Skip to content

Phase E5: Replace the Compute dialog #498

Closed

Description

@neuromechanist

Goal

After this phase the "Compute instructions" dialog on dataset pages tells the truth: how to request compute access, how to run a job on this dataset with the nemar compute commands, and where to see it afterwards.
It no longer points at an s3://nemar/... prefix or at the Neuroscience Gateway for a one-click flow that never shipped, and website#6 no longer carries the NSG connector as a pending deliverable.

Routing

Scope

In: the compute dialog and its button in src/components/ActionBar.astro, a pure instruction builder with tests, the dialog's script hook, the website#6 retarget.
Out: the Download dialog (it already lists the local and file routes; do not copy them into the compute dialog); the about page's description of the SDSC and NSG partnership (a separate wording decision for the owner once compute is live); submitting jobs from the page.

Design references

Design section 7 (CLI forms: nemar compute request-access, nemar compute submit ... --dataset nm000104, status|logs|results|cancel, the specification and inline header), section 1 (compute access is requested once; limits are the owner's).
Closest code: src/components/ActionBar.astro (the Download dialog's dl-cmd rows with data-copy-text and installUrl; the current compute dialog and its data-compute-stub script at the foot), src/lib/host.ts (appUrl for a link from the marketing host to the app host), src/lib/use-this-data.ts (a tested builder of command strings from dataset facts), src/lib/data-base.ts.

Changes

  • src/lib/compute-instructions.ts (pure): from { datasetId, version } build the command rows (request access; submit a script against <id>@<version>; a folder with an entry file; list and status), each as a label, a command and a one-line explanation; with no version it returns a not_available result that names why. Quote nothing that is not in the merged CLI's help. Tested with the shapes ActionBar is given.
  • src/components/ActionBar.astro: replace the data-compute-dialog body. Contents: one sentence on what NEMAR compute is (jobs run on a supercomputer under NEMAR's allocation; access is requested once and an admin grants it; the first step is a link to Settings via appUrl("/settings#compute-access", host)); the command rows with the Download dialog's copy buttons and styles; a link to the person's compute page (appUrl("/compute", host)) and to the CLI install page (installUrl); a link to the compute terms page once published, and the same link text only then; limits and balance live in nemar compute limits and nemar compute balance, not as numbers in the dialog (they are the owner's policy and change without a deploy). Remove the s3://nemar/{datasetId} prefix text, the npm install -g nemar-cli block, the "coming soon" note and the website#6 link. Rename data-compute-stub to a name that is not a stub and update the script that opens the dialog.
  • Do not repeat the Download dialog's local and file routes; link to it if the planner wants a pointer.
  • Anonymous deposits and datasets with no reachable data-plane copy: show the not-available text from the builder rather than commands that would fail.
  • Retarget website#6 ("Phase 5: Live-features architecture + cutover") with gh at implementation time: edit its body so the NSG connector deliverable and the .context/nsg_connector_architecture.md reference are struck as superseded by Epic: Website compute surfaces #493 (website) and Epic: Third-party compute: access, user API, nemar compute, partner contract nemar-cli#2021 (nemar-cli), leave the viewer, accessibility, performance and cutover deliverables untouched, and add a comment naming this pull request. Do not close the issue unless every remaining deliverable is verified done; say in the PR what was checked. The dialog's old link to Phase 5: Live-features architecture + cutover #6 disappears with the "coming soon" note.
  • Capture a screenshot of the new dialog in light and dark for the PR description.

Tests (write these first)

  1. compute-instructions.test.ts: a dataset with a version yields the rows with the id and version in the submit command; no version yields not_available; ids of each band (nm, ds, on, xx) pass through unchanged; no row contains s3://, npm or nsgportal. Fails if the old S3 prefix returns.
  2. A parity test (pure) that the command names in the builder appear in a captured nemar compute --help-all fixture from the merged CLI, so a renamed flag is caught: the fixture is real CLI output with the capture command in the test file.
  3. bun run build succeeds and a grep of the built dataset page for website/issues/6 and s3://nemar finds nothing.
  4. Browser check on a dataset page with a version and one without: the dialog opens from the button, copy buttons work, the Settings and compute links go to the app host from the marketing host and stay relative on the app host, both themes, phone width, Escape closes, focus returns to the button.

