Extensions Catalog

Auto-generated by packages/core/registry.ts. Do not hand-edit.

For install package choices, see docs/profiles.md. For extension authoring, see docs/building-extensions.md.

Name Avenue Category Status Summary Profiles
autonomous-loop First-party workflow beta Autonomous loop mode. /loop activates; agent re-queues itself after each turn via pi.sendUserMessage followUp. /loop stop (operator) or the loop_complete tool (model) exits; an aborted or errored turn stops it. Max iterations guarded by PI_KIT_LOOP_MAX (default 20). balanced, long-horizon, self-improving, engagement
autonomy-run First-party workflow beta Operator command /autonomy: create, plan, start and manage autonomous runs (the packages/autonomy engine) from inside pi. A run starts only after you confirm its resolved boundary on screen, or from a contract that already carries an authorisation for exactly that boundary; it is a command, not a tool, so the model cannot start, steer, authorise or promote a run. long-horizon
branch-lab First-party execution beta Worktree/branch manager for isolated pi sessions. Creates, lists, switches, discards, and merges git branches for risky or experimental task isolation. long-horizon, self-improving, engagement
compress First-party workflow beta Instant deterministic /compress: compacts through pi's native pipeline but builds the summary from user asks and assistant replies only, with no model call. balanced, long-horizon, self-improving, engagement, lite
conductor First-party planning beta Durable engagement lifecycle, independent validator dispatch, and verifier-board-gated phase transitions. self-improving, engagement
context-sieve First-party memory beta Single authority for model-visible system-prompt assembly. Collects contributions from other extensions via .pi/ctx-contributions/ files, enforces a token budget, and injects them as system-prompt and hidden-message context; deliberately leaves pi's native compaction summarization untouched. balanced, long-horizon, self-improving, engagement, lite
control-center First-party ui beta One navigable control centre for Luna: /menu (and alt+m) opens a searchable hub of levels and pages — profile, effort, thinking, security, autonomy, work, session, connect and system. Rows show their current value where a process registry publishes one, fall back to the typed command, and hold back any change that lowers oversight until a confirmation the terminal alone can answer. balanced, long-horizon, self-improving, engagement, lite
delegation-guard First-party safety beta The single launch contract for child pi sessions. Every delegation path (subagents, workflows, completion review, conductor specialists and validators) gets its extension arguments and environment here: mandatory protections and task governance always load, a missing protection refuses the launch, the launch reserves against the shared effort budget, and the child verifies before any tool runs that everything required registered itself, otherwise it blocks all tools and exits. balanced, long-horizon, self-improving, engagement, lite
dual-review First-party execution experimental Spawns a second pi reviewer agent on git diffs, as a read-only child governed by delegation-guard (the parent's protections, the effort ledger; nothing starts without the guard). /review or dual_review tool. Set DUAL_REVIEW_MODEL to route to a different model; uses the primary model if unset. self-improving
effort First-party workflow beta Execution-effort control: five tiers (E1 Minimal to E5 Exhaustive) that set how deep and broad the agent works and how much it delegates, independent of thinking level, profile and permissions. /effort picks and saves the tier, one prompt contribution replaces the last at each user turn, and a shared ledger enforces concurrent/total/scout child limits at every depth. balanced, long-horizon, self-improving, engagement, lite
engagement-governance First-party safety beta Engagement-specific governance layer. Composes tool-firewall with scope/ROE enforcement, action-card approval, and evidence chain of custody. engagement
finish-reason-retry First-party safety beta Makes pi's built-in auto-retry recognise 'Provider finish_reason: error' (raw OpenAI-completions providers, e.g. OpenRouter) as a transient error, and makes an exhausted context window visible: when pi auto-compaction is off an overflow error says so and how to recover, a reply clamped by the window headroom is explained from the model's real maxTokens/contextWindow, and a failed compact-and-retry names the way out. Also the turn doctor: one bounded steer (two per user message) when a model returns an empty turn, tool-argument errors rewritten to name the bad fields and the required keys, and a 402 that says how many output tokens the account can afford is retried once with that maximum. Provider failures are classified once (credit, rate limit, server) and offered to provider-router's fallback list. Loaded in subagent children, where the parent otherwise sees only a raw provider error. balanced, long-horizon, self-improving, engagement, lite
