Skip to content

Pane interaction verification: select/interact shell, history, activity, and approvals (#186) ​

Status: verified locally on Linux/tmux; unit suite, full fake suite, size matrix, multi-location, empty-catalog fake suite, and latency probe pass; the real-host run is a post-merge step because the live host still runs main Date: 2026-11-09 Issue: #186 Spec: #178 Tickets: #180, #181, #183, #184, #185 ADRs: 0006 Base: integration/workspace-interaction at 3c65604 (includes #180/#181/#183/#184/#185/#187, #188, and #189) Branch: verify/pane-interaction Worktree: /home/tristan/dev/ensemble-186 Final merge commit: a33059a (verified tree)

This record is the release evidence for the pane-interaction delivery:

  • One select/interact state machine owns every move: h/l/j/k choose a pane, Enter enters it, Esc returns to selection, chat Enter is transcript navigation, i is composer entry, and h from chat reaches Agents first.
  • Every pane uses the same focused white / unfocused grey contract, including the separate AGENTS (n) and WORKSPACES (n) sections, chat, Work, and approvals. ANSI capture proves the selected heading/border (38;5;255) against the subdued border (38;5;238), muted heading (38;5;140), and faint heading (38;5;8).
  • Space b l / Space b r are the only column switches, pane visibility is durable across a TUI restart, closing a focused column returns focus to chat, reopening selects the column without entering content, and F3 is retired.
  • HISTORY partitions tasks, so one group can appear in ACTIVE and HISTORY with disjoint cards. A never-seen newest group expands, a former auto-expanded newest rolls up, a collapsed group promoted by a later finish stays collapsed with a fresh marker, older groups are compact rollups, Enter expands, x dismisses, D clears the view, and h/l cross page edges. HISTORY cards order by finish, newest first, across task and native subagent cards.
  • Chat keeps one aggregate activity line (Asking $Name…, Working…, and the +N more overflow) with the Work pane open; no per-agent row returns.
  • Approvals are only the right column's small summary and full-height inbox. v/Enter open full, Esc returns to split, the left footer and chat modal are gone, and V is retired. Full approvals also works in the 80×24 drawer.
  • Verification found and fixed four focus/scroll defects with regression tests: entering Workspaces left the rail cursor on an agent, closing Settings dropped a chat interaction back to selection, the host leader layer fired top-level l before a pending Space b prefix resolved (and the which-key popup clipped its body), and a former auto-expanded HISTORY group never rolled up.

The full fake suite ran on commit 00ed668, the pane-interaction content immediately before the final 3c65604 (#189) merge. npm run verify, the size matrix, multi-location, latency, and the empty-catalog suite ran on the merged tree a33059a; #189 touches service/session continuity, not the pane shell.

Environment ​

ItemValue
OSDebian GNU/Linux 13 (trixie), Linux 6.12.115+deb13.1-amd64 x86_64
Terminal hosttmux 3.5a
Nodev26.12.1
npm12.2.1
OpenCode (fake harness)opencode v2.2.6
Live service hostmain; not restarted or reconfigured
Worktree plugin/home/tristan/dev/ensemble-186 at the verification commit
macOS/HerdrNot available; not run
Hosted CINot run

Commands and results ​

CommandResult
npm run verifyPass — strict type check; 1530 tests passed, 0 failed, 0 skipped, 0 todo, 47.7 s.
ENSEMBLE_LIVE_REQUIRE=1 npm run test:live:fake (ENSEMBLE_LIVE_ARTIFACTS=/tmp/opencode/artifacts-186-fake)Pass on 00ed668 — full fake scenario including the #180/#181/#183/#184/#185 pane-interaction checks, doctor, and log summary. 53 live: lines; artifacts under /tmp/opencode/artifacts-186-fake.
ENSEMBLE_LIVE_REQUIRE=1 npm run test:live:sizesPass — 80×24, 100×30, 120×40, and 160×50, NO_COLOR=1, the CPU/load probe, pane heading/border colours, drawer durability/F3 removal, drawer full-approvals geometry, and page-edge j/k/h/l. Console copy under /tmp/opencode/artifacts-186-sizes/console.log.
ENSEMBLE_LIVE_REQUIRE=1 npm run test:live:multilocationPass — 100/100 jobs, 137/137 task rows succeeded, one service.start, 0 task.failed, 0 failure records, 1 warn-level feed.gap, the fake ask_person tool probe approved and replied, and doctor Mismatches: none. Artifacts under /tmp/opencode/artifacts-186-multi.
npm run test:live:fake:emptyPass — empty model catalog, agent wizard empty-state, init routing, and solo guard. Console copy under /tmp/opencode/artifacts-186-empty/console.log.
npm run test:live:latencyPass (measurement only) — rail j/k min 9 ms, p50 9 ms, p95 12 ms, max 12 ms (n=10); delegation card min 190 ms, p50 1940 ms, p95 1956 ms, max 1956 ms (n=5). Console copy under /tmp/opencode/artifacts-186-latency/console.log.
Real host-catalog matrix (npm run test:live / test:live:all)POST-MERGE — the live host runs main; the host was not restarted or reconfigured
git diff --checkPASS
Prototype source checkPASS — no prototype files in the tree or diff

Focus-state matrix ​

The pure matrix is pinned by src/v2/tui/panefocus.test.ts; the live checks exercise every visible combination at 160×80 and the 80×24 drawer.