Acceptance criteria

  • No instruction in the dialog works only by accident of an old backend: each command exists in the merged CLI.
  • No numbers from the owner's policy are hard-coded.
  • The dialog is not shown with commands for a dataset the data plane cannot serve.
  • website#6 is retargeted and its remaining deliverables untouched.

Gates

bun run typecheck && bun run lint && bun run test && bun run build; then the browser check.

Rules that bite here

  • The data-plane rule: never send a person to a storage URL or an S3 prefix; the stable routes are the data plane and the CLI.
  • Scoped Astro styles do not cross components: the Download and compute dialogs share class names only because both live in ActionBar.astro.
  • Do not state limits, prices or allocation sizes: they are the owner's and may change.
  • Anonymous deposits must not gain a route that discloses their repository (the ActionBar's ghCloneUrl rule).
  • Close issues by hand; Closes #N does nothing on staging.

Escalate to the owner if

  • The owner wants NSG kept as an option, or wants the about page's partnership text changed now.
  • The compute terms page is still unpublished at release time and the dialog's terms link would 404.

When merged

Update the epic issue and AGENTS.md's architecture map; confirm in the epic issue that website#6 was retargeted and what remains open on it.

Activity

  1. neuromechanist commented on Nov 7, 2026

    @neuromechanist
    ContributorAuthor

    Plan: Phase E5, replace the dataset page Compute dialog and retarget website#6

    Planner output for the Sonnet implementer. Everything below was read from the merged code: the website epic branch feature/issue-#493-epic-compute-web (E1 and E2 merged, up to 442be8b) and the nemar-cli Epic D branch at e0685a25 (D1 to D4). Where this plan and the issue text disagree, the merged code is the fact and the difference is called out.

    0. Findings that change the issue's assumptions

    1. The dataset page cannot know the visitor's access state. src/pages/dataset/[id].astro is a marketing-host page with Cache-Control: public, s-maxage=300, and the session cookie is scoped to the app host (ADR 0006). So the dialog is one state-agnostic dialog; the five states are handled by sending the person to the surfaces that do know (Settings, /compute, the CLI), not by branching the markup. A client-side probe is impossible (no cookie on that host) and a per-visitor server render would break the edge cache.
    2. There is no nemar compute jobs. The merged command is nemar compute list. The full set is balance, cancel, limits, list, logs, queue, request-access, results, status, submit (from nemar compute --help-all). Design section 7 and this issue's wording ("list and status") are right; any jobs in older text is not.
    3. A dataset version is mandatory in --dataset. nemar compute submit refuses a versionless dataset: "A dataset needs its version, like nm000104@1.0.2, so the job can be run again on the same data." The page's selected version therefore goes into the command.
    4. Eligibility is the backend's, not the page's. POST /compute/jobs runs the same gate the pipelines use (resolveRun): the dataset must be active, public, not anonymous, first published, not withdrawn or tombstoned, have a version tag, and be a real nm (below the reserved fixture band) or on dataset. Sandbox, exemplar and ds ids are refused. The backend answers with one dataset_not_eligible ("A dataset the job names is not available to compute.") and does not say which term failed. There is no public read that exposes eligibility (see section 9, issue A2).
    5. No public availability read exists. Every /compute/* route needs a credential. The only way for a marketing-host page to know "compute is open" is the existing build-time rule from ADR 0030 (computeRequestsOpen(COMPUTE_REQUESTS_ENABLED, apiBase())). Reuse it; do not invent a second switch (section 5).
    6. website#6 was already narrowed by the owner (its second comment, 2026-08-03): remaining scope is exactly the NSG connector, and the stated condition is "keep open until either the NSG connector is built, or the Compute dialog stops pointing here". The viewer, accessibility and cutover deliverables are recorded there as done or superseded. This changes what "retarget" means (section 8). .context/nsg_connector_architecture.md does not exist in the repo; it is only text in Phase 5: Live-features architecture + cutover #6's body.
    7. The issue's test 3 (grep the built dataset page for s3://nemar) would fail for the wrong reason. src/lib/use-this-data.ts:526 legitimately renders s3://nemar/<id>/zarr/{store.zarr} in the Use this data section of the same page. Scope the grep to the compute dialog element only. Also the site builds with output: "server", so there is no built HTML to grep; the check is a source scan plus a curl against the dev server (section 7).
    8. ADR 0030 says "the marketing host links to docs, not to a compute surface." This phase deliberately adds links from a marketing-host dialog to the app host (/settings#compute-access, /compute). They are navigation only and render nothing from compute on the marketing host. The ADR 0030 addendum for E5 must say so (section 6).

