From 84e26b6c95c7d70a30d1b18ff230ebe1858bb805 Mon Sep 17 00:00:00 2001 From: Aleksandr Shevchuk Date: Sun, 15 Nov 2026 17:29:42 +0200 Subject: [PATCH 1/2] fix(author): route 3 tries the two files wherever Node runs, whatever the code tool says about the internet Since 0.10.0 step 3 took "no internet" from the chat's own tools and then skipped route 3. Grok's code tool is documented as having no network, yet it fetches the organisation's site (v24.15.0, 200), so Grok handed over the JSON source unless the build archive was attached. Now only having no code tool means no npm; route 3 holds wherever Node runs and is tried even where the tools say otherwise or npm gave up. Haiku: 3 of 3 runs with npm hanging and the web reachable fetched both files and built packages that pass the verifier; 2 of 2 with nothing reachable gave the json block, which builds and passes. Closes #211 --- design/author.md | 50 +++++++++++++------ skills/h5p-author/SKILL.md | 15 +++--- .../h5p-author/references/build/two-files.md | 3 +- 3 files changed, 43 insertions(+), 25 deletions(-) diff --git a/design/author.md b/design/author.md index 7902fea7..3c492585 100644 --- a/design/author.md +++ b/design/author.md @@ -888,13 +888,13 @@ to 3,000. Twenty-nine things the code does not say by itself: code stands for, such as MIT, is `U` (cubic's review): Haiku 3 of 3, each built. The brief keeps what the model chooses and says it chose (the material, who it is for, the size, the language) and how to write well; it asks nothing. - - Step 3 finds two answers once, and both the route and the verifier use them. **Internet**: none - with no code tool, or where the chat's own tools say its sandbox has none, and then nothing is - run to check; otherwise taken as there. **npm**: none without internet; otherwise - `npm view … --fetch-timeout=10000 --fetch-retries=0`, run beside the first file reads, as it - waits on nothing. Then the first route that holds: the build archive attached (route 1, - `references/build/archive.md`), npm (route 2, `npm.md`), Node with internet (route 3, - `two-files.md`), anything else (route 4, `browser.md`). Routes 1 to 3 hand over by + - Step 3 finds one answer once, and both the route and the verifier use it: whether npm reaches + its registry. With no code tool it does not; otherwise `npm view … --fetch-timeout=10000 + --fetch-retries=0` answers, run beside the first file reads, as it waits on nothing. Then the + first route that holds: the build archive attached (route 1, `references/build/archive.md`), + npm (route 2, `npm.md`), Node (route 3, `two-files.md`, tried even where the chat's tools say + its sandbox has no internet or npm gave up: *A code tool that says it has no internet still + tries the two files*, below), anything else (route 4, `browser.md`). Routes 1 to 3 hand over by `built-here.md`: the build's problems and the file's name, the verifier, run only where npm answered, the link, written in the verifier's command line, and the hand-over. - Route 4 runs no link command: the link's commands are written once, in `built-here.md`. It @@ -921,16 +921,15 @@ to 3,000. Twenty-nine things the code does not say by itself: - Check 1 at first ran `find / -name 'h5p-author-build*.tar.gz'` wherever Node ran. On a computer's whole disk it ran past the code tool's 110 seconds, and the run with npm working took npm for broken. It now asks whether the user attached the archive, and `find` is route 1's. - - A code tool that says its sandbox has no internet should cost no check. Under three wordings - ("answer it from what you know for certain"; "take their word and run nothing"; the two - bullets, the first ending "run nothing to check") Haiku ran the npm check anyway in 4 of 6 - such runs, 1 of 2 under the last. The check is held to ten seconds and leads to the same - route, so a run that ignores the bullet loses those seconds and nothing else; a model that - keeps to it, as stronger ones are expected to, loses none. Not measured outside Haiku. + - A code tool that said its sandbox had no internet cost no check until 0.10.3. Under three + wordings ("answer it from what you know for certain"; "take their word and run nothing"; the + two bullets, the first ending "run nothing to check") Haiku ran the npm check anyway in 4 of 6 + such runs, 1 of 2 under the last. A model that kept to it skipped route 3 as well, which + Grok did: *A code tool that says it has no internet still tries the two files*, below. - Route 3 read "Node runs here and can fetch a file from the web": a run told it had "no web tool" - took that for no internet and wrote the page without trying the two files. It reads "Node + took that for no internet and wrote the page without trying the two files. It read "Node runs here, with internet", the answer found once, and the run after it fetched both files and - built. + built. Since 0.10.3 it reads "Node runs here", and says to try it whatever the tools say. - A run called the Skill tool and got the skill installed on this computer, an older text: the lab's prompts now say the reader is the skill. Haiku also read a route's file before its check now and then, at most 1.5 KB more. @@ -1001,6 +1000,27 @@ to 3,000. Twenty-nine things the code does not say by itself: `.source.json` file, equal to the block, and the archive line, and none of the four without one did. The route's file is 1.5 KB. The build page still reads `#source={…}`, which costs nothing, and routes 1 to 3 keep the compressed link their command prints. +- **A code tool that says it has no internet still tries the two files** (2026-11-15, #211). From + 0.10.0, step 3 took "no internet" from the chat's own tools and then skipped route 3. Grok on the + web (the owner, a free account) then gave the `json` block unless the build archive was attached, + though on 2026-11-13 under 0.9.6 it had fetched the two files and built. Its code tool