Strict engineering discipline and agentic orchestration for the pi coding agent. Multi-agent workflows, durable goals, and verifier-gated delivery. pi μ½λ© μμ΄μ νΈλ₯Ό μν μ격ν μμ§λμ΄λ§ λμ¬μ΄νλ¦°κ³Ό μμ΄μ νΈ μ€μΌμ€νΈλ μ΄μ . λ€μ€ μμ΄μ νΈ μν¬νλ‘, μ§μν λͺ©ν, κ²μ¦ κΈ°λ° μ λ¬.
____ ___ _ ____ _ _ ____ ___
| _ \ / _ \ / \ / ___| | | | | _ \_ _|
| |_) | | | |/ _ \| | | |_| | | |_) | |
| _ <| |_| / ___ \ |___| _ | | __/| |
|_| \_\\___/_/ \_\____|_| |_| |_| |___|
ROACH PI
Durable Goals | Multi-Agent | Verifier-Gated
/setup /clarify /goal /loop
Every line is on GitHub. No hidden prompts, no secret system instructions, no obfuscated behavior. Tools, event hooks, skill injections, and agent prompts are all plain TypeScript and Markdown.
λͺ¨λ μ½λκ° GitHubμ 곡κ°λμ΄ μμ΅λλ€. μ¨κ²¨μ§ ν둬ννΈ, λΉλ° μμ€ν μ§μμ΄, λλ νλ λμ μμ. λꡬ, μ΄λ²€νΈ ν , μ€ν¬ μ£Όμ , μμ΄μ νΈ ν둬ννΈκ° λͺ¨λ μΌλ° TypeScriptμ MarkdownμΌλ‘ μμ±λμ΄ μμ΅λλ€.
The footer displays real-time metrics every session: prompt cache hit rate, context usage bar, active tools, branch, and model. See exactly how your context is being utilized.
νΈν°μ λ§€ μΈμ μ€μκ° λ©νΈλ¦μ΄ νμλ©λλ€: ν둬ννΈ μΊμ μ μ€λ₯ , 컨ν μ€νΈ μ¬μ©λ λ°, νμ± λꡬ, λΈλμΉ, λͺ¨λΈ. 컨ν μ€νΈκ° μ΄λ»κ² μ¬μ©λλμ§ μ νν νμΈνμΈμ.
No magic. Read the source and see exactly what the agent does. Skill rules are plain Markdown. Agent prompts are plain text. Every event hook is auditable.
λ§λ² μμ. μμ€λ₯Ό μ½κ³ μμ΄μ νΈκ° μ νν 무μμ νλμ§ νμΈνμΈμ. μ€ν¬ κ·μΉμ μΌλ° Markdown, μμ΄μ νΈ ν둬ννΈλ μΌλ° ν μ€νΈ, λͺ¨λ μ΄λ²€νΈ ν μ΄ κ°μ¬ κ°λ₯ν©λλ€.
We prefer straightforward solutions over elaborate orchestration. Fewer moving parts, fewer failure modes. See our contributing guidelines.
볡μ‘ν μ€μΌμ€νΈλ μ΄μ λ³΄λ€ λ¨μν ν΄κ²°μ± μ μ νΈν©λλ€. μμ§μ΄λ λΆλΆμ΄ μ μμλ‘ μ€ν¨λ μ μ΅λλ€. κΈ°μ¬ κ°μ΄λλΌμΈμ μ°Έμ‘°νμΈμ.
The extension provides a streamlined workflow: clarify β goal. /goal creates, activates, continues, and verifies automatically.
μ΅μ€ν
μ
μ λ¨μν μν¬νλ‘λ₯Ό μ 곡ν©λλ€: λͺ
νν β goal. /goalμ΄ μλμΌλ‘ μμ±, νμ±ν, μ΄μ΄κ°κΈ°, κ²μ¦μ μ§νν©λλ€.
Live-style runtime preview showing command execution order, verifier gates, and evidence checkpoints.
Live-style runtime preview showing command execution order, verifier gates, and evidence checkpoints.
00:05 /setup -> bootstrap complete. 00:28 /clarify -> goal contract locked. 00:53 /goal -> runtime active. 01:24 /goal subgoal -> queue accepted. 02:15 /loop -> verifier PASS pending.
00:05 /setup -> bootstrap complete. 00:28 /clarify -> goal contract locked. 00:53 /goal -> runtime active. 01:24 /goal subgoal -> queue accepted. 02:15 /loop -> verifier PASS pending.
