Menu
Adversarial verify-subagent dispatch protocol. Use after any plan's final step to get VERDICT before declaring completion.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Evidence-gated completion check. Use before committing, creating PRs, or claiming work is done. Pairs with verification-spec at task start.
Systematic knowledge extraction from open-source projects. Use when: user shares a repo/link to study, says 偷师/steal/学习/研究, or wants to analyze another project's patterns for adoption.
Launch a real interactive claude CLI session in a target directory (typically a worktree) with a pre-seeded prompt. Use when the current session cannot absorb the task context — e.g., N parallel worktree jobs that would blow the context window, or when you need the user to watch progress in a visible terminal. Each spawned session runs independently in its own Windows Terminal tab.
FileNotFound recovery protocol. Before reporting a missing-file error, suggest near-misses by basename similarity (difflib threshold 0.4) so typos and stale paths fail loudly with actionable hints.
Paste-by-reference guard for Edit calls. When new_string would echo >30 lines of existing file content, use {{file:path:start:end}} reference instead and expand before submission.
File-IO protocol for parent ↔ subagent communication. Parent writes input.txt + context.json, subagent appends to output.txt with [ROUND END] sentinels. Used with Monitor tool for live observation.
| name | verify-subagent |
| description | Adversarial verify-subagent dispatch protocol. Use after any plan's final step to get VERDICT before declaring completion. |
You are the verify-subagent dispatcher. Your job is to dispatch an adversarial subagent after every plan's final step, collect its VERDICT, and handle PASS/FAIL/PARTIAL branches. You do NOT declare the plan complete — the VERDICT does.
After the plan's final implementation step, read verify_sop.md content and dispatch:
Agent(
subagent_type="claude-sonnet",
system_prompt=<contents of .claude/skills/verify-subagent/verify_sop.md>,
prompt="验证以下产物:\n\n{deliverable description}\n\n产物路径:\n{list of files created/modified}\n\n验证上下文:\n{relevant plan step verify commands}"
)
The deliverable description must include:
Wait for the subagent response. Scan for the literal string VERDICT: followed by PASS, FAIL,
or PARTIAL. Do not interpret partial matches — only the exact literals count.
| VERDICT | Action |
|---|---|
VERDICT: PASS | Declare plan complete. Reference the VERDICT as evidence. |
VERDICT: FAIL | Enter fix loop: fix the specific failure listed, re-dispatch verify-subagent. Max 2 iterations. |
VERDICT: PARTIAL | Treat as FAIL unless failures explicitly do not affect the plan's goal. Justify in writing if accepting PARTIAL. |
Iteration 1: Fix → re-dispatch → check VERDICT
Iteration 2: Fix → re-dispatch → check VERDICT
If still FAIL after 2 iterations: STOP, report to owner with full failure details.
Never declare done after a failed fix without re-running verify-subagent.
When reporting dispatch results, include:
[verify-subagent] Dispatched for: {deliverable}
[verify-subagent] VERDICT: {PASS|FAIL|PARTIAL}
[verify-subagent] Details: {summary of what was checked}
If PASS: proceed to completion declaration with the VERDICT as evidence. If FAIL/PARTIAL after max iterations: report all details to owner, list what passed and what failed.
VERDICT: line as a passing verification.constraints/verdict-required.md for full scope).