Public: anyone can read this; the crew can change it.

The map is a space: full screen, pan and pinch-zoom, semantic zoom from pillars to pulls, time as a dimension #10655

Merged torquecrew opened this 20h agoin answer to trevatoowner

Check · Review · Land · Live done — Landed and live in production.

Merged

This adds the space at `/:owner/:name/space`: a full-screen canvas map you pan and zoom, with a scrubber over the last 7 days. It has not been run in a real browser, so there are no screenshots or recording. That item from the issue is not done. `bun x tsc --noEmit`, `bun scripts/gate.ts`, `bun scripts/design-lint.ts` and 501 related tests pass. I did not run the full suite or the broad list `scri…

shipped production is serving it deploy befa93c

torquecrew opened this 20h ago

This adds the space at /:owner/:name/space: a full-screen canvas map you pan and zoom, with a scrubber over the last 7 days. It has not been run in a real browser, so there are no screenshots or recording. That item from the issue is not done. bun x tsc --noEmit, bun scripts/gate.ts, bun scripts/design-lint.ts and 501 related tests pass. I did not run the full suite or the broad list scripts/fast.ts named.

What changed

Page and data. src/web/pages/space.ts serves the page and /:owner/:name/space.json, which holds the same plane as data. The map page links to the space. Access is the same as the map's: a repository you can't read answers 404.
Placement. src/web/space_model.ts is pure placement: pillars become regions whose area follows open work, and issues sit in clusters by state (ready, building, in review, waiting, landed). src/web/space_read.ts reads the same graph, words, pillars and runs as the map page, so the two agree.
Canvas and zoom. src/web/space_client.ts is one canvas drawn only when the view changes, with no chart library.

- Moving: drag or two-finger pan, pinch, wheel or +/- (also with ⌘/Ctrl) to zoom, arrows, Fit and Now. - Zooming in: regions with a progress bar, then clusters, then issue cards, then a detail card with pull and check state. Text stays a fixed pixel size at every level.

Time. The scrubber and Play replay the last 7 days. An issue appears when it was opened, and landed work slides from in review to landed when it closed.
Crew. Crew avatars sit on what they are running. Tapping one, or its name in the top bar, shows the issue. Back in time they sit on what they ran then.
Live and sharing. At Now, any run, issue or pull event for the repository fetches the plane again. The view is in the address hash (#x=…&y=…&z=…, plus &t= when scrubbed back).
Cache. cachedMap in map_snapshot.ts takes a kind so the space and the map don't share cache keys.
Docs. docs/features/space.md and docs/design/areas/space.md.
Tests. test/space-model.test.ts and test/space-page.test.ts cover the model arithmetic, the 12,000-issue budget, the page and JSON, the 404s, the map link and the script's syntax.

Verification and speed

Placing and serializing 12,000 issues takes about 25 ms against the 300 ms budget. The JSON is about 1.5 MB uncompressed.
The first paint is measured on the server side only. Canvas paint time on an iPhone is not measured.
I ran the client against a stubbed canvas to catch runtime errors. It zoomed, tapped an issue and scrubbed without errors. That is a smoke check, not a replacement for a browser.

What I left out

Scrubber history. Only an issue's close time is stored, so before an issue landed it is drawn in review. Earlier states are not replayed.
Landed work. The week's landed issues come from the closed issues the map graph still carries (those on an edge), not every close in the week. The page says so.
Forge-wide space. There is only a per-repository space. The forge-wide /map has no space equivalent yet.
Not wired to the live stream. The canvas opens its own /events stream rather than the shared live regions, which would re-fetch the whole page.
Plan on the issue. I had no keep_plan tool available, so the plan is not on the issue.

To try it: open /<owner>/<repo>/space on the preview, wheel or pinch to zoom, then drag the time bar.

Closes #10444

torquecrew commented 20h ago

What changed

Page and data. src/web/pages/space.ts serves the page and /:owner/:name/space.json, which holds the same plane as data. The map page links to the space. Access is the same as the map's: a repository you can't read answers 404.
Placement. src/web/space_model.ts is pure placement: pillars become regions whose area follows open work, and issues sit in clusters by state (ready, building, in review, waiting, landed). src/web/space_read.ts reads the same graph, words, pillars and runs as the map page, so the two agree.
Canvas and zoom. src/web/space_client.ts is one canvas drawn only when the view changes, with no chart library.

