Structured Output from Tools

PROJECT CONTEXT:
My pipeline parses agent output with regex and hope, and breaks weekly. Enterprise agent frameworks ground everything in typed transformations: schema-validated outputs with repair loops and explicit fallbacks.

GOAL:
Move all model-to-system boundaries onto schemas: defined output types, validation, a bounded repair loop, and fallbacks that never crash the pipeline.

REQUIREMENTS:
- Every model output consumed by code has a schema (Pydantic/JSON Schema/equivalent)
- Native structured-output API modes used where the provider supports them
- A bounded repair loop: validation errors fed back for one correction attempt
- Explicit fallbacks per output type when repair fails
- Validation telemetry: pass/repair/fallback rates tracked per output type

CONSTRAINTS:
- No regex-parsing of prose at system boundaries - schema or it does not ship
- Fallbacks must be safe and visible, never silent invented defaults

DEFINITION OF DONE:
- All boundary outputs validate against their schemas in tests
- Malformed model output triggers repair and recovers in the common case
- Repair failure produces the typed fallback plus an alert, not a crash
- Telemetry shows validation rates per output type in normal operation

COMMON FAILURES TO AVOID:
- Regex scraping of markdown that breaks on every format drift
- Schemas that exist but are not enforced at runtime
- Unbounded repair loops burning tokens on unfixable output
- Silent fallbacks that fabricate data downstream systems trust