owner releases the worker, its saved conversation can be archived or deleted normally. Ordinary client-controlled threads keep their existing behavior.

Amazon Bedrock authentication

If model_providers.amazon-bedrock.aws.credential_export is configured, Bedrock setup and Bedrock login return an error without changing configuration or saved credentials. Remove the exporter configuration before selecting another credential source. aws.credential_export and aws.profile cannot be configured together.