Skip to main content

STAGE 0.5.7

Released: 2026-07-22

STAGE 0.5.7 closes an artifact-consumer ambiguity in visual review and keeps STAGE's own project map durable across routine verification work. It keeps the 0.5 Practitioner Kernel and all supported artifact schemas compatible.

Why This Patch Exists​

A qualified dogfood probe loaded the installed 0.5.6 visual-review skill and correctly rejected a universal capture matrix. It then proposed an external contact-sheet generator and thin export harness for generic "human and AI review." The skill obeyed the phrase "named consumer" syntactically while recreating the detached workflow the method had already rejected.

A second probe of the 0.5.7 candidate refused detached exports, but still audited the entire presentation layer and turned "every visual object" into an Almanac expansion plan. That exposed a sequencing bug: project inventory ran before the skill required one bounded visual question.

The first correction moved the question gate earlier, but a follow-up probe only narrated the stop. It then proposed Almanac subjects, deterministic sweeps, agent driving, and optional export before asking for scope. The final correction makes the unbounded branch terminal and constrains its response.

The same maintenance cycle also showed that STAGE's own project map should not copy release tags and last-run outcomes from their canonical release records. Orientation and invocation history have different update lifecycles.

Changes​

  • Define a decision-bearing artifact consumer as a concrete decision or operation whose portability, retention, