dsh-turn-budget
DeepSeek Harness 的每轮步骤、工具调用和提供商令牌预算采用失效关闭机制
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-turn-budget@0.1.0Turn Budget 以 dsh-turn-budget 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
dsh-turn-budget is a fail-closed resource governor for DeepSeek Harness. It limits model steps, root and nested tool executions, and exact provider-reported token usage independently for every Agent turn.
The plugin uses public Harness extension points only:
agent/pre-step rejects a model request that would exceed the step or provider-token ceiling.tools/pre-execute denies a tool body before dispatch when the turn has consumed its tool-call allowance.tool/result and turn/end records preserve the resulting policy outcome in the canonical Session log.This is a circuit breaker, not a billing system. It does not estimate currency, rewrite tool arguments, replace sandboxing or approvals, or terminate an uncooperative model/tool operation in the middle of a step.
git clone https://github.com/Nunchakus888/dsh-turn-budget.git
dsh plugin --profile web add ./dsh-turn-budget
dsh --profile web --dump-config
dsh --profile web
The bundled profile layer enables a conservative baseline of 24 model steps and 40 tool executions per turn. Override the complete row in the profile's cordis.patch.yml when a deployment needs different ceilings:
- update:
id: turn-budget
config:
maxStepsPerTurn: 16
maxToolCallsPerTurn: 32
maxProviderTokensPerTurn: 250000
Every configured limit must be a positive safe integer. At least one limit is required.
| Limit | Enforcement point | Counted work | Outcome |
|---|---|---|---|
maxStepsPerTurn | Before the next model request | Entered requests in the current turn | The proposed step is rejected; the turn closes as blocked. |
maxToolCallsPerTurn | Before each tool body | Root calls plus Code Mode/nested dispatches carrying the Agent | Excess calls return a policy error naming the exceeded budget; the model retains one opportunity to finish without another tool. |
maxProviderTokensPerTurn | Before the next model request | Latest provider usage per step, including cache read/write buckets | The proposed step is rejected after the reported total reaches the ceiling. |
Token enforcement is exact only when the provider reports usage. Missing provider usage is not estimated. A provider can exceed the configured ceiling inside its current response; the plugin prevents the following request because the public lifecycle exposes no preemptive token stream budget.
The tool ledger is process-local and keyed by live Agent identity. Cold recovery closes interrupted turns before they can continue, so a process restart cannot resume an old in-flight ledger.
pnpm install
pnpm run check
The integration tests boot the real Agent Loop with the published Harness testkit. They cover step rejection, pre-dispatch tool denial, per-turn reset, final-response allowance, provider-token accounting, and duplicate usage replacement.
MIT