Double Check

Double-Check Skill

Execute this workflow whenever the skill is invoked. You (the agent executing this skill) ARE the main model. All script paths below are relative to this skill's directory (the directory containing this SKILL.md file) — resolve them to absolute paths before running, e.g. cd "$(dirname SKILL.md)" or use <skill-dir>/scripts/....

0. Identify the main model

• Read $PI_MODEL (e.g. deepseek-v4-flash) and $PI_PROVIDER (e.g. deepseek) from the environment.
• If $PI_MODEL is unset, ask the user or infer from pi --list-models / current session. Record it for the final report.

1. Complete the task with the main model

• Do the task normally with your full tools and best effort, exactly as you would without this skill.
• Write your complete, self-contained answer to a temp file answer.txt (use mktemp and keep the path). For code/file tasks, the answer must include the key code, diffs, file paths, and a summary of what was done, so the checker can actually review it.
• Write the user's original task (verbatim) to a temp file task.txt.

2. Pick a check model (must be different from the main model)

• Run: <skill-dir>/scripts/pick-check-model.sh "$PI_MODEL"
• It prints the chosen model as provider/model on stdout (e.g. minimax-cn/MiniMax-M3).
• Selection rules: MUST differ from the main model → prefer a different provider → prefer the provider's latest model (MiniMax → MiniMax-M3, DeepSeek → deepseek-v4-flash) → else any other model.
• If it exits non-zero (no different model available), tell the user transparently that no different model is configured, deliver your answer anyway with a warning, and stop. Never fake a check.

3. Check the answer with the check model

• Run: <skill-dir>/scripts/run-check.sh <task.txt> <answer.txt> <check-model> (optional 4th arg: thinking level, default low — the checker models are reasoning models and review more reliably with thinking on; use high/medium for the strictest review, off for the fastest).
• It prints a single-line JSON verdict: {"verdict":"pass"|"fail","score":0-100,"issues":[...],"suggestions":[...],"explanation":"..."}
• Save the verdict to round-<n>.json for the final report.
• If the command fails or hangs (use a bash timeout of ~300 s), retry once; if it still fails, note the check failure and continue with the best answer you have.

4. Loop: revise and re-check (max 3 rounds)

round = 1
loop:
  verdict = check(answer, round)          # step 3
  if verdict == "pass": break             # step 3 of the requirement -> success
  if round >= 3: break                    # forced exit after 3 failed rounds
  # step 4 of the requirement: feed the check result back to the main model (= you)
  revise answer.txt as the main model, addressing every issue/suggestion from the verdict
  (checker feedback is guidance, not final truth — use your judgment and your tools)
  round += 1
  go to loop

• On a fail, you MUST visibly revise answer.txt (re-read the issues/suggestions and rework the answer), then re-run step 3. Do not just resubmit the same answer.
• After the 3rd failed round, force exit the loop and keep the best revised answer.

5. Report

Compose the final answer to the user:

1. Final answer — the (possibly revised) best answer.
2. Double-Check Report (short, at the end):
   • Main model: <provider>/<model> (from $PI_PROVIDER / $PI_MODEL)
   • Check model: <provider>/<model> (the one actually used)
   • Rounds: N / 3, with per-round summary: round → pass/fail → top issues (if any)
   • Status: "✅ passed check" / "⚠️ forced exit after 3 failed rounds — unverified, remaining issues: …"

Be transparent: never claim a check passed when it did not. If the loop was forced out, clearly mark the result as unverified and list the remaining issues.

Example

SKILL_DIR=<absolute path to this skill directory>
task=$(mktemp -t dc-task.XXXXXX); answer=$(mktemp -t dc-answer.XXXXXX)
# (write the user's task into $task, your answer into $answer)
CHECK_MODEL=$("$SKILL_DIR/scripts/pick-check-model.sh" "$PI_MODEL")
"$SKILL_DIR/scripts/run-check.sh" "$task" "$answer" "$CHECK_MODEL"