maintain-rpiv-pi-models

name	maintain-rpiv-pi-models
description	Maintain models used for rpiv-pi skills and agents
modified	2026-07-21T17:14:15+0900

maintain-rpiv-pi-models

This skill maintains the per-user model override file at chezmoi/dot_config/rpiv-pi/private_models.json (chezmoi source path; installs to ~/.config/rpiv-pi/private_models.json).

That file maps AI/LLM models to rpiv-pi agents, stages, skills, and presets, overriding rpiv-pi's built-in defaults. It is the single place to adjust which models are used by the agent framework's various components.

Current model policy

Prefer task-specific model assignment rather than using one model everywhere. Use expensive models only for high-intelligence architecture/design work, and use lighter or specialized models for routine coding tasks.

Current model-role mapping

Task type	Preferred model
Default	aihubmix-oc/deep-deepseek-v4-flash
Visual/frontend design	aihubmix-am/xiaomi-mimo-v2.5
Research / implementation / agent work	aihubmix-oc/deep-deepseek-v4-flash
Planning / code review / revise	aihubmix-am/cc-minimax-m3
Validation / slice verification	aihubmix-oc/coding-xiaomi-mimo-v2.5-pro
Blueprint	aihubmix-oc/deep-deepseek-v4-flash
Architecture / design	aihubmix-oc/qwen3.8-max-preview
Claim verification	aihubmix-oc/coding-xiaomi-mimo-v2.5-pro
Commit generation	aihubmix-oc/deep-deepseek-v4-flash (ship preset: thinking: minimal)

Low-hallucination models

Use aihubmix-am/cc-minimax-m3 and aihubmix-oc/coding-xiaomi-mimo-v2.5-pro for tasks requiring rigid adherence to instructions. Both have low hallucination rates, so they are good choices when the model must follow constraints, preserve facts, and avoid inventing details.

• aihubmix-am/cc-minimax-m3: planning, code review, and revise
• aihubmix-oc/coding-xiaomi-mimo-v2.5-pro: validation and slice verification

Fast-response model

Use openrouter/nex-agi/nex-n2-mini for things requiring rapid responses. It is the fastest model in the current configuration, so it is best suited for lightweight tasks such as discovery, exploration, claim verification, and quick turnaround where speed matters more than deep reasoning.

Current usage

openrouter/nex-agi/nex-n2-mini

Use for lightweight tasks:

• skills.discover
• skills.explore

aihubmix-oc/deep-deepseek-v4-flash

Use for default, research, implementation, blueprint, commit, and agent support:

• Default model
• stages.commit
• skills.commit
• stages.research
• stages.blueprint
• agents.codebase-analyzer
• agents.artifact-code-reviewer
• agents.artifact-coverage-reviewer
• agents.scope-tracer
• skills.create-handoff
• skills.resume-handoff
• skills.implement
• skills.blueprint
• skills.research
• skills.browser-tool
• skills.hunk-review
• presets.build.stages.commit
• presets.build.stages.implement
• presets.build.stages.research
• presets.build.stages.blueprint
• presets.ship.stages.commit (thinking: minimal)
• presets.ship.stages.implement
• presets.ship.stages.blueprint

aihubmix-oc/qwen3.8-max-preview

Use sparingly for expensive high-intelligence architecture/design work:

• stages.architecture-review
• stages.design
• skills.design

Do not use qwen3.8 for commits, routine research, planning, blueprint, validation, slice verification, claim verification, or code review unless there is a specific reason.

aihubmix-oc/coding-xiaomi-mimo-v2.5-pro

Use for validation, claim verification, and slice verification (all adversarial code-grounded tasks):

• agents.claim-verifier
• agents.slice-verifier
• stages.validate
• skills.validate
• presets.build.stages.validate
• presets.ship.stages.validate

aihubmix-am/cc-minimax-m3

Use for planning, code review, and revision:

• stages.code-review
• skills.code-review
• skills.plan
• presets.build.stages.code-review
• presets.build.stages.revise

aihubmix-am/xiaomi-mimo-v2.5

Use for visual/frontend design:

• skills.frontend-design

This is the non-pro Xiaomi MIMO model. It is the cheapest model in the current configuration with image input and visual understanding, so use it for any UI, UX, or visual-related work.

Maintenance rules

• Keep skills.frontend-design on aihubmix-am/xiaomi-mimo-v2.5 for visual work.
• Keep commit generation on aihubmix-oc/deep-deepseek-v4-flash with minimal or no thinking for speed. Use thinking: minimal for the ship preset commit stage.
• Keep qwen3.8 reserved for architecture and design work (not blueprint).
• Use aihubmix-oc/coding-xiaomi-mimo-v2.5-pro for validation and slice verification.
• Use aihubmix-am/cc-minimax-m3 for code review, planning, and revise.
• Use aihubmix-oc/deep-deepseek-v4-flash for research, implementation, handoffs, and agent support.