/setup
β Configure Settings
β μ€μ ꡬμ±
Run once after installation to apply recommended settings. Sets quietStartup: true so the extension's custom ROACH PI banner replaces the default startup listing.
μ€μΉ ν ν λ² μ€ννμ¬ κΆμ₯ μ€μ μ μ μ©ν©λλ€. μ΅μ€ν
μ
μ 컀μ€ν
ROACH PI λ°°λκ° κΈ°λ³Έ μμ λͺ©λ‘μ λ체νλλ‘ quietStartup: trueλ₯Ό μ€μ ν©λλ€.
/clarify
β Resolve Ambiguity
β λͺ¨νΈμ± ν΄κ²°
The agent asks dynamic, context-aware questions one at a time, using codebase exploration only when implementation or uncertain technical context makes it necessary. Ends with a structured Goal Contract that defines scope, constraints, success criteria, and evidence requirements.
μμ΄μ νΈκ° λμ μ΄κ³ λ¬Έλ§₯μ λ§λ μ§λ¬Έμ νλμ© λ¬»κ³ , ꡬνμ΄λ λΆνμ€ν κΈ°μ λ§₯λ½ λλ¬Έμ νμν λλ§ μ½λλ² μ΄μ€λ₯Ό νμν©λλ€. λ²μ, μ μ½, μ±κ³΅ κΈ°μ€, μ¦κ±° μꡬμ¬νμ μ μνλ ꡬ쑰νλ Goal Contractλ‘ λλ©λλ€.
/goal
β Activate Durable Execution
β durable μ€ν νμ±ν
Starts or continues a durable runtime that tracks goals, subgoals, todos, evidence, blockers, verifier receipts, and automatic continuation. Completion requires an independent verifier PASS.
λͺ©ν, νμ λͺ©ν, todo, μ¦κ±°, μ°¨λ¨ μμ, κ²μ¦ μμμ¦, μλ μ΄μ΄κ°κΈ°λ₯Ό μΆμ νλ durable runtimeμ μμνκ±°λ μ΄μ΄κ°λλ€. μλ£μλ λ 립 verifier PASSκ° νμν©λλ€.
/goal subgoal
β Subgoal Queueing
β νμ λͺ©ν νμ
For complex multi-step tasks. Queue subgoals with dependencies and required evidence, then let the goal runtime advance after verifier-guarded completion.
볡μ‘ν λ€λ¨κ³ μμ μ μν΄. μμ‘΄μ±κ³Ό νμν μ¦κ±°κ° μλ νμ λͺ©νλ₯Ό νμ λ£κ³ , verifier-guarded μλ£ ν goal runtimeμ΄ μ§ννκ² ν©λλ€.
/loop
β Session Loop Scheduler
β μΈμ
루ν μ€μΌμ€λ¬
Schedule recurring prompts at fixed intervals. Cron-style timing β fires on schedule regardless of execution state. Supports 5s, 10m, 2h, 1d formats. Up to 100 concurrent jobs with per-job error isolation and automatic cleanup on session shutdown.
κ³ μ κ°κ²©μΌλ‘ λ°λ³΅ ν둬ννΈλ₯Ό μ€μΌμ€λ§ν©λλ€. Cron μ€νμΌ νμ΄λ° β μ€ν μνμ κ΄κ³μμ΄ μ€μΌμ€λλ‘ μ€νλ©λλ€. 5s, 10m, 2h, 1d νμμ μ§μν©λλ€. μ΅λ 100κ° λμ μμ
, μμ
λ³ μλ¬ κ²©λ¦¬, μΈμ
μ’
λ£ μ μλ μ 리.
π‘ Tip: The ask_user_question tool is always available β the agent will ask clarifying questions autonomously whenever it detects ambiguity, even outside /clarify mode.
π‘ ν: ask_user_question λꡬλ νμ μ¬μ© κ°λ₯ν©λλ€ β /clarify λͺ¨λ μΈμμλ λͺ¨νΈμ±μ κ°μ§νλ©΄ μλμΌλ‘ μ§λ¬Έν©λλ€.
18+ bundled agents for exploration, execution, planning, and review. Delegate tasks via single, parallel (max 12 tasks, 10 concurrent), or chain modes with cycle detection and depth guards.
