Preserve Codex encrypted reasoning across Claude turns

[mulfyx]

Summary

CCP already requests reasoning.encrypted_content from the Codex Responses API, but the Claude translation previously discarded it:

• outgoing Claude thinking blocks received an empty signature;
• assistant thinking blocks were ignored when the next Claude request was translated back into Responses input.

As a result, encrypted reasoning could not survive stateless multi-turn replay when previous_response_id was disabled.

This PR round-trips the Codex reasoning item through a CCP-owned, namespaced Claude thinking signature:

Codex { id, encrypted_content }
  -> Claude thinking.signature
  -> next Codex reasoning input item

The encrypted payload is preserved byte-for-byte. CCP does not decrypt or interpret it.

Protocol rationale

OpenAI documents that reasoning items should be included in subsequent input when application code manages conversation history manually. Requesting reasoning.encrypted_content makes the opaque reasoning payload available for this purpose:

• OpenAI Responses API reasoning documentation
• Codex requests reasoning.encrypted_content
• Codex reasoning item representation

Claude's extended-thinking protocol provides an opaque signature field. During streaming, Anthropic specifies that signature_delta is emitted immediately before content_block_stop:

• Claude extended thinking documentation

This PR uses that field only as a transport for a namespaced CCP envelope. Foreign, malformed, or oversized signatures are ignored.

Prior art

Other Codex-to-Claude adapters use the same general round-trip pattern.

CLIProxyAPI translates Claude thinking signatures into Codex encrypted_content, requests encrypted reasoning from Responses, and maps it back into Claude thinking blocks:

• Claude signature -> Codex encrypted reasoning
• Requesting reasoning.encrypted_content
• Codex encrypted reasoning -> Claude signature
• Handling summary and signature-only thinking blocks

The pi-mono multi-provider adapter independently stores complete Responses reasoning items in thinkingSignature and restores them on the next request:

• Restoring the reasoning item
• Saving it in thinkingSignature
• Requesting encrypted reasoning

Implementation

• Add a versioned ccp:codex:v1: signature envelope containing the reasoning item ID and encrypted payload.
• Capture id and encrypted_content from Responses streaming events.
• Emit signature_delta before content_block_stop.
• Support signature-only thinking blocks when Codex returns encrypted reasoning without a visible summary.
• Restore CCP-owned signatures as Responses reasoning items in their original transcript position.
• Preserve encrypted content exactly without decoding or modifying it.
• Ignore foreign, malformed, and oversized signatures.
• Support both buffered and streaming translation paths.
• Include signed thinking blocks in token estimation and request summaries.

Validation

• cargo test --all-targets: 559 passed, 0 failed
• rustfmt 1.96 --check: passed
• Tested against the real Codex API through an isolated CCP instance on port 18766.
• CCP_CODEX_PREVIOUS_RESPONSE_ID=0 was used to force full transcript replay.
• The second upstream request contained exactly one restored reasoning item.
• Its ID and encrypted payload matched the first response byte-for-byte.
• The two-turn request completed successfully.
• /v1/messages/count_tokens accepted the signed transcript.
• The active proxy on port 18765 was not touched.

Scope

This PR only changes encrypted-reasoning round-trip behavior. It does not change HTTP transport, previous_response_id handling, Docker configuration, authentication, or diagnostic logging.

[raine]

Thanks