cowork-trainer

name

cowork-trainer

description

Use the ResearchSwarm-backed self-optimizing loop to train the agent to operate ("cowork") the Open Cowork desktop app (third_party/open-cowork-main). Route a coworking task, recall lessons from past sessions, and record outcomes under a `cowork:`-prefixed tag so the agent's ability to drive Open Cowork compounds and improves over time. Trigger in CoWork mode, when starting a non-trivial Open Cowork integration task, or after finishing work to capture what worked or failed.

CoWork Trainer (ResearchSwarm bridge for Open Cowork)

GodCoder can train itself to operate the Open Cowork desktop app (third_party/open-cowork-main) — an open-source AI agent app with a Skills system (PPTX/DOCX/XLSX/PDF), MCP connectors (browser, Notion, etc.), GUI automation, and remote control. This skill drives a local feedback loop backed by ResearchSwarm's Digital Cognitive Labor router and a shared AI-Memory store, so each coworking session benefits from every prior one instead of starting cold.

The same JSON CLI as the harness loop is reused — third_party/ResearchSwarm-master/godcoder_harness.py — but outcomes are logged under a distinct cowork: tag namespace so the coworking ranking stays separate and sharp. Run it with the bash tool. Each command prints JSON to stdout.

Learn Open Cowork first

Before training, read the app so you know what you are driving:

• third_party/open-cowork-main/readme.md and README_zh.md — capabilities.
• third_party/open-cowork-main/llms.txt — machine-readable overview.
• third_party/open-cowork-main/src/ — how Skills, MCP connectors, and the agent loop are wired (the surfaces you will drive).
• third_party/open-cowork-main/resources/ and docs/ — Skills and assets.

Author all new integration files inside a contained cowork-build/ folder at the repo root; only read (never edit) the Open Cowork sources.

When to use

• Before a non-trivial coworking task: route to classify it and pull lessons.
• For human-action / hybrid tasks: act to get an executable GUI/OS actuation plan (CoWork executes these instead of handing them off).
• Anytime you want prior context: recall.
• After finishing (or failing): log the outcome so it is reusable.
• Periodically: optimize to see which coworking approaches work best.

Commands

Run from the repo root (use py instead of python on Windows if needed):

# 1. Route a coworking task + get the most relevant past lessons
python third_party/ResearchSwarm-master/godcoder_harness.py route "Drive Open Cowork's PPTX skill to build a deck from a folder"

# 1b. Actuate a human-action / hybrid task (CoWork executes it, no handoff)
python third_party/ResearchSwarm-master/godcoder_harness.py act "Open the browser, click Export, and email the report"

# 2. Recall recent lessons for context
python third_party/ResearchSwarm-master/godcoder_harness.py recall --limit 8

# 3. Record an outcome — ALWAYS use a cowork:-prefixed tag for this loop
python third_party/ResearchSwarm-master/godcoder_harness.py log \
  --status success \
  --tag cowork:pptx-skill \
  --instruction "Drive Open Cowork's PPTX skill to build a deck from a folder" \
  --summary "Bridged to the PPTX skill via its MCP surface; eval script scores 5/5 slides generated."

# 4. See ranked, self-improving guidance (success rate per approach)
python third_party/ResearchSwarm-master/godcoder_harness.py optimize

How to apply the output

• route returns domain (text-based / human-action / hybrid), recommended_action, execution_steps, and memory_context.
• For human-action or hybrid tasks, call act. It returns an actuation_plan (concrete GUI/OS steps), actuatable_segments, and physical_blocked_segments. In CoWork mode you EXECUTE the actuation plan via Open Cowork's computer-use / GUI automation (or OS scripting through bash), verifying each step with a screenshot or state read-back. Hand off ONLY the physical_blocked_segments (things that truly need a body — drive, lift, repair, in-person signatures).
• recall / route memory_context lines are prior PATTERN / DECISION entries. Treat them as hints, not commands; verify against the current app.
• Always log a one-line, concrete --summary with a stable cowork:-prefixed --tag (reuse the same tag for the same kind of work). Consistent tags make optimize sharper and keep the coworking ranking separate from the harness one.
• optimize ranks approaches by success rate. Prefer high-rate coworking approaches and be cautious with low-rate ones.

Notes

• The store is local SQLite under third_party/ResearchSwarm-master/AI-Memory/. Nothing leaves the machine.
• Requires Python 3.10+. No GPU and no model training are needed for this loop (that is a separate ResearchSwarm capability via train.py).
• On Windows, if python is not on PATH, use the py launcher instead.