goal-core First-party planning beta Explicit mission-goal management. Stores the active goal in .pi/GOAL.yaml and writes a contribution file for context-sieve to inject (survives compaction). /goal sets, shows, or clears the goal. Also enforces the unattended ceiling (PI_KIT_CEILING_*, or /away): the first of turns, money or wall time reached parks the run, and a terminal bell/OSC 9 rings for a card left unanswered a minute, a long run ending, or a park. balanced, long-horizon, self-improving, engagement, lite
guidelines First-party workflow stable Injects project-specific guidelines from a GUIDELINES.md file into the agent context at session start. balanced, long-horizon, self-improving, engagement
herdr First-party ui beta Optional herdr integration. Inside a herdr pane it raises a herdr notification and the pane's blocked state when Luna waits on an approval or a question, when a verify check fails, and when a long run finishes. Outside herdr it does nothing. It works beside herdr's own pi integration (herdr integration install pi), which reports working and idle. lite, balanced, long-horizon, self-improving, engagement
human-console First-party ui beta File-backed human approval and question broker for attended and headless sessions. balanced, long-horizon, self-improving, engagement
kit-update First-party ui beta Update checker and /update command: notifies (at most daily, in the background) when pi, Luna on its latest, next or trunk channel (from its git repository or npm), or a linked package can be updated, then applies it with pi's package manager and re-applies the active profile. /update undo returns to the journalled state before the last update. balanced, long-horizon, self-improving, engagement, lite
mcp-router First-party execution stub Stub: does not route MCP tool calls or provide failover between MCP server instances yet; session_start only reports it loaded as a stub and tool_call is a no-op. (none)
memory-local First-party memory experimental Superseded by memory-vault (which imports its store once); kept for operators who load it explicitly. Do not load it next to memory-vault: both register memory_search and pi keeps only one. Persistent memory stored in ~/.pi/agent/memory-local with keyword fallback and optional LM Studio semantic embedding search. Automatically recalls relevant memories into a context-sieve contribution on non-trivial input, in addition to manual memory_search. (none)
memory-mem0 First-party memory experimental Mem0-compatible memory tools backed by the Luna memory MCP server by default, with legacy mem0 REST support behind PI_KIT_MEMORY_BACKEND=mem0. self-improving
memory-vault First-party memory beta Durable memory in an Obsidian-compatible vault (default ~/.pi/vault): memory_save/search/forget tools, /remember and /memory, automatic per-turn recaps by a configurable cheap model (Projects//Recaps), promotion of flagged decisions/preferences/gotchas to deduplicated notes, and relevance-gated recall injected as a hidden message (never the system prompt). Replaces memory-local and imports its store once. balanced, long-horizon, self-improving, engagement
orchestrator First-party planning beta Delegation policy over the guarded subagent tool: classifies requests, writes budgeted context-sieve contributions, honours user and effort restrictions, and requires trusted verifier completion. Shipped roles are resolved directly; project role copies are not materialised. balanced, long-horizon, self-improving, engagement
progress-guard First-party workflow beta Auto-boost against loops: detects repeated actions, A→B→A oscillation, and read-without-progress; prompts self-reflection; auto-nudges when an autonomous-loop is armed (no env var needed); escalates a persistent loop to recovery-orchestrator. balanced, long-horizon, self-improving, engagement, lite
provider-router First-party execution beta Opt-in model router that switches between two operator-configured models (hot_path_model and strong_model) by task class from PI_KIT_ROUTING_POLICY; a no-op when the policy is unset. The policy may also list up to three fallback models: on a credit, rate-limit or server failure reported by finish-reason-retry the session moves to the next one that is not cooling down and returns afterwards. It does not select providers by cost or manage local model runtimes. long-horizon, self-improving, engagement
recovery-orchestrator First-party workflow beta Deep root-cause recovery mode that progress-guard escalates into. On a persistent stuck signature it writes a recovery report scaffold and steers the §2 flow (fresh scouts + forked top-10 + synthesis + primary/backup plan + delegated fix) via a context-sieve contribution. Non-destructive; bounded attempts per signature. long-horizon, self-improving, engagement
