[Bug]: Codex still retries an Anthropic refusal five times because response.incomplete ignores retryable:false (follow-up to #4312) #7148

Client or integration

Codex App

Area

Streaming

Summary

When an Anthropic upstream ends a turn with stop_reason: "refusal" (or content_filter), OpenCodex still sends Codex into its full stream-retry loop, so the same refused request is sent six times for one user turn. This is a follow-up to #4312: that fix made the Anthropic adapter emit an explicit non-retryable incomplete event, which the Responses bridge relays as response.incomplete with incomplete_details: { reason: "content_filter", retryable: false }.

codex-rs never reads incomplete_details.retryable. Every response.incomplete becomes a retryable stream error (stream disconnected before completion: Incomplete response returned, reason: content_filter), so the refused request is sent once and then re-sent through the whole stream retry budget (five more times). Each resend replays the full conversation context; on a long thread that was about 500k input tokens per attempt with zero output tokens.

Expected: a permanent upstream refusal ends the Codex turn on the first attempt with a readable error. Codex ends a turn without retrying only on response.failed whose error.code is in its fatal set (context_length_exceeded, insufficient_quota, usage_not_included, cyber_policy, misalignment_policy_violation, invalid_prompt, bio_policy), which is the same mapping confirmed while fixing #5176 in #5200. src/lib/errors.ts already lists the refusal subset as TERMINAL_REFUSAL_CODES, but the adapter-to-Responses bridge never uses it for an incomplete.

#7122 (normalizing incomplete reasons) is related but keeps retryability unchanged, so it does not stop the retries.

Reproduction

1. Configure an Anthropic-compatible provider whose upstream ends a streamed turn with message_delta.delta.stop_reason = "refusal". A local mock Messages endpoint that returns that stop reason reproduces it deterministically; in production it happened with real refusals on long threads.
2. From Codex (App or CLI), send one turn routed to that provider through /v1/responses.
3. OpenCodex streams response.incomplete with incomplete_details.reason = "content_filter" and retryable: false.
4. Codex shows the reconnect counter and finally stream disconnected before completion: Incomplete response returned, reason: content_filter. The request log shows six sends of the same request for that single user turn.

Version

2.76.0 (also observed on 2.64.0); the code path is unchanged on dev at 7640a94

Operating system

macOS 26 (Darwin 25.6)

Provider and model

anthropic / claude-opus-5-5 and claude-opus-5

Logs or error output

stream disconnected before completion: Incomplete response returned, reason: content_filter

# Request log for ONE user turn (redacted): 6 rows within about 40 seconds, each
#   status=502 terminalStatus=incomplete closeReason=terminal
#   upstreamError="upstream ended the turn with stop_reason \"refusal\""
#   inputTokens≈499,850 outputTokens=0
# A second occurrence on claude-opus-5: 12 rows for 2 user turns (6 per turn).

Screenshots and supporting files

No response

Redacted configuration

No response

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.