AI MM

AI-assisted code review, planning, and stabilization through the host-level Rust llm-router service.

ai-mm owns review orchestration and prompts. It does not choose API models, construct vendor clients, hold provider credentials, calculate vendor pricing, or implement fallback cascades. Every LLM call goes to llm-router as a semantic stage: or profile: intent.

Architecture

ai review / ai plan / rig-assist
        |
        | stage:review | stage:audit | profile:local_only | ...
        v
LLMRouterProvider
        |
        | POST http://127.0.0.1:4000/v1/chat/completions
        v
Rust llm-router
        |
        | profile resolution, model/provider selection, retries,
        | fallbacks, request quirks, cost and provenance
        v
provider adapter selected by llm-router

Responses without the router provenance block are rejected. This prevents a proxy, misconfigured endpoint, or direct vendor service from silently bypassing routing policy.

Prerequisite

Install and run the router from ~/Development/llm-router:

command -v llm-router
llm-router --version
curl -fsS http://127.0.0.1:4000/health

Provider credentials and exact model policy belong in ~/.config/llm-router/, not in the ai-mm environment.

Install

./run install

This installs the ai command and its Python environment under ~/.local/.

Review

# Router-owned normal review
git diff | ai review --model stage:review

# Normal single review: fail-closed DeepSeek V4 Flash profile
git diff | ai review --model deepseek --focus verification

# Parallel semantic council seats; each seat is independently routed
git diff | ai review --model mm --focus review

# Explicit opt-in/offline local review; never an automatic fallback
git diff | ai review --model local

# Explicit router profile
git diff | ai review --model profile:kimi --reasoning-effort max

--model is retained as the CLI option name for compatibility, but its value is a router selector, not an API model ID. Supported forms are:

• stage:<intent> — let the router resolve the stage;
• profile:<name> — explicitly request a router-owned profile;
• deepseek — stable stage:audit selector; llm-router currently resolves it to fail-closed profile:deepseek_v4_flash_direct;
• local, kimi, and commercial — convenience profile selectors;
• mm, all, fast, local, and max — orchestration groups.

Exact API model names are rejected.

Review groups

Group	Router intents
mm	stage:review, stage:audit, stage:adversarial
all	mm plus stage:architect
fast	stage:review with a fast effort hint
local	profile:local_only
max	profile:kimi with a max effort hint

The group controls parallel review seats only. It never names providers or fallback models; the router remains authoritative for each seat.

Normal review is one deepseek seat after all edits are complete. Use mm, all, or an iterative loop only when the operator explicitly requests extra review. Local review is opt-in/offline tooling and must not be inserted into the critical path.

The historical lmstudio alias has been removed intentionally because stale automation used it to put unbounded local Qwen reviews on the critical path. Use local explicitly when an operator requests an offline review.

Focus

Use --focus to tune the review prompt:

• review
• verification
• security
• performance
• architecture
• testing
• general

Timeouts

--per-model-timeout is a client deadline for one complete router task. The default is 600 seconds so the router has time to execute its bounded cascade. Individual attempt timeouts remain router-owned.

--local-model-timeout is accepted only for backward compatibility and no longer changes routing or provider timeouts.

Planning

ai plan "Refactor billing" --model stage:planning --output-format json
ai plan "Refactor billing" --model mm --depth deep --rounds 3 --strict

The planning council uses semantic planning, architecture, and adversarial stages. Structured response validation and synthesis remain in ai-mm; execution routing remains in Rust.

Rig assistance

ai rig-assist --mode plan --model local --input packet.json
ai rig-assist --mode recovery --model deepseek --input packet.json

Rig assistance accepts only local and deepseek. Returned JSON remains schema-constrained and advisory-only.

Configuration and health

The only ai-mm connection setting is:

export LLM_ROUTER_BASE_URL="http://127.0.0.1:4000/v1"

The default already points there. Use either command to inspect the connection:

ai config
ai check-models

ai check-models calls the router health endpoint; it does not send a paid test completion.

Optional settings:

default_models:
  plan: stage:planning
  review: stage:review
review_per_model_timeout_seconds: 600
cache_ttl_hours: 24

Router provenance and cost

llm-router returns the resolved profile, provider, served model, escalation state, and cost source. ai-mm exposes that metadata on ReviewResult.metadata and records the router-reported cost. There is no second pricing table or cost estimate in this repository.

Development

./run test
./run lint

Architecture tests fail if feature code imports direct vendor providers, contains API model selection, or expands the config surface beyond the router connection.

Migration notes

Version 0.8 removes direct OpenAI, Anthropic, DeepSeek, Alibaba, Ollama, and LM Studio adapters from ai-mm. It also removes local provider fallback logic and exact model groups. Configure those providers and their cascades in llm-router instead.