    1. What the CLI actually offers (quoted, so the dialog invents nothing)

    From bun run src/index.ts compute <cmd> --help-all at e0685a25:

    Command Quoted behavior
    nemar login "Sign in with your browser (shortcut for 'auth login')"
    nemar compute request-access [--purpose <text>] [--accept-terms] "Ask an admin for compute access (one-time)." "You ask once; asking again while the request is open does nothing, and a request that was answered cannot be made again." "In a terminal the command asks for whatever you did not pass." "The request tells you which are missing and how to set each."
    nemar compute balance "It works in every access state: without a grant it says so and how to ask for one." Shows service units granted, held, charged, available per pool.
    nemar compute limits "what one job may ask for on each partition offered to you ... how many jobs you may have ... at once, and how long your results are kept."
    nemar compute queue "Show recent queue waits per partition"
    nemar compute submit [path] [args...] "Run a script, a folder or a container image on NEMAR compute." --dataset <id@version> "A dataset the job reads (repeatable), like nm000104@1.0.2". Help examples: nemar compute submit analysis.py --wait; nemar compute submit ./study --entry run.py --dataset nm000104@1.0.2 --cores 8 --mem 16G --time 2h; nemar compute submit analysis.py --dry-run. "--dry-run packs and scans and prints what would be sent, and sends nothing about the job (it needs no sign-in)." "Running submit again makes a SECOND job and a second hold." "--wait follows the job until it ends. Ctrl-C stops waiting and does NOT cancel the job." Resource flags (--partition --nodes --cores --mem --gpus --time) exist and have defaults.
    nemar compute list [--state --limit --all], status <job>, logs <job> [--tail n] list your jobs; show one; print its log
    nemar compute results <job> [--download <dir>] [--renew] "Lists the files a finished job made, and with --download fetches them into , checking each file's size and SHA-256". Links are bearer links that expire.
    nemar compute cancel <job> [--yes] cancel one of your jobs

    Access states (website COMPUTE_ACCESS_STATES, mirrored from nemar-cli): none, requested, active, suspended, denied, revoked. The CLI's refusals at submit are no_compute_access ("Your account does not have compute access.") and suspended ("Compute is paused for your account."), insufficient_balance, over_concurrency, dataset_not_eligible, version_not_found, version_untagged, too_many_files.

    How a job reads a dataset: the job names id@version; the platform stages that version from the data plane before the entry runs. The CLI help says nothing about where the files appear inside the job or where outputs go, so the dialog does not say. That is a gap in the CLI's help, filed as nemar-cli issue A3 (section 9), not filled in here.

    2. Decision: the dialog, per access state

    One dialog, same markup for every visitor. It never claims the visitor has access. The per-state behavior lives where the state is known:

    Visitor is What the dialog gives Where the state is then handled
    Not signed in Step 1 nemar login; the Settings and compute links go to the app host, which sends them to sign-in /compute redirects to /login?next=/compute; the CLI says "not signed in"
    No access (none) Step 2: link "Request compute access" to /settings#compute-access and the command nemar compute request-access Settings card (E1) has the form; submit is refused with "Your account does not have compute access."
    requested Same dialog; step 2 says "You ask once" Settings card and /compute show "An admin is reviewing your request." (compute_access.requested.*); asking again does nothing
    active Steps 3 to 5 are the working path /compute shows balances and the jobs list; nemar compute balance
    suspended Same dialog /compute and Settings show compute_access.suspended.body; submit says "Compute is paused for your account."
    denied or revoked Same dialog compute_access.closed.body: answered, cannot be asked again

    A single sentence in the dialog points at this: "Where your access stands is on your compute page" with the /compute link (label from the existing key compute_access.granted.link, "Open your compute page"). The CLI half is nemar compute balance ("works in every access state").

