Ensemble v5.1 wizard verification — 2026-11-02
Ticket: #49, parent spec #48. Tested from feat/v51-49, based on integration commit abbdb90.
Environment
Working directory: /tmp/opencode/ensemble-worktrees/49-wizard.
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. Pinned host types 2.2.1 and tested host 2.2.5 remain distinct evidence. macOS delivery of the universal pair is by construction of the key family (plain Ctrl+letter is delivered by stock terminals); the live checks below ran on Linux, and no live macOS run is claimed.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (dev dependencies already pinned by ticket #43). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 127 passed, 0 failed, 0 skipped. |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
npm run verify | Exit 0; check plus 127 passing tests. |
Live checks are reported separately from unit checks; all three executed and asserted rather than skipping.
What each check proves
- Universal wizard navigation. The fake scenario opens the New-agent wizard and asserts the footer reads
Ctrl+N nextandCtrl+P back, then sendsCtrl+N(pane advances toAgent title) andCtrl+P(returns toAgent name). The wizard layer bindsctrl+n/ctrl+pdirectly and is deliberately not wrapped in the platform Ctrl→Cmd transform, because stock macOS terminals never deliver Cmd chords; unit tests pin the constants and prove the generic transform would rewrite them, documenting why the bypass must stay. The searchable help applies the same exemption so those two entries stay truthful on macOS while other app shortcuts still show Cmd. - Live model options and explained empty state. Unit tests cover option derivation from a changing catalog, option equality, choice preservation/fallback, and the empty-state copy. The fake scenario (one model configured) walks to
Main modeland continues. The new empty-catalog scenario (models: {}) asserts0 optionsplusNo models available …on the model step and that Escape still cancels the wizard; the catalog's failure reason reaches the snapshot through a new optionalmodelsErrorfield. Service tests prove the field is populated for both an empty and a failing catalog and that credential-shaped text (keyword pairs, Bearer tokens,sk-keys) in the host error is redacted. The empty explanation is gated on the catalog's option count, not the filter result, so filtering a healthy catalog never claims a catalog failure. - Snapshot output contract. Live runs initially failed with
rpc.invalid_output: Expected string at ["configError"]: the host RPC validator rejects optional fields present withundefined. The snapshot payload now omits absent optional fields instead of serializing them as undefined; that fixed the live RPC and keeps the schema honest. - Centered overlay. Unit tests cover the shared geometry helper: bounded width (48–96), horizontal and vertical centering, and recomputation for a changed viewport. The wizard overlay reads renderer dimensions from a resize-fed signal, so it recenters after a renderer resize; content stays left-aligned and the hint row is horizontally centered inside the box. The live checks assert text and behavior, not pixel positions.
- Help consistency. The searchable help lists
Ctrl+P/Ctrl+Nfor wizard steps; the help test asserts the new tokens.
Cleanup and data safety
Both fake scenarios isolate HOME and XDG directories, use the loopback fake provider with no API key, and clean their temporary directories and tmux sessions on success. No real workspace data was read or modified by the live checks. npm run test:live uses the harness's isolated project/HOME as before.
Remaining uncertainty
- No live macOS run: the Ctrl+letter family is the only one stock macOS terminals reliably deliver, but terminal-specific overrides (kitty protocol, custom key bindings) are outside this verification.
- Centering is verified through the geometry helper plus behavior, not pixel-level pane captures; extremely small viewports are clamped to the 48-column floor and may still hide some content by design.
- The empty-catalog scenario proves the product's empty state; the reporter's original empty catalog on their machine may also have had a provider configuration cause, which this work explains but does not repair.
Ensemble v5.2 conversation fidelity (ticket #51) — 2026-11-02
Ticket: #51, parent spec #50. Worked in /tmp/opencode/ensemble-worktrees/52-51-fidelity on branch feat/v52-51, based on integration/ensemble-v5.2 at 8df78d1 (merge was a no-op: the tip was already contained, Already up to date).
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. Host-versus-runtime distinction: pinned peer types are @opencode/plugin 2.2.1, while the executed host binary is 2.2.5; behavior below is tested evidence on 2.2.5, not a claim about the declaration surface.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; dependencies already pinned). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 148 passed, 0 failed, 0 skipped. |
npm run verify | Exit 0; check plus the same 148 passing tests (re-run after the integration merge). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
Live checks are reported separately from unit checks; all three executed and asserted rather than skipping.
What each check proves
- Durable DM transcript (unit). New
src/v2/dm.tsseam:mapNativeChatkeeps only non-empty user/assistant text;mergeDmTranscripttreats stored per-binding lines as authoritative forrole: "user"(native user content — echoes, onboarding steer prompts — never renders), keeps native assistant replies, preserves queued flags, and sorts stably by(createdAt, user-first, id). Fivedm.test.tscases pin the merge, including the phantom-line and exactly-once cases.chat.sendnow appends the user line to thev2/dm/<workspaceKey>/<agentID>/store synchronously at send time (monotonic per-service clock so same-millisecond sends keep order), reusing the same id for thedmQueueentry;flushDmQueueclears the queued flag on send/fail;moveHomemigrates the transcript; agent deletion prunes it;pruneWorkspace/resetremove it. Service tests prove: the line renders immediately; it survives a restart through the store load path with a native echo present (exactly["Persist me", "OK"]); a queued-while-busy line stays visible withqueued: truethen clears; store tests prove reload ordering and pruning. - Leader routing (unit + handler).
routing.test.tspins total routing: unmentioned → leader only, single mention replaces default, fan-out → exactly the mentioned, archived members excluded, unknown names reported viaunknownMentions. Service tests prove an unmentionedmessage.sendcreates exactly one subtask assigned to the leader with the existing system line→ Nova: <title>intact, and that routing ignores recent activity (after$Sagework, an unmentioned message still routes to Nova).lastActiveAtwas audited: written for rail ordering/unread counts, never read byvisibleRecipients/sendMessage/createSubtask. - Routing observability without redesign. Group-chat surface title now reads
#name · dir · leader <Name>;ensemble_workspace_info/ensemble_list_workspacesreturn the leader and members as{id, name}pairs (plussynthesison workspace_info);/init,Shift+I, and the shortcut help now label workspace initialisation as initialisation, not routing (Init workspace files (initialisation, not routing)). No init semantics changed. - Wizard exclusive focus + Enter (unit + live).
openWizardnow closes the completion popup, blurs and deactivates the composer, and every competing key layer (global, composer-cancel, completion, settings, approvals, help) additionally requireswizard() === null. Enter on non-text steps routes through the new purewizardEnterAction(kind, final)helper (select/continue/submit/text, unit-tested inwizard.test.ts); the handler closes completion, owns Enter on choice/pick/multi including the final step, and text steps still return early to the field (Enter= continue,Shift+Enter= newline). The fake live scenario walks the full workspace wizard (n, directory Enter, name type + Enter, members Enter = continue, leader Enter, synthesisYesEnter) and asserts#LiveSiteappears, the surface readsleader FakeAgent, the snapshot persistssynthesis: true,$ agents(completion title) never appears, and the DM user linefake-rich-fixturerenders exactly once. The fake provider already answers every DM prompt (short reply / rich fixture), so the DM send completes in-scenario. - Untouched surfaces.
test:live:fake:emptystill passes (catalog/model-step behavior unchanged); the real-modeltest:livepasses with no new assertions affected.
Cleanup and data safety
All live scenarios isolate HOME/XDG, use the loopback fake provider or the harness project (no API key material in this repo), and clean tmux sessions and temp dirs on success. Unit tests use temp dirs and the in-memory fake context. No real workspace data was read or modified. Issues #50/#51/#52 stay open; no issue was closed.
Remaining uncertainty
- Native session user-echo shape/timing on the real host is now irrelevant by construction (native
role: "user"is never rendered for DMs), but the monotoniccreatedAttiebreak only orders lines created throughchat.send; externally seeded DM rows sharing a timestamp still fall back to id order. - The live
Escapedismissal of the host move-home confirm worked on host 2.2.5; host dialog keybindings are host behavior and could differ on other versions. - No live macOS run; the wizard key family is unchanged (plain Ctrl+letter), so the v5.1 macOS reasoning still applies unmodified.
Ensemble v5.2 in-house surfaces and motion (ticket #52) — 2026-11-02
Ticket: #52, parent spec #50. Worked in /tmp/opencode/ensemble-worktrees/52-52-surfaces on branch feat/v52-52, based on integration/ensemble-v5.2 at c3a2d6c (merge of the tip was a no-op: already up to date). Blocked-by ticket #51 is merged at that tip; no #51 behavior (DM transcript, routing, wizard exclusivity) was altered except where a migrated overlay integrates with it.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. Host-versus-runtime distinction: pinned peer types are @opencode/plugin 2.2.1, while the executed host binary is 2.2.5; behavior below is tested evidence on 2.2.5, not a claim about the declaration surface.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 160 passed, 0 failed, 0 skipped (148 pre-existing + 12 new). |
npm run verify | Exit 0; check plus the same 160 passing tests (re-run after the integration merge). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Live checks are reported separately from unit checks; all three executed and asserted rather than skipping. A grep -rn "context.ui.dialog." src/ over the repo returns zero matches (every file count 0); 13 context.ui.toast call sites remain, which stay native per the spec.
What each check proves
- Dialog migration (unit + live). New
src/v2/tui/overlays.tsseam pins the confirm-modal keys (Enter/y confirm, Escape/n cancel, everything else trapped), the result-modal keys (dismiss/scroll), single-choice resolution, and inline skill-filter editing. Newsrc/v2/tui/motion.tsseam pins the streaming frame cycle, caret blink, leader pulse, motion-active gating, and the one-shot empty mark. All 22 inventory call sites now land on Ensemble surfaces: model/variant/delegate-target/reassign-target/edit-field/approval choices on the generic single-select picker overlay (newchoicemode reusing the existing picker search/filter); memory text, role values, delegate titles/instructions, and ask_person answers on the existing multiline text prompt (Enter submit, Shift+Enter newline); move-home, reset handoffs, discard-settings, archive/delete workspace, archive/delete agent, and both paired native-session questions on the new in-house confirm modal (Enter yes, Esc no); long task results on the new scrollable in-house result modal (arrows/pgup/pgdn scroll, Enter/Esc/q dismiss); workspace search on the workspace picker; skill filtering inline (typing appends, Backspace deletes, Enter/Esc done, gated key layers). RESET still requires the exact text. Escape cancels, Enter confirms/selects everywhere, and every opener refuses while a wizard is open so overlays never steal Enter/focus from the wizard-exclusivity gate (#51). - Motion (unit + live). The frame clock runs at 4 fps only while busy, streaming, or composing and resets to the still frame otherwise; the interval is owned by the effect with
onCleanup, so nothing runs after finalize/unmount/route-leave. The streaming header keeps the literalstreaming…anchor plus a rotating frame (live-pinned as one ofstreaming… ◐/◑/◒/◓while working, absent after finalize); the composer caret blinks only while focused; the leader marker in the rail pulses while active and rests on○when idle; new-message reveal is color-only (no text change, no layout shift) settling after 1.5 s; empty states show the static○───○───○mark once. No Timeline dependency was added. The fake scenario asserts the animated frame appears with the anchor and thatstreaming…disappears at finalize; count/absence assertions (rich heading, code marker, quote, table cell ×1) stayed green because no animated characters enter asserted regions. - Untouched surfaces. The wizard, DM transcript, routing,
/init, and Shift+I behavior are unchanged;test:live(real host) andtest:live:fake:emptypass with only additive assertions (the empty mark).
Live failures fixed during the run
- The confirm modal first rendered without an explicit height, so its title row collapsed under the message/footer lines (caught by the preserved-pane inspection: only the footer and a message fragment were visible). Fixed with an explicit height on the centered box, matching the other overlays.
- The pre-existing picker overlay renders word gaps as non-breaking spaces, so literal
"Edit field"/"enter confirm · esc cancel"assertions could never match (verified via pane hexdump:Edit c2 a0 field). The component was left untouched (v5.1 behavior) and the script matches the non-breaking-space forms.
