Assert Agent Story

Mimir metadata: slug do-assert-agent-story; always_apply: false (link to BPE-02 / MIN-04 / TFK-02 / DTA-19).

When a plan, iteration manifest, or SAO §17 row declares Agent proof scenarios (PRF-*, agent_proof_script[], or agent_proof_command), you MUST prove orchestration with pytest integration tests using CAP-004 ScriptedLLM and the shared trace helper — not by asserting chat prose, grepping logs/app.log alone, or "the agent loop exists" checklists.

Requirements

1. For each PRF row mapped in SAO §17, a test asserts run trace beats (workflow, tools offered/called, domain effects, blackboard/plan state as applicable) via tests/support/agent_story.py → assert_agent_story(...).
2. Cover happy and reject paths when the PRF is marked adverse or has a reject sketch (test_*_agent_story_happy / test_*_agent_story_reject).
3. Use ordered LLM scripts from tests/fixtures/llm_scripts/<prf-id>/ injected through the agent factory — only CAP-004 ScriptedLLM may stand in for the LLM in integration tests.
4. Agent-proof tests ship in the same commit as the CAP/worker behavior they prove — never a deferred "agent testing pass".
5. Never assert on raw secrets, tokens, or passwords in trace or log lines.
6. Do not mock domain services, tools, or brokers in @agent_proof tests — real DB and real tool registry per Test Trophy.

Relationship to log story

• Log story (do-assert-log-story): proves logging beats at decision points (caplog).
• Agent story: proves control-plane behavior (trace + domain) under scripted LLM turns.
• Both may apply to the same slice; both must green in the same commit when declared.

Anti-patterns

• Behavior test green while PRF trace absent or weakened to "agent completed"
• FakeLLM that returns [] on exhaustion instead of raising (hides missing script turns)
• Mocking ToolExecutor or domain services in @agent_proof integration tests
• Skipping checkpoint.agent_proof_command when the manifest lists PRF proofs
• Using live LLM in PR merge gate (belongs to lane 3/4 — see TFK-01 §6)