AI NewsProductivityAnnouncement
Stalemap tells a coding agent which dependencies moved since its training cutoff
Stalemap is a Node tool that reads a project's dependencies, compares each version against the model's training cutoff, and writes a short block of breaking changes into AGENTS.md or CLAUDE.md so a coding agent stops producing calls that reference an older API.
Image: GitHub
Why it mattersThe failure mode this fixes is the one that costs the most engineer time with a coding agent today, a confident call to an API the library already dropped or renamed, and the author's benchmark says it moves Claude Sonnet 5.5 from 68 percent to 89 percent on breaking-change tasks.
A coding agent writes code against the APIs it saw during training, and the APIs keep moving. Stalemap is a small Node tool that reads a JavaScript or TypeScript project, diffs the installed versions of each dependency against the version the model would have seen at its training cutoff, and writes the breaking changes into the project's AGENTS.md or CLAUDE.md so the agent stops producing calls that no longer exist.
Umair, a GitHub user with 128 followers, released the project earlier this week. The repository is three days old and has 98 stars. The description on GitHub is one sentence: "Tell your coding agent which dependencies changed after its training cutoff. Diffs types and changelogs, writes a short block into AGENTS.md / CLAUDE.md."
What it reads and what it writes
Stalemap works in three stages. A pre-session step builds the context block in AGENTS.md or CLAUDE.md with the API changes it has found. An edit-time guard runs the project's own tsc --noEmit after the agent writes a file and correlates each compile error to the dependency that moved. An on-demand search_changes command lets the agent query the diff in natural language when it needs a specific function.
The sources are the installed packages' .d.ts files, their release notes and their upgrade guides. The release notes themselves are summarised by Claude at install time, so the AGENTS.md block is short enough for the agent to carry in context. The README limits support to JavaScript and TypeScript projects with one package.json, and lockfiles for npm, pnpm, yarn and bun.
The benchmark the author ran
The README publishes a pre-registered benchmark called Held-out v4. It is 66 tasks sampled from breaking changes released between July and October 2026, each graded on hidden tests the agent cannot see. The agent harness is a headless Claude Code session.
For Claude Sonnet 5.5, the author reports the agent alone passed 68 percent of the tasks, the agent with a plain tsc --noEmit hook passed 73 percent, and the agent with Stalemap passed 89 percent. Claude Sonnet 4.5 is also measured. The author says the tasks were written by a separate session that never saw the Stalemap code, and the build was frozen before any result existed, which is what the pre-registered label means. The confidence intervals come from a two-level bootstrap over tasks and runs.
These are the author's own figures on the author's own benchmark, and the tasks are JavaScript and TypeScript. A reader who runs an agent against those stacks has a tool worth testing on their own repository, with the caveat that the 20-point delta is on breaking-change tasks and tells no story about greenfield code.
Source
Umair/stalemap on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
