mirror of
https://github.com/CopilotKit/CopilotKit.git
synced 2026-09-14 16:26:20 +08:00
Alem Tuzlak
23a3b24a01
Adds real working Shared State (Read+Write) and Sub-Agents demos to 15 showcase integrations, mirroring the canonical langgraph-python and google-adk reference implementations. Fills rows previously empty on the showcase coverage dashboard. Packages: ag2, agno, claude-sdk-python, claude-sdk-typescript, crewai-crews, langgraph-fastapi, langgraph-typescript, langroid, llamaindex, mastra, ms-agent-dotnet, ms-agent-python, pydantic-ai, spring-ai, strands. (built-in-agent landed independently on main as PR #4321 — its variant is canonical; this PR no longer touches it.) Per-package deliverables: framework-native backend agents (preferences-injection middleware/callback + set_notes tool; supervisor + 3 sub-agents wired as tools with running -> completed /failed delegation log); frontend page.tsx + preferences-card.tsx / notes-card.tsx for SSRW and delegation-log.tsx for subagents — wired to useAgent({ updates: [OnStateChanged] }); manifest entries; runtime route registration + per-package agent server config; real QA scripts. Includes targeted hardening fixes from a 7-agent code-review loop: - Sub-agent failure paths now correctly emit status: "failed" (previously hardcoded "completed" or unreachable in mastra/strands/langgraph-fastapi/langgraph-typescript/ag2) - Parallel-tool-call delegation race fixed in langgraph-fastapi (Annotated[list, add]) and langgraph-typescript (concat reducer) - Silent data loss eliminated in claude-sdk-python/claude-sdk-typescript/crewai-crews — empty JSON.parse catches now log + emit error events - ms-agent-dotnet set_notes writes to per-thread slot via AsyncLocal - mastra working-memory writes are deterministic via src/mastra/tools/working-memory.ts helper - spring-ai tool-call envelope ids match supervisor's tc.id() and AG-UI event ordering reordered; CopyOnWriteArrayList for parallel-call safety - Stack trace + raw error message leaks scrubbed across 8+ Next.js routes — log server-side with errorId + return generic envelope - Sub-agent calls no longer block event loops in ag2 (asyncio.to_thread), langroid (llm_response_async), pydantic-ai (async run + async tools) - langroid lru_cache cross-request contamination dropped - Numerous smaller items: claude-sdk-python invalid model id, Callable annotation, /health endpoint exposed; crewai-crews supervisor no longer resets delegations every turn; pydantic-ai snapshot uses model_dump() CI fixes folded in: - crewai-crews test_forwarded_props: extend the stubbed ag_ui_crewai.endpoint module to expose add_crewai_flow_fastapi_endpoint and add stub agents.shared_state_read_write / agents.subagents modules - generate-catalog test: bump crewai-crews wired-cell expectation 28 -> 30; replace hardcoded total-wired count with an invariant (wired + stub + unshipped = 737) plus a lower-bound floor - oxfmt run on the qa/shared-state-read-write.md files in mastra + spring-ai Rebased onto latest main (post showcase/packages -> showcase/integrations rename + post built-in-agent landing). Original blitz history preserved at the blitz-pre-rebase-snapshot tag. Known follow-ups (deferred to follow-up PR): - agno sync sub_agent.run() blocks event loop (perf only) - ms-agent-python asyncio thread-fallback fragility - llamaindex initial-state coercion when UI clears state - Manifest highlight audit (langgraph-typescript headless-complete, langgraph-fastapi byoc-* missing route.ts highlights) - agno hitl-in-chat declared in demos but not features; duplicate /demos/hitl-in-chat route - langgraph-typescript server.mjs graphSpec only registers 3 graphs vs 23 in langgraph.json (pre-existing) - mastra hitl legacy demo missing from features list - claude-sdk-python agents/agent.py line 474 also has the legacy claude-opus-4-5 default - PARITY_NOTES vs manifest mismatches for hitl-in-app across spring-ai/agno/ag2 (pre-existing) - spring-ai a2ui-fixed-schema missing from generative_ui list
90 lines
3.4 KiB
Markdown
90 lines
3.4 KiB
Markdown
# QA: Sub-Agents — Spring AI
|
||
|
||
## Prerequisites
|
||
|
||
- Demo is deployed and accessible
|
||
- Agent backend is healthy (check `/api/health`)
|
||
- `OPENAI_API_KEY` is set on the Spring backend
|
||
|
||
## Architecture under test
|
||
|
||
- Spring's `SubagentsController` runs a per-request supervisor agent that
|
||
exposes three tools — `research_agent`, `writing_agent`, `critique_agent`
|
||
— each backed by its own `ChatClient` call with a dedicated system
|
||
prompt (research vs. writing vs. critique).
