DEV Community

Shaolin Zhou
Shaolin Zhou

Posted on Edited on

Vibe Coding made code cheap. It didn't make "done" trustworthy.

Vibe Coding made code cheap. It didn't make "done" trustworthy.

When AI agents write most of your code, "it's finished" becomes a claim you have to verify. I ran into this while building a multi-physics simulation engine with AI coding agents: I had no reliable way to see which "done" features actually had verification evidence behind them.

Tasks lived in chat context that scrolled away. A new idea mid-session had nowhere to land. And nothing connected a feature to the test or benchmark that was supposed to prove it works.

That is a project management problem, and engineering projects solved it long ago with task hierarchies, state machines, and traceable verification. So I applied that discipline to agent-driven development and built Lite ADLM.

What it is

Lite ADLM is a local-first task tree that humans and AI agents work on together:

  • Plain JSON files on disk. No database.
  • A pure-Python CLI that uses only the standard library, so there is nothing to install.
  • An optional real-time board (Node.js + Socket.IO), so you and the agent see the same data. You can add or reprioritize tasks while the agent works, and it sees them on its next query.
  • Five lists: architecture, features, tests, benchmarks, bugs.
  • Relations between items: depends-on, verifies, blocks.
  • A gaps command that audits which features have no recorded verification.

Agent writes tasks, board updates in real time

Try it in 30 seconds

You need Python 3 (I tested on 3.12). Nothing else.

git clone https://github.com/alainpdeai/lite-adlm-en.git lite-adlm
cd lite-adlm
python scripts/litetask_cli.py init
python scripts/litetask_cli.py add --category func --title "First Task" --agent "AI Agent"
python scripts/litetask_cli.py list --category func
Enter fullscreen mode Exit fullscreen mode
[ ] First Task  P=Medium  S=open
Enter fullscreen mode Exit fullscreen mode

Now ask the audit what's missing:

python scripts/litetask_cli.py gaps
Enter fullscreen mode Exit fullscreen mode
GAP Feature not verified by test: 1 item(s)  (owner=func referrer=test relation=verifies)
   - [func] First Task
Enter fullscreen mode Exit fullscreen mode

To close the gap, add a test item and link it to the feature with a verifies relation:

python scripts/litetask_cli.py add --category test --title "Verify First Task" --agent "AI Agent"
python scripts/litetask_cli.py link --category test --title-or-id "Verify First Task" \
  --relates func:<func_id>:verifies
Enter fullscreen mode Exit fullscreen mode

Run gaps again and that feature no longer appears.

What the audit found on my own project

I tracked 220+ task items (including subtasks and archived ones) while building the simulation engine. When I ran gaps on the active task tree, it listed 69 items (including subtasks) with no recorded verification link. 37 of them were already marked done.

That does not mean the code was untested. It means the evidence wasn't traceable: nothing in the task graph connected those features to a test or benchmark. Before the audit, that was a hidden assumption. After it, it was a visible backlog.

What it doesn't do (honest limitations)

  • The verification rule is advisory, not enforced. A rule in the skill file tells the agent not to close a feature until a test or benchmark links to it. The tool itself allows open → done directly. Illegal jumps such as open → verified are rejected, but the "no verification, no done" part relies on the agent following the rule, and gaps audits it afterwards. That is exactly why the audit command exists.
  • It uses a local trust model. The server binds to 127.0.0.1 by default, and the data API trusts caller-supplied identity. Don't expose it to an untrusted network.
  • It's not a Jira or GitHub Issues replacement. It's for one developer (or a small group) working locally with agents, not for enterprise permissions or managed cloud collaboration.
  • I've only tested it with Claude Code. The Cursor and Aider configs in the README are example templates I haven't verified.

Design choices

  • Files over a database, so the task tree is something you can read, diff, and back up.
  • Standard library only for the CLI, so it works anywhere Python does.
  • Optimistic concurrency on the HTTP API (baseRev and 409 Conflict) to prevent lost updates, and atomic temp-file replacement in CLI mode to prevent corrupted files.
  • A test suite you can run yourself. There are 13 fast graders; with init run and the server started, they pass 13/13 on my machine.

Feedback welcome

If you run long sessions with Claude Code, Cursor, or any agent that can run shell commands, I'd love to hear what feels clunky, especially around connecting your agent and the verifies workflow.

Top comments (0)