devflow-native

mcp
Security Audit
Warn
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 7 GitHub stars
Code Warn
  • fs module — File system access in .github/workflows/release.yml
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Repo-local evidence gates and handoffs for Codex, Claude Code, and shell sessions.

README.md

Devflow Native

Devflow Native README banner

한국어 문서

Stop AI coding agents from saying "done" when your repo has no evidence.

Devflow Native is a repo-local evidence gate, handoff layer, and reviewed
mistake-memory loop for Codex, Claude Code, and shell sessions. It does not
write code for you. It records what agents changed, what they actually
verified, which mistakes they keep repeating, and what the next session should
pick up.

Agent: done.
Devflow: not yet.
- no gate evidence recorded
- no review evidence recorded
- repeated mistake candidate: skipped the repo's PowerShell-safe command rule
Next: run the configured gate or record why it was skipped.

Why This Exists

AI coding agents are getting better at generating code. The next bottleneck is
often trust and continuity:

  • What changed in the last session?
  • Which tests, typechecks, builds, or reviews actually ran?
  • What failed or was skipped?
  • Which repo docs and project rules should the next agent trust?
  • Is it really safe to say the task is done?
  • Did a short maintainer command such as ㄱㄱ or 끝내 mean continue,
    finish, review, or hand off?

Long context, chat history, and session compaction help, but they are not the
same as project-local workflow state. Devflow keeps that state in the repo so a
new Codex, Claude Code, shell, or human review session can resume without
rediscovering everything from scratch.

From Mistakes To Repo Rules

Devflow does not just remember that a session happened. It records repeated
agent mistakes such as shell mismatch, unsafe commands, skipped setup, encoding
issues, path handling failures, or wrong finish claims.

A mistake can stay local as evidence, become a promotion candidate after it
repeats, and only then be reviewed into AGENTS.md, a Devflow skill, or a hook
rule. The next agent session starts with that repo-specific lesson instead of
repeating the same failure.

How It Fits

Devflow is not trying to replace the tools around it.

Tool layer What it owns
Codex / Claude Code Run coding agents inside the repo.
Claude hooks / Codex skills Add host-specific automation and instructions.
Superpowers Teach workflow habits such as TDD, debugging, planning, and review.
TaskMaster-style tools Track tasks and agent work queues.
Devflow Native Record repo-local evidence, block unsafe finish claims, promote reviewed mistake rules, and generate next-session handoffs.

In plain terms: Devflow asks the repo to remember enough evidence that the next
agent does not have to guess where the previous one stopped, and the current
agent cannot honestly claim "done" without proof.

What It Does

  • Creates a .devflow/config.json project contract with gates and review policy.
  • Installs and checks local Codex/Claude plugin, hook, and MCP harness files.
  • Shows repo status, changed files, work/session state, gates, and latest handoff.
  • Records review evidence and gate evidence before work is called done.
  • Provides finish --dry-run to check whether a task can honestly be claimed complete.
  • Generates the prompt the next agent session should continue from.
  • Captures repeated agent mistakes, aggregates repeated observations, and promotes
    durable repo-local rules, skills, or hooks only after review evidence.

Quick Try

The important path is agent-native setup. Devflow ships local harness files for
both Codex and Claude Code: plugin manifests, skills/commands, hooks, and MCP
configuration. A global npm install gives you the devflow CLI; the agent
experience starts when a target repo has the Devflow harness installed and the
agent host is restarted or reloaded if required.

Open Codex or Claude Code in the target repo and paste:

Install Devflow Native for this repository.

Inspect the repo first. Preserve existing AGENTS.md, CLAUDE.md, README, tests,
and project rules. Use npx devflow-native@latest if devflow is not already
installed.

Initialize the Devflow scaffold when missing, install only missing Codex/Claude
harness files, run doctor/status/harness health, and tell me exactly what files
changed and whether I need to restart the agent host.

Maintainer shorthand is part of the intended workflow. The prompt hook maps
short commands into workflow intent and gives the agent the next Devflow actions
to run:

  • ㄱㄱ, 진행해, 계속, continue, next, go -> continue/start from
    devflow status --json and devflow prompt latest.
  • 끝내, 마무리, 완료, finish, done -> run devflow finish --guided
    and follow any review or gate blockers before claiming completion.
  • 다음 세션 프롬프트 줘, 여기까지, handoff -> create a handoff with
    devflow status --json and devflow prompt next.
  • 리뷰, pr, pull request -> request and record review evidence.
  • html, 리포트, 보드 -> inspect status first and generate an artifact only
    when it is explicitly useful.

When a prompt contains mixed signals, Devflow uses this priority:
finish > handoff > review/pr > artifact > continue. The agent still has to
inspect repo state, run required gates, and record review evidence before
claiming that work is complete.

What to expect after Quick Try:

  • devflow works as a CLI through npx or a global npm install.
  • The target repo gets local .devflow/ project state and, when confirmed,
    local plugins/devflow/ harness files for Codex and Claude Code.
  • Codex can install Devflow from this repo's .agents/plugins/marketplace.json
    and load its bundled skills, MCP server, and hooks after restart or a new
    thread.
  • Claude Code can install Devflow from the same repo marketplace and load its
    bundled skills, commands, MCP server, and hooks after restart or
    /reload-plugins.
  • Future sessions in that repo receive compact Devflow context automatically.

