DevClaw

One agent. One group chat. A full dev team.

Add the orchestrator to a Telegram group, point it at a GitLab repo, and it spins up an isolated dev team: a PM (the orchestrator itself) that manages the backlog, a DEV sub-agent that writes code, and a QA sub-agent that reviews it. Add another group chat — another team. Each runs independently with its own task queue, workers, and session state.

DevClaw is the OpenClaw plugin that makes this work.

Why

OpenClaw is great at giving AI agents the ability to develop software — spawn sub-agents, manage sessions, work with code. But running a real multi-project development pipeline exposes a gap: the orchestration layer between "agent can write code" and "agent reliably manages multiple projects" is brittle. Every task involves 10+ coordinated steps across GitLab labels, session state, model selection, and audit logging. Agents forget steps, corrupt state, null out session IDs they should preserve, or pick the wrong model for the job.

DevClaw fills that gap with guardrails. It gives the orchestrator atomic tools that make it impossible to forget a label transition, lose a session reference, or skip an audit log entry. The complexity of multi-project orchestration moves from agent instructions (that LLMs follow imperfectly) into deterministic code (that runs the same way every time).

The idea

One head agent acts as the PM across all your projects. It reads task backlogs, decides priorities, and delegates work. For each task, it spawns (or reuses) a DEV sub-agent to write code or a QA sub-agent to review it. Every Telegram group is a separate project — the PM keeps them completely isolated while managing them all from a single process.

DevClaw gives the PM four tools that replace hundreds of lines of manual orchestration logic. Instead of following a 10-step checklist per task (fetch issue, check labels, pick model, check for existing session, transition label, update state, log audit event...), it calls task_pickup and the plugin handles everything atomically.

How it works

graph TB
    subgraph "Group Chat A"
        direction TB
        A_PM["🎯 PM (head agent)"]
        A_GL[GitLab Issues]
        A_DEV["🔧 DEV (sub-agent)"]
        A_QA["🔍 QA (sub-agent)"]
        A_PM -->|task_pickup| A_GL
        A_PM -->|spawns / sends| A_DEV
        A_PM -->|spawns / sends| A_QA
    end

    subgraph "Group Chat B"
        direction TB
        B_PM["🎯 PM (head agent)"]
        B_GL[GitLab Issues]
        B_DEV["🔧 DEV (sub-agent)"]
        B_QA["🔍 QA (sub-agent)"]
        B_PM -->|task_pickup| B_GL
        B_PM -->|spawns / sends| B_DEV
        B_PM -->|spawns / sends| B_QA
    end

    subgraph "Group Chat C"
        direction TB
        C_PM["🎯 PM (head agent)"]
        C_GL[GitLab Issues]
        C_DEV["🔧 DEV (sub-agent)"]
        C_QA["🔍 QA (sub-agent)"]
        C_PM -->|task_pickup| C_GL
        C_PM -->|spawns / sends| C_DEV
        C_PM -->|spawns / sends| C_QA
    end

    AGENT["Single OpenClaw Agent"]
    AGENT --- A_PM
    AGENT --- B_PM
    AGENT --- C_PM

It's the same agent process — but each group chat gives it a different project context. The PM role, the workers, the task queue, and all state are fully isolated per group.

Task lifecycle

Every task (GitLab issue) moves through a fixed pipeline of label states. DevClaw tools handle every transition atomically — label change, state update, audit log, and session management in a single call.

stateDiagram-v2
    [*] --> Planning
    Planning --> ToDo: Ready for development

    ToDo --> Doing: task_pickup (DEV)
    Doing --> ToTest: task_complete (DEV done)

    ToTest --> Testing: task_pickup (QA)
    Testing --> Done: task_complete (QA pass)
    Testing --> ToImprove: task_complete (QA fail)
    Testing --> Refining: task_complete (QA refine)

    ToImprove --> Doing: task_pickup (DEV fix)
    Refining --> ToDo: Human decision

    Done --> [*]

Session reuse

Sub-agent sessions are expensive to start — each new spawn requires the agent to read the full codebase (~50K tokens). DevClaw preserves session IDs across task completions. When a DEV finishes task A and picks up task B on the same project, the plugin detects the existing session and returns "sessionAction": "send" instead of "spawn" — the orchestrator sends the new task to the running session instead of creating a new one.

sequenceDiagram
    participant O as Orchestrator
    participant DC as DevClaw Plugin
    participant GL as GitLab
    participant S as Sub-agent Session

    O->>DC: task_pickup({ issueId: 42, role: "dev" })
    DC->>GL: Fetch issue, verify label
    DC->>DC: Select model (haiku/sonnet/opus)
    DC->>DC: Check existing session
    DC->>GL: Transition label (To Do → Doing)
    DC->>DC: Update projects.json, write audit log
    DC-->>O: { sessionAction: "send", sessionId: "...", model: "sonnet" }
    O->>S: sessions_send (task description)

Model selection

The plugin selects the cheapest model that can handle each task:

Complexity	Model	When
Simple	Haiku	Typos, CSS, renames, copy changes
Standard	Sonnet	Features, bug fixes, multi-file changes
Complex	Opus	Architecture, migrations, security, system-wide refactoring
QA	Grok	All QA tasks (code review, test validation)

Selection is based on issue title/description keywords. The orchestrator can override with modelOverride on any task_pickup call.

