Offline qualification
What the fixtures and compatibility checks prove.
Adapted from toolkit · f5911f6Monarchy's integration suite follows a request through the Secretary and back to the Queen. It uses the real SQLite record layer, the real Conductor adapter and the real supervision code. A fixture supplies native CLI responses so the tests can drop a reply or kill a process at the awkward moment.
That is offline qualification. It does not launch a coding agent or prove that a live Conductor host behaves exactly like the fixture. A separate read-only native pilot passed on November 16, 2026 (UTC). Its receipt records the narrower guarantees observed on the actual host.
The roundtrip
tests/test_qualification.py registers the
owner's request with a task, brief, authority reference and existing 00-Agent
binding. From there it exercises:
- Prepare a dispatch without sending it, then deliver it through the adapter.
- Read the matching native message back and retain its original identifiers.
- Ingest a 00-Agent report from a different workspace using native sender identity.
- Let the Secretary handle routine progress quietly and save a decision for the Queen when the 00-Agent needs a choice.
- Restart the runtime, reopen the same pending presentation, record the owner's answer and prepare its delivery as a separate obligation.
- Restart again, deliver the answer, ingest the 00-Agent's completion report and check independent review evidence against the exact source revision.
- Restart once more and find the verified result in the Queen's inbox. Reading it leaves it pending; explicit processing clears that presentation.
The fixture's owner and reviewer references stand in for trusted host inputs. The test checks their record lifecycle. It does not authenticate a human or assess the quality of a real code review.
The failures that matter
| Scenario | Required result |
|---|---|
| The process dies after native acceptance, before saving the receipt | The saved operation remains in_flight; restart reads the original message back without sending again |
| A send times out and its message is temporarily invisible | The operation stays unknown until later readback confirms it; repeated recovery sends nothing |
| The Queen misses a notification | The durable outcome still appears when she opens the inbox |
| A presentation is reopened after restart | It has the same identity and remains unprocessed |
| The same native report arrives again | Its immutable report identity produces one stored event, decision and outcome |
| Triage saves an outcome and fails before acknowledging its event | The next pass resumes the claim and reuses the saved outcome |
| A review belongs to an older source revision | Verification refuses it, even when another completion report exists |
| A 00-Agent says done while an owner question is unanswered | Completion and presentation acknowledgment leave the decision pending |
| An answer is recorded but has no delivery operation yet | Verification remains blocked until the answer has been delivered |
| An answered question is revised before its prepared message is sent | The old answer cannot be sent as current authority |
| A question changes during native preflight | The record layer refuses send admission before the native write |
| A report comes from a retired assignment | Native validation rejects it before it enters the current inbox |
| A 00-Agent receives a skill archive without the source checkout | Selected skills, references and exact project policy files remain readable and hash-verifiable |
The process-death case uses a separate Python process and os._exit after the
fixture persists native acceptance. It deliberately skips the clean database
close and the adapter's final receipt. The native fixture keeps its own file so
it survives independently of the runtime process.
Duplicate report handling is an identity guarantee inside the runtime. It is not a claim of exactly-once execution across a network or inside an agent.
Run it
From the repository root, with Python 3.9 or newer. The full suite also needs Bash and jq for the isolated upstream outcome test:
python3 -m unittest discover -s tests -p test_qualification.py -v
python3 -m unittest discover -s tests -v
git diff --checkThe focused command runs the integration scenarios. The full command also runs the observer, record, transport, supervision, package and upstream-contract tests. All fixtures use temporary directories. No native mutation, credentials, model call or shared project state is required.
The regression tests were written before their corresponding fixes. In the first assembled run, the full roundtrip passed, while three tests exposed gaps: an undelivered answer did not block verification, a superseded answer could still be sent, and a revision change during preflight did not fence the send. Keep those cases when changing either the supervisor or the record layer.
What comes from First Mate
These are behavioral contracts adopted from First Mate at
427056c4181c41074ad24f051f79919330df0fa5, with a different record and transport
implementation:
| Upstream contract | Monarchy check |
|---|---|
| Actor-scoped wake claims and acknowledgment | Save the outcome before consuming the claimed event; replay after a crash preserves the obligation |
| Separate presentation and processing | Opening the Queen inbox does not acknowledge its contents |
| Durable delivery before the notification | A missing notification or uncertain reply cannot erase the pending work or justify another send |
| Decision holds separate from task completion | 00-Agent completion and inbox processing cannot supply an owner answer |
Monarchy intentionally keeps unanswered decisions after a 00-Agent reports completion. It does not copy First Mate's task-type-specific terminal clearing rules. The integration fixture exercises Monarchy's equivalent contracts. A separate upstream test runs First Mate's pinned outcome store in an isolated temporary home and checks that presentation leaves processing outstanding. Neither test certifies compatibility with its Pi or tmux hosts. The adoption plan records the broader boundary.
What still needs a live host
The native pilot records actual dispatch, cross-workspace reports, delayed readback recovery and independent result verification. The native owner decision and answer path still needs qualification with trusted host input. The fixture covers that record lifecycle without supplying human authentication.
The CLI cannot express explicit fast mode off during session creation. The adapter therefore holds that creation path; existing-session reads and recovery remain available. A live test must use a profile the adapter can express and must record any native fields it cannot verify. It must not silently choose a different profile to make the test pass.
The current Queen inbox is a pull interface. These tests verify that pending
outcomes survive a missed notification, and that open or unknown prompt state
requests a hold. They do not establish an automatic native prompt latch or a
notification service. Such a host integration needs its own qualification.
The skill scenarios are also separate: their not_run status means
no model has been evaluated against them. Package tests prove that the selected
instructions arrive intact, not that an agent will follow them.
Finally, process restart is not machine-loss recovery. SQLite receipts depend on the underlying storage. Backup, restore to a new host, unattended scheduling, real project acceptance and authorization for external actions remain separate requirements. Check what Monarchy has proved before entrusting it with more work.