Skip to content

v6 verification — nine-scene motion parity, sizes, reduced motion, macOS/Herdr ​

Status: verified locally on Linux/tmux; hosted CI not run; macOS/Herdr checklist manual and unexecuted Date: 2026-11-07 Specification: #113 Tickets: #114–#127 Branch: ticket/127-verification Base: integration/ensemble-v6 at e49a09e

This is the specification-level verification record for v6. It covers the nine motion scenes, the Work-card size matrix, the one/two/three-card load probe, reduced motion, NO_COLOR, and the automated regression gates. It does not close #113 or #127. The macOS/Herdr checklist remains manual and unexecuted.

Linux/tmux is not macOS/Herdr evidence. Everything below was observed on Linux with tmux. No macOS/Herdr host ran any part of this pass, so Braille font coverage, Herdr key interception, native-session key ownership, toast appearance, and actual color/opacity rendering on macOS remain unverified.

Environment ​

ItemValue
OSLinux 6.12.113+deb13-amd64 (Debian 6.12.113-1), x86_64
Terminal hosttmux 3.5a
Nodev26.12.1
npm12.2.1
OpenCodeopencode v2.2.5
Git2.47.3
macOS/HerdrNot available; not run
Plugin type surface@opencode/plugin 2.2.1
TUI packages@opentui/core 0.7.0, @opentui/solid 0.7.0, solid-js 1.9.12, TypeScript 7.0.2
Temp root/tmp/opencode; /tmp had 15G free, so no alternate TMPDIR was needed

The repository type-checks against @opencode/plugin 2.2.1 while the live checks exercise the installed OpenCode 2.2.5 host. These are different version surfaces, as recorded in the runtime guide.

Commands ​

All commands ran sequentially on this branch. No two live suites ran at once and no hosted CI job was triggered.

CommandResult
npm run verifyPass — strict type check, then 785 tests passed, 0 failed, 0 skipped, 0 todo (duration_ms 15252.40).
npm run test:live:fake:allPass — fake 207s, fake-empty 9s; 2 passed, 0 failed, 0 skipped, 0 not run in 216s.
npm run test:live:latencyPass (measurement only) — raw samples 8 8 8 8 9 15 9 8 9 8 ms; min 8, p50 8, p95 15, max 15, mean 9.0, n=10.
npm run test:live:sizesPass — 80×24, 100×30, 120×40, 160×50 Work matrix; 1/2/3-card load probe; NO_COLOR=1 still-motion check.
npm run test:live:preflightPass — node v26.12.1, opencode v2.2.5, tmux 3.5a; fake scope does not require fd.
bash -n scripts/live-size-matrix.sh scripts/live-check-fake.shPass — no shell syntax errors.
git diff --checkPass — no whitespace errors.

The latency result is comparable to the existing baseline, not a new threshold. Ticket #106 recorded a p50 of 9ms for the same rail j/k probe; this pass measured p50 8ms. The probe is machine- and tmux-specific and reports no pass/fail gate.

Nine-scene coverage ​

The full test:live:fake scenario covers the integrated surfaces in one deterministic run. Unit seams cover the motion decisions that a text pane capture cannot observe.

SceneAutomated evidenceResult
Shell + Overlayoverlays.test.ts presence curve plus the narrow app-wiring.test.ts presence guard; fake suite help, picker, confirm, result, approvals, settings, and wizard overlayspass
Railchrome.test.ts, rail.test.ts, and the narrow app-wiring.test.ts Braille guard; fake suite rail labels, counts, and searchpass
Transcripttranscript-motion.test.ts settle/error beats; fake suite rich Markdown, finalization, completion order, scroll, End, and jump controlpass
Composer + Statuscomposer.test.ts, skill-invocation.test.ts, and the narrow app-wiring.test.ts planner/status guards; fake suite mode labels, hints, drafts, queues, and single-channel validationpass
Work + Initworkrows.test.ts, tide.test.ts, motion.test.ts, and the narrow app-wiring.test.ts wiring guards; fake suite Work cards, tide frames, recent unmount, and init phase orderpass
Settingssettings-feedback.test.ts ticket #124; fake suite four tabs, memories, backups, save-then-close, and the saved toastpass
Wizardwizard.test.ts step rail and submit toast; fake agent/workspace wizards, validation, memories, step rail, and dirty cancelpass
Overlaysmotion.test.ts, approvals.test.ts, and the narrow app-wiring.test.ts approvals guard; fake approvals resolve and overlay open/closepass
Feedbacktoast.test.ts plus the narrow app-wiring.test.ts host-fallback guard; fake suite copy toast top-right with composer focus retainedpass

The full fake suite also re-ran the existing composer, transcript-scroll, native-handoff, and overlay regressions. Concrete passing log lines included ticket #119 copy toast is top-right and leaves keys with the composer, work-list o opens the selected work's native session, ticket #80 command line, completion, history, and transcript search verified, slice E help/settings/approvals chrome verified, and settings p promoted the selected memory through the TUI approvals inbox. The F2 round-trip and native prompt checks also passed in the same run. The original ticket #127 diff was verification harness and documentation only; the post-review remediation below records the later product changes and reruns.