PaneSelectInteractClosed
Agentsh from chat; white heading/border; h/l columnEnter; row cursor; enter openSpace b l; focus returns to chat; no AGENTS (
Workspacesj from Agents; white WORKSPACES; Agents greyEnter; workspace row cursor; enter openclosed with the left column
Chatl from the left column; white title; enter enterEnter transcript navigation; j/k scrollchat is never closable
Composeri from chatINSERT │ COMPOSEREsc/Ctrl+G returns to chat
Workl from chat; h/l columnEnter; card cursor; j/k cardsSpace b r; focus returns to chat
Approvals (small)j from Work; APPROVALS 1 pending │ enter openEnter opens full inboxlast row resolved or right column closed
Approvals (full)Esc from full returns to the small selectionj/k move, Enter resolve, esc splitEsc closes to split; right-column close hides it

ANSI evidence ​

Captured with tmux capture-pane -e -p -J and decoded by the harness. The focused pane heading and border are white; unselected panes keep the border, muted, or faint tone.

StateHeading/borderCaptured SGR
Agents selectedAGENTS (n) + section borderwhite 38;5;255
Agents selectedWORKSPACES (n) + section borderfaint 38;5;8 / border grey 38;5;238
Chat selectedchat title + chat borderwhite 38;5;255
Chat selectedWork header + Work bordermuted 38;5;140 / grey 38;5;238
Work selectedWORK ACTIVE + right borderwhite 38;5;255
Work selectedchat title + chat bordermuted 38;5;140 / grey 38;5;238
Approvals selectedAPPROVALS 1 pending + right borderwhite 38;5;255
Approvals unfocusedsmall APPROVALS 1 pending + top bordermuted 38;5;140 / grey 38;5;238
Work cardsfocused card borderwhite 38;5;255; neighbours 38;5;238

Space b, F3, and durable visibility ​

CheckEvidence
Which-key submenuSpace b shows LEADER │ which-key │ Space b, Toggle the left pane, Toggle the right pane, and the status hints l left │ r right; Esc cancels.
Left toggleSpace b l closes the left column, chat reclaims the width, and focus recovers to chat; reopening selects Agents in select mode.
Right toggleSpace b r closes Work, focus recovers to chat; reopening selects Work in select mode.
F3 retiredWith Work closed, F3 leaves WORK ACTIVE absent and does not steal focus.
DurableClosing both columns survives a TUI restart; both remain closed and chat owns the narrowed shell; reopening selects each column in select mode.
80×24 drawerThe drawer starts open, Space b r closes/reopens it, F3 does nothing, and the closed drawer survives a restart.

HISTORY, activity, and approvals ​

AreaLive evidence
Task-level splitA two-task dispatch group keeps its live card in ACTIVE (1 group · 1 active · 1 done) while its cancelled card is in HISTORY (1 group · 1 active · 1 done).
Newest expandedWith two finished groups, the newest renders ▾ GROUP 1 · 1 done and its card; the older renders ▸ GROUP 2 · 1 done with no card.
Rollup/expandA former auto-expanded newest rolls up when a newer group arrives; G + Enter expands the older row and reveals its card.
Dismiss/clearx dismisses the focused group and persists ui.dismissedHistoryGroups; D asks, clears the view to WORK HISTORY │ none yet, and terminal task rows remain in the snapshot.
Page edgesAt 160×50, l on the last page-one card moves to page two and h on the page-two card returns to page one; j/k keep their existing page edges.
Activity summaryWith live work and the Work pane open, chat shows one Working… line and later one Asking $FakeAgent… line; no │ working… per-agent row is rendered.
Approvals small/fullA pending ask shows APPROVALS 1 pending │ v open under Work; Enter from the small selection opens APPROVALS 1 pending │ 1/1 full height, Work is hidden, Esc returns to split.
Retired surfacesThe rail slice contains no APPROVALS footer; V no longer opens a chat approvals modal; v opens the right-column inbox from chat.
Drawer approvalsAt 80×24, v opens the full-height inbox in the drawer (APPROVALS 0 pending), Work is covered, the heading is white, and Esc returns to split. Pending resolution is exercised at 160×80; pending-in-drawer is listed as a skip below with unit-seam layout evidence.

Workspace-scoped approvals. The inbox deliberately filters pending approvals to the active workspace (approvalsForWorkspace), so one workspace cannot see or resolve another workspace's decisions. Spec #178 did not ask for that boundary; it is a safety decision and is covered by the pure approvals tests.

