stack/agents/darkwing/work/queue-53-review/review-r2.md

Row 53, console tokens and shell restyle, round 2 review (Darkwing)

Issue #1542, request comment 27156, packet comment 27157, queue revs 362-364. Base 2d4d4e7d. Candidate manifest sha256 78c3aef963b7e35500c9f8afa6341057f1a6372ced70915d175e507b5fac8406, the same 16 files under packages/webui. The packet manifest and my snapshot of the checkout both check 16 OK. round2.patch applied to my round 1 snapshot gives the round 2 files byte for byte, so the delta is those five files and nothing else.

Verdict: request changes. All five of my round 1 items are fixed for the cases I listed, and Sage's strip test is there and works. Every suite is green and 35 of 37 mutants die. Two refusal gaps are left, and both show data from before a board refusal:

• A board 403 followed by any other failed read brings the refused session table and the "Waiting on you" card back on screen. Filbert found the same thing (their C4, comment 27163).
• A board 403 while a conversation is open leaves the conversation and its history on screen. I should have caught this in round 1. My probes then never opened History, so they didn't.

Both fixes are a few lines each.

Method

• Detached worktrees at 2d4d4e7d, cand and mutwt, each with the 16 candidate files staged and checking 16 OK. mutwt still checks 16 OK after the mutant runs (r2/mut/manifest-after.txt).
• r2/gate.sh: the packages/webui suite, build-tokens.mjs --check, then every scripts/test-*.sh with DOCKER_HOST=unix:///nonexistent.sock, 18:21:44Z to 18:24:57Z.
• 37 mutants (r2/mut/mutate.py, r2/mut/run.sh), each against the full webui suite, 18:25:08Z to 19:00:50Z. M01 to M20 are round 1's, with M11 rewritten for the strip. M21 to M37 are new and aim at the round 2 code.
• test-release with Docker, finished 19:01:30Z. Then the S5 browser test three times on the unmutated candidate (r2/out/s5-rep-*.txt), see note 3.
• Probes in Chromium through the candidate's tests/browser.mjs (r2/probes/probe2.test.mjs, output r2/probes/probes-cand.txt). The first test is my round 1 probe set again, a fake board and fake bus verbs, plus P9. The second, P10, runs the real control-board from the candidate's tests/history-fixture.mjs behind a small proxy that can answer every request with 403, the way control-board's Host check does. No tracker request.
• Read the five changed files against round 1's required list, Filbert's comment 27143 and IMPLEMENTING.md.

Node v26.8.1, TMPDIR=~/darkwing-scratch/r53b/tmp.

Suites

Suite	Result
packages/webui (node)	27/0
build-tokens --check	rc 0, "tokens.css is current"
test-auth	15/0
test-conductor	17/0
test-config	24/0
test-discord	66/0
test-extension-package	18/0
test-foundation	44/0
test-queue	27/0
test-release	4/0 without Docker, 14/0 with it
test-task, Docker unreachable	26/0

As in round 1, I didn't run test-task with real Docker because it makes live model calls. Dewey's run gave 98/0. Filbert's gate also ran the other node packages, all green.

Round 1 items

Item	Result
Required 1, board refusal	Fixed for a single refusal. The inspector closes, the tree reads "Not read: the board refused the read.", the counts show "–", and the footer, #status and #fresh-board all say not read (P1, P2)
Required 2, bus refusal	Fixed. forget() clears last and the Inbox count from a view, from the board route and from pkOpen. The palette lists only the four views and the section list has no counts (P3, P4, P4b). P7 now shows no task rows
Required 3, tests	Done for the inspector. Not for an open conversation, see Required 2 below
Required 4, M01	Done. no-bus-host runs over kept rows and M01 dies
Required 5, M08	Done. Markup in a question and a task title, and M08 dies
Filbert C3	Done. "Board not read: refused"

The notes: the strip is now one row below 760px (note 1). The toast stays in the accessibility tree while off, at 1×1 with clip-path (note 4, P8). M09 and M16 are tested and die. Note 2 (freshness hidden below 760px) and note 5 (Seen 403 wording) are unchanged, and that's fine.

The strip test

Sage asked me to check it. shell.test.mjs sets 360px and then 400px and asserts four things at each width:

• every #sections a has the same offsetTop;
• the list's computed overflow-x is auto or scroll;
• no label wraps (one client rect, scrollWidth within the box);
• the page doesn't scroll sideways.

All three of Sage's properties are pinned. Three mutants confirm it: M11 (back to wrap and visible) fails "one row at 360px", M37 (wrap with overflow-x:auto) fails the same, and M36 (one row with overflow-x:visible) fails "strip scrolls at 360px". M36 is caught only by the computed-style check, because the fixture's labels fit at 360px and the page never overflows. That check is the right one to have.

Probes

Fixture as in round 1: one board with two sessions (project secretproj, task "Kept fixture task"), one inbox decision and one task. P10 uses the control-board fixture's session, whose history holds "Secret question from before the refusal" and "Secret answer from before the refusal".

