VibeFlow OS
English · Français
You talk. The team ships. You validate.
A Claude Code plugin for spec-driven agentic development. One agent reads your intent, a team
of agents challenges, plans, builds and proves the work, and machine gates check it.
Nothing is done without proof, nothing merges without you.
How it works · Proof · Team · Memory · Labs · Install · Manual
https://github.com/user-attachments/assets/31709de8-a2c0-4e7c-b78e-8328f57e3f43
Install
claude plugin marketplace add picmakpro/vibeflow-os
claude plugin install vibeflow
Then, inside Claude Code, run /vibeflow-install: it detects the scope, lets you pick modules,
and recaps every dependency before writing anything. Update with /vf-update. Codex and
kimi-code are covered below; details in INSTALL.md.
New here? The User Manual walks you through installing, understanding and
running VibeFlow, without ever opening .planning/ or docs/.
How it works
Say "add Google sign-in". vibeflow-head reads the intent and launches the team. The request
is challenged with questions until it is clear, then it becomes a spec, a plan with tests first,
atomic commits, and a verdict from a judge that did not write the code. Every step leaves a file
on disk, so a session can die without the project losing ground. The last step is yours.
→ The cycle, step by step
Done means proven
A green that cannot turn red proves nothing. Tests are written red first, check-*.sh gates run
in CI, and the judge is never the author. The repo holds itself to the same rule: 116 suites run
on every push, and a "fresh lab" job installs the baseline into an empty lab that must pass its
own gates with zero intervention.
→ The machine gates
A team, not a chatbot
"Do steps 3 to 5, I'm back tomorrow." Past a threshold, a mission manager takes over on the
team-kernel: it plans the battle as a graph, runs workers in parallel and reads typed reports
instead of prose. Anything that challenges intent or security freezes the node until you
decide, even at 3 AM.
→ How a long mission holds together
Memory that holds
The context window fills and degrades; the disk does not. PROJECT.md, ROADMAP.md,
STATE.md, indexed registries and agent memory carry the project from one session to the
next. Every session restarts from a fresh disk, never from a compacted context, and every token
has a job: sonnet workers, digests, on-demand loading.
→ Anatomy of an installed lab
A lab for any domain
VibeFlow is not dev-only. It builds labs, governed workspaces for design, content, growth,
business or your own domain, on the same kernel and the same gates. /vf-new-lab creates one,
and its skills are built by skill-creator with an eval loop.
→ What is a lab?
Modules
18 modules, each versioned with its own CHANGELOG.md. conductor is the mandatory baseline;
then you pick a dev lab from a preset (dev, dev-mobile, dev-audite, dependencies resolved
for you) or a tailor-made domain lab. Each module's README is its full documentation, with the
same structure everywhere.
→ Module catalog ·
commands ·
skills ·
agents
Other runtimes
Claude Code is the reference. Install and usage are also measured end to end on two others:
| Runtime | Install | What you give up |
|---|
| Codex | codex plugin marketplace add picmakpro/vibeflow-oscodex plugin add vibeflow@vibeflow-os | Hooks. Needs multi_agent_v2 and a trusted repository, both fail silently otherwise; set VF_RUNTIME=codex if claude is also installed. |
| kimi-code | No install channel: load an agent with kimi --agent-file <path> | Hooks. |
Every loss is declared at install time rather than hidden. Full list:
Installing outside Claude Code.
Trust
- Source-available: public code and history, see LICENSE.
- Auditable: bash +
jq, every script covered by its suite, idempotent install with a
backup before any overwrite.
- Zero plugin-level hooks: nothing runs until you invoke it. Module governance hooks are
written by
/vibeflow-install, in plain sight.
- Anti-hallucination by design: routing relies on a factual index generated from disk;
incomplete modules are flagged
proposable:false, never sold.
Releases
Semver per module, and a tagged root version on every release (check-release-tag is a gate).
The three latest are below; the full history lives in CHANGELOG.md.
| Version | Date | Change |
|---|
v2.73.1 | 2026-10-29 | Lock guard judged in the worktree, gsd-run.sh rc 64 (patch). guard-driver-lock.sh resolves the directory a cd <worktree> && git commit … lands in and judges that checkout's index instead of the main tree's untracked files (same-repository checkouts only; ;, subshell, variable, ` |
v2.73.0 | 2026-10-29 | Product bundle (minor). New module product-team-bundle v1.0.0: product team (manager, BRIEF analyst, PRD PM, prd-gate judge with threshold 80, check-prd.sh gate), vibeflow-product front door and /vf-product skill, visible to the product and solo roles. Stays proposable: false until a real run. planning-core v2.9.0 (product template), dev-orchestrator v2.29.0 (the head points to /vf-product at the product role). |
v2.72.0 | 2026-10-29 | GitHub issues closure (minor). driver-lock works under uutils coreutils (Ubuntu 26 runner: ln -sn, mv -T before the BSD fallback); framework-version.sh drift compares the lab to the plugin cache instead of to itself; vf-optional-agents frontmatter key for agents of optional modules; gsd-run.sh launches gsd-tools by path (rc 64 without an engine) and ensure-deps.sh derives the project scope from ./.claude/gsd-core; managers no longer busy-wait on workers; reindex.sh keeps the registry mode; sourced libs are no longer counted as non-executable; persistent design-plugin exclusion (VF_DESIGN_SKIP, .vf-design-skip); the install engine backs up and reports overwritten local patches and announces the settings.local.json routing (conductor v1.50.3, dev-orchestrator v2.28.5, consolidator v1.11.4, infrastructure-audit v1.3.7, design-orchestrator v1.5.16, mobile-test v1.0.5, 10 doc patches). Closes #47, #52, #81, #165, #167. 114 suites. |
v2.71.2 | 2026-10-23 |
- ADR-032 — 4-pillar memory consolidation system
- ADR-035 — AI-Safe software architecture doctrine
- ADR-053 — Swarm layer: driver lock + DAG + typed reports
- ADR-055 — Altitude boundary lab planning ↔ dev engine
- ADR-056 — Runtime support vigilance
- ADR-057 — Tooled boundaries with third-party bricks
- LRN-101 — "Minimal agent + composable skills" pattern
- LRN-106 — Audit before fix
Main lab (private): vibeflow-lab — structural changes are tested there before release.
Authors
- @picmakpro — creator of the VibeFlow methodology and repo owner. Laid the project's foundations and remains the guardian of its doctrine: the governance backbone (
conductor, planning-core, consolidator), the scriptural hooks and guards that keep every lab honest. VibeFlow's identity — governance enforced by tools, not by prose — is his.
- Samuel Neveu — @samuel-neveugall — the project's driving force day to day: main contributor and release driver. Built the entire development side (
dev-orchestrator, design-orchestrator, mobile-test-team), led the agentic pivot and the team-kernel, and steers the framework's evolution — including its migration onto the @opengsd/gsd-core engine.
Contributing
Contributions are welcome — no need to be a collaborator: fork, branch, open a pull request.
Read CONTRIBUTING.md (workflow, tests, licensing of contributions), the
Code of Conduct, and report vulnerabilities privately as described in
SECURITY.md.
License
Source-available under a proprietary license — see LICENSE. Public code and
history; training students get a private-reuse right; redistribution and resale prohibited.
The skill-creator module reuses original Anthropic content under the MIT license.
