Skip to content

Clarify replan admission after compute quota is exhausted #8571

Open

Description

@xiaobaifly7

Product boundary to decide

A positive compute allocation whose current allowance is fully spent can still produce should_run=true / run_now for vision_checkpoint_missing. Ordinary delivery remains prohibited. Please clarify whether this replan exception is intended: compute quota is described as the allocation of automatic agent time, so exhausted quota is easy to interpret as a stop on further model work.

This is an observed product behavior, not a Windows-specific diagnosis or an Agent command failure. Confirmed with synthetic source packets on main beff0f3e55d1f70eceebbd8e78c11f1cff34643b; the behavior was also observed on release 2.3.0.

Minimal counts-only reproduction

Use a disposable active Goal with one registered Agent and exactly one blocked advancement Todo, zero executable Todos and no due monitor, user gate or owed prior settlement. Use the public synthetic status/checkpoint shapes from examples/control_plane/quota-replan-decision-plane-smoke.py.

Set compute positive, allowance 1 slot, spent 1 slot, and quota state throttled. Compare the same source with zero versus one retained checkpoint reporting a missing per-Agent Vision baseline; call build_quota_should_run with explicit generic_cli scheduler context.

Input / output No missing checkpoint One missing-baseline checkpoint
Allowed / spent slots 1 / 1 1 / 1
Blocked / executable Todos 1 / 0 1 / 0
Quota state throttled throttled
Effective action throttled_skip autonomous_replan_required
should_run false true
Agent delivery allowed false false
Quiet no-op allowed true false
Scheduler action backoff_until_state_change run_now

These are source decision results. No model turn, spend, fabricated receipt or sustained host loop was executed for this reproduction.

Existing ownership and related work

Existing issue searches for quota/replan exhaustion and throttling found no issue specifically deciding this boundary.

Requested maintainer outcome

Decide whether all fresh replan attempts must wait for renewed compute allowance, or whether replan has an intentional exception with its own visible budget/accounting boundary. Encode and test that decision at the existing typed quota owner and describe it clearly to callers. Preserve owner Stop/zero-compute authority, exact-identity settlement recovery and legitimate no-debit closeouts; do not resolve this by charging every guard read or inventing settlement.

Activity

  1. BigDataDZ commented on Nov 14, 2026

    @BigDataDZ
    Contributor

    独立代码勘察与提案(供维护者决策参考)

    现状代码事实(main 1cdd9bbef):

    • 配额耗尽时普通交付被禁(throttled),但 loopx/control_plane/quota/run_decision.ts:140-160 的 replan 覆盖在 replan_decision_allowed 成立时仍给 should_run=true + AUTONOMOUS_REPLAN_REQUIRED,调度上下文随之输出 run_now / active_work。
    • 该 replan 义务源自 vision_checkpoint_missing frontier gap(outcome_continuity.py:181-219),其收尾动作是"有界 Vision patch 或显式退休"。

    矛盾点:compute quota 定义为"自动 agent 时间的配额",耗尽即无自动时间可用;而 run_now 授权的是一次需要 agent 判断的写入。两个自洽方向:

    • A. 耗尽即停:quota exhausted(throttled 且无余量)时压制 replan 唤醒 → blocked-only wait,missing baseline 保留为显式 pending,配额恢复后 replan 自动恢复。语义最一致;代价是 baseline 修复最多延迟到下一配额周期。
    • B. 控制面例外(我推荐,等价于 A + 一条免扣额控制面通道):若修复可以 no-debit 完成(unchanged-reason patch / 显式退休,不需要模型时间),允许一次有界控制面收尾;任何需要模型判断的路径仍等配额。

    无论选哪个,建议 active_work 的 cadence 标签都不应出现在该场景,避免被读作可执行任务进展。

    建议的验收形状:

    1. allowance 1 / spent 1 + missing baseline → 非 run_now;若允许控制面例外,effective_action 单独分类且 zero-debit;
    2. 配额恢复后 → replan 恢复为现行行为;
    3. 普通 throttled 交付禁令保持不变。
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions