Vibeflow

Vibeflow

Stop vibe-coding blind. Spec it, then ship it.

AI agents write code fast — but without specs, they write the wrong code fast. Vibeflow adds a thinking layer before coding: you define what to build with specs, guardrails, and quality gates. The agent implements following your project's real patterns.
Works with Codex, Claude Code and Hermes Agent through one shared plugin, plus Cursor (rules + skills) and GitHub Copilot (prompts + agents).

3 commands to start

Run analyze once, then gen-spec → implement for each feature. That's it.

The pipeline
Think before you code. Build from specs. Verify against DoD.

analyze → discover → gen-spec → implement → audit

Utility: quick (fast-track small tasks), teach (update knowledge base), stats (audit trends, plus the loop's numbers when you use it).

The execution loop (opt-in)

implement --loop splits the work between sessions that cannot grade
themselves: the session you call becomes the orchestrator, an executor
implements each part in a fresh session, an independent audit checks it in
another one, and at most two fix rounds follow, each re-audited on its own
diff. With your OK, asked once per feature, every round is committed on the
feature branch, and the loop ends in company mode: a pull request whose body
carries the evidence (verdicts, DoD, findings, probes, commands run), opened
after the suite runs in the CI's environment when your machine differs from it
(when that run is not possible, the PR opens anyway and says why). The agent
never merges, deploys, migrates a shared environment or reads production.

The loop is off by default and runs only when you turn it on:
implement --loop for one run, or loop: true in .vibeflow/loop/config.md
for the project (--no-loop turns it off for one run). Requirements:

- The…