fix: litellm-compat for dynamiq integration tests (v0.4.8)

[vitalii-dynamiq]

Summary

Three drop-in gaps prevented dynamiq's test fixtures from passing against arcllm even though direct API calls worked. After this PR dynamiq's full suite passes: 1149 unit + 1066 integration, 0 failures (was 281 integration failures).

What was wrong

1. Exception positional args

litellm's exception classes take (message, llm_provider, model, ...) positionally. arcllm made these keyword-only after message. Tests construct errors as RateLimitError("msg", "bedrock", "amazon.titan") — that raised takes 2 positional arguments but 4 were given.

• ArcLLMError: provider/model/status_code now positional after message; llm_provider stays keyword-only as the litellm-name alias.
• RateLimitError: accepts (message, provider, model) positionally.
• ProviderAPIError: detects litellm shape (status_code, message, ...) by type — first int positional becomes status_code.
• BadRequestError (renamed from InvalidRequestError to match the canonical litellm/OpenAI name; InvalidRequestError stays as alias): accepts (message, model, provider) per litellm and (message, provider, model) per arcllm. Disambiguates by checking `SUPPORTED_PROVIDERS`.

2. Streaming chunk serialisation

`Choice.model_dump()` omitted `.delta`. dynamiq's streaming callback reads `chunk["choices"][0]["delta"]["content"]` from the serialized dict, so it `KeyError`'d on every streamed event.

3. `token_counter` overhead

Counts now follow OpenAI's per-message formula (3 per message + 1 for `name` + 3 priming) so totals match litellm's. The previous sum-of-fields undercount made dynamiq's history-summarisation logic preserve more context than the model could actually accept.

4. `ModelResponse` shape (rolled in from earlier work)

• `choices` defaults to `[Choice()]` so fixtures doing `ModelResponse()["choices"][0]["message"]["content"] = ...` work.
• `stream: bool = False` added so `ModelResponse(stream=True)` is accepted (litellm idiom).
• `Choice.delta` added so streaming fixtures can set delta on the same `Choice` class litellm uses for both chat and stream modes.

Test plan

• arcllm unit + parity tests: 783 passed (8 pre-existing Ollama integration failures unrelated)
• `ruff check` + `ruff format --check` + `mypy --strict` + `pyright` on changed files: clean
• dynamiq unit tests: 1149 passed, 0 failed
• dynamiq integration tests: 1066 passed, 0 failed (was 281 failed before this PR)
• Manual exception construction smoke (positional + keyword + litellm-shape) verified for AuthenticationError, RateLimitError, ProviderAPIError, BadRequestError

---

Note

Medium Risk
Medium risk because it changes public exception constructors and token-counting semantics, which can affect downstream error handling and context-window logic; intended to improve litellm/dynamiq compatibility but may alter behavior for existing callers relying on the old signatures/counts.

Overview
Improves LiteLLM/dynamiq drop-in compatibility by relaxing exception constructor signatures (more positional-arg tolerant), renaming InvalidRequestError semantics to BadRequestError (keeping aliases), and adding litellm-style positional parsing to ProviderAPIError/RateLimitError.

Fixes streaming fixture compatibility by serializing Choice.delta in Choice.model_dump() and extending response shapes (Choice.delta, ModelResponse(stream=...)). Updates token_counter() to apply OpenAI chat per-message overhead (plus priming/name tokens) so message-based counts match litellm/OpenAI. Also bumps version to 0.4.8 in __init__ and pyproject.toml.

^{Reviewed by Cursor Bugbot for commit 6692eeb. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6692eeb. Configure here.}

[cursor]

Single-arg BadRequestError skips documented disambiguation heuristic

Low Severity

When only arg2 is provided (without arg3), the code unconditionally treats it as provider via kwargs.setdefault("provider", arg2). However, the class docstring states the heuristic checks SUPPORTED_PROVIDERS to disambiguate — "if the second positional looks like a provider name, treat it as provider; otherwise treat it as model." This means BadRequestError("msg", "gpt-4") (litellm pattern where the second positional is a model) incorrectly assigns provider="gpt-4" instead of model="gpt-4".

 

^{Reviewed by Cursor Bugbot for commit 6692eeb. Configure here.}