[HTExplicit]

Problem

Codex conversations on /v1/responses end with continuation_state_unavailable and cannot be resumed. Production, 2026-10-04/05: 48 such failures in three conversations, all with recovery.not_attempted_reason = ambiguous_encrypted_carriers.

What happened in the reported conversation (Ops records of API key 101, group 53):

1. 21:38:45 its OpenRouter account ran out of credit and failover moved the conversation to an account on api.openai.com, a different upstream organisation. Reasoning ciphertext minted through the first organisation cannot be verified by the second.
2. For the next 38 requests the upstream answered 400 invalid_encrypted_content, recovery removed every reasoning ciphertext and the retry succeeded. cache_skipped_items was 0 each time: the rejected-ciphertext memory accepts at most 64 hashes per call and received the whole history (230+ reasoning items) in one call, so nothing was remembered and nothing was stripped before sending.
3. 21:57 the history gained a Codex multi-agent agent_message whose body is ciphertext in a content part (content[].encrypted_content). The policy compares the number of all encrypted_content keys in the body with the number of reasoning ciphertexts; they now differed, the rejection was treated as ambiguous and no retry was made. 12 attempts failed the same way.

The guard is from 2026-09-07 and was written for compaction items. It is not a regression of the recent releases; session stickiness did not change either. The upstream error has param: null and names the rejected item only in its message: The encrypted content for item rs_… could not be verified.

Change

• Selection. A rejection whose message names, by id, a reasoning item that carries ciphertext in the sent body is no longer ambiguous when other ciphertext carriers exist. It is recovered like any unindexed rejection: every reasoning ciphertext is removed, one retry, same account. Compaction items and agent_message content parts are never modified. Without a named reasoning item the refusal stays. The policy receives one boolean (named_cipher), not the message.
• Memory. skipRejectedHistory and the write in TryRecover call the store in batches of OpenAIReasoningStateMaxLookupEntries inside the existing shared I/O budget. After the first recovery on an account the old ciphertexts are stripped before sending, so later requests are not rejected and ciphertext minted on the new account is kept.
• Diagnostic. history.nested_encrypted counts ciphertext held in content parts. Backend JSON only.

Unchanged: recognised codes, the single retry, no account switch, protected statuses, the global switch, WS and Messages.

Limits

• The id match needs reasoning ids on the wire: API-key accounts and auto-passthrough. The OAuth / Setup Token transform removes reasoning ids before sending, so a history with another ciphertext carrier is still refused there.
• If an agent_message body was itself minted by another organisation, the retry is rejected again and the request ends as before. That ciphertext cannot be removed without losing the message.
• The memory keeps one entry per remembered ciphertext per source (156 bytes, 32768 entries overall, LRU). Long histories now use it; eviction falls back to reject-and-retry.

Verification

• go test ./internal/codexruntime/recovery/
• go test -tags unit ./internal/service/ -run 'TestOpenAIContinuationDiagnostic|TestOpenAIReasoningRecovery|ReasoningRecovery|RejectedReasoning|ProjectReasoningCipher|SystemPrompt.*Reasoning'
• go test -tags unit ./internal/repository/ -run 'ReasoningState|RejectedReasoning'
• New cases: named reasoning item with an agent_message present (HTTP error and response.failed envelope), no id in the message, an amsg_ id, a longer identifier with the same prefix; the agent_message item is byte-identical after the strip; a 150-item history is written and read in three batches against a store that refuses more than 64 per call.
• Reviewed from four angles with each finding checked by a second reader; the doc sentence was narrowed as a result. No real model calls were made.

🤖 Generated with Claude Code