    3. Decision: dialog content and layout

    Button: keep the label "Compute instructions". It stays true (this is instructions, the one-click flow never shipped), it does not suggest a launcher, and it needs no copy churn. Only the hook changes: data-compute-stub becomes data-compute-open; the CSS class .compute-dialog is shared with the Download dialog and the cite dialog comments and is not renamed.

    Dialog (<dialog class="compute-dialog" data-compute-dialog aria-labelledby="compute-dialog-heading">, <h2 id="compute-dialog-heading">Compute on this dataset</h2>), ready state:

    • Intro (local label): "NEMAR compute runs a script, a folder of code or a pinned container image against this dataset, from the command line. Jobs run on a supercomputer under NEMAR's allocation." Then the access paragraph verbatim from the shared key compute_access.section.lede ("A separate grant from upload access. An admin reviews each request against export-control and local-jurisdiction rules and the purpose you state, and a grant lets you use compute within the limits an administrator sets."). No other sentence about what access means.
    • Ordered steps, each with the Download dialog's dl-cmd row (<pre><code> plus .dl-cmd__copy with data-copy-text, so the existing delegated copy handler is reused unchanged):
      1. Install the NEMAR CLI and sign in. Link "Install the NEMAR CLI" to installUrl (the existing constant; no install command is repeated, which also drops the old npm install -g, against the project's Bun rule and a second copy to drift). Row: nemar login.
      2. Ask for compute access, once. Link labeled with compute_access.request.cta ("Request compute access") to appUrl("/settings#compute-access", host). Row: nemar compute request-access. One local sentence: "You ask once, and the request tells you what your account is still missing."
      3. Check the job first. Row: nemar compute submit analysis.py --dataset {id}@{version} --dry-run. "A dry run packs and checks your job and prints it. It sends nothing about the job and needs no sign-in."
      4. Run it on this dataset. Rows: nemar compute submit analysis.py --dataset {id}@{version} --wait and, for a folder, nemar compute submit ./study --entry run.py --dataset {id}@{version}. Local sentences: "Name the dataset with its version so the job can be run again on the same data." "--wait follows the job; Ctrl-C stops waiting and does not cancel it." "What one job may ask for is in nemar compute limits; what is left of your balance is in nemar compute balance." No flag values or numbers.
      5. Follow it and collect the results. One multi-line row, in the Download dialog's multi-line style: nemar compute list / nemar compute status <job> / nemar compute logs <job> / nemar compute results <job> --download ./results / nemar compute cancel <job>. Local sentence: "<job> is the id submit prints. Your jobs also appear on your compute page." (true: E2's jobs list is merged). Then nemar compute queue with the shared key compute_page.queue_note ("It describes the queue of the whole platform, not yours, and it is a recent average, not a promise."), which is the only statement about speed.
    • Links row: "Open your compute page" (appUrl("/compute", host), label compute_access.granted.link) and "Install the NEMAR CLI".
    • Close button and corner X as today.

    How the dataset id is used: --dataset <id>@<version> with the page's datasetId and version (the selected version, so a visitor on ?v= gets that pin). The id passes through unchanged for every band; the version is validated against ^v?\d+\.\d+\.\d+$ and passed through as the page holds it (the CLI and Worker accept bare, v-prefixed and latest; the help example is bare).

    Honesty rules (all enforced by tests, section 7): the dialog never says compute is free, available, fast or that an allocation will be granted; it states no limit, price, balance or allocation size (they are the owner's and change without a deploy); "an admin reviews" comes from the shared key; every command is in the captured CLI help; it never mentions S3, a storage URL, rclone or the Neuroscience Gateway.

