diff options
| author | Claude <noreply@anthropic.com> | 2026-11-19 01:38:59 +0000 |
|---|---|---|
| committer | Claude <noreply@anthropic.com> | 2026-11-19 01:38:59 +0000 |
| commit | fc515815fa14ab96bebc363a5f42e79a7f571158 (patch) | |
| tree | 796871d943efb9f92635f2cb62a34498d82d300e /plugins/code-modernization/README.md | |
| parent | 4d9854e0873b9d068fb444c8b14d00278ce2402a (diff) | |
code-modernization: look up whether a target version is supported before proposing itcode-modernization/supported-target-versions
Commands proposed language and framework versions from the model's memory, which is older than today, so a plan could name a release that is already out of support. A shared rule (references/target-versions.md) now has each command look versions up, record them in a Target versions table with the support end date and source, keep a version the person chose with a warning, and mark a version that could not be checked as unverified instead of guessing. Version 1.1.0.
Diffstat (limited to 'plugins/code-modernization/README.md')
| -rw-r--r-- | plugins/code-modernization/README.md | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/plugins/code-modernization/README.md b/plugins/code-modernization/README.md index e9da7047..d76a8dba 100644 --- a/plugins/code-modernization/README.md +++ b/plugins/code-modernization/README.md @@ -67,7 +67,7 @@ Each module gets one verdict, computed by `scripts/proof_pack.py` from those fil | If you want | Run | What happens | | --- | --- | --- | -| The same technology on a newer version (.NET Framework to .NET 8, Java 8 to 17, Spring Boot 2 to 3) | `modernize-uplift <name> [source-version] [target-version]` | Keeps your code and fixes only what the new version breaks, driven by a catalog of the breaking changes this code actually hits. One pilot unit first with its lessons written down, then batches. The proof is the same test suite run on both versions. | +| The same technology on a newer version (.NET Framework to modern .NET, Java 8 to the current LTS, Spring Boot 2 to the current release) | `modernize-uplift <name> [source-version] [target-version]` | Keeps your code and fixes only what the new version breaks, driven by a catalog of the breaking changes this code actually hits. One pilot unit first with its lessons written down, then batches. The proof is the same test suite run on both versions. | | A new technology, one module at a time, while the old system keeps running | `modernize-transform <name> [module] [target-stack]` | A plan you approve, tests that pin the old behavior, an idiomatic rewrite, and proof: old and new run on the same inputs and a script compares them. | | A rebuild on a new architecture | `modernize-reimagine <name> [target-vision]` | A spec mined from the code, an architecture that is reviewed and approved, then services scaffolded with executable acceptance tests. | @@ -133,11 +133,12 @@ Helpful but optional (run `preflight` to check them all): [`scc`](https://github - **Analyzed code is untrusted input.** A hostile codebase can plant comments like "ignore previous instructions", a README that tells tools to run a script, or file names with shell syntax. Agents treat file content as data, list the instruction-shaped text they found and never follow it (the legacy code's own build and tests run only where a command needs its behavior, in a scratch copy when it can), verification agents re-derive every rule and finding from the cited code, and `brief` is a human approval gate before anything is built. Treat discovery artifacts from untrusted code with the same skepticism as the code. - **Secrets stay out of shared artifacts.** Discovered credentials are masked (`AKIA****`) and inventoried in a gitignored `SECRETS.local.md` (or `~/.modernize/<name>/` outside git); `harden` keeps credential-removal hunks in a separate gitignored patch. `--show-secrets` puts raw values in the quarantine file only. - **The old system is run only where it is safe.** Baselines come from the legacy code running locally or against a test environment you named. Production or third-party services, new accounts and real data are off limits unless the plan you approved names them, a token in a recorded response is replaced before anything is saved, and every command stops the servers it started. +- **Target versions are looked up.** The model's memory is older than today, so a target version is checked when a command proposes it: Claude's web tools or a shell command read the vendor's support page, endoflife.date or your package registry, under your usual permissions. Only product names and versions are sent. The result is a **Target versions** table with a support end date and its source. Offline, versions are marked `unverified` and listed for a person to confirm. A version you named is kept, with a warning if it is out of support. - **Trying it on a live repository.** `preflight` and `assess` change no source file. Preflight's smoke test compiles one file and, where there is a build system, restores and builds one small project, which writes build output wherever the build normally does. ## Telemetry -The plugin counts how it is used, so the next version can be better. It sends **whole numbers only**: never code, file or system names, paths, prompts or anything you typed. Counts of 100 or more are rounded to two significant figures (17,727 is sent as 18,000), so a number says roughly how big, not exactly which system. It sends them through Claude Code's own telemetry, so nothing is sent to Anthropic when that is off. The plugin itself makes no network call and adds no identifier; the one file it writes is a small `telemetry-state.json` of hashes in its own data folder, so the same counts are not sent twice. +The plugin counts how it is used, so the next version can be better. It sends **whole numbers only**: never code, file or system names, paths, prompts or anything you typed. Counts of 100 or more are rounded to two significant figures (17,727 is sent as 18,000), so a number says roughly how big, not exactly which system. It sends them through Claude Code's own telemetry, so nothing is sent to Anthropic when that is off. The counting itself makes no network call and adds no identifier; the one file it writes is a small `telemetry-state.json` of hashes in its own data folder, so the same counts are not sent twice. Five small hooks do it: |