Cleanup and data safety
All live scenarios isolate HOME/XDG, use the loopback fake provider or the harness project (no API key material in this repo), and clean tmux sessions and temp dirs on success; two preserved-failure runs were inspected then removed (tmux kill-session plus rm -rf of the fixture dirs). No real workspace data was read or modified. Issues #50/#51/#52 stay open; no issue was closed.
Remaining uncertainty
- The scrollable result modal and the ask_person/delegate/reassign/router pickers are covered by unit-tested key/decision helpers plus code review, but no live run opens them (the fake scenario has no tasks, approvals, or ask_person prompts); the confirm modal, choice picker, and multiline prompt are the flows exercised live end to end.
- No live macOS run; key handling uses the same plain key names as the #51 paths, so the v5.1 macOS reasoning still applies unmodified.
- Motion frame glyphs (
◐◑◒◓▌●○) are plain text renderables verified through tmux pane capture on this host; other terminals may substitute glyphs but the literal anchors (streaming…,COMPOSER,leader) are glyph-independent.
Ensemble v5.2 review fixes — 2026-11-02
Worked in /tmp/opencode/ensemble-worktrees/52-review-fixes on branch feat/v52-review-fixes, based on integration/ensemble-v5.2 at 6ddb9cc. No product changes; surgical fixes only for the three findings from the two-axis review of the #51/#52 merge.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1. Linux. npm ci run once in the worktree (no node_modules was present).
What was fixed
- Durable DM ordering clock.
EnsembleV2Servicenow re-seeds its in-memorylastDmAtfromV2Store.dmMaxCreatedAt()after every store load/sync (start, snapshot, diagnostic, invoke) and after move/reset, so a post-restartdmNow()(max(now(), lastDmAt + 1)) always exceeds storedcreatedAtvalues. New regression test proves a DM sent after restart orders after a future-dated stored line with strictly increasing timestamps; the test fails on the pre-fix code. - Shared
v2/dm/ingest. The duplicated scan/dedupe/sort blocks inV2Store.loadandV2Store.syncnow both call onemergeDmScan()helper (dedupe by id per binding key, deterministic(createdAt, id)sort). Behavior unchanged; a new store test exercises the sync path against lines written by another store instance. - Atomic, idempotent
moveDmTranscript. The move unions storage-prefix rows with in-memory lines, writes all new keys first, then deletes the old prefix; same-key moves are a no-op and retrying after an interruption converges to one copy with no loss (worst case is a transient cross-binding duplicate). New store tests cover the move, the empty-transcript no-op, and retry after a simulated between-phases interruption.
No reviewer minor notes required code changes beyond the above; the recorded live-coverage gap for the result modal / ask_person / delegate pickers is unchanged by design, and the shared verification file was not renamed.
Command results
| Exact command | Result |
|---|---|
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 165 passed, 0 failed, 0 skipped (160 pre-existing + 5 new). |
npm run test:live:fake | Exit 0; fake live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; other live scenarios were not run (the fake run passed, so no comparison run was needed).
Remaining uncertainty
- The new clock test pins ordering through the service seam with a future-dated seed; real-host clock skew beyond the tested margin follows the same
max()logic by construction but was not live-exercised. - The move-interruption test simulates the between-phases state in storage; an actual process kill mid-move was not exercised.
Ensemble v5.3 ticket #54 verification — 2026-11-02
Ticket: #54, parent spec #53 (reference only; not closed or modified). Worked in /tmp/opencode/ensemble-worktrees/53-54-queue on branch feat/v53-54, based on integration/ensemble-v5.3 at ef37e2a (tip; no merge needed, verified via merge-base).
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0. |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 186 passed, 0 failed, 0 skipped (165 pre-existing + 21 new). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including the new group work ran to replied with derived status (["replied"]) line; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
What each check proves
- Derived lifecycle (unit). New
src/v2/queue.tsseam pinsderiveWorkStatusover persisted fields: first-attempt pending, valid-lease working, expired-lease retrying, exhausted-attempts failed with reason preserved, deadline-exceeded failed, finished success replied, stopped rows cancelled (explicitstoppedAtplus legacy error-text fallback), and terminal/active partitioning. 11 tests. - Store (unit).
sync()now ingests subtasks, messages (seq-sorted, nextSeq never moves backwards), approvals, memories, and the DM queue, so a second location/service view sees the same queue; load recovery clears claims/leases without duplicating rows and backfills idempotency keys in memory. 2 new tests. - Immediate start (service red→green).
agent-created delegation starts work with no further RPCholds the parent run open on a gate, invokes thedelegatetool, then asserts the child reachesrepliedfrom snapshot polls alone — no invoke/RPC in between. Failed before the fix (child stuck pending until an unrelated invoke), passes after (schedule()after every work-changing mutation plus repump-safepump()). - Visible refusal (service red→green). Non-leader
delegatenow posts a system refusal to the group chat in addition to the tool-result string, creates no subtask, and changes no queue state. Failed before (no chat line), passes after. Depth/cap/paused/unavailable refusals post visible lines the same way. - Idempotency, lease, exhaustion, handoff, status facts (service red→green). Same-delegation resubmission returns the existing item (
Already delegated…, one row); a stale-claimed row re-queues and runs exactly once (attempts 2→3, one row); attempts-exhausted rows retire as failed with the preserved reason plus a visibleSubtask failedline; handoff registration persists a durable terminal work item (workKind: "handoff", idempotency key) and dedupes re-registration;task_statusexposes derivedstatus/attempts/queueKey/leaseUntil/deadlineAt. - Routing (unit). A new pure test pins that recipient resolution uses explicit
$/!naming only — bare prose names never assign, repetition never weights — alongside the pre-existing service-level activity-ignores tests, which all still pass unchanged. - Live. The fake scenario sends one group message on the
#LiveSitesurface and polls the snapshot RPC until the leader subtask readsreplied, asserting the derivedstatusandqueueKeyare exposed — proving immediate start and status visibility through the real service plus fake provider with no extra TUI action. The real-host and empty-catalog runs pass unmodified otherwise.
Schema / state-derivation decisions
- Extended the subtask schema in place (
queueKey,workKind,claimedBy,leaseUntil,deadlineAt,retryAfter,finishedAt,stoppedAt) instead of a parallel work-item table: one source of truth, no join or drift, sync/load/recovery reuse the existing store machinery, and v5.2 board/TUI readers keep working (storedstateis a write-through cache;deriveWorkStatusis canonical and overlaid asstatuson snapshot subtasks for ticket #56). - Dedupe scope: idempotency suppresses only in-flight duplication (same key, non-terminal row). Finished rows do not block new work — re-asking completed work is legitimate new work. Handoff registrations dedupe including terminal rows (registration is the effect; re-registering posts nothing twice). Delegation keys scope to the delegating turn (
parentIDin the key); handoff keys exclude it so registration stays idempotent across turns while still recording the caller run as correlation. - Adaptations from OpenBot mechanics (documented in
queue.ts): 5 attempts, 60s lease, ~20s heartbeat, 5-minute deadline, 24h finished retention — taken unchanged; delegation uses the same 5-minute deadline. No cross-replica sweep or 2s poll loop (single-process service): stale leases are reaped on every schedule pass and on snapshot/diagnostic reads (the TUI polls every 250ms), which bounds reap latency without spawning runs from a read path. Transient run exceptions retry with 5s/60s backoff inside budget and deadline; model-availability failures keep the existing user-mediated interrupted-approval path. Handoff hops and agent-to-agent turns reuse the same record/key/derivation foundation; ticket #55 extends handoff caps on top of it. - GLOSSARY untouched: the visible-state vocabulary (
pending/working/repliedplus derivedretrying/failed/cancelled) is surfaced by ticket #56's work-row UI, which owns the language update if needed.
Remaining uncertainty
- The heartbeat renewal path (20s interval) is covered by construction and read-path assertions on
claimedBy/leaseUntil, not by a live 20s+ lease test; real-host clock skew beyond the tested margins follows the sameDate.now()comparisons but was not live-exercised. - No live macOS run; the live proof ran on Linux via tmux pane capture.
- Cross-instance subtask/message convergence is unit-tested at the store seam (writer/reader over shared storage); a live two-service-instance run was not exercised — single-process deployment (SharedEnsemble) remains the supported layout.
Ensemble v5.3 inline work-progress UI (ticket #56) — 2026-11-02
Ticket: #56, parent spec #53 (reference only; not closed or modified). Worked in /tmp/opencode/ensemble-worktrees/53-56-workui on branch feat/v53-56, based on integration/ensemble-v5.3 at f1d0b34 (tickets #51/#52/#54 merged). Blocked-by ticket #54 merged at that tip; no #54 coordination logic changed.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. Host-versus-runtime distinction: pinned peer types are @opencode/plugin 2.2.1, while the executed host binary is 2.2.5; behavior below is tested evidence on 2.2.5, not a claim about the declaration surface.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; no node_modules was present). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 192 passed, 0 failed, 0 skipped (186 pre-existing + 7 new work-row/split tests − 1 removed leader-pulse test). |
npm run verify | Exit 0; check plus the same 192 passing tests. |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including group work ran to replied with derived status (["replied","replied"]); no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
What each check proves
- Board retired, width reclaimed (unit + live).
TaskBoard/TaskGroupsdeleted fromcomponents.tsx; the main layout is rail + chat only (messageWidth = viewWidth − railWidth − 10). Newworkrows.tsseam pins glyphs (○/◐/●/✕/⊘), status labels, row text ({glyph} {title} · {agent} · {status}), active-first ordering with capped terminal rows, and theWORK · …counts line. Live assertsSUBTASKSabsent on the empty surface (fake + empty scenarios). - Inline work rows with assignee + lifecycle (unit + live). The group chat surface renders a
WORKsection oforderedWorkItemsrows with per-row derived status viaderiveWorkStatus(never stored state alone); working/retrying rows append an animatedstreamingFrame. The fake scenario requires· FakeAgent · workingthen· FakeAgent · replied, proving the loading→replied transition with the assigned agent on the row. - Active-work surface with concurrent items (unit + live). The single
liveLineis replaced byliveMessagesfrom the puresplitLiveFinalizedpartition (unit-tested: two concurrent live rows stay distinct and ordered; finalize moves each id to exactly one partition — no duplicates, no stale lines). The TUI renders one… · streaming… {frame}box per live message with per-item text. Live sends two group messages (fake-queue-alpha,fake-queue-beta), waits for 2/2 snapshotreplied, asserts both rows present with exactly 2· FakeAgent · repliedoccurrences andstreaming…absent afterwards. - Immediate visibility (live). Both probes start and finish with the TUI sitting on the group surface: no navigation, no native session opened. The 250 ms snapshot poll +
rpc.ensemble.v2.*event invalidation path is unchanged; the inline section reads the same snapshot, so new states appear where the user is watching. No gating required navigation. - Motion correction (unit + live). Removed the rail leader pulse (
leaderGlyph/LEADER_FRAMESdeleted; rail shows a static· leader) and the one-shot new-message reveal (recentIDtint, timer, and effect deleted). Kept functional motion only: working/streaming frames on active work rows and the live header, composer caret while focused, status transitions. The frame clock stays idle-quiescent (4 fps only while busy/live/composing, still frame otherwise, interval owned by the effect withonCleanup). Live still asserts the stablestreaming… ◐/◑/◒/◓frames while working and their absence after finalize; no Timeline dependency added; no pixel assertions. - Compatibility preserved (unit + live). Wizard focus exclusivity (every competing key layer still requires
wizard() === null; Enter ownership untouched), in-house overlays (picker/prompt/confirm/result all exercised live), centered geometry, DM transcript durability (fake-rich-fixtureDM line exactly once),$-mention routing, and queue semantics (derived status, idempotency keys) all pass unmodified. The fake script keeps every v5.1/v5.2 assertion (rich Markdown, single markers, DM lines, wizard Enter/Ctrl+P/Ctrl+N, dialogs/help, stale checks after leaving — extended with work-row/section absences). - Language rule. No
@Nameintroduced anywhere (grepclean); agent mentions stay$Name,@remains file-only. TheAssigned to $AgentName/Asking $AgentName…/Answer from $Agentcopy is untouched per the ticket — plain attribution only, left for v5.4 (#59).
