SkillPedia Skills Encyclopedia
中文 EN

Skills 2026-09-08 18:33 UTC

watch-skillv0.1.4

Version [v0.1.4] committed by oxbshw. Update Category: Documentation & Guidelines Update. Commit details: Watch Skill 1.4.3 and DeepWatch 0.1.4: show the workspace, and metadata people can find it by (#39). This update improves usage instructions, deployment guides, and developer docs.

oxbshw f1317c8 oxbshw/watch-skill

CommitRaw commit message

Watch Skill 1.4.3 and DeepWatch 0.1.4: show the workspace, and metadata people can find it by (#39)

* docs: show the workspace instead of only describing it

The DeepWatch half of the README asked a reader to take three things on trust
that a screenshot settles in a second.

**Onboarding.** The first screen now appears where the install commands end, so
"what do I get" is answered before the next section. It is also the screen that
carries this product's one unusual idea, and the prose next to it says why the
panel reads two numbers rather than one percentage: installed and proved are
different facts, and a fraction blurs them.

**Role Bindings.** The four-step "connect a model, and prove the connection"
table now ends in the screen it describes — Chat marked Ready against a named
provider and model, with Visual perception and Speech to text saying *Not
configured* beside it. That is the point the paragraph was making in words: a
role with nothing assigned never falls back to another role's model.

**Sources.** The privacy paragraph claimed capture is a separate consent. The
Sources screen shows every recordable source tagged Local and Not requested,
with the permission requested at first use, which is the claim rather than a
restatement of it.

The walkthrough also opens with its own arc — recording, question, repair,
proof, and the record the next day — so the five steps read as one job rather
than five headings.

And the images say which build each came from. They are not all from one: the
onboarding shot is the current gallery at 1.4.2 / 0.1.3, gated against the
engine it was taken with, and the Settings shots are from the 1.4.0 / 0.1.1
candidate whose surface this release did not change. Printing them together
without saying so would be the easiest false claim on the page.

* release: Watch Skill 1.4.3 and DeepWatch 0.1.4 — metadata people can find this by

Every published package carried three shared keywords and one or two of its
own: enough to find a package you can already name, and useless to anyone
searching for what it does. npm weights `keywords` heavily and this was the
whole of it.

`scripts/gen-package-keywords.mjs` now owns them for all twenty, on two rules.
Every term has to be true of that package — `ocr` belongs on the bundle that
brings OCR, not on the brand tokens, whatever the word is worth in traffic. And
anything true of all twenty lives in one shared list, so a rename happens once.
It runs as a check by default and fails on a package it has no entry for,
because a package nobody wrote keywords for is a package nobody finds.

PyPI: seven keywords naming the ingredients (`ffmpeg`, `yt-dlp`) become
twenty-four naming the job — video-understanding, agent-observability, mcp,
speech-to-text, verification, evidence, local-first. Six more classifiers, and
the sidebar links PyPI renders now include Getting started and Release proof.
`Typing :: Typed` was written and then removed: there is no `py.typed` marker,
so it would have been a claim with nothing behind it.

`llms.txt` said the `@deepwatch/*` packages are not published and that
`npx @deepwatch/cli` does not work. That has been false since 0.1.2 and it is
the file crawlers read. It now describes both products, both registries, and
what the Desktop build is not.

The gallery is retaken against the engine this release builds, and the two
shots that do not show their subject still say so.