Work-card size matrix ​

npm run test:live:sizes starts the isolated fake host, creates three agents and one workspace, and restarts the TUI at each geometry with ENSEMBLE_LIVE_TUI_WIDTH/ENSEMBLE_LIVE_TUI_HEIGHT. At each size it delegates one fake-queue-hold task per agent, waits for WORK 3 active, and checks the header, visible cards, overflow, focus cursor, and composer.

SizeLayout observedAssertionsResult
80×24Compact fallback: one one-line row, +2 more, no card headers or card hint rowHeader WORK 3 active · 0 recent │ +2 more; compact row size-80x24-a │ SizeAgent1 │ working; cursor › on that row after Tab; Esc returns NORMAL │ Group chat; composer Message… visiblepass
100×30Card mode: one full card, +2 moreThree active tasks exceed the card-height budget, so workCardLayout renders one card; card hint row present; cursor › on the first card; composer visiblepass
120×40Card mode: three cards, no overflowThree Work card headers, no +N more, card hint row, cursor on the first card, Esc returns to group chat, composer visiblepass
160×50Card mode: three cards, no overflowSame card checks as 120×40pass

The 100×30 result is the implemented height policy, not a skipped check: three active cards cannot fit the available rows, so one card plus +2 more is shown. Two active tasks at 30 rows fit two cards (covered by workrows.test.ts and by the layout rule).

An exploratory run of the full test:live:fake scenario at 120×40 stopped at the existing finalized rich reply assertion because the shorter viewport did not show the Rich Markdown Fixture heading. That scenario is calibrated for its 160×80 default and its transcript assertions are viewport-sensitive; the targeted size matrix above is the size evidence. The toast placement assertion in the full scenario was made width-relative in this ticket because the scenario is size-parameterized.

One, two, and three working cards ​

At 160×50 the size matrix starts one long-running task at a time and measures the TUI opencode process plus tmux pane updates at each count.

Working cardsTUI CPU (0.9s sample)tmux outputResult
134.4%pane changed across capturespass
236.7%pane changed across capturespass
336.7%pane changed across capturespass

Repeat runs during this pass observed roughly 28–63% for the same counts, so these are single-sample observations, not a threshold. The measurement covers the TUI process only; the OpenCode server and fake provider are separate processes. assert_pane_updates proves the tmux output is still repainting while the tide runs.

Reduced motion and NO_COLOR ​

SettingEvidenceResult
ENSEMBLE_REDUCED_MOTION=1Full fake suite restarts the TUI with this setting, sees ⠋ FakeAgent │ working, finds none of ⠙⠹⠸⠼⠴⠦⠧⠇⠏▪▹▌ across eight captures, gets one completion line, and sees a byte-identical idle screen after 1.5spass
NO_COLOR=1Size matrix restarts at 120×40 with NO_COLOR=1, sees ⠋ SizeAgent1 │ working, keeps the working state word and composer, and finds none of the non-zero Braille frames across eight capturespass
Unit coveragemotion.test.ts, tide.test.ts, workrows.test.ts, chrome.test.ts, settings-feedback.test.ts, transcript-motion.test.ts, and toast.test.ts pin frame 0, still tide frames, static settings feedback, and reduced approval/transcript emphasispass

Missing manual checks: text capture cannot prove color suppression, contrast, opacity, or font coverage. Those remain on the macOS/Herdr checklist sections 5, 7, and 8.

