Skip to content

PATCH_NOTES.md

Latest commit

 

History

History
67 lines (54 loc) · 3.31 KB

PATCH_NOTES.md

File metadata and controls

67 lines (54 loc) · 3.31 KB
Raw

Codex Audit patch notes

2026-09-05 — Astra usage windows and visual reports

  • Added frozen-window, per-chat/model/effort usage audits with parent-tree rollups.
  • Added offline HTML dashboards and Markdown/JSON reports with cache and credit breakdowns.
  • Updated standard-rate estimates and documented subscription, service-tier, and attribution limits.
  • Added regression coverage for modern and legacy usage records and window boundaries.

2026-08-02 — Luna v2 activation and Sol-led orchestration

  • Added an opt-in, rollback-safe guide for exposing Luna to native multi-agent v2 through a validated custom model catalog.
  • Added a Sol-led senior mode: Sol-low for routine orchestration and delegation; Sol-high/xhigh for initial specs/docs, senior advice, broad QA, and hard or high-blast-radius decisions.
  • Routed direct focused code/adjustment packages to Luna-xhigh, lower-level interpreted problems to Terra-medium, and deeper issues to Terra-high/xhigh.
  • Added native lifecycle and nested-spawn proof gates. CLI success alone is not treated as proof that Codex Desktop accepted Luna for native v2.

Validation

  • 31 unit tests passing.
  • Efficient orchestrator skill validation and git diff --check passing.
  • Independent Sol-low policy QA: GO after rollback and capability-proof fixes.

2026-07-11 — calibration, live monitoring, and center-out routing

This release adds the complete calibration and live-monitoring workflow for long-running Codex orchestration:

  • Added audit-codex-token-routing calibration scorecards with per-agent token, credit, duration, context, compaction, coordination, and rework signals.
  • Added explicit acceptance/rework/defect ledgers and matched task-family comparisons. Quality claims remain gated on evidence rather than inferred from task_complete.
  • Added monitor-codex-token-routing, which continuously refreshes a compact agent snapshot, Markdown report, redacted calibration JSON, local dashboard, and monitor status file.
  • Added hypothesis-first monitoring. When quality evidence is incomplete, the monitor now reports a provisional route candidate, confidence, blockers, and a concrete matched next test instead of refusing to reason from the data.
  • Added live --baseline support for phase-over-phase deltas.
  • Added center-out routing guidance: Luna-xhigh control plane for clear spec-gated work; Luna-low for mechanical inventory/extraction; Luna-high for ordinary implementation/tracing; Terra-medium/high for depth, reconciliation, and meaningful integration QA; Sol-low for judgment QA; and Sol-high only for material architecture, security, release, or conflicting-evidence judgment.
  • Added explicit controller/worker route handling, a default two-worker cap, and a hard four-worker maximum.
  • Hardened read-only SQLite polling, malformed session handling, SQL identifier validation, output collision protection, and monitor failure preservation.

Validation

  • 31 unit tests passing.
  • All three skills validate.
  • Ruff and Python compilation pass.
  • Live redacted monitor smoke emits a hypothesis and next test.
  • Privacy scan is clean.

Safety contract

The tools remain read-only and recommendation-only. A working hypothesis is allowed and expected; automatic route/default promotion still requires explicit quality evidence and user review.