@lebronj/pi-suite

Slim Pi suite for WildClawBench / Claw-Eval. Default install profile is Lenovo ModelFactory DeepSeek; the bench profile (TEAM_PROFILE=zhizengzeng) provisions GPT-5.5 as the main model plus Gemini vision tools.

Evolution helpers and subagents are not loaded during evaluation. See docs/bench/LEADERBOARD.md.

Install

npm install -g --ignore-scripts @earendil-works/[email protected]
pi install npm:@lebronj/pi-suite
pi install npm:pi-web-access
pi install npm:@lebronj/pi-lsp

Or bootstrap Pi, write the DeepSeek provider, and install the slim suite. The script asks for the API key on the terminal; it does not ship a key.

curl -fsSL https://registry.npmjs.org/@lebronj/pi-suite/-/pi-suite-0.1.40.tgz | tar -xzO package/scripts/bootstrap.sh | bash

Defaults written to ~/.pi/agent/models.json / settings.json:

• Provider: lenovo-deepseek-v4-flash
• Base URL: https://modelfactory.lenovo.com/service-large-600-1777255649450/llm/v1
• API: openai-completions
• Default model: DeepSeek-V4-Flash-0731

Override with TEAM_BASE_URL, TEAM_MODEL, or TEAM_API_KEY. Old claude-code.club / gpt-5.5 leftover: TEAM_PROFILE=legacy (see scripts/bootstrap.legacy.sh).

Bench profile: GPT-5.5 + Gemini vision

TEAM_PROFILE=zhizengzeng ZHIZENGZENG_API_KEY=sk-... bash scripts/bootstrap.sh

Writes the zhizengzeng provider (https://api.zhizengzeng.com/v1) with:

• gpt-5.5 / gpt-5.5-pro — main control, Responses API (chat completions rejects tools + reasoning_effort for these models), text+image input.
• gemini-3.1-pro-preview — vision fallback model, chat completions.
• ~/.pi/agent/media-tools.json — key + base URL for the gemini_vision tool below.

What Is Included

• update_plan for long-horizon checklists.
• Benchmark control (extensions/bench-control.ts): records the Pi/suite/tool runtime manifest, enforces explicit search-query limits across turns, checks requested artifacts at 25%/50%/75% of the working budget, and uses PI_BENCH_DEADLINE_EPOCH to steer the agent into artifact finalization before the outer timeout. Finalization blocks new searches, remote vision calls, installations, clones, and downloads.
• Media tools (extensions/media-tools.ts), backed by Zhizengzeng's Google-native gateway (/google/v1beta):
  • gemini_vision(paths, question) — images, whole videos (audio track included, timestamp-aware, MM:SS), and audio files. Oversized videos are transcoded down with ffmpeg; frame sampling is the last resort. smartCrop=true zooms into the relevant image region (two-pass crop-and-reask). startSeconds/durationSeconds clip long videos.
  • image_contact_sheet — build a numbered grid for pixel-based batch classification instead of inferring classes from filenames.
  • video_frames — extract frames to files: fixed interval, exact timestamps, or scene-change detection.
  • image_crop — crop/resize an image with ffmpeg.
  • media_probe — ffprobe metadata (duration, resolution, codecs, fps).
  • gemini_vision uses the configured model only, one bounded retry with exponential backoff, a 30-second default request timeout, a 65-second whole-tool deadline, and a cross-call circuit breaker so a failed visual backend cannot consume the task budget indefinitely.
  • Requires ffmpeg/ffprobe on PATH. Key resolution: ZHIZENGZENG_API_KEY env, then ~/.pi/agent/media-tools.json. Known gateway limits: OpenAI-format video_url does not deliver video; the Files API upload returns 500 — hence inline + transcode.
• Safety gate (extensions/safety-gate.ts), model-agnostic bench rails:
  • Blocks catastrophic filesystem/Git commands, credential harvesting, runtime package installation, executing workspace SKILL.md, and cloning unreviewed code into auto-loaded skill/plugin directories.
  • Scans a repository for likely secrets before Git mutation/publication and blocks without revealing secret values.
  • Preserves existing files beside the original as *.preexisting.* before write/edit or common shell overwrites, blocks writes through symbolic links, treats overwrite authorization as negation-aware, and protects fixed/human-only lines from modification.
• Vendored memory, bench-slim by default: memory_read + lexical memory_search only. Set PI_MEMORY_BENCH=0 to restore write/curator/share tools. For harness runs also set PI_MEMORY_FINALIZE=0 and PI_MEMORY_SKILL_DRAFTS=off to disable shutdown finalization noise.
• Companions installed by bootstrap: pi-web-access, @lebronj/pi-lsp.
• Web fetch fallback (pi-web-access): bootstrap writes ~/.pi/web-search.json with fetchRouting.allowRemoteHostedProviders: true and providers: [http, firecrawl, jina, tinyfish, search1api], so pages that fail plain http/Readability (403 / anti-bot / JS-rendered) fall back to jina (r.jina.ai). Search routing defaults to serper with jina fallback. Set SERPER_API_KEY / JINA_API_KEY in that file for the fallback tiers.
• Reward-only self-evolution loop (bench/evolve/): evaluates the harness on a task set for multiple rounds, feeds back only a clamped scalar score/status plus agent-owned traces and artifact manifests (never grader details), and evolves bench/workspace/ (system prompt append, memory card, tool trims, skills) with evidence-backed change manifests falsified by next-round flips. See bench/evolve/README.md.
• /bench command (extensions/bench.ts), the in-pi switch for the loop: /bench tasks lists tasks, /bench run [task ...] starts a one-off eval in the background, /bench evolve [N] starts the self-evolution loop, /bench shows status/results, /bench stop kills the run. Operator-only: disabled in the eval profile, and bench child processes never register it (PI_BENCH_CHILD=1).

