map: connections read as a route, and zoom shows more instead of bigger

[Yushangjinghong2]

The map looked unfinished in three ways, all seen on the stable web trial (iclr-eval-v5-0917-0359).

Connections were invisible. A map without recorded dependencies has only "the order work happened in" between its cards, and that was drawn as sparse grey dots. It is now a continuous slate line that strengthens toward the card it arrives at, with a small anchor where it leaves its card and a lighter swept arrowhead. The dash is kept for "related work", which is an association and not a path. Step links inside an open card use the same line.

One view mixed three kinds of card. Each card chose its overview tier from its own on-screen width, so headed cards sat next to bare grey tiles. The tier now comes from the zoom alone.

Zooming was ugly. Between overview and detail a card was its small layout magnified: 40px type, the title cut off under the toolbar, the footer spilling out of the frame. Type now lives in a band of screen sizes (a 13–18px title), so a larger card says more at the same size. Content is clipped inside the card, and the title ends on a whole line that fits the card's height.

Also: card chrome (hairline, corner, state rail, shadow) is measured on the screen and is the same at every zoom; a card that keeps earlier parts shows a second sheet behind it; zooming into open canvas no longer enters detail mode with no card in focus, which dimmed the whole map around nothing.

The new look is stated in one stylesheet, frontend/web/src/map/design.css, loaded last. Layout, geometry and the camera's framing are unchanged.

Later rounds on this branch

Ink on paper. The palette, serif titles, paper grain, ink washes and far hills, a brushed stroke for each route (brush.ts), a course row on each card (plan, execution, review, result), a segmented progress bar in the header, frosted controls.

A small card says what its task is about, with a line budget computed from the room it has, so nothing is sliced through.

Zooming is smooth again. The redesign had cut zooming to a quarter of its frame rate. Same harness, 72 wheel ticks in headless Chromium on a ten-task map:

build	fps	frames over 50 ms	layout	style
dev before this branch	50	5	0.8 s	2.6 s
PR #119	46	7	1.0 s	2.5 s
this branch before the fix	12.5	196	6.4 s	4.4 s
this branch now	45–46	7–8	0.3 s	1.3 s

Type and chrome were sized against the live zoom, so every card was laid out again on every frame. They are now sized against the zoom in quarter-octave steps (zoomStep.ts), and a moving map keeps the step it had: a gesture scales one picture, and text is set again when the movement rests. The live zoom is no longer published as an inherited custom property. The washes lost their 36px blur, grain and hills lost their blend modes and filters, labels lost their backdrop blur, and frosted panels turn to paper while the map moves.

A line says what it carries. Every line at overview read "same study". A line that can only name its kind now shows those words when the reader points at one of its tasks. For the rest, POST /api/map-lines/{source}/{name} (argus/webapi/map_lines.py) takes the pairs of tasks the map draws and has the map model write, once per map, what one task handed the next ("figures and tables for the paper"), or nothing where the records show no connection. A note annotates an existing line and never adds one or changes its kind, so the layout does not depend on it. One pass on a ten-task map: about 3.5k tokens in, half a cent, 30 s with gemini-3.8-flash; usage is recorded under map-summary. The trial portal's allow-list and analytics route templates know the new path.

A task reads as its stages again. Inside a task every stretch of tool activity had become a card of its own ("40 steps of work", "said what comes next"): a one-round task was twelve cards and nearly every link between them said "then". Tool activity is now folded into the round of work or review it happened in (the card's footer gives the count, opening it lists the actions), so a task is again goal, plan, work started, a round, its review, the changes asked for, the next round, the result, with the relation named on each link. Steps are named by their stage; a failed tool call no longer turns a step red; a finished task no longer shows "in progress" on the moment it was planned or taken up. Tests: submap.test.ts (folding, a round under way, activity with no round), submapReadability.test.ts; documentation in docs/research-map.md.

A card says what its task found, in the reader's language. A task's full explanation is written only when a reader opens it, and not at all once the project's daemon has stopped, so cards showed the planner's text: an imperative English title, the first words of a specification, and the execution's state three times. POST /api/map-cards/{source}/{name} (argus/webapi/map_cards.py) has the map model write a short title and a sentence per task in one request per map, from the task's goal and the last record its work left (about 3.5k tokens, half a cent, 7-20 s for ten tasks with gemini-3.8-flash; recorded under map-summary). The full explanation stays on demand and takes over the card once it exists. With words on it a card no longer prints the specification under the title and says its state once. A step's detail no longer shows harness instructions or raw JSON arguments and lists eight actions with a count of the rest. Tests: tests/webapi/test_map_cards.py, macroCardCopy.test.tsx, submap.test.ts; documentation in docs/research-map.md.

A stopped project can still be explained. An explanation is still written only for what a reader opens, but no longer held back when the project's daemon is not running (on a trial, every project). With gemini-3.8-flash the writing then failed about half the time because a brief section came back as an object of named parts; that shape is now put right before the schema check (the model's own paragraphs, joined; length limits still apply). Tests: mapCopyLifecycle.test.tsx, tests/webapi/test_reader_brief_shapes.py.

Card words and line notes are sized for the reader's language. They were already written in the interface's language and kept per language, but the limits were sized for Chinese and applied by cutting ("Formulate three long-context LLM researc"). Limits and the lengths asked for are now per language, the schema is wider than what is shown so one overlong phrase does not discard the rest, and text that overruns ends at a sentence, clause or word, never inside a word or a number. Tests: tests/webapi/test_map_cards.py, tests/webapi/test_map_lines.py, relationLabels.test.ts.

Tests added: zoomStep.test.ts, mapLineNotes.test.ts, mapBrush.test.ts, updated relationLabels.test.ts; tests/webapi/test_map_lines.py (one write per map, omitted pairs remembered, a settling task re-asks only its lines, failed attempts cool down), tests/trial/test_map_lines_route.py (portal allow-list, analytics route template). Documentation: docs/research-map.md describes line notes and the stepped zoom.

Checked: tsc clean, web suite 1544 passed, tests/webapi and tests/trial passed, check_artifacts matches 0.1.8; light, dark and phone screenshots against a copy of the trial sessions; the zoom harness above against the deployed build.

🤖 Generated with Claude Code

[Yushangjinghong2]

Completed the reader cleanup in ef418ad; release artifacts are in 368e8c0.

• Removed reader provenance disclosures and stopped generating card source snapshots and redundant relation evidence.
• Added on-demand question context: opening the map or reading an existing explanation does not retain a source; selecting Ask about this step prepares only that explanation and the current task records. Existing saved questions remain readable.
• Kept full explanation generation behind an explicit reader action, including stopped projects, and preserved language-specific card/line copy.
• Added visible API errors and context-switch race coverage instead of silently regenerating explanations.

The existing trial is deployed at 368e8c0. Browser verification confirmed no provenance UI, no background full-explanation/source calls on overview, successful question preparation for the selected task, legacy cache responses without the removed fields, and an explicit 422 for a missing explanation. For the same five cached cards, the map-copy response decreased from 67,855 to 37,068 bytes without regenerating their explanations.

Validation: frontend typecheck and all 1,502 frontend tests pass; 198 targeted backend tests pass; release build/artifact checks pass. The broader backend suite is not wholly green: gateway concurrency/deadline and socket-forwarding timeout cases failed outside the changed code. Their failures were not hidden or addressed with unrelated changes.