Remaining uncertainty
- Screenshot assumption (recorded in
motion.ts). The screenshots identifying the decorative animations were unavailable during research, so the removal set is interpreted as the rail leader pulse and the newest-message reveal highlight. Confirm the target set before changing any other motion; functional motion is kept regardless. - GLOSSARY untouched. The visible-state vocabulary (
pending/working/repliedplus derivedretrying/failed/cancelled) is surfaced by the work-row UI; the glossary line still names the three primary states. Left for the spec owner to update if needed. - Concurrency bound of the live proof. The fake run proves two sequential live items finalize without duplicates plus unit coverage of the concurrent partition; two truly simultaneous live rows (two agents running at once) were not staged live — per-agent serialization plus one fake agent makes that racy to capture deterministically.
- No live macOS run; frame/caret glyphs are verified through tmux pane capture on this host only.
Ensemble v5.3 ticket #55 verification — 2026-11-02
Ticket: #55 (handoff protocol and agent-to-agent delivery), parent spec #53 (reference only; not closed or modified). Worked in /tmp/opencode/ensemble-worktrees/53-55-handoff on branch feat/v53-55, based on integration/ensemble-v5.3 at f1d0b34 (ticket #54 merged), then merged the integration tip 11a9d43 (ticket #56 merged) before reporting — all results below are post-merge re-runs unless noted. Blocked-by ticket #54 completed first; no #54 behavior was redesigned.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 211 passed, 0 failed, 0 skipped (186 pre-existing at the #54 tip + 11 new handoff pure tests + 8 new handoff service tests + 6 net new from the merged #56 tip). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including group work ran to replied with derived status (["replied","replied"]); no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping. Pre-merge runs (check exit 0; 205 passed; all three live suites passing) were re-executed after merging 11a9d43 with the results above; the merge itself was conflict-free except a mechanical auto-merge in src/v2/types.ts (see below).
What each check proves
- Typed handoff records (unit). New
src/v2/handoff.tsseam pins structural frontmatter validation (never a substring check:---markers plus parsedensemble: handoffand a non-emptytask), refusal reasons for missing markers/wrong value/missing task, recipient resolution (id exact-match, then one case-insensitive name; archived/unknown/ambiguous refused, bare prose never resolves), an idempotency key scoped to workspace+recipient+typed instruction (asking-turn independent),$Name-only copy with correlation, and the direct-message envelope (addressing, correlation, per-conversation ordering bycreatedAt, queued/sent/failed delivery state). 11 tests. - Relay exactly once (service). A handoff from Nova to Sage creates one pending
workKind: "handoff"hop with correlation/sender/recipient/depth, postsHanded to $SagewithtaskID+correlationIDlinkage, runs through the existing pump/lease/heartbeat path, and lands exactly oneRelayed answer.reply attributed to Sage carrying the hop correlation;answeredByIDis recorded. Resubmission after delivery still dedupes (one row,Already handed to $Sage).task_statusexposesworkKind, the typedhandoffrecord, and the derivedstatus. - Failure notices where asked (service). Attempt-exhausted and deadline-overdue seeded hops retire through
reapAndRetirewith the reason preserved and aHandoff to $Sage failed: <reason> (ref <corr>)system notice in the asking workspace; a hop whose binding cannot start (brokensession.create) fails throughfailTaskwith the same notice shape. No@Nameappears in any group-chat text asserted. - Refusals visible, caps enforced (service). Malformed files, unknown agents, and missing recipients are refused with a visible
Handoff refused: …system line plus the reason string (no subtask created). Depth (maxDepth) and per-run (maxDelegations) caps refuse the same way. A cancelled hop readscancelledfromtask_statuswith correlation intact (stop semantics inherit the queue's terminal derivation). - Agent-to-agent envelope + approval (service). Member DMs deliver with
(ref handoff_…)correlation; DMs to outside-roster agents create anoutside_rosterapproval with a visibleApproval needed: contact $Echoline instead of sending, and approval resolution delivers the message once with the approved correlation. Cross-workspace group posts stay refused; unknown/archived DM targets stayAgent not found.Approved cross-boundary handoffs create a typed hop through the same queue with the approved correlation. - Live (unchanged scenarios, still green). The #56-extended fake scenario (two sequential group probes to
replied, inlineWORKrows, DM durability, wizard/overlay/help assertions) passes unmodified post-merge, proving the handoff/DM changes did not regress scheduling, snapshot observability, routing, wizard focus, or dialogs. Handoff initiation is an agent tool call, not a TUI keystroke, so the TUI scripts cannot trigger it: relay/failure legs are covered at the service seam driving the realpump()/runTask()/lease machinery that the live queue probe exercises. No live scenario was added or weakened.
Schema / state-derivation decisions
- Extended the subtask row in place (
handoff: {correlationID, senderID, recipientID, depth, parentID?, path?, answeredByID?}) plusMessage.correlationID,DmQueueEntry.{fromID,correlationID}, andChatMessage.correlationID— additive optional fields only; store sync/load/recovery and older readers are untouched. - Dedupe scope: handoff hops dedupe including terminal rows (registration/hop is the effect; re-asking posts nothing twice). Delegation dedupe (live rows only) is unchanged from #54.
- One leg, not two: the recipient's run posts its reply in the asking workspace transcript, which is the relay (attributed via message
agentID, linked bycorrelationID); exhaustion posts the failure notice there. No scratch session, no second work item, no join — one source of truth on the existing queue/lease/retry/deadline machinery. - Approval reuse: outside-roster handoffs and cross-boundary DMs resolve through the existing
outside_rosterapproval kind with the correlation carried in the payload; member paths send/run immediately. No cross-workspace leaks (group cross-posts still refused; approved outside DMs deliver to the home binding). - Docs:
docs/implementation.mdtool paragraph updated to the new protocol; GLOSSARY left untouched (language update owned by the work-row UI ticket per #54 notes).
Remaining uncertainty
- The heartbeat renewal path (20s interval) is covered by construction plus read-path assertions, not by a live 20s+ lease test; same standing caveat as #54.
- Test synchronization note: the store binding lands before the run registers its session, so a tool call made between the two reads no asking run (depth/caps see depth 0). Tests wait for the run in
snapshot().state.runsbefore exercising depth/cap paths. This is pre-existing queue behavior (delegation shares it), not introduced by this ticket; flagged here so a future ticket can decide whether schedule-time or tool-time should close the window. - No live macOS run; the live proof ran on Linux via tmux pane capture.
- Cross-instance convergence remains unit-tested at the store seam only; single-process deployment (SharedEnsemble) stays the supported layout.
Ensemble v5.3 group-chat peer turns (ticket #57) — 2026-11-02
Ticket: #57, parent spec #53 (reference only; not closed or modified). Worked in /tmp/opencode/ensemble-worktrees/53-57-peerturns on branch feat/v53-57, based on integration/ensemble-v5.3 at 0ccb070 (tickets #54 queue, #55 handoff, #56 work UI merged). Blocked-by tickets #54/#55 completed first; no #54 queue or #55 handoff semantics were redesigned.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 230 passed, 0 failed, 0 skipped (211 pre-existing + 6 new peer-turn pure tests + 13 new peer-turn service tests). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including `peer turn ran to completed with asking/answered attribution ("user |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
What each check proves
- Group peer-turn states (unit). New
src/v2/peer-turn.tsseam pins the group-visible projection of the derived lifecycle: pending→queued, working→running, retrying→waiting, replied→completed, failed/cancelled→failed; plus theworkKind === "turn"guard,$Name-only refusal copy, archived-mention resolution (never silent), and outside-roster mention resolution (non-member, non-self, non-archived). 6 tests. - User mentions on the durable queue (service).
$Sage please summarizecreates exactly oneworkKind: "turn"item (no leader fallback), runs torepliedthrough the existing pump/lease path, and recordsturn: {correlationID, askerKind: "user", answeredByID: <sage>}; the asking user line, the→ Sage:task line (carrying the correlationID), and the single answering reply from Sage are linked in transcript order with no second relay leg and no@Nameanywhere. Unmentioned messages still route only to the leader asdelegationkind. A$Sage $Ghostmessage creates the Sage turn and postsPeer turn refused: $Ghostwithout aborting it; a purely unknown or archived mention posts the visible refusal and failsunknown_memberwith no subtask and no leader fallback. Two parallel$Sagementions serialize per agent (secondstartedAt >=firstrepliedAt, one attempt each). - Agent-to-agent asks from the group chat (service). An agent
send_messageworkspace post naming$Sagecreates a peer turn withaskerKind: "agent"/askerID, answered once by the recipient with correlation linkage and noRelayed answerdouble-render. Depth (maxDepth) and per-run (maxDelegations) caps refuse visibly with zero turns queued; self-mentions never create self turns; outside-roster$Echofollows the existingoutside_rosterapproval (payload.peerTurn,Approval needed: contact $Echo) and the approved turn runs once with the approved correlation. 13 service tests. - Recovery and cancellation (service). A seeded interrupted
turnrow resumes after restart through the interrupted-approval path with attempts preserved (2→3), its idempotency key intact, exactly one row, and the single correlated answer attributed. Cancelling a turn records the terminal state with the correlation intact (task_statusreadscancelledfor the handoff-analogous path; peer-turn derivation maps cancelled to groupfailed). - Live (extended fake scenario). After the existing two delegation probes (still
· FakeAgent · replied×2, unchanged), the TUI accepts the$agent completion, sends$FakeAgent peer-check, and the turn runs to completed with no navigation or native session:· FakeAgent · replied×3,peer-checkvisible, noPeer turn refused, no@FakeAgent,streaming…settled. The snapshot asserts oneturnrow,askerKind: "user",answeredByID === assigneeID, and exactly one answering message carrying the turn correlation ("user|true|1"). The real-host and empty-catalog runs pass unmodified otherwise. - Incidental streaming-race fix. The new timing exposed a latent race: a
session.text.deltaarriving between run registration and live-row registration was dropped, which deterministically failed the pre-existing streaming unit test. The live slot now registers alongside the run with the opening row filling it in (buffered deltas flush on fill); behavior is otherwise unchanged and the pre-existing test passes unmodified.
Schema / derivation decisions
- Extended the subtask row in place (
turn: {correlationID, askerKind, askerID?, answeredByID?}) plus correlation on the→task line and recipient reply messages — additive optional fields only; store sync/load/recovery and older readers untouched. Group states are a pure projection (peerTurnStatus) ofderiveWorkStatus, never a stored status. - One leg, not two: the recipient's reply in the asking transcript IS the relay (attributed via message
agentID, linked bycorrelationID); exhaustion/cancellation posts the peer failure notice there. Same one-source-of-truth shape as #55 hops, minus the handoff file. - Routing preserved: user text with no mention syntax still delegates to the leader only; mentions replace the default; recent activity never decides (existing activity-ignores tests pass unchanged);
/initand Shift+I untouched (initialisation only). - Dedupe scope: peer turns dedupe live rows only (same key while live returns the existing turn); finished turns never block re-asks. Outside-roster approvals dedupe including terminal rows (registration is the effect).
- Copy: concise
Peer turn refused: $Name — <reason>lines and the existing→ Name:task line carry the mechanics; the exactAssigned to $AgentName/Asking $AgentName…/Answer from $AgentNamewording stays owned by v5.4 (#59). No@Nameintroduced in copy, help, tests, or live scenario. - Docs:
docs/implementation.mdtool paragraph and mention bullets updated to the peer-turn protocol; GLOSSARY left untouched.
Remaining uncertainty
- The heartbeat renewal path (20s interval) is covered by construction plus read-path assertions, not by a live 20s+ lease test; same standing caveat as #54/#55.
- A two-agent live exchange (one agent visibly answering another agent's ask in the TUI) was not staged — the live proof covers a user mention to the single fake agent; agent-to-agent asks are covered at the service seam driving the real pump/lease machinery. Two truly simultaneous live rows remain covered by unit partition tests only (same standing note as #56).
- No live macOS run; the live proof ran on Linux via tmux pane capture.
- Cross-instance convergence remains unit-tested at the store seam only; single-process deployment (SharedEnsemble) stays the supported layout.
