Skip to main content

A declared token budget, and a checklist executors can't skip

· 3 min read
Carefree Investments LLC
Anchor maintainers

Two scattered disciplines — "will this even fit?" and "did I check the basics before diving in?" — are now one mechanical gate: every task spec declares a token budget up front, and mythos-core rule 13 makes every executor print a fixed pass/fail checklist before doing any work.

The gap it closes​

Rules 1, 10, and 11 already asked a model to restate the goal, judge whether the task fit its tier, and flag a poor fit — but as separate prose scattered across the system prompt, easy for a small model to skim past under pressure. And nothing declared, up front, whether a task's context plus provided material would actually fit the window it was headed for; a model found out by running out of room mid-task.

Rule 13 PRE-FLIGHT collects the entry checks into one fixed 6-item block a model prints before any other output:

  • Goal restated?
  • Acceptance criteria present?
  • Files-in-scope listed?
  • Budget declared and fits?
  • Tier fit OK (rule 11)?
  • Task small enough for this tier (rule 10)?

Any FAIL routes to that item's already-defined response — rule 1's clarifying question, rule 11's SUGGEST-ESCALATE, rule 10's right-size question, or a decompose-back-to-planner request for files-in-scope/budget gaps — and stops. Rules 1–12 are untouched; this is a pure addition.

Where the numbers come from​

The task-spec template now opens with a ## Budget section:

## Budget
- Context window: <n tokens>
- Output ceiling: <n tokens>
- Spec + provided context exceeding this budget means the task is decomposed
  wrong — reject back to the planner, never truncate silently.

Those numbers are never the model's own guess. scripts/prompt_tuner.py fills them from the registry:

python scripts/prompt_tuner.py "fix the login bug" --target h100-executor

--target names a registered endpoint in scripts/endpoints.yaml; the tuner reads its max_context and computes an output ceiling (context minus spec minus provided context minus a safety margin). Omit --target, or point at an endpoint with no max_context set, and both fields read unspecified — an honest gap, not an invented number.

scripts/orchestrate.py enforces the other half at dispatch time: before each attempt, it checks the assembled prompt against the picked endpoint's max_context. A prompt that's already too big is marked failed-budget and rejected outright, never truncated — an oversized prompt means the task was decomposed wrong, and that's the planner's call to fix, not a retry.