remote-control First-party ui beta /remote: drive pi from the pi remote Android app over the LAN or Tailscale. Links phones with a one-time code plus approval at the terminal, and shares this session with them. balanced, long-horizon, self-improving
save First-party memory beta /save writes a check-in snapshot of the session (pi's compaction summary) to disk or a vault project without compacting, and copies every compaction summary to the same log. balanced, long-horizon, self-improving, engagement, lite
secret-guard First-party safety beta Secrets/protected-path backstop (in the balanced and long-horizon profiles). Blocks reads of credential files, writes/edits to secret files, protected engagement/repo paths, or secret content (key material, tokens), and bash commands that leak or exfiltrate them (cat/cp/mv/base64/curl). Protected-path list kept in parity with engagement-governance via verify.ts. balanced, long-horizon
self-improvement First-party execution beta The /improve self-improvement cycle: mine traces → propose a reviewable AGENTS.md diff (never a live model, never auto-committed). Applies only when explicitly armed (/improve arm [id]), which applies a proposal already written for review rather than mining again, and only to the allowlisted AGENTS.md 'Learned notes' section. self-improving
session-helpers First-party ui beta Session-control helpers: /clear (new session, optional --handoff note), all-or-nothing /profile switching (snapshots every config file, verifies marker, extension list and firewall, rolls back byte-for-byte on any failure, /profile status prints the effective configuration), /compaction status and toggles backed by the real effective compaction state (published on globalThis, warned once per session), and a lean-ctx binary preflight. balanced, long-horizon, self-improving, engagement, lite
skill-forge First-party execution beta Deterministically mines trace-ledger sessions to synthesise a candidate SKILL.md into .pi/skill-proposals/ (never a live model, never auto-installed), scores Luna against the eval harness for a real regression baseline, and archives superseded skills. Powers the self-improvement loop. self-improving
skill-router First-party workflow beta Progressive disclosure for skills: only core skills are listed in the system prompt; the rest (disable-model-invocation) are reachable through skill_search (BM25 over name/description/triggers) and deterministic trigger hints — one line per matching hidden skill, at most two per prompt, never repeated. /skills lists what is listed vs on-demand. balanced, long-horizon, self-improving, engagement, lite
spec-plan First-party planning stable Encourages spec-before-code discipline: warns when large write/edit operations lack a corresponding plan or spec file in the project. balanced, long-horizon, self-improving, engagement
task-graph First-party planning beta Persistent DAG task board shared by the orchestrator and implementer subagents. Planner emits tasks with dependencies; task_next hands the next unblocked task to an implementer; task_update/task_complete track progress; a claim not renewed within its lease returns to pending, and task_clear empties the board. Stored in .pi/task-graph.json. long-horizon, self-improving, engagement
tool-capture First-party execution beta Lossless tool I/O capture for audit and SIEM: every tool call's exact arguments and output (blocks and errors included) as rotating, gzip-compressed JSONL with content-addressed blobs for large values, joined to the firewall audit and trace-ledger by action hash. Self-manages its disk budget and never breaks the agent. /capture shows status. balanced, long-horizon, self-improving, engagement
tool-firewall First-party safety beta Autonomy gate for every tool call: parses shell commands into their real effects, tiers them (low/medium/high/critical) with session history, auto-allows routine work, judges medium actions in auto mode, learns from operator decisions and denies critical ones. balanced, long-horizon, self-improving, engagement, lite
trace-ledger First-party execution beta Structured action ledger. Records every tool call and result to .pi/trace.jsonl, warns on repeated reads, records each model request's tokens, cache split and cost, and exposes /trace and /usage (where the tokens and money went; warns when the cache hit rate drops). The facts layer for efficiency and anti-loop tooling. balanced, long-horizon, self-improving, engagement, lite
verifier-board First-party execution beta Definition-of-done gate, not tied to a test runner. Checks, reviewers, and validators record verdicts (record_verdict); verdict_status shows verdicts plus goal, todo, task progress and the latest verification report. Stored in .pi/verdicts.json. long-horizon, self-improving, engagement, lite
verify-gate First-party execution beta General completion verification. /verify and verify_completion run the project check (if any) plus an isolated read-only reviewer that checks the goal, todos, tasks, and requests, and record verdicts. PI_KIT_VERIFY_ON_TURN=1 runs the check after edits. balanced, long-horizon, self-improving, engagement, lite