Ensemble v5.3 review fixes — 2026-11-02
Worked in /tmp/opencode/ensemble-worktrees/53-review-fixes on branch feat/v53-review-fixes, based on integration/ensemble-v5.3 at 5a25aaf (all four v5.3 tickets merged). Focused pass over the ten review findings; no new product scope. Mention syntax stays $Name only (@ file-only) throughout.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree.
Command results
| Exact command | Result |
|---|---|
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 246 passed, 0 failed, 0 skipped (230 pre-existing + 16 new). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including the unchanged work-row and peer-turn assertions; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
Disposition per finding
- Idempotency keys and retention — fixed.
workKey/handoffWorkKeynow scope actor, workspace, kind, a hashed run prefix (never the raw parent id), the addressed recipient, and the canonical trimmed task/constraints/expected/instruction body (queue.ts,handoff.ts; one sharedstableHashhelper). Relay/failure-notice correlations derive from the work key outside the fan-out prefix (correlationIDForKey); relay/notice legs remain messages, not work rows, so control legs never consume the run budget. Suppression honors the row lifetime for all work kinds including terminal rows (24h retention expiry ends it);task.retry/task.restartre-queue the same row and a different parent/run hashes a different key, giving two explicit re-ask paths. Tests: key-shape/canonicalization/derivation units, terminal-row suppression with retry re-run (service seam), rewritten handoff key test; the old "finished rows don't block new work" wording is replaced in code comments anddocs/implementation.md. - Handoff caps — fixed. New
HANDOFF_MAX_DEPTH = 1/HANDOFF_MAX_PER_RUN = 3enforced intoolHandoffwith distinct refusal copy; only handoff hops count toward the per-run budget. Peer turns and delegations keep the generic settings. Test rewritten to the new semantics (3 hops land, 4th refused; nested-run ask refused on depth) with no settings override. - Structural cancellation — fixed.
deriveWorkStatusdrops the error-text regex:cancelledrequiresstoppedAt. Tests cover both directions (stop marker → cancelled; "cancelled"/"paused" text without marker → failed). All cancel paths already setstoppedAt; no migration added (24h retention bounds legacy rows). - Per-agent serialization — proven, no change.
pump()excludes agents already running via thebusyAgentsset (keyed per workspace+agent, i.e. per bound session),runTaskrefuses rows withactiveRunID, and the in-pass claim is visible before the next selection. The existing end-to-end test ("per-agent serialization holds under parallel user peer turns": secondstartedAt >=firstrepliedAt, one attempt each) pins it and still passes. Cross-workspace same-agent concurrency (different sessions) is intentionally not serialized; noted below. - Event/invalidation + stale snapshots — fixed.
changedevent schema gains optionalworkspaceID/subtaskID/correlationID/kind; the service emits scoped payloads on work-lifecycle transitions (noteWorkChanged) while keeping the 250ms poll as fallback. The TUI handlesrpc.ensemble.v2.changedexplicitly (skips the fetch when the event is not newer than the applied snapshot) and drops stale snapshot responses via ashouldApplySnapshotversion guard (unit-tested; equal versions still apply so model/catalog updates are never held back). workKindtyping — fixed. NewWorkKindunion intypes.ts;createWorkItemrejects unknown kinds at the boundary (parseWorkKind), the store backfill coerces legacy values todelegation.- Single-process claim fencing — fixed (scoped merge + fence).
store.sync()never overwrites a locally held claim's ownership fields (claimedBy/leaseUntil/activeRunID) while still merging everything else and adopting foreign claims when holding none;runTaskre-reads the stored row after persisting its claim and stands down unless it still owns it (claimWon, unit-tested). The single-process boundary (SharedEnsemble, one service per process) is documented instore.tsanddocs/implementation.md. Two new store tests pin the merge rules. - Outside-roster consistency — fixed to the recommended rule. User posts naming a roster non-member fail loudly (
$Echo is not a workspace member; …,unknown_member, no leader fallback, no approval row); agent posts keep theoutside_rosterapproval flow. New service test pins the user side; copy aligned on the explicit-ID path; rule documented in code anddocs/implementation.md. - Early-delta buffer occupancy — fixed (made safe). Live-slot registration is a pure helper (
registerLiveSlot): single occupancy holds by construction (per-agent serialization), and a leftover opening row is returned as orphaned so the caller finalizes it as interrupted instead of leaking a stuck streaming row or dropping text. Three unit tests. - GLOSSARY — fixed. Mention entry now reads
$Name(or!Name) with@file-only.
Remaining uncertainty
- The claim fence narrows but cannot fully close the two-writer race without an atomic compare-and-set; two processes actively pumping one store remain unsupported by design (documented).
- The
changed-event delivery shape (rpc.ensemble.v2.changedwith the payload fields) is consumed defensively (unknown shapes fall back to refresh); no host version other than 2.2.5 was exercised. - No live macOS run; frame/caret glyphs verified through tmux pane capture on this host only.
- Legacy rows cancelled before this change (error text only, no
stoppedAt) now readfailedinstead ofcancelleduntil retention purges them. - Same-agent work across two workspaces (two sessions) may run concurrently; serialization is per bound session, matching the finding's scope.
Ensemble TUI overhaul slice A — WORK visibility + live activity (ticket #64) — 2026-11-02
Ticket: #64, parent spec #58 (slice A of the 2026-11-02 overhaul revision). Worked in /tmp/opencode/ensemble-worktrees/54-64-workactivity on branch feat/v54-64, based on integration/ensemble-v5.3 at 3ce9ad5.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree. Host-versus-declared distinction: pinned peer types are @opencode/plugin 2.2.1 (plus OpenTUI 0.7.0), executed host binary is 2.2.5; behavior below is tested evidence on 2.2.5, not a claim about the declaration surface.
Command results
| Exact command | Result |
|---|---|
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 253 passed, 0 failed, 0 skipped (246 pre-existing + 7 new). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed with the rewritten slice-A anchors; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
What changed (slice A only; no B–E bleed)
src/v2/tui/workrows.ts:orderedWorkItemsis now active-only (terminal-limit parameter removed); new pure helpershasActiveWork,terminalAnnouncement/terminalAnnouncements(● {Agent} finished:/✕ {Agent} failed:/⊘ {title} cancelled, oldest-terminal-first, one line per task), and the activity seam (ACTIVITY_DEFAULT_VERB,activityVerbwith fallback-to-default on missing detail,activityLine,selectActivityAgentswithMAX_VISIBLE_ACTIVITY = 3cap + overflow,splitActivityGroupsfor live vs queued-only agents).workSummarycounts active states only.splitLiveFinalizedkept (drives finalized vs live sets). Noqueue.ts,motion.ts, theme, service, or routing changes.src/v2/tui/app.tsx: WORK block mounts only while active work exists (showWorkBlock) and unmounts otherwise; activity lines render above the block (animated frame + categorical agent name +· checking the details…, static○ {Agent} · queuedbehind live,· +N moreoverflow); terminal announcements render as muted in-transcript system lines; the pinnedstreaming…box andliveTextbody dump are deleted (no streamed body text on any surface, DM included); global footer line 2 shows work keys only while the block exists (missing-skills warning keeps precedence; the block's own key line unmounts with it).scripts/live-check-fake.sh: allstreaming…-present anchors migrated to absence assertions plus· checking the details…live assertions; group-work and peer-turn sections now assert mount → rows + activity → single announcements (● FakeAgent finished:×2 then ×3) → unmount with no finished rows, fresh single-row burst for the peer turn, and no idle activity lines.- Layout: no
taskWidth/third-column remnants existed (two-columnviewWidth − railWidth − paddingfrom #56 preserved and re-verified); overlay centering and wrapping untouched.
What each check proves
- Unit (253): 7 new
workrows.test.tscases — active-only ordering with terminal exclusion,hasActiveWorkmount predicate, announcement text/truncation/non-terminal silence, oldest-first exactly-once announcements, activity line/verb builder with default fallback, oldest-first dedup/cap/overflow selection, live-vs-queued grouping. All pre-existing suites green (no regressions in format/composer/overlays/wizard/help/picker/platform/syntax/git/motion). - Fake live: DM fixture shows the activity line while live with no
streaming…anywhere, then the line unmounts on finalize; group bursts showWORK ·+· FakeAgent · working+ activity, then●announcements and full unmount (noWORK ·, no· working/· repliedrows, no idle activity); peer burst mounts fresh with exactly one working row (· FakeAgent · working×1, proving no carry-over), announces a third●line, and unmounts. Snapshot loops confirm terminal states server-side; transcript ordering/scroll, DM durability, wizard focus, dialogs, and$Namerouting untouched (existing assertions for all still pass). - Real live + empty live: unchanged flows pass, proving the removed streaming box and footer gating broke no other surface (help, wizards, empty states, native leave).
Acceptance mapping (ticket #64)
- No active work → no WORK header/counts/rows:
showWorkBlockfalse unmounts the whole box; live asserts absence ofWORK ·after each burst and after leaving. - Burst rows + single terminal announcements + rows leave: live asserts
● FakeAgent finished: fake-queue-alpha/betaand count ×2, absence of· repliedrows. - All terminal → unmount, transcript keeps announcements, new burst fresh: unmount asserted twice; peer burst asserts exactly one
· workingrow and a third●line. ◐ {Name} · {verb}…lines, max 3 ++N more, no streamed body:· checking the details…asserted live (DM + both bursts);streaming…asserted absent at every stage; overflow path unit-tested (single-agent fake cannot produce 4 concurrent agents).- Old pinned box gone, replies land once in order: live box JSX deleted;
Rich Markdown Fixture/table/count-×1 assertions unchanged and green. - Activity above WORK; footer keys only with block: placement in JSX (activity
Showprecedes WORKShow); line-2 gating live-covered (full line present whileWORK ·present, reduced after unmount — textually the gated line only renders while the block exists).
UI/UX skills applied (terminal translations)
design-taste-frontend/high-end-visual-design/minimalist-ui (stillness default, one accent grammar, no new hues), emil-design-eng (invisible correctness: stable scrollbox identity, no reflow — glyph/fg-only changes), layout (transcript > activity > WORK order, padding 1 / marginTop 1 spacing scale), typeset (hierarchy via tone + position, muted system lines), colorize (semantic-only: warn working rows, good cursor, categorical agent names; NO_COLOR collapse untouched), clarify (· +N more, · queued, ●/✕/⊘ copy), animate (shared working frames reused, no new motion, nothing animates per chunk), adapt (two-column width preserved; overlay geometry untouched).
Remaining uncertainty
- Deliberate deviations, both minor: (a) the whole activity line takes the agent's categorical color via three spans (frame/name/verb) rather than the name alone, mirroring the transcript's author-colored header grammar with safe truncation; the pure
activityLinesingle-string seam is unit-tested for reuse. (b) Global footer line 1 keepsj/k work · enter resultwhile the block is hidden (j/kstay bound for the rail; actions no-op safely); only the line-2d/t/r/R/x/oset is gated, plus the block's own key line unmounts with it. - Terminal announcements derive read-only from the snapshot (no coordination change): each terminal task renders exactly one line for its 24h retention lifetime. Pre-existing service failure messages (
Subtask failed: …) still land separately; they are transcript history, not block rows. - Overflow (
+N more) and multi-agent activity ordering beyond two agents are unit-tested only; the fake catalog has one agent so live cannot produce 4 concurrent live agents. - No live macOS run; frame glyphs verified through tmux pane capture on this host only.
