Prompt caching on Codex was not reliably reusable through the proxy because repeated chat-completions requests could reach the upstream without the same continuity envelope. In practice this showed up most clearly with OpenCode, where cache reads worked in the reference client but not through CLIProxyAPI, although the root cause is broader than OpenCode itself.
The proxy was breaking continuity in several ways: executor-layer Codex request preparation stripped prompt_cache_retention, chat-completions translation did not preserve that field, continuity headers used a different shape than the working client behavior, and OpenAI-style Codex requests could be sent without a stable prompt_cache_key. When that happened, session_id fell back to a fresh random value per request, so upstream Codex treated repeated requests as unrelated turns instead of as part of the same cacheable context.
This change fixes that by preserving caller-provided prompt_cache_retention on Codex execution paths, preserving prompt_cache_retention when translating OpenAI chat-completions requests to Codex, aligning Codex continuity headers to session_id, and introducing an explicit Codex continuity policy that derives a stable continuity key from the best available signal. The resolution order prefers an explicit prompt_cache_key, then execution session metadata, then an explicit idempotency key, then stable request-affinity metadata, then a stable client-principal hash, and finally a stable auth-ID hash when no better continuity signal exists.
The same continuity key is applied to both prompt_cache_key in the request body and session_id in the request headers so repeated requests reuse the same upstream cache/session identity. The auth manager also keeps auth selection sticky for repeated request sequences, preventing otherwise-equivalent Codex requests from drifting across different upstream auth contexts and accidentally breaking cache reuse.
To keep the implementation maintainable, the continuity resolution and diagnostics are centralized in a dedicated Codex continuity helper instead of being scattered across executor flow code. Regression coverage now verifies retention preservation, continuity-key precedence, stable auth-ID fallback, websocket parity, translator preservation, and auth-affinity behavior. Manual validation confirmed prompt cache reads now occur through CLIProxyAPI when using Codex via OpenCode, and the fix should also benefit other clients that rely on stable repeated Codex request continuity.
When assistant has tool_calls but no text content, the translator
emitted an empty message into the Responses API input array before
function_call items. The API then couldn't match function_call_output
to its function_call by call_id, returning:
No tool output found for function call ...
Only emit assistant messages that have content parts. Tool-call-only
messages now produce function_call items directly.
Added 9 tests for tool calling translation covering single/parallel
calls, multi-turn conversations, name shortening, empty content
edge cases, and call_id integrity.
Add util.SanitizeClaudeToolID() to replace non-conforming characters in
tool_use.id fields across all five response translators (gemini, codex,
openai, antigravity, gemini-cli).
Upstream tool names may contain dots or other special characters
(e.g. "fs.readFile") that violate Claude's ID validation regex.
The sanitizer replaces such characters with underscores and provides
a generated fallback for empty IDs.
Fixes#1872, Fixes#1849
Made-with: Cursor
The instructions restore logic was originally needed when the proxy
injected custom instructions (per-model system prompts) into requests.
Since ac802a46 removed the injection system, the proxy no longer
modifies instructions before forwarding. The upstream response's
instructions field now matches the client's original value, making
the restore a no-op.
Also removes unused sjson import.
Closesrouter-for-me/CLIProxyAPI#1868
Claude's Tool Search feature (advanced-tool-use-2025-11-20 beta) adds
defer_loading field to tool definitions. When proxying Claude requests
to Codex or Gemini, this unknown field causes 400 errors upstream.
Strip defer_loading (and cache_control where missing) in all three
Claude-to-upstream translation paths:
- codex/claude: defer_loading + cache_control
- gemini-cli/claude: defer_loading
- gemini/claude: defer_loading
Fixes#1725, Fixes#1375
Add support for Claude's "adaptive" and "auto" thinking modes using `output_config.effort`. Introduce support for new effort level "max" in adaptive thinking. Update thinking logic, validate model capabilities, and extend converters and handling to ensure compatibility with adaptive modes. Adjust static model data with supported levels and refine handling across translators and executors.
Non-spark codex models (gpt-5.3-codex, gpt-5.2-codex) stream function call
arguments via multiple delta events followed by a done event. The done handler
unconditionally emitted the full arguments, duplicating what deltas already
streamed. This produced invalid double JSON that Claude Code couldn't parse,
causing tool calls to fail with missing parameters and infinite retry loops.
Add HasReceivedArgumentsDelta flag to track whether delta events were received.
The done handler now only emits arguments when no deltas preceded it (spark
models), while delta-based streaming continues to work for non-spark models.
Some Codex models (e.g. gpt-5.3-codex-spark) send function call arguments
in a single "done" event without preceding "delta" events. The streaming
translator only handled "delta" events, causing tool call arguments to be
lost — resulting in empty tool inputs and infinite retry loops in clients
like Claude Code.
Emit the full arguments from the "done" event as a single input_json_delta
so downstream clients receive the complete tool input.
The OpenAI Chat Completions translator was silently dropping
response.function_call_arguments.delta and
response.function_call_arguments.done Codex SSE events, meaning
tool call arguments were never streamed incrementally to clients.
Add proper handling mirroring the proven Claude translator pattern:
- response.output_item.added: announce tool call (id, name, empty args)
- response.function_call_arguments.delta: stream argument chunks
- response.function_call_arguments.done: emit full args if no deltas
- response.output_item.done: defensive fallback for backward compat
State tracking via HasReceivedArgumentsDelta and HasToolCallAnnounced
ensures no duplicate argument emission and correct behavior for models
like codex-spark that skip delta events entirely.
- Standardized the handling of `stop_reason` and `finish_reason` across Codex and Gemini responses.
- Restricted pass-through of specific reasons (`max_tokens`, `stop`) for consistency.
- Enhanced fallback logic for undefined reasons.
- Replaced all instances of `bytes.Clone` with direct references to enhance efficiency.
- Simplified payload handling across executors and translators by eliminating unnecessary data duplication.
Google official Gemini Python SDK sends thinking_level, thinking_budget,
and include_thoughts (snake_case) instead of thinkingLevel, thinkingBudget,
and includeThoughts (camelCase). This caused thinking configuration to be
ignored when using Python SDK.
Changes:
- Extract layer: extractGeminiConfig now reads snake_case as fallback
- Apply layer: Gemini/CLI/Antigravity appliers clean up snake_case fields
- Translator layer: Gemini->OpenAI/Claude/Codex translators support fallback
- Tests: Added 4 test cases for snake_case field coverage
Fixes#1426
