docs(CLAUDE.md): discourage unprompted subagents and dynamic workflows (#24027)

v5-next counterpart of #24026.

Adds an `<agent_and_workflow_restraint>` block to the root `CLAUDE.md`
telling Claude to do work inline in the current session and not spawn
parallel subagents or launch dynamic workflows unless the user
explicitly asks.

## Why

Operators have reported burning through their token budget from a single
prompt that quietly fanned out — in one case a "summarize recent ZK
advancements" query started ~30 agents, and another exhausted a 5h
budget spinning up subagents. Parallel agents and dynamic workflows
multiply spend (≈2x for one helper, far more for a swarm) and the user
can neither see the fan-out coming nor stop it. This appears to be a
current tendency of Fable. The guidance reasserts: handle
search/summarize/research/multi-file edits inline, reserve subagents for
explicit user requests or a single read-heavy isolation case, and never
start a dynamic workflow by default.

Passes the repo's `<editorial_test>`: the line would have prevented the
~30-agent fan-out on an ordinary research prompt described above.

An equivalent shared rule is being added in the claudebox repo so it
applies to every managed session.

---
*Created by
[claudebox](https://claudebox.work/v2/sessions/7d5ecfdd5f37c5cd) ·
group: `slackbot`*

Author: ludamad

CLAUDE.md

@@ -83,6 +83,10 @@ Tests should validate behavior, not mock call-count. Prefer `expect(result).toEq
 Before writing a new helper, utility, or component, search for an existing one with Grep or Glob. Reuse or refactor to a shared module; do not introduce a parallel implementation.
 </reuse_before_writing>
 
+<agent_and_workflow_restraint>
+Do the work in this session by default. Do not spawn parallel subagents (the Agent/Task tool) or launch dynamic workflows (the Workflow tool) unless the user explicitly asks for it. Each extra agent multiplies token spend — roughly 2x for one helper and far more when a request fans out to many — and the user cannot see the fan-out coming or stop it; a single prompt that quietly started ~30 agents has exhausted an operator's budget. Searching the codebase, summarizing, researching, and ordinary multi-file edits are inline work: run the tool calls yourself. Reach for a subagent only when the user requested orchestration, or when one clearly-scoped read-heavy helper genuinely needs isolation from the main context — prefer a single agent over many, and never start a dynamic workflow by default. If a task would benefit from parallel agents but the user has not asked, either do it directly or describe the multi-agent option and ask before spending the budget.
+</agent_and_workflow_restraint>
+
 <preserve_todos>
 Preserve existing `// TODO`, `// TODO(name)`, and `// NOTE:` comments unless the current task is to resolve them. A "tidy up" refactor that deletes another author's deferred-work markers destroys context that is not recoverable from git history.