Skip to content

Child token budgets below one iteration's cost are no longer refused (#736 removed the floor; no replacement or changelog note) #2694

Closed

Description

@drewstone

supervise no longer refuses a child token budget that is too small to finish one iteration.

#685 (commit 012762f, 2026-08-01) added this check. #736 (merged 2026-08-04) removed it in the sub-commit "fix(supervise): remove guessed worker token floor":

  • src/runtime/supervise/budget-floor.ts was deleted;
  • below-runtime-floor and PI_TOKEN_FLOOR no longer appear on main (checked 2026-11-17 at fc92d9d);
  • the CHANGELOG does not mention the removal.

Why it matters (the evidence from #685's own commit message)

  • Parents wrote child budgets of 2,000, 8,000, 12,000 and 20,000 tokens.
  • Six pi children on tangle-router/glm-5.2 settled between 31,211 and 62,616 input tokens.
  • In run proof-native-tools-20260801b, a child wrote its artifact correctly, then overran a 12,000-token budget by spending 76,657. The run recorded it as a failure: a false negative.

What main offers instead (checked in the docs, 2026-11-17)

  • The budget ledger reconciles measured spend after the turn. Overspend returns a BudgetViolation.

  • preflightSpawn can refuse a spawn before reservation, but:

    • SpawnRefusalCause is closed to model-route | bridge-full | unmountable-tool | unprovisionable-profile;
    • SpawnPreflightContext does not carry the child's token budget.

    So a caller cannot write a budget-floor check through the supported gate.

Ask

  • If dropping the floor was intended (one measured pairing should not be a global guess), document the decision in the CHANGELOG.
  • Say how a caller should catch an unsatisfiable child budget before dispatch. For example, expose the child's budget to SpawnPreflightContext and allow a budget refusal cause.

tangle.tools has a published post (2026-08-09) that still describes below-runtime-floor. That post is being rewritten from this evidence.

Activity

  1. drewstone commented on Nov 17, 2026

    @drewstone
    MemberAuthor

    Closing. The floor's removal was deliberate and right. #2695 (merged as 53d6ad6) documents what Runtime does today and adds the CHANGELOG line that 0.129.0 lacked. It changes docs only, so there is no release and no consumer bump.

    Deliberate. #736 removed the floor in aafd2a1, "remove guessed worker token floor". The same commit rewrote the agent-graphs skill to give the reason: "Do not turn one run's cumulative spend into a universal harness minimum; Runtime enforces the caller's conserved pool, not guessed per-harness floors."

    Right.

    • 31,211 was the lowest cumulative settlement of six pi children on one model, each running 1 to 4 iterations. It was not the cost of one iteration.
    • The table applied that number to every pi child on every model, and every other harness was null.
    • After refactor(runtime): enforce canonical model execution boundary #736, the exact AgentProfile decides model, prompt, tools and context. Those set the cost of a turn, so a harness-keyed constant refuses lean profiles and admits heavy ones.

    The false negative from #685 has been recovered since then.

    On catching an undersized budget before dispatch: preflightSpawn still does not see the budget. The documented path is to size perWorker and authored budgets from measured runs of the same profile, and to raise the budget through a same-key continuation when a child runs out. The tangle.tools post that describes below-runtime-floor is queued for a rewrite with these facts.

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