Skip to content

feat(event-log): correlation overview, zoom band and unplaced list (Phase 10b, #840) - #1004

Open
adamgell wants to merge 82 commits into
feat/event-logs-correlation-view-839from
feat/event-logs-correlation-overview-840
Open

adamgell wants to merge 82 commits into
feat/event-logs-correlation-view-839from
feat/event-logs-correlation-overview-840

Conversation

@adamgell

Copy link
Copy Markdown
Owner

Closes #840. Part of #823. Stacked on Phase 10 (#839, the Correlation view): this PR's base is feat/event-logs-correlation-view-839, and Phase 10's PR is not open yet. This branch also contains Phase 6b (#835), which Phase 10 does not, so the diff against that base shows 6b's 36 commits as well as this lane's 38 until 6b merges; review this lane's commits, which are the ones that are not 6b's. The range also holds seven merge commits (of 6a, 6b and Phase 10) and one recapture commit.

What it adds

Three additions to the Correlation view, each reading the shared correlation model Phase 10 introduced.

  • EvtxCorrelationOverview.tsx and evtx-correlation-overview.ts (new): a 48 px overview strip of every chain's span, with a draggable and resizable zoom band that sets the lane window of the plot, by mouse and by keyboard.
  • EvtxUnplacedList.tsx and evtx-unplaced.ts (new): the log lines the timeline could not place, behind a "Not placed (n)" row, with a cap of 100 and the omitted count, "Copy as text", and "Open in Log Viewer".
  • EvtxCorrelationView.tsx: mounts the overview and the "Not placed" row, and replaces Phase 10's sticky loss flag with a one-shot loss record (evtx-correlation-selection-loss.ts, new).
  • EventDiagnosisPanel.tsx: the Insights card's unplaced line and its "Review unplaced lines" link.
  • useEvtxCorrelation.tsx, evtx-correlation-store.ts, evtx-correlation-view-model.ts, evtx-correlation-test-fixtures.ts: the unplaced list in the shared model, and the store's "unplaced open" selection kind (never persisted).
  • EvtxCorrelationLanes.tsx: one pure tickLabels shared by the lane axis and the overview.
  • evtx-visual-tokens.ts: a strengthBars token role, its inventory row and a band surface key. No token values change.
  • evtx-timeline-paging.ts and UnifiedTimelineView.tsx: one exported constant for the engine's analysis timeline preview limit.
  • e2e/screenshots/event-logs.spec.ts: the Correlation capture asserts the overview strip and the absence of unplaced items.

