Skip to content

feat(pages): publish the deck and the instrument - #26

Merged
WWStoryMode merged 4 commits into
mainfrom
docs/journal-day-5
Sep 12, 2026
Merged

WWStoryMode merged 4 commits into
mainfrom
docs/journal-day-5

Conversation

@WWStoryMode

Copy link
Copy Markdown
Owner

Puts the deck and the instrument on the web, so the generated movement can be reviewed by anyone without a GPU.

A session file carries its motions and opens with no service, no GPU and no network — so a static build gives a reviewer the whole loop except generation: open a poem, play it, scrub it, read the four notation registers, rate it in review mode.

The site this produces

/BodyPrompt/ landing page — new
/BodyPrompt/presentation/…deck.html the deck as delivered — unchanged
/BodyPrompt/journal/… the journal, served verbatim — unchanged
/BodyPrompt/app/ the instrument — new

One bug, fixed regardless of any deployment

start() ended if (restored) showCurrent(); else void draftLine(); — so a first-time visitor generated on arrival and failed, because there was no service.

Drafting into a dead service is never useful, on a laptop or a static host. refreshCapabilities now reports whether the service answered instead of returning silently, a fresh start only drafts when a model is actually ready, and otherwise the hint says generation needs a local service and that IMPORT works without one.

Silence there was the real problem: with no service the model dropdown kept its default labels and nothing on screen said generation was unavailable. That is the v0-stub.md rule applied to a deployment — the screen must not imply a capability that is not there.

?session= — a link that opens a poem

The parked roadmap item, needed now that docs/ is served verbatim and the session files are same-origin fetchable:

…/app/?session=../journal/data/2026-09-08/poem-writing-007a.json

Two decisions worth review:

  • It is handled inside start(), not in the boot-flag block, because it has to beat the autosave restore. A returning visitor's own poem must not silently win over the poem they were sent.
  • sessionUrl() confines it to this site. A boot flag is a URL somebody else can choose, so a relative path resolves and anything reaching another origin is refused loudly. Three tests cover that, including protocol-relative (//evil.example) and scheme-downgrade attempts.

Failure is reported in the session bar exactly as a failed IMPORT is — a link that quietly opens the wrong poem is worse than one that says it will not be followed.

All fourteen Day 4 sessions

Byte-identical, 39 MB raw / ~14 MB compressed. Not a curated five, because the day's findings are comparisons that only exist across the variants: the seated collapse is reproducible because three seeds agree, and the locomotion finding is six bakes rather than one.

Their README carries every seed and net-travel figure, and two warnings found while writing it:

  • ⚠ poem-writing-004 holds SnapMoGen's line drafts, but its bake is Kimodo's — carried over from 003c, because SnapMoGen reports can_stitch_poems: false and cannot bake at all. Pressing play shows Kimodo. This is the same trap Day 1's session files carried.
  • One intermediate bake of 002 was overwritten during the session, so the 0.018 m wrist span quoted in Day 4 §3 is real, was measured, and cannot be reproduced from these files.

The repository's first CI

.github/workflows/pages.yml builds rather than committing dist/. A checked-in bundle drifts from source silently, and a deployed app older than the code is exactly the drift v0-stub.md exists to prevent — nobody would notice for weeks.

pnpm, not npm: npm install genuinely fails here on the pre-existing typescript-eslint peer conflict; the pnpm lockfile resolves cleanly. Verified both.

The assembly copies docs/ first and whole, then adds the app beside it, and asserts .nojekyll, the deck and the built app are all present before uploading.

⚠ The risk, and the order to merge in

Pages is currently in legacy mode serving /docs. Switching the source to GitHub Actions hands every URL to this YAML — including the deck published today. If the assembly is wrong, that deck 404s.

Suggested order:

  1. Run the workflow via workflow_dispatch on this branch — it builds and uploads without deploying.
  2. Merge.
  3. Switch Pages source to GitHub Actions (Settings → Pages).
  4. Check the deck URL first, then the landing page, then the app in a private window so no IndexedDB session masks the fresh-visitor path.

Verification already done

The artifact was assembled exactly as the workflow does and served under a real /BodyPrompt prefix. The landing page, the deck, the app, its hashed JS bundle and a session JSON all return 200.

npm test 76 → 78, tsc --noEmit clean, vite build clean with the base applied.

What this does not do

It does not make the models available to anyone. Generation still needs the GPU and three workers, and the deployed instrument now says so rather than appearing broken.

🤖 Generated with Claude Code

https://claude.ai/code/session_014yJSPiRjB5WvTH2x4LB8V6

WWStoryMode and others added 4 commits September 12, 2026 10:53
Adds the deck presented on 2026-09-09 so it can be read outside the room
it was shown in. One self-contained HTML file, sixteen slides, arrow
keys to navigate; the only thing it fetches is two typefaces, and
without a network it falls back to Georgia and stays readable.

Committed byte-identical to the file that was presented, and it is not
maintained. The research moved the following day, and a deck edited
later to match current findings would stop being a record of the talk
and become a second, quieter version of the journal -- the drift
docs/v0-stub.md exists to prevent, wearing a different costume. Where
the deck and the journal disagree the journal is the record, and the
README beside it names the two places that already applies: the deck
presents the five prompt levels without saying an LLM wrote the derived
rungs, and its model table still lists Language of Motion as an open
question when all three models were real by Day 4.

