This protocol preserves the private/public boundary: raw provider responses, authorized source material, cookies, transcripts, drafts, and command exports remain in the parent-private submission/evidence/ workspace. Only the verifier, redacted schema, and this procedure belong in the public repository. Passing the verifier proves consistency; retained cloud exports prove provenance.
Phase 1: preflight
Before creating or changing application state, confirm an active gcloud identity and matching project. Record the deployed Cloud Run revision for both services, the exact Agent Engine resource, Firestore database, Pub/Sub topic/subscription, secret metadata (never values), and deployed health response. HARMONIA_MOCK_AI and HARMONIA_MOCK_X must be unset, as must every emulator setting. Any missing or unauthorized prerequisite stops the run.
Also record the ready operations composite index, the subscription retry/dead-letter policy, and
both scheduler jobs (harmonia-durable-autonomy and harmonia-durable-recovery). Confirm the worker
health response exposes the expected protocol/compiler and bounded recovery settings without any
credential, tenant content, prompt, or artifact body.
The evidence runner path above is relative to the private parent workspace, not this public repository.
Phase 2: capture-awaiting-approval
Use Harmonia’s normal dashboard or chat to ingest one authorized public YouTube URL. Do not call an internal stage or decision endpoint by hand. Wait until the persisted job reaches awaiting_approval, then run the read-only collector without --require-complete. Preserve each W3C traceId and its lineage across web ingress, Pub/Sub stage events, ADK cognition, usage, and the pending action. A later operator decision is a separate inbound request and may correctly begin a new trace.
At this checkpoint, the collector must not approve, publish, retry, or manufacture a receipt. An incomplete run is expected and must not be presented as a completed vertical slice.
Operator checkpoint
The signed-in operator reviews the real drafts and clip candidates in the normal dashboard and makes the approval decision there. This manual checkpoint is intentionally outside the collector. Rejection ends the candidate run honestly; it is never rewritten as approval.
Phase 3: capture-complete
After the normal workflow persists a human approval, a finalized pre-effect claim, effect receipt, and independent verification, capture again:
The complete capture must correlate the approval action, finalized effect claim, idempotency key, effect receipt, verification readback or artifact digest, cognition usage, cost records, and trace lineage. The claim must follow durable human approval, precede effect execution, own the same idempotency key, operation, and action trace as the effect, and finalize against the original receipt. Cognition must share a trace with a lifecycle event; approval, claim, effect, and verification must share their action trace; replay must be a distinct traced attempt. A missing or late claim, zero trace ID, verification timestamp before effect execution, receipt mismatch, unknown observed cost, mock provenance, or raw private field fails closed.
The redacted bundle also requires replay evidence: a distinct replay operation after the original effect must return already_applied and reference the original receipt.
Restart and ambiguity proof
In a separate authorized test job, interrupt the worker after a persisted boundary and before its
next commit. Preserve the pre-interrupt operation epoch and lease, allow or invoke bounded recovery,
and record the replacement claim’s newer epoch. Attempt the old finalize call and retain the rejected
fence result. Redeliver the same source event and retain the canonical duplicate response.
For effects, inject failure only at the supported deterministic fault boundary. A failure before
provider dispatch may return to prepared; a failure after dispatch must remain unknown until
provider reconciliation or the operator-resolution workflow supplies digest-verified evidence.
Never perform the provider call merely to manufacture failure evidence, and never relabel unknown as
failed. Keep raw records and command output in the private parent evidence directory.
Replay proof
Replay the already-approved action only through Harmonia’s Prove duplicate suppression audit control. The normal claim state machine must return already_applied and reference the original receipt; the replay route has no authority to execute. Capture that durable replay observation and the external readback privately.
Google Calendar evidence
Calendar sync is useful supporting product evidence, but it does not replace the required publishing or content-pack effect in the vertical-slice bundle.
To claim Calendar integration in the demo, capture privately:
- the Google consent screen showing only
calendar.app.created;
- the connected workspace state with credentials redacted;
- one operator-triggered event creation and its verified link/timestamp;
- one Harmonia reschedule that becomes
update_required, followed by an operator-triggered update and a newer verified ETag/timestamp; and
- one operator-triggered removal followed by Google API
404 or 410 read-back.
Retain the content-item document before and after each action, sanitized Calendar API request/response metadata, and trace correlation outside the public repository. Offline tests, a green connection badge, or an event visible only in the Harmonia UI are not live-provider proof. Do not include OAuth tokens, calendar IDs tied to a personal account, personal event data, or private extended-property values in public artifacts.
The collector itself performs neither the original action nor the replay. Do not edit a failed bundle to make it pass. Preserve command status, verifier output, and unresolved gaps exactly as observed. Last modified on August 28, 2026
