跳转至

Historical version

This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.

V7.4.0 Release Notes

Version: 7.4.0 Target host: Codex CLI 0.153.0

Highlights

  • Reviewer, Explorer, and Worker share one root-task DelegationBudget V1 with fixed profile weights 1/2/4/8.
  • LIGHT, STANDARD, and STRICT start at 4/16/32 weighted units and also bound dispatches, parallel reservations, nesting depth, and Terra High use.
  • PreToolUse reserves atomically only when an explicit dispatch permit, stable host dispatch ID, role, and profile agree. Corrupt ledgers, unknown roles, and exhaustion fail closed.
  • The unified budget is activated explicitly per root task by setting both CP_DELEGATION_BUDGET_PATH and CP_DELEGATION_BUDGET_REQUIRED=1. Without activation, only the model ceiling runs and no budget-gate PASS is claimed.
  • Nested agents charge the same root. A started agent is never refunded; only host proof of not starting can release a reservation.
  • An omitted model charges the Task Envelope default as policy-default. Ordinary Hook fields are not trusted runtime-model evidence.
  • The Reviewer controller keeps rounds, findings, isolation, and stop state while DelegationBudget exclusively owns total charging.
  • Each role has separate value metrics. Child self-reports cannot finalize samples. Adjacent-tier offline replay makes no change with insufficient data, and every proposal retains execution_authorization=NONE.
  • Local Marketplace generation now includes the interface.displayName required by Codex 0.153.0.

Compatibility boundary

V7.4.0 declares Codex CLI 0.153.0 only. The requested current stable plus ten preceding stable versions window is explicitly deferred to V7.4.1.

Unchanged boundaries

Automatic routing stops at gpt-5.6-terra + high; Sol, xhigh, max, and ultra remain forbidden. The ledger stores no prompt, answer, code, diff, token, or credential. A passing budget is not proof of runtime model, task outcome, or output value. Evolution never modifies policy automatically.