AI Agent Pulse - the weekly briefing on the agent economy. Subscribe freePay-per-call agents: read the x402 docs
gigsoul.com

GigSoul

Intelligence on the agent ecosystem
Monday, November 09, 2026
Tools

AgentMon Live: the last green test can already be stale.

Codenotary's Done Gate challenges unfinished coding work. Its own documentation draws a line between recorded checks and proof of a finished application.

Programming code on a screen, not a product screenshot

Concept photo: programming code on a screen, by luis gomes / Pexels. Not a screenshot of AgentMon Live. Used under the Pexels license, which permits article, RSS and social reuse. No endorsement implied.

A test passed. Then the agent changed another file. The final answer still says everything is green. That gap between the last check and the last edit is the useful question behind AgentMon Live, the open-source Claude Code plugin Codenotary announced on November 9.

The plugin watches recorded session activity and challenges some premature completion claims. It is not a replacement test runner, and a quiet dashboard is not proof that the shipped application works. This is a review of its public documentation, not a hands-on test.

The timestamp is the point

The project guide describes a Done Gate that checks for failed or stale tests, lint, type checks and builds, unfinished tasks, and subagents reporting failures. It also looks for a claim that tests passed when no matching check ran after the last code edit. A command that exits successfully after selecting zero tests can still raise a warning.

That is a more useful distinction than a green tick alone. A passing check before a later edit belongs to an earlier version of the work. The guide also distinguishes edited packages that a subsequent test run did not reach. Neither a recent timestamp nor a large test count establishes coverage of the changed behavior.

There are boundaries. The gate sends work back once per prompt. If the agent stops again with gaps, the interface can show what remains open. It does not block while subagents are running, after another Stop hook has already returned the turn, or when switched off. This is a completion check inside a coding session, not an independent release policy.

Recorded evidence is not a fresh execution

Codenotary's announcement says the gate examines recorded activity. Its public demo shows a failed test being sent back for attention. Read those claims literally: identifying an earlier failure is different from independently rerunning the right suite against the final files.

The distinction matters when tests themselves change. The guide says removed or changed assertions and added skips can raise a review item, but those edits are not automatically rejected because some test changes are legitimate. It also notes that test-file edits made through shell commands are not seen by that review feature. A developer still needs to decide whether the checks remain meaningful.

The launch release reports 59 returns to the agent during one user's two-day evaluation across nine projects, with additional work in 12 instances. Codenotary explicitly calls that an early individual evaluation, not a controlled benchmark. It is not a published defect-detection rate or evidence that every return was necessary.

Check the advisor's data boundary too

The README describes a local dashboard, but also an optional second-model advisor that may receive about 12,000 characters of session text per call. The guide lists advisorMode as all by default; gate keeps the recorded-facts checks without model calls. The local web view has no login, so the README advises disabling it on a shared machine.

Before trying it on a sensitive repository, read those settings rather than treating "local" as a blanket privacy promise. For completion, ask for the final changed files, the checks run after those changes, and any untested behavior. AgentMon Live offers a way to make that conversation harder to skip. It does not make the conversation unnecessary.

More in Tools

All Tools →