review-loop
bởi microsoft
Vòng lặp đánh giá lặp đi lặp lại sử dụng bất kỳ tác nhân thấu kính nào chống lại bất kỳ mục tiêu nào (kế hoạch, mã, thiết kế). Sử dụng khi bạn muốn chạy một chu kỳ đánh giá tập trung: tạo ra một người đánh giá mới…
npx skills add https://github.com/microsoft/trieste --skill review-loopReview Loop
Use this skill to run an iterative review cycle against any target using any lens agent.
Inputs
The user specifies:
- Lens — which lens/reviewer to use (e.g.,
adversarial-lens,security-lens,conservative-lens,speed-lens,usability-lens). - Target — what to review (a plan, file(s), module, design document, diff, etc.).
If either is ambiguous, ask the user before proceeding.
Procedure
Step 1 — Gather Context
Read the target material and any surrounding context needed for a meaningful review. This includes:
- The target itself (plan text, file contents, etc.)
- Relevant WF specs, token definitions, and pass pipeline context
- Relevant tests or prior review findings
Step 2 — Spawn Reviewer Subagent
Spawn a fresh subagent with the specified lens. Provide it:
- The full target material (verbatim — do not summarize)
- Sufficient surrounding context for the reviewer to assess the target meaningfully
- Instructions to produce structured findings per the lens's Expected Output format
- Instructions to explicitly state if the target is clean (no findings)
The subagent must be a fresh instance with no prior context from this conversation.
Step 3 — Present Results
Present the reviewer's findings to the user along with a proposed remediation plan:
- List each finding with its severity/likelihood
- For each finding, propose a specific remediation (code change, plan amendment, verification step, etc.)
- If the reviewer found no issues, report that the target passed review cleanly
Step 4 — Apply Remediations
With the user's guidance:
- Apply agreed-upon remediations (edit files, amend plans, add tests, etc.)
- Skip or defer any remediations the user rejects
- If the user modifies a remediation, apply their version
Step 5 — Loop or Exit
Ask the user whether to:
- Loop — go back to Step 2 and run the review again on the updated target with a fresh subagent
- Exit — the review cycle is complete
Guardrails
- Every review iteration uses a fresh subagent — no context contamination between iterations.
- Do not summarize or filter the target material before passing it to the reviewer. Provide it verbatim.
- Do not apply remediations without user confirmation.
- The loop exits only when the user says so, not when the reviewer returns clean. The user may want to run additional iterations even after a clean pass.