fix(c-generative-ui): three polish followups (flat args, loop cap, chip text)

[blove]

Summary

Three small followups to PR #436 that polish the c-generative-ui demo.

1. render_spec flat args

gpt-5 reasoning='minimal' frequently emits render_spec args as {"elements": ..., "root": ...} instead of the wrapped {"spec": {"elements": ..., "root": ...}} the tool signature required. Pydantic rejected the first call; the LLM retried successfully; but the user saw an ugly Error invoking tool 'render_spec' with kwargs ... spec: Field required block in the chat.

Fix: flatten the tool signature to render_spec(elements: dict, root: str). Matches the spec's actual schema and matches what the LLM naturally emits. Tool body just wraps them back into {"elements": elements, "root": root} before serializing. Zero retries observed in the smoke run.

2. Loop-termination prompt tightening

The agent loop was running render_spec + the 4 data tools TWICE per first turn — wasted LLM calls. Existing prompt said "return without further tool calls" but gpt-5 ignored it.

Fix: add explicit "Critical: if render_spec and the data tools have already been called this turn, you are DONE. Return with no tool_calls. Do NOT call render_spec again. Do NOT re-call the data tools." Smoke now shows exactly 1 render_spec + 4 data tools per first turn.

3. Welcome chip text aligned with aviation theme

Chip said "Show me a Q3 sales dashboard with three metrics" but the backend is aviation-themed. When the user clicked the chip, the LLM saw "Q3 sales" + an aviation tool catalog, got confused, and wrote a fictional sales dashboard description in prose instead of calling tools.

Fix: chip values now match the backend — "Show me a dashboard of airline operations." and "Filter to only the cancelled flights." (Replaces the broken second chip "Create a contact form" which was never relevant to this demo.)

Files

• cockpit/chat/generative-ui/python/src/graph.py — flatten render_spec signature
• cockpit/langgraph/streaming/python/src/dashboard_graph.py — umbrella mirror
• cockpit/chat/generative-ui/python/prompts/dashboard.md — tighten loop instructions + update tool signature reference
• cockpit/langgraph/streaming/python/prompts/dashboard.md — umbrella mirror
• cockpit/chat/generative-ui/angular/src/app/generative-ui.component.ts — chip text

Test plan

• Real-LLM smoke against per-cap (3-turn): 1 render_spec call, 4 data tools, 0 Pydantic errors, clean conversational summary
• pnpm nx run cockpit-chat-generative-ui-python:build — green
• pnpm nx run cockpit-langgraph-streaming-python:build — green
• pnpm nx run cockpit-chat-generative-ui-angular:build — green
• CI

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cacheplane	Ready	Preview, Comment	May 18, 2026 8:17pm