OpenAI-compatible chat adapter ignores configured timeout and hard-codes 120s #703

Issue

Title

OpenAI-compatible chat adapter ignores configured timeout and hard-codes 120 seconds

Description

CECLI appears to ignore the configured timeout value for OpenAI-compatible chat providers.

I am using a custom OpenAI-compatible provider with a long-running non-streaming model response.

My config includes:

timeout: 3000

model-providers:
  very_slow_response_model:
    api_base: "http://192.168.10.89:9000/v1"
    requires_api_key: false
    supports_stream: false

CECLI debug output confirms that the configured timeout is propagated into the request kwargs:

"timeout": 3000.0

However, the request is abandoned at approximately 120 seconds, after which CECLI prints:

Retrying in 0.2 seconds...

The upstream proxy is still processing the request at that point and later completes successfully.

Example timing:

t = 0s
Request sent from CECLI

t ≈ 120s
CECLI:
Retrying in 0.2 seconds...

t ≈ 160s
Upstream proxy:
claude.completed duration_ms=160324 exit_code=0
chat_completion_response model=xxx
usage | xxx | in 32078 out 13033

The upstream proxy itself has a 300-second timeout, so it is not the source of the 120-second cutoff.

I traced this to:

cecli/helpers/llms/domains/chat.py

which contains:

DEFAULT_TIMEOUT = 120.0

and uses it directly in both the non-streaming and streaming paths:

async with make_client(timeout=DEFAULT_TIMEOUT, verify=VERIFY_SSL) as client:
    resp = await client.post(...)

This means the higher-level configured timeout is effectively ignored by the OpenAI-compatible chat adapter.

Expected behavior

The adapter should honor the configured/request timeout, for example:

timeout = float(kwargs.get("timeout", DEFAULT_TIMEOUT))

async with make_client(timeout=timeout, verify=VERIFY_SSL) as client:
    ...

This should be applied to both:

• chat_complete()
• chat_stream()

Secondary retry issue

After the 120-second timeout triggers a retry, CECLI also crashes in the retry path with:

TypeError: cannot pickle '_thread.RLock' object

Traceback points to:

cecli/models.py
cecli/helpers/config_utils.py

where deep_merge() calls copy.deepcopy() on request kwargs:

kwargs = deep_merge(kwargs, {"allowed_openai_params": ["tools", "tool_choice"]})

and:

merged = copy.deepcopy(dict1)

So there appear to be two separate issues:

1. OpenAI-compatible chat requests are hard-limited to 120 seconds despite a larger configured timeout.
2. The retry path can fail with _thread.RLock during deepcopy().

Environment

CECLI version: 1.6.1
Python: virtual environment on Windows
Provider type: custom OpenAI-compatible
Streaming: disabled
Configured timeout: 3000 seconds
Observed client cutoff: ~120 seconds
Upstream successful completion: ~160 seconds

The same provider/model works correctly for shorter requests.

Version and model info

No response