AgentSkillsCN

loop-review-skill-until-fixed-point

对目标技能进行迭代式审查,直至达到“固定点”。在所有评审者均给出“无问题”的评价之前,持续进行多次审查与反馈循环。

SKILL.md
--- frontmatter
name: loop-review-skill-until-fixed-point
description: Iterate /review-skill on a target until fixed point. Runs review passes until all reviewers return NO ISSUES.

Loop Review Skill Until Fixed Point

Run /review-skill on a target document repeatedly until fixed point—when all reviewers return NO ISSUES.

Core Concept

code
/review-skill <target> --auto
         │
         ▼
    ┌─────────┐
    │ Issues? │
    └────┬────┘
         │
    yes ─┴─ no
     │      │
     ▼      ▼
  repeat   done (fixed point)

Fixed point = the document is both correct AND unambiguous. No reviewer can find anything to flag.


Arguments

ArgRequiredDescription
<target>yesPath to SKILL.md to review

State

yaml
max_iterations: 10        # Safety limit
iteration_count: 0        # Current iteration
target: "<from args>"     # Target SKILL.md path
history: []               # Minimal: just iteration outcomes

History Entry Schema

Each iteration appends a minimal entry:

yaml
- iteration: 1
  fixed_point: false
- iteration: 2
  fixed_point: false
- iteration: 3
  fixed_point: true   # All reviewers returned NO ISSUES

The orchestrator does NOT track per-reviewer metrics. That detail stays inside /review-skill where it belongs. This prevents context leak between iterations.


Phase: Initialize

  1. Parse target path from arguments
  2. Set iteration_count = 0
  3. Set max_iterations = 10
  4. Confirm target exists

Phase: Loop

Isolation requirement: Each iteration must be context-isolated. The orchestrator must NOT carry synthesis, triage details, or "recurring theme" narratives between iterations. This prevents context leak that would bias subsequent reviews.

code
while iteration_count < max_iterations:
    iteration_count += 1

    1. Run: /review-skill <target> --auto
       (This is a FRESH invocation—no prior iteration context)
    2. Capture only the exit status: "fixed point" or "issues remain"
    3. Append to history file: { iteration, fixed_point: bool }
    4. Output: "Iteration {N}: {'fixed point' | 'issues remain'}"
    5. If fixed point → exit loop
    6. Else → continue loop (do NOT summarize or analyze findings here)

The orchestrator is intentionally stateless. All synthesis, triage, and issue-addressing happens INSIDE /review-skill --auto. The orchestrator only observes the outcome.

Exit Conditions

ConditionAction
/review-skill reports fixed pointExit with success
iteration_count >= max_iterationsSafety limit hit, ask user how to proceed

Phase: Report

Present final state:

markdown
## Loop Complete

| Metric | Value |
|--------|-------|
| Target | {target} |
| Iterations | {iteration_count} |
| Fixed point reached | yes/no |

If fixed point reached:

{target} reached fixed point after {N} iterations.

If max iterations hit:

Safety limit reached after {max_iterations} iterations without convergence.


Begin now.