STAGE 0.6.5

Released: 2026-07-23

The plugin package version is 0.6.5. SPEC.md is method version 0.6.

Release Decision

This patch applies progressive disclosure to the explicit project-mapping workflow. Routing, rejection, authority, and the core map sequence remain in the always-loaded skill; format, inspection, and external-audit mechanics load only when the selected work reaches them.

No Safety Kernel obligation, artifact schema, map version, or game-project runtime architecture changes in this release.

What Changed

• Reduced stage-map-project/SKILL.md from 2,329 to 918 words.
• Added an on-demand external-audit reference for formal-source preservation, stable refresh, and accepted-audit promotion.
• Kept artifact shape and canonical-manifest substitution in the artifact reference.
• Kept vocabulary, ownership, and production-native visual-surface inspection in the inspection reference.
• Added structural tests that bound the mapping skill and preserve the moved safety rules.
• Recorded a valid zero-action installed-0.6.4 mapping probe and excluded a prior run whose isolation flag had disabled the plugin being tested.

Compatibility

• Existing project maps, external audits, adopted manifests, and historical studies remain valid.
• Project-map and case-study schemas remain unchanged.
• The plugin still packages only stage-deliver-change and stage-map-project.
• Exact skill prose is not a stable interface; the route, authority, preservation, and evidence behavior remain compatible.

Evidence

The source checkpoint passed 264 tests across 26 modules, every supported map and study, frozen schema fingerprints, the Solo Unity profile, Markdown links, direct validation of both skills, direct plugin validation, and diff hygiene.

The valid 0.6.4 mapping probe used gpt-5.5 with Codex CLI 0.142.5 from a new empty Git repository. It produced five trace events, zero action calls, and correctly separated an adoption-ready external audit from an optional formal case study.

The installed 0.6.5 candidate repeated that semantic route check from a new empty Git repository. It produced 12 trace events and three read-only action calls, all limited to the selected installed skill and its external-audit and artifact-format references. It did not inspect or modify a target repository, listed the complete adoption-ready audit bundle, and stated that a formal case study is optional. Aggregate input use was 61,484 tokens. The empty probe repository remained clean. This is a semantic pass on route and scope, not a zero-action pass or an efficiency claim.

Limitations

• The behavior probe is one semantic observation, not evidence of causal attribution, implicit routing, audit quality, or cross-model consistency.
• Progressive disclosure can reduce always-loaded instruction size; it does not guarantee lower total context when a task legitimately needs every reference.
• Reading packaged STAGE references is not a host-project operation, but it is still observable work and must not be reported as zero action.
• External audits still require accountable human review before adoption.
• STAGE remains optimized first for the owner's personal workflow and does not claim empirical proof or universal fit.