pcq

pcq is the contract for agent-run ML experiments. This repository hosts the contract specification under spec/ and the reference Python implementation under src/pcq/. Install the reference impl: uv add pcq (Apache-2.0).

The contract turns a project with cq.yaml into a reproducible experiment unit. The reference Python implementation loads config, resolves output paths, captures metrics, writes standard artifacts, finalizes run evidence, and exposes JSON/JSONL/MCP surfaces that coding agents, CI jobs, notebooks, and services can consume. See spec/IMPLEMENTATIONS.md for the registered implementation list (Python reference + CQ Go production worker today) and the procedure for adding yours.

pcq is not a training framework, model zoo, adapter matrix, or CQ-only client. Use PyTorch, Hugging Face Trainer, Lightning, sklearn, TabPFN, PyCaret, XGBoost, shell scripts, remote jobs, or project-local research code. The contract is the integration layer.

pcq does not operate the model.
pcq operates the experiment boundary.

SITE | INTRODUCTION | V4_DIRECTION | VISION | AGENT_OPERABILITY | RUN_RECORD | AGENT_OPERATING_GUIDE | CHANGELOG

Contract specification (single source of truth): spec/INDEX.md | SPEC | CQ_YAML_RUNTIME_CONTRACT | JSON_CONTRACTS | STRICTNESS | CQ_MCP_SPEC | VERSIONING | CONFORMANCE | schemas/ (auto-exported via scripts/export_schemas.py)

Case studies (external evidence): mnist-dogfood | tabular-dogfood | mcp-dogfood | cq-worker-dogfood

Agent-readable site files: llms.txt, llms-full.txt, agent-manifest.json.

Identity

pcq = open-source experiment evidence/control library
cq  = managed execution + orchestration + dashboard + agent loop

CQ service is one managed consumer of the contract. pcq remains useful without CQ: locally, in CI, in notebooks, and inside third-party orchestrators.

Related MCP server: CatoBot autoexperiment MCP Server

Why pcq

• Framework-neutral — keep the training stack that fits the problem.

• Agent-readable — use JSON/JSONL instead of terminal scraping.

• Agent-verifiable — validate source, config, environment, metrics, artifacts, and run records.

• Agent-operable — run, observe, validate, describe, compare, lineage, and iterate through stable commands.

• Service-ready — CQ can consume the same contract for managed execution and automatic experiment loops.

What's New (v4.4 – v4.6)

Three agent-fillable metadata fields were added to run_record.json across the last three minor releases, making each run's evidence richer with zero extra code in most cases.

attribution — who

Records the human author, the AI committer, and the operator that launched the run. Coding agents (Claude Code, Codex) fill this automatically from their identity context.

pcq.attribution(
    author={"kind": "human", "id": "alice"},
    committer={"kind": "agent", "id": "claude-code"},
    operator="ci-runner-42",
)

Spec: spec/SPEC.md § Attribution

worker_spec — where

Records CPU model, core count, GPU kind/VRAM, total memory, and OS. Called with no arguments for a full auto-detection pass.

pcq.worker_spec()          # 자동 감지 — 인수 불필요

Spec: spec/SPEC.md § Worker Spec

fingerprint — what

Records dataset modality, task kind, sample count, size class, domain, and PII-safe summary statistics. Accepts a NumPy/pandas array or DataFrame and infers most fields.

pcq.fingerprint(X, y, modality="tabular")

Spec: spec/SPEC.md § Fingerprint

All three fields are optional — existing runs remain valid. When present they appear as first-class evidence in run_record.json and are surfaced through pcq describe-run --json.

Reproducibility Substrate

3개의 선택적 필드로 독립 재현이 가능한 substrate를 run_record.json에 제공한다. pcq는 검증하지 않는다 — 검증을 가능하게 만든다.

상세 스키마, PHI 게이트(R5), integrity 확장, R8 한계 문장: spec/SPEC.md § Reproducibility Pack

Note: code content sha proves WHAT code was recorded, not THAT it produced these outputs. See SPEC.md R8.

Note: pcq records claims, not judgments — intent is a recorded assertion (a fact about what was claimed), not a pcq verdict on success.

