Skip to content

v7 verification — dispatch groups, durable feed, latency, scale, and durability ​

Status: verified locally on Linux/tmux; hosted CI not run; macOS/Herdr manual checklist unexecuted Date: 2026-11-07 Specification: #130 Ticket: #143 Branch: ticket/143-verify-v7 Base: integration/ensemble-v7 at e64c348 Worktree: /home/tristan/dev/ensemble-t143

This is the specification-level verification record for v7. It covers the dispatch-group lifecycle, Work-card visuals, handoff/delegation card latency, idle and working CPU, streaming, large-history bootstrap, bounded client memory, fault recovery, the size matrix, reduced motion, NO_COLOR, and the macOS/Herdr status. It does not close #130 or #143. The parent specification is a version milestone and may be closed only by the documented verification-and-closure rule: every child issue closed and this record linked from the closing comment.

Linux/tmux is not macOS/Herdr evidence. Everything below was observed on Linux with tmux. No macOS/Herdr host ran this pass, so Herdr key interception, macOS font coverage, and human-perceived color/opacity remain unverified. See the macOS/Herdr checklist, section 9 (NOT RUN).

Environment ​

sh
export PATH=$HOME/.nvm/versions/node/v26.12.1/bin:$HOME/.local/bin:$PATH
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 v2.2.6
Git2.47.3
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 (16G tmpfs, 14G free)
macOS/HerdrNot available; not run
Hosted CINot run

npm ci installed 376 locked packages in 4s. The repository type-checks against @opencode/plugin 2.2.1 while the live checks exercise the installed OpenCode 2.2.6 host; these are different version surfaces, as recorded in the runtime guide.

Commands and results ​

All live suites ran sequentially; no two live suites ran at once and no hosted CI job was triggered.

Exact commandResult
npm run verifyPass — strict type check, then 946 passed, 0 failed, 0 skipped, 0 todo (duration_ms 24287.25).
npm run test:live:fake:allPass — fake 193s, fake-empty 9s; 2 passed, 0 failed, 0 skipped, 0 not run (202s total).
npm run test:live:sizesPass — 80×24, 100×30, 120×40, 160×50 group/layout/focus/composer checks; idle and 1/2/3-card CPU; NO_COLOR=1 still-motion check.
npm run test:live:latencyPass (measurement only) — rail p50 9 ms, delegation-card p50 150 ms; raw samples below.
npm run test:live:preflightPass — Node v26.12.1, OpenCode v2.2.6, tmux 3.5a; fake scope needs no fd.
node --check scripts/fake-provider.mjsPass — no provider syntax errors.
bash -n scripts/live-check-fake.sh scripts/live-size-matrix.sh scripts/live-latency-probe.sh scripts/live-matrix.sh scripts/live-preflight.sh scripts/lib/live-harness.shPass — no shell syntax errors.
git diff --checkPass — no whitespace errors.
Targeted node --test --experimental-strip-types files belowPass — diagnostics recorded in the measurement tables.
Hosted CINot run — local gates only.

Latency samples ​

ProbeRaw samples (ms)Summary
Rail j/k repaint (n=10)9 9 15 9 8 9 15 9 9 9min 8, p50 9, p95 15, max 15, mean 10.1
Delegation card appearance (n=5)188 154 144 150 142min 142, p50 150, p95 188, max 188, mean 155.6

The rail measurement boundary is tmux send-keys → first capture showing the moved cursor. The card boundary is the task.delegate RPC invoke → first capture showing the Work card title. Both include tmux capture overhead and are machine-specific; they are observations, not thresholds.

CPU observations ​

npm run test:live:sizes records one 0.9 s TUI-process sample per count at 160×50. The TUI process only; the server and fake provider are separate.

StateTUI CPU
Idle (after 1.5 s settle)1.1%
1 working card8.9%
2 working cards8.9%
3 working cards30.0%

These are single short samples, not precise regression budgets. The pre-#142 baseline recorded in ticket #142 was 27.8%, 25.6%, and 24.4%; the 1- and 2-card samples are materially lower, while the single 3-card sample is noisy and is not treated as a regression conclusion. The harness does not gate on these numbers.

Group lifecycle and Work pane ​

The pure and service seams are covered by src/v2/dispatch-group.test.ts, src/v2/dispatch-group-service.test.ts, src/v2/completion-batch-service.test.ts, src/v2/tui/workgroups.test.ts, src/v2/tui/workrows.test.ts, and src/v2/tui/tide.test.ts. The live fake scenario and size matrix cover the user-visible lifecycle.

