feat(provider): forward reasoning effort across OpenAI and Anthropic providers

[rodboev]

Summary

SkillSpector now applies SKILLSPECTOR_REASONING_EFFORT across its OpenAI-compatible and native Anthropic provider paths instead of silently dropping the setting when users switch from ChatOpenAI-based providers to native ChatAnthropic.

Closes #283

Implements the maintainer-requested Anthropic expansion from the July 20 review.

Root cause

The shared OpenAI-compatible constructor already forwarded SKILLSPECTOR_REASONING_EFFORT, but both native Anthropic providers built ChatAnthropic directly and bypassed that seam. The proxy rewrite also lacked a regression test for nested output_config.effort, so the PR had no proof that the Anthropic request body preserved the setting after URL and auth rewriting.

Diff Notes

• keep the shared OpenAI-compatible forwarding path as pass-through for non-empty literals
• add an Anthropic-specific resolver that trims input, preserves unset or blank defaults, and accepts low, medium, high, xhigh, and max
• wire the validated value into AnthropicProvider and AnthropicProxyProvider
• add direct proxy rewrite coverage for nested output_config.effort
• update the env-var docs so the setting is no longer described as OpenAI-only

Scope

This change covers the shared OpenAI-compatible path plus the two native Anthropic providers. OpenAI-compatible providers still pass non-empty literals through unchanged. Native Anthropic providers accept low, medium, high, xhigh, and max and reject other non-empty values at the provider boundary. Bedrock and CLI-provider behavior stay out of scope for this PR.

Verification

• uv sync --all-extras passed; checked 151 packages.
• Focused provider and proxy tests passed through the repo-configured Windows runner for python -m pytest tests/unit/test_providers.py tests/unit/test_anthropic_proxy_provider.py; pytest reported 106 passed, 9 skipped.
• uv run ruff check src/skillspector/providers/chat_models.py src/skillspector/providers/anthropic/provider.py src/skillspector/providers/anthropic_proxy/provider.py tests/unit/test_providers.py tests/unit/test_anthropic_proxy_provider.py passed with All checks passed!.
• uv run ruff format --check src/skillspector/providers/chat_models.py src/skillspector/providers/anthropic/provider.py src/skillspector/providers/anthropic_proxy/provider.py tests/unit/test_providers.py tests/unit/test_anthropic_proxy_provider.py passed with 5 files already formatted.

[keshprad]

Could you please expand this PR to support the native Anthropic provider paths as part of this PR, rather than leaving them out of scope?

SKILLSPECTOR_REASONING_EFFORT is provider-neutral, but currently it only affects providers using ChatOpenAI. ChatAnthropic definitely supports model reasoning through adaptive thinking and explicit effort configuration.

Could we translate the common SkillSpector setting into the native Anthropic request format in this PR?

• AnthropicProvider should map the setting to the corresponding ChatAnthropic effort configuration.
• AnthropicProxyProvider is also built on ChatAnthropic, so it should use the same configuration where the proxy accepts those fields. Please add coverage confirming that the proxy request rewrite preserves output_config.effort.
• Anthropic-specific values should be validated at the provider boundary because its accepted effort levels are low, medium, high, xhigh, and max, rather than the full OpenAI value set.

Otherwise, the same environment variable silently has no effect when users switch from an OpenAI-compatible provider to a native Anthropic provider.

References:

• Anthropic effort and adaptive thinking
• LangChain ChatAnthropic reasoning configuration

[keshprad]

FYI, I edited the comment: Let's remove Bedrock from scope for this PR since it seems Bedrock has a wide range of such parameter names including thinking, output_config.effort, budget_tokens.

We can come back to this support in a followup PR or future work based on community feedback. @rodboev Do you agree with that approach?

[keshprad]

Requesting changes to include native Anthropic support in this PR. Please map SKILLSPECTOR_REASONING_EFFORT for both AnthropicProvider and AnthropicProxyProvider, validate the Anthropic-specific accepted values at the provider boundary, and add coverage confirming that the proxy rewrite preserves output_config.effort. See the linked conversation comment for the detailed rationale and references.

[rodboev]

Agreed on keeping Bedrock out of this PR. I limited the expansion to the two native Anthropic paths covered by the formal review.

1. I added Anthropic-specific validation for SKILLSPECTOR_REASONING_EFFORT, accepting low, medium, high, xhigh, and max, while treating unset or blank input as provider default. The OpenAI-compatible pass-through behavior stays unchanged.
2. AnthropicProvider and AnthropicProxyProvider now pass accepted values through ChatAnthropic's effort configuration. The existing model, credentials, endpoint, timeout, and token-limit settings stay intact.
3. I added a direct proxy rewrite assertion showing that nested output_config.effort survives the URL, authentication, and body rewrite, and I updated the env-var docs and PR description so they no longer describe the setting as OpenAI-only.

I left Bedrock parsing, translation, tests, and documentation out of this PR. Its provider-specific reasoning controls can be handled separately if that support is requested later. CLI provider behavior also stays outside this change.

[keshprad]

[P1] Could we remove the Anthropic-specific validation and keep SKILLSPECTOR_REASONING_EFFORT as a provider pass-through setting?

The behavior should be consistent across providers:

• If the environment variable is unset or blank, omit the argument and preserve the provider default.
• If it is non-empty, trim and forward it unchanged.
• Let LangChain or the upstream provider validate whether the value is supported by the selected model.

Maintaining an Anthropic-specific allowlist in SkillSpector creates an unnecessary capability matrix that can become stale as models and supported effort levels evolve. It is also inconsistent with the OpenAI-compatible path, which already passes non-empty values through.

Please remove the local Anthropic enum validation and its invalid-value tests. Retain coverage for trimming, pass-through, omission when unset or blank, and preservation of output_config.effort through the proxy rewrite. The documentation should note that supported values depend on the selected provider and model.

[keshprad]

[P2] Add the new variable to the test_llm_utils environment cleanup.

_LLM_ENV_VARS omits SKILLSPECTOR_REASONING_EFFORT, despite promising to clear all LLM-related variables. I reproduced a deterministic failure with:

SKILLSPECTOR_REASONING_EFFORT=invalid ... pytest \
  tests/unit/test_llm_utils.py::TestCredentialResolution::test_get_chat_model_returns_native_anthropic_client

Add the variable to that fixture’s cleanup tuple.

[keshprad]

Two small changes and this looks good to me. Thanks @rodboev!