feat: make parallel_tool_calls injection configurable for codex-api-key channels

[ypq123456789]

Problem

The codex executor hardcodes "parallel_tool_calls": true in the request template sent to upstream APIs. This causes 400 errors on third-party OpenAI-compatible providers that strictly validate parameters and do not support this field without tools being specified.

Error from upstream:

Invalid value for 'parallel_tool_calls': 'parallel_tool_calls' is only allowed when 'tools' are specified.

Current Behavior

CPA injects parallel_tool_calls: true into every codex request regardless of whether the upstream supports it. This is hardcoded in the binary — there is no config option to disable it.

From CPA logs:

[debug] [codex_executor.go:866] request error, error status: 400, error message: Invalid value for 'parallel_tool_calls': 'parallel_tool_calls' is only allowed when 'tools' are specified.

Impact

Any third-party OpenAI-compatible API that strictly validates request parameters (e.g. evomap.ai, and likely others) cannot be used as a codex-api-key channel. The alias model mapping feature works correctly, but the injected parallel_tool_calls causes the upstream to reject the request.

Channels that are lenient about extra parameters (songsongai, zicc, etc.) work fine because they silently ignore the unsupported field.

Suggested Solutions

Any of the following would resolve this:

1. Per-channel config option: Add a parallel-tool-calls: false field to codex-api-key entries to disable injection for specific channels
2. Capability-based: Respect the supports_parallel_tool_calls model capability flag — if the upstream model doesn't declare support, don't inject it
3. Global config option: A top-level disable-parallel-tool-calls flag
4. Strip on error: If upstream returns 400 for parallel_tool_calls, retry without it

Environment

• CPA Version: 7.1.59
• Config: codex-api-key with alias model mapping
• Upstream: OpenAI-compatible API (non-OpenAI provider)

[ypq123456789]

Workaround: Use payload.filter to strip parallel_tool_calls

CPA already has a built-in payload filter system that can remove parameters from requests before sending to upstream. Add this to your config.yaml:

payload:
  filter:
    - models:
        - name: "*"
          protocol: "codex"
      params:
        - "parallel_tool_calls"

This strips parallel_tool_calls from all codex requests before they reach the upstream API. Tested and confirmed working on v7.2.16 with evomap.ai (which was returning 400 due to strict parameter validation).

Test results:

• Before filter: 400 "Invalid value for parallel_tool_calls"
• After filter: 200 OK, all 18/18 requests succeeded

The filter is applied per-channel based on protocol and model matching, so you can target specific upstream providers if needed rather than stripping globally.