RPC Runner Requirements

• Use Pi 0.84.3 or newer. Bootstrap pins 0.84.3 by default; set PI_VERSION=latest to opt into a newer release after validating extension compatibility.
• In multi-turn sessions, wait for agent_settled before sending the next normal prompt. agent_end only marks one low-level run and may be followed by retry, compaction, or queued continuation.
• If a message must be submitted while Pi is busy, send RPC prompt with streamingBehavior: "followUp" (or "steer" when interruption is intentional). Do not retry a normal prompt in a tight loop after Agent is already processing.
• Keep the outer task timeout and each prompt timeout separate in logs and result status. A 600-second prompt timeout inside a 7200-second task budget is a prompt timeout, not a 7200-second Pi run.
• After any timeout, abort the active run and wait for settlement or replace the Pi process before reusing the session. Multi-role handoffs must not reuse a still-busy process.

Not installed or loaded:

• autogoal and skill-creator (repository-only evolution helpers)
• pi-subagents, pi-mcp-adapter, pi-mono-figma

Do not also list those companions inside this package manifest; duplicates conflict.

Autogoal (evolution only)

The repository-only evolve profile can load /autogoal <task> for bounded autonomous coding runs. It is excluded from evaluation and from the published leaderboard package.

Useful commands:

/autogoal <task>
/autogoal status
/autogoal pause
/autogoal resume
/autogoal checkpoint optional reason
/autogoal drop

Behavior:

• 60% context: prepare and stay concise.
• 75% context: write a structured checkpoint soon.
• 85% context: checkpoint and switch to a new session.
• Completion requires current-state evidence: changed files read after edits and a passing validation command.
• Subagents are optional and budgeted; worker subagents must use worktree isolation.
• Run artifacts are written under ~/.pi/agent/workflow-runs/autogoal-<run-id>/.

Update Plan

The update_plan tool gives Pi a Codex-style visible execution checklist for non-trivial tasks. It supports init, start, done, drop, rm, append, and note, shows active plan progress in the UI, and injects guidance to use it for 3+ step tasks or user-provided checklists.

Useful commands:

/plan-status
/plan-clear

Default Model Setup

Bootstrap writes the Lenovo DeepSeek provider above. The API key is entered at install time and stays on the user's machine. Do not publish a shared key in this package.

Old team endpoint leftover (main):

TEAM_PROFILE=legacy bash scripts/bootstrap.legacy.sh

Memory And Versioning

@jhp/pi-memory works without qmd for core memory features:

• memory_write
• memory_read
• memory_edit
• scratchpad
• memory_curate

memory_search automatically falls back to local lexical matching without qmd or embeddings. The bootstrap script installs qmd and initializes its collection when Bun is available, but does not run the time-consuming embedding step. Run qmd embed only when semantic search is needed. If Bun is missing, install qmd later:

bun install -g https://github.com/tobi/qmd
qmd collection add ~/.pi/agent/memory --name pi-memory
qmd embed

Memory versioning is enabled by default. It snapshots the resolved memory root and resolved disabled skill-draft root into the local evolution repo, commits local changes automatically, and leaves push manual by default. Standalone Pi resolves to ~/.pi/agent/memory and ~/.pi/agent/skill-drafts; Multica-connected runs can resolve to ~/multica_workspaces/<workspace_id>/.pi/agents/<agent_id>/memory and skills/drafts. memory_curate also scans yesterday's daily log into REVIEW.md when learning is enabled and the daily file changed since the last scan.

For local multi-agent self-evolution, @jhp/pi-memory now supports:

• PI_MEMORY_DIR, PI_SKILL_DRAFTS_DIR, PI_AGENT_ROOT, MULTICA_WORKSPACE_ID, MULTICA_AGENT_ID, and MULTICA_WORKSPACES_ROOT resolvers.
• Agent root initialization with isolated memory/, skills/drafts, skills/generated, inbox/, shared-cache/, profile/, feedback/, and sync_queue/ directories.
• /memory-review plus startup and memory_curate pending proposal reminders.
• A Local Curator Manager registry/dirty-root API for one local manager to process many agent roots safely.
• Share candidate, downflow receive, sync upload/pull, profile generation, Local Curator Manager tools, and feedback JSONL helpers. Server downflow is per-Agent delivery, not broadcast, and local delivery never overwrites formal memory or auto-enables skills.

The external memory curator service uses a systemd user timer when available, with cron fallback. When the service points at a vendored TypeScript CLI under node_modules, the launcher uses Bun or tsx instead of plain Node so Node 22 can run it reliably.

Useful commands:

/memory-version-status
/memory-version-snapshot optional reason
/memory-version-list
/memory-version-restore <snapshot-id> [memory|skill-drafts|all]
/memory-version-push

Memory evolution is local-only by default and does not configure a shared remote. If a user wants backup sync, set PI_EVOLUTION_REMOTE to their own private repo before bootstrap/setup, or add a personal remote later with git -C ~/.pi/agent/evolution remote add origin <url>. Set PI_EVOLUTION_AUTO_PUSH=1 only if automatic remote sync is desired.

Update

pi update

Only update Pi itself:

pi update --self

Only update installed Pi packages:

pi update --extensions