codexforge

Autonomous open-source maintainer agent platform powered by Claude Agent SDK and MiniMax M2.7 via the Anthropic-compatible endpoint.

codexforge helps repository maintainers operate at scale with a team of specialist agents that triage issues, investigate bugs, draft fixes, review patches, and compose release notes — with human approval gates on every write action.

Why codexforge

Large open-source portfolios suffer from backlog sprawl, inconsistent responses, and lost context across repositories. codexforge turns that noise into a structured workflow:

• Unified triage across multiple repositories
• Consistent, high-quality drafts for responses, summaries, and release notes
• HITL approval gates for any destructive or public-facing action
• Full audit trail and observability for every agent step
• Reusable SKILL.md expertise and scenario-based evaluation

Architecture at a glance

codexforge is organized as a small control plane over a team of specialist subagents:

• triage-agent — classify and prioritize issues
• investigator-agent — reproduce bugs and gather context
• coder-agent — propose patches with rationale
• reviewer-agent — assess proposed changes
• scribe-agent — draft responses, release notes, and summaries

The control plane wires them through the Claude Agent SDK runtime, providing hooks, sessions, skills, and MCP server integration on top of MiniMax M2.7.

See ARCHITECTURE.md for details.

Model configuration

codexforge uses MiniMax M2.7 through the Anthropic-compatible endpoint. All runtime components read these environment variables:

export ANTHROPIC_BASE_URL="https://api.minimax.io/anthropic"
export ANTHROPIC_AUTH_TOKEN="<YOUR_MINIMAX_API_KEY>"
export ANTHROPIC_MODEL="MiniMax-M2.7"

No Anthropic API key is required. MiniMax's Anthropic-compatible surface is the officially documented integration path for Claude Agent SDK and related tools.

Install

uv pip install -e ".[dev]"

or with plain pip:

python -m pip install -e ".[dev]"

Quick start

codexforge health
codexforge triage --repo JithendraNara/example --issue 42
codexforge eval

See RUNBOOK.md for operational guidance and DEMO.md for a scripted walkthrough.

Repository layout

.
├── ARCHITECTURE.md
├── EVALS.md
├── RUNBOOK.md
├── DEMO.md
├── pyproject.toml
├── src/codexforge/
│   ├── __init__.py
│   ├── cli.py
│   ├── config.py
│   ├── runtime/
│   │   ├── client.py
│   │   ├── orchestrator.py
│   │   ├── subagents.py
│   │   ├── hooks.py
│   │   ├── permissions.py
│   │   └── session_store.py
│   ├── workflows/
│   │   ├── triage.py
│   │   ├── investigation.py
│   │   ├── coding.py
│   │   ├── review.py
│   │   └── release.py
│   ├── adapters/
│   │   └── github.py
│   └── telemetry/
│       ├── logging.py
│       └── tracing.py
├── skills/
│   ├── issue-triage/SKILL.md
│   ├── bug-reproduction/SKILL.md
│   ├── code-fix-proposal/SKILL.md
│   ├── pr-review/SKILL.md
│   └── release-notes/SKILL.md
├── evals/
│   ├── scenarios.json
│   └── run_eval.py
├── tests/
└── .github/workflows/ci.yml

Safety model

Every write action — GitHub API mutation, file modification, shell command, push — passes through the approval hook chain. codexforge refuses to execute uncontrolled commands, never writes outside configured repositories, and logs every decision for audit.

License

MIT.

Status

This is an active flagship build. The roadmap tracks phase-level progress in EVALS.md and the architecture.