Getting started/End-to-end walkthrough

End-to-end walkthrough

A realistic fake run showing scoped prompt, submit, polling, terminal result, transcript, decomposition, read-only oracle, evidence tiers, cost reporting, partial behavior, and corrective hermes_respond.

1. Scoped prompt

markdown

Update the docs safety page.

## Acceptance
- The page warns that bridge access is operator-grade access.
- The page says unauthenticated MCP initialize must fail.
- The page says authenticated MCP initialize must succeed.
- The page reports cost as known, unknown, or unreconciled.

Each acceptance bullet becomes a user-requirement claim.

2. Submit and poll

satellite runbook

hermes_submit(prompt, caller="docs-satellite")
=> task_id: task_01SAFEWALK, status: pending
# wait 30s, then poll every 120s
hermes_status(task_01SAFEWALK) => running
hermes_status(task_01SAFEWALK) => completed

3. Terminal result and cost

json

{
  "task_id": "task_01SAFEWALK",
  "status": "completed",
  "session_id": "sess_01SAFEWALK",
  "result_text": "Updated the safety page and ran pnpm docs:validate successfully.",
  "cost": { "estimated_cost_usd": null, "cost_source": "unavailable", "note": "unknown cost is not free" }
}

Result prose is a claim. Cost reporting must say known, unknown, or unreconciled; unknown cost is not free.

4. Transcript evidence

satellite runbook

hermes_transcript(session_id="sess_01SAFEWALK")
=> T2 transcript fetched
=> terminal rows include: pnpm docs:validate exit_code=0

5. Claim decomposition

json

[
  {"id":"req-1","kind":"user_requirement","text":"operator-grade access warning exists"},
  {"id":"tool-1","kind":"tool_execution","text":"pnpm docs:validate exited 0"},
  {"id":"assert-1","kind":"assistant_assertion","text":"The safety page was updated"}
]

6. Read-only oracle and confidence

Claim type	Oracle expectation	Evidence tier
Tool execution	Match transcript tool rows and exit codes.	T2
Assistant assertion	Check file contents or command output, not prose.	T2/T3
User requirement	Trace to original acceptance bullets.	T2/T3

7. Failed or partial behavior

json

{
  "decision": "PARTIAL",
  "confidence": "T0",
  "reason": "Only result prose was inspected.",
  "corrective_action": "Fetch transcript evidence or rerun verification."
}

A failed task is terminal. If there is no resumable session, submit a fresh task.

8. Corrective hermes_respond

markdown

hermes_respond(session_id="sess_01SAFEWALK", message="Verification found cost reporting missing. Add an explicit unknown/unreconciled cost sentence, rerun docs validation, and report command output.")