BehaviorEvidenceResult
One command fans out to one dispatch groupservice-handler test; fake fake-queue-alpha/beta; size matrix 3-recipient fan-outpass
Second command while the first is active appends a second groupfake v7 two-group block; size matrix 120×40 and 100×30 (WORK 2 groups, GROUP 1, GROUP 2)pass
Complete group clears as one unit after the shared 1.2 s holdunit hold/reload tests; fake block; size matrix group removalpass
Needs-you overrides the hold and persists until the decisionlive fake fake-ask-person approval (NEEDS YOU beyond 2.2 s); unit terminal-pending-approval test; review-fix hold-after-resolution testpass
Failed and cancelled cards stay until the group completesunit failed/cancelled members; live cancelled cards; live failed card from rejected ask (below)pass
Canonical terminal event, no duplicate generic failureterminalAnnouncement tests; service failure dedupe tests; live ⊘ … cancelled and ✕ … failed exactly-once checkspass
Header names groups/active/done instead of recentunit header tests; live WORK 1 group · … · 0 donepass
Newest-group priority under height pressureunit 100×30 priority; size matrix 100×30 summarized old grouppass
Reload during the hold preserves the remaining holdworkgroups.test.ts pure reload testpass
Cursor/actions only target non-terminal cardsunit terminal/overflow tests; live Work focus and action probespass

The live fake v7 block creates two overlapping commands, verifies both group labels and cards, cancels the older group, verifies the younger group survives the older group's removal, and asserts one ⊘ v7-group-one fake-queue-hold-long cancelled event. It then emits a native ensemble_ask_person tool call from the deterministic provider, verifies the real approval, the NEEDS YOU treatment, persistence beyond the hold window, resolution, and normal completion. A second ask is rejected to produce one ✕ FakeAgent failed: User rejected the request. terminal event and a FAILED card. The final fake run logged:

text
live: v7 two-group append, independent removal, and canonical cancellation events verified
live: v7 needs-you approval persisted, resolved, and cleared live
live: v7 failed transition, FAILED card, and single terminal event verified live

Work-card visuals ​

ClaimEvidenceResult
Header order is name, spacer, glyph, right-aligned statusworkCardHeader tests; live card capturespass
Name is pale blue in every state; status never truncates; name truncates firstpure header/tone tests; palette() dark/light pale-blue token tests; live text capture; color not machine-verifiedpass
WORKING: grey status, yellow Braille/glyph, yellow left barpure workCardTones; live text capture; color not machine-verifiedpass
REPLIED: green dot, green status, green barpure tones; live text capture; color not machine-verifiedpass
NEEDS YOU: yellow diamond, yellow status, yellow barlive NEEDS YOU card; pure tones; color not machine-verifiedpass
FAILED red; CANCELLED faint; QUEUED muted; RETRYING reads as workingpure tone tests; live cancelled/failed cardspass
Selection is a focus border, not a green gutter; rounded border neutralWorkCardView selection tests; live focus capturepass
3-column, 2-column, compact widths preserve the header160×50/120×40 card rows, 100×30 overflow, 80×24 compact rows in the size matrixpass
Reduced motion and NO_COLOR preserve state words/final glyphsfake reduced-motion run; size matrix NO_COLOR frame-0 checkpass

Visual review. Pure tone/header tests cover the palette-role mapping (pale-blue name, yellow WORKING/NEEDS YOU, green REPLIED, focus border), and live text captures cover header order, state words, and glyph placement. No automated ANSI color assertion was implemented. A human perceptual review of dark/light themes and macOS font rendering was not performed; the manual checklist remains the authority for that claim. Text capture alone is not proof of human-perceived color.

Durable feed, fault injection, and streaming ​

The feed is described in ADR 0002 and the implementation guide. Fault injection is covered at the pure client seam (src/v2/tui/change-feed.test.ts), the store seam (src/v2/change-feed.test.ts), the service seam (src/v2/service.test.ts), and the client-store seam (src/v2/tui/client-store.test.ts).

