跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
其他技能

delegation-tiering

Decide whether to spawn a subagent, and on which model tier and reasoning effort. Use when planning a fan-out, choosing a subagent model, writing a workflow script's opts.model, authoring an agent definition, setting a repo's cost posture, or when a delegation decision is non-obvious.

JakeSelby · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add JakeSelby/model-citizen --skill delegation-tiering

技能说明


name: delegation-tiering description: Decide whether to spawn a subagent, and on which model tier and reasoning effort. Use when planning a fan-out, choosing a subagent model, writing a workflow script's opts.model, authoring an agent definition, setting a repo's cost posture, or when a delegation decision is non-obvious.

Delegation and model tiering

The operative defaults live in ~/.claude/rules/delegation.md. This is the reasoning, the evidence, and the cases that file is too short to carry.

Research date 2026-09-16. Where a number is vendor-run or unreplicated it says so. Treat every tier boundary below as extrapolation unless it names a Claude-tier measurement.

Runtime mapping

A shared role names one of four capability classes, strongest first — frontier, strong, standard, light — in its contract's tier: line, and each adapter's bindings.json maps the classes it has qualified to native models in a tiers table. The class is a statement about the work; the table is the only place a provider's model names appear. An unmapped class resolves to the nearest stronger mapped class and otherwise inherits the session model — never downward, because a weaker model than the role asked for is a silent failure. Both adapters map all four. Claude Code's table uses version-free aliases. Codex has none — every id carries a version and keeps resolving after its successor ships — so citizen tiers check reads the catalog Codex fetches from its provider and flags a mapped model that is gone, superseded or out of order. tiers.<runtime>.<class> in your config remaps a class in one line; on a provider that lacks these ids, override a role's model to inherit in role_bindings.

Provider model names and benchmark examples below describe their original evaluation context; they are not cross-provider capability equivalences. With no qualified cheaper mapping, inherit the session model and report the gap. Codex does not interpret Claude model aliases. Role authority remains subject to native restrictions; read-only defaults are not proof of confinement.

The headline

Model tier is the third-best cost lever. Reasoning effort beats it and prompt caching beats both. The most decision-relevant number in the corpus, Anthropic-run on a SWE-bench Pro subset, priced as billed:

ConfigurationSolved$/solved task
Opus 5, default effort91.7%$1.01
Opus 5, low effort84.0%$0.25
Fable 5.1, low effort88.6%$0.54
Sonnet 5, default effort77.4%$0.84

Opus 5 at low effort beats Sonnet 5 at default by 6.6 points at 3.4x lower cost per solved task. Sonnet 5 at default is strictly dominated. Drop effort before you drop tier.

Gate 0 — should this be a subagent at all?

In order. First "no" ends it.

  1. Which currency binds? Dollars on API billing; the rate-limit window on a subscription. Cheaper models take more round-trips for the same outcome — one measured tiered run came in 59.4% cheaper while burning more total tokens (15.26M vs 14.84M). On a subscription, most dollar reasoning is the wrong objective function.
  2. Does the working state exceed one context window, or will the orchestrator take many more turns after this? If the work is one dependent chain that fits in one context, the orchestrator pays for a plan, a handoff and a merge that a single model gets for free.
  3. Can you bound the return? You cannot predict a compression ratio, so cap the numerator: name a token ceiling in the brief. Below roughly 10:1 the delegation stops paying.
  4. Does it need back-and-forth, share context with adjacent phases, or is latency binding? All three are don't-delegate signals. A non-fork subagent inherits nothing — no history, no prior reads, no skills, no output style, no memory.

Gate 1 — the brief contract

Agent count correlates −0.021 with quality. Information-transfer coverage correlates 0.614–0.952. Invest in the handoff, not the headcount. Every brief names:

  • the file list or search scope — the subagent does not choose what to look at
  • the return schema and a token cap
  • what the subagent must not decide
  • the output shape, per voice-and-format.md

A vague brief to a frontier model beats a sharp brief to a cheap one far less often than the reverse.

Checks travel with the work

A check that lives outside the model — a governance or trust call hosted by an MCP server, an approval gate, a licence or secret scan — binds the delegated path exactly as it binds the supervised one. A subagent's tool list is usually narrower than its spawner's, so a check the spawner runs by habit is silently skipped the moment the action moves into a subagent.

  • Name the checks in the brief. Every check the spawner would have to run before an action the brief asks for — commit, push, send, deploy — is listed with the action it guards.
  • The subagent makes the call itself when it holds the tool, and obeys the answer as the spawner would: a clear go proceeds, anything else stops.
  • When it cannot make the call, or the answer is not a clear go, it does not act. It finishes the work that needs no check, leaves the guarded action undone, and returns it as a pending action: the exact command, the check it could not run, and why.
  • Each level repeats this. The spawner makes the call if it can and then performs or re-dispatches the action; if it cannot, it passes the pending action to its own spawner. Only the top session prompts the user, so the user sees one question, from the session they are talking to.
  • Pre-clearing is the same chain run early. A spawner that can run the check before dispatch may do so, and says in the brief which action was cleared, at what level, and for which branch or target. A clearance covers that action only; anything wider goes back up.
  • **An unreachable check is reported,
阅读完整来源 ↗