νμ, μ€ν, κ³ν, 리뷰λ₯Ό μν 18κ° μ΄μμ λ²λ€ μμ΄μ νΈ. μ¬μ΄ν΄ κ°μ§ λ° κΉμ΄ κ°λμ ν¨κ» λ¨μΌ, λ³λ ¬(μ΅λ 12κ° μμ , 10κ° λμ), μ²΄μΈ λͺ¨λλ‘ μμ μ μμνμΈμ.
Karpathy rules are auto-injected into code-writing agents so implementation stays surgical, evidence-based, and scoped to the active goal.
Karpathy κ·μΉμ΄ μ½λ μμ± μμ΄μ νΈμ μλ μ£Όμ λμ΄ κ΅¬νμ΄ νμ± λͺ©ν λ²μ μμμ μΈκ³Όμ μ΄κ³ μ¦κ±° κΈ°λ°μΌλ‘ μ μ§λ©λλ€.
/goal tracks active subgoals, evidence, blockers, and verifier receipts. Completion only lands after reviewer-verifier independently returns PASS.
/goalμ νμ± νμ λͺ©ν, μ¦κ±°, μ°¨λ¨ μμ, κ²μ¦ μμμ¦μ μΆμ ν©λλ€. μλ£λ reviewer-verifierκ° λ 립μ μΌλ‘ PASSλ₯Ό λ°νν λ€μλ§ νμ λ©λλ€.
Custom ROACH PI ASCII banner, branded footer with directory, branch, model, context bar, cache hit rate, and active tools display.
λλ ν 리, λΈλμΉ, λͺ¨λΈ, 컨ν μ€νΈ λ°, μΊμ μ μ€λ₯ , νμ± λꡬ νμκ° μλ 컀μ€ν ROACH PI ASCII λ°°λ λ° λΈλλλ νΈν°.
Microcompaction truncates old tool results. Phase-aware summarization preserves workflow state across compaction. Phase and goal document survive context limits.
λ§μ΄ν¬λ‘μ»΄ν©μ μ΄ μ€λλ λꡬ κ²°κ³Όλ₯Ό μλ¦ λλ€. λ¨κ³ μΈμ μμ½μ΄ μ»΄ν©μ μ λ°μ κ±Έμ³ μν¬νλ‘ μνλ₯Ό 보쑴ν©λλ€. λ¨κ³ λ° λͺ©ν λ¬Έμκ° μ»¨ν μ€νΈ νκ³λ₯Ό μ΄κ³Όν©λλ€.
Schedule recurring prompts with /loop 5m check status. Cron-style fixed intervals, per-job error isolation, AbortController cancellation, timeout protection, and automatic cleanup on session shutdown. Up to 100 concurrent jobs.
/loop 5m check statusλ‘ λ°λ³΅ ν둬ννΈλ₯Ό μ€μΌμ€λ§ν©λλ€. Cron μ€νμΌ κ³ μ κ°κ²©, μμ
λ³ μλ¬ κ²©λ¦¬, AbortController μ·¨μ, νμμμ 보νΈ, μΈμ
μ’
λ£ μ μλ μ 리. μ΅λ 100κ° λμ μμ
.
Bundled LLM skill rulesets: clarification, durable goal execution, verifier-guarded review, simplify, systematic debugging, Karpathy discipline, Rob Pike optimization, and more.
λ²λ€ LLM μ€ν¬ λ£°μ : λͺ νν, durable goal μ€ν, verifier-guarded 리뷰, λ¨μν, 체κ³μ λλ²κΉ , Karpathy λμ¬μ΄νλ¦°, Rob Pike μ΅μ ν λ±.
Built-in search acceleration for find and grep through a native FFF engine. Adds multi_grep, frecency-aware ranking, git-aware indexing, optional replacement of @ file autocomplete, and built-in fallback to standard find/grep behavior when the native layer is unavailable.
λ€μ΄ν°λΈ FFF μμ§μ ν΅ν find / grep λ΄μ₯ κ²μ κ°μ. multi_grep, frecency κΈ°λ° λνΉ, git μΈμ μΈλ±μ±, μ νμ @ νμΌ μλμμ±μ μΆκ°νλ©°, λ€μ΄ν°λΈ λ μ΄μ΄λ₯Ό μΈ μ μμ λλ κΈ°λ³Έ find/grep λμμΌλ‘ ν΄λ°±ν©λλ€.
Deep-dive docs for every feature: Workspace Memory, Compaction, Team Mode, Code Review, Sandboxed Bash, Worktree Isolation, and Power Tools.