Pending approvals in the 80×24 drawer (unit seam). The live size matrix does not hold a pending approval in the populated drawer fixture, so no live 80×24 pending pass is claimed. The width-independent layout is covered by src/v2/tui/workapprovals.test.ts: at workspacePaneWidths(80) (workWidth ≥ 30), split geometry keeps a positive Work body and a positive small-approval row budget, and full geometry gives workBodyHeight = 0 with a positive inbox row budget (approvalRows ≥ 2 at workPaneBodyHeight(24)). The 160×80 live checks cover pending split/full/open/resolve and the Esc return; only the pending-in-drawer combination is a unit-only seam.

Size matrix ​

npm run test:live:sizes restarted the isolated fake TUI at each approved size with a fresh five-group/six-task fixture and checked the shell before the fixture started (so the hold tasks stay active through every functional probe).

SizeResult
80×24Drawer starts open; Space b r closes/reopens; F3 does nothing; closed drawer survives restart; five compact rows, page edge to page 2/2, whitespace drawer full approvals, NO_COLOR untouched by this size.
100×30Three panes, compact Work rows, group labels, budget, focus, page indicator, exact right edge, white focused / grey unfocused pane headings and borders.
120×40Same compact contract at the middle breakpoint.
160×50Three panes, five groups/six tasks in two-column cards over two pages, hjkl, gg/G, j/k and h/l page edges, right edge, white/grey card and pane focus.
NO_COLOR=1120×40 restart keeps frame-0 glyphs and the state word; no animated frame appears.

The load probe recorded idle TUI CPU 6.7% and 16.7%, 21.1%, and 26.7% for one, two, and three working cards; those numbers are observations, not thresholds.

Latency measurement ​

ProbeSamples (ms)Summary
Agents rail cursor j/k9, 9, 9, 9, 9, 12, 9, 10, 9, 9min 9, p50 9, p95 12, max 12, mean 9.4 (n=10)
Delegation card appearance190, 1956, 1953, 1931, 1940min 190, p50 1940, p95 1956, max 1956, mean 1594.0 (n=5)

The first card sample measured a warm path; later samples include the full patch/repaint path. These are measurements, not pass/fail thresholds.

Multi-location ​

npm run test:live:multilocation kept one service (serviceStarts: 1) across the measured run: 100/100 jobs, 137/137 task rows succeeded on attempt 1, 0 task.failed, 0 failure records, and one warn-level feed.gap. The fake ask_person tool probe created a real pending approval, resolved it, and the linked task ended replied/succeeded. The dispatch-budget relay and doctor finished with Mismatches: none.

Product fixes found by verification ​

DefectFixRegression test
Entering Workspaces left the row cursor on an agent, so the next Enter opened a DM.Mark the section cursor move as deliberate in ensureRailSectionCursor.app-wiring: a rail pane entry claims the row cursor.
Closing Settings from chat interaction returned to pane selection.Restore the prior pane select/interact mode.app-wiring: settings restore keeps the mode.
A pending Space b prefix still let the host leader layer run top-level l (limits).Disable the host leader layer while a prefix is pending; the raw key machine resolves the submenu.app-wiring: pending prefix yields the host layer.
The which-key popup allocated two border rows but not the header, clipping its body.Height is lines + 3.app-wiring: popup row budget.
A former auto-expanded newest HISTORY group stayed expanded when a newer group arrived.Track explicit user expansion separately; only manual choices persist, otherwise the newest group owns the default expansion.workpane.test: former auto-expanded newest rolls up; user-expanded older group stays open.

Post-merge deterministic rerun ​

The first CI run on eb3ebd3 failed in the fake suite after the #189 session-continuity merge: the reused FakeAgent session queued the rejection ask behind the previous compaction, and the fake provider's short reply did not match the host's structured compaction template, so compaction retried and blocked later probes. The harness now waits for the queued rejection task, and the fake provider returns the host's required summary template. The post-merge run ENSEMBLE_LIVE_REQUIRE=1 npm run test:live:fake:all passed: fake 457s, fake-empty 9s, 2 passed, 0 failed, 0 skipped. No product code changed for this rerun; the deterministic fake + empty suite is the post-merge gate.

Skips ​

  • Pending approvals in the 80×24 drawer are not automated in the size matrix: the populated size fixture queues the ask behind its active tasks. The width-independent pending split/full/resolve path passes at 160×80, and the 80×24 drawer proves the same full-height surface, white heading, Work-covered layout, and split return on the empty inbox. The unit seam adds positive split/full row budgets at the drawer width (see the unit-seam note above); no live 80×24 pending pass is claimed. Manual evidence is the /tmp/opencode/artifacts-186-sizes/console.log drawer section.
  • Workspace-scoped approvals are intentional, not a skipped cross-workspace check: another workspace's pending rows are filtered out by design and the boundary is covered by approvalsForWorkspace tests. Cross-workspace leakage would be a defect, not a missing capability.
  • The real host-catalog matrix is post-merge: the user's live host runs main and was not restarted or reconfigured. This is an environment limit, not a skipped assertion.
  • macOS/Herdr behavior is unavailable on this Linux/tmux host; the existing platform checklist remains manual.
  • The harness records transient states that the fake provider cannot hold deterministically (for example the single resolving frame) as unit-seam evidence; the live checks cover the durable before/after states.

Released under the MIT License.