pi-autoresearch

Autonomous experiment loop for pi

Install · Usage · How it works

Try an idea, measure it, keep what works, discard what doesn't, repeat forever.

An extension for pi — an AI coding agent that runs in your terminal. pi-autoresearch gives pi the tools and workflow to run autonomous optimization loops: try an idea, benchmark it, keep improvements, revert regressions, repeat.

Inspired by karpathy/autoresearch. Works for any optimization target: test speed, bundle size, LLM training, build times, Lighthouse scores.

pi install npm:pi-autoresearch

Examples:

/autoresearch optimize unit test runtime, monitor correctness
/autoresearch model training, run 5 minutes of train.py and note the loss ratio as optimization target
/autoresearch export
/autoresearch off
/autoresearch clear

To avoid conflicts with other pi extensions, override or disable these shortcuts in <agent-dir>/extensions/pi-autoresearch.json. <agent-dir> is the active pi profile config directory (usually ~/.pi/agent, or PI_CODING_AGENT_DIR when set):

{
  "shortcuts": {
    "fullscreenDashboard": "ctrl+shift+y"
  }
}

Use null to skip registering a shortcut. Omitted shortcuts keep their defaults.

• Dashboard widget — always visible above the editor: a full results table with columns for commit, metric, status, and description.
• Confidence score — after 3+ runs, shows how the best improvement compares to the session noise floor. ≥2.0× (green) = likely real, 1.0–2.0× (yellow) = above noise but marginal, <1.0× (red) = within noise.
• Fullscreen overlay — Ctrl+Shift+F opens a scrollable full-terminal dashboard. Shows a live spinner with elapsed time for running experiments.

autoresearch-create asks a few questions (or infers from context) about your goal, command, metric, and files in scope — then writes two files and starts the loop immediately:

autoresearch-finalize turns a noisy autoresearch branch into clean, independent branches — one per logical change, each starting from the merge-base. Groups must not share files, so each branch can be reviewed and merged independently.

autoresearch-hooks (optional) helps author .auto/hooks/before.sh and .auto/hooks/after.sh for a session. It ships with ten reference scripts in skills/autoresearch-hooks/examples/ (external search, learnings journal, native notifications, anti-thrash, idea rotation, and more) — the skill handles the contract, you pick the inspiration. The core autoresearch loop has no hook awareness.

All session files live in a single .auto/ subfolder at the working-directory root — one folder to preserve across reverts, gitignore, and clean up. (Legacy flat autoresearch.* files are still read for in-flight sessions.)

pi install npm:pi-autoresearch

cp -r extensions/pi-autoresearch ~/.pi/agent/extensions/
cp -r skills/autoresearch-create ~/.pi/agent/skills/

/skill:autoresearch-create

The agent asks about your goal, command, metric, and files in scope — or infers them from context. It then creates a branch, writes .auto/prompt.md and .auto/measure.sh, runs the baseline, and starts looping immediately.

The agent runs autonomously: edit → commit → run_experiment → log_experiment → keep or revert → repeat. It never stops unless interrupted.

Every result is appended to .auto/log.jsonl in your project — one line per run. This means:

• Survives restarts — the agent can resume a session by reading the file
• Survives context resets — .auto/prompt.md captures what's been tried so a fresh agent has full context
• Human readable — open it anytime to see the full history
• Branch-aware — each branch has its own session

/skill:autoresearch-finalize

The agent reads .auto/log.jsonl, groups kept experiments into logical changesets, proposes the grouping for your approval, then creates independent branches from the merge-base. Each commit includes metric improvements in the message. Groups must not share files, so branches can be reviewed and merged independently.

• Widget — full results table, always visible above the editor
• Ctrl+Shift+F — fullscreen scrollable dashboard overlay (config key: shortcuts.fullscreenDashboard)
• /autoresearch export — open a live browser dashboard with chart and share card
• Escape — interrupt anytime and ask for a summary

The extension is domain-agnostic infrastructure. The skill encodes domain knowledge. This separation means one extension serves unlimited domains.

┌──────────────────────┐     ┌──────────────────────────┐
│  Extension (global)  │     │  Skill (per-domain)       │
│                      │     │                           │
│  run_experiment      │◄────│  command: pnpm test       │
│  log_experiment      │     │  metric: seconds (lower)  │
│  widget + dashboard  │     │  scope: vitest configs    │
│                      │     │  ideas: pool, parallel…   │
└──────────────────────┘     └──────────────────────────┘

Two files keep the session alive across restarts and context resets:

.auto/log.jsonl   — append-only log of every run (metric, status, commit, description)
.auto/prompt.md   — living document: objective, what's been tried, dead ends, key wins

A fresh agent with no memory can read these two files and continue exactly where the previous session left off.

