Disinto

Autonomous code factory — disinto.ai

A mining robot, lost and confused, builds a Disinto from scrap —
a device so powerful it vaporizes three-quarters of a mountain on a single battery.
— Isaac Asimov, "Robot AL-76 Goes Astray" (1942)

Point it at a Codeberg repo with a Woodpecker CI pipeline and it will pick up issues, implement them, review PRs, and keep the system healthy — all on its own.

Architecture

cron (*/10) ──→ supervisor-poll.sh    ← supervisor (bash checks, zero tokens)
                 ├── all clear? → exit 0
                 └── problem? → claude -p (diagnose, fix, or escalate)

cron (*/10) ──→ dev-poll.sh        ← pulls ready issues, spawns dev-agent
                 └── dev-agent.sh   ← claude -p: implement → PR → CI → review → merge

cron (*/10) ──→ review-poll.sh     ← finds unreviewed PRs, spawns review
                 └── review-pr.sh   ← claude -p: review → approve/request changes

cron (daily) ──→ gardener-poll.sh  ← backlog grooming (duplicates, stale, tech-debt)
                  └── claude -p: triage → promote/close/escalate

systemd ──→ matrix_listener.sh   ← long-poll daemon for human replies
              └── dispatches thread replies → supervisor/gardener

all agents ──→ matrix_send()     ← status updates, escalations, merge notifications

Prerequisites

Required:

• Claude CLI — claude in PATH, authenticated
• Codeberg account with an API token — disinto reads issues, opens PRs, posts comments, and merges via the Codeberg API
• A second Codeberg account for the review bot — reviews posted under a separate identity so the dev-agent doesn't review its own PRs (REVIEW_BOT_TOKEN)
• Woodpecker CI — local instance connected to your Codeberg repo; disinto monitors pipelines, retries failures, and queries the Woodpecker Postgres DB directly
• PostgreSQL client (psql) — for Woodpecker DB queries (pipeline status, build counts)
• jq, curl, git

Optional:

• Matrix homeserver (Dendrite or Synapse) — real-time notifications, escalation threads with human-in-the-loop replies
• Foundry (forge, cast, anvil) — only needed if your target project uses Solidity
• Node.js — only needed if your target project uses Node

Setup

# 1. Clone
git clone ssh://git@codeberg.org/johba/disinto.git
cd disinto

# 2. Configure
cp .env.example .env

Edit .env with your values:

# Target repo
CODEBERG_REPO=yourorg/yourproject              # Codeberg org/repo slug
CODEBERG_API=https://codeberg.org/api/v1/repos/yourorg/yourproject
PROJECT_REPO_ROOT=/path/to/your/project        # local clone of the target repo

# Auth tokens
CODEBERG_TOKEN=...          # main account — or put it in ~/.netrc
REVIEW_BOT_TOKEN=...        # separate Codeberg account for code reviews

# Woodpecker CI
WOODPECKER_SERVER=http://localhost:8000
WOODPECKER_TOKEN=...
WOODPECKER_DB_PASSWORD=...
WOODPECKER_DB_USER=woodpecker
WOODPECKER_DB_HOST=127.0.0.1
WOODPECKER_DB_NAME=woodpecker

# Tuning
CLAUDE_TIMEOUT=7200         # max seconds per Claude invocation (default: 2h)

# 3. Install cron (staggered to avoid overlap)
crontab -e
# Add:
#   0,10,20,30,40,50 * * * * /path/to/disinto/supervisor/supervisor-poll.sh
#   3,13,23,33,43,53 * * * * /path/to/disinto/review/review-poll.sh
#   6,16,26,36,46,56 * * * * /path/to/disinto/dev/dev-poll.sh
#   15 8 * * *                /path/to/disinto/gardener/gardener-poll.sh

# 4. Verify
bash supervisor/supervisor-poll.sh   # should log "all clear"

Directory Structure

disinto/
├── .env.example          # Template — copy to .env, add secrets + project config
├── .gitignore            # Excludes .env, logs, state files
├── lib/
│   ├── env.sh              # Shared: load .env, PATH, API helpers, matrix_send()
│   ├── ci-debug.sh         # Woodpecker CI log/failure helper
│   ├── matrix_listener.sh  # Matrix long-poll daemon (dispatches replies)
│   └── matrix_listener.service  # systemd unit for the listener
├── dev/
│   ├── dev-poll.sh       # Cron entry: find ready issues
│   └── dev-agent.sh      # Implementation agent (claude -p)
├── review/
│   ├── review-poll.sh    # Cron entry: find unreviewed PRs
│   └── review-pr.sh      # Review agent (claude -p)
├── gardener/
│   ├── gardener-poll.sh  # Cron entry: backlog grooming
│   └── best-practices.md # Gardener knowledge base
├── planner/
│   ├── planner-poll.sh   # Cron entry: weekly vision gap analysis
│   └── planner-agent.sh  # Updates AGENTS.md, creates backlog issues (claude -p)
├── vault/
│   ├── vault-poll.sh     # Cron entry: process pending dangerous actions
│   ├── vault-agent.sh    # Classifies and routes actions (claude -p)
│   ├── vault-fire.sh     # Executes an approved action
│   ├── vault-reject.sh   # Marks an action as rejected
│   └── PROMPT.md         # System prompt for vault agent
└── supervisor/
    ├── supervisor-poll.sh   # Supervisor: health checks + claude -p
    ├── PROMPT.md         # Supervisor's system prompt
    ├── update-prompt.sh  # Self-learning: append to best-practices
    └── best-practices/   # Progressive disclosure knowledge base
        ├── memory.md
        ├── disk.md
        ├── ci.md
        ├── codeberg.md
        ├── dev-agent.md
        ├── review-agent.md
        └── git.md

Agents

Agent	Trigger	Job
Supervisor	Every 10 min	Health checks (RAM, disk, CI, git). Calls Claude only when something is broken. Self-improving via best-practices/.
Dev	Every 10 min	Picks up backlog-labeled issues, creates a branch, implements, opens a PR, monitors CI, responds to review, merges.
Review	Every 10 min	Finds PRs without review, runs Claude-powered code review, approves or requests changes.
Gardener	Daily	Grooms the issue backlog: detects duplicates, promotes tech-debt to backlog, closes stale issues, escalates ambiguous items.
Planner	Weekly	Updates AGENTS.md documentation to reflect recent code changes, then gap-analyses VISION.md vs current state and creates up to 5 backlog issues for the highest-leverage gaps.
Vault	Every 30 min	Safety gate for dangerous or irreversible actions. Classifies pending actions via Claude: auto-approve, auto-reject, or escalate to a human via Matrix.

Design Principles

• Bash for checks, AI for judgment — polling and health checks are shell scripts; Claude is only invoked when something needs diagnosing or deciding
• Pull over push — dev-agent derives readiness from merged dependencies, not labels or manual assignment
• Progressive disclosure — the supervisor reads only the best-practices file relevant to the current problem, not all of them
• Self-improving — when Claude fixes something new, the lesson is appended to best-practices for next time
• Project-agnostic — all project-specific values (repo, paths, CI IDs) come from .env, not hardcoded scripts