reaches + the organisation's site: asked to run `node -v` and `curl` on the build tool's address, it + answered `v24.15.0` and `200`. xAI documents that tool as having no access to external networks. + So nothing decides "no internet" now but having no code tool: npm is checked wherever one runs, + and route 3 holds wherever Node runs, tried even where the tools say otherwise or npm gave up. A + sandbox that truly reaches nothing loses the ten-second check and a failed fetch, about a + second here. On Haiku, each run told that its code tool's sandbox has no internet: + - npm hanging, the web reachable (as Grok): 3 of 3 took route 3, fetched both files, built + packages that pass the verifier here, and said they were unverified. One of them also tried + the verifier's `npx` first, which hung for the code tool's 110 seconds: the hand-over's rule + runs it only where npm answered. + - Nothing reachable: 2 of 2 gave the `json` block, which built and passed, the `.source.json` + file equal to it and the archive line. + + The lab's npm must be cut for real: a run root with a `package.json`, a `.npmrc` naming a dead or + hanging registry and a cache of its own. Without the cache, npm answered from this computer's + cache with no network ("cache stale", exit 0); without the `package.json`, a run working in a + subfolder reached the real registry. Four runs spoiled that way were not counted. - **The skill names no product** (2026-11-12, the owner's call). Its routes were told apart by the apps they were measured in; they now say what each needs: a shell that reaches the npm registry, Node and a download where npm cannot install, or nothing that runs or fetches. What a diff --git a/skills/h5p-author/SKILL.md b/skills/h5p-author/SKILL.md index 4f531cb8..b2e6261b 100644 --- a/skills/h5p-author/SKILL.md +++ b/skills/h5p-author/SKILL.md @@ -91,14 +91,9 @@ H5P.Video, H5P.Text. ## 3. Build Where a command runs here, work in a folder of your own, not the one the user named, with the -source as `source.json`. Two answers pick the route, and later whether the verifier runs. Find -each once: - -- **Internet**: none where you have no code tool, or your tools say this sandbox has none; run - nothing to check. Otherwise take it that there is. -- **npm**: it cannot reach its registry where there is no internet. Otherwise run this, early, - beside the files you read, since it waits on nothing. It prints a version, or gives up within - ten seconds. +source as `source.json`. Whether npm reaches its registry picks the route, and later whether the +verifier runs. With no code tool it does not. Otherwise run this once, early, beside the files you +read, since it waits on nothing. It prints a version, or gives up within ten seconds. ```bash npm view @missing-elements/h5p-author version --fetch-timeout=10000 --fetch-retries=0 @@ -110,7 +105,9 @@ next. Tell the user which you take. 1. The user attached the build archive, `h5p-author-build.tar.gz`, to a message or to this chat's files, and Node runs here: [route 1](./references/build/archive.md). 2. npm reaches its registry: [route 2](./references/build/npm.md). -3. Node runs here, with internet: [route 3](./references/build/two-files.md). +3. Node runs here: [route 3](./references/build/two-files.md), which fetches two files from the + web. Try it even where your tools say this sandbox has no internet, or npm gave up: such a + sandbox has fetched them. 4. Anything else: [route 4](./references/build/browser.md), the user's browser. ## 4. Verify and hand over diff --git a/skills/h5p-author/references/build/two-files.md b/skills/h5p-author/references/build/two-files.md index c27192be..40c1a68c 100644 --- a/skills/h5p-author/references/build/two-files.md +++ b/skills/h5p-author/references/build/two-files.md @@ -2,7 +2,8 @@ Fetch the build tool and the bundle of the source's own `type`, the libraries a build of that type reads, at the addresses https://missing-elements.github.io/h5p/skill/ lists. Where the sandbox -downloads only what a browsing tool has opened, open each address with that tool first. Below, +downloads only what a browsing tool has opened, open each address with that tool first. Where +neither file arrives, this route has failed. Below, `` is that machine name, such as `H5P.Blanks`: another type's bundle lacks what this one needs. From 69e743a0b1d46c8a9bbdd16bb1a6c56107457329 Mon Sep 17 00:00:00 2001 From: Aleksandr Shevchuk Date: Sun, 15 Nov 2026 17:36:35 +0200 Subject: [PATCH 2/2] docs(author): the step 3 check keeps an answer from npm's cache, and why Refs #211 --- design/author.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/design/author.md b/design/author.md index 3c492585..dc7ac085 100644 --- a/design/author.md +++ b/design/author.md @@ -1020,7 +1020,11 @@ to 3,000. Twenty-nine things the code does not say by itself: The lab's npm must be cut for real: a run root with a `package.json`, a `.npmrc` naming a dead or hanging registry and a cache of its own. Without the cache, npm answered from this computer's cache with no network ("cache stale", exit 0); without the `package.json`, a run working in a - subfolder reached the real registry. Four runs spoiled that way were not counted. + subfolder reached the real registry. Four runs spoiled that way were not counted. The skill's + check still takes an answer from npm's cache (cubic's review on #212): with no network, route 2 + then builds from that same cache, as one of those runs did after about 70 seconds of npm's retries. + `--prefer-online` still falls back to the stale cache, and `npm ping` would say nothing of the + package. - **The skill names no product** (2026-11-12, the owner's call). Its routes were told apart by the apps they were measured in; they now say what each needs: a shell that reaches the npm registry, Node and a download where npm cannot install, or nothing that runs or fetches. What a