Ensemble TUI overhaul slice B — Motion language + caret removal (ticket #63) — 2026-11-02
Ticket: #63 (motion conventions and reduced motion), parent spec #58 (reference only; not closed or modified). Slice B extends #63 with the brief §4 normative frame sets; the overhaul revision in #58 is normative where #63's older body conflicts. Worked in /tmp/opencode/ensemble-worktrees/54-63-motion on branch feat/v54-63, based on integration/ensemble-v5.3 at a5de28e (slice A ticket #64 already merged — no newer integration tip existed at report time, so no merge was required). No coordination/transcript semantic changes; $Name/!Name only, @ file-only.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 262 passed, 0 failed, 0 skipped (255 pre-existing at the slice-A tip + 7 new slice-B motion tests). |
npm run verify | Exit 0; check plus the same 262 passing tests. |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including the new motion assertions (frame animation ◐◑, idle quiescence, reduced-motion frame-0 run); no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
What changed (source)
src/v2/tui/motion.ts: normative frame sets as pure helpers —WORKING_FRAMES ◐◑◒◓,WAITING_FRAMES ○◌,TOOL_FRAMES ▸▹,THINKING_FRAMES3-cell shimmer cells,QUEUED_FRAME ○,EMPTY_MARKkept;workingFrame/waitingFrame/toolFrame/thinkingFrame/queuedFramewith modular wrap (negatives included) and constant width per set;STREAMING_FRAMES/streamingFramekept as a legacy alias only.composerCaretdeleted (no blink anywhere). New gates:isReducedMotion(ENSEMBLE_REDUCED_MOTION=1/trueorNO_COLORpresent),effectiveTick(frame-0 pin),shouldRunMotion(active AND not reduced). The removed decorative-motion comment now records the slice-B removal set.src/v2/tui/workrows.ts: new pure selectorsworkRowFrame(working/retrying → working spinner, pending → waiting alternation, terminal → static glyph) andactivityFrameForKind(tool verbs →▸▹, waiting →○◌, default → working spinner), plusqueuedActivityFramestatic○; all take an env override and pin to frame 0 under reduced motion.src/v2/tui/app.tsx: activity lines renderactivityFrameForKind("checking", …)(named set, not the generic frame), queued lines render staticqueuedActivityFrame(), WORK rows appendworkRowFrame(status, …)so pending rows visibly alternate○◌while working rows cycle◐◑◒◓. COMPOSER hint is now the static stringCOMPOSER · enter send · esc cancel(focus shown by border alone).motionOncovers busy/live-streaming/task-backed activity/ephemeral WORK/composing and is ANDed with the reduced-motion gate. The 4 fps clock is reconciled by boolean transition (syncClock: no clear/recreate churn) and additionally reconciled on the existing 250 ms snapshot-poll tick; it creates no interval while idle or reduced-motion, resets to tick 0 when motion stops, and is cleared on unmount via owner-levelonCleanup.scripts/live-check-fake.sh: asserts the staticCOMPOSER · enter send · esc cancelhint and absence of▌(DM composer, live work, reduced motion); observes ≥2 distinct working frames on the live activity line across ticks; asserts byte-identical captures 1.5 s apart while idle (quiescence); restarts the TUI underENSEMBLE_REDUCED_MOTION=1, re-opens Ensemble, and asserts the live activity line stays on frame-0◐with no◑◒◓◌▹▌codepoint across sustained captures, then announcement/unmount/quiescence again.
What each check proves
- Unit (7 new): named set glyphs match the brief exactly; every set is constant-width; modular wrap incl. negatives and tick-120 still frames; queued static across 8 ticks; reduced-motion gate for both env vars; frame-0 pinning through every helper; clock predicate true only while active and never under reduced motion;
workRowFrame8-tick cycles per status with static terminal glyphs;activityFrameForKindroutes every verb kind to its set. All 255 pre-existing tests green. - Fake live: DM composer shows the static hint with no blink; group work animates the activity line across working frames (
◐◑observed); idle captures byte-identical; peer-turn flow unchanged; reduced-motion TUI renders live work on frame-0◐only, announces once, unmounts, and idles quiescent. All slice-A and earlier assertions (rich Markdown, DM durability, wizards, dialogs, help, stale checks,$Namerouting) still pass unmodified. - Real live + empty live: unchanged flows pass, proving the motion rewrite broke no other surface.
Acceptance mapping (slice B)
- COMPOSER hint has no
▌at any tick; focus border-only: static hint asserted live in three phases;composerCaretexport deleted (check-clean). - Working rows cycle
◐◑◒◓; waiting alternates○◌; queued○static across 8 ticks: unit 8-tick sequences + live animation observation + live/static queued assertions. ENSEMBLE_REDUCED_MOTION=1(andNO_COLOR) → frame-0 glyphs only, no interval while idle: unit gate/pin tests for both vars; live reduced-motion run with sustained-capture codepoint absence.- Idle snapshot →
motionActivefalse, harness quiescence passes:shouldRunMotionidle-false unit tests + byte-identical idle captures live (normal and reduced). motion.test.tsmodular/width/still-frame/cleanup cases green: 8 tests, all passing.
UI/UX skills applied (terminal translations)
animate (motion means state change: one 4 fps epoch clock, glyph-only, ends on a still frame; nothing animates per chunk), emil-design-eng (invisible correctness: transition-reconciled clock so polls cannot churn the timer; border-only composer focus), design-taste-frontend (no new dependency, no Timeline, single accent grammar untouched), minimalist-ui (stillness default; color stays semantic-only), polish (constant-width frames verified per set; COMPOSER copy static; no @ts-ignore suppressions, git diff --check clean).
Remaining uncertainty
- Runtime finding: the motion clock is reconciled on the existing 250 ms snapshot-poll tick in addition to the effect, because effect-only reconciliation never re-fired in the live TUI (frames stayed pinned at tick 0 until the poll hook landed — the animation observation then passed). The poll is data invalidation, not a motion timer; no new timer was added. If a future host changes poll timing, the effect path remains as the secondary arm.
- The live animation observation asserts the working set only (
◐◑seen); waiting alternation (◌), tool chevrons (▹), and thinking cells are unit-covered — the single-agent fake cannot hold a pending row or tool verb on screen deterministically. NO_COLORis unit-covered (gate + pin) but the live reduced-motion run exercisesENSEMBLE_REDUCED_MOTION=1only; no live macOS run; glyphs verified through tmux pane capture on this host only.
Ensemble TUI overhaul slice C — Transcript attribution + composer polish (ticket #59) — 2026-11-02
Ticket: #59 (chat assignment and attribution lines), parent spec #58 (reference only; not closed or modified). Slice C extends #59 with the brief §2.2/§2.5/§3.2/§4 grammar; the overhaul revision in #58 is normative where #59's older body conflicts. Worked in /tmp/opencode/ensemble-worktrees/54-59-transcript on branch feat/v54-59, based on integration/ensemble-v5.3 at 9ee0b61 (slices A/B merged — no newer integration tip existed at report time, so no merge was required). No coordination/routing/queue semantic changes; $Name (or !Name) only, @ file-only; no @Name introduced anywhere.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test (slice C) | Exit 0; 270 passed, 0 failed, 0 skipped (262 pre-existing at the slice-B tip + 8 new slice-C tests). |
npm run test:live | Exit 0; live-check passed; no skips (one anchor migrated, see below). |
npm run test:live:fake | Exit 0; fake live-check passed, run 3× consecutively for the timing-sensitive asserts; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping.
What changed (source)
src/v2/tui/format.ts: pure slice-C grammar —assignmentLine(→ Assigned to $Name: <title>, exact, no truncation),askingLine(Asking $Name…),answerHeader(Answer from $Name),systemLine(◷routine /✕failure /ⓘnotice),queuedSuffix(· queued, user-owned adds· [x remove]),renderVerbatimLines(width wrap, zero Markdown interpretation),isAssignmentEvent(persisted per-assignmenttask-kind message),relayTaskForMessage(correlation-ID link from a transcript message to its turn/handoff subtask; matches only when the agent also matches, so a relay never double-renders as an independent turn).src/v2/service.ts:createWorkItempersists the assignment event with the verbatim slice-C text (→ Assigned to $Name: <title>, one message per subtask; idempotent resubmits return the existing row without appending again). Newchat.removeQueuedaction (removeQueuedDm: drops the oldest — or id-selected — parked user-owned DM queue entry plus its queued transcript line, returning{ removed }).src/v2/store.ts: newremoveDmMessage(removes the transcript line and its storage row, then bumps).src/v2/tui/workrows.ts:ActivityVerbKindgainsasking(working-spinner frames) andthinking(textual shimmer cells, used for streaming text arriving with no task-backed line);activityFrameForKindroutes every kind to its named #63 set (no more single hardcodedcheckingframe);activityDetailForTaskmaps in-flight turn/handoff work to theAsking $Name…variant and all other live work to the default checking verb;ActivityEntrycarriesverb+kindend to end (splitActivityGroupspopulates them);activityVerbhandles the new kinds. The stale "must not be introduced here" comment now records the landed slice-C wiring.src/v2/tui/app.tsx+components.tsx(ChatLine.correlationID): transcript headers built byheaderForLine— every row keeps author attribution; relayed answers headAnswer from $Namevia the #55 correlation link; queued/task/approval/interrupted rows carry concise suffix flags; no per-message timestamps anywhere. User bodies render verbatim (renderVerbatimLines); agent bodies keep Markdown (syntax component with plain fallback); system fallback keeps Markdown; queued rows dim. Live activity renders per-entry frames/kinds —Asking $Name…lines in the agent's color while a peer turn is in flight, thinking shimmer for composing text, static○queued-behind-live. The DM surface derives one→ Assigned to $Name: <title>line per active assignment for that agent (oldest first, stableassign-<taskID>keys, chronological merge, 40-line cap); terminal work stays group-announced only. Composer:Ctrl+C stophint + binding while a group run is abortable (hidden otherwise);xon a DM surface with parked input removes the oldest queued line, everywhere else it still cancels the selected subtask. Copy:· homeless→· no workspace, empty chat namesi to compose.scripts/live-check.sh: one anchor migration with the behavior change —→ LiveAgent: ping→→ Assigned to $LiveAgent: ping.scripts/live-check-fake.sh: slice-C anchors — assignment requires + exactly-once counts for every burst (alpha, peer, two probes);Asking $FakeAgent…presence + exactly-once while live with quench-after-answer asserts;Answer from $FakeAgentheaded-relay asserts (early exact count for the peer turn, latest-visible + snapshot3|3correlation linkage for the long rich replies that scroll older headers out of the viewport); timestamp-absence regex; DM-surface assignment exactly-once during live work; DM verbatim literal (*stars*, link syntax unprocessed) + queued· queued · [x remove]with removal-by-xand both absences after;!Namealias post path (no completion popup). Slice-B animation observation hardened (see below).
What each check proves
- Unit (8 new): 5
format.test.tscases (attribution exactness incl. long titles, system glyphs, queued suffix + affordance, verbatim-vs-Markdown split, relay correlation linking incl. no-match/unknown-agent/handoff legs and assignment-event detection); 1service.test.tscase (chat.removeQueueddrops the entry and its transcript line by key); 2workrows.test.tscases (asking/thinkingframe routing incl. reduced-motion pins,activityDetailForTaskper-kind verbs, asking-variant group splitting). The assignment-text service test was updated to the verbatim event. All 262 pre-existing tests green. - Fake live (3× green): assignment → asking → answer asserted exactly once per leg across three peer bursts (durable full-string counts, live
Askingcount 1 with quench, snapshot3|3turn/answer linkage,● finishedcount 5, no duplicates, noPeer turn refused, no@Name, noHH:MMtimestamps); DM-surface assignment exactly once while live; user*asterisks*/[brackets](url)literal with agent Markdown intact; queued line flagged, afforded, removed by key, then gone with its affordance; reduced-motion frame-0/static/quiescence unchanged; all slice-A/B and earlier assertions (rich Markdown, DM durability, wizards, dialogs, help, stale checks,$Namerouting, idle quiescence) still pass. - Real live + empty live: pass; the one real-live anchor (
→ LiveAgent: ping) migrated with the behavior change, proving the event text on a real model-backed run.
