Skip to content

About

Local-first Python framework for generating, executing, validating, and packaging agent training data with executable state, tool evidence, lineage, replay checks, and release artifacts.

Resources

Security policy

Stars

36 stars

Watchers

2 watching

Forks

Repository files navigation

Agent Data Synthesis

简体中文

Agent Data Synthesis generates executable Agent Episodes from three local Domains: Contacts, Mobile Messages, and Workspace Tasks. The Agent-first engine owns bounded model calls, isolated task execution, deterministic admission, private SQLite state, replay, and sanitized dataset collections. Shadow quality review is the default; deterministic admission does not imply human approval.

Quick start

Python 3.13 or newer and uv are required.

uv sync
uv run python main.py --help
uv run python scripts/validate_docs.py
uv run python -m unittest

main.py provides run, resume, replay, create-review-queue, and import-review-labels commands. A run requires a validated JSON configuration and either scripted fixture responses or a separately authorized remote provider. No provider is called by default. Outputs go to the chosen directory under artifacts/: demonstrations.jsonl, negatives.jsonl, manifest.json, quality_report.json, provider_usage.json, and run_report.json.

The old synthesis/ and awm_runtime/ implementations and their commands were removed at cutover. Earlier artifacts and profiles are historical; regenerate with the current interface instead of reading them through compatibility code.

Acceptance status

The frozen three-Domain live cohort has 40 Agent Episodes per Domain. An AI review found 106 pass, 13 fail, and 1 uncertain. It is diagnostic and does not satisfy direct-human dataset acceptance. The operator directed cutover work to proceed under an explicit exception. See current work for the exact decision and remaining work. No unreviewed Episode is labeled human-approved.

Documentation

About

Local-first Python framework for generating, executing, validating, and packaging agent training data with executable state, tool evidence, lineage, replay checks, and release artifacts.

Resources

Security policy

Stars

36 stars

Watchers

2 watching

Forks

Releases

Packages

Contributors

Languages