Refactor capi.py: provider pattern for endpoint-specific logic

[anticomputer]

Replace the StrEnum + three scattered match/case blocks with a single APIProvider dataclass registry. Endpoint-specific behavior (catalog path, response parsing, tool-call detection, headers) now lives in one place per provider. Copilot-Integration-Id header is no longer leaked to non-Copilot endpoints.

[Copilot]

Pull request overview

This PR refactors the CAPI integration to centralize endpoint-specific behavior behind an APIProvider registry, reducing scattered endpoint conditionals and preventing Copilot-only headers from being sent to other providers.

Changes:

• Introduces an APIProvider dataclass + provider registry to encapsulate per-endpoint logic (catalog path, parsing, tool-call detection, headers).
• Updates model catalog + tool-call capability helpers to delegate behavior to the selected provider.
• Updates the agent client construction to only attach provider-specific headers (e.g., Copilot-Integration-Id) when appropriate, and adjusts tests accordingly.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
src/seclab_taskflow_agent/capi.py	Replaces endpoint enum/match blocks with provider registry and provider-based model/tool-call logic.
src/seclab_taskflow_agent/agent.py	Uses provider-derived defaults and ensures provider-specific headers are only sent to matching endpoints.
tests/test_capi_extended.py	Updates/expands tests to cover provider registry and new tool-call capability semantics.
tests/test_api_endpoint_config.py	Removes enum-based assertions and adds provider base URL resolution coverage.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[p-]

We just changed the default models this morning

seclab-taskflow-agent/src/seclab_taskflow_agent/agent.py

Line 44 in 474b44f

default_model = "gpt-4.1"

[anticomputer]

Rebased on main, addressed feedback:

Default models (p-): Updated all provider defaults to gpt-4.1 to match main.

base_url unused (copilot-reviewer): list_capi_models() now uses provider.base_url as the canonical URL instead of the raw parameter. Known providers always resolve to their configured base URL.

_DEFAULT_PROVIDER dead code (copilot-reviewer): Removed. get_provider() constructs the fallback inline.

check_tool_calls default (copilot-reviewer): Keeping bool(model_info) as the base default. The previous heuristic ("gpt-" in model.lower()) was worse — it missed o1, o3-mini, o3, and any non-gpt model. Defaulting to False would break custom endpoints that don't publish capability metadata. This only affects CLI -l listing, not runtime behavior.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (4)

src/seclab_taskflow_agent/capi.py:169

• httpx.URL(base).join(provider.models_catalog) will drop the last path segment when base has a path and doesn’t end with / (urljoin semantics). With the OpenAI provider’s base_url="https://api.openai.com/v1" and models_catalog="models", this will likely request https://api.openai.com/models instead of .../v1/models. Consider either (a) ensuring provider base URLs intended for joining end with a trailing slash, or (b) making models_catalog an absolute/fully-qualified path for providers like OpenAI (e.g., v1/models) so the final catalog URL is correct.

        r = httpx.get(
            httpx.URL(base).join(provider.models_catalog),
            headers=headers,

src/seclab_taskflow_agent/capi.py:66

• APIProvider.check_tool_calls() currently defaults to return bool(model_info), which makes list_tool_call_models() treat every catalog entry as tool-call capable for providers that don’t override this (notably the OpenAI provider). That contradicts the list_tool_call_models docstring (“only models that support tool calls”) and regresses prior OpenAI filtering behavior. Suggest adding an explicit OpenAI provider implementation (or other reliable rule) so non-chat models (embeddings/audio/etc.) don’t get incorrectly reported as tool-call capable.

    def check_tool_calls(self, _model: str, model_info: dict) -> bool:
        """Return True if *model* supports tool calls according to its catalog entry."""
        # Default: optimistically assume support when present in catalog
        return bool(model_info)

src/seclab_taskflow_agent/capi.py:124

• For unknown endpoints, get_provider() returns APIProvider(name="custom", base_url=url) which inherits the default default_model="gpt-4.1". This silently changes the previous behavior where unknown endpoints forced an explicit default model, and it can cause confusing runtime failures for custom/self-hosted providers that don’t have a gpt-4.1 model. Consider setting a sentinel/empty default_model for the custom provider (or requiring COPILOT_DEFAULT_MODEL when name=="custom") to preserve the explicit configuration requirement.

def get_provider(endpoint: str | None = None) -> APIProvider:
    """Return the ``APIProvider`` for the given (or configured) endpoint URL."""
    url = endpoint or get_AI_endpoint()
    netloc = urlparse(url).netloc
    provider = _PROVIDERS.get(netloc)
    if provider is not None:
        return provider
    # Unknown endpoint — return a generic provider with the given base URL
    return APIProvider(name="custom", base_url=url)

src/seclab_taskflow_agent/capi.py:197

• list_tool_call_models() behavior changed to use the provider registry (and now depends on each provider’s check_tool_calls implementation), but there are no direct tests covering list_tool_call_models for the built-in providers (Copilot/GitHub Models/OpenAI). Adding tests here would help catch regressions like over-including non-tool-capable models for OpenAI or incorrect catalog URL construction.

def list_tool_call_models(token: str, endpoint: str | None = None) -> dict[str, dict]:
    """Return only models that support tool calls."""
    models = list_capi_models(token, endpoint)
    provider = get_provider(endpoint)
    return {
        mid: info
        for mid, info in models.items()
        if provider.check_tool_calls(mid, info)
    }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

src/seclab_taskflow_agent/capi.py:188

• The catalog URL is constructed with httpx.URL(base).join(provider.models_catalog). When base_url includes a path segment like /v1 (OpenAI) or a custom endpoint with versioned paths, URL.join() treats the base path as a file and can drop the last segment (e.g. https://api.openai.com/v1 + models becomes https://api.openai.com/models). This will break list_capi_models() for those endpoints. Consider storing models_catalog as an absolute path (e.g. /v1/models) or adjusting URL construction so it reliably appends/overrides paths as intended for each provider.

        r = httpx.get(
            httpx.URL(base).join(provider.models_catalog),
            headers=headers,

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[anticomputer]

@kevinbackhouse heads up — this PR has merge-order conflicts with #186, #189, #195 (and mildly #191, #203, #205). Those PRs fix linter rules in the StrEnum/to_url()/match-case code that this refactor deletes entirely. If they land first we'd just be resolving conflicts against dead code. Cleanest path is to merge this first, then rebase the lint-fix PRs — the capi.py hunks become no-ops and the agent.py changes should still apply.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.