Installation

uv add pcq
# Optional — to expose pcq as MCP tools to agent runtimes:
uv add 'pcq[mcp]'

pyproject.toml:

[project]
dependencies = ["pcq"]              # core only
# or:
dependencies = ["pcq[mcp]"]         # core + Model Context Protocol server

Docker (MCP server only)

A minimal container image is also published; it packages pcq[mcp] from PyPI and runs pcq mcp serve on stdio.

docker build -t pcq .
docker run -i --rm pcq          # MCP client attaches to stdin/stdout

The image is intentionally scoped to the MCP server surface — for pcq run, pcq describe-run, pcq agent install and other CLI subcommands, install pcq directly with uv add pcq instead.

For a tag, branch, or private fork:

[tool.uv.sources]
pcq = { git = "https://github.com/playidea-lab/pcq.git", tag = "v4.1.0" }

The PyPI distribution, import name, CLI command, GitHub repository, runtime workspace, and JSON contract namespace are all pcq. Runtime contract names from CQ remain stable: cq.yaml, CQ_CONFIG_JSON, and cq://.

Minimal Contract

cq.yaml declares the run:

name: sklearn-baseline
cmd: uv run python train.py
configs:
  output_dir: output
  seed: 42
  strictness: 3
  monitor: eval_acc
  mode: max
metrics:
  - epoch
  - eval_acc
artifacts:
  - output/
inputs: {}

train.py can use any framework:

import pickle

import pcq
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

cfg = pcq.config()
out = pcq.output_dir()
pcq.seed_everything(cfg.get("seed", 42))

x, y = load_iris(return_X_y=True)
x_train, x_eval, y_train, y_eval = train_test_split(
    x,
    y,
    test_size=0.25,
    random_state=int(cfg.get("seed", 42)),
    stratify=y,
)

model = RandomForestClassifier(random_state=int(cfg.get("seed", 42)))
model.fit(x_train, y_train)
eval_acc = float(model.score(x_eval, y_eval))

with (out / "model.pkl").open("wb") as f:
    pickle.dump(model, f)

history = [{"epoch": 0, "eval_acc": eval_acc}]
pcq.log(**history[-1])
pcq.save_all(history=history, artifacts={"model": "model.pkl"})

No sklearn adapter is required. The same pattern works for HF Trainer, Lightning, XGBoost, TabPFN, PyCaret, shell commands, or custom code.

Agent Command Surface

Read and validate the project:

pcq resolve --json
pcq inspect . --json
pcq validate . --strictness 2 --json

Run the project:

pcq run --path . --json
pcq run --path . --jsonl
pcq run --path . --events output/events.jsonl --json

Validate and summarize outputs:

pcq validate-run output --strictness 3 --json
pcq describe-run output --json
pcq compare-runs old_output new_output --json
pcq lineage output --json

Iterate:

pcq apply-plan experiment.plan.json --json

Agent rule: prefer JSON/JSONL surfaces over scraping human output. pcq reports facts; the agent or service chooses policy.

Standard Artifacts

A completed run should produce:

• config.json

• metrics.json

• manifest.json

• run_summary.json

• run_record.json

• validation_report.json

run_record.json is the canonical completion object. It combines execution, source, environment, input identity, metric schema, artifact manifest, agent provenance, validation, and summary evidence.

Agent Runtime Assets

pcq can install its canonical agent instructions and skill into a project. Package installation itself never mutates project agent files.

pcq agent install --target codex --path .
pcq agent install --target claude --path .
pcq agent install --target both --path . --dry-run --json
pcq agent status --target both --path . --json

To also wire the project for MCP-aware agents (Claude Code, Codex), install pcq[mcp] and pass --mcp:

uv add 'pcq[mcp]'
pcq agent install --target claude --path . --mcp     # writes .mcp.json
pcq mcp serve                                         # stdio (default)

This exposes 14 mcp__pcq__* tools (resolve_project, validate_run, describe_run, compare_runs, ...) so agents call pcq directly without subprocess parsing. See MCP Integration.

Reusable assets:

• templates/AGENTS.pcq.md

• skills/pcq/SKILL.md

v4 Direction

v4 clarifies the product boundary:

• contract-first workflow, not a 3-tier training API

• project-local training code, not built-in production catalogs

• contract scripts, not framework adapters

• run evidence validation, not recipe ownership

• JSON/JSONL facts, not prose parsing

See pcq v4 Direction.

Development

uv run ruff check src/ tests/ scripts/
uv run python -m compileall src/pcq
uv run pytest tests/ -q
bash scripts/release-smoke.sh

License

Apache-2.0.