smartway_ce — frozen-config constraints

Additional HARD constraints for any architect / adas-subagent loop running on this graph. These are user-fixed config decisions that the meta-LLM must NOT modify; the architect's design space is the rest of the graph + nodesets.

planner_llm config is FROZEN

The planner_llm node (id planner_llm, type llmCall) has been deliberately configured by the user. Do NOT modify any of these fields:

• config.profile — set to "gpt-5-mini" (gpt-5 family)
• config.temperature — set to 1.0 (required by gpt-5 family)
• config.max_tokens — set to 6000
• config.response_format — set to "json_object"
• config.mode — set to "single_turn"
• config.template — set to "{plan}" (passthrough from assemble_prompt.prompt)

Allowed: modify planner_llm.config.ports, system_prompt, or the template only if you also rewire the upstream assemble_prompt node to produce the new prompt shape — and only when the design genuinely requires it. Otherwise leave planner_llm exactly as it stands.

Reason: the user is benchmarking gpt-5-mini specifically; swapping the planner LLM defeats the experiment.

Wrong changes — forbidden for this graph

A patch.intent that does any of these to the existing planner_llm node is forbidden:

1. Change planner_llm.config.profile (e.g. swap to gpt-4o).
2. Change planner_llm.config.temperature (e.g. set to 0.0).
3. Change planner_llm.config.max_tokens (e.g. set to 600).

Allowed: NEW llmCall nodes

You may add new llmCall nodes (critic, verifier, summarizer, etc.) with any profile / temperature combination that the cheat sheet permits — those are under the architect's control, not frozen.