|
||
- Every tool invocation appends a `Delegation` entry
|
||
`{ id, sub_agent, task, status, result }` to `state.delegations` and
|
||
emits a `STATE_SNAPSHOT` event, so the UI's live "delegation log"
|
||
panel grows entry-by-entry while the supervisor works.
|
||
|
||
## Test Steps
|
||
|
||
### 1. Page load
|
||
|
||
- [ ] Navigate to `/demos/subagents`.
|
||
- [ ] Verify the delegation log card renders
|
||
(`data-testid="delegation-log"`).
|
||
- [ ] Verify it shows the empty-state message
|
||
("Ask the supervisor to complete a task...").
|
||
- [ ] Verify `data-testid="delegation-count"` reads "0 calls".
|
||
- [ ] Verify the chat panel is visible on the right with the
|
||
"Give the supervisor a task..." placeholder.
|
||
|
||
### 2. Single delegation
|
||
|
||
- [ ] Send: "Research the basics of magnesium supplementation. Just one
|
||
delegation, no writing or critique."
|
||
- [ ] While the supervisor is running, verify
|
||
`data-testid="supervisor-running"` is visible.
|
||
- [ ] Verify exactly one `data-testid="delegation-entry"` appears with
|
||
sub-agent label "Research" and status "completed".
|
||
- [ ] Verify `data-testid="delegation-count"` reads "1 calls".
|
||
|
||
### 3. Full pipeline
|
||
|
||
- [ ] Click the "Write a blog post" suggestion (or send the equivalent
|
||
message manually).
|
||
- [ ] Watch the delegation log grow in real time:
|
||
- [ ] First entry: Research (sub_agent = `research_agent`)
|
||
- [ ] Second entry: Writing (sub_agent = `writing_agent`) — should
|
||
contain a polished one-paragraph draft.
|
||
- [ ] Third entry: Critique (sub_agent = `critique_agent`) — should
|
||
contain 2–3 actionable critiques.
|
||
- [ ] Verify each entry's "Task" line shows what the supervisor passed
|
||
to that sub-agent.
|
||
- [ ] Verify final supervisor message is concise (a short summary, not a
|
||
replay of all three sub-agent outputs).
|
||
|
||
### 4. Live snapshot ordering
|
||
|
||
- [ ] Verify the entries appear incrementally — the writing entry should
|
||
not show up at the same time as the research entry; if it does, the
|
||
backend is batching state snapshots.
|
||
|
||
### 5. Multi-turn
|
||
|
||
- [ ] After the pipeline completes, send: "Now do the same for the
|
||
benefits of cold exposure training."
|
||
- [ ] Verify three more entries are appended (count = 6).
|
||
- [ ] Verify older entries remain visible (delegation log is append-only
|
||
within a thread).
|
||
|
||
### 6. Error handling
|
||
|
||
- [ ] Send an empty message — UI should handle gracefully.
|
||
- [ ] If a sub-agent fails (transient OpenAI error), verify the entry's
|
||
`status` field reads "failed" and its result starts with
|
||
"Sub-agent failed:".
|
||
- [ ] No console errors during normal usage.
|
||
|
||
## Expected Results
|
||
|
||
- Page loads within 3 seconds.
|
||
- First delegation entry appears within 10 seconds of sending the
|
||
message; subsequent entries appear as each sub-agent ChatClient call
|
||
completes.
|
||
- "Supervisor running" indicator clears as soon as the run finishes.
|
||
- Final delegation count matches the number of sub-agent calls the
|
||
supervisor made (typically 1 or 3).
|
||
- No UI errors, no broken layout, no overlap between log and chat panes.
|