Lean context for delegate_task children: let a worker skip the orchestration instructions and orchestration tools

[EmilioGinzo]

Problem

Every Claude thread T3 starts gets the same appended system prompt and the same t3-code MCP
server, whether it is the orchestrator a human talks to or a leaf child started with
delegate_task to do one task. For a leaf child almost all of it is unused.

We log every request our agents send. Over three days (2,429 sessions, 32,742 requests):

	delegated children (566 sessions)
first request, mean	30.0k tokens
the task brief itself	335 tokens (1.1%)
T3 appended text (orchestration, choose-workspace, visuals, ACP fallback, runtime_info, PR linking)	~8.4k chars, ~3.1k tokens
t3-code deferred tool names	~83 names, ~1.0k tokens
children that called delegate_task	25
children that called t3_thread_launch	0
children that called link_pull_request	5

These tokens are written to the cache at every child start and re-read on every later request of
the child (12,439 requests in the window), and they also pull the child's attention toward
orchestration it is not supposed to do.

What we can and cannot do today

With a second Claude provider instance for workers (launchArgs: --strict-mcp-config --disallowedTools "...", an env marker read by a SessionStart hook) and delegate_task target.providerInstanceId pointing at it, a real child's first request went from 30.0k to 12.7k
tokens. Of those 12.7k, T3's appended text is now the largest single block: 3.3k tokens (26%),
plus 1.2k for the t3-code tool names. What we cannot remove from outside T3:

• the orchestration text appended in ClaudeAdapterV2.ts (systemPrompt.append); launch
  arguments cannot change it ([Bug]: --append-system-prompt-file in Claude launch arguments is silently ignored #13934);
• the orchestration and worktree tools of t3-code, which a leaf child does not use.

Proposal

1. A provider-instance setting for T3's injected instructions, e.g.
   runtimeInstructions: "full" | "worker" | "none" (default full). worker keeps
   <runtime_info> and PR linking, drops the orchestration, choose-workspace, visuals and ACP
   blocks, and drops the orchestration and worktree capabilities from the thread's MCP
   credential, so tools/list shrinks too. This follows the existing browserToolsAvailable
   pattern (mcpSession.ts, ProviderSessionManager.prepareMcpSession).
2. Or, per task: an optional instructions: "worker" on delegate_task, applied when
   appThread.lineage.creationSource === "mcp". makeClaudeQueryOptions has no child flag
   today, but its only caller (openQuery) has turnInput.appThread.lineage, so passing it in
   is small.
3. Export a child marker in the launch environment (as asked in Expose thread lineage to delegated children in the launch env (T3_THREAD_KIND / T3_PARENT_THREAD_ID / T3_TASK_ID) #12286), so user hooks can
   adapt as well.

Either 1 or 2 would cover our case; 1 needs no change to orchestrator behaviour.

Related

• Blank-slate threads: start an agent without any injected context #19790 / feat: blank-slate threads that start without injected context #19793 (blank-slate threads; PR closed unmerged): all-or-nothing per thread. A worker
  still needs PR linking and the tools it is given.
• Setting to not attach the t3-code MCP server (and its prompt text) to agent sessions #17205: global opt-out of the t3-code server and its text.
• Expose thread lineage to delegated children in the launch env (T3_THREAD_KIND / T3_PARENT_THREAD_ID / T3_TASK_ID) #12286: child marker in the launch env.
• Working toward greater efficiency in orchestration context #17253: orchestration context overhead.
• [Bug]: Claude Code threads get T3 instructions meant for other agents #16465 / fix(server): Claude threads no longer get ACP and Codex tool instructions #16482: ACP fallback text sent to Claude threads.
• [Bug]: With agent browser/device access off, the agent still sees every preview_* and device_* tool #17202: tool catalog not filtered by capabilities.
• [Bug]: --append-system-prompt-file in Claude launch arguments is silently ignored #13934 / fix(server): Claude append-system-prompt launch arguments reach the model #20413: --append-system-prompt in launch args never reaches the model.