web-console First-party ui beta /console: start, stop and check the bundled Pi Console web UI in packages/web-ui, launched as a detached local server. balanced, long-horizon, self-improving, engagement
auto-commit-on-exit Vendored execution experimental Opt-in (no profile enables it): commits uncommitted changes when the pi session shuts down, using the last assistant message as the commit message. Never stages the .pi/ directory. (none)
caveman Vendored workflow beta Compresses the agent's conversational output (to the human and to sub-agents) via a context-sieve contribution. Auto-off for report/doc writing (session-flip). Never compresses file, code, or report content. balanced, long-horizon, self-improving, engagement, lite
custom-footer Vendored ui stable Status bar: three densities (light, default, heavy) built from prioritised segments that survive narrow terminals, showing identity, model and thinking, profile and effort tier, context headroom, live agents, and the boundary, compaction and failure warnings first. Costs are labelled measured, estimated or unknown. Also live activity, tips and a todo checklist, and an ASCII cat whose face shows the agent's state (/footer cat on off). /footer status is the accessible detail view; /footer light
dirty-repo-guard Vendored safety stable Prompts (or blocks in non-interactive mode) when switching or forking sessions with uncommitted git changes. self-improving
git-checkpoint Vendored execution stable Snapshots the working tree to a shadow ref (refs/luna/checkpoints) when a turn finds it changed, skipping credential files, so /undo and /fork can restore code state; never touches the branch, index or stash. balanced, long-horizon, self-improving, engagement, lite
handoff Vendored workflow stable Appends a timestamped handoff note to HANDOFF.md on session shutdown or via /handoff command. LM-powered summary is a future enhancement. balanced, long-horizon, self-improving, engagement
notify Vendored ui stable Sends a native desktop notification when the pi agent finishes a turn and is waiting for input. Supports Windows Terminal, Kitty, and OSC 777 terminals. balanced, long-horizon, self-improving, engagement, lite
plan-mode Vendored planning stable Read-only exploration mode: /plan toggle or --plan flag restricts tools to read-only set. Write/edit/bash are blocked until plan mode is exited. self-improving
protected-paths Vendored safety stable Blocks write/edit/bash writes to protected paths: secrets and agent-control surfaces under the coding firewall policy, git's hooks/, config and info/ under both policies, plus .git/, node_modules/ and agent definitions under the engagement policy (PI_KIT_PROTECTED_PATHS adds entries). Optional allowlist mode (PI_KIT_WRITE_ALLOWLIST) blocks writes to anything NOT allowlisted — the enforcement point for dream/print-mode runs. balanced, long-horizon, self-improving, engagement, lite
subagent Vendored planning beta Delegate tasks to specialized subagents in isolated context windows. Single and parallel (concurrency-limited, optional background) modes over JSON-mode pi subprocesses that load only the extensions the delegation guard grants (the parent's protections first). Built-in kit roles are always available; roles may set model, thinking, tools, skills (preloaded) and extensions. Esc cancels children. Also runs declarative workflows (packages/kit/workflows, ~/.pi/agent/workflows, .pi/workflows): deterministic step chains with a shared run directory, parallel groups, review gates and resume. The workflow tool runs JavaScript workflow scripts in a sandbox in the background (agent/parallel/pipeline hooks, journal and resume) and posts the result as a follow-up. A run given a result schema loads a child-only StructuredOutput tool and the host validates the submitted value itself. balanced, long-horizon, self-improving, engagement
todo Vendored workflow stable Registers a 'todo' tool for the LLM to manage a persistent TODO.md file (list, add, start, done, toggle, remove, clear) with an in-progress state; every action returns the whole list. Also registers /todos. balanced, long-horizon, self-improving, engagement, lite
trigger-compact Vendored workflow stable Compacts context when token usage crosses a fixed threshold (default 100k, configurable) - only while pi auto-compaction (compaction.enabled) is on, only when it is below pi's own window-minus-reserve trigger, and never in a one-shot child session. A compaction that interrupts a live run is resumed once. Registers /trigger-compact and /compact-threshold [amount off
