fix(copilot): infer nested prompt schema from output definitions

[lesandiz]

Summary

• build the Copilot prompt schema recursively from nested output: definitions instead of flattening to top-level fields only
• include nested object properties, required keys, and array item schemas in the prompt-facing schema used for initial guidance and parse recovery
• add regression coverage for recursive schema generation and the actual prompt sent to Copilot

Validation

• uv run --python 3.13 pytest tests\test_providers\test_copilot.py -q
• uv run --python 3.13 conductor run C:\Users\lester.copilot\session-state\9a4adc0c-6650-46e6-9540-061d5b29a7ef\files\conductor-nested-output-repro.yaml --input topic="AI Harness Engineering"

Context

• no exact open issue or open PR was found for this bug
• related prior work: docs: add web dashboard section to README with screenshot #48 flattened an example workflow schema, but did not fix provider-side schema inference

[lesandiz]

@microsoft-github-policy-service agree

[jrob5756]

Hello - Thank you so much for your contribution to conductor. One small recommendation. If you can add this check, I'd be happy to merge.

[jrob5756]

Missing recursion depth limit — These three new recursive methods (_build_prompt_schema → _build_prompt_field_schema → _build_prompt_schema / _build_prompt_item_schema → _build_prompt_item_schema) have no depth guard. A pathologically deep schema would cause a RecursionError crash.

The Claude provider already protects against this with _max_schema_depth = 10 and checks at the top of its corresponding methods (_build_json_schema_properties and _build_single_field_schema).

Suggested fix: add a depth parameter (default 0) to all three methods and raise ValidationError if depth > MAX_SCHEMA_DEPTH, matching Claude's pattern:

MAX_PROMPT_SCHEMA_DEPTH = 10

def _build_prompt_schema(self, schema: dict[str, OutputField], depth: int = 0) -> dict[str, Any]:
    if depth > self.MAX_PROMPT_SCHEMA_DEPTH:
        raise ValidationError(...)
    ...

[lesandiz]

Thanks for reviewing Jason. I applied your suggestion, following existing pattern in claude.py.