Purpose
Turn a rough proposal into a bounded plan whose assumptions, risks, authority limits, and verification gates are visible before implementation begins.
When it fires
- The user asks to review, harden, or stress-test a plan.
- The work crosses product, data, security, design, runtime, or deployment boundaries.
- A wrong decision would be expensive or difficult to reverse.
Do not invoke the full review for a trivial, reversible edit unless the user asks for it.
Inputs
- Desired user outcome and explicit non-goals.
- Current repository or system evidence.
- Time, cost, compatibility, privacy, and authority constraints.
- Known failure history and required verification commands.
Workflow
- Restate the outcome, success criteria, non-goals, and unresolved assumptions.
- Select review depth proportional to reversibility and impact.
- Run four lenses: product value, engineering integrity, design quality, and developer/operator experience.
- Give every finding an evidence reference, confidence level, and consequence.
- Reconcile contradictions; never silently average incompatible recommendations.
- Produce the smallest coherent execution sequence with rollback and promotion gates.
- Stop for human direction before destructive, costly, public, or credential-bearing actions.
When the same agent performs every lens, disclose that the review was sequential rather than independent.
Output contract
Return:
decision: build, revise, defer, or stop.scope: user-visible outcome, non-goals, and affected surfaces.assumptions: verified, inferred, and unresolved items kept separate.reviews: findings from each lens with severity and evidence.execution_plan: ordered, independently verifiable steps.verification: automated, manual, security, accessibility, and deployment gates as applicable.artifacts: plan, decision, rollback, and verification records with stable references.failure_conditions: conditions that force revise, defer, or stop.verification_receipts: one receipt per applicable gate withgate,result,evidence, andowner.rollback: safe recovery path and ownership.human_gates: decisions the agent may not make.
Return decision: build only when every gate applicable before implementation has a passing verification receipt. Record later-stage gates as pending; when their decision point arrives, a failed or missing applicable receipt blocks promotion and requires revise, defer, or stop.
Tools & MCP
Use repository search, tests, type/lint gates, preview deployments, and an approved browser connector when relevant. Do not enable connectors, spend money, publish, merge, or mutate external systems merely because a plan mentions them.
Quality bar
- No claim that AI makes execution free or removes operational ownership.
- No speculative expansion without a stated user outcome.
- Evidence and inference are distinguishable.
- Every high-risk step has a verifier and rollback.
- The plan can be handed to another operator without hidden context.
Example
Input: “Add a self-service agent marketplace with portable plugins.”
Good output: a revise decision that separates catalog, package, installer, evaluation, and promotion contracts; identifies marketplace approval and credential handling as human gates; and gives exact API, schema, preview, security, and rollback checks.
Built on SIP — Starlight Intelligence Protocol Vertical: starlight-agent-skills · portable capability layer