Blank-slate threads: start an agent without any injected context

[colonelpanic8]

TL;DR: I'd like to start a thread with a blank slate: the agent gets none of the context T3 Code, my user config, or the project would normally inject. That means no T3 instructions or t3-code tools, and no CLAUDE.md/AGENTS.md, skills, hooks, memory, or user MCP servers. You get the stock model and harness. Would you accept a per-thread toggle for this? I have an implementation ready.

Problem

Every thread gets a stack of context I can't turn off for one conversation:

• From T3 Code: the t3-code MCP server (about 80 tools on main), plus <runtime_info>, <pull_request_linking>, and the orchestration instructions.
• From my user config: ~/.claude/CLAUDE.md / ~/.codex/AGENTS.md, skills, plugins, hooks, auto-memory, and MCP servers from settings or claude.ai.
• From the project: CLAUDE.md / AGENTS.md and .claude/ settings.

Most of the time that is what I want. Sometimes I want the opposite:

• asking a model something without my agent workflow steering the answer;
• checking whether a behavior comes from my config or from the model;
• getting a second opinion that isn't primed by the project's instructions;
• comparing how the stock harness handles a task.

Today the only way is to leave T3 Code, or to edit global config and remember to undo it.

Proposal

A Blank slate toggle in the composer of a new thread, next to the access and plan controls.

• It is chosen before the first message and stored on the thread. It can't change after the thread exists, because the context is fixed at session start. Forks, subagent children, and resumed or restarted sessions inherit it.
• An existing blank-slate thread shows the toggle as pressed, so you can tell it apart.

For a blank-slate thread:

	What is withheld
All providers	The t3-code MCP server, and every T3 instruction block (runtime info, PR linking, orchestration, browser and device tools)
Claude	settingSources: [] (no CLAUDE.md, settings hooks, plugins, or .mcp.json), skills: [], autoMemoryEnabled: false, disableAllHooks, --strict-mcp-config (left out when a managed-mcp.json exists), claude.ai connectors off, --no-chrome
Codex	project_doc_max_bytes = 0 (no project AGENTS.md), skills.include_instructions = false, memories.use_memories = false
Cursor	Only the team and mdm policy layers

Administrator policy (Claude managed settings, Cursor team/MDM) still applies.

Known gaps

• Codex has no override for $CODEX_HOME/AGENTS.md or for the MCP servers in the user's config.toml. Swapping CODEX_HOME would also drop auth.
• ACP agents still get T3's plan/default mode text, which plan mode needs in order to work.
• The harness's own context stays: Claude Code still adds git status and its built-in system prompt. "Blank" means no user, project, or T3 additions, not no system prompt.

Related

• Setting to not attach the t3-code MCP server (and its prompt text) to agent sessions #17205: opt out of the t3-code MCP server and its prompt text, globally. Blank slate is the per-thread version, and it also covers user and project context.
• Allow custom PR watchers without conflicting runtime instructions #15370: opt out of the PR-watching instructions.
• [Bug]: Claude Code threads get T3 instructions meant for other agents #16465: Claude threads get T3 instructions meant for other agents.
• [Feature]: Configurable instruction files injected into the system prompt, per provider and per project #4490: configurable instruction files.

Evidence

The same prompt in the same project: a normal Claude thread vs. a blank-slate one. The project's CLAUDE.md says "End every reply with the exact word PINEAPPLE".

Normal thread	Blank slate

[colonelpanic8]

Implementation: #19793

[juliusmarminge]

Note

Grok responding on behalf of Julius.

Open PR #19793 already implements this blank-slate toggle. I've linked it so this discussion closes when that lands. Related ideas #17205 (global MCP opt-out) and #19757 (quick no-project chat) stay separate.