    States other than ready (decided by a pure builder, section 4):

    • closed (compute not open in this build): same dialog shell, no commands. Body is the shared key compute_access.terms_unpublished.body ("Compute access is not open yet: the compute terms are not published. Please check back later.") plus one local sentence: "To work with this data now, use one of the routes in Download." No link to the old routes; no copy of them.
    • not_available (anonymous deposit, unpublished or no version, id outside nm/on, malformed version): no commands, one sentence: "NEMAR compute cannot read this dataset." and the same Download pointer. One user-facing sentence for every reason, matching the backend, which also does not say why. The reason word exists only in the builder's result for tests. Anonymous deposits get no route that discloses a repository: the dialog never reads githubUrl.

    4. Files and the pure builder

    New / changed (Sonnet implementer; atomic commits, tests first):

    1. scripts/capture-compute-cli-help.mjs: NEMAR_CLI_DIR=<Epic D checkout> bun scripts/capture-compute-cli-help.mjs runs bun run src/index.ts --no-color <cmd> --help-all for login and compute plus each of the ten subcommands, writes test/fixtures/compute/cli-help/<name>.txt, and records the nemar-cli commit in a header comment in an index.json. Same style as capture-compute-read-fixtures.mjs. Real output, never hand-written.
    2. src/lib/compute-instructions.ts (pure, no imports from Astro):
      • buildComputeInstructions(input: { datasetId: string; version: string | null; anonymous: boolean; unpublished: boolean; open: boolean }): ComputeInstructions, a discriminated union: { kind: "closed" }, { kind: "not_available"; reason: "anonymous" | "unpublished" | "no_version" | "id_band" | "version_shape" }, { kind: "ready"; datasetId; version; steps: Step[] }. Order of precedence: closed, then the not_available reasons, then ready.
      • Each step: { id, heading, rows: { label, command, explanation }[] }. command is the exact copyable string. A small exported commandWords(command) returns { sub: string[]; flags: string[] } for the parity test.
      • id_band rule: ^(nm|on)\d{6}$ only. This is a UI pre-filter on facts the page already holds, not a copy of the backend rule; the backend stays the authority (dataset_not_eligible, which the CLI prints). The reserved-band ceiling is deliberately not reimplemented (A2 asks the backend to expose eligibility).
    3. src/lib/compute-labels.ts: add a datasetDialog group for the local labels (title, step headings, row labels, the short explanations, link labels, the two fallback sentences). Update the file header (it says "of the /compute page"). No sentence about access or limits is typed here.
    4. src/components/ActionBar.astro: new props anonymous: boolean and unpublished: boolean (the page already holds anonymousDeposit and unpublished; githubUrl === null is no longer used as a proxy). Frontmatter computes open = computeRequestsOpen(COMPUTE_REQUESTS_ENABLED, apiBase()) and the builder result, plus appUrl(path, Astro.url.hostname) for the two links. Replace the compute dialog body; delete .compute-dialog__soon* CSS; rename the button hook and the script selector; add aria-labelledby; give each new <pre> tabindex="0" (scrollable region on a phone). No other dialog is touched.
    5. src/pages/dataset/[id].astro: pass anonymous={anonymousDeposit} and unpublished={unpublished} to ActionBar.
    6. Docs inside the repo: AGENTS.md architecture map gets compute-instructions.ts and the cli-help fixtures; ADR 0030 gets an addendum (section 6).

    No new account-copy.ts key is needed (section 6), so nemar-cli's shared/contract/account-copy.ts and the drift test are untouched.

    5. Decision: degradation when compute is off or the release gate is not met

    Reuse the one switch ADR 0030 already owns: computeRequestsOpen(COMPUTE_REQUESTS_ENABLED, apiBase()). In the production build it is false until the owner publishes the terms and flips COMPUTE_REQUESTS_ENABLED in a reviewed change; against any non-production backend (staging, local) it is true. The consequences are the ones we want:

    • A production release made before the production API serves /compute shows the closed dialog, never commands that fail. This is the ADR 0026-style gate, expressed as the existing flag instead of a request; the epic's "check with a request first" step before promotion stays.
    • The dialog opens on the same day and in the same change as the Settings request form. One switch, no drift between "you may ask" and "here is how to run".
    • The switch is read on the server at render time and the page is edge-cached for 300 seconds; flipping the constant is a deploy, so the cache turns over with it.

