Follow the code.
See the gates.
Explore two recorded harness outcomes: a change accepted in an isolated repository, and a defect blocked after passing tests.
A guided replay of local development results, not a live agent run. This example uses Companion and Orchestrator; they are optional and independent. Use either, both or neither in your own workflow.
Choose an outcome to see what passed, why work stopped and what the evidence supports. These controls replay a summary; they do not run an agent or change code.
- Context
- Plan
- Code
- Validate
- Review
- Integrate
- Code
- Qwen3 Coder Next · LM Studio
- External validation
- Passed · tree unchanged
- Independent review
- GPT-6.1 · refused
- Integration
- Blocked
A child’s cancellation exception was replaced by the wrong error. The review caught it before integration.
Recorded local tests · October 10, 2026. One development fixture. Scope and evidence
Every stage has a job.
Context
Relevant sources and reviewed decisions ground the task.
Plan
Companion makes the objective, phases and acceptance criteria visible.
Code
Orchestrator dispatches a native agent inside the declared scope.
Validate
An external command checks the proposed change and records proof.
Review
A separate native session reviews the candidate and can refuse it.
Integrate
Configured gates decide whether the reviewed candidate can be integrated.
A checked result, with a defined scope.
The accepted run used Haiku 4.5 for coding and Sonnet 4.6 for review. A test-local launcher supplied Claude CLI controls. The Qwen3 Coder Next run used LM Studio through Codex tools; GPT-6.1 reviewed it and refused an exception-handling defect.
Both used one exposed development fixture. Model support, permissions and execution controls depend on the adapter. These runs show the gates operating, not a general reliability score.