fix(dispatch): give every Grid Dispatch lane its own selector strip

[Juliusolsson05]

Problem

In Grid Dispatch, the first (leftmost) lane of every row had no compact index strip. Every other lane had one. Changing the agent in the first lane took two gestures (click into the lane to move focus, then pick from the row's index), while every other lane took one click on its strip. The first lane is also the widest by default, so it's where agents get swapped most.

How it got that way, corrected after review:

• Before Grid Dispatch, one sidebar index always wrote lane 0, so lane 0 correctly had no strip.
• feat(dispatch): grid dispatch mode with per-row projects and no self-filling slots #687 (524671b2) made the index per-row and made it fill whichever lane of its row is focused. For exactly that reason it gave every lane a strip.
• fix(dispatch): wake a hibernated agent before placing it in a Grid Dispatch lane #691 (1eb9a472, a lane-wake fix) then removed the first lane's strip as a side change, arguing that the row's index "IS its selector". That was already false, because the index followed focus.

Change

• Every Grid Dispatch lane now renders its DispatchMiniList strip, including the first lane of each row. A strip always writes its own lane and focuses it, through the waking selectTiledLaneSession path. The index is unchanged and still follows focus.
• Single-lane rows get a strip too. A strip that appeared only when a row gained a second lane would shove the existing lane's content 46px sideways mid-rearrangement. One rule for every lane also keeps the chip column in the same place, which is what makes it scannable.
• Empty-lane hint:
  • It is one sentence in every lane: "Pick an agent from the strip, or press ⌥↓ for the top of the index".
  • It is not shown at all when the row offers no agents, for example a row bound to projects that currently have none. There the strip is empty and ⌥↓ does nothing.
  • The gate uses rowScopedRows, the same filter the strip and ⌥↓ use.
• DispatchMiniList's gridRow prop type now names projectTabIds, the field rowScopedRows actually filters on, instead of the legacy projectTabId.

Design notes

• Why not revert the index to lane-0-only. The follow-focus index was a deliberate choice (clicking a row's index means "I am working in this row now"). A new test pins that premise. If the index ever stops following focus, that test fails first, and the first lane's strip should be reconsidered.
• Why the old bound-row hint branch was deleted rather than "fixed". 524671b2 withheld "for the top of the index" from project-bound rows because ⌥↓ then walked the global list. 29ecd829 later made ⌥↓ walk the focused row's own filtered list (tiledRowScopedRows), which makes the promise true in bound rows. The branch read the legacy projectTabId, which normalizeGridShape folds away, so it was dead and already fell through to the correct copy.

Review

Two orchestrated reviewers ran through the Agent Code orchestration MCP, both on Codex: one on correctness and behavior, one on tests, comments and conventions. A Claude reviewer was attempted twice but never received its prompt (#854), so the second seat moved to Codex. Verdicts were Yes and With fixes. There were no Critical findings. Everything valid is resolved on this branch:

Finding	Resolution
Regression history attributed to the wrong commit (both reviewers)	Corrected in the layout's WHY comment, the plan, this body, and #850 (318ad66f)
Strip-click test only used flat lane 0, where laneIndex, column and start coincide	Now a table over lanes 0, 2 and 3 with focus elsewhere (19dbccd1)
Strip gridRow wiring untested	New test: each strip gets its own row's binding
Expand test only proved a handler was present	Now asserts toggleDispatchRowExpandedParent(rowIndex, parentId) per strip
Row-index test comment overclaimed	Reworded; new test pins that the index fills the focused lane of its row
Strip lookup could silently match more than one	stripSelecting requires exactly one match
Empty-lane hint promised a pick in rows with no agents (pre-existing)	Hint gated on the row offering an agent (6ee33cc0); test was RED first

Mutation-checked: I temporarily applied each realistic miswiring to the layout and confirmed the matching test fails:

• the strip writes column or start instead of its own lane;
• every strip is handed grid.rows[0];
• expand is hardcoded to row 0;
• selectIntoRow always targets start.

Not changed (known limitation, follow-up): the fixed 46px strip plus the percentage-only LANE_MIN_FRACTION (0.08 of the lane region) can leave a lane that was dragged to its minimum with no visible content. This already applied to every other lane; this PR extends it to the first lane. The fix is a minimum-content-width rule for lanes and belongs in its own change.

Tests

gridDispatchLayout.renderer.test.tsx renders the real TiledDispatchLayout over the recorded dispatch-global-d23 fixture.

• Every lane has a strip. 2×2 → 4 strips, [3, 1] → 4. This replaces the two tests that pinned "first lane has no strip". It failed on main with 2.
• Each strip writes its own lane and focuses it, wherever focus is. Lanes 0, 2 and 3. The lane-0 case failed on main, where lane 0 had no strip.
• Each strip gets its own row's project binding.
• Expand addresses the strip's own row.
• A row's index fills the focused lane of that row.
• No hint is shown when the row offers no agents. This case uses an unbound row as the contrast. It failed before 6ee33cc0.

Verification

Node 24 (.nvmrc):

• Dispatch renderer suite: 6 files, 44 tests pass.
• tsc -b: exit 0 (node and web projects).
• CI on the first push (0ef2b17e): quality-gate (full suite, coverage, package verification) and minimum-node-fixture-gate both passed.
• Full npm test locally: could not complete. macOS killed it for low memory (load 99–169, swap about 12 of 13 GB, alongside other agent sessions and the running app).
  • vitest related on the changed components passed: 5 files, 30 tests.
  • Of the files that failed during that partial run, the ones outside this change's import graph were re-run one worker at a time. 4 of 7 pass.
  • store.test.ts and lazy-prose still time out, but only under this machine's load; both pass in CI.
  • imageAttachment.test.ts fails a machine-local provenance check because a cited ~/.claude session file was deleted. CI skips that check.

I didn't launch the app, per the repo rule.

Risk

Low. The change is purely presentational: the first lane of each row is 46px narrower, like every other lane already was, and a hint is hidden in rows that offer nothing. There are no state shape, persistence, keybinding or command changes.

Unrelated CI note: main at 48ea6912 went red in "Enforce coverage baseline" on src/remote-client/src/WebSocketSessionFeed.integration.test.ts. The same test passed in that job's "Test system" step, and it passed on this PR's first run, so it looks timing-sensitive under coverage instrumentation.

Fixes #850

🤖 Generated with Claude Code

https://claude.ai/code/session_01J8PfoJxJ1cvJ9akXsWSVZB

[Juliusolsson05]

Review round complete. Two orchestrated Codex reviewers ran: correctness gave Yes, tests and conventions gave With fixes, and neither found anything Critical. All valid findings are resolved in 19dbccd1 (strip tests pinned to lane, row binding and expand target, each checked by mutation), 318ad66f (corrected #687/#691 regression history) and 6ee33cc0 (no empty-lane hint in rows that offer no agents). One finding is not adopted: the narrow-lane content limitation predates this PR and applies to every lane, so it's recorded in the description as a follow-up. The description has the full table.