    A probe (like ADR 0026's) was rejected: no public compute route exists, the marketing host has no session, and a per-render upstream call would hit the cached page's hot path. If nemar-cli adds a public availability read (A1), a later change can replace the constant for both Settings and this dialog; ADR 0030's hand-off already describes that for Settings.

    6. Copy: shared keys vs local labels

    Rule (ADR 0030, nemar-cli ADR 0045): a sentence about what access is, what a state means, or what a limit is comes from account-copy.ts; a label or a description of this surface's own layout is local.

    • Reused shared keys, zero new: compute_access.section.lede (what access is), compute_access.request.cta (link label), compute_access.granted.link (compute page link label), compute_access.terms_unpublished.body (closed state), compute_page.queue_note (queue is not a promise).
    • Local labels (COMPUTE_LABELS.datasetDialog): the intro sentence, step headings, row captions, the explanations that paraphrase the quoted CLI help, the install link label, the two fallback sentences. Each explanation maps to a quoted help line in section 1.
    • If the implementer finds an access sentence they want that is not in the contract, stop and file a nemar-cli issue to add it to both repos; do not type it. Extend the existing source-scan in the dashboard UI test so it also covers compute-instructions.ts and the compute dialog block of ActionBar.astro for long sentences outside the labels file and the copy table.
    • ADR 0030 addendum "the dataset dialog (phase E5)": (1) the dialog is state-agnostic because the dataset page is cached and anonymous; (2) it reuses computeRequestsOpen as its gate; (3) the Neuroscience Gateway route is dropped from the dialog and the s3 prefix instructions are removed for good (data-plane rule); (4) amends the sentence "the marketing host links to docs, not to a compute surface": from this phase the dataset dialog links to the app host's /settings#compute-access and /compute, navigation only; (5) the id-band pre-filter is interim and the backend's dataset_not_eligible is the authority. Receipts: website#498, Epic: Website compute surfaces #493; nemar-cli#2021, #2024.

    7. Tests (written first, Vitest, no mocks)

    1. src/lib/compute-instructions.test.ts (pure, real-shaped inputs, the exact shapes ActionBar passes): a published nm dataset with version yields the exact command strings with id and version; ids of each band nm, on pass through unchanged and ds, xx give not_available/id_band; null version, an unpublished flag, an anonymous flag, and a malformed version (abc, 1.0) each give their own reason; v1.0.2 and 1.0.2 both pass; open: false gives closed even for a good dataset; and no string in any result (commands, labels, explanations) contains s3://, npm, nsgportal, rclone, aws , http, a digit-bearing limit (--cores, --mem, --time, --partition, --gpus) or the words free, fast, guaranteed, unlimited. Mutation check: reintroduce s3:// in a label and the test must fail.
    2. test/compute-instructions-cli-parity.test.ts: reads test/fixtures/compute/cli-help/*.txt (capture command named in the file header) and asserts, for every command in every ready row, that each subcommand is listed in compute.txt and each long flag appears in that subcommand's help text, and that the id@version example form appears in compute-submit.txt. Mutation check: change --dataset to --data or list to jobs and the test fails. Also asserts the ten subcommand names the dialog may use are a subset of the fixture's command list.
    3. test/compute-dialog-ui.test.ts (source scan, like compute-dashboard-ui.test.ts): ActionBar.astro has no data-compute-stub, s3://, nsgportal, npm install, website/issues/6, compute-dialog__soon, "coming soon"; it contains data-compute-open, aria-labelledby="compute-dialog-heading", the appUrl("/settings#compute-access" and appUrl("/compute" calls, computeRequestsOpen(COMPUTE_REQUESTS_ENABLED, and does not read githubUrl inside the compute block; every ACCOUNT_COPY[...] key it uses exists; dataset/[id].astro passes anonymous and unpublished.
    4. flags.test.ts already pins the production-closed default; add one assertion that the dialog's gate is that function (covered by test 3).
    5. Gates: bun run typecheck && bun run lint && bun run test && bun run build.
    6. Replacement for the issue's built-page grep (the site is server-rendered): with the dev server pointed at a non-production backend, curl -s localhost:4321/dataset/<id> | sed -n '/data-compute-dialog/,/<\/dialog>/p' contains none of s3://, nsgportal, website/issues/6, npm install; the same with the production default shows only the closed text.