λͺ¨λ κΈ°λ₯μ λν μ¬μΈ΅ λ¬Έμ: μν¬μ€νμ΄μ€ λ©λͺ¨λ¦¬, μ»΄ν©μ , ν λͺ¨λ, μ½λ 리뷰, μλλ°μ€ Bash, μν¬νΈλ¦¬ 격리, νμ ν΄.
pi is a terminal-based AI coding agent that helps you read, write, and debug code directly from your terminal. It supports extensions, custom tools, themes, and multiple AI providers. GitHub β
piλ ν°λ―Έλμμ μ§μ μ½λλ₯Ό μ½κ³ , μ°κ³ , λλ²κΉ νλ ν°λ―Έλ κΈ°λ° AI μ½λ© μμ΄μ νΈμ λλ€. μ΅μ€ν μ , 컀μ€ν λꡬ, ν λ§ λ° μ¬λ¬ AI νλ‘λ°μ΄λλ₯Ό μ§μν©λλ€. GitHub β
No. All engineering discipline skills are bundled with this extension. Just install the extension and run /setup β everything is ready out of the box.
μλμ. λͺ¨λ μμ§λμ΄λ§ λμ¬μ΄νλ¦° μ€ν¬μ΄ μ΄ μ΅μ€ν
μ
μ λ²λ€λ‘ ν¬ν¨λμ΄ μμ΅λλ€. μ΅μ€ν
μ
μ μ€μΉνκ³ /setupμ μ€ννμΈμ β λͺ¨λ κ²μ΄ μ¦μ μ¬μ© κ°λ₯ν©λλ€.
Single: One-off tasks sent to a specific agent.
Parallel: Dispatch multiple agents at once (max 12 tasks, 10 concurrent). Each runs independently.
Chain: Sequential pipeline where each step receives the previous step's output via {previous} placeholder.
All modes include cycle detection, depth limits (max 3), and concurrency control.
λ¨μΌ: νΉμ μμ΄μ νΈμ 보λ΄λ μΌνμ± μμ
.
λ³λ ¬: μ¬λ¬ μμ΄μ νΈλ₯Ό ν λ²μ λμ€ν¨μΉ(μ΅λ 12κ° μμ
, 10κ° λμ). κ°κ° λ
립μ μΌλ‘ μ€ν.
체μΈ: κ° λ¨κ³κ° μ΄μ λ¨κ³μ μΆλ ₯μ {previous} νλ μ΄μ€νλλ‘ λ°λ μμ°¨ νμ΄νλΌμΈ.
λͺ¨λ λͺ¨λμ μ¬μ΄ν΄ κ°μ§, κΉμ΄ μ ν(μ΅λ 3), λμμ± μ μ΄κ° ν¬ν¨λ©λλ€.
The harness injects focused implementation guardrails into code-writing agents: read before writing, keep changes scoped, verify assumptions, and record evidence before completion.
νλ€μ€λ μ½λ μμ± μμ΄μ νΈμ μ§μ€λ ꡬν κ°λλ μΌμ μ£Όμ ν©λλ€. μ°κΈ° μ μ μ½κ³ , λ³κ²½ λ²μλ₯Ό μ§ν€λ©°, κ°μ μ κ²μ¦νκ³ , μλ£ μ μ μ¦κ±°λ₯Ό κΈ°λ‘ν©λλ€.
It overrides the built-in find and grep tools with a native FFF-backed implementation and adds multi_grep. In both mode it also replaces @ file autocomplete suggestions. Use /fff-health to inspect engine status and /fff-mode tools-only if you only want the tool overrides.
κΈ°λ³Έ find / grep λꡬλ₯Ό λ€μ΄ν°λΈ FFF κΈ°λ° κ΅¬νμΌλ‘ κ΅μ²΄νκ³ multi_grepλ₯Ό μΆκ°ν©λλ€. both λͺ¨λμμλ @ νμΌ μλμμ±λ λ체ν©λλ€. μμ§ μνλ /fff-healthλ‘ νμΈνκ³ , λꡬ overrideλ§ μνλ©΄ /fff-mode tools-onlyλ₯Ό μ¬μ©νμΈμ.
Use /clarify to produce a Goal Contract, then /goal to create the durable runtime entry. Add subgoals when the work needs decomposition.
/clarifyλ‘ Goal Contractλ₯Ό λ§λ λ€μ /goalλ‘ durable runtime νλͺ©μ λ§λμΈμ. μμ
λΆν΄κ° νμνλ©΄ νμ λͺ©νλ₯Ό μΆκ°νμΈμ.