loop-review-skill-until-fixed-point
npx skills add https://github.com/corygabrielsen/skills --skill loop-review-skill-until-fixed-point
Agent 安装分布
Skill 文档
Loop Review Skill Until Fixed Point
Run /review-skill on a target document repeatedly until fixed pointâwhen all reviewers return NO ISSUES.
Core Concept
/review-skill <target> --auto
â
â¼
âââââââââââ
â Issues? â
ââââââ¬âââââ
â
yes ââ´â no
â â
â¼ â¼
repeat done (fixed point)
Fixed point = the document is both correct AND unambiguous. No reviewer can find anything to flag.
Arguments
| Arg | Required | Description |
|---|---|---|
<target> |
yes | Path to SKILL.md to review |
State
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:
- 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
- Parse target path from arguments
- Set
iteration_count = 0 - Set
max_iterations = 10 - 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.
while iteration_count < max_iterations:
iteration_count += 1
1. Run: /review-skill <target> --auto
(This is a FRESH invocationâno 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 â exit loop
6. Else â 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
| Condition | Action |
|---|---|
/review-skill reports fixed point |
Exit with success |
iteration_count >= max_iterations |
Safety limit hit, ask user how to proceed |
Phase: Report
Present final state:
## 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.