菜单
Spawn a pi subagent to handle a specific task in a separate process. Use when the user asks to delegate work to a subagent, run a task in parallel, or says "spawn a subagent", "use a subagent", or "run this with pi".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Autonomous developer iteration loop driven by a thin orchestrator that dispatches coder/reviewer/diagnoser subagents. Pick a mode — fix (make tests green), feature (implement one or more features), optimize (push a metric), debug (root-cause a failure), ab-test (race candidate implementations), or prototype (explore design variants) — define a verification signal, confirm the plan with the user, then loop implement → verify → review → commit with all state kept in loop-log.md. Use when the user states a concrete engineering goal or says "keep iterating", "loop on this", "autoloop", or "iterate until fixed".
Break a plan, spec, or PRD into independently-grabbable issues on the project issue tracker using tracer-bullet vertical slices.
Turn the current conversation into a PRD and publish it to the project issue tracker — no interview, just synthesis of what you've already discussed.
Move issues and external PRs through a state machine of triage roles — categorise, verify, grill if needed, and write agent-ready briefs.
Keeps the tswift project website in sync with runtime implementation changes — coverage numbers, feature tables, WASM binary, and content. Use when features are implemented, coverage changes, the runtime is updated, or the user asks to update/sync/refresh the website.
Use when completing tasks, implementing major features, or before merging to verify work meets requirements
| name | pi-subagent |
| description | Spawn a pi subagent to handle a specific task in a separate process. Use when the user asks to delegate work to a subagent, run a task in parallel, or says "spawn a subagent", "use a subagent", or "run this with pi". |
Spawn a separate pi process as a subagent, wait for it to finish, and read its output.
# 1. Write the prompt to a file
cat > /tmp/my-prompt.md << 'EOF'
<your task here>
When finished, write exactly `<<<DONE>>>` on its own line at the very end.
EOF
# 2. If unsure what model/provider to use, list available options first
pi --list-models
# 3. Spawn the subagent (cd into the project root so tools resolve correctly)
cd /path/to/project
pi --print \
--provider "anthropic" \
--model "claude-opus-4-5" \
--name "my-subagent" \
"$(cat /tmp/my-prompt.md)" \
> /tmp/my-output.md 2>&1 &
SUBAGENT_PID=$!
# 4. Poll for the termination token
for i in $(seq 1 80); do
sleep 15
if grep -q '<<<DONE>>>' /tmp/my-output.md 2>/dev/null; then
echo "✅ Done after $((i*15))s"
break
fi
echo "⏳ $((i*15))s — $(wc -l < /tmp/my-output.md) lines so far…"
done
# 5. Read the output
cat /tmp/my-output.md
| Flag | Purpose |
|---|---|
--print | Non-interactive: process prompt and exit |
--provider "<provider>" | Provider to use (e.g. anthropic, openai-codex) |
--model "<model>" | Model to use |
--name "…" | Names the subagent session |
--cwd — pi has no --cwd flag. Use cd /path before the command.--no-tools) so the subagent can read files and run git/bash.<<<DONE>>> as its last line so you can poll without guessing when it's finished."$(cat /tmp/prompt.md)" to pass it; avoids shell quoting issues.cd into the project root first so relative paths in tools resolve correctly.# Anthropic (Claude)
pi --provider "anthropic" --model "claude-opus-4-5" …
pi --provider "anthropic" --model "claude-sonnet-4-5" …
# OpenAI Codex (GPT)
pi --provider "openai-codex" --model "gpt-5.5" …
# Quick sanity check
echo "say hi" | pi --provider "anthropic" --model "claude-sonnet-4-5" --no-session --print