Loading...
Loading...
Token/quota throttling rules for multi-agent orchestration. Before any Agent tool call or Workflow dispatch—regardless of whether the user mentioned saving tokens, even a single subagent counts—read this skill first before acting. Also triggered when the user says "save tokens", "quota", "don't burn my limit". Trigger words: workflow, multi-agent, fan-out, subagent, dispatch, orchestration, save tokens, quota, token optimizer. English triggers: "save tokens", "quota", "don't burn my limit", "token optimizer".
npx skill4agent add tingyulu/myr2d2 token-optimizerAdapted from kieiken/ultracode-token-optimization (MIT), adapted for all Claude environments and generalized in 2026-07. Core insight: Saving tokens is not about reducing unit price—it's about eliminating "rework loops" and "context flooding"— The edit→review→edit-again→review-again loop, and flooding the main orchestrator with raw diffs/logs, are the real biggest wastes.
agent()parallel()pipeline()| Role | Assignment | Rationale |
|---|---|---|
| Main orchestrator (session itself) | Follows the session, don't touch | Only does judgment and arbitration |
| Executor (writes code/edits files) | Mid-tier (e.g., | Quality is sufficient, quota-friendly |
| Reviewer (fault-finding/review) | Mid-tier, default | Only upgrade to flagship for high-risk areas (see below) |
| Mechanical work (summarization/formatting/grep aggregation) | Low-tier (e.g., | Don't use expensive models for work that needs no judgment |
| Arbitration/final judgment | Not specified (inherits session) | Only exception; must add a comment explaining why |
agent()modelThe criterion is "if the session runs more expensive than mid-tier, it applies"—not a specific model name; no need to change this when new models are released.
settings.jsonenv.CLAUDE_CODE_SUBAGENT_MODEL=sonnetsettings.json.claude/settings.local.jsonschemabudget.remaining()pipeline()parallel()agent()agent()labelresumeFromRunIdisolation: 'worktree'agent()modelpipeline()schemaagent()labelbudgetbudget.remaining()