Suggest-next: recommend experiments from accumulated knowledge

[susiejojo]

Problem

After #271 and #273, we can extract knowledge from campaigns and accumulate it in a cross-campaign registry. But there is no way to use that accumulated knowledge when planning the next campaign.

A user starting a new campaign still has to manually:

1. Remember which campaigns ran before
2. Open individual campaign files to check what was tried
3. Figure out which dead-ends to avoid
4. Identify which frontiers or interactions are worth pursuing
5. Decide how to frame the research question

This is exactly the retrieval problem the registry was built to solve.

Desired Outcome

Two deliverables:

1. scripts/retrieve_wiki_context.py

A Python script that builds an entity-scoped knowledge subgraph from the registry and per-campaign files. Given a set of campaigns, entities, and a research intent, it:

• Loads relevant principles, dead-ends, frontiers, and interactions from the specified campaigns
• Filters to entries related to the specified entities
• Returns structured context suitable for LLM consumption

python scripts/retrieve_wiki_context.py \
  -c campaign-1 campaign-2 \
  -e "Entity Name 1" "Entity Name 2" \
  -i "research intent description"

2. /suggest-next Claude skill

A skill that, given a research intent (natural language), retrieves relevant prior knowledge and recommends how to frame the next campaign.

Flow:

1. Load ~/.nous/wiki/registry.json, match intent to relevant campaigns and entities
2. Run retrieve_wiki_context.py to assemble focused context
3. Synthesize top 3 recommended campaign framings, each with:
   • A suggested research question
   • Which entities/concepts it builds on
   • Which frontiers/interactions it addresses
   • Which dead-ends to avoid
   • A composite score (Novelty, Foundation, Impact, Testability, Efficiency)
   • Predicted cost and model configuration

Output: Writes markdown to ~/.nous/wiki/suggestions/<date>-<slug>.md

Constraints

• Requires registry.json to exist (produced by /index-wiki from feat(wiki): add /index-wiki skill for cross-campaign registry #273)
• Reads per-campaign files from ~/.nous/wiki/campaigns/<name>/ (produced by /post-campaign from feat(wiki): post-campaign knowledge extraction and visualization #271)
• No modifications to existing orchestrator code
• No modifications to existing skills (post-campaign, index-wiki)

Context

See Discussion #269 for how this fits into the broader persistence strategy. This is PR 3 of 4 -- it provides the retrieval and recommendation engine that /visualize-registry (PR 4) will orchestrate per entity cluster.