[madskronborg]

The Flow used its space poorly and was slow to use, so this polishes it after hands-on use in an isolated dev Desktop against the laptop's and vm-mk's real hosts. No control was pressed on a real Run.

• Map: each Run's Steps now wrap in a grid under a one-line header (title, Machine · execution, Started by). There is no horizontal scroll, and no sticky label covers a glyph. Long Step names are cut to one line, but the tooltip and the spoken name keep the whole name. The inspector sits beside the map once the Flow is at least 1024 px wide. The selected Step stays in view when you select it and when the map resizes.
• Speed: a running glyph now turns in 12 steps instead of smoothly. With the same record and 5 running glyphs, renderer CPU went from ~40% to 14–17% and GPU process CPU from ~27–30% to 2–3%. A running Task's Flow reads agents' delivery ledgers only when a finished Run needs its Reopened check. On a copy of vm-mk's state, taskFlowOf went from ~200 ms to ~43 ms (median of 7).
• Correctness: Flow action icons render; they were missing from Nuxt Icon's client bundle. A Run replayed into Steps after it launched no longer draws its first Step twice.
• Inspector: times read as clock times ("11:22", "23 Oct, 09:00") instead of ISO instants. "Started by" now sits with the Step's facts.
• Board card (steered in by the human): the branch and pane location each take one line, cut with the full text on hover, and the pane line leads with the tab. The Flow strip's label wraps under the glyphs instead of being cut. At a ~330 px column, a card went from 334 px to 231 px tall, with no status or action removed.

The before and after screenshots and timings are on mk-laptop: the measurements are in ~/.local/state/herdr/plugins/cego.collie/runs/run-3df1ab2e/handoffs/flow-polish-measurements.md and the screenshots in handoffs/screenshots/. The host-side gains (ledger reads, replayed Steps) reach a Machine only once its host runs this release.

Verifications: the gate's check passed on 2cb95eb5, the commit before this head. On this head (a2c8ad99), card-tests and desktop-typecheck passed. Every other recorded verification ran on an earlier tree.

Verification this run is held to:

• check: /home/mk/.local/share/mise/installs/bun/latest/bin/bun run check (in /home/mk/.herdr/worktrees/.collie/madskronborg/test-and-polish-the-new-flow-feature-in-3u9a28)

The board card commit (a2c8ad99) came after the review and after the gate, so its own tests checked it, not a reviewer. It changed:

• BoardCard.vue: tighter padding; the branch and pane are each cut to one line, with a title; the Flow strip wraps.
• paneLine in desktop/src/shared/terminal-input.ts and usePanes.ts: the pane line reads tab · workspace.
• test/desktop-card-pane.test.ts, which failed first.
• The pane sentence in docs/using.md.

Non-blocking finding left open:

• test/desktop-icon-bundle.test.ts:31 (minor), "Record a failing baseline for the new icon-bundle test". The record already holds one: icon-bundle ran with --expect fail on e53cd719, before the fix, and failed (exit 1). It names the five missing icons (message-square-plus, terminal, message-circle-question, scroll-text, file-check).

Not settled by the run

• assumed in build: The worktree was cut at 0.91.0, before the Flow landed, so I fast-forwarded the branch to origin/master e53cd71 (0.94.0, which contains Flow Show Task work as a Flow across Desktop, CLI and chat #195) before starting. There is no merge commit.
• assumed in build: No planner was live, so I wrote the spec and tickets in the run's plan dir and chose the test seams myself.
• assumed in build: Ticket 03 changed while I built it. Serving the record's Flow from the latest board build made a record opened right after Pause/Continue show the old Flow for up to one build (~5 s on vm-mk), which 4 host tests caught. A profile showed 147 of taskFlowOf's ~200 ms went to reading every agent's delivery ledger, so the board now reads the ledgers lazily, only once a card has a finished leader. A running Task's Flow dropped from ~200 ms to ~43 ms on vm-mk's state.
• assumed in build: Ticket 06 changed while I built it. Skipping the step-less launch would have dropped the Step's real agent and model. Such a launch is now attributed to the entered Step whose key equals its operation.
• assumed in build: Ticket 02 (stepped glyph rotation) is proved by measurement, not by a unit test. Same record and same 5 running glyphs: renderer CPU ~40% -> ~14-17%, GPU process ~27-30% -> ~2-3%. CSS animation timing can't be observed by the unit tests the repo allows.
• assumed in build: Tickets 04 and the resize follow-up are layout changes with no logic to unit-test. They are verified by screenshots at 1000 px, at the default 1200 px window (673 px Flow), and at 1900 px (inspector beside the map), in handoffs/screenshots/after-*.
• assumed in build: Inspector times use the en-GB 24-hour format ('11:22', '23 Oct, 09:00', '23 Oct 2026, 09:00'), matching clockTime's existing 'HH:MM' style.
• assumed in build: Out of scope, left as found: vm-mk's 4-5 s board builds, which predate the Flow (lastActivityAt/writtenIn walks); the dev-only updates-stream resubscribe loop in a checkout Desktop; and Up/Down moving between Runs rather than between the visual lines of a wrapped Run.
• assumed in build: The host-side gains (tickets 03, 06) only reach a Machine once its host runs this release. The dev Desktop used for screenshots talked to hosts running 0.94.0.
• assumed in build: docs/using.md was updated for the new map layout (invariant 7). The other tickets change no documented behaviour.

🤖 Generated with Claude Code