Probe	Candidate
P1 board 403 with the inspector open	#board-view display:none, inspector hidden. Fixed
P2 board 403	tree, footer, #status and #fresh-board say not read. Fixed
P3 bus human-required after a read	the palette lists only Board, Inbox, Tasks and Agents. Fixed
P4, P4b the same refusal	no Inbox count or BLOCKING chip, on a bus view and on the board page. Fixed
P6 not-configured after a good read	banner ref, no task rows
P7 refusal, then outcome-unknown	banner err, no task rows
P8 the toast before any copy	display:flex, 1×1, empty, in the tree
P9 good read, board 403, then a 500	#board-view is display:block again with both session rows and the waiting card ("agent1 … secretproj … Kept fixture task … Input needed: kept text"). The banner says "No board data loaded", #fresh-board "the read failed", the footer still "refused"
P9b then a good read	2 rows, footer "waiting 1 · working 1". Recovery works
P10 History open, then every board request 403	#conversation stays shown with the title, the meta line and both secret messages in #conv-log. Beside it, #board-refused says "Nothing to show. The board refused the read, so Console shows nothing rather than a guess."

P9: on a 403, error() sets data = null but leaves the markup in #sessions, #waiting and #seen, and render() returns at once when !data. The table hides only through body.board-refused #board-view. The next error() that isn't a 403 removes that class and the old markup shows again. The README line "A board refusal drops the last scan and everything drawn from it" is not true after that second failure.

P10: #conversation sits outside #board-view (index.html line 17), so the board-refused rule doesn't reach it. convForm() runs and the reply form says "This session is no longer in the board scan", but the history read before the refusal stays on screen. Filbert's T7 predicts this from the code; P10 shows it with the real control-board.

Mutants

35 of 37 killed (r2/mut/summary.txt). Every kill is an assertion or a wait timeout, not a load error.

Mutant	Change	Result
M01	refusal() drops not-configured and no-bus-host	killed (was survived)
M02	error(): refused = false	killed
M03	no body.board-refused toggle	killed
M04	the refused: code not in <code>	killed (was survived)
M05	no stale words in the bus freshness line	killed
M06	System sets data-mode	killed
M07	toast text not escaped	survived
M08	palette label not escaped	killed (was survived)
M09	:focus-visible offset 0	killed (was survived)
M10	focus outline 1px	killed
M11	strip back to wrap, overflow-x visible	killed (was survived)
M12	no section icons	killed
M13	accept() doesn't clear board-refused	killed
M14	task-link ↑/↓ reversed	killed
M15	/icons.svg served as text/plain	killed
M16	no tabular-nums	killed (was survived)
M17	storage key changed	killed
M18	no body.board-refused #board-view rule	killed
M19	refusal shows kept rows	killed
M20	board refusal uses banner err	killed
M21	forget() does nothing	killed
M22	the board route's inbox read swallows a refusal	killed
M23	a view's refusal doesn't call forget()	killed
M24	pkOpen ignores a refused refetch	killed
M25	a board 403 keeps selected	killed
M26	fresh() loses the "refused" wording	killed
M27	a board 403 keeps the footer	killed
M28	a board 403 keeps #status	killed
M29	a board 403 keeps the project tree	killed
M30	a board 403 keeps the three heading counts	survived
M31	freshBus loses the refused state	killed
M32	.toast:not(.on) is display:none again	killed
M33	forget() keeps the Inbox count	killed
M34	forget() keeps last	killed
M35	board freshness in local time	killed
M36	strip one row, overflow-x:visible	killed
M37	strip wraps, overflow-x:auto	killed

M07 survives as in round 1. The toast text comes from validated refs, and Dewey declined it with a reason. I accept that.

M30 survives because the three counts sit inside #board-view, which is hidden while refused. Today that makes it close to equivalent. After P9 it isn't: the counts come back into view with the rows. Filbert's T7 asks for the same assertion. A P9 test that checks the counts kills it.

Required

1. P9, Filbert's C4. After a board 403, a later failure that isn't a 403 must not show the refused scan. Clear #sessions, #waiting and #seen in error() on a refusal, or hide #board-view whenever data is null. Test: a good read, a 403, then a 500, and assert no session row, no waiting card and the three counts not numeric.
2. P10. A board 403 closes an open conversation, or hides #conversation under body.board-refused. Closing is cleaner, because the history then leaves the DOM. Test: open History, refuse the board, and assert #conversation is hidden and #conv-log is empty or hidden. history-fixture.mjs already gives you the conversation; a proxy that answers 403 is about 20 lines (mine is in r2/probes/probe2.test.mjs).

Then fix the README line on refusals if the wording no longer matches.

Notes (not blocking)

1. I agree with Filbert's N4 to N7. N6 (a bus read in flight when forget() runs puts its rows back) is real by the code, and it only restores a read that succeeded, so it can wait.
2. P1 hides the inspector, but its text stays in the DOM. With the inspector hidden that is fine; if you take the "clear on refusal" route for Required 1, clearing it too costs one line.
3. In the M05 run, S5 also failed at bus-browser.test.mjs line 124 ("BODY" !== "H1"). M05 only changes the bus freshness words, so that failure isn't the mutant's doing. S5 passed in the gate, in the 36 other mutant runs and three times alone afterwards. The cause is a race in render() that base 2d4d4e7d has too. The 10 s refresh timer isn't cleared when a navigation starts. If it fires during a navigated read, the refresh (navigated = false) takes gen, the navigation's own render returns early, and nobody focuses the new h1 or announces it. Under the mutant load a read took long enough to hit that window. Not this row's regression and not blocking. clearTimeout(timer) at the top of render(), or carrying navigated into the next render, would close it. Worth its own small row.

Files

• r2/candidate-manifest.sha256: copy of Dewey's.
• r2/gate.sh, r2/out/: suite runs, summary.txt, test-release-docker.txt, the S5 repeats and r1-to-r2.diff.
• r2/mut/: mutant definitions, runner, diffs, outputs, summary.txt and the manifest check after the runs.
• r2/probes/: the probe test and its output.