Evidence rules

  1. Unplaced items are never positioned on any timeline. A source guard and behaviour tests pin it. The list's caption uses the spec's sentence only when the list is complete and every item is a log line; otherwise it reads "Entries without timestamps are listed, never guessed onto the timeline."
  2. Unknown is never zero. unplacedList is null for no run, and { items: [], total: null } for a run whose count is unknown, so a type decides, not a sentinel. An unknown count shows "Not placed (count unknown)". Each list pairs with its own page's total (knownTotal). The "Not placed" row is the last option of the same listbox, renders with no chains and in the counts-only stage, and the Insights card shows "unknown", never 0, while loading. A card with unplaced items is never hidden by the zero-edge rule.
  3. "Open in Log Viewer" arms a target only when the Log view can consume it. It is on log rows only, and is enabled only when, read from the log store at click time: the open mode is exactly single-file; the open file is the row's file; the loaded entries hold that exact line, and it is untimed (a timed entry at that number proves the list is stale); line numbers in the loaded entries never decrease; and the correlation data is not updating. Otherwise it is disabled with a stated reason. The accessible name starts with the visible label ("Open in Log Viewer: {file}, near line {n}", WCAG 2.5.3). It sets the pending target, then switches to the Log view, with no reload.
  4. No message column. UnplacedItem carries no message (Event Logs: carry a bounded message on each unplaced timeline item #965).
  5. The overview domain is the union of each drawn chain's default lane window. Its caption says it is the extent of the chains shown, not of the session. Bars are 6 px; an unselected chain's fill is at 40% with a full-opacity 1 px outline in its colour (a deviation, following Phase 10's 55% marker precedent).
  6. Band keys. Arrows move the band by 10% of its width, at least 1 s. Shift with arrows resizes it symmetrically, clamped between 10 s and the domain. Home resets it. Escape is spent with preventDefault, so it never reaches the workspace's window handler and never clears the time range. The handles are not tab stops, and there is no click-to-recenter. A band equal to the default is stored as no override. A domain change during a drag re-clamps or ends it.
  7. The overview never writes the transient time range. A spy test pins it.
  8. Height. The bordered box (bars and band) is exactly 48 px, with the caption above it and the tick row below. The total height is a minimum, because the caption wraps. Every new text size comes from logListFontSize through getLogListMetrics.
  9. Axis labels. One pure tickLabels serves the lane axis and the overview and adds the date when the window crosses a calendar day, on 6b's label rules and its one year rule. The band's value text is tickLabels plus the width, so the band slider and the histogram slider (which reads ranges with formatTimeRange) read differently, and in local mode across a DST fall-back the two ends of a range can read alike, where the width tells them apart.
  10. LOG chip colour. The LOG chip uses bg3, not the spec's bg4, because bg4 is 4.36:1 in solarized-dark (a spec errata candidate).
  11. Live regions. Exactly two role=status regions exist in every state, including the unplaced pane. The pane replies through the view's existing action regions, and "Copy as text" is disabled while the data is updating, with the same reason the chain's Copy gives.
  12. One preview limit. The engine's MAX_ANALYSIS_TIMELINE_PREVIEW_ITEMS is one constant in evtx-timeline-paging.ts, pinned to the Rust source, for the unplaced, edge and coverage-gap previews. The per-edge evidence cap is its own named display cap. Duplicate strength rows and the duplicate strength type are removed. This edits Phase 10 code (UnifiedTimelineView.tsx, the chain model).
  13. Changes to Phase 10's view behaviour. (1) The lost-chain sentence is spoken once, where it was re-spoken. (2) With zero chains the visible notice drops "Another chain is shown." and the spoken sentence drops "showing another". (3) The notice disappears when the chain returns or the pane is shown. (4) pickUnplaced is removed; the option and the keyboard call showUnplaced. (5) A loss is recorded only from a current model, where Phase 10 recorded it from an updating one. A sentence counts as said only when a publication made after the loss includes it, and never on an announcement that does not carry it.
  14. Residual of the announcer. With a model present, a second loss whose composed announcement is identical to the current publication (same sentence, same count, within about 5 s) is merged into the first, because a live region cannot repeat identical text.
  15. Forward note. openCorrelationView(chainId) selects through the store and does not clear the loss record. Any lane that selects a chain from outside the view must route through a path that clears it.
  16. Marker clicks are not gated on updating; they resolve against the records loaded at click time.
  17. Trailers. The commit trailers name Claude Sonnet 5.5, the model that wrote them, and were not rewritten.

Scope

38 commits of this lane over 32 files, plus one recapture commit (the replies JSON), against the issue's 5. The CLAUDE.md issue-lane exception applies; each commit has at most 5 files. The files: the overview (EvtxCorrelationOverview.tsx, its two test files, evtx-correlation-overview.ts and its test); the unplaced list (EvtxUnplacedList.tsx, evtx-unplaced.ts and their tests); the loss record (evtx-correlation-selection-loss.ts and its test); the shared model (useEvtxCorrelation.tsx, evtx-correlation-store.ts, evtx-correlation-view-model.ts, evtx-correlation-chains.ts, evtx-correlation-test-fixtures.ts and their tests); the view, lanes and card (EvtxCorrelationView.tsx, EvtxCorrelationLanes.tsx, EventDiagnosisPanel.tsx, UnifiedTimelineView.tsx, EventLogWorkspace.correlation.test.tsx and their tests); the preview constant (evtx-timeline-paging.ts and its test); the tokens (evtx-visual-tokens.ts and its test); evtx-correlation-copy.test.ts; and e2e/screenshots/event-logs.spec.ts.

Known limits

Reviews

  • Gate 1: review at ebb8db4b (2 findings), verifications at a2e7d59c (2) and c6aa991e (zero). Gate 2: review at 9919392e (2), verifications at 017c161e (1), fabaddd7 (1) and 6ee96ae3 (zero). Gate 3: d1451a48 (zero). Merge review of Phase 10's ec4dec57 at 0faa34b7: zero findings, contract conformance met.
  • The last zero-finding verification is 0faa34b7. After it: 6b at cee16160 and Phase 10 at 99e6ae8a were merged in. The first merge conflicted in EventLogWorkspace.tsx (two hunks) and in the engine replies. In the workspace it kept Phase 10's provider-wrapped layout and took 6a's rail and dock resize rules (the range props and proposeResize handlers) on top; the replies took this lane's file and were recaptured after the second merge (86849370, provenance fields only). The second merge conflicted only in the replies.