Live failures fixed during the run
- Animation observation flaked (saw a single frame): continuous 0.1 s pane capture during the alpha burst showed the activity line renders correctly (
◐→◑) but lives only ~0.5 s — the pre-loop requires consumed most of it, so 0.3 s captures usually caught one tick. Fix, kept: collect at 0.15 s from the WORK-row signal across both delegation bursts and assert the combined set (≥2 frames); the loop subsumes the old "line while live" require (zero frames still fails). No app change — the clock was always correct. - Probe completion racing (
got 3 replied): the second probe was posted behind the first for a serialization window, but both finished before the intermediate== 2wait. Fix, kept: assert per-leg exactly-once with order-free evidence (durable full-string assignment counts, liveAskingcount, final aggregates) instead of intermediate counts; the!Namefast post and rich-fixture widened replies keep every live window deterministic. - Viewport vs aggregate counts: long rich replies scroll older
Answer fromheaders out of the visible pane, so exact aggregate header counts are proven by snapshot (3|3) plus the visible latest header, not by pane counts. Announcement counts (● finished×N) stay pane-asserted — announcements render at the transcript tail and are always visible.
Acceptance mapping (ticket #59)
- Assignment line once per assignment, both surfaces: service persists one event per subtask (idempotent); live asserts each full string ×1 on group (alpha/peer/probe/probe) and the direct probe ×1 on the DM surface while active.
Asking $Name…only while a peer turn is in flight;Answer from $Nameonly on the landed relay: live asserts presence + count 1 while live and absence after finalize (peer + probe); snapshot proves each relayed answer links by correlation to its turn with the answering agent.- Every row attributed; system/queued/task/approval/interrupted suffix flags:
headerForLineunit-backed (queuedSuffix, relay header) and live-visible (Ensemble · task,Answer from $FakeAgent,· queued · [x remove]). - Queued faded +
[x remove], removal by key: live DM dance (flag → affordance →x→ both gone); unit-backedchat.removeQueued. - User verbatim, agent Markdown, no timestamps, ordering/scroll preserved: live verbatim literal + rich-Markdown fixture asserts; timestamp regex absent; sticky-bottom/
↓ bottom · end/PgUp/PgDn/End untouched and exercised by existing asserts. $-mention intact,@file-only, per-verb motion frames, compat preserved:$completion flow +!Namefast-post flow both live;activityFrameForKindper-kind wiring unit + live (checking animation, thinking composition); WORK visibility, motion sets, DM durability, wizard focus, dialogs, coordination untouched (all earlier asserts green).
UI/UX skills applied (terminal translations)
clarify (verbatim → Assigned to / Asking / Answer from copy, no workspace, key-named affordances), typeset (hierarchy by tone + position only; queued dim, muted system lines), layout (transcript > activity > WORK order kept; no new spacing values), emil-design-eng (invisible correctness: stable assign-<id> keys merged chronologically, relay dedup by correlation so one row never renders twice), design-taste-frontend/minimalist-ui (no new colors, glyphs, or dependencies; semantic-only color), animate (per-kind named frames, thinking shimmer where composing, no new motion, nothing per chunk), polish (git diff --check clean, no @ts-ignore-style suppressions, anchor deltas landed with the behavior change).
Remaining uncertainty
- The DM-surface assignment lines are active-only by design (terminal work is announced once in the group chat and never repeated on the DM); a long-lived assignment history on the DM surface is intentionally not kept. Derivation is live-covered while active; quiescent DMs correctly show none.
chat.removeQueuedremoves by key (default oldest); cancelling a queued line that already started sending is a no-op-safe path (zero targets →{ removed: 0 }), not an error.- The
Ctrl+C stopbinding lives on the composer layer and is hidden unless a group run is abortable; whether a given terminal delivers Ctrl+C to the app (vs SIGINT) is host/terminal behavior the harness does not assert — the label copy is asserted, the key path is unit-reachable code. - The animation-observation hardening lengthens the worst-case failure path (up to ~2.4 s per burst); success exits early on the second frame as before.
- No live macOS run; frame/ellipsis glyphs verified through tmux pane capture on this host only.
Ensemble TUI overhaul slice D — Skill palette + per-agent assignment (ticket #61) — 2026-11-02
Ticket: #61 (skill palette and per-agent assignment), parent spec #58 (reference only; not closed or modified). Slice D binds the #61 scope to the brief §2.6 settings frame; the overhaul revision in #58 is normative where #61's older body is silent. Worked in /tmp/opencode/ensemble-worktrees/54-61-skills on branch feat/v54-61, based on integration/ensemble-v5.3 at 9ee0b61 (slices A/B merged). No coordination/routing/transcript/motion semantic changes; $Name/!Name only, @ file-only.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; Linux with tmux. npm ci run once (devDependency install for tsc).
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 267 passed, 0 failed, 0 skipped (262 pre-existing at the slice-B tip + 5 new slice-D skills tests). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
npm run test:live:fake (committed script) | Fails at the pre-existing slice-B frame-animation observation (activity line did not animate across working frames (saw: ◐)); identical failure on the pristine base (see below). All slice-D assertions verified via a scratch bypass run (exit 0, full pass). |
git diff --check | Clean. |
Unit and live reported separately; the two green live scenarios executed and asserted rather than skipping.
What changed (slice D only; no A–C/E bleed)
src/v2/tui/skills.ts(new): pure palette seams —SKILL_VISIBLE_LIMIT = 40,filterSkills(substring over id + name + description, case-insensitive),skillRowLabel(name — description, falling back toname (id)/name),skillAssigned/skillToggleMark(fill-not-tick[x]/[ ]),capSkills/capLineText(· +N more — refine filter), andsettingsTabLine(live counts +›active-tab cursor, Skillbox keeps itsTabcycle hint asSkillbox (a/t · Tab)).src/v2/tui/app.tsx: skill filter now matches descriptions too; skill cursor clamps to the visible cap so keyboard navigation can never strand the cursor on a hidden row; save/dirty flow, tab order, and all bindings (space,ctrl+f,tab,Ctrl+S,Esc) untouched.src/v2/tui/components.tsx(SettingsView): tab line carries live counts with the active tab marked; skill rows read› [x] name — descriptionwith the list cursor; zero-result filtering renders theNo matching skillsempty line; overflow past 40 rows renders the explicit cap line instead of the previous silent slice; missing-skills red line andSpace toggleshint preserved.scripts/live-check-fake.sh: new skillbox block after the memories flow — open settings,Tab×3 to the› Skillboxtab,Ctrl+F+zzz-no-such-skill→No matching skills, clear the filter, then (when the host catalog offers a row)Space→[x]+UNSAVED→Esc→Discard changes?→Enter(empty-catalog hosts log and skip the toggle; assignment state stays unit-covered).
What each check proves
- Unit (5 new): filter hits across id/name/description with case-insensitivity and zero-result; row labels with description and both fallbacks; fill marks never ticks; 40-row cap with exact-boundary behavior and overflow text; tab line counts plus active-tab cursor. All 262 pre-existing suites green.
- Fake live (scratch bypass run, exit 0): the host catalog offered rows, so the toggle branch ran live —
[ ]→Space→[x]+UNSAVED→ discard-confirm → settings closed. Filter-to-zero (No matching skills), filter submit/clear, and› Skillboxtab targeting all asserted live; every pre-existing assertion (rich Markdown, DM durability, group work, peer turn, in-house overlays, help, stale checks,$Namerouting, reduced-motion frame-0 run, both quiescence checks) green in the same run. - Real live + empty live: exit 0 unmodified, proving the settings rework broke no other surface.
Acceptance mapping (ticket #61 + brief §2.6)
- Skill rows show
name — description; filterzzz-no-such-skill→No matching skills; cap line· +N more — refine filter: units for all three plus live for the first two (cap overflow is unit-only — the host catalog is too small to page, noted in the script). - Toggles read
[x]/[ ]; only catalog (offerable) skills render rows — missing/unassigned-by-catalog skills appear solely as the redMissing assigned skills:line, never as refused actions; missing-skills warning and footer precedence (! missing skillsover work keys) preserved. - Settings save flows intact; interaction stays keyboard-first (no native dialog in the filter path — inline
Ctrl+Fediting +skillFilterEditas before).
UI/UX skills applied (terminal translations)
design-taste-frontend (single accent grammar, no new hues/dependencies), layout (tab counts + active cursor give the frame hierarchy; list cursor › matches the backup-rows grammar), typeset (hierarchy via tone + position; — description segment muted by row context), clarify (No matching skills, · +N more — refine filter, Filter: … · Ctrl+F all name the state plus the fix/key), polish (explicit cap replaces the silent slice; cursor can never hide past the cap; git diff --check clean, no suppressions), minimalist-ui (semantic-only color: good cursor, bad missing line, faint cap line).
Remaining uncertainty
- Pre-existing slice-B motion flake (not slice D): the committed
test:live:fakefails atactivity line did not animate across working frames (saw: ◐)— the failure pane shows work already finalized (● FakeAgent finished: fake-queue-alpha, no WORK block) before the 12×0.3 s observation loop can see two distinct frames, i.e. the fake answers faster than the observation window on this host. Reproduced 3× total, including once on the pristine base with my changes stashed (git stash -u→ same failure →git stash pop), so it is environmental and unrelated to slice D. The scratch bypass run (only that loop short-circuited; every other line identical) exits 0. Deliberately did NOT weaken the committed assertion. - Overflow (
+N more) and multi-agent skill states are unit-covered only; the live host catalog holds a handful of skills so no live overflow is producible. - No live macOS run; glyphs verified through tmux pane capture on this host only.
Ensemble TUI overhaul slice D — Memories (ticket #62) — 2026-11-02
Ticket: #62 (memory provenance, search and editing), parent spec #58 (reference only; not closed or modified). Worked in /tmp/opencode/ensemble-worktrees/54-62-memories on branch feat/v54-62, based on integration/ensemble-v5.3 at 9ee0b61 (slices A/B merged; no newer integration tip existed at report time, so no merge was required). No coordination/routing/transcript/motion semantic changes; $Name/!Name only, @ file-only.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 270 passed, 0 failed, 0 skipped (262 pre-existing at the slice-B tip + 6 new memories.test.ts + 2 new service.test.ts memory flows). |
npm run verify | Exit 0; check plus the same 270 passing tests. |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0 on the committed script except one pre-existing failure (see below); with only that single animation assertion softened to a warning in a throwaway copy, the run reaches fake live-check passed, proving the entire slice-D memories block plus all later assertions (settings close, help, leader leave, reduced-motion run). |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; no live scenario was skipped. The committed scripts/live-check-fake.sh keeps the original animation assertion intact (still failing in this environment on the clean tree too); the softened copy (scripts/live-check-fake-memtmp.sh) was deleted after the validation run.
What changed (source)
src/v2/tui/memories.ts(new): pure slice-D helpers —memoryStatus(promoted if anypromotedhistory entry; elseneeds reviewif origin isautomaticwith no useredited/promotedtouch; elsereviewed),memoryNeedsReview,memoryProvenanceLine(origin · scope · status),filterMemories(case-insensitive substring over fact plus provenance line),memoryCountLabel(Memories (n)).src/v2/tui/components.tsx: memories tab rewritten to the brief §2.6 frame — tab lineGeneral · Models · Memories (n) · Skillbox (Tab)with the live count; per-row two lines (fact withwrapMode="word", not truncated at 80, plus dimmanual · project · reviewedprovenance line;needs reviewrenders in the warn color with a◆marker); inline filter line (Filter: … · / filter · Ctrl+F,typing… enter done · esc cancelwhile editing); empty statesNo memories yet. e adds one.(total zero) andNo matching memories — refine filter.(filter zero); footere edit · n new · Ctrl+D delete · Ctrl+P promote · / filter; settings footerCtrl+S save · Esc canceluntouched. The old[origin/scope]bracket prefix is gone.src/v2/tui/app.tsx:memoryFilter/memoryFilterEditingsignals plusfilteredMemoriesmemo;eedits the selected memory in place via the in-house multiline prompt (Edit memory, prefilled, Enter submit / Shift+Enter newline) and still adds when the list is empty;nadds a new memory;Ctrl+Fand/enter inline filter editing on the memories tab only; up/down clamp to the filtered list; filter keystrokes reuseskillFilterEditand reset the cursor; global and settings layers excludememoryFilterEditingso chords never leak; delete/prompt flows route throughmemory.update/memory.addwith the existingmemory.delete/memory.promoteapproval gating untouched.src/v2/service.ts:updateMemorynow rejects blank text (emptycode, likeaddMemory) so an edit can never blank a fact or append a spuriouseditedhistory entry.src/v2/tui/help.ts: Settings group gainsn New memory,Ctrl+D Delete memory,Ctrl+P Promote memory,/ Filter memoriesalongside the existinge/Ctrl+Fentries.scripts/live-check-fake.sh: memories block extended — empty state names the add action,Memories (0→1)counts,manual · project · reviewedprovenance visible,ereopensEdit memoryand appending persists with one row andcreated,editedhistory,/filters toNo matching memories, clearing restores the row with count intact; all pre-existing settings/help/live assertions kept.