pi-lens External (reference) external — MIT; Node >=22.19.0. Static review of the supplied 4.4.4 package only; no package code executed or TUI/model-provider validation. Native @ast-grep/napi ^0.50.0, @ast-grep/cli 0.50.0 (dependency lifecycle platform binary setup), bundled core/vendor tree-sitter WASM and optional web-tree-sitter 0.25.10; no runtime typescript dependency. No preinstall/install/postinstall in this manifest; prepare is a source-build hook (build, grammar download, git hooks, lock stamping, loader warming), not a registry-install hook. Registers 13 pi tools: adds symbol_search, project_report, effective_config, lens_diagnostic_mark and pi_lens_activate_tools; removes ast_grep_dump, ast_dump and standalone lsp_diagnostics. Exact firewall entries allow analysis/config/discovery, ask for structural replacement, LSP navigation and diagnostic suppression; path checks include paths arrays and analysisRoot (tests/tool-firewall-smoke.ts, tests/firewall-shell-smoke.ts). Diagnostics also run in process through session/write/edit/turn hooks, outside tool-call isolation. Profiles select the companion as a whole; balanced does not filter its registered tools. The earlier dependency warning applies to 3.8.63: 4.4.4 declares optional peers pi-coding-agent=* and pi-tui=*, typebox=^1.0.0. pi 1.3 warns on host packages in dependencies, not this manifest, though typebox still differs from the recommended wildcard. The kit defaults PI_LENS_DISABLE_LSP_INSTALL=1 and PI_LENS_DISABLE_TOOL_INSTALL=1; 4.4.4 honours both, including managed refreshes. Existing tools keep working. Missing non-vendored grammar downloads are separately trust-gated and not disabled by these switches; hash checking depends on a matching grammars.lock.json entry. See docs/supply-chain.md for the registration diff and limits. balanced, long-horizon, self-improving, engagement
pi-lean-ctx External (reference) external — Apache-2.0. OPT-IN, in no profile: it needs the external lean-ctx Rust CLI (cargo install lean-ctx or brew; not bundled, not in this package's supply chain) and, in its default additive mode, registers extra model-visible tools even when that CLI is absent (ctx_shell, ctx_read, ctx_ls, ctx_find, ctx_grep, lean_ctx, ctx_edit; checked against 3.11.4 in a real pi 1.3.0 TUI with a scratch agent directory). When the CLI is present, its MCP bridge also registers whatever tools the lean-ctx MCP server lists (extensions/mcp-bridge.ts registerMcpTool). Those tools are not inspected the way the built-in ones are: secret-guard matches the built-in names bash/write/edit, so ctx_shell and ctx_edit are outside its checks, and tool-firewall sees them only as unknown tools (policy default: ask) rather than as shell commands. Enable it only for trusted repositories and expect an approval prompt per call; it is installed by install.ts --all only when its binary is on PATH. 3.11.4 static review against 3.10.5: no dependencies, no install lifecycle scripts (prepack/pretest only build the vendor bundle), the same seven registered tools and the same peers (typebox ^1.0.0, pi-tui *, pi-coding-agent >=0.74.0); source changes are SPDX headers, one parameter description and the version constant. Its self-contained extensions/vendor/mcp-sdk.cjs (MCP Client and StdioClientTransport) is rebuilt from @modelcontextprotocol/sdk ^1.32.1 instead of ^1.29.0. GHSA-22jm-h49p-29qw (1.24.0 to 1.31.0) affects only servers that pass a taskStore, so the bundled client was never exposed to it, but a scanner matching the SDK version would flag the old bundle; OSV and npm audit cannot see a vendored bundle, so moving the pin is the only way to update it. Verified to load and degrade gracefully without the CLI. Replaces 3.10.5. (none)
pi-readseek External (reference) external — MIT. @jarkkojs/readseek (native) + xxhash-wasm. Hash-anchored read/edit/grep + structural code maps + AST search — a lean file-I/O primitive for navigating large repos cheaply. Registers replacement tools (manual). OSV reports no advisories for a fresh 0.4.26 install. 0.12.0 (latest, published 2026-10-23) was evaluated against its tarball and in pi 1.3.0 with a scratch agent directory, and not adopted: it is a different tool surface, not an update. It no longer replaces read/edit/write/grep by default; it adds 16 tools (readSeek_show, _tree, _meta, _assets, _syntax, _check, _caption, _objects, _ocr, _search, _grep, _def, _refs, _edit, _rename, _write) beside the built-ins, which more than doubles the active tool-definition size measured in pi (7,033 to 15,845 characters) and works against the lite profile. Its readseek.overrideTools setting is read only from ~/.pi/agent/settings.json and .pi/settings.json, ignoring PI_CODING_AGENT_DIR. readSeek_check rules and readSeek_search rule accept @path file references that the firewall's path fields do not see, so allow entries would let them read credential files unchecked. readSeek_caption/_objects/_ocr run locally installed vision models. The licence is Apache-2.0 and the linux-x64 AppImage is LGPL-3.0-or-later; it depends on @jarkkojs/readseek-api (bundling @sinclair/typebox 0.34.52 and @bgotink/kdl). Adopting it needs exact firewall entries for all 16 tools, @path handling in the classifier and a decision on the lite surface; 0.4.26 stays pinned until then. long-horizon, self-improving, engagement, lite