Stop your AI coding agent from saying "done" without evidence

Rocky shi — Mon, 02 Nov 2026 05:28:32 +0000

If you use an AI coding agent every day, you have probably seen this loop:

You ask for a bug fix.
The agent starts editing files right away.
A few minutes later it says "Done, the fix works."
You check, and nothing was actually run.

The agent is not lying on purpose. It just has no structure that forces it to say what it is about to do, wait for you, and separate what it read from what it ran.

The three habits that cause most of the pain

Jumping straight to edits. The plan, the guesses and the changes all land in one go. By the time you see the plan, the files are already different.

Mixing plans with guesses. "The root cause is X" sounds the same whether the agent reproduced the bug or just read one file.

Reporting "done" without evidence. A static read of the code and a real test run get reported with the same confidence.

None of these are model problems you can prompt away once. They are workflow problems, so they need a workflow.

What a fix looks like

What I wanted from an agent is boring and specific:

Tell me what you are about to do before you do it.
Stop at the key points and let me confirm.
Keep "checked by reading" apart from "proven by running".
Never push, tag or release on your own.
And please, do not turn a simple question into a ceremony.

That is what task-lifecycle does. It is an open-source, MIT-licensed, skills-only plugin for AI coding agents.

How task-lifecycle works

It routes by intent. Every reply opens with a short pre-check (4 items on a light turn, 7 when it is about to execute). Plain questions, lookups and trivial edits get ordinary help. Only substantive work (a feature, a bug fix or a documentation change) becomes a governed task, and it tells you first.

Three flows with clear stages. Feature work goes through stages A01–A10, bug fixes through B01–B10, docs through DOC01–DOC10. Each formal artifact waits for your confirmation by default.

You pick how deep and how strict. When a formal plan needs it, you choose a depth (fast / standard / thorough / custom) and a review intensity (light / balanced / strict). No option skips the required quality checks.

Honest evidence. Design or static checks are reported separately from real-environment evidence. Anything that was not run is not marked as passed.

You keep the dangerous buttons. There is an opt-in Auto card: after you confirm A01 and A02, one card can run A03–A07. It can pre-authorize at most a local commit. Push, tag, release and final acceptance always stay with you, and the plugin never runs Git on its own.

It also does read-only review of code, pull requests and plans, with findings backed by evidence.

Process notes stay out of your repo. By default they go into a sibling .task-lifecycle/ folder next to the project, not into your delivery.

What it is not

It is one SKILL.md plus reference rules loaded on demand. No server, no background service, no database, no remote API.

It is also not a sandbox or a test runner. What it can do depends on the tools and permissions your host gives the agent. I think it is better to say that up front than to oversell it.

Try it

It is not on any marketplace yet; you install it from the repository. It is tested on Cursor, Codex and Grok Build.

Cursor: place the repository at ~/.cursor/plugins/local/task-lifecycle, run Developer: Reload Window, and check that the task-lifecycle skill shows up under Customize.

Codex:

codex plugin marketplace add devcodex-labs/task-lifecycle
codex plugin add task-lifecycle@devcodex-task-lifecycle

Grok Build:

grok plugin install devcodex-labs/task-lifecycle --trust

A successful install does not prove the host loaded it, so start a new session and send something like:

Use Task Lifecycle to plan a bugfix: reproduce, find the root cause, then propose a fix.

You should see the pre-check first, a note that it is entering a governed Bugfix task, and then a stop at the first point that needs your confirmation.

Over to you

How do you keep your coding agent from claiming success it never verified? Rules files, hooks, a checklist in the prompt? I would like to hear what works for you.

Repo, issues and pull requests: https://github.com/devcodex-labs/task-lifecycle

I post about task-lifecycle and AI coding agent workflows on X: https://x.com/rockyshi_93

I used nodejs+react to build a full-stack development framework called nodestack, which is very simple and fast

Rocky shi — Mon, 22 Jul 2024 09:47:20 +0000

nodestack It is a full-stack framework for JavaScript developers. It is designed to be very simple, allowing you to create high-performance, high-quality programs with minimal code, all for the purpose of improving development efficiency.node、hapi、react、mongodbetc. to build, use esbuild As a compilation tool.

Advantage:

No configuration required
Very simple, small, safe, fast, and ready to use
Supports automatic code updates, no need to restart the server with tools such as nodemon after each code modification.
Compile speed is increased by 20-100 times using esbuild.
Powerful routing function.
Supports ES5\ES6+ at the same time

This is a new project, so if you run into problems, feel free to ask!

You can view the relevant documentation here https://github.com/rockyshi1993/nodestack

DEV Community: Rocky shi