🪲 BUG-#28: Fix GenericProvider race between complete() and reload()

[FernandoCelmer]

Description

Fixes a race condition in GenericProvider where a concurrent reload() could mutate url, model, headers, and other config fields mid-read inside complete(). Introduces a _ConnectionSnapshot frozen dataclass and a threading.Lock so complete() takes a single consistent snapshot before doing any I/O.

Motivation and Context

Closes #28. CodeLoop.ask() documents it is safe to call while run() is mid-turn on another thread, but GenericProvider had no protection against non-atomic config mutation during reload().

Types of changes

• Bug fix

Checklist

• Self-review done
• Tests added
• CHANGELOG updated
• Docs updated

[FernandoCelmer]

The _ConnectionSnapshot pattern is a clean solution for eliminating the torn-read problem. Two issues are worth addressing before merge.

[FernandoCelmer]

[Blocking] _ConnectionSnapshot is missing request_builder

Problem — _ConnectionSnapshot is documented as "a consistent, point-in-time copy of everything reload() can mutate", but request_builder is absent from the dataclass. reload() reassigns self.request_builder inside the lock, and complete() currently calls self.request_builder(...) inside the same lock window — so for the call itself it is correct. But the snapshot abstraction is incomplete: any future helper that accepts a _ConnectionSnapshot and expects snap.request_builder will get an AttributeError, and any new code path that reads self.request_builder outside the lock will silently introduce a race.

Failure scenario — A developer adds def _build_body(self, snap: _ConnectionSnapshot, ...) and writes snap.request_builder(...). This compiles fine, fails at runtime with AttributeError: _ConnectionSnapshot has no field request_builder.

Fix — Add the field to the snapshot and populate it in _snapshot_locked():

@dataclass(frozen=True)
class _ConnectionSnapshot:
    ...
    request_builder: Callable  # add this

def _snapshot_locked(self) -> _ConnectionSnapshot:
    return _ConnectionSnapshot(
        ...
        request_builder=self.request_builder,
    )

Then replace self.request_builder(system_prompt, messages, tools, config.model) in complete() with config.request_builder(...).

---

[Suggestion] Lock is held during request body serialisation

Problem — complete() calls self.request_builder(system_prompt, messages, tools, config.model) inside with self._lock. Building the request body (serialising tool schemas and message arrays) is CPU-bound work. Holding the shared lock for this duration unnecessarily delays any concurrent reload() call.

Failure scenario — A caller passes a large tools list with complex nested schemas. reload() on another thread blocks for the full serialisation duration instead of just the microseconds needed to copy scalar fields.

Fix — Take the snapshot under the lock, release, then build the body:

def complete(self, system_prompt, messages, tools, ...):
    with self._lock:
        config = self._snapshot_locked()
    # lock released — body build is now contention-free
    body = config.request_builder(system_prompt, messages, tools, config.model)
    ...