Known limitations and residual risks ​

  • No macOS/Herdr evidence. Braille font coverage, Herdr key interception, native-session key ownership, toast appearance, and actual color/opacity rendering were not observed on macOS. The checklist is extended for v6 and marked unexecuted.
  • Text capture is not visual verification. The record observes glyphs, labels, counts, and layout text. It does not prove exact color, contrast, opacity, anti-aliasing, or smoothness, and no screenshots were captured.
  • The full fake scenario remains 160×80-calibrated. At 80×24 and 100×30 the targeted size script covers Work layout, focus, overflow, and the composer, not all nine scenes. The 120×40 exploratory full-scenario failure is recorded above.
  • CPU samples are noisy. They are one 0.9s sample per card count on this host, exclude the server process, and have no threshold.
  • Recent-card boundary. The eight-second constant and boundary are unit-tested; the live suite observes the recent window close, not the exact millisecond boundary. The size script waits out the window before each restart.
  • Tide variety is unit-tested more strongly than live. The fake suite observes Braille animation and Work-card tide rows; the six-variant shuffle-bag, distinct fixed-width strips, and stable per-subtask assignment are proven in tide.test.ts and workrows.test.ts, not enumerated live.
  • Toast size coverage. The live toast placement check runs at the default 160×80 geometry. The size matrix does not trigger a toast at 80×24/100×30.
  • Hosted CI was not run. The local gates above passed; no GitHub Actions run is claimed.

Post-review remediation (fix/v6-review) ​

The v6 two-axis code review found test-coupling, duplicated state, and spec divergences after #127 merged. The remediation branch re-ran the gates:

CommandResult
npm run verifyPass — strict type check, 787 tests passed, 0 failed, 0 skipped.
npm run test:live:fake:allPass — fake 207s, fake-empty 9s; 2 passed, 0 failed, 0 skipped.
npm run test:live:sizesPass — 80×24, 100×30, 120×40, and 160×50 Work fit/overflow/focus/composer checks; 1/2/3-card load probe (TUI CPU 31.1%, 45.6%, 55.6%); NO_COLOR=1 still-motion check.
npm run test:live:latencyRecorded — min 9ms, p50 9ms, p95 21ms, max 21ms, mean 10.2ms, n=10.
git diff --checkClean.

Behavior and evidence changes in the remediation:

  • app-wiring.test.ts and the wizard source guard were reduced to narrow wiring checks; whole-region slices and copy pins were removed. The scene table above now points at the pure seams that own those behaviors.
  • Settings save closes the surface again and confirms through one app toast; the fake scenario asserts the close and the top-right toast.
  • Composer rejection keeps the draft and one statusline message; the extra app toast was removed. The fake scenario asserts a single message occurrence.
  • The Work cursor and actions only walk the visible active prefix; the overflow rule is unit-tested.
  • Needs-you cards freeze on the tick where they became blocked; the freeze is unit-tested at the tide and card seams.
  • The fake scenario now also asserts distinct tide frames, the wizard step rail, TUI approvals resolution, and the saved-settings toast and placement.
  • Finalization/error color emphasis, overlay opacity, and one-shot settings row flashes remain unit-covered because tmux text capture observes text, not ANSI color or opacity. The live overlay checks still prove open/close and focus behavior, and the approvals check proves the settled row.
  • Init phase state, the eight-second retention, tide clamps, row feedback, card construction, transcript settle branches, unknown-command handling, and init phase copy were consolidated to one source each.

Remaining gaps are unchanged: the full fake scenario stays 160-wide-calibrated, and macOS/Herdr remains unexecuted. The targeted size matrix was re-run on the remediation code and passed all four geometries, the one/two/three-card load probe, and the NO_COLOR still-motion check.

Post-v6 follow-up (#129) ​

The follow-up fixed two production issues found after v6 delivery:

  • Agent tool handlers now refresh session bindings before resolving the calling session, independent of the global storage version. New workspace bindings carry kind/directory, existing bindings are repaired, and workspace config-version reconciliation now notices a missing/incorrect version.
  • The Work pane now uses a width-aware grid: three columns on wide terminals, two on medium, compact rows when the width/height cannot support cards. Cards keep a state-colored left accent, uppercase agent/state, and per-card gapped tide strips. Four tasks at 120×40 form a 2×2 grid.

Verification for this follow-up:

CommandResult
npm run verifyPass — strict type check, 789 tests passed, 0 failed.
npm run test:live:sizes (120x40)Pass — three Work card headers, fitting/overflow, Work focus, and composer checks.
git diff --checkClean.

The full size-matrix script remains environment-flaky in its final NO_COLOR phase on this host; the 120×40 Work-grid assertions passed before that phase. The shared binding regression is covered by the new store test and the service-level workspace_info test. macOS/Herdr remains unexecuted.

Tracker state ​

  • #113 and #127 remain open. This record does not close them.
  • Tickets #114–#126 were integrated on integration/ensemble-v6; this branch is based on its tip e49a09e and merges the current tip again before handoff.
  • No push was made and no issue was closed from this branch.

Released under the MIT License.