Verification

At head 86849370709a294086a4bbe669c95d5cc367a7a2:

  • npx tsc --noEmit: exit 0.
  • npx vitest run --maxWorkers=3: exit 0, 212 files, 5069 tests, no "Unhandled Errors". The resolved workspace's suites (EventLogWorkspace* and use-row-capped-size: 9 files, 143 tests) also passed before the first merge was committed.
  • node e2e/fixtures/capture-event-log-replies.mjs --check: exit 1 after the merges because the engine source had changed; after the recapture, only engineCommit and engineSourceSha256 changed, and --check is exit 0.
  • git diff feat/event-logs-correlation-view-839 --check: exit 0. No U+2013 or U+2014 in added lines, in the log's added lines or in commit messages. No real account name or machine identifier in added lines or in git log -p feat/event-logs-correlation-view-839..HEAD.
  • Every commit in the range is signed.
  • src-tauri and crates are untouched against the base, so no Rust gates apply. E2E and screenshots were not run locally (the Vite port is shared).

Screenshots

event-logs-*.png are regenerated, and event-logs-correlation.png gains the strip. Design check: with the fixture the overview strip is mostly empty (five chains over about 15 h, four inside 30 min, so the bars cluster at the right edge); that follows from the domain rule above, not a defect. The new overview assertions are in e2e/screenshots/event-logs.spec.ts. PNGs are regenerated in Main's capture session before merge.

🤖 Generated with Claude Code

adamgell and others added 30 commits October 17, 2026 11:49
…er than a day (#835)

formatTimeRange already printed Oct 16 22:00 for each end of a range that crosses a day. The
histogram axis needs the same reading for one instant, so the form is exported and the range label
uses it.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ry (#835)

A pure module with no clock: the extent of the timed records is the axis (a capped load must not
draw its unread stretch as zero), the buckets are half open so a brush over whole buckets selects
exactly their events, and untimed records are counted apart and placed in none. Edges come from an
injected zone clock, so the DST and half-hour paths are tested without the host's zone. Every
bucket is drawn at the same width and, below a day, lasts the same real time.

The width ladder extends the spec's with seconds and long day rungs (the spec's reaches 48 to 96
buckets only for spans of about 3 hours to 90 days) and never exceeds 96 buckets.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ver (#835)

Which buckets a gesture selects, what a click or Enter does (a second click on the sole bucket
clears, like the Top event IDs rows), and the keyboard model: arrows and Home/End move a cursor,
Shift extends a draft, and Shift on an end bucket of the applied range moves that edge.

The wash is drawn behind the bars, and its opacity is solved from the resolved tokens: the largest
of 60% down to 10% at which every bar, the border, the cursor outline and the focus ring keep 3:1
on it. That is the spec's 60% in six themes and 25% in high contrast, where the cyan selection fill
at 60% puts Error at 1.01:1 and Critical at 1.25:1. The borders use colorNeutralStrokeAccessible
where the spec names Foreground3 (2.34:1 on the dracula wash).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… tooltip (#835)

Stacked bars of equal width from the token map's level colours, the five-row legend with the zone,
seven axis labels, a visually hidden table of every bucket's counts, and one controlled tooltip
that is hidden from assistive technology because the table carries the same strings.

It draws nothing while a load runs, says so when nothing is timed, and adds a note for events it
cannot place and for coverage gaps: an empty bar over a capped load is not a zero. Axis and legend
text use colorNeutralForeground2 where the spec names Foreground3 (3.37:1 on solarized-dark and
3.03:1 on dracula, under the 4.5:1 text needs).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…#835)

Drag, click, or Left/Right/Home/End/Shift/Enter on one focusable slider named Time range. The
gesture works on the buckets it started on (a live tail cannot regrid them under it) and sets the
range once, on release or Enter, from whole bucket edges. A click on the bucket that is already
the whole range clears it; Escape cancels a pending selection first and passes up to clear the
applied range second.