Host-native install commands:

codex plugin marketplace add Sungblab/devflow-native
codex plugin add devflow@devflow-native-local

claude plugin marketplace add Sungblab/devflow-native
claude plugin install devflow@devflow-native-local

MCP-only fallback:

codex mcp add devflow -- npx --yes devflow-native@latest mcp stdio
claude mcp add devflow -- npx --yes devflow-native@latest mcp stdio

The fallback exposes tools only. The full Devflow experience requires the
plugin or repo-local harness so skills, commands, and hooks are loaded too.

Manual Fallback

npx devflow-native@latest --help
npx devflow-native@latest init --confirm
npx devflow-native@latest harness install --confirm
npx devflow-native@latest harness health
npx devflow-native@latest status --simple

harness install --confirm keeps generated plugins/devflow/ harness files
local by default by adding them to .gitignore. Use --repo-visible only when
the target repository should commit those plugin files as part of its public
development workflow.

For repeated local use, a global install is still fine:

npm install -g devflow-native
devflow harness health

Global npm install is not the same thing as registering an agent plugin. It
only makes the devflow command available everywhere. Each Codex or Claude Code
environment still needs the Devflow harness/plugin registration that the Quick
Try prompt or devflow harness install --confirm verifies.

To update an existing install:

devflow update
npm install -g devflow-native@latest
devflow --version
devflow harness health

For one-off use without changing a global install, run:

npx devflow-native@latest --version
npx devflow-native@latest update

What Devflow Does Not Do

  • It is not an autonomous coding agent.
  • It does not replace Codex, Claude Code, Superpowers, git, tests, or PR review.
  • It does not treat HTML dashboards or generated artifacts as source of truth.
  • It is not tied to one agent runtime or one workflow methodology.

Core Loop

Codex or Claude Code opens the repo
  -> Devflow session-start hook injects compact repo context

Maintainer says "continue" or "next"
  -> Devflow prompt hook classifies workflow intent
  -> the agent starts from status, active work, and handoff state

Maintainer says "finish" or "review"
  -> Devflow finish flow checks docs impact, gates, risks, and next prompt
  -> completion evidence is recorded in .devflow/state/events.jsonl

An agent repeats a repo-specific mistake
  -> Devflow detects or records the mistake in .devflow/mistakes.json
  -> repeated high-confidence observations become promotion candidates
  -> promote --dry-run shows patch candidates without editing durable files
  -> review evidence gates promote --apply into AGENTS.md, a Devflow skill, or a hook/config rule

Runtime state such as .devflow/state/ and .devflow/next-prompt.md is local
by default. Public project contracts such as .devflow/config.json can be
committed when a repository wants to adopt Devflow as part of its workflow.

Dogfood Smoke

Devflow Native is dogfooded against OpenCairn, a larger Windows/PowerShell
monorepo, to keep the product honest on mature-repo adoption instead of only
greenfield demos.

Latest local smoke, run from C:\Users\Sungbin\Documents\GitHub\opencairn-monorepo:

devflow harness inspect --json
devflow harness health --json
devflow gates run docs-check --work local-work --json
devflow finish --json

Observed result on 2026-05-29:

  • Codex and Claude harness targets reported ready.
  • Harness health reported status: ok; plugin manifests, MCP config, hook
    scripts, and review.required passed.
  • docs-check passed when recorded against local-work.
  • finish kept canClaimDone: false until review evidence was recorded, which
    is the intended guardrail.

This is a smoke record, not a performance benchmark. It proves the current
harness can install, inspect, run hooks, record gate evidence, and block unsafe
finish claims in a real repository.

Common Commands

devflow --help
devflow update
devflow doctor --platform windows-powershell --json
devflow status --simple
devflow finish --guided
devflow prompt latest
devflow mistakes detect --platform windows-powershell --command "node script.mjs << 'EOF'" --stderr "ParserError: Missing file specification after redirection operator." --record --json
devflow mistakes promote --id powershell-bash-heredoc-redirection --target agents --dry-run --json
devflow mistakes review --id powershell-bash-heredoc-redirection --status approved --summary "Repeated PowerShell heredoc correction is repo-relevant." --json
devflow mistakes promote --id powershell-bash-heredoc-redirection --target skill --apply --json
devflow mistakes rules --json
devflow harness health
devflow mcp stdio

Documentation

Repository Structure

packages/core     shared product model, local state, gates, handoff contracts
packages/cli      terminal command surface over core contracts
packages/mcp      MCP handler and stdio transport over the same contracts
packages/adapters agent/session history adapters
plugins/devflow   dogfood Codex and Claude Code plugin drafts
docs              product, architecture, roadmap, examples, and public notes
.devflow          dogfood project contract; runtime state is gitignored

Status

The current v0.1 foundation release includes the npm package, CLI, MCP handler,
repo-local Codex/Claude plugin drafts, hooks, finish guard, and the first
review-gated repeated-mistake promotion loop. Hosted sync, richer artifact
generation, broader adapter coverage, and more detector families are later work.

Research notes, paper drafts, evaluation fixtures, and non-public data live in
a separate private repository. This public repository contains product
implementation and public product documentation only.

Reviews (0)

No results found