What: Replace Dagre layout engine with ELK (Eclipse Layout Kernel) for
workflow graph visualization in the dashboard.
Why: Dagre lacks support for edge label placement and orthogonal edge
routing, causing condition labels to overlap with nodes. ELK provides
proper label positioning, better edge routing, and more compact layouts.
Changes:
- packages/workflow-dashboard/package.json: add elkjs dependency
- packages/workflow-dashboard/src/components/workflow-graph/use-layout.ts:
rewrite layout from Dagre to async ELK with layered algorithm,
orthogonal routing, reduced spacing for compactness
- packages/workflow-dashboard/src/components/workflow-graph/condition-edge.tsx:
use ELK-computed label positions, show all labels including FALLBACK,
switch to getSmoothStepPath for all edges
- packages/workflow-dashboard/src/components/workflow-graph/workflow-graph.tsx:
wrap in ReactFlowProvider, add fitView on async layout change,
key-based remount for layout stability
- packages/workflow-dashboard/src/components/workflow-list.tsx:
left-right layout (info left, graph right), fix toggleExpanded
React 18 batching bug, increase graph container height
Add ModeratorTable syntax, AdapterFn/AdapterBinding types, lazy init
pattern, bundle import restrictions, and descriptor requirements.
Knowledge from smoke test discoveries — these are the most common
mistakes when writing workflow bundles.
小橘 <xiaoju@shazhou.work>
Root cause: executeThread awaited gen.next() without racing against
the abort signal. When a workflow bundle awaited a long setTimeout
between yields, the engine could not respond to kill until the
Promise resolved — causing the kill test to flake when the thread
completed before kill arrived.
Fix: Promise.race gen.next() with an abort listener so kill takes
effect immediately, even mid-yield. Also move the bundle's delay
to after the first yield (between planner and coder) to ensure the
thread is killable while running.
Closes#209
- Add AgentSocket Durable Object (holds one WS per agent name)
- Add /ws/connect route with GATEWAY_SECRET auth
- Add ws-client.ts with auto-reconnect (exponential backoff 1s-30s)
- serve defaults to WS mode (no cloudflared needed)
- Keep --tunnel-url and --no-tunnel as fallback options
- Endpoints list merges KV heartbeat + DO WebSocket status
Testing: #211
- Add RoleFn<T>, AdapterFn, AdapterBinding types to workflow-protocol
- Mark AgentFn, AgentFnResult, AgentBinding as @deprecated
- Refactor createWorkflow to accept AdapterBinding instead of AgentBinding
- Adapter returns typed meta directly — no more extract call in workflow loop
- Add buildThreadInput (ThreadContext-based), keep buildAgentPrompt as deprecated wrapper
- Update template bundle-entries to wrap AgentFn as AdapterFn
- Update solve-issue tests to use AdapterFn directly
- cmdInitWorkspace now resolves full paths via resolve() instead of
requiring a single segment name
- mkdir uses recursive: true for nested paths (e.g. ./a/b/workflows)
- Setup interactive prompt retries on existing directory instead of exiting
- Update tests: nested paths are now valid, add accepts-nested-path test
- Show available models with numbered labels in multi-column layout
- User can pick by number or type model name directly
- Print selected model with arrow confirmation
- Add providers.yaml with 18 preset providers (international + China + local)
- Add preset-providers.ts to load and cache YAML presets
- Refactor interactive setup to show numbered provider list
- Only prompt for manual name/URL when choosing Custom
- YAML-driven: add new providers without code changes
Users can now type bare model names (e.g. 'qwen-plus') or paste
model IDs with vendor prefixes (e.g. 'MiniMax/MiniMax-M2.7') —
the provider prefix is normalised automatically.
Filter out speech/embed/image/video/audio/tts/asr/ocr/rerank models
from the /models listing. Display remaining models in a responsive
multi-column grid that adapts to terminal width.
Raw mode receives pasted text as a single chunk. Iterate
per-character so every char gets a '*'. Also fix backspace
to erase the visual '*' from the terminal.
1. Mask API key input with * characters (raw mode)
2. Fetch and list available models from provider /models endpoint
3. Workspace prompt: fill path directly (default skip), not y/n
4. Add .gitkeep to workflows/ in init workspace scaffold
- Add intro line explaining what's being configured
- Provider: explain it's a label for the LLM service
- Base URL: explain it's OpenAI-compatible, show examples
- API key: clarify it's for this provider
- Default model: show format with dynamic provider name
- Workspace prompt: clearer wording
xingyue
xiaoju
xiaomo