Fumadocs on TanStack Start

Agent output is not complete when the runtime says it is done. It is complete only when required checks pass, review gates are satisfied, and approved artifacts are promoted.

Principle

Agent output is not complete when the runtime says it is done. It is complete only when required checks pass, review gates are satisfied, and approved artifacts are promoted.

Lifecycle

Runtime result
  -> collect artifacts/diff/logs
  -> run automated checks
  -> perform spec compliance review
  -> perform code/docs quality review
  -> perform sandbox/security review
  -> human approval decision
  -> promote selected changes
  -> write audit event

Verification result draft

schema_version: uh.verification-result.v0
mission_id: mission-2026-05-13-docs-spine
status: passed # passed | failed | blocked | waived
checks:
  - name: docs-tree-exists
    type: command
    command: find docs -type f | sort
    status: passed
  - name: spec-compliance
    type: review
    reviewer: qa-agent
    status: passed
    notes: All required docs exist and map to handoff criteria.
findings:
  - severity: info
    message: AgentFS backend remains interface-level for MVP.
approvals:
  - gate: human-promotion
    status: pending

Review gates

Spec compliance — does output satisfy the mission and linked issue/spec?
Quality — is the implementation/docs clear, maintainable, and minimal?
Verification — did required checks run and pass?
Security/sandbox — did the runtime stay within permitted boundaries?
Human promotion — should changes become canonical?

Promotion record draft

schema_version: uh.promotion.v0
mission_id: mission-2026-05-13-docs-spine
sandbox_id: sandbox-abc123
decision: promoted
approved_by: Lalo
promoted_at: 2026-05-13T00:00:00Z
changes:
  - docs/README.md
  - docs/glossary.md
audit_event_id: audit-...

Promotion policies

human-approved — default. Requires explicit human approval.
review-agent-approved — allowed only for low-risk generated docs or tests when configured.
auto-promote-on-green — deferred; too risky for MVP.
manual-only — harness records evidence but a human applies changes.

Verification and Promotion Lifecycle