Fault / behaviorEvidenceResult
Dropped eventrevision heartbeat + changes.since recovery test (a changes.since response recovers a dropped event as patches or a reset); live patch pathpass
Duplicate patchduplicate batch ignored; overlapping redelivery keeps only post-cursor changes; duplicate message patch is harmlesspass
Reordered patcholder batch ignored; older reordered stream never rewinds a rowpass
Malformed patchunknown/bogus/missing-identity batches ask for a resync; Work-only malformed repairpass
Reset / ring evictionevicted cursor, cursor ahead, and foreign-version writes return reset; scoped Work resyncpass
Epoch change / service restartrestart changes epoch and resets the feed without losing work; full snapshot fallbackpass
Durability before publicationentity + meta persisted before the publish callback; checkpoint has no version/patchpass
Compatibility fallbackprotocol-2 clients keep the legacy changed event and snapshot path; protocol-3 only applies entity patchespass
Streaming finalizationephemeral stream never bumps the version; missed stream recovered by final message patch; final text replaces the live overlay exactly oncepass
Crash recoverylast checkpoint reloads, streaming clears, row marked interruptedpass

Live defect found and fixed in this pass ​

The live gap-recovery path exposed one real integration defect: an HTTP changes.since response returned 500 rpc.invalid_output at entries[0].entities[0].message because a message entity carried explicit undefined optional fields (agentID, updatedAt, and peers). The TUI caught the error and fell back to a full snapshot, so no data was lost, but the intended cheap gap path was unusable.

Fixed in src/v2/service.ts: changesSince now applies the same JSON sanitization as the snapshot/bootstrap boundaries. Regression coverage lives in src/v2/service.test.ts (changes.since output stays JSON-safe at the RPC boundary), which walks the handler response for explicit undefined values. A live probe against OpenCode 2.2.6 now returns 200 with contiguous entries and the expected message/subtask/workspace/agent entities.

Large-history bootstrap, client memory, and Tier 3 gate ​

The large-history test seeds 10,000 durable messages and 602 retained tasks (600 long-terminal rows plus one visible group), then calls the scoped bootstrap handler over durable storage:

MeasureObservedBudget
Bootstrap payload26,379 B< 128 KiB
Bootstrap latency11.7 ms< 500 ms
Transcript rows in payload40one active window
Work rows in payload2visible group only
Tier 3 gateenabled (messages >= 1000)recorded with inputs

The client-memory benchmark applies 5,000 message patches plus 1,000 invisible terminal tasks:

MeasureObservedBudget
Loaded message rows240CLIENT_MESSAGE_LIMIT (240)
Loaded Work rows1 (active)visible projection
5,000 patch applications109.1 ms< 5,000 ms

Tier 3 gate decision: enabled for the measured large workspace and left disabled for small workspaces. src/v2/scoped-read.ts is the single policy source: history ≥ 1,000 messages, retained tasks ≥ 500, last snapshot payload ≥ 512 KiB, or last snapshot build ≥ 120 ms. The response carries gate, scoped, counts, window, and scopeVersions; a disabled gate keeps the full snapshot path. The scoped path remains a disposable read model over durable storage, never a new authority.

Targeted performance measurements ​

These are the implementation-ticket benchmarks rerun on this tip:

MeasureObservedBudget / comparison
10k-message handoff action11.0 msbeats one snapshot clone (16.42 ms)
10k-message delegation clones0 full-state cloneszero
Explicit snapshot clone count1exactly one
10k-message foreign-version sync99.5 ms< 500 ms
5k-task waiting-work reconciliation314.7 ms< 2,000 ms
5,000 bounded client message patches109.1 ms< 5,000 ms

Skips and explicit limitations ​

  • macOS/Herdr: NOT RUN. No macOS/Herdr host was available. The manual checklist section 9 records the v7 group, card, recovery, reduced-motion, and color checks as NOT RUN. Linux/tmux evidence does not satisfy them.
  • Hosted CI: NOT RUN. No GitHub Actions run is claimed.
  • Human visual review: NOT PERFORMED. Pure tone tests and live text captures verify palette-role wiring and structure; human perception of dark/light themes and macOS font rendering remains manual.
  • Live true-failed coverage: the fake scenario produces a real failed transition by rejecting an ask_person approval; provider/transport exhaustion remains covered at the service seam rather than by a slow live retry fixture.
  • CPU samples are noisy: one short sample per card count on this host; they establish direction, not a precise budget.
  • Network/tool failures: no real model or external provider was used; the fake provider is deterministic test infrastructure.

Tracker state ​

  • #131–#143 are closed. The parent #130 was closed after every child and this record were delivered, per the documented rule.
  • integration/ensemble-v7 is the delivered integration branch.
  • A post-v7 retry follow-up is tracked in #144 and recorded in the retry-stall verification.
  • That follow-up raises the animation cadence to 8 fps Braille / 15 fps tide; the CPU table above reflects the earlier 4/10 fps cadence.

Released under the MIT License.