Summary

Phase 2 Batch C subset of React UI v2.0 — the data composition layer that turns the primitives from Batch A/B into hooks that components can consume directly. After this lands, only Tasks 28 (useApprovalsQueue) and 29 (useAgentDefinitions) remain in Phase 2; both are independent and dispatch in parallel as a final mini-batch.

• Task 26 — useSessionFull(sid) (web/src/state/useSessionFull.ts) — composes useReducer(sessionReducer) + apiFetch('/sessions/{sid}/full') for bootstrap + useEventSource('/sessions/{sid}/events') for live deltas. Exposes { state, isLoading, error, refresh }. SSE gated on bootstrapped flag so no event lands before initial state. 3 tests.
• Task 27 — useSessionList() (web/src/state/useSessionList.ts) — fetch /sessions + subscribe /sessions/recent/events for session.created/session.status_changed/session.agent_running deltas. SessionSummary shape kept local to the hook (not promoted to types.ts until Phase 3 needs it). 3 tests.
• Task 30 — useUiHints() (web/src/state/useUiHints.ts) — @tanstack/react-query cache-once for /config/ui-hints. staleTime: Infinity, gcTime: Infinity since UiConfig is loaded at app startup and immutable. 2 tests.
• Task 31 — selectedRef (web/src/state/selectedRef.tsx) — React Context selection model with { kind: 'agent'|'tool_call'|'message'|null, id?: string }. Two-context split (value + setter) so components that only call useSetSelected() don't re-render on selection change. Throws when consumed outside Provider. 4 tests.

Validation

• cd web && npm install → 0 vulnerabilities
• npm run typecheck → exit 0
• npx vitest run → 58 passed / 0 failed (Phase 2 Batch B's 46 + 12 new)
• npm run build → clean

Nothing touches src/runtime/ — no dist/app.py regen.

Design notes

• Hooks compose existing primitives (apiFetch, useEventSource, sessionReducer) — no logic duplication.
• useSessionList uses setState rather than the reducer because cross-session deltas don't need vm_seq watermarking (each session has its own seq scope).
• selectedRef deliberately uses Context rather than Zustand to avoid the dep. The two-context split (SelectedRefContext + SetSelectedRefContext) keeps re-renders narrow when only the setter is consumed.
• useUiHints doesn't share the _helpers/render.tsx test wrapper — it inlines its own QueryClient with gcTime: 0 for test isolation. Intentional.

Out of scope (next batch)

• Task 28 useApprovalsQueue() — derived from useSessionList(), filters pending-approval tool calls.
• Task 29 useAgentDefinitions() — react-query cache-once for /api/v1/agents.

Both will land in a final Phase 2 Batch D (parallel, ~10 min total) before Phase 3 (Shell components) starts.

Test plan

• Vitest: 58/58 pass
• Typecheck: exit 0
• Build: clean
• Pytest backend stays green (no backend changes)
• CI: lint + type + test + sonar + bundle + skill-lint
• CI: SonarCloud quality gate

🤖 Generated with Claude Code