deep-model-router
🔌PluginSungmin-Cho/deep-suite
Installation
/plugin marketplace add Sungmin-Cho/deep-suite/plugin install deep-model-router@claude-deep-suite
|Deterministic model/effort/review router — classifies delegated work and returns a RouteDecisionV1; large_context and latency_sensitive rebind worker_balanced to the Claude balanced seat; 1.5.0 adds fail-closed Grok envelope, artifact, and session-evidence contracts plus verified read-only reviewer seat profiles; 1.5.1 verifies grok-hosted grok.to_claude / grok.to_openai from Darwin.
Part of
Sungmin-Cho/deep-suite
More from this repository10
Evidence-Driven Development Protocol — a single-command Brainstorm → Research → Spec → Plan → Implement → Test → Integrate workflow with TDD enforcement, governed specification and methodology authority, adaptive risk-based review and verification, per-phase model selection (Claude Code / Codex), and a hard separation between planning and coding.
Independent Evaluator for AI coding agents. v2.8.1 unifies GROK_AUTHORIZED_MODEL so the router gate and the bridge argv assertion cannot disagree.
Cross-project semantic memory — harvests sibling-plugin artifacts into reusable memory cards (rule-based plus LLM distillation) and surfaces task-specific briefs, with opt-in cross-runtime hook capture and multi-pass redaction.
Cross-plugin harness diagnostics + suite telemetry — aggregates receipts, health reports, and harnessability metrics emitted by sibling deep-suite plugins into one dashboard, with optional OTLP export.
Exact worker contracts with bounded timeout fallback and journaled wiki mutation
Loop Engineering control plane over the deep-suite — discovers work, routes to sibling deep-* plugins as maker/checker episodes, keeps durable lock-safe loop state, and hands off to fresh sessions autonomously with proof-gated terminal states and proposal-only external actions.
Document gardening + authoring — v1.6.1 fixes the native Windows atomic-write identity false-positive (path-lstat dev:0); scan/garden/audit for Claude Code and Codex.
Goal condition compiler — evaluates a long-running task, reshapes it to fit, scouts prerequisites, and compiles a ready-to-paste native /goal condition for Claude Code and Codex, with multi-plugin synergy recipes.
Harness layer for Claude Code & Codex: plan-first development, independent review, durable long-running orchestration, and knowledge capture — ten composable plugins.
Autonomous Experimentation Protocol — a measured outer loop that proposes, runs, and selects code changes against a fitness metric, with lineage and session history, for Claude Code and Codex. v3.6.3 parses Grok PreToolUse envelopes in the readonly guard so Grok file and shell tools are classified instead of denied as malformed.