The overlay is drawn from the store's range, so the Time chip, Escape, a load start and a refused
commit all keep it in step. Layers paint wash, bars, edges, cursor in document order, and a test
reads that order from the rendered DOM and composites it in every theme: every level bar keeps
3:1 on the wash. The tooltip is drawn by the component, not Fluent's Tooltip, whose document-level
Escape handler marks the key handled and would stop the workspace's Escape from clearing a range
while a tooltip showed.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The histogram throttles its records the way the Insights rail does. Importing the hook from the
rail's module made the histogram depend on the rail, and three workspace tests that replace that
module with a stub lost the hook. The rail re-exports both names, so no importer changes.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… clock still while brushing (#835)

The histogram sits between the chip row and the coverage banner (spec 5.1), in every view, with no
panel toggle. It is given the events the filters leave WITHOUT the time range, from a context input
that is stable while a range is set, so brushing cannot change the chart it selects from. Without
a range the visible set is that context, so there is still one pass over the records.

Transient criteria no longer move the clock: a time range or a record scope left the nowEpoch memo
dependencies, so a relative window (1 h to 30 d) does not slide when the operator brushes. This
changes Phase 6a's record 17, which had them re-anchor it.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…t spec 5.1 and 8.5 (#835)

The 56 px bar area, the 84 px legend column and the 16 px axis, each to within 2 px, its position
between the chip row and the body, and the fixture's 72 buckets of 20 minutes. Not run in this
lane (the capture harness shares port 1420 across lanes); the first run is Main's capture session.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… that fits, not a two year bucket (#835)

chooseBuckets kept the rung nearest 72 even when its count was over 96, then rejected it and fell
straight to a multiple of 365 days. Every jump wider than 2x has a window of spans that hit this:
a live channel's 3 h 15 min charted as ONE bucket named 'Events per 730 days'. Only rungs that fit
in 96 buckets are candidates now, and the multiple is taken only when every rung is over 96.

Tests: the six jump windows (2 s to 5 s, 2 min to 5 min, 2 d to 7 d, 14 d to 30 d, 30 d to 90 d,
90 d to 365 d), the 3 h 15 min live case, the 205.8 day sample, and a dense property check that
the choice is a ladder rung in 48 to 96 wherever one exists and the finest that fits where none.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ST fall-back (#835)

A first record in the repeated hour floors to a wall time that resolves to its FIRST occurrence, an
hour before the record, so up to 60 empty buckets led the chart over a stretch that was never read.
The first edge now steps forward while its bucket still ends before the first record.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…#835)

Ruling R4 says the tooltip always names errors, warnings and info. An empty bucket read 'no events',
which the same string carried into the slider's value text and the hidden table. The zone suffix
stays, because evtx-time requires a time to be labelled with its zone; both are differences from
spec 8.5's tooltip format.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…#835)

The 2 px floor included the 1 px separator, so a lone Critical event in a tall bucket drew as a 1 px
red line under a 1 px line of surface, easy to miss for the rarest and most severe level. The floor
is now 2 px of colour (3 px with the separator) and fitScale absorbs the extra height.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
formatDayClock and formatTimeRange take withYear: 'Oct 16 2026 22:00', and a range always shows
day and year at both ends, so a range a year wide cannot read with its start equal to its end.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…n crosses one (#835)

The axis, the figure name and every bucket's text (the tooltip, the slider's value text and the
hidden table) printed month and day only, so buckets in different years read alike, and on the
365 day rung a bucket read with its start equal to its end. The histogram now carries showYear:
the domain crosses a calendar year in the zone, or the buckets are a year wide or more.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…835)

Nothing imports useThrottledValue from the rail: the histogram and every test import evtx-throttle.
AGENTS.md says to remove obsolete paths, not to keep a compatibility layer for them.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The Deviates-from comment said Foreground3 measures 3.08:1 in solarized-dark, 'under the 3:1 a
boundary needs'. 3.08 passes; only dracula (2.34:1) fails. A test now pins that only dracula needs
the deviation, so the evidence cannot be misstated again.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…in local mode (#835)

The screenshot spec leaves the store in its default local mode and pins only the browser's zone,
so timeZoneLabel('local') prints UTC+00:00; the assertion expected the word UTC and would have
failed in the first capture session before the screenshot was taken. The capture's mode is
unchanged, and a unit test pins the string the assertion matches. Not run in this lane.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ives way to the edges (#835)