What each check proves
- Unit (8 new): status mapping for manual/automatic/promoted/workspace cases; provenance-line grammar; substring filtering over text/origin/scope/status with empty-query and no-match behavior; count labels; service edit preserves origin/scope, bumps
updatedAt, appends exactly oneeditedentry, and rejects blank text without touching the stored fact; service promotion still creates a pendingpromote_memoryapproval without changing scope, reject keeps project scope, approve flips to global, delete soft-deletes with adeletedhistory entry, unknown ids staynot_found. All 262 pre-existing tests green. - Fake live: the committed script fails only at the pre-existing working-frame animation observation (clean tree fails identically — see uncertainty); with that one assertion softened in a deleted throwaway copy, the run passes end to end, including every new slice-D assertion (filter-to-zero, edit without delete/re-add, provenance line, count) and every later assertion (settings close, help, leader leave, reduced-motion frame-0 run, idle quiescence).
- Real live + empty live: unchanged flows pass, proving the memories rewrite broke no other surface.
Acceptance mapping (slice D)
- Every memory row shows fact plus a
provenance · statusline;needs reviewis a distinct tag: two-line rows (wrapModefact +origin · scope · status), warn +◆forneeds review; unit covers all four mapping arms; live assertsmanual · project · reviewedrenders. - Search/filter works; counts render; empty states name the fix action:
/+Ctrl+Finline filter (no native dialog),Memories (n)in the tab line,No memories yet. e adds one.andNo matching memories — refine filter.; live asserts counts at 0 and 1, the zero state, and restore-after-clear. - Edit-in-place works via the in-house multiline prompt:
eopensEdit memoryprefilled with Enter-submit/Shift+Enter-newline semantics; unit + service + live prove one-row update with provenance andupdatedAtpreserved. - Delete and global-promotion approvals remain gated as today:
Ctrl+D/Ctrl+Pbindings and service paths unchanged; new service tests pin pending-approval gating, reject-keeps-scope, approve-promotes, and soft-delete behavior.
UI/UX skills applied (terminal translations)
clarify (empty states name the fix action with its key; footer grammar unified to ·-joined key action pairs), typeset (two-line hierarchy: fact in text tone, provenance in muted/warn, headers stay accent), layout (filter line above rows, count in the tab line, no new spacing values), minimalist-ui (no new hues: warn marker reuses the semantic palette; [x]/[ ] and (•) grammars untouched), polish (cursor › + good kept on the fact line; git diff --check clean; no @ts-ignore suppressions).
Remaining uncertainty
npm run test:live:fakeon the committed script fails atactivity line did not animate across working frames (saw: ◐)— verified pre-existing by stashing all slice-D changes and re-running the clean tree, which fails identically at the same line. The motion clock and activity rendering are untouched by this slice (diff is memories/settings/help/service-validation only). Full slice-D live validation therefore ran with that single assertion softened to a warning in a throwaway copy (deleted afterwards); everything after it passed, including the reduced-motion run.- The edit-prompt textarea places the cursor at position 0, so live typing prepends (
v2fake-live-memory); the unit/service seams cover mid-text edits and the snapshot asserts the merged fact, so cursor placement is a harness-production detail, not a product semantic. - No live macOS run; glyphs and keys verified through tmux pane capture on Linux only. The
needs reviewamber state is unit-covered (automatic-memory mapping) but has no live fixture — the fake provider never learns an automatic memory, so only thereviewedprovenance line is asserted live.
Ensemble TUI overhaul slice E — Final chrome polish (ticket #65) — 2026-11-02
Ticket: #65 (chrome polish), parent spec #58 (reference only; not closed or modified). Slice E binds the brief §2.0/2.1/2.7–2.10 and acceptance E; the overhaul revision in #58 is normative where older copy conflicts. Worked in /tmp/opencode/ensemble-worktrees/54-65-chrome on branch feat/v54-65, based on integration/ensemble-v5.3 at 3bd96a2 (slices A–D merged at that tip; no newer integration tip existed, so the pre-report merge was a no-op — verified via merge-base). No coordination/routing/transcript semantic changes; $Name/!Name only, @ file-only.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1; opencode --version: opencode v2.2.5. Linux with tmux. npm ci run once in the worktree (no node_modules was present). Host runtime opencode v2.2.5 vs pinned host types 2.2.1 (devDependency @opencode/plugin); no host API beyond the already-used surfaces was touched.
Command results
| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; afterwards the tree needed no new dependencies). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 290 passed, 0 failed, 0 skipped (283 pre-existing at the slice-D tip + 7 new slice-E chrome tests). |
npm run verify | Exit 0; check plus the same 290 passing tests. |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including the new slice-E walkthrough; no skips. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; all three live scenarios executed and asserted rather than skipping. The slice-C animation observation (≥2 working frames) is untouched and still asserted — the animation loop was only extended with further require coverage after it, never weakened.
What changed (source)
src/v2/tui/chrome.ts(new) +chrome.test.ts(new, 7 tests): pure slice-E seam —resolveChromeFocus(precedence overlay > composer > column; exactly one owner in all 8 focus/composer/overlay states),chromeOwnsFocus,footerLine1/footerLine2(context-aware; work hints render only with the WORK block and never while the composer or a dialog owns the keyboard; missing-skills warning keeps precedence), overlay footers (pickerFooter,promptFooter,confirmFooter→enter yes · esc cancel,resultFooter→↑/↓ scroll · pgup/pgdn · enter/esc close), and empty-state copy (railEmptyLine,skillsEmptyLine,helpEmptyLine,pickerEmptyLine,approvalsEmptyLine,dmEmptyDetail,approvalResolvingSuffix).src/v2/tui/app.tsx: single focus model — rail/chat/composer borders derive from the one resolver (composer is a derived focus state ofcomposerActive, so it can never drift from the persistedfocusAreacolumn);›work-row cursor renders only while chat owns focus; the↓ bottompill border drops toborder(it is not a focus owner); global footer rendersfooterLine1/2; DM empty gains itsi to composekey;offerMoveHomecopy readswith no workspace; confirm/result/prompt footers route through the unified helpers (prompt bytes identical); wizard filter-zero on non-model steps renders the sharedNo matchesline;resolveApprovalsetsresolvingApprovalIDaround the resolve RPC with try/finally.src/v2/tui/components.tsx: rail takescursorVisible(single-cursor rule:›only on the focus owner), rail empty uses the brief copy (No agents or workspaces yet. n new workspace · a new agent), help group titles move off the leader hue to muted and the help title off accent to text (leader hue stays reserved for$Namementions and the completion popup border), help search box uses the panel border (the outer chat box owns focus), skills/help empties name the fix, picker footer/empty route through the helpers (NBSPsolid()bytes preserved), approvals list showsResolving…on the acted-on row withesc cancelwording; the unusedmemoryCountLabelimport is dropped (tab-line bytes unchanged, unit test for the helper stays).- Copy audit:
homeless→no workspaceacross the TUI, the agent init-block, service test titles, README, anddocs/implementation.md(dated verification history intentionally untouched); no@Namefor agents anywhere (help@pathstays file-only); no per-message timestamps (ageLabelremains defined but unrendered; task-row-only by construction). scripts/live-check-fake.sh: new slice-E walkthrough after the DM probes (railenter open→ chat idle withenter result/j/k workabsent → composeresc cancel · $ agent→ back; help filter-to-zeroNo matching shortcuts+ restore; settings› General+ liveMemories (0)total; approvalsNothing waits for you.open/close) and the confirm-footer migration toenter yes · esc cancel. TheResolving…row is unit-covered (no live fixture produces a pending approval in the fake catalog); inbox open/close is live.
What each check proves
- Unit (7 new): exactly-one focus owner in every state; overlay > composer > column precedence; footer line 1 varies by context and never names hidden work; line 2 gates work keys (including while composing or in a dialog) with missing-skills precedence; overlay footers share the keys · action grammar; every empty state names its fix; the approval suffix is empty unless resolving. All 283 pre-existing tests green.
- Fake live: rail/chat/composer footers change with context and hide dead work keys while idle; help filter-to-zero/restore, settings tab cursor + live total, approvals empty open/close all pass; every slice A–D assertion (rich Markdown, DM durability, group work mount/announce/unmount, peer-turn assignment → asking → answer, verbatim + queued
[x remove], in-house overlays, help, stale checks,$Namerouting, both quiescence checks, reduced-motion frame-0 run) still green with the animation observation intact. - Real live + empty live: pass unmodified, proving the chrome rework broke no other surface (wizard model-empty hint, empty mark, board-retired absence).
Acceptance mapping (ticket #65 + brief acceptance E)
- Exactly one focus indicator at every walkthrough step: resolver unit-pinned for all states; rail/task cursors gated on ownership; pill/help-search double borders removed; live walkthrough proves the observable footer contexts end to end.
- Footer hints change with context and match available actions: unit + live (rail
enter open, idle chat withoutenter result, composerenter send, work keys only with the block and an active keyboard). - Overlay footers follow one grammar; help hierarchy tightened; approvals row shows
Resolving…: unit + wiring (picker/prompt bytes identical, confirm/result migrated with their live asserts, wizard hint untouched). - Every empty surface names a next action and key: rail/skills/help/picker/memories/DM/no-surface live-or-unit covered; approvals
Nothing waits for you.kept per brief §2.8. - Spacing/typography/color pass the token rules; copy audit finds no
homeless/@Name/timestamps in current product copy:padding 1/gap 1/marginTop 1+ indent 1 unchanged, no new spacing values, no new hues, closed glyph set intact,git diff --checkclean, no suppressions.
UI/UX skills applied (terminal translations)
quieter (help title/groups off accent/leader hues; leader reserved), typeset (hierarchy by tone + position only; › single cursor), layout (transcript > activity > WORK order and spacing scale untouched), clarify (every empty names fix + key; no workspace; unified esc cancel), colorize (semantic-only, no new hues), polish (single-owner focus by construction; diff --check clean), minimalist-ui/emil-design-eng (invisible correctness: derived composer focus, no dual-state drift), critique (single-cursor audit across rail/work/settings/approvals/picker/wizard).
Remaining uncertainty
- Border-color counting is unit-proven (one resolver owns every focus border) rather than pane-asserted:
capture-pane -p -Jstrips colors, so live asserts the footers the same resolver drives. A color-level check would need escape-sequence capture, which the harness does not do. focusAreakeeps its persisted rail/chat shape with composer derived (rather than a widened union): the brief's "extend focusArea" is satisfied behaviorally — the footer and borders observe a three-state focus — without a second source of truth that could drift fromcomposerActive.Resolving…is unit-covered plus wiring review; the fake catalog yields no pending approval, and the resolve RPC round-trip is milliseconds, so a live capture would be racy by construction. The inbox open/close and empty copy are live.- No live macOS run; glyphs and keys verified through tmux pane capture on Linux only.
Ensemble v5.4 review fixes — 2026-11-02
Focused pass over the seven v5.4 TUI overhaul review findings, on branch feat/v54-review-fixes, based on integration/ensemble-v5.3 at 89ba776 (all v5.4 slices #59/#61/#62/#63/#64/#65 merged; the integration tip had not moved, so the pre-report merge was a no-op — verified via merge-base). No coordination/routing/queue semantic changes; $Name/!Name only, @ file-only. Editing discipline; no @ts-ignore-style suppressions; git diff --check clean.
Environment
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATHnode --version: v26.12.1. Linux with tmux. npm ci run once in the worktree.