    Browser check (real browser, both themes, desktop and phone width ~390 px):

    • Dataset with a published nm version, non-production backend: button opens the dialog; steps show the page's id and version; copy buttons copy exactly the row text (including the multi-line row) and show "Copied"; the clipboard text has no placeholder other than <job>.
    • A dataset with no published version, an anonymous deposit (nm099998 on staging, never published under a name), and an xx exemplar: the not_available sentence, no commands, no repository link.
    • Production-default build (no PUBLIC_API_BASE_URL): closed text, no commands.
    • Marketing host: Settings and compute links are absolute to the app host; app host or a single-host preview: relative. Signed out, the link lands on sign-in; check whether the #compute-access hash survives the sign-in redirect and record the result (do not fix here).
    • Escape closes; focus returns to the "Compute instructions" button; Tab order is intro, links, copy buttons, Close; the corner X works; the Download and Citations dialogs are unchanged.
    • Phone: no horizontal page scroll; each <pre> scrolls inside itself and is keyboard-focusable; copy buttons are at least 44 px tall on touch; 16 px gutter.
    • Both themes: row, badge and link colors are tokens only (no raw colors added); contrast reads at AA.
    • Screenshots of the ready dialog in light and dark for the PR description.

    8. Removal of the old content and the retarget of website#6

    Remove from ActionBar.astro: the NSG step and its nsgportal.org link, the s3://nemar/{datasetId} text, the npm install -g nemar-cli block, "Just the files" (duplicates the Download dialog's direct route), the .compute-dialog__soon "coming soon" note and the link to website#6, the CSS for .compute-dialog__soon*. Nothing else in the repo referenced these (grep for data-compute-stub, nsgportal, issues/6 finds only ActionBar.astro; s3://nemar elsewhere belongs to Zarr/Use this data and stays). No dead test exists for the old dialog. .context/website-bible.md line about NSG describes the partnership for the about page and stays (the issue puts the about page out of scope).

    website#6 (Phase 5: Live-features architecture + cutover), exact edit, done by the implementer with gh after the PR is open:

    • Title unchanged.
    • Body: add this block as the first lines, and strike two fragments. Do not touch the DNS cutover, accessibility, performance, light/dark or viewer deliverables.
      • Banner: > **Retargeted (website#498):** the NSG connector deliverable is superseded by NEMAR compute, tracked in nemarOrg/website#493 (web) and nemarOrg/nemar-cli#2021 (backend and CLI). The other deliverables were closed out or overtaken on 2026-08-03, see the comments. The dataset page Compute dialog no longer links here.
      • Deliverable 2: wrap the whole bullet Design doc .context/nsg_connector_architecture.md: ... in ~~ ~~ and append Superseded by nemarOrg/website#493 and nemarOrg/nemar-cli#2021.
      • Deliverable 3: Wire Vis / ~~Compute~~ placeholder UI ... strike only the word Compute and append (Compute: superseded, see above).
      • "Out of scope" line: strike only or NSG connector.
    • Comment on Phase 5: Live-features architecture + cutover #6 after the PR merges into the epic branch: "Retargeted by (website#498). The NSG connector is superseded by NEMAR compute (website#493, Epic: Third-party compute: access, user API, nemar compute, partner contract nemar-cli#2021). The dataset page dialog now describes nemar compute and no longer links here. This issue stays open until that dialog reaches production."
    • Do not close Phase 5: Live-features architecture + cutover #6 in E5. The owner's own reopening comment set the closing condition as "the Compute dialog stops pointing here", and production still shows the old link until the staging to main release. The driver closes it by hand after the release, once curl of a production dataset page shows no website/issues/6 and the owner agrees (open question 3). Record in the PR what was verified: the other deliverables are already recorded as done or superseded in Phase 5: Live-features architecture + cutover #6's comments; no Lighthouse run against the original thresholds was ever made and the owner's comment says that belongs in its own issue if it still matters.

