Sister CODEC & Reasoning Profiles

Per-sister model configuration and reasoning effort profiles.

Reasoning Effort Policy

All sisters default to high reasoning — DeepSeek V4 Flash handles it efficiently with minimal token overhead.

Sister Reasoning Rationale
Tai (default) high Primary builder, general operations
Rei high QA audit, code review, risk analysis
Nei high Quartermaster, digests, organization
Mai high Support ops, logistics, creative problem-solving

Note: xhigh was the predecessor to high (renamed June 2026). high is the current canonical value and does NOT carry the 9x hidden token multiplier that xhigh had. All sisters on high use the standard token budget.

Model Config

  • Primary: deepseek-v4-flash via OpenCode Go
  • Fallback: OpenRouter free tier models
  • Provider format: OpenAI chat-completions

Cost Guardrails

  • high reasoning is the global standard — DeepSeek Flash is efficient enough that this doesn’t drive cost
  • Cron jobs use free/cheapest models where possible
  • Non-essential crons paused when cost is a concern

Per-Sister System Prompts

Sister Tokens Role
Tai ~455 Primary, builder/operator
Rei ~195 QA, auditor
Nei ~216 Quartermaster
Mai ~320 Support, wildcard

Prompts stored in agent.personalities.<name> in each profile’s config.yaml.

next Home