since-cutoff – skill Lập trình cho agent AI

since-cutoff: dependency API changes after the model's training cutoff

Overview

A coding model learns each library's API as it was at its training cutoff, but the project's lockfile keeps moving. This skill runs the since-cutoff CLI (MIT, PyPI since-cutoff, pinned here to release 0.4.1). For each pinned dependency, the CLI takes the latest release on or before the model's cutoff and compares that release's public API with the pinned release. The comparison is static (griffe), so no package code runs. It then reports the changed APIs that the project's code uses, with the files that use them. If the user agrees, it writes short notes into AGENTS.md or CLAUDE.md, each tagged with the evidence behind it (for example [diff]). scan makes no model calls and needs no API key. This adapts the [upstream skill](https://github.com/MohammadHijjawi97/since-cutoff/tree/f2108c13c112dce0ca3597b5a2b79ca01afa7b80/skills/since-cutoff) under its [MIT license](https://github.com/MohammadHijjawi97/since-cutoff/blob/f2108c13c112dce0ca3597b5a2b79ca01afa7b80/LICENSE).

When to Use

• Use when the user asks whether you know the versions of their Python dependencies, or what changed since your training data.
• Use when code keeps failing on a renamed, moved or removed function, class or parameter of a pinned library.
• Use after a dependency upgrade, before writing code against the upgraded library.
• Do not use for projects that are not Python.

How It Works

1. Work from the project root, the directory that has pyproject.toml, requirements*.txt or a lockfile (uv.lock, poetry.lock, pdm.lock, pylock.toml, Pipfile.lock).

2. Check the runner with since-cutoff --version. If it is missing or does not report 0.4.1, ask the user before running uvx [email protected], which downloads that release from PyPI the first time. Do not install anything else or change the user's environment.

3. Run the read-only scan and name the model you are: uvx [email protected] scan --model <provider>:<model> (for example anthropic:claude-sonnet-4-5 or openai:gpt-5.4). Without --model, it reads the model from the coding agent's settings and prints where it found it. The scan reads the project's files, PyPI and models.dev. It sends none of the project's code anywhere.

4. Summarise the result for the user: the model and its cutoff, which dependencies changed, and which changed APIs the code uses, with the files that use them.

5. Write notes only if the user asks for them. Run uvx [email protected] sync --dry-run and show the user the diff it prints. Run uvx [email protected] sync --yes only after they agree. It writes one marked block into AGENTS.md (or into CLAUDE.md if that is the file the project uses) and leaves the rest of the file unchanged. Run after later upgrades, it keeps the block in step with the lockfile. Exit code 4 means the block was edited by hand: tell the user, and pass --force only if they say so. uvx [email protected] unapply removes the block.

6. Treat the notes as reference facts about the versions the project pins, and check them when you write code that uses those libraries.

7. Optional: since-cutoff run measures which of the changes the model actually gets wrong. It sends prompts to the model provider the user picks (package names, versions, public API signatures and generated tasks, never the project's source code) and uses their credits or usage, so ask first.

Examples

uvx [email protected] scan --model anthropic:claude-sonnet-4-5

Output on the project's sample app, abridged to one package:

Your code uses 2 APIs that changed after claude-sonnet-4-5's training cutoff (2025-07-31)

anthropic 0.60.0 -> 1.8.0 (0.60.0 was the latest release at the cutoff; pyproject.toml pins 1.8.0)
  Messages.create: temperature, top_k and top_p were removed                          uses this API
    app/main.py   calls create
    Note: This is illustrative diff output, not current Anthropic API guidance.
          Verify removed parameters against the pinned SDK and service documentation;
          do not bypass removal by injecting them through extra_body/extra_query. [diff]

Limitations

• Python projects only.
• It compares the shape of the public API: names, signatures, parameters and deprecation markers. A change in behaviour behind an unchanged signature is invisible to it.
• The training cutoff only decides which changes to look at. It does not show what the model memorised; run measures that.
• It needs network access to PyPI, and to models.dev for model cutoffs. The first scan downloads the releases it compares. Results are cached locally.

Cutoff metadata and static API differences are hypotheses to verify against the exact pinned release; they do not prove model knowledge or service acceptance. Review the pinned runner/dependency download behavior before use; static analysis of downloaded package data is distinct from installation/runtime execution. Do not claim zero execution/network disclosure solely from this entry.

Security Notes

scan and status do not change project files: they write only to their own git-ignored .since-cutoff/ folder and a local cache. sync writes one marked block to AGENTS.md or CLAUDE.md, and only after the user agrees. No package code and no model-written code is executed. Credentials are needed only for the optional run.