Forced colors repaint the surface as Canvas but the bars keep their own colours, so a bar could sit
at about 2:1 on a black Canvas, and the wash's 3:1 proof (made on the theme's surface) no longer
held. Each segment gets a CanvasText outline and separator under the media query, the wash is not
drawn, and the edge borders, forced to a system colour, carry the range.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… cursor's text is a line in the caption (#835)

At the keyboard cursor the tooltip sat over the Time chip and the 'Showing n of m' readout for as
long as the plot had focus, and nothing dismissed it (Escape clears the range, not the tooltip).
Hovering still shows it. A focused plot shows the cursor bucket's text inline in the figure's caption
line: persistent content, not an overlay. The slider's value text carries the same string.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…835)

While a Shift selection was pending, the anchor index came from the grid that selection froze, but
the press then froze the live chart. If a tail had regridded the chart in between, the anchor was a
bucket of one grid and the drag ran on another, so its first edge was not under the pointer.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…a load runs (#835)

The slider's role and tab stop were spread only while the chart was ready, so a load that started
with the plot focused dropped it from the tab order and focus fell to the body, to start again from
the top of the page. While loading (or catching up on a throttled arrival) the same element now
keeps role, tabIndex and name with aria-disabled, and takes up where it was when the load ends.
With nothing to chart and no load it still offers no slider.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… an alias of it (#835)

INSIGHTS_DEBOUNCE_MS was a second name for DERIVATION_THROTTLE_MS that nothing else imported.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…side it (#835)

While loading the plot is a disabled slider, and a slider's children are presentational (ARIA 1.2),
so the polite 'Loading events...' status nested in it might not have been exposed to a screen
reader whose focus was elsewhere. The message is now a layer over the plot in a box they share.
The slider keeps its value text 'Loading events'.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…om the chart that follows it (#835)

Since the plot stays focusable through a load, nothing reset what a gesture held when the chart went
away: a pending Shift selection kept the OLD grid frozen, so after the load the bars, axis and value
text showed the previous set's data and Enter committed a range from it; a drag in progress went on;
and focus that arrived during a load pinned the cursor to an empty chart's first bucket.

When the chart stops being ready for a load (or the first events of the next set are arriving) the
frozen grid, the drag, the anchor and the cursor are dropped, in the same render. The cursor is no
longer stored on focus: with no keys held it is derived from the chart, so it starts at the newest
bucket once the chart is ready.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…or the keyboard only (#835)

The figcaption mounted when the plot took focus and unmounted on blur, so with no notes the figure
changed height by a caption line each time: a drag on the plot moved the grid down, and clicking a
grid row blurred the plot and moved it back up before mouseup. A pointer press also focused the plot
and showed the line, which the keyboard line should not do.

The caption is always present: the notes on a line of their own above a cursor slot of fixed
height, so focus and blur change only the slot's text. The cursor line shows on Tab focus or after
a key and is cleared by a press.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…nvasText (#835)

The segments were outlined for forced colors but the legend's colour key kept its level colour with
no ring, so the Critical swatch (#881337) was about 2:1 on a black Canvas, the filter row's level
dots being ringed for exactly this reason. The swatches get the same CanvasText ring.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ever reads start equal to end (#835)

timeRangeNeedsYear says when a range must carry the year: either end is not in the current year in
the chosen zone, the ends fall in different years, or the range is 365 days or longer. A range of a
year or more names it whatever the caller asked, and a range under a second adds milliseconds, so
no label reads with its two ends alike.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…#835)

The chip, its tooltip and its remove button's name formatted the range without a year, while the
histogram that sets it now names the year on a multi-year chart. A bucket brushed there read
'Time: brushed Mar 3 00:00-Apr 2 00:00' beside a value text that said 2025, and a 365 day bucket
read with its start equal to its end.

The year is named when either end is not in the current year in the chosen zone, the ends fall in
different years, or the range is 365 days or longer. Tests that pin a label now fix the clock (Date
only) or compute the label with the same rule, so they do not change on 1 January.