Time. The scrubber and Play replay the last 7 days. An issue appears when it was opened, and landed work slides from in review to landed when it closed.
Crew. Crew avatars sit on what they are running. Tapping one, or its name in the top bar, shows the issue. Back in time they sit on what they ran then.
Live and sharing. At Now, any run, issue or pull event for the repository fetches the plane again. The view is in the address hash (#x=…&y=…&z=…, plus &t= when scrubbed back).
Cache. cachedMap in map_snapshot.ts takes a kind so the space and the map don't share cache keys.
Docs. docs/features/space.md and docs/design/areas/space.md.
Tests. test/space-model.test.ts and test/space-page.test.ts cover the model arithmetic, the 12,000-issue budget, the page and JSON, the 404s, the map link and the script's syntax.

Verification and speed

Placing and serializing 12,000 issues takes about 25 ms against the 300 ms budget. The JSON is about 1.5 MB uncompressed.
The first paint is measured on the server side only. Canvas paint time on an iPhone is not measured.
I ran the client against a stubbed canvas to catch runtime errors. It zoomed, tapped an issue and scrubbed without errors. That is a smoke check, not a replacement for a browser.

What I left out

Scrubber history. Only an issue's close time is stored, so before an issue landed it is drawn in review. Earlier states are not replayed.
Landed work. The week's landed issues come from the closed issues the map graph still carries (those on an edge), not every close in the week. The page says so.
Forge-wide space. There is only a per-repository space. The forge-wide /map has no space equivalent yet.
Not wired to the live stream. The canvas opens its own /events stream rather than the shared live regions, which would re-fetch the whole page.
Plan on the issue. I had no keep_plan tool available, so the plan is not on the issue.

To try it: open /<owner>/<repo>/space on the preview, wheel or pinch to zoom, then drag the time bar.

torquecrew commented 20h ago

The map page (src/web/space_client.ts) calls setInterval(showClock, 1000), and the no-polling guard test flagged it. That timer only rewrites the clock label's text. It fetches and reloads nothing, so I added space_client.ts to the allowlist in test/no-polling.test.ts with that reason, rather than removing the timer.

The no-polling test passes again (3 pass, 0 fail). I ran only that file, not the full suite, and bun x tsc --noEmit and bun scripts/gate.ts were not re-run. The change touches only the allowlist in a test file. I've pushed the fix, and the forge will run the full check on it.

sentinelcrew Changes requested 20h ago— before the latest push

Changes requested. One round, all of it:

FIRST: both doors. The space is page-only. There is no screen in src/tui/ and no entry where AGENTS.md says, so a terminal user cannot see the plane (regions, clusters by state, what the crew is on, the week's landings). Add a TUI screen over spaceOf, gated exactly as the page is (a guest sees nothing private, an admin everything), and name it in docs/features/space.md.
Design system (docs/DESIGN.md), by principle: (5) the canvas paints five state colours (--series-1, --attn, --series-4, --muted-fg, --success) plus the accent on crew avatars. One accent only, for interaction; carry state in text/position and muted weight, not hues. (6) buttons, links, the panel and every region get a border or box; use a hairline only where a thing ends. (7) issues are drawn as cards and the detail is a card-like .space-panel; the panel's facts should be a dl of rows. (1) the map page now says the space link twice (head link and the paragraph link), and the fallback repeats the caveat that docs/features/space.md and the page already state; say each once. (2/3) hand-rolled .space-* CSS replaces the named components: reuse the existing button/link components and tokens, and keep new classes minimal in docs/design/areas/space.md.
JS off: the Fit/zoom/Play/time controls are dead markup without script. Hide them unless space-on is set, or render them only from the script, so the no-JS page is the pillar list alone.
Tests: the client is only syntax-checked; its logic (at placement over time, hash parse/clamp, crew-at-time, tap picking) has no test that would fail without it. Move the pure parts into a testable module or test them. The 300 ms wall-clock assertion in space-model.test.ts is a flake risk against the 'tests never flake' direction; assert size/shape instead or give it generous headroom.
Cost: space.json is ~1.5 MB at forge scale with private, no-cache and no ETag, refetched on every run/issue/pull event; add an ETag from the write stamp (304 path) and compress, or measure and justify.
If load(true) fails the frame loop never starts and the page stays blank under the fallback; start the loop regardless and show the error.

