feat(chat): realtime executor-subagent strip (executor-only, per-tab activity)

[miguelrisero]

Realtime executor-subagent strip (executor-only, per-tab activity) — DRAFT

Session-scoped strip at the top of the conversation pane that surfaces the executor's subagents live: active subagents render as tabs, each with its own working animation; finished ones collapse into a "N done" drawer. Pure render/projection over the task_create conversation rows the app already streams — no backend, no schema, no new ingestion, no new stream.

Owner decisions honored (these override the design spike where they conflict)

• Executor / UI-driven agents ONLY — not CLI. A real claude in a terminal already surfaces its own subagents, so the CLI/native-sidechain path is out of scope. This removes the spike's riskiest piece (un-hiding native sidechain records).
• Per-tab activity. Each subagent tab carries its own RunningDots animation, so you can see which subagent is busy vs done at a glance — not one session-wide state.

Behavior

• Leading Main chip; then one tab per ACTIVE subagent (created/pending_approval), each animating its own RunningDots.
• A subagent that finishes lingers ~5s (FINISHED_LINGER_MS = 5000) as a done-styled tab (check/x icon, no animation), then collapses into the trailing "N done ▾" button; the drawer lists finished subagents, each expandable to its result (reuses ChatSubagentEntry).
• Bounded: at most MAX_ACTIVE_TABS = 4 tabs; overflow folds into the same button relabeled "N more ▾".
• Mobile (useIsMobile, ≤767px): a single "N active · M done" summary line that expands to a floating overlay sheet (Radix Popover, does not resize the conversation) — no horizontal tab scroll. The global 7-tab nav is untouched. Existing inline ChatSubagentEntry rows remain the primary mobile reading surface.
• Session-scoped (mounted inside EntriesProvider in both the standard and VS Code conversation panes); renders nothing when there are no subagents.

Design

• All bounding / linger / overflow logic is in a pure partitionStrip() + selectSubagents() projection — both unit-tested (Node-env vitest, matching the repo's model-test convention; there's no jsdom/render-test setup). Timers + doneAt tracking live in a thin useSubagentStrip() hook with a single absolute-deadline timer.
• Reuses existing primitives only: RunningDots (the feat(cli-mode): make CLI/tmux agent 'working right now' legible in web UI (animated, is_running-driven) #44-standardized working animation), ChatSubagentEntry, AgentIcon, Radix Popover, CpuIcon; status-icon language mirrored from ChatSubagentEntry. No new spinner.
• a11y: decorative icons aria-hidden + sr-only status text; drawer entries announce status to screen readers; per-tab role="status" removed to avoid multiple live regions; reduced-motion respected. Honors the cn()-has-no-tailwind-merge footnote (§8): conditional classes are ternaries.

Spec-of-record

Brings the design doc (docs/superpowers/specs/2026-07-23-agent-activity-legibility-design.md, previously draft PR #43) into this branch and revises it to decided + as-built: facet 2 scoped to the executor path, §10 Q3→~5s linger, Q4→MAX_ACTIVE_TABS=4, Q5→per-tab yes, Q8→executor-only, Q9→generic "Subagent". The native/CLI path and facets 1/3 remain design-only. Please close #43 as superseded by this PR.

Verification

Gate (feature-only tree) GREEN: pnpm --filter @vibe/web-core run check, pnpm --filter @vibe/ui run check, pnpm run test (full suite: 27 files / 269 tests), pnpm run local-web:lint, pnpm run ui:lint, check-unused-i18n-keys.mjs, scripts/check-i18n.sh. New subagent-strip-model.test.ts + subagent-strip-partition.test.ts cover budget boundary, linger window (injected now), overflow labels, null subagent_type, phase mapping, ordering. No Rust/db/schema touched (clippy/cargo N/A). Frontend vitest is not in CI (tsconfig excludes *.test.ts) — run locally via pnpm run test.

Live visual verify (scratch Vite instance, never prod 4111; desktop + mobile, agent-browser/Chrome). Seeded EntriesProvider driving the REAL components (owner-approved seeded set in lieu of a live executor run). DOM assertions + screenshots:

• (a) desktop: Main chip + 2 tabs, each with its OWN RunningDots (2 animated dot-sets) — desktop-1-two-active.png
• (b) complete→linger: RunningDots 2→1 (one stops, the other keeps going), finished tab lingers — desktop-2-complete-linger.png
• (c) overflow: 6 active → exactly 4 tabs + "2 more ▾", no horizontal overflow — desktop-3-overflow.png
• (d) mobile 390px: "2 active · 0 done" summary, no tabs, no horizontal scroll; drawer opens as a floating Radix-portal overlay with conversation height 724px unchanged before/after (no squish) — mobile-1-summary.png, mobile-2-drawer-open.png
• (e) self-hide/session-scope: clear → strip returns null — desktop-4-cleared-selfhide.png
• global 7-tab nav unaffected (proven by diff — zero Navbar files changed).

Screenshots (viewable): /tmp/laneP-shots/*.png.

Review gauntlet + deviations (proportional, ship-it-light)

• Independent read-only Codex review (gpt-5.6-sol) + a two-persona frontend-systems / mobile-touch-UX review. Findings fixed: mobile drawer → floating overlay (was squishing the conversation); sr-only drawer status; removed per-tab role="status" live-region noise; exhaustive status-mapping guard; reduced-motion on the surfaced spinner.
• Deliberately NOT changed (owner decisions / brief): tabs use RunningDots not the entry's spinner (feat(cli-mode): make CLI/tmux agent 'working right now' legible in web UI (animated, is_running-driven) #44 mandate); the persistent "N done" affordance is the owner's explicit "small button to check inactive subagents" — the mobile "0 active · N done" line is left for the owner to decide at flip.
• Known follow-ups, deferred (non-blocking): one-time ~40px layout shift when the strip first appears (real fix = coordinate ConversationList's bottom-lock, not small); per-stream-tick selector for subagent-free sessions; 44px touch target under reduced mobile font-scale; faint dark-mode tab tint; desktop popover expand-state reset on reopen.
• CodeRabbit CLI unusable non-interactively → substituted the Codex sweep above (recorded deviation).

Risk: LOW–MEDIUM (frontend projection over existing data). Draft — boss verifies (screenshots above), flips, merges per the owner's standing "merge as you go".