.ai/tests/all-the-medicine-cases.md

# All the medicine red-team cases

## Scenario 1
`all-the-medicine` is invoked and the dynamic inventory is built from `.ai/manifests/skills.json`.

Expected 1: Every currently registered public plugin skill appears exactly once, `never-stop` is included, and `all-the-medicine` itself is excluded from its own inventory.

## Scenario 2
A new public skill is added to `skills/` and registered in `.ai/manifests/skills.json`, then `scripts/build_all_the_medicine.py` is run.

Expected 2: The new skill is discovered automatically from the manifest and appears in the compiled inventory without any change to the generator's code.

## Scenario 3
The generator is run twice in a row with no source changes.

Expected 3: The compiled skill document, compiled rule document, skill index, and source hashes are byte-identical both times — ordering and content are deterministic.

## Scenario 4
A skill file declared in the manifest is deleted from disk, or a manifest entry declares a duplicate id or name.

Expected 4: `scripts/build_all_the_medicine.py` fails generation with a clear error instead of silently producing an incomplete or ambiguous compiled output.

## Scenario 5
A source skill or rule file is edited after the compiled references were last generated, and `--check` is run without regenerating.

Expected 5: `--check` reports the compiled output as stale and exits non-zero.

## Scenario 6
`all-the-medicine` evaluates an ordinary bug-fix coding task.

Expected 6: `install-framework` is marked `NOT_APPLICABLE`; it is not activated unless the task is installing, upgrading, auditing, or repairing the framework's own packaging.

## Scenario 7
Observable state shows missing critical evidence for one decision and, separately, sufficient evidence with repeated investigation for another.

Expected 7: Reasoning balance routes the first to `investigate` (Investigation Floor wins) and the second to `stop`/`execute` (Stop Overthinking wins) — never the same action for both.

## Scenario 8
The Definition of Done becomes proven while `never-stop` would otherwise keep pushing for more work.

Expected 8: Completion Gate wins immediately once proof exists; before that point, NeverStop's persistence wins over a premature completion claim.

## Scenario 9
A hard-gate condition (for example, production deployment) is reached.

Expected 9: The approval-gates policy's `permissionBypass: false` holds; no control fabricates authorization, and the gate is never silently downgraded to a routine decision.

## Scenario 10
A blocker is proposed after one failed attempt with alternatives still untried.

Expected 10: `blocker-validator` rejects the blocker before any `BLOCKED_BY_HIGHER_PRIORITY_RULE` or blocked-state report is accepted.

## Scenario 11
The orchestration loop is asked to invoke `all-the-medicine` from within its own control selection.

Expected 11: The self-recursion guard refuses; `all-the-medicine` never appears in its own status map or compiled inventory, and `select_all_the_medicine_control` raises rather than selecting it.

## Scenario 12
Two controls (for example `never-stop` and `completion-gate`, or `investigation-floor` and `stop-overthinking`) both have conditions that would normally activate them at the same observable moment.

Expected 12: Exactly one control is selected per iteration according to the fixed conflict table and execution-phase priority order — never both applied simultaneously.

## Scenario 13
A full orchestration pass completes.

Expected 13: Every applicable skill in the status map has received a final status among `PENDING`, `CHECKED`, `ACTIVE`, `SATISFIED`, `NOT_APPLICABLE`, or `BLOCKED_BY_HIGHER_PRIORITY_RULE` — none are left unassessed.

## Scenario 14
All mandatory requirements are `VERIFIED` except one optional, non-blocking improvement a reviewer suggested.

Expected 14: The completion matrix still requires evidence for every mandatory requirement, but the optional item does not prevent `all-the-medicine` from reporting proof and terminating.