devflow

---

The Problem

AI coding agents start every session with amnesia. They don't remember what you decided yesterday, can't review their own code before pushing, and have no isolation between concurrent tasks. You end up babysitting: re-explaining context, manually running checks, juggling branches, and losing hard-won decisions to the void between sessions.

The Solution

devflow composes 6 independent tools into one CLI that runs alongside your AI agent. One command (devflow init) sets up persistent memory across sessions, git worktree isolation per feature, local AI-powered code review, process discipline via skills, session orchestration, and self-hosted tracing — all running on your machine, nothing phoning home. You stay in control. Agents get better context, guardrails, and memory.

What Makes It Different

• Memory that persists across sessions — Hindsight's 3-tier memory (mental models, observations, facts) means your agent recalls past decisions, patterns, and mistakes without you repeating them. 29 MCP tools for recall, retain, and reflect.
• One command, six layers — devflow init installs tools, configures MCP servers, registers hooks, sets up skills, and seeds memory. No manual wiring. Idempotent — safe to re-run.
• Zero build dependencies — Pure Bash CLI. No Node, Python, or Go build step. Works with what's already on your macOS dev machine.
• Agent-agnostic — Works with Claude Code, OpenCode, or any tool that reads CLAUDE.md and speaks MCP. Skills and hooks adapt to your CLI.
• Local-first, privacy-first — Memory, code review, and observability all run on your machine. Langfuse is self-hosted. No data leaves your laptop.

---

Architecture

devflow orchestrates 6 independent layers. Each tool works standalone; devflow wires them together.

Layer	Tool	What It Does	Runtime
1	Hindsight	3-tier persistent memory via MCP (mental models, observations, facts)	Local daemon (uvx)
2	Agent Deck	TUI session wrapper with Conductor auto-monitoring, web dashboard	Homebrew
3	Worktrunk	Git worktree lifecycle — wt step copy-ignored eliminates cold starts	Homebrew
4	Code Review	Local pre-push AI review via individual markdown check rules	claude / opencode
5	CLAUDE.md + Skills	Process discipline baked into agent config. 18 slash commands	Files
6	Langfuse	Multi-agent tracing, MCP call spans, cost tracking. Self-hosted	Docker

graph TD
    CLI["devflow CLI"]

    subgraph L1 ["Layer 1 — Memory"]
        HS["Hindsight<br/>:8888 API · :9999 UI"]
    end

    subgraph L2 ["Layer 2 — Sessions"]
        AD["Agent Deck"]
        COND["Conductor<br/>auto-monitor"]
        WEB["Web UI<br/>:8420"]
    end

    subgraph L3 ["Layer 3 — Isolation"]
        WT["Worktrunk<br/>git worktrees"]
    end

    subgraph L4 ["Layer 4 — Code Review"]
        CR["devflow check<br/>.devflow/checks/*.md"]
        RV["devflow review<br/>local or PR/MR URL"]
    end

    subgraph L5 ["Layer 5 — Process"]
        SK["18 Skills<br/>slash commands"]
        HK["3 Hooks<br/>auto-guards"]
    end

    subgraph L6 ["Layer 6 — Observability"]
        LF["Langfuse<br/>:3100 UI"]
    end

    CLI --> HS
    CLI --> AD
    CLI --> WT
    CLI --> CR
    CLI --> RV
    CLI --> SK
    CLI --> LF
    AD --> COND
    AD --> WEB
    HK -->|"guards"| SK

    classDef mem fill:#7c3aed,color:#fff,stroke:#5b21b6
    classDef sess fill:#3b82f6,color:#fff,stroke:#1e40af
    classDef iso fill:#059669,color:#fff,stroke:#047857
    classDef rev fill:#d97706,color:#fff,stroke:#b45309
    classDef proc fill:#be185d,color:#fff,stroke:#9d174d
    classDef obs fill:#0891b2,color:#fff,stroke:#0e7490
    classDef cli fill:#374151,color:#fff,stroke:#1f2937
    classDef cond fill:#f59e0b,color:#fff,stroke:#d97706

    class HS mem
    class AD,WEB sess
    class WT iso
    class CR,RV rev
    class SK,HK proc
    class LF obs
    class CLI cli
    class COND cond

Development Workflow

From feature request to merged PR — the full lifecycle managed by devflow: