djobs 0.8.1

Crash-proof task memory for Claude Code, Cursor, Cline, Copilot agents via MCP.

djobs

Crash-proof workflow state for AI coding agents. djobs gives Claude Code, Cursor, Cline, Copilot, and other AI coding agents durable, resumable task memory over MCP — so long, multi-file work survives IDE crashes, context loss, or session interruptions instead of being lost or redone. It ships MCP tools, agent instructions, and a VS Code sidebar; the runtime installs and manages itself.

---

The Problem

AI coding agents (GitHub Copilot, Cursor, Cline, Claude Code) often run multi-file tasks: add docstrings to 40 files, migrate a framework version, batch-refactor an API. These can take minutes.

When the IDE crashes, the chat disconnects, or you accidentally close the window — all in-flight progress is lost. The agent's state lives only in chat history. You have to start over and guess which files were already done.

The Fix

djobs gives your agent three tools that solve this:

Tool	What it does
enqueue_task	Save each file as a durable task — survives any crash
complete_task	Mark a file done after the agent edits it
resume_session	On next chat, find all unfinished files instantly

You: "Add docstrings to all 20 files in src/"

  Agent calls enqueue_task for each file  ← checkpoint saved
  Agent edits file 1 → complete_task      ✅
  Agent edits file 2 → complete_task      ✅
  ...
  Agent edits file 12 → complete_task     ✅
  💥 VS Code crashes

You reopen VS Code, start a new chat: "hi"

  Agent calls resume_session              ← finds 8 incomplete tasks
  Agent edits file 13 → complete_task     ✅
  ...
  Agent edits file 20 → complete_task     ✅
  Done — zero files lost, zero files re-done.

Everything is stored in a local SQLite file. No Redis, no Docker, no cloud service.

How is this different from Celery / RQ / Dramatiq? Those are general-purpose task queues built for backend workers and high throughput. djobs is purpose-built for AI coding agents: it speaks MCP natively, optimizes for crash-recovery and human-inspectable audit trails over raw throughput, and runs with zero infrastructure — one local SQLite file, no broker, no daemon.

Maturity — early but tested. 291 passing tests, CI across Python 3.11–3.13, SQLite and optional PostgreSQL backends. Marked Alpha while the public API stabilizes; the core enqueue → complete → resume flow is stable and used daily.

---

Quick Start

djobs is workflow state for your AI agent — not a dependency of your app. It works in any repo (Python, JS, Go, Rust, docs) because the queue is a tool the agent uses, not a library your project imports. djobs installs and manages its own runtime; you just pick how to set it up.

1. VS Code / GitHub Copilot (recommended)

Install the djobs extension from the Marketplace, then run djobs: Set up / Repair djobs from the Command Palette (or click Set up djobs when the extension offers).

That one step installs the runtime, wires the MCP server, installs the agent instructions, and adds the task sidebar. No terminal, no manual config.

2. Any MCP agent (Claude Code, Cursor, Cline, …)

One command wires the current project for any MCP-compatible agent:

djobs init

It writes .vscode/mcp.json, installs the agent guidance block in .github/copilot-instructions.md, runs djobs doctor, and prints next steps. It auto-detects the right interpreter, so the wiring works even in a JavaScript, Go, or Rust repo with no Python environment.

djobs is not only an MCP tool. It also installs agent instructions so coding agents proactively call resume_session, enqueue_task, complete_task, and fail_task during long or risky work — you don't have to remember to tell them.

Installing the djobs runtime (only if setup asks)

The VS Code extension installs and repairs the runtime for you. If you are setting up outside VS Code, install it once — like git or ripgrep, globally, not per project:

pipx install djobs   # isolated global install (recommended)
# no pipx? ->  pip install djobs   (or: python -m pip install --user djobs)

Then run djobs init in any project.

Granular commands

djobs init is the recommended path, but each step is also available on its own:

djobs install-mcp           # write only .vscode/mcp.json
djobs install-instructions  # write only the agent guidance block
djobs doctor                # diagnose an existing setup

Verify the setup at any time:

djobs doctor
# [OK  ] djobs package: v0.7.3 ...
# [OK  ] queue db (global default): ~/.djobs/global.db — exists, writable
# [OK  ] mcp.json wiring: command='djobs-mcp' — found
# [OK  ] agent guidance block: present in .github/copilot-instructions.md

Options and manual setup

# One-command setup with full auto-approve (agent can enqueue/complete/fail without prompts)
djobs init --full-approve

# Re-run setup, overwriting an existing mcp.json
djobs init --force

# Also write .agent.md (for agents that read it) in addition to copilot-instructions
djobs init --instructions-target all

# Just the wiring, safe default (read-only tools auto-approved)
djobs install-mcp

# Or wiring with full auto-approve
djobs install-mcp --full-approve

# Just the agent guidance block (no mcp.json changes)
djobs install-instructions                 # -> .github/copilot-instructions.md
djobs install-instructions --target agent-md  # -> .agent.md
djobs install-instructions --target all       # -> both
djobs install-instructions --print            # print the block, write nothing

Or add to .vscode/mcp.json manually. After a global install the djobs-mcp console script is on your PATH, so the wiring is identical on every OS:

