Let a project append to, replace or turn off the injected orchestration instructions

[kkolodziej7]

What I'm building

A team orchestration layer that runs on top of T3 Code. It has roles (CTO, Tech Lead, engineers), an issue journal, and routing through its own MCP server. Every repo's AGENTS.md carries the team's rules for placing work, picking models and waiting on results.

Problem

T3 Code injects a fixed T3 Code orchestration block into every session that has the t3 MCP. Claude gets it in the system prompt and Codex gets it as the <t3_code_orchestration_instructions> prefix. In 0.1.5 the text comes from constants (orchestrationInstructions, t3OrchestrationSystemPrompt), and the only gate is hasT3Mcp. Nothing in settings can change it. The customizable instruction fields that exist (sourceControlWritingStyle.customInstructions, branchNameInstructions, review guides) cover git text only.

A team with its own workflow ends up with two instruction sources that contradict each other. These are real conflicts we hit:

• New threads. The block says "t3_thread_launch … only when the user asks for a new thread. 'Subagent', 'delegate', and 'parallel' never mean one." Our rule is that every assignment is its own top-level t3_thread_launch in the target product's project, so that our router sees it.
• delegate_task. The block routes background work, a separate checkout or another provider's model to delegate_task. We deny delegate_task in .claude/settings.json, because its children stay in the caller's project and skip our routing.
• Model switching. The block says "never switch unasked". Our rule moves recon to another model when a vendor's quota window is nearly used up.

Agents then spend tokens reconciling the two sources on every decision, or follow the wrong one. Turning off the t3 MCP is not a workaround, because we need its tools (t3_thread_launch, t3_thread_read, t3_thread_configure, …).

Related: #19908 (a carve-out for repo-mandated review workflows), #24967, #16465. Each of them patches one sentence of the block. A setting would let a team fix all such cases without a release.

Request

A per-project setting, or a global one with a per-project override, modeled on Source control writing style, that can:

• (a) append custom text after the orchestration block;
• (b) replace the block with custom text;
• (c) turn the block off while the t3 MCP tools stay attached.

Per-section toggles (Subagents, Threads, Background work, Visuals, PR linking) would also do.

The minimum is (c). A team that writes its own AGENTS.md can describe the t3 tools itself, and the tool descriptions already carry most of the mechanics.