Its statistics slide was checked against the record before committing.
All eleven figures -- the five level means, the five percentages rated
at or above 3, and the 1.43-point Q->A step -- match Day 3 section 6
exactly, on the same 315-judgement census. What the slide necessarily
leaves out is in the README: the rating was not blind, level order was
confounded with rating order, one rater, and a retest that moved 8 of 21
judgements upward.

One thing this does not do on its own: GitHub serves .html in a
repository as source, so the deck renders as markup until Pages is
switched on. With Pages serving from /docs on main it would be at
wwstorymode.github.io/BodyPrompt/presentation/ -- recorded in the README
rather than assumed, since enabling it is a repository setting.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014yJSPiRjB5WvTH2x4LB8V6
Pages enabled from /docs on main, so the deck is a page rather than a
view of its own markup:

  wwstorymode.github.io/BodyPrompt/presentation/
    2026-09-09-lecture-performance-deck.html

docs/.nojekyll turns the build step off. The deck then reaches a browser
byte-for-byte as it was presented, with nothing able to rewrite it on
the way -- which is the point of committing it unchanged in the first
place.

The tradeoff, stated in the README: markdown under docs/ is served as
plain text rather than rendered. Those files are meant to be read on
github.com, where the .md links between them resolve; under Jekyll they
would render prettily and every one of those links would 404, since the
files become .html. Serving the directory verbatim keeps one canonical
place to read the journal instead of two that disagree.

Nothing became newly public: the repository was already public, and
Pages only changes how it is rendered.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014yJSPiRjB5WvTH2x4LB8V6
A session file carries its motions and opens with no GPU, no service and
no network -- so a static build gives a reviewer the whole loop except
generation: open a poem, play it, scrub it, read the four registers,
rate it. This publishes that.

Three code changes, one of which is a bug independent of any deployment.

  The instrument no longer generates on arrival. start() ended with
  "if (restored) showCurrent(); else void draftLine()", so a first-time
  visitor with no autosaved session tried to generate and failed.
  Drafting into a dead service is never useful, on a laptop or a static
  host. refreshCapabilities now reports whether the service answered
  rather than failing silently, and a fresh start only drafts when a
  model is actually ready; otherwise the hint says generation needs a
  local service and that IMPORT works without one. Silence there would
  have left the model list reading like a working instrument.

  ?session=<relative path> opens a poem from a link, the parked roadmap
  item, needed now that docs/ is served verbatim and the session files
  are same-origin fetchable. It is handled inside start() rather than in
  the boot-flag block, because it has to beat the autosave restore -- a
  returning visitor's own poem must not silently win over the poem they
  were sent. sessionUrl() confines it to this site: a boot flag is a URL
  someone else can choose, so a relative path resolves and anything
  reaching another origin is refused, loudly. Three tests cover that,
  including protocol-relative and scheme-downgrade attempts.

  vite base is /BodyPrompt/app/ for builds only; the dev server stays at
  the root so existing habits and boot-flag URLs are unaffected.

The site gains a landing page -- docs/index.html, hand-written, no build
step -- which states plainly that nothing here generates motion and that
opening a saved session needs nothing at all. Its journal link points at
GitHub rather than at the Pages copy, because .nojekyll means the .md
files are served as text and their inter-links only resolve there.

All fourteen Day 4 sessions are committed byte-identical, 39 MB raw and
around 14 MB compressed. Not a curated five: the day's findings are
comparisons that exist only across the variants -- the seated collapse
is reproducible because three seeds agree, and the locomotion finding is
six bakes rather than one. Their README carries the seeds, the net
travel, and two warnings: poem-writing-004 holds SnapMoGen's line drafts
but its bake is Kimodo's, the same trap Day 1's files carried; and one
intermediate bake of 002 was overwritten during the session, so a figure
quoted in Day 4 section 3 cannot be reproduced from these files.

.github/workflows/pages.yml is this repository's first CI. It builds
rather than committing dist/, because a checked-in bundle drifts from
source silently and a deployed app older than the code is the drift
v0-stub.md exists to prevent. pnpm, not npm: npm install fails here on a
pre-existing typescript-eslint peer conflict while the pnpm lockfile
resolves cleanly. The assembly copies docs/ first and whole, then adds
the app beside it, and asserts .nojekyll, the deck and the built app are
all present -- because once Pages is switched from /docs to workflow
mode this YAML owns every URL the site has, including the deck published
today.

Verified by assembling the artifact exactly as the workflow does and
serving it under a /BodyPrompt prefix: the landing page, the deck, the
app, its hashed JS bundle and a session JSON all return 200.

npm test 76 -> 78; tsc and vite build clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014yJSPiRjB5WvTH2x4LB8V6
workflow_dispatch cannot reach a workflow that is not yet on the default
branch -- GitHub returns 404, because it only registers a workflow once
main carries it. So the trigger that lets this be proved before main
depends on it is pull_request, not dispatch.

The deploy job is guarded off for pull requests: a PR runs install,
tests, build and the artifact assembly with its three assertions, and
stops before touching the live site. Publishing still happens only from
main.

Better than a one-off dispatch in any case -- every future change that
touches the frontend or this file now proves the site still assembles
before it can be merged.

Adds a step printing the assembled tree and its size, so a reviewer can
see what would be published without downloading the artifact.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014yJSPiRjB5WvTH2x4LB8V6
@WWStoryMode
WWStoryMode merged commit 9f9c02c into main Sep 12, 2026
2 checks passed
@WWStoryMode
WWStoryMode deleted the docs/journal-day-5 branch September 12, 2026 10:31
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant