3b-forge

Personal workspace for the 3b cross-agent plugin — built and maintained for Claude Code, Codex, Gemini CLI, and (in principle) any future AI agent host that can read markdown-based skills.

Single Source of Truth layout: the plugin's agent prompts, skill playbooks, and (optional) Python engine live once, and are shared across every host via per-host manifests. No sibling duplication.

Supported platforms

macOS is the only tested platform for the installer and shell-based tooling (installer/setup.sh, scripts/check-3b-drift.sh, plugins/3b/skills/check-symlinks). Linux is likely to work — the scripts are bash + POSIX tools — but it is not part of the test matrix yet. Windows is not supported: the installer relies on bash, ln -s symlinks, and $HOME/.claude layout assumptions. WSL may work by proxy but has not been verified.

Platform-specific install payloads (e.g., .ps1 scripts, junction helpers) are not shipped. See installer/README.md for install prerequisites.

What ships today

Path	Purpose	Status
plugins/3b/	The consolidated 3b plugin. Ships /3b:interview (Socratic requirement interview) with a conversational layer (SKILL.md + 7 agent prompts) AND an optional Python engine (engine/) for programmatic integrators.	v0.0.1 — pre-release
todos.md	Near-term backlog for the forge.	active
CHANGELOG.md	Release history.	active

Layout

3b-forge/
├── plugins/
│   └── 3b/                          # the consolidated cross-agent plugin
│       ├── .claude-plugin/          # Claude Code manifest
│       ├── .codex-plugin/           # Codex manifest
│       ├── skills/interview/        # /3b:interview SKILL.md + tool mappings
│       ├── agents/                  # SSoT role prompts (7)
│       ├── commands/                # slash-command stubs
│       ├── engine/                  # optional Python layer (interview_plugin_core)
│       └── README.md
├── todos.md                         # near-term backlog
├── CHANGELOG.md
├── LICENSE                          # MIT
└── README.md                        # this file

Why ONE plugin with two layers, not two sibling plugins

The plugin keeps both prompt-heavy and engine-heavy surfaces inside one installable as two layers:

• Conversational layer — skills/interview/SKILL.md + the 7 prompts under agents/. Zero runtime deps. Works on any AI agent host that reads markdown skills. This is the default path.
• Programmatic layer — engine/ holds a Python package (interview_plugin_core) for CLI / server / automation integrators who need numeric ambiguity scoring (0–1 scale, 40/30/30 weighting), file-locked InterviewState persistence, and a pluggable LLMAdapter protocol. 60+ pytest-asyncio tests. Loads prompts from the shared agents/ directory — no duplication.

Roadmap

v0.0.1 → v0.1.0 (current focus)

Graduation criteria documented in plugins/3b/README.md. Outstanding items:

• Validate perspective-rotation decision table in SKILL.md §B.6 via golden transcripts (greenfield + brownfield).
• Add per-dimension observable-signal rubric to seed-closer.md.
• Document session-continuity transcript convention (path under projects/*/actives/ + frontmatter schema).
• 2+ golden transcript fixtures under plugins/3b/fixtures/.
• Cross-host install flow verified end-to-end on Claude Code + Codex.

Post-v0.1.0 (speculative)

• Additional /3b:<skill> members as they emerge from practice — candidates: /3b:simplify, /3b:wrap, /3b:review, /3b:brainstorm. Each new skill joins plugins/3b/skills/<name>/, reuses the existing agents/ where relevant.
• Publish interview_plugin_core wheel to PyPI with force-include bundling agents/ into the wheel so pip-installed users don't need the filesystem SSoT path.
• Optional MCP wrapper around engine/ for Claude Code plugin hosts that want server-side persistence; would load the same engine, no new code.

Install

Pre-release. Install paths below are the intended shape; verify before publishing.

Claude Code users:

claude plugin marketplace add brandonwie/3b-forge
claude plugin install 3b

Codex users: discovery via .codex-plugin/plugin.json inside plugins/3b/. Concrete command varies by Codex CLI version.

Gemini CLI: plugin format still evolving; best-effort support via the portable SKILL.md format.

Programmatic (Python) integrators:

cd plugins/3b/engine
uv sync --extra dev
uv run python -m pytest -q

See plugins/3b/engine/README.md for the LLMAdapter contract and the round-driving loop.

Upstream

The interview skill forks from Q00/ouroboros. Upstream carries the original Socratic methodology, five-perspective model, and numerical ambiguity-scoring design. The 3b consolidation adds the ontologist perspective, the cross-agent manifest layer, and the two-layer architecture.

Contributing

Personal workspace. Issues welcome for discussion; direct PRs are not the expected contribution pattern.

License

MIT. See LICENSE.

See also

• Ouroboros — upstream source.
• 3b (private, personal) — the knowledge system this harness is adjacent to.