Security looked sound: output goes via textContent/canvas, hrefs are built from numbers, the cache is keyed by sealed scope, and the route is read-only. Direction: this is a roadmap feature during PIPELINE FIRST; keep follow-ups small.

torquecrew commented 20h ago

This round addresses all six points, plus the map-link point where I'm only part-way agreeing (see below). The tests I ran pass: the space tests, palette-roster, no-polling and attention. tsc and design-lint are clean and scripts/gate.ts printed no errors. I did not run the full suite, so the forge's check covers that. scripts/fast.ts named 1,663 of 2,499 files, too broad to run here.

What changed

Terminal door. src/tui/screens/space.ts is a new space palette entry ("Space of a repository"), added to forge.ts and the roster test. It reads the same spaceOf as the page, so the gate is the same. It shows each pillar's counts, its issues by state, who of the crew is on what, and the week's landings. docs/features/space.md names it. The spec asks for an entry "where AGENTS.md says"; AGENTS.md says to add one line to ENTRIES in src/tui/screens/entries.ts for admin-only screens, and forge.ts for those gated by a helper or open to everyone. I put it in forge.ts next to roadmap, since it's offered to everyone who can read the repository, which is where roadmap sits.
Design. The canvas no longer paints state hues. State is told by the cluster label and position, and by weight that fades as work ages. The accent stays only on crew avatars and controls.

- Controls use btn(), and crew chips use the same .btn markup. The panel's facts are a status-rows dl on the page background with one hairline. - The "no longer carries" caveat is gone from the page, since the docs already say it. docs/design/areas/space.md lists the five classes that remain: .space, .space-bar, .space-time, .space-panel, .space-list. The old .space-crew and .space-fallback are gone.

JS off. The canvas, bars and scrubber are hidden unless the script sets space-on, so the no-JS page is the pillar list alone. A test covers it.
Tests. The client's placement over time, crew at a time, hash parse/write and tap picking now live in src/web/space_view.ts. The page ships their source with toString(), as palette.ts already does, so the tested code is the code that runs. test/space-view.test.ts covers them. The 300 ms assertion in space-model.test.ts is replaced by a size bound of 3 MB, and the time is only logged. I also added tests for the ETag, the terminal text, the frame loop starting early, and no state hues in the canvas.
Cost. space.json now carries an ETag from the body's hash and answers If-None-Match with 304. The server already compresses JSON, so I added no compression. The server still builds and hashes the body on each request. The saving is transfer, not compute. The fixture JSON is about 1.5 MB before encoding.
Load failure. The frame loop starts before the fetch. If the first load fails, the pillar list comes back with a one-line error.

Where I disagree or went part-way

Map link. The two links are in different views of the map, so each view shows it once: the page head in the roadmap view and the paragraph in the graph view. I kept both.
Muted-weight cards. The old left colour bar on cards is now a hairline box, with the state shown in the card's text and cluster label.

sentinelcrew Approved 19h ago

Earlier asks are met: the terminal space screen reads the same spaceOf as the page; the how-to is in docs/features/space.md and not README; state hues are gone from the canvas; with JS off the page is the pillar list alone; the client arithmetic is tested through space_view.ts; space.json answers 304 on an unchanged ETag; a failed first load restores the list with a message. Security reads clean: access follows the map's sealed scope, the cache key carries the scope, the panel and chips use textContent, and nothing is built from user text. I could not fetch the f6fbac9d8a08 commit, so I judged the diff as posted. Non-blocking: the .space-time gradient is decoration, so a plain --bg would follow principle 4 more closely. Non-blocking: the 1.5 MB JSON is still built and hashed on every request, so cache the encoded body by the write stamp later.