fix: slug->UUID canonicalisation in paper_flow + repair magic resolver structured-output schemas

[keola808hunt-dot]

Summary

Fixes two openai-agents SDK structured-output failures that prevented paper_flow and the magic natural-language resolver from running end-to-end. Both stem from QuantMind's rich Pydantic models (UUID-keyed trees, discriminated-union inputs, ModelSettings) meeting the SDK's strict-schema requirements.

1. paper_flow — 94 uuid_parsing errors (the crash)

With strict_json_schema=False, the model emits human-readable node-id slugs ("root", "intro", "methodology") into the UUID-typed Paper/TreeNode fields → 94 uuid_parsing validation errors. The slug tree is internally coherent (parent/child/citation refs all consistent) — the model's behaviour is correct for one-shot tree generation; it just collides with the storage type.

The domain UUID is load-bearing (tested node-id uniqueness, dict[UUID, TreeNode] JSON round-trips, stable dedup identity), so rather than weaken it to str, this adds a slug-tolerant extraction boundary:

• quantmind/knowledge/_extraction.py — pure canonicalize_tree_ids() maps each distinct slug to one UUID and rewrites every id slot (nodes keys, node_id, parent_id, children_ids, root_node_id, citation anchors), passing values that are already UUIDs through unchanged.
• PaperExtraction(Paper) — a model_validator(mode="before") running the canonicalizer; paper_flow defaults output_type to it. The domain model is untouched.

2. magic resolver — un-schemable ModelSettings + strict-mode rejection (latent)

ResolvedFlowConfig embeds BaseFlowCfg.model_settings: ModelSettings, whose callable fields cannot be JSON-schema'd (PydanticInvalidForJsonSchema), and the discriminated-union / knowledge schemas trip strict mode's additionalProperties guard. This was hidden because the existing resolver tests mock Runner.run.

Two layers, both required:

• SkipJsonSchema on model_settings — it's an execution knob (set programmatically), never LLM-populated, so it is skipped during schema generation. Lets the schema build at all.
• AgentOutputSchema(..., strict_json_schema=False) on the resolver output type — accepts the additionalProperties the union/knowledge models emit (same pattern as paper_flow).

3. Portable local-file provenance

_fetch_and_format records local paths via Path.as_posix() so provenance is consistent cross-platform (fixes a Windows-only test).

Tests & verification

• New: tests/knowledge/test_extraction.py (slug→UUID + UUID passthrough); ResolverOutputSchemaTests (real, non-mocked schema build exercising both magic layers); a fallback-coverage test.
• Updated: test_output_type_override_propagated and test_basemodel_renders_json_schema to the post-fix contracts; a non-strict wrap-guard assertion on the resolver.
• Full suite: 238 passing. ruff and basedpyright clean.
• Live verification (real models, not mocks): paper_flow on arXiv 2606.05138 (the exact paper that 94-errored) builds a valid Paper; the magic resolver via gpt-4o-mini parses natural language into a correct ArxivIdentifier + config. The strict-mode layer of the magic fix was caught by the live run — the mocked tests could not surface it.

Notes

• No new dependencies (SkipJsonSchema ships with pydantic); the lockfile is unchanged.
• DOI input remains an intentional NotImplementedError (documented follow-up).

🤖 Generated with Claude Code

[github-actions]

This pull request has been inactive for 14 days and is now marked stale. It will be closed in 7 days if there is no further activity. Comment or remove the stale label to keep it open.