    9. What the backend and CLI do not offer (issues to file in nemarOrg/nemar-cli, not synthesized)

    • A1. A public, unauthenticated, cacheable availability read (for example requests_open, terms_published, terms_version) so marketing-host pages and Settings can drop the baked COMPUTE_REQUESTS_ENABLED constant and the terms-version transcription.
    • A2. A per-dataset compute-eligibility fact on a public read (the landing or catalog payload), derived from the same predicate as resolveRun, so the website never copies the band and flag rules. Interim: the nm/on pre-filter above.
    • A3. The job environment is undocumented for users. nemar compute submit --help-all does not say where a job finds the staged dataset, where it writes outputs, or what the working directory is, and docs.nemar.org has no compute command page yet (the Epic D notes list it as not shipped). The dialog cannot teach a script author how to read the data until this is documented. Ask for the help text and a docs page; once the docs page exists, add a "How jobs work" link to the dialog.
    • A4. (Low priority, only if the owner wants it) The job specification supports an include glob list per dataset, but --dataset and the inline header expose only id@version. The dialog does not mention subsets.
    • Not a nemar-cli issue but an owner item: no compute terms page is published, so the dialog has no terms link (the issue says to add it "once published"). The Settings form carries the terms acceptance.

    10. Risks

    • The captured help fixture ages: a later nemar-cli flag rename passes our tests until the capture is rerun. Mitigation: the capture script and commit are named in the fixture index; rerun it at the epic-to-staging step.
    • The dialog can show commands for a dataset the backend then refuses (nm reserved fixtures, withdrawn, tombstoned). The CLI prints the backend's honest sentence; fixed properly by A2.
    • The gate is a build constant: if the owner flips COMPUTE_REQUESTS_ENABLED but the production API is not serving /compute, the dialog opens with commands that fail. Same risk as the Settings form and the same pre-promotion request check.
    • Cached page: a flag flip reaches visitors after the deploy and up to 300 seconds of edge cache.
    • <job> placeholders are copied literally. Accepted and labeled.
    • The #compute-access anchor may be lost across the sign-in redirect for signed-out visitors; the browser check records it.
    • Staging runs single-host, so the absolute-vs-relative link behavior of the marketing host is only checkable by unit-testing appUrl (already tested) plus reading the production HTML after release.
    • The claim "a supercomputer under NEMAR's allocation" is the one sentence not backed by CLI help; it comes from the epic text and the public partnership description. See question 2.

    11. Open questions for the owner

    1. Keep the Neuroscience Gateway as one neutral sentence? Default in this plan: dropped from the dialog; the about page's partnership text is untouched.
    2. May the dialog say "Jobs run on a supercomputer under NEMAR's allocation" without naming the machine or the host institution? Default: yes, unnamed. Say no and the sentence becomes "Jobs run on NEMAR compute."
    3. Close website#6 after the production release (default), or keep it open as a placeholder for a future connector? The reopening comment on Phase 5: Live-features architecture + cutover #6 favors closing once the dialog stops pointing there.
    4. Is a pre-filter to nm and on ids acceptable until A2, or should the dialog show commands for every published dataset and let the CLI refuse? Default: pre-filter.
    5. Keep the button label "Compute instructions"? Default: yes.
    6. When the compute terms page exists, give the URL and its link text for the dialog; until then there is no terms link.
    7. A1, A2 and A3 are filed as nemar-cli issues by the driver at the start of E5 unless you prefer otherwise.

    12. Routing note

    Implementer: Sonnet. No Opus checkpoint (nothing here touches authorization or money; the dialog only prints commands). Review: one Sonnet pass over the diff, with the parity and honesty tests as the focus. Depends on: nothing in code from #496 (the dialog links only to /compute, whose jobs list is merged in E2); the CLI's command names are merged. When merged: update the epic issue, AGENTS.md's map and ADR 0030, and confirm in the epic what the #6 edit was.

  2. neuromechanist commented on Nov 7, 2026

    @neuromechanist
    ContributorAuthor

    Merged in #506 (ea21483). website#6 body retargeted and left open until the production release.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions