Skip to main content

STAGE 0.5.10

Released: 2026-07-22

STAGE 0.5.10 reduces duplicated plugin instructions, adds bounded inspection for opt-in Codex behavior traces, and narrows external-study claims to the evidence actually collected. It keeps the 0.5 Practitioner Kernel and every supported artifact schema compatible.

Why This Patch Exists​

The 0.5.9 route-rejection probes exposed useful agent behavior, but their JSONL traces had to be counted and summarized manually. The plugin also repeated its full stopping and authority policy in each UI launcher prompt even though the same policy already lived in the skill body loaded by that prompt. Finally, an external-study conclusion claimed that STAGE could prevent evidence-scope overclaiming when the collected maps only showed that mismatches could be made explicit.

Changes​

  • Add an opt-in trace inspector that counts action calls and token use, checks explicit final-message literals, and refuses incomplete traces.
  • Keep model-bound behavior probes outside the deterministic release gate and leave semantic judgment with the maintainer.
  • Separate target-maintainer review status from the STAGE researcher's disposition toward an external study.
  • Replace unsupported prevention language with source-bounded claims about surfacing ownership assumptions and evidence-scope mismatches.
  • Reduce each skill launcher prompt to one route-specific invocation sentence; keep authority, stopping, preservation, and evidence policy canonical in SKILL.md.

Practitioner Impact​

Ordinary use is unchanged. The primary skill still starts game changes directly, optional mapping and visual-review skills remain explicit-only, and route rejection still takes the zero-inspection terminal branch when no cleanup is authorized.

For maintainers, launcher metadata now has one job: invoke the correct skill. The skill body owns workflow policy, and behavior traces can be inspected with a deterministic helper after an explicitly chosen model probe.

Compatibility​

The plugin package version is 0.5.10; SPEC.md is method version 0.5. Project-map and case-study artifact versions 0.1 and 0.3 remain supported. No game-project migration is required.

Verification​

The release is checked through:

  • the complete STAGE release gate;
  • direct Codex plugin and all skill-package validators;
  • repository-local project-map, case-study, schema, and Markdown-link validation;
  • focused launcher-metadata and canonical-policy regression tests; and
  • an opt-in fresh installed-plugin rejection probe recorded outside the repository.

The rejection probe used the installed 0.5.10 candidate with gpt-5.5 and Codex CLI 0.142.5 against clean Lanternworks revision 541b2747096bb358daeff31b5d21994421582498. It produced four trace events, one agent message, zero command or tool actions, no repository inspection, and one concrete project-native outcome question. The target revision and working tree were unchanged.

Limitations​

Shorter launcher prompts remove duplicated instructions; they do not by themselves establish a material reduction in whole-task token use. The trace format reports aggregate usage rather than attributing tokens to skill, system, tool, or conversation context. Behavior probes remain model-, prompt-, plugin-, and toolchain-bound. This probe used 21,630 input tokens and missed its exploratory 20,000-token ceiling despite passing the structural zero-inspection checks. The configured gpt-5.6-sol model could not be used because Codex CLI 0.142.5 requires an upgrade for that model, so the explicit gpt-5.5 result is not a cross-model comparison.

Last updated on