tooling-policy

harness-config registry

Policy registry for model selection, reasoning effort, local LLM discovery, and safe retrieval roots consumed by sync and platform adapters.

Overview

config/tooling-policy.json provides a single place for harness-specific model defaults (model id, reasoning_effort, personality/effort, text_verbosity) and local LLM provider profiles (lmstudio shape for codex/opencode). It also lists docs_domains for retrieval allow-listing and trusted_roots for path safety.

Used by sync projections (e.g. OpenCode model blocks, Grok, Codex), wagents platform modules, and any tooling that needs consistent model policy without hard-coding.

What It Controls

• model_defaults: keyed by harness short name (codex, copilot, gemini, opencode, grok); each declares default model, reasoning params, and personality/verbosity hints.
• local_llm_providers: profiles such as lmstudio with base_url_env, aliases, default_model, and per-target harness config (wire_api, provider_id, npm package, models map).
• docs_domains: safe domains for doc lookup / fetch tools (openai, github, anthropic, mcp spec).
• trusted_roots: array of repo paths considered safe for certain operations (currently just ${REPO_ROOT}).

Harness Surfaces

• OpenCode: provider_model, reasoning_effort, reasoning_summary, text_verbosity, model blocks in generated opencode.json / ~/.config/opencode/opencode.json
• Codex: model + reasoning_effort in ~/.codex/config.toml merges
• Grok: default + web_search model routing in .grok/config.toml
• Copilot / Gemini: effort_level and model guidance
• Local LM Studio: injected only when LM_STUDIO_API_BASE (or alias) present; used by both Codex and OpenCode adapters.

Changes here are canonical; sync or platform renderers project the values.

Key Fields

Field	Description
model_defaults.<harness>	Object with model, reasoning_effort, personality, effort_level, provider_model, text_verbosity, etc.
local_llm_providers.lmstudio	Discovery + wire config for local OpenAI-compatible server
docs_domains	Retrieval safety list
trusted_roots	Filesystem roots treated as first-party/trusted

AGENTS.md Sections

• §2.5 OpenCode Project Plugins: details model defaults (“repo and live OpenCode config set root model: "openai/gpt-5.5" … reasoning xhigh”) that are sourced from this policy.
• §2.7 Bucket B: “tool policy” listed among harness registries that remain canonical under config/.
• Grok and Codex platform notes reference config tomls that receive values from here.

Related

• config/grok-config.toml, generated ~/.grok/config.toml
• opencode.json (repo) + home merge
• wagents/platforms/*.py adapters
• Local provider env: config/grok-env.sh, source config/grok-env.sh

AGENTS.md §2.5 OpenCode model and plugin defaults driven from policy.

AGENTS.md §2.7 Classification of tool policy as Bucket B canonical config.

Grok doctor wagents grok doctor surfaces effective policy + env.

Other harness configs sync-manifest, mcp-registry, tooling-policy.

Full policy (config/tooling-policy.json)

config/tooling-policy.json

{
  "version": 1,
  "instruction_source": "${REPO_ROOT}/instructions/global.md",
  "model_defaults": {
    "codex": {
      "model": "gpt-5.5",
      "reasoning_effort": "xhigh",
      "personality": "pragmatic"
    },
    "copilot": {
      "model": "gpt-5.4",
      "effort_level": "high",
      "continue_on_auto_mode": false
    },
    "gemini": {
      "model": "gpt-5.4",
      "effort_level": "xhigh"
    },
    "opencode": {
      "provider_model": "openai/gpt-5.5",
      "manage_model_options": true,
      "reasoning_effort": "xhigh",
      "reasoning_summary": "auto",
      "text_verbosity": "low"
    },
    "grok": {
      "default": "grok-composer-2.5-fast",
      "web_search": "grok-4.20-multi-agent"
    }
  },
  "local_llm_providers": {
    "lmstudio": {
      "name": "LM Studio (local)",
      "base_url_env": "LM_STUDIO_API_BASE",
      "base_url_env_aliases": [
        "LMSTUDIO_API_BASE"
      ],
      "default_base_url": "http://127.0.0.1:1234/v1",
      "default_model": "local-model",
      "codex": {
        "provider_id": "local-lmstudio",
        "profile": "local-lmstudio",
        "wire_api": "responses"
      },
      "opencode": {
        "npm": "@ai-sdk/openai-compatible",
        "models": {
          "local-model": {
            "name": "LM Studio Local Model"
          }
        }
      }
    }
  },
  "docs_domains": [
    "developers.openai.com",
    "docs.github.com",
    "docs.anthropic.com",
    "modelcontextprotocol.io"
  ],
  "trusted_roots": [
    "${REPO_ROOT}"
  ]
}

---

View source on GitHub