Skill: Agent Task Eval

Capability Domain: TEST_INTEGRATION
Technology Stack: pytest, golden-tasks, optional-deepeval
Linked Activities: TFK-04, TFK-05, TFK-09, DTA-19

Overview

Patterns for lane 4 (task performance quality): frozen tasks with structured ground truth.

Principles

1. Structured oracle is source of truth.
2. Same harness as lane 1; live provider at temp=0.
3. Not a PR merge gate — promotion band vs last certified baseline.
4. Deepeval optional overlay for fuzzy SC-01 only.

Golden task catalog

See artifact 56 Part 4.6 — TASK-* under tests/fixtures/agent_tasks/.

Makefile / CI

make test-agent-quality gated on AGENTS_ENABLED; separate agent-eval.yml.

References

Artifact 56 Part 4.6; DTA-19 §16.