Phase 2.1: SSE service layer
- Export apiUrl() from client.ts for shared URL construction
- Rewrite runs.ts: startRun, resumeRun, cancelRun, getRunStatus,
  connectStream with typed SSE events and Last-Event-ID support
- All request bodies wrap input in { input: ... } per server schema

Phase 2.2: RunSlice state machine
- Full event dispatch for all 7 GraphEvent types
- terminalReceived flag prevents onerror race after graph_completed
- Module-closure cleanup ref (not in Zustand state)

Phase 2.3: Run button + validation
- Run/Stop button in CanvasHeader with status-driven states
- validateGraph checks Start/End presence, orphans, LLM prompts
- Auto-save before run with error checking
- RunInputDialog for JSON input

Phase 2.4: Run panel + node highlighting
- Sheet extended with side="bottom" (openTransform/closedTransform)
- RunPanel with event timeline, auto-scroll, status banners
- Active node pulse via useNodeId() + useRunStore in BaseNodeShell

Phase 2.5: Reconnection + resume
- Exponential backoff (1s→2s→4s, max 3 attempts)
- Status polling fallback (completed/running/paused/error)
- reconnecting concurrency guard prevents parallel chains
- ResumeForm for human-in-the-loop paused runs
- Fix stale URL in gw-frontend skill

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Validate JSON input is a plain object (not string/array/null)
- Reset terminalReceived in resumeRun to prevent stale guard
- Add Number.isFinite guard on lastEventId query param
- Extract duplicate formatDuration to utils/format.ts

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The canvas saves schema_json without the top-level `id` field (it's
stored as a separate DB column). But validate_schema requires `id`.
Inject graph.id into the schema dict before calling build_graph and
validate_schema in both the run and validate endpoints.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The default state only has a "messages" field. The previous default
"result" caused a 422 from the server because the output_key must
reference an existing state field.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

LLM node: when input_map is empty (default), use messages from state
instead of building an empty HumanMessage. Gemini rejects empty content
with ValueError. Falls back to "Begin." if state messages are also empty.

Also include tracebacks in structured JSON logs — the JSONFormatter
previously discarded exc_info from logger.exception calls.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

gemini-1.5-flash was retired from the API. Update the default
provider to gemini and model to gemini-2.0-flash.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace raw node IDs with human-readable labels, add timestamps,
show provider/model for LLM nodes, and remove redundant graph_completed row.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Replace recursive reconnection with iterative loop + .catch() safety net
- Fix _handleStreamError type signature (void | Promise<void>)
- Add stale-state guards (terminalReceived/reconnecting) after each await
- Truncate node output display at 2000 chars to prevent DOM bloat
- Reset RunInputDialog state on reopen
- Clear ResumeForm input after submit
- Update gw-frontend skill with Phase 2 patterns and accurate RunSlice shape

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>