This touches EvtxFilterChips.tsx, a file outside the lane's earlier diff (file 16 of the ledger).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…range inside a leap year, and block-level caption lines (#835)

Three mutants of the previous fixes survived: the chip comparing with the real clock instead of the
one it was given (the file pins the real one to the same year), the 365 day clause of the year rule
(a range whose ends share the current year), and the notes sharing the cursor line as a span. Each
now has a test that fails for it.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…#835)

A label names the year when it is not entirely within the current year in the chosen zone, crosses
a calendar year, or spans 365 days or more. The histogram's labelsNeedYear was a second copy that
only knew the last two, so a chart of a single past year read 'Mar 3 00:00' on its axis, tooltip,
value text and table while a bucket brushed on it gave a chip that said 'Mar 3 2024'.

buildHistogram now sets showYear from timeRangeNeedsYear (evtx-time) over the charted domain, with
the clock injectable, and the copy is deleted. The histogram tests pin Date as the chip tests do.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
adamgell and others added 26 commits October 17, 2026 20:35
…d pane (#840)

- The overview sits between the header (with its marker-shape legend) and the lane plot. Its band is
  the lane window: the view passes it to buildChainPlot, as undefined for the chain's default
  window (null would mean no window and draw nothing), and resets it when the chain changes. A live
  rebuild of the same chain keeps it in absolute time. It never touches the time range.
- "Not placed (n)" is the last option of the same listbox, so the arrow keys, Home and End reach it,
  and selecting it replaces the detail area with the unplaced list. It is drawn when there is no
  chain at all and in the counts-only stage, from data.unplacedList; null is no row, and an unknown
  count reads "Not placed (count unknown)".
- The reconcile effect never calls selectChain while the pane is open, so "Review unplaced lines"
  before any chain is selected keeps it open. A pane left with nothing to list is left for the
  chain, and said once. The plot is not built under the pane.
- The pane replies through the view's two action regions: exactly two role=status regions hold in
  every state, and a failed copy is an alert.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…e placed and links to review it (#840)

Phase 4 deferred the unplaced line and "Review unplaced lines" (spec 8.13 item 6, 8.11a) to this
phase; the deferral comment is removed.

- The line reads counts.unplacedList, the newest session's, never the model side that can be an
  older session's. A complete list uses the existing unplaced summary wording with the kind split; a
  partial one says the engine's count and not the preview's length; an unknown count, or counts that
  are not in yet, say the number is not known and are drawn muted. None of them is ever a 0.
- "Review unplaced lines" calls openUnplacedList, which shows the Not placed list in the Correlation
  view and leaves the rail tab and the chain selection alone. It is offered for a counted non-zero
  and for an unknown count over a run, and not while the counts are loading.
- A session known to hold no edge still has no card, unless it has unplaced items to review: the
  zero-edge rule no longer hides the link to them.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…dule imports (#840)

panWindow, resizeEdge, bandStepMs, overviewPaletteFor and the types OverviewBar, OverviewBars and
BandKeyResult had no importer outside the module but tests. They are private. The tests reach
panning and resizing through dragBand, the key step through reduceBandKey, and the palette through
a reader that takes the colors straight from the token map (tokenPalette in the fixtures), which is
also a better test: it no longer measures the code against its own palette.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…re private (#840)

LogViewerLines and OpenAtLine (evtx-unplaced.ts) and the two components' props interfaces had no
importer outside their module but tests. They are private; the tests take a component's props from
ComponentProps<typeof Component>.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
#840)

Neither is imported outside the module that declares it: tickLabels' options are written inline by
its callers, and the hook's list type is reached as EvtxCorrelationData["unplacedList"].

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…en themes (#840)

The EVT chip (the selection triplet), the LOG chip (bg3 with Foreground2, where the spec's bg4 is
4.36:1 in solarized-dark), the header, the caption and the row text are measured on the surface in
every theme. The chips set backgroundColor, as the rest of the workspace does.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…#840)

bandValueText and the tick labels compare a label's year with the real
clock when no nowMs is given, and the component calls them that way. Two
overview tests asserted exact clock text with no pinned clock, so they would
fail from 2027 when those strings gain the date and year. Date is faked for
the whole file (timers stay real), as the Lanes axis tests already do.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
#840)

The button showed "Open in Log Viewer" but was named "Open <file> in the
Log Viewer near line n", which does not contain the visible words in order,
so a speech-input user saying the words on the button missed it (WCAG 2.5.3
Label in Name, level A). The name is now "Open in Log Viewer: <file>, near
line n", built from one shared label constant that the button also shows.
R4's "near line n" stays. Tests pin the new name in the pure and component
layers and read the name against the button's visible text, enabled and
disabled.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ed (#840)