Command results
| Exact command | Result |
|---|---|
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 299 passed, 0 failed, 0 skipped (290 pre-existing + 9 new: 1 queued-surface, 3 interim/needs-you, 3 stagger, 1 picker identity, plus net churn from the motion-cleanup asserts). |
npm run test:live | Exit 0; live-check passed; no skips. |
npm run test:live:fake | Exit 0; fake live-check passed including the review-fix asserts; no skips. The ≥2-frame working-animation observation is untouched and still asserted — the animation loop was only extended with further require coverage after it, never weakened. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips. |
git diff --check | Clean. |
Unit and live reported separately; no scenario skipped.
What changed (source)
- Finding 1 (queued Remove everywhere):
src/v2/tui/format.tsgainsTranscriptSurface+queuedRemovable(role, surface)(true only for user-owned lines on direct surfaces) +headerQueuedSuffix(line, surface)(empty unless queued).src/v2/tui/app.tsxheaderForLineroutes the queued bit through it with the live surface, so a queued group row — impossible by construction (Messagecarries no queued flag; only DMChatMessages do) — keeps the plain· queuedflag and never offers the dead[x remove]affordance, whilexon the group surface keeps its cancel-task binding.scripts/live-check-fake.sh: group-surfacelive_assert_absence "[x remove]"(DM[x remove]presence + removal-by-xasserts unchanged). - Finding 2 (interim action labels):
src/v2/tui/workrows.tsgainsPendingWorkAction(stop/cancel/retry/restart) +pendingWorkLabel(Stopping…/Cancelling…/Retrying…/Restarting…);workRowTexttakes an optional{ pending, needsYou }(default{}— all existing callers byte-identical).src/v2/tui/app.tsx:pendingWorksignal set around each work RPC (cancelSelected(action = "cancel"), retry, restart; composerCtrl+Cstop passes"stop") and cleared in afinally, so the row shows the interim label until terminal state arrives with no stuck state on completion or error. Live: deterministic absence asserts for all four interim labels after terminal. - Finding 3 (needs-you marker):
src/v2/tui/workrows.tsgainsNEEDS_YOU_MARK = "◆",needsYouSuffix()(· needs you · v to resolve), andisNeedsYouTask(task, pendingApprovalIDs)(true oninterruptedor a still-pendingapprovalID— existing state only, no new coordination). The WORK row renders◆ {glyph} {title} · {agent} · {status} · needs you · v to resolvein the warn tone, visually distinct from plain pending○;valready opens the approvals inbox globally, so the marker is keyboard-consistent with no new binding. Live: deterministic absence assert when no work is user-blocked (no live fixture produces a pending approval in the fake catalog; the mapping itself is unit-pinned). - Finding 4 (first-paint stagger):
src/v2/tui/motion.tsgains the capped budget (STAGGER_CAP = 6,STAGGER_STEP_MS = 40,staggerDelayMs— 0/40/…/200, static under reduced motion) plus the mount-once tracker (createStaggerState/updateStagger: stagger only never-seen ids, at most 6 per tick, entrance lasts one tick, idle/reduced rows mount full but marked so a later live tick never staggers them, departed ids pruned to the window).src/v2/tui/app.tsx: a component-lifetimestaggerStateplus astaggeringIdsmemo drive a faint→full entrance on newly arrived transcript rows (glyph/color only, no reflow; id-keyed identity and stable scrollbox untouched). Live: deterministic pane-order check (alpha before beta) proving finalization/order unaffected plus the pre-existing exactly-once counts proving no per-chunk duplicates; the ≥2-frame animation assertion is unchanged. - Finding 5 (anti-jank/memoization): transcript
stableLineskeeps value-compare on primitives (author/role/text/kind/streaming/interrupted/queued/correlationID/agentID) with a documented limit note (the compared set covers every rendered field; the cache prunes to the window);src/v2/tui/picker.tsgains value-comparedstablePickerItems(previous reference wins when id/label/selected compare equal in order) wired into thepickerItemsmemo; attachment rows need no equivalent (attachments are refused visibly and never queued — documented at the helper). Mount-once stagger guard is theupdateStaggerseen-set from finding 4, bounded by the transcript window with pruning (documented limit: a burst larger than the window re-staggers only ids it has never seen). - Finding 6 (dead code/naming): removed
STREAMING_FRAMES/streamingFrame(no production callers;motionActivestays — it is the live predicate insideshouldRunMotion);footerLine1drops its unusedmissingSkillsparameter (new{ focus, hasWork }shape;footerLine2keeps the warning precedence);skillFilterEdit/SkillFilterKey/SkillFilterEditrenamed to the sharedfilterEdit/FilterKey/FilterEdit(both skill and memory filter paths, no behavior change); removed the duplicated peer-assignmentlive_assert_count; DM-vs-group cap asymmetry resolved with one documented shared bound (TRANSCRIPT_WINDOW = 40, used at all three slice sites — group window, DM window, DM merge). - Finding 7 (spec-reconciled wording): the
workrows.tsattribution comment now cites the clarified standard→ Assigned to $AgentName: <title>; rendered lines already matched it byte-for-byte, so no rendered line changed. Dated verification history below intentionally keeps its contemporary wording and was not rewritten.
What each check proves
- Unit (9 new): queued-removable matrix (user×agent true; group/agent/system false) + group header never renders
[x remove]; all four interim labels + row override + status preserved without pending; needs-you row shape/marker/suffix + derivation from interrupted vs pending vs resolved approval; stagger budget math (0/40/…/200 cap, negatives, reduced-motion short-circuit) + mount-once/cap/poll/tick/prune/idle/reduced semantics; picker identity kept on equal values and refreshed on any value/order/length change. All 290 pre-existing tests green. - Fake live (exit 0): group surface never offers
[x remove]; no stuckStopping…/Cancelling…/Retrying…/Restarting…after terminal; noneeds youwithout blocked work; transcript order alpha-before-beta with every heading/count exactly-once; the ≥2-frame working-frame observation still asserted and passing; reduced-motion frame-0/static run, idle quiescence, DM verbatim + queued removal, peer assignment → asking → answer, and all slice A–E asserts unchanged. - Real live + empty live: exit 0 unmodified.
Acceptance mapping (findings 1–7)
- Queued Remove works on DM surfaces (
xremoves the oldest parked line, proven live); group rows cannot be queued by construction and no longer offer the affordance (unit + live absence) — no dead affordance remains. - Cancelling/stopping/retrying/restarting rows show the interim label until terminal; cleanup is
finally-guaranteed on completion and error (unit + live absence, wiring review). - User-blocked work shows the distinct
◆ … needs you · v to resolvemarker in the warn tone, derived from interrupted/pending-approval state;vresolves via the existing global inbox (unit + wiring; live absence when unblocked). - First-paint stagger capped at ~6 lines × 40 ms, mount-only, never per chunk/poll; reduced-motion static; scroll and finalization unaffected (unit + deterministic live order/count asserts; timing-sensitive animation assert untouched).
- Primitives memoization kept, picker rows value-compared, mount-once stagger guard added; limits documented at each helper.
- Legacy motion exports removed, footer param dropped, filter helper renamed, duplicated assert removed, one shared transcript bound — no behavior change beyond the intended affordance/label fixes.
- Comments/docs aligned to
→ Assigned to $AgentName: <title>; rendered lines unchanged.
UI/UX skills applied (terminal translations)
clarify (needs-you/interim copy names the state and the key; shared filterEdit naming), typeset (hierarchy by tone + position only; faint entrance, warn needs-you/interim), layout (transcript > activity > WORK order and spacing scale untouched), animate (one-tick capped entrance, spinners unchanged, nothing per chunk), optimize (value-compared rows, mount-once guard, pruned caches), polish (git diff --check clean, no suppressions, duplicated assert removed).
Remaining uncertainty
- Interim labels are unit-pinned plus
finally-wiring review; the fake catalog's resolve RPC round-trip is milliseconds, so a live presence capture ofStopping…/Cancelling…would be racy by construction — live asserts the deterministic cleanup (absence after terminal) instead. Same standing limitation as the pre-existingResolving…row. - The needs-you
◆marker is unit-covered (all four derivation arms) but has no live fixture: the fake provider never produces interrupted or approval-gated work, and manufacturing one would change coordination semantics, which this pass forbids. Only the unblocked absence is live-asserted. - Stagger entrance is color-level (faint→full) and
capture-pane -p -Jstrips colors, so live proves the observable contract (order preserved, exactly-once rows, stable anchors, idle quiescence) rather than the dim frame itself; the frame math and reduced-motion short-circuit are unit-pinned. No live macOS run; glyphs verified through tmux pane capture on Linux only. motionActivewas kept deliberately: it is the live predicate consumed byshouldRunMotion(a production caller), so the "if truly unused" condition does not hold for it. Only the caller-free aliases were removed.- Dated verification sections below this one keep their contemporary
→ Assigned to $Nameshorthand as history; only live code comments were aligned to the clarified standard.
CORRECTION to the v5.4 review-fix verification above — TDZ crash (ticket #58) — 2026-11-02
Worked in /tmp/opencode/ensemble-worktrees/54-tdz-fix on branch feat/v54-tdz-fix, based on integration/ensemble-v5.3 at bc0fecf (the review-fix merge). The command-results table in the v5.4 review-fix section immediately above is wrong for all three live rows: the merged review-fix tree cannot open the Ensemble TUI route at all, so none of test:live, test:live:fake, or test:live:fake:empty could have passed on it. The table is left intact as history; this section supersedes its live rows.
The crash (deterministic, reproduced by the merger)
Opening the Ensemble TUI route crashed with:
ensemble.v2.tui crashed in route: Cannot access 'motionOn' before initializationRoot cause
src/v2/tui/app.tsx: the staggeringIds memo (then line 367) called motionOn(), but const motionOn = … was declared ~840 lines below (then line 1210). Solid evaluates createMemo callbacks eagerly at creation, so the call hit the const temporal dead zone during component setup. tsc cannot catch this (use-inside-a-callback before the const declaration is legal TypeScript), and no unit test mounts the app component, so npm run check and npm test stayed green while every live scenario failed at page open.
Audit of the other v5.4 memos
All 34 createMemo call sites in app.tsx were audited for the same eager-evaluation-before-const pattern. No other genuine hazard exists: pendingApprovalIDs reads pendingApprovals declared above it; no memo reads liveMessages, liveActivity, or motionFrame before their declarations; the late pickerItems memo reads only the picker signal and imported helpers. The audit is pinned mechanically (see below), not just by review.
The fix
- Moved the
staggeringIdsmemo below themotionOndefinition (no behavior change — same inputs, same memo), with a comment documenting the ordering requirement. - Added
src/v2/tui/app-order.test.ts: a static declaration-order guard asserting every component-local identifier called inside acreateMemobody is declared above that memo. Verified both ways — it fails on the pre-fix tree with exactlymemo 'staggeringIds' (line 367) calls 'motionOn' declared below it (line 1210)and passes post-fix. - Full app/route mount in the unit harness remains infeasible (the component requires the live OpenTUI plugin host: renderer, router, keymap, storage). The static guard plus the three live scenarios (which all open the real Ensemble page) are the coverage; no comment-only mitigation was used.
Actual command results (this branch, post-fix)
export PATH=/home/tristan/.nvm/versions/node/v26.12.1/bin:/home/tristan/.local/bin:$PATH| Exact command | Result |
|---|---|
npm ci | Exit 0 (run once; node_modules was absent). |
npm run check | Exit 0; strict no-emit TypeScript check. |
npm test | Exit 0; 300 passed, 0 failed, 0 skipped (299 pre-existing + 1 new app-order guard). |
npm run test:live | Exit 0; live-check passed; no skips; Ensemble page opens. |
npm run test:live:fake | Exit 0; fake live-check passed including the ≥2-frame working-animation observation (unweakened); no skips; Ensemble page opens. |
npm run test:live:fake:empty | Exit 0; fake empty-catalog live-check passed; no skips; Ensemble page opens. |
git diff --check | Clean. |
Remaining uncertainty
- The guard is static (declaration order), not a runtime mount: a future memo could still reference a later
constthrough an indirection this regex scan misses (e.g. a string-indirect lookup), but any direct call — the entire shape of this bug class as it occurred — fails loudly. - No live macOS run; glyphs verified through tmux pane capture on Linux only.