{
  "servers": {
    "djobs": {
      "type": "stdio",
      "command": "djobs-mcp",
      "autoApprove": [
        "health", "resume_session", "check_task", "list_tasks", "audit_log"
      ]
    }
  }
}

Per-project venv (portable, for repos that commit mcp.json)

If you'd rather keep djobs inside each project's virtual environment — e.g. so a checked-in mcp.json resolves to whatever .venv each collaborator has — run djobs install-mcp --portable to emit a relocatable interpreter hint:

{
  "servers": {
    "djobs": {
      "type": "stdio",
      "command": "${workspaceFolder}/.venv/Scripts/python",
      "args": ["-m", "djobs.mcp_server"],
      "autoApprove": [
        "health", "resume_session", "check_task", "list_tasks", "audit_log"
      ]
    }
  }
}

On macOS / Linux the path is ${workspaceFolder}/.venv/bin/python. You can also pin any interpreter explicitly with djobs install-mcp --python /path/to/python.

Security note: The default autoApprove list only includes read-only tools. If you want your agent to enqueue/complete/fail tasks without confirmation prompts, add "enqueue_task", "complete_task", and "fail_task" to the array — but understand that this allows the agent to mutate queue state without asking.

---

Making Your Agent Use djobs Automatically

After installing, your agent has the MCP tools available — but it won't use them unless you tell it to. Add the following to your agent instructions (e.g. .github/copilot-instructions.md or any .agent.md):

At the start of every session, call resume_session to find unfinished work.
For multi-file tasks, enqueue each file as a durable task and call complete_task after each edit.

What this gives you:

• On every new chat, unfinished work is automatically surfaced
• For multi-file tasks (>3 files), each file is tracked as a durable task
• After editing each file, progress is recorded
• If a session crashes, the next chat auto-resumes from where it stopped — no questions asked

You can also use the djobs install-instructions CLI command to add the guidance block automatically.

---

See It In Action

The animated SVG above shows the full demo. To run it yourself:

pip install djobs
# if you cloned the repo:
python examples/run_migration_demo.py

20 files enqueued → 12 completed → crash → resume → 8 remaining finished. Zero data loss.

---

What Else Can It Do?

Beyond the three core tools, djobs also provides:

• audit_log — "What did the AI do yesterday?" Full event history across sessions.
• check_task / list_tasks — Inspect individual tasks or list by workspace.
• health — Queue depth by status at a glance.
• djobs doctor — One-shot setup check: confirms djobs is installed, the queue DB is writable, and .vscode/mcp.json is wired correctly. Run it (or "djobs: Diagnose Setup" in VS Code) whenever something feels off.
• Multi-agent coordination — Several agents share one queue: claim_task (atomic, exclusive), heartbeat_task, release_task, task dependencies (depends_on), resource locks (resource_key), and an agent registry (register_agent / agent_heartbeat / list_agents).
• Web dashboard — djobs dashboard serves a read-only, cross-agent view of queue health, every task, and the live agent fleet at http://127.0.0.1:8787 (stdlib only, no extra deps). Local-only by design: no authentication, binds to 127.0.0.1; for remote access use an SSH tunnel rather than exposing a public interface.
• Retry with backoff — Failed tasks can retry automatically.
• Dead letter queue — Tasks that exhaust all retries are preserved for review.

For the full architecture, Python library API, PostgreSQL backend, configuration reference, and comparison with other tools, see docs/INTERNALS.md.

---

Development

git clone https://github.com/jhuang-tw/djobs.git
cd djobs
python -m venv .venv && .venv/bin/activate
pip install -e ".[dev]"

pytest -q              # 291 tests (18 skipped without Postgres)
ruff check src/ tests/ # lint

See CONTRIBUTING.md for guidelines.

---

VS Code Extension

djobs includes a VS Code sidebar extension for visual workflow control:

• Workflow dashboard — tasks grouped by workflow and action type with progress indicators
• Start tracked workflow — copies an agent prompt that resumes first and creates durable djobs tasks before multi-step edits
• Resume from any step — expand a workflow, pick a file, and the previous steps are auto-accepted
• Skip / Archive — mark tasks done without editing, or archive stale workflows
• Evidence trail — see what the agent actually changed in each completed task

Install the VSIX from vscode-ext/ or build it yourself:

cd vscode-ext && npm install && npm run package
# Install: code --install-extension djobs-X.Y.Z.vsix

---

Roadmap

• Durable workflow state (enqueue → complete → resume)
• Audit trail — "what did the AI do?"
• MCP server with 14 tools
• pip install djobs && djobs install-mcp — two-command setup
• Published on PyPI
• complete_task evidence field — agent records what it changed
• VS Code sidebar — workflow dashboard, resume from any step, skip/archive
• Start Tracked Workflow prompt — nudges agents to resume/enqueue before editing
• CLI workflow control — djobs skip, djobs accept-before, djobs archive-workflow
• Multi-agent coordination — shared-queue claim, dependencies, resource locks, agent registry
• Web dashboard — djobs dashboard cross-agent global view
• Published on VS Code Marketplace
• Status bar badge + notification alerts

---

License

MIT