Allow configuring default instructions or a default skill per provider model

[occasl]

Problem

I use a route skill that classifies each incoming request and escalates harder work to stronger models via T3's delegate_task. It pays off most on threads that start on the cheapest models:

• Claude Haiku (claudeAgent / claude-haiku-5-5)
• GPT-6 Luna (codex / gpt-6-luna)
• Gemini 3.8 Flash (antigravity / gemini-3.8-flash-*)

T3 Code has no way to attach default instructions or a skill to a specific model, so I type /route at the start of every message on those models.

What T3 already has

On main (b5b6c8c), T3 can already deliver instructions across providers, but none of it is user-configurable per model:

• packages/provider-core/src/server/orchestrationInstructions.ts defines T3_CODE_ORCHESTRATION_INSTRUCTIONS and injects instructions into threads for each provider when T3's MCP tools are present, via t3OrchestrationSystemPrompt, t3OrchestrationPromptForFirstRun, and t3AcpPromptWithInstructions. The text is fixed.
• packages/contracts/src/settings.ts has sourceControlWritingStyle.customInstructions, the only user-editable instructions setting. It applies to commit and PR writing only.
• packages/shared/src/projectSettings.ts has defaultModelSelection, which picks a project's default model but carries no instructions.

Proposal

A user setting for default instructions and/or a default skill, keyed by provider instance and model ID, with globs such as antigravity / gemini-3.8-flash-*.

• Scope: applies to every thread running on a matching model, including a thread that switches to that model mid-conversation.
• Delivery path: the same mechanism T3 uses for its orchestration instructions, so it reaches every provider, including Antigravity, which doesn't load rules files today.
• Per-turn injection: a reminder on each turn, not only at thread start. In my testing, Haiku ignored a one-time CLAUDE.md instruction on the first turn but followed a per-prompt hook reminder.
• Defaults and opt-out: empty by default, so nothing changes for existing users. Optionally, an escape prefix to skip it for one message.

Example configuration (the shape is illustrative; key names are made up):

{
  "modelInstructions": [
    {
      "provider": "claudeAgent",
      "model": "claude-haiku-5-5",
      "skill": "route"
    },
    {
      "provider": "codex",
      "model": "gpt-6-luna",
      "instructions": "Classify the request first and delegate complex tasks via delegate_task."
    },
    {
      "provider": "antigravity",
      "model": "gemini-3.8-flash-*",
      "instructions": "Classify the request first and delegate complex tasks via delegate_task."
    }
  ]
}

Current workaround

• Claude: a model-conditional rule in global CLAUDE.md ("if your model ID starts with claude-haiku-, invoke the route skill") loaded into context, but Haiku ignored it on the first turn. A Claude Code UserPromptSubmit hook that injects the reminder on every prompt works, but only for Claude threads.
• Codex: global AGENTS.md applies to every Codex model, so it can't target gpt-6-luna alone.
• Antigravity: threads load no AGENTS.md or GEMINI.md rules ([Bug]: Antigravity sessions load no AGENTS.md/GEMINI.md rules (global or workspace), unlike Claude and Codex #15283), so there's no file-based workaround.

Related

• Configurable model routing rules for Orchestrator subagent delegation #15182: configurable routing rules for which model delegate_task picks. This proposal is the other half: instructions that follow the model the thread itself runs on.
• GitHub Copilot-Style Custom Agents with Lead-Subagent Delegation, Dynamic Model Switching, and Project Defaults #17924 / PR feat(agents): add project custom agents #18225: project custom agents with a fixed model and instructions, chosen per thread. This setting follows the active model instead.
• [Bug]: Antigravity sessions load no AGENTS.md/GEMINI.md rules (global or workspace), unlike Claude and Codex #15283: Antigravity sessions don't load AGENTS.md or GEMINI.md rules.
• Run a configurable prompt or skill when settling a thread #19004: run a prompt or skill when a thread settles. That's an end-of-thread hook, not per-model, per-turn instructions.

Environment

• T3 Code 0.1.1-nightly.20261017.3238
• macOS arm64
• Providers: Claude, Codex, Antigravity

---

🤖 Generated with Claude Code