What this test proves
One colima invocation runs the canonical Hermes testbed test.sh, then asks real agents in that fixture to execute AT9, AT10, and AT11. Every step records an independent verdict, execution continues after failure, and the aggregate passes only when all four panels pass.
Evidence layout
The driver writes exactly one run directory under site/reports/integration/colima/<run>/. Step runners own only the step.json and raw sidecars in the directory passed by the driver. scripts/integration/render_colima.py is the sole renderer for step panels, the run page, integration.json, and the sibling envelope. The four steps always run in the sequence above.
Changes
The revision sections below identify each driver revision that produced committed evidence.
Flow
Steps
| step | action | observable | evidence |
|---|---|---|---|
| 1 | Run `colima-hermes-skills` in the fixture | the seven skill reports and fixture health checks have a retained verdict | `steps/NN-hermes-skills/step.json` and raw sidecars |
| 2 | Run `colima-task-start` in the same fixture | task-start notification, ordering, and reminder cases have a retained verdict | `steps/NN-task-start/step.json` |
| 3 | Run `colima-assignment` in the same fixture | assignment/task-pass cases have a retained verdict | `steps/NN-assignment/step.json` |
| 4 | Run `colima-prompt-handoffs` in the same fixture | the fixture agent's public task-event and doctor observations have a retained verdict | `steps/NN-prompt-handoffs/step.json` |