`src/testing.rs` — the mock toolkit [feature: testing]

Fakes for every moving part, so agent logic runs end-to-end with zero network. Compiled only under the feature — no production impact.

Key items

• MockApiClient — scripted model: new(model), with_responses(vec![...]), with_text_response, with_tool_call, with_stop_reason, with_error. Emits the real provider event sequence; usage fixed 50/25; last script entry repeats forever; set_model works (model-switch tests); response_format served with the canned text (parse-failure tests); tool_constraint rejected loudly.
• MockResponse { text, tool_call, stop_reason }, MockToolCall { id, name, input } — the script shapes.
• MockTool — knobs for dispatch behavior: with_result, with_error, with_delay, with_concurrency_safe, with_read_only, with_schema. Ignores its input.
• Fixtures: test_config(), test_message, test_assistant_message, test_tool_use_message.
• EnvGuard — snapshot/set/restore environment variables across a test, serialized between tests, restores even on panic.

Behavior notes

• The mock ignores request content entirely — “the model should have seen X” assertions belong to observers or the audit trail.
• with_responses(vec![]) is a no-op (keeps the default script).

Deep dive: Testing.