fix(c-generative-ui): use gpt-5 + minimal reasoning for planner LLM

[blove]

Summary

gpt-5-mini ignored the "EXACTLY ONE tool" directive added in PR #363 and kept calling all four data tools on every filter follow-up. The prompt rewrite was strict and explicit ("call EXACTLY ONE tool", "Do NOT call the other tools") but the model's default reasoning prefers thoroughness over literal directive-following.

Fix

Split the LLMs in dashboard_graph.py / graph.py:

• _llm (gpt-5-mini) — unchanged for generate_shell + respond. Cheap, good enough for prose + JSON-spec emission.
• _planner_llm (gpt-5, reasoning_effort="minimal") — bound to tools, used in plan_tools. gpt-5 follows directives more precisely; reasoning_effort="minimal" suppresses the "let me be thorough" deliberation that drives the fan-out.

Test plan

• Standalone smoke against the planner directly:

  llm = ChatOpenAI(model='gpt-5', reasoning_effort='minimal').bind_tools(ALL_TOOLS)
  await llm.ainvoke([sys, HumanMessage("Filter to cancelled flights only")])
  # → ['query_recent_disruptions']

• Chrome MCP end-to-end with backend running real LLM:
  • "Show me the dashboard" → 4 stat cards + 2 charts + 1 grid populate
  • "Filter to cancelled flights only" → backend AI tool_calls: ['query_recent_disruptions'] (one tool, not four), data grid updates to 3 cancelled rows
• CI
• Cost note: planner now uses gpt-5 per dashboard turn — one extra-effort call per follow-up. Worth it for the directive adherence; can revisit if cost becomes material.

Files

• cockpit/langgraph/streaming/python/src/dashboard_graph.py — add _planner_llm; swap _llm_with_tools source
• cockpit/chat/generative-ui/python/src/graph.py — same in standalone

🤖 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 16, 2026 5:54pm