An uncomfortable verdict first: asking 'what is the difference between these two rc versions' is the phrasing most likely to produce fabrication. RC changelogs are often thin or unpublished, so with no reliable basis the model supplies something plausible.
Ask yourself first: is there checkable evidence
| Checkable evidence | Where | What it answers |
|---|---|---|
| The engines field on npm | engines per version at registry.npmjs.org | Whether the node requirement changed (genuinely checkable) |
| Publish times on npm | the registry time field | Which shipped first (corrects 'bigger version is newer') |
| Dependency versions | each @deepseek-ai package version under dependencies | Whether internal components moved generation |
| CHANGELOG and release notes | the repo's releases page | Official notes on changes (often incomplete for rc) |
| Measurement | Install both and run each once | Behavioural difference, the only trustworthy behavioural evidence |
The checkable part: hard evidence
| Item | 0.2.1-rc.5 | 0.2.2-rc.1 |
|---|---|---|
| engines.node | not declared | >=24.9.0 |
| Published | 2026-10-29 | 2026-11-04 |
| bin | dsh to lib/bin.js | dsh to lib/bin.js |
| dsh-hmr and friends | same generation | the 0.2.2-rc.1 generation |
This part belongs on a page, because it comes from files rather than memory.
The uncheckable part: make the AI admit ignorance
I want to know how @deepseek-ai/dsh behaves differently in 'plugin loading' between 0.2.1-rc.5 and 0.2.2-rc.1. Answer in this format: 1. Differences you can confirm from publicly checkable sources: one line each with the source (which file, page, or field) 2. Parts you cannot confirm and can only infer: label them INFERRED and state the basis 3. What you do not know: write 'no reliable information'; do not fill with 'may' or 'usually' Do not output anything like 'rc versions generally...'
The load-bearing sentence is do not fill with may or usually. Without it you get three paragraphs of true-but-empty talk about rc versions being more stable, which looks like an answer and says nothing.
Most reliable: stop asking, start measuring
# Run the same plugin-loading task once per version and record actual behaviour npm i -g @deepseek-ai/[email protected] && dsh --version npm i -g @deepseek-ai/[email protected] && dsh --version
A behavioural difference counts only once you run it. Asking an AI gives you what others said; measuring gives you what happens here. Config also changes across the upgrade, so pair this with the config-diff page.
General principles for this class of question
- Do not ask what can be looked up: engines, publish times, dependency versions come faster and truer from the registry.
- For behaviour, run it: there is no reliable memory of a specific rc's runtime behaviour.
- Require confirmed-versus-inferred separation; it is the only instruction that makes an answer honest.
- Stop at 'usually': that word marks where it started filling blanks.