State management

All project state lives in a single memory/projects.json file in the orchestrator's workspace, keyed by Telegram group ID:

{
  "projects": {
    "-1234567890": {
      "name": "my-webapp",
      "repo": "~/git/my-webapp",
      "groupName": "Dev - My Webapp",
      "baseBranch": "development",
      "dev": {
        "active": false,
        "sessionId": "agent:orchestrator:subagent:a9e4d078-...",
        "issueId": null,
        "model": "haiku"
      },
      "qa": {
        "active": false,
        "sessionId": "agent:orchestrator:subagent:18707821-...",
        "issueId": null,
        "model": "grok"
      }
    }
  }
}

Key design decision: when a worker completes a task, sessionId and model are preserved (only active and issueId are cleared). This enables session reuse on the next pickup.

All writes go through atomic temp-file-then-rename to prevent corruption.

Tools

task_pickup

Pick up a task from the GitLab queue for a DEV or QA worker.

Parameters:

• issueId (number, required) — GitLab issue ID
• role ("dev" | "qa", required) — Worker role
• projectGroupId (string, required) — Telegram group ID
• modelOverride (string, optional) — Force a specific model

What it does atomically:

1. Resolves project from projects.json
2. Validates no active worker for this role
3. Fetches issue from GitLab, verifies correct label state
4. Selects model based on task complexity
5. Detects session reuse opportunity
6. Transitions GitLab label (e.g. To Do → Doing)
7. Updates projects.json state
8. Writes audit log entry
9. Returns structured instructions for the orchestrator

task_complete

Complete a task with one of four results.

Parameters:

• role ("dev" | "qa", required)
• result ("done" | "pass" | "fail" | "refine", required)
• projectGroupId (string, required)
• summary (string, optional) — For the Telegram announcement

Results:

• DEV "done" — Pulls latest code, moves label Doing → To Test, deactivates worker
• QA "pass" — Moves label Testing → Done, closes issue, deactivates worker
• QA "fail" — Moves label Testing → To Improve, reopens issue, prepares DEV fix cycle with model selection
• QA "refine" — Moves label Testing → Refining, awaits human decision

queue_status

Returns task queue counts and worker status across all projects (or a specific one).

Parameters:

• projectGroupId (string, optional) — Omit for all projects

session_health

Detects and optionally fixes state inconsistencies.

Parameters:

• autoFix (boolean, optional) — Auto-fix zombies and stale state
• activeSessions (string[], optional) — Live session IDs from sessions_list

Checks:

• Active worker with no session ID (critical)
• Active worker whose session is dead — zombie (critical)
• Worker active for >2 hours (warning)
• Inactive worker with lingering issue ID (warning)

Audit logging

Every tool call automatically appends an NDJSON entry to memory/audit.log. No manual logging required from the orchestrator agent.

{"ts":"2026-02-08T10:30:00Z","event":"task_pickup","project":"my-webapp","issue":42,"role":"dev","model":"sonnet","sessionAction":"send"}
{"ts":"2026-02-08T10:30:01Z","event":"model_selection","issue":42,"role":"dev","selected":"sonnet","reason":"Standard dev task"}
{"ts":"2026-02-08T10:45:00Z","event":"task_complete","project":"my-webapp","issue":42,"role":"dev","result":"done"}

Installation

# Local (place in extensions directory — auto-discovered)
cp -r devclaw ~/.openclaw/extensions/

# From npm (future)
openclaw plugins install @openclaw/devclaw

Configuration

Optional config in openclaw.json:

{
  "plugins": {
    "entries": {
      "devclaw": {
        "config": {
          "glabPath": "/usr/local/bin/glab"
        }
      }
    }
  }
}

Restrict tools to your orchestrator agent only:

{
  "agents": {
    "list": [{
      "id": "my-orchestrator",
      "tools": {
        "allow": ["task_pickup", "task_complete", "queue_status", "session_health"]
      }
    }]
  }
}

Requirements

• OpenClaw
• Node.js >= 20
• glab CLI installed and authenticated
• A memory/projects.json in the orchestrator agent's workspace

License

MIT