Create .auto/config.json in your pi session directory to customize behavior:

{
  "workingDir": "/path/to/project",
  "maxIterations": 50
}

The loop is designed to run unattended across context limits. When pi's auto-compaction summarizes the older portion of the conversation, autoresearch detects the resulting idle and re-prompts the agent to re-read .auto/prompt.md, the tail of .auto/log.jsonl, .auto/ideas.md, and git log before continuing. All progress is persisted in those files, so the post-summary turn rehydrates from the source of truth instead of relying on whatever survived compaction. No tuning required — if pi's auto-compaction is enabled (the default), this just works.

After 3+ experiments in a session, pi-autoresearch computes a confidence score — how the best improvement compares to the session's noise floor. This helps distinguish real gains from benchmark jitter, especially on noisy signals like ML training, Lighthouse scores, or flaky benchmarks.

How it works:

• Uses Median Absolute Deviation (MAD) of all metric values in the current segment as a robust noise estimator.
• Confidence = |best_improvement| / MAD. A score of 2.0× means the best improvement is twice the noise floor.
• Shown in the widget, fullscreen dashboard, and log_experiment output.
• Persisted to .auto/log.jsonl on each result for post-hoc analysis.
• Advisory only — never auto-discards. The agent is guided to re-run experiments when confidence is low, but the final keep/discard decision stays with the agent.

Create .auto/checks.sh to run correctness checks (tests, types, lint) after every passing benchmark. This ensures optimizations don't break things.

#!/bin/bash
set -euo pipefail
pnpm test --run
pnpm typecheck

How it works:

• If the file doesn't exist, everything behaves exactly as before — no changes to the loop.
• If it exists, it runs automatically after every benchmark that exits 0.
• Checks execution time does not affect the primary metric.
• If checks fail, the experiment is logged as checks_failed (same behavior as a crash — no commit, revert changes).
• The checks_failed status is shown separately in the dashboard so you can distinguish correctness failures from benchmark crashes.
• Checks have a separate timeout (default 300s, configurable via checks_timeout_seconds in run_experiment).

Drop executable scripts in .auto/hooks/ to run code at iteration boundaries. Hooks are transparent to the agent — the agent calls tools and sees results; hooks run alongside without any agent-facing surface.

• .auto/hooks/before.sh — fires before every iteration (at /autoresearch activation and at the end of every log_experiment, after after.sh). Use for prospective work: fetch research, prime context for the next attempt.
• .auto/hooks/after.sh — fires at the end of every log_experiment. Use for retrospective work: annotate learnings, send notifications.

Contract:

• Must be executable (chmod +x). Preserved on revert — the entire .auto/ folder survives (as do legacy autoresearch.* artefacts).
• Stdin — a JSON object on a single line. Shape depends on the stage (see below). Extract fields with jq.
• Stdout is delivered to the agent as a steer message (capped at 8 KB). Empty stdout = silent.
• Non-zero exit or >30s timeout surfaces an error steer to the agent.
• Each fire appends a {"type":"hook",…} entry to .auto/log.jsonl for observability.

before.sh stdin (on fresh activation last_run is null):

{
  "event": "before",
  "cwd": "/path/to/workdir",
  "next_run": 6,
  "last_run": {
    "run": 5, "status": "discard", "metric": 42.1,
    "description": "…",
    "asi": { "hypothesis": "…", "next_focus": "…" }
  },
  "session": {
    "metric_name": "total_ms", "metric_unit": "ms", "direction": "lower",
    "baseline_metric": 40.7, "best_metric": 33.5,
    "run_count": 5, "goal": "optimize sort speed"
  }
}

after.sh stdin:

{
  "event": "after",
  "cwd": "/path/to/workdir",
  "run_entry": {
    "run": 6, "status": "discard", "metric": 38.9,
    "description": "…",
    "asi": { "hypothesis": "…", "learned": "…" }
  },
  "session": { "metric_name": "total_ms", "direction": "lower", "baseline_metric": 40.7, "best_metric": 33.5, "run_count": 6, "goal": "…" }
}

Agent signal. The agent writes description and asi.* fields in its log_experiment calls for its own future-self reasoning. The hook opportunistically mines whichever fields the agent naturally uses — asi.hypothesis, asi.next_focus, description, etc. There is no dedicated "hook input" field; the agent is unaware the hook exists.

Examples. Reference scripts for both stages live at skills/autoresearch-hooks/examples/ — external search, qmd document search, persistent learnings, native notifications, git tagging, anti-thrash, idea rotator, hypothesis reflection, context rotation. Copy one to your session's .auto/hooks/ directory, adapt, chmod +x.

1. Install pi — follow the instructions at pi.dev
2. An API key for your preferred LLM provider (configured in pi)

Autoresearch loops run autonomously and can burn through tokens. Two ways to cap spend:

• API key limits — most providers let you set per-key or monthly budgets. Check your provider's dashboard.
• maxIterations — cap experiments per session in .auto/config.json:

  {
    "maxIterations": 30
  }

MIT