Point the e2e CLI at the server under test, and roots at the dock

[pcarrier]

Why

CI's e2e job has been red on main since e149dfc (2026-08-16). Two specs fail deterministically, on every PR, so nothing can go green — including #289, whose only failing check is this one.

Both failures are stale tests, not product regressions.

panel-subscriptions — the oracle was a different server

The spec shells out to blit client list. start-servers.sh puts its server on a mktemp -d socket and exports BLIT_SOCK to its own process tree only, so the CLI resolved the default socket instead, started a server of its own there, and read that empty one's client list. Hence Expected length: 2 / Received: [] — the browser's subscriptions were on the other server all along.

It passes locally because reuseExistingServer hands back the developer's already-running gateway, and that one does front the default server.

The script now publishes its socket path in e2e/.e2e-socket for as long as its servers live; the spec reads it and passes BLIT_SOCK to the CLI. With no file (someone else's gateway) the CLI's own resolution is still correct, and a stale file whose socket is gone is ignored rather than used to start a third server.

roots-entry — asserted an entry that was deliberately removed

SwitcherOverlay.tsx says it outright: palette, font, workspace roots and remotes are not switcher entries, because the status bar and left dock already reach them in one click. The spec still asserted a Workspace roots row in Cmd+K. It now holds the app to the contract it actually has — the ⚙ beside the workspace-root selector opens the dialog, and nothing else does — and the comment in createKeyboardShortcuts.ts that still advertised the switcher entry is corrected.

Also

playwright.config.ts gains the CHROMIUM_BIN escape hatch playwright.dev.config.ts already had, so this exact suite can be reproduced outside CI on NixOS (the npm-bundled browser cannot find libglib there). Inert in CI.

Verification

Full suite, locally, CI=1 with an isolated XDG_CONFIG_HOME (so the run sees one remote, as CI does):

• before: 10 failures, including both of these
• after: both pass; 5 consecutive pair runs green

The other local failures (composition-chip, font-size-sync, text-prediction) fail identically with and without this change and do not reproduce in CI — they are this machine, not the branch.

[indent]

PR Summary

Fixes the two e2e specs that have kept CI's e2e job deterministically red on main since 2026-08-16, blocking all merges. Both were stale tests rather than product regressions. (Note: the branch was force-pushed and rebased onto main, dropping the four unrelated media commits that were present in an earlier head; the PR is now purely e2e/test changes.)

• panel-subscriptions.spec.ts: start-servers.sh runs its server on a private mktemp socket exported only to its own process tree, so the CLI resolved the default socket and read an empty server's client list. The script now publishes its socket path to a new gitignored e2e/.e2e-socket, and the spec reads it to point blit at the right server, with fallbacks for a reused developer gateway (no file) and a stale/dead socket. Now that the CLI reaches the real server the spec's two cat sessions are live, so an afterAll(closeAllTerminals) was added to tear them down — a leftover focused cat was echoing typed input and failing later text-prediction specs.
• roots-entry.spec.ts: rewritten to assert the actual contract — workspace roots opens from the ⚙ beside the dock's workspace-root selector and is deliberately not a Cmd+K switcher entry.
• playwright.config.ts: adds the CI-inert CHROMIUM_BIN escape hatch (already in playwright.dev.config.ts) so the suite runs on NixOS.
• createKeyboardShortcuts.ts: corrects a stale comment that still advertised the removed switcher entry.

Issues

Review closed.

_{View session}

[github-actions]

🔗 Preview: https://blit-ly6lg2p4h-indent.vercel.app

[github-actions]

Coverage

Crate	Lines	Functions	Regions
alacritty-driver	73.7% (840/1139)	77.5% (69/89)	77.1% (1363/1767)
browser	0.0% (0/825)	0.0% (0/69)	0.0% (0/1404)
cli	37.8% (5196/13744)	43.8% (511/1168)	38.6% (8077/20900)
compositor	34.9% (5961/17103)	49.3% (495/1005)	34.6% (8182/23649)
desktop	77.6% (4164/5366)	70.4% (367/521)	74.0% (5772/7804)
fonts	85.3% (756/886)	89.9% (71/79)	86.4% (1485/1719)
fssync	92.6% (5518/5961)	94.4% (501/531)	92.8% (10186/10981)
gateway	34.0% (669/1966)	40.2% (68/169)	31.4% (1029/3279)
git	87.6% (4478/5111)	89.6% (362/404)	87.3% (7130/8163)
guest	83.2% (2211/2658)	80.6% (250/310)	81.9% (3708/4529)
lsp	77.2% (2688/3483)	79.4% (262/330)	74.9% (4210/5619)
proxy	19.2% (172/898)	20.5% (26/127)	21.0% (293/1392)
remote	91.4% (15583/17054)	94.2% (1082/1149)	88.7% (25259/28474)
sd-notify	72.8% (67/92)	100.0% (6/6)	82.4% (108/131)
server	55.9% (30939/55362)	63.1% (2656/4209)	57.2% (46915/81960)
ssh	32.2% (165/512)	48.2% (27/56)	31.4% (261/830)
upsidedown	31.4% (391/1247)	27.8% (55/198)	34.8% (797/2287)
webrtc-forwarder	8.5% (238/2805)	10.7% (22/205)	6.3% (289/4595)
webserver	64.4% (1250/1941)	67.6% (173/256)	66.6% (2099/3151)
Total	58.8% (81286/138153)	64.4% (7003/10881)	59.8% (127163/212634)

[blacksmith-sh]

Found 12 test failures on Blacksmith runners:

Failures

Test	View Logs
[chromium] › tests/ mobile-keyboard.spec.ts:200:5 › a tap while the keyboard failed to rise retries instead of hiding	View Logs
[chromium] › tests/ pane-tools.spec.ts:311:7 › Parked terminal does not resurrect › closing the session tha t displaced a parked one shows it, not an empty pane	View Logs
[chromium] › tests/ text-prediction.spec.ts:102:7 › host text prediction › a Backspace through the field de letes one character, not none or two	View Logs
[chromium] › tests/ text-prediction.spec.ts:136:7 › host text prediction › a proposal shows in a chip on th e cursor's line and is not forwarded	View Logs
[chromium] › tests/ text-prediction.spec.ts:184:7 › host text prediction › with prediction off the field st ays empty, as it always was	View Logs
[chromium] › tests/ text-prediction.spec.ts:68:7 › host text prediction › the session runs a shell that edi ts its own line	View Logs
[chromium] › tests/ text-prediction.spec.ts:88:7 › host text prediction › typed characters reach the pty ex actly once	View Logs
host text prediction/ a Backspace through the field deletes one character, not none or two	View Logs
host text prediction/ a proposal shows in a chip on the cursor's line and is not forwarded	View Logs
host text prediction/the session runs a shell that edits its own line	View Logs

...and 2 more test failures. View all on Blacksmith

^{Need help on this PR? Tag @codesmith-bot with what you need.}