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.
A direct command skips the review prompt. Inspect the source before running it.
Pi terminal agent workspace — install, /commands, sessions, compaction, settings.json, AGENTS.md, skill discovery, precedence, session JSONL, /tree, providers/auth, models.json, credential resolution, platform setup, editor (@file, Tab, images, !bash), message queue (steering, follow-up), built-in tools (read/bash/edit/write/grep/find/ls, --tools, --no-tools), print mode (-p, piped stdin), SYSTEM.md/APPEND_SYSTEM.md, CLI flags (-c, -r, --session, --fork, --no-session), @files args, resource flags (--no-extensions, --no-skills), model shorthand (provider/id, name:thinking), all slash commands, env vars, philosophy (no MCP/sub-agents/plan-mode by design). Use for pi config, skills, sessions, compaction, providers, models, tmux, Termux, Windows, tools, modes, system prompt, message queue, or "does pi support X".
compatibility
Host agent with read access to workspace files and the pi-mono corpus; Pi includes available_skills in the system prompt only when the read tool is among selected tools (see pi-mono/packages/coding-agent/src/core/system-prompt.ts).
Pi CLI and workspace
Answer only from pi-mono/ sources listed below. If something is not in the tree, say you cannot confirm from the corpus.
pi-mono/packages/coding-agent/README.md §Philosophy — what core intentionally omits (no MCP, no sub-agents, no permission popups, no plan mode, no built-in todos, no background bash); everything buildable via extensions.
Invariants
Skill name collisions: first registered name wins; later paths emit collision diagnostics (pi-mono/packages/coding-agent/src/core/skills.ts). Path order is assembled in resource-loader.ts then fed to loadSkills.
Resource precedence rank (lower = earlier in sorted package-manager lists): project local (0), project auto (1), user local (2), user auto (3), package (4) — pi-mono/packages/coding-agent/src/core/package-manager.ts.
Merged skill paths: mergePaths([...cliEnabledSkills, ...enabledSkills], additionalSkillPaths) — CLI paths appear before package-manager lists; see pi-mono/packages/coding-agent/src/core/resource-loader.ts.
<available_skills> is appended to the system prompt only when the read tool is among selected tools — pi-mono/packages/coding-agent/src/core/system-prompt.ts.
Default system prompt already points models at packaged docs paths (readmePath, docsPath, examples) — pi-mono/packages/coding-agent/src/core/system-prompt.ts. This skill adds workspace and precedence detail, not a duplicate README.
Session format version is v3 (tree-based id/parentId); buildSessionContext assembles compaction summary + branch summary + messages after the compaction point — see pi-mono/packages/coding-agent/docs/session.md.
Credential resolution order: --api-key flag > auth.json > environment variables > models.json custom keys — see pi-mono/packages/coding-agent/docs/providers.md.
SYSTEM.md / APPEND_SYSTEM.md: Replace the default system prompt with .pi/SYSTEM.md (project) or ~/.pi/agent/SYSTEM.md (global). Append without replacing via APPEND_SYSTEM.md at the same locations. Context files and skills are still appended after override — pi-mono/packages/coding-agent/README.md §Context Files.
Message queue: Enter queues a steering message (delivered between tool calls); Alt+Enter queues a (delivered after the agent finishes all work). Escape aborts; Alt+Up retrieves queued messages. Settings: and ( default vs ); (, , ) — §Message Queue, .
Workflows
Find where a skill is discovered: Walk docs/skills.md locations, then cross-check package-manager.ts auto-discovery and settings.jsonskills arrays.
Explain shadowing: Combine resourcePrecedenceRank ordering with mergePaths / loadSkills "first name wins" using file citations only.
User forced load: /skill:name expansion — pi-mono/packages/coding-agent/src/core/agent-session.ts.
Debug compaction: Read compaction.md algorithm; check reserveTokens / keepRecentTokens in settings.json; trace chained compactions via firstKeptEntryId.
Add custom model: Follow models.md minimal example for Ollama/vLLM/LM Studio; check compat flags for non-standard OpenAI-compatible servers.
Session archaeology: Parse .jsonl using the switch example in session.md; navigate branches via /tree per tree.md.
Configure providers: Check providers.md for subscription OAuth (/login) vs API-key flow; see auth.json layout and env-var table.
Platform setup: For tmux, read tmux.md (extended-keys); for Windows, windows.md (shellPath); for Android, termux.md; for terminal quirks, terminal-setup.md.
Shell aliases: Read shell-aliases.md for shellCommandPrefix to make pi's bash tool see dotfile aliases.
Develop pi from source: Read development.md for clone/build/test and pi-test.sh runner.
Control tools: Use --tools read,grep,find,ls for read-only mode; --no-tools to disable all built-in tools (extension tools still work). Default is read,bash,edit,write.
: or for CI/scripts. Combine with for machine-readable JSON-line output.
Anti-patterns
Do not invent MCP or sub-agent behavior as "built into core"; check pi-mono/packages/coding-agent/README.md philosophy section for what core omits.
Do not claim exact merge behavior without citing resource-loader.ts and skills.ts.
Do not describe compaction cut-point behavior from memory; cite compaction.md algorithm section.
Do not guess credential resolution order; cite providers.md for the exact precedence.
follow-up
steeringMode
followUpMode
"one-at-a-time"
"all"
transport
"sse"
"websocket"
"auto"
pi-mono/packages/coding-agent/README.md
pi-mono/packages/coding-agent/docs/settings.md
Built-in tools: Default four: read, bash, edit, write. Additional available: grep, find, ls. Control via --tools <list> (e.g., --tools read,grep,find,ls for read-only) and --no-tools (disables all built-in; extension tools still work) — pi-mono/packages/coding-agent/README.md §Tool Options.
Editor features: @ fuzzy-searches project files; Tab completes paths; Ctrl+V pastes images (Alt+V on Windows); !command runs and sends output to LLM; !!command runs without sending — pi-mono/packages/coding-agent/README.md §Editor, pi-mono/packages/coding-agent/docs/keybindings.md.
Print mode: -p/--print for non-interactive stdout output. Reads piped stdin: cat README.md | pi -p "Summarize". Combine with --mode json for JSON-line output — pi-mono/packages/coding-agent/README.md §Modes.
@files CLI arguments: Prefix files with @ on CLI: pi @screenshot.png "What's in this?", pi @code.ts @test.ts "Review". Included as part of the initial message — pi-mono/packages/coding-agent/README.md §File Arguments.
Resource control flags: --no-extensions, --no-skills, --no-prompt-templates, --no-themes disable auto-discovery. -e/--extension, --skill, --prompt-template, --theme for explicit loading. Combine --no-* with explicit flags for exact control (e.g., --no-extensions -e ./my-ext.ts) — pi-mono/packages/coding-agent/README.md §Resource Options.
Model shorthand: --model provider/id (e.g., openai/gpt-4o), --model name:thinking (e.g., sonnet:high), --models <patterns> for Ctrl+P cycling, --list-models — pi-mono/packages/coding-agent/README.md §Model Options.
Philosophy (what pi intentionally omits): No MCP (use CLI tools or extensions), no sub-agents (use tmux or extensions), no permission popups (use container or extension), no plan mode (use files or extension), no built-in to-dos, no background bash (use tmux). Everything is buildable via extensions — pi-mono/packages/coding-agent/README.md §Philosophy.
Non-interactive mode
pi -p "prompt"
cat file | pi -p "Summarize"
--mode json
Custom system prompt: Place .pi/SYSTEM.md in a project or ~/.pi/agent/SYSTEM.md globally to replace the default prompt. Use APPEND_SYSTEM.md at the same locations to append without replacing.