Open in Log Viewer arms the pending scroll target and then switches to the
Log view. setActiveView refuses a workspace that is not available, and a
target armed with no Log view to consume it would scroll a later load, which
R4 forbids. The Log view is always available in a real build (the backend
workspace list starts with it), but the switch is now checked: when the
active view is not the Log view afterwards the target is cleared and the
refusal is said in the action region. A test enables only the Event Logs
workspace and clicks.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…tent note (#840)

The caption row had a fixed height, no wrap and hidden overflow, so at a
large log list font size with the side panes open, "Spans the chains shown,
not the whole session." (which R5 requires the strip to say) could be cut
off. The row now has a minimum height, wraps, and keeps the separator dot
with the hint so a wrapped line never starts with it. The strip's total
height is therefore a minimum: 83 px at f=13 and 97 px at f=20, on one line.
A test reads the styles that would clip (jsdom has no layout).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…840)

EMPTY_UNPLACED_LIST (added by this lane) and makeCounts (re-signed by it)
were exported but imported nowhere in the repository; both are used only
inside the fixtures module. An import-statement audit of the modules this
lane added or changed finds no other export without an importer.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ange alone on the band (#840)

The overview tests checked that Escape on the band is spent with
preventDefault against a stand-in for the workspace's handler. The
workspace correlation test already renders the real lazy view, the real
hook and the real useClearTimeRangeOnEscape, so this drives the actual
chain: Escape on the band keeps the 6a time range, moving the band never
writes it, and Escape elsewhere still clears it.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…as no override (#840)

Condition 4 says the plot gets `undefined` for the default window. A drag
cancelled with Escape (or any move that lands exactly on the default)
restored the band to a concrete window equal to the default, so the plot
stopped following the default when a live rebuild changed it. The view's
single entry point for the overview's window now stores no override when
the incoming window equals the chain's default at both ends. Tests: the
plot is built with an undefined window before a gesture and after a
cancelled drag or a key move back onto the default; a rebuild that moves
the default is followed after a cancelled drag; and a window equal at only
one end (either handle, on the first chain and on a middle chain) stays a
concrete override.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
)

Phase 10's selectionLost and this lane's unplacedLost were sticky flags
that only a pick cleared: their sentences were spoken again at every later
model or updating change, said "showing another" or "showing a chain" with
zero chains, and "Nothing is listed under Not placed" after the pane was
shown again or a later session listed items again.

This is the pure half of the fix: a loss is a record made once, at the
reconcile that found it. What is said is derived from the record and from
what exists now: only while it is still true, naming a chain only when one
is shown, and only until the polite region has published it. The view uses
it in the next commit.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… true (#840)

Review F1. unplacedLost (this lane) and selectionLost (Phase 10) were
sticky flags that only a pick cleared. Their sentences were spoken again at
every later model or updating change, said "showing a chain" or "showing
another" when there were zero chains, and said "Nothing is listed under Not
placed" after the rail showed the pane again or a later session listed
items again.

Both are now one event record, made at the reconcile that found the loss.
What is said is derived from the record and from what exists now: it is
dropped when it stops being true (pane shown, items listed again, chain
back) or on a pick of a chain; it names a chain only when one is shown; the
polite region gets it until it has published it once, and the visible
notice for a lost chain stays only while it is true. The loss is also
recorded only once the model is current, as the replacement is.

Tests: zero chains, the rail reopening the pane, a later session listing
items again, the chain coming back, said once across a later model and an
update, a pick dropping it, and the items going away again with the pane
closed, for both the pane and the chain.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ing (#840)

Review F2. While the data is updating, the list the pane shows belongs to
an earlier session than the one being built, and the copied text carries no
mark of that. Open in Log Viewer already waited for this reason; Copy did
not. It now uses disabledFocusable with the reason the chain's Copy gives,
spoken (aria-description) and shown (title). The reason is one constant,
shared with the chain's actions.

The audit of every action in the pane, the overview strip and the card that
acts on the model on show is in the lane report.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… an earlier text (#840)

Review fv1. The receipt marked a loss announced when the throttled
announcement contained its sentence. That text can be a publication made
before the record existed, and one kind's zero-chain wording is a prefix of
its chain-shown wording, so a second loss of the same kind inside the 5 s
window was marked as already said and never spoken.

