feat(orchestrator): decouple summarizer model from design

[kalantar]

Summary

• Per-iteration gate summaries previously read models.design, but the two phases use different transports: DESIGN goes through claude -p (which can carry the [1m] 1M-context tag) while gate summaries go through the LLM API via an OpenAI-compatible proxy that may not serve the same tag.
• Add a dedicated models.summarizer knob (default claude-opus-4-6), expose it in campaign.schema.yaml, and have run_iteration resolve the summarizer model via _model_for("summarizer") instead of "design".
• DESIGN and EXECUTE_ANALYZE behavior is unchanged. The existing summarizer role plumbing in dispatch.py, campaign.py, and llm_dispatch.py already expected this — the previous models.design lookup was a leftover from when the two phases shared a model.

Files changed

• orchestrator/defaults.yaml — new models.summarizer default.
• orchestrator/schemas/campaign.schema.yaml — schema entry so users can override per-campaign.
• orchestrator/iteration.py — one-line swap: _model_for("design") → _model_for("summarizer") for the summarizer's LLMDispatcher.

Ripples considered

No graphify-out/graph.json is present in the repo, so no automated ripple detection. Manual sweep across orchestrator/ and docs/ for models.summarizer / models.design / summarizer shows no other call sites that need updating.

Test plan

• On a campaign that does NOT set models.summarizer: confirm gate summaries run with claude-opus-4-6 (the new default).
• On a campaign that sets models.summarizer: <some-proxy-served-tag> in campaign.yaml: confirm gate summaries use that override and DESIGN keeps its own models.design.
• On a campaign whose models.design is a [1m] variant the proxy doesn't serve: confirm DESIGN still works and summaries are no longer broken by the proxy mismatch.

[mtoslalibu]

Review — Small fix + docs needed

Good fix for a real problem (proxy can't serve [1m] tags). Config/schema/defaults look good.

The issue

The llm_dispatcher at line 384 is shared — it's the fallback for DESIGN/EXECUTE when there's no CLI. This PR changes its model, which accidentally changes the fallback model too.

The fix — do what the reporter does

The reporter in campaign.py (line 94) already solves this correctly:

resolved = _resolve_model(campaign, "report", model)
dispatcher = LLMDispatcher(work_dir=work_dir, campaign=campaign, model=resolved)
dispatcher.dispatch("extractor", "report", ...)

It creates its own fresh LLMDispatcher with its own model. Nobody else shares it.

The summarizer should do the same. Instead of changing the model on the shared llm_dispatcher, create a dedicated instance:

# Line 384 — leave this alone (DESIGN/EXECUTE fallback)
llm_dispatcher = LLMDispatcher(work_dir=work_dir, campaign=campaign, model=_model_for("design"))

# New — dedicated instance for gate summaries, same pattern as reporter
summary_dispatcher = LLMDispatcher(work_dir=work_dir, campaign=campaign, model=_model_for("summarizer"))

Then pass summary_dispatcher to _generate_gate_summary() at lines 439 and 548.

Both reporter and summarizer use LLMDispatcher — the only difference is the reporter creates its own instance while the summarizer currently piggybacks on a shared one. This fix makes them consistent.

Documentation needed

Please add a section to the campaign configuration docs (or README) that explains the model configuration for users:

1. What models exist and their defaults:

Role	Default	Transport	Purpose
models.design	claude-opus-4-6	claude -p (CLI)	Plans experiments, reads codebase
models.execute_analyze	claude-sonnet-4-6	claude -p (CLI)	Implements and runs experiments
models.summarizer	claude-opus-4-6	API proxy	Generates per-iteration gate summaries
models.report	claude-sonnet-4-6	API proxy	Generates final campaign report

2. How to override in campaign.yaml:

models:
  design: "claude-opus-4-6[1m]"       # CLI — supports [1m] context tags
  execute_analyze: "claude-sonnet-4-6"
  summarizer: "claude-opus-4-6"        # API proxy — must be proxy-served
  report: "claude-sonnet-4-6"          # API proxy — must be proxy-served

3. The key distinction users need to understand:

• Agents (design, execute_analyze) use claude -p CLI — support [1m] tags, tool use, deep context
• LLMs (summarizer, reporter) use the OpenAI-compatible API proxy — must use tags your proxy actually serves. Don't use [1m] variants here.

This helps users understand why models are configured separately and prevents the exact proxy mismatch bug this PR fixes.

Summary

What	Status
Config + schema + defaults	Good
iteration.py wiring	Needs fix — create a separate summary_dispatcher
User-facing documentation	Needed — model table, override example, agent vs LLM explanation

With those two additions, good to merge.