[sarah-chitty-1]

What

A characterization test of prompt_agent's user side, so the rework in #265 can show it leaves that side unchanged. Each test records every request a conversation makes, in order, and what it stores. The requests cover the agent, the user sim or a human peer, agent-trigger decide, env trigger state, and the firing-log readback. The records cover the stored conversation and the env_trigger_snapshots, agent_trigger_firings, usersim_turn_outputs and agent_trigger_state metadata.

The cases:

• The LLM user sim: a reply, done (its message stays unrecorded), malformed JSON, a reply missing done, extra fields, a timeout (its task is cancelled), a task that fails, is canceled or is rejected, and a reply with no parts.
• Agent triggers: one fires and speaks for the user, one ends the conversation with a closing message, or none fires and the user sim answers.
• A human peer: named by user_a2a_url, and the configured default.
• How a run ends: a conversation closed from outside, an agent failure, and the turn limit.

Worth a look

• Test only: no code changes.
• Peers are answered through an httpx MockTransport and the conversation goes to the real local store, so the test pins requests and stored records, not function calls. A refactor that keeps the behaviour passes it unchanged. Ids are normalized: the conversation's is <ctx>, and the others are <idN> in order of appearance.
• The env in the agent-trigger tests offers only triggers/v1, so a change that reads an env's clock wouldn't show up in these request logs.

Tests

• The new file: 17 passed.
• Full unit suite (tst/unit, packages/agentenv-protocol/tests): 10442 passed, 2 failed. Both failures are in test_mcp_cli_builder.py, whose generated scripts run this machine's system python3, which has no click. They're unrelated to this change.
• ruff check: clean. uv lock --check: ok.

🤖 Generated with Claude Code

Confidence Score: 5/5

This test-only PR appears safe to merge.

Summary

Adds tests that pin down requests and stored records for prompt_agent conversations.

• Prompt-agent tests now pin down simulator turns and run endings.
• Prompt-agent tests record when agent triggers speak for the user.
• Prompt-agent tests cover replies from an explicitly chosen or default human peer.

_{Reviews (2) · Last reviewed commit: "review: the user-side test pins the user..." · Reviewed by Greptile}