feat: research director — periodic strategic redirection (opt-in)

[aktasbatuhan]

An optional meta-agent that periodically reads the evolving population and writes one strategic directive into the steering channel for the next interval — aimed at breaking the local plateaus the population loop gets stuck in (autocorrelation parked at ~C 1.52 across every roster/prompt/config; that's a strategy bottleneck, not a tunable knob).

Design

Grounded in how the autoresearch loop it's modeled on actually works — a single agent in a keep/discard loop whose quality comes from a good instruction skill + a structured memory log, not multi-turn deliberation:

• kaievolve/research_director.py — ResearchDirector: gathers the experiment log (current best, trajectory, top programs + the HMRD notes their authors left), sees its own previous directive and whether best-so-far improved (keep vs pivot), makes one LLM call, writes research_directive.md, and keeps memory across fires in research_log.md. Fully fail-safe — any error is swallowed so it can never break the evolution loop.
• skills/research-director/SKILL.md — the director's instructions, including the full KaiEvolve mechanism (island MAP-Elites, LLM-mutated programs, frozen evaluator) and how its directive reaches the agents (folded into the steering brief that rides atop every generation prompt).
• Wiring — fires on its own cadence (research_director_interval, decoupled from migration so it gets enough shots), and the directive is folded into the existing steering channel in _create_database_snapshot, so it reuses plumbing that already reaches every worker.
• Opt-in: prompt.research_director_enabled (default off) + research_director_interval.

Verification

• test_research_director.py (5): directive written + rendered into the steering block, skill+task+experiment-log present in the prompt, keep/pivot state advances across fires, and the fail-safe (a throwing DB yields None, not a crash).
• Full suite green (347 tests).

Single call per fire to start (faithful to the source's per-decision simplicity); multi-turn / web-search grounding / per-island targeting are natural follow-ups if the A/B shows uplift.

🤖 Generated with Claude Code