The record now captures how many publications the polite region had made
when it was created. The view counts publications itself (the shared
throttle does not report them and is not touched): the region text changes
exactly when it publishes. A loss counts as said when a publication made
after it carries its sentence, or when the announcement composed with it is
identical to what the region already holds (the region would not publish it
again, so the sentence would otherwise sit in every later announcement).

Tests: a second loss of the same kind inside the window, for the chain and
for the Not placed pane; the zero-chain wording after the chain-shown one
that contains it, and the reverse; an identical announcement that is neither
published again nor sticky; and one sentence said once when the data starts
updating in the same step as the publication.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
#840)

Review fv2. The identical-announcement rule marked a loss said whenever the
composed announcement equalled what the region held, without checking that
it carried the sentence. With no model the announcement is the no-analysis
text, the reading text, or empty, and never carries it, yet the record stays
true; when that text matched the region the loss was marked said although it
was never spoken, and a session returning with the replacement shown
recorded no new loss.

The rule now also requires that the composed announcement carries the
sentence. Every branch that composes without it (no model, in its three
texts) leaves the record unsaid until the first later publication that
carries it. Tests: the pure rule for the three texts; the view for each
text, then a session returning with the replacement shown, and for the Not
placed pane. The count of the publications is now mirrored into its ref in a
layout effect instead of during render, so it is the count of the commit the
effects run in.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Merges the clean Phase 10 head (engine-aware matcher, malformed-identity
mirror, recaptured engine replies with the vectors tests) into GATE 2 of 10b.
Never a rebase.

One textual conflict, e2e/screenshots/event-logs.spec.ts: both lanes pin the
page clock at the same instant in the same place with different comments.
One setFixedTime call is kept, with both rationales (the year in day labels,
now also the overview strip's ticks; and the workspace's single nowEpoch
read plus the warning not to add a live-stream capture here).

Everything else merged without conflict. Phase 10 did not change
EvtxCorrelationView.tsx, so 10b's one-shot loss record stands in place of the
selectionLost flag that Phase 10 carried; the matcher, the mirror and the
shared fixtures are Phase 10's, unchanged.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… absence of unplaced items (#840)

The event-logs correlation capture is the Correlation overview capture: its
PNG gains the strip, which sits between the chain header and the lane plot.
Before the picture the test asserts the strip is drawn (48 px, above the
plot, five bars for the five chains with the selected Exact one solid, five
ticks, the sentence that its extent is only the chains shown) and that its
zoom band is one named slider and the strip's only tab stop. After the
picture it drives the band from the keyboard: a key moves it, Home brings it
back, and Escape on it keeps the brushed time range.

The shared fixture is not changed (R6). It holds no unplaced item, so the end
of the test only asserts absence: no Not placed option (the list keeps its
exact name), no pane, no unplaced line on the Insights Correlations card
once its counts are loaded, and no link to review them. The rows with items
are covered by component tests with engine-shaped items.

Not run: the screenshot capture (Main regenerates the PNGs). Verified with
--list only, and by replaying the captured engine replies through the real
chain builder: five chains, all positioned, five bars.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Merges the re-cleaned Phase 10 head: the correlation detail column keeps
max-content rows so the lane plot is not squeezed (Z1, DETAIL_COLUMN_STYLE,
data-testid="correlation-detail"), and the unified timeline strip keeps a
three-row floor with its height in evtx-timeline-strip.ts (Z2). Never a
rebase.

One textual conflict, e2e/screenshots/event-logs.spec.ts, in the
event-logs correlation test after the screenshot: 10b's band keyboard steps
(ArrowRight, Escape, Home) against Phase 10's short-window layout
assertion. Both are kept, the band steps first, then Phase 10's block
unchanged. One line is added to Phase 10's block: the overview strip sits
in the same column, so its box must still measure 48 px in the short window.

The views merged without conflict. The overview strip is a child of the
detail column (the grid with max-content rows), so it keeps its content
height and the column scrolls; the Not placed pane is another child of it.
10b's one-shot loss record and Phase 10's floors do not touch each other.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…rview-840

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…rview-840

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…d Phase 10 (#840)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@coderabbitai

coderabbitai Bot commented Oct 22, 2026

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review: 1
  • main

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Comment @coderabbitai help to get the list of available commands.

@github-actions github-actions Bot added enhancement New feature or request feature New feature labels Oct 22, 2026
@adamgell adamgell mentioned this pull request Oct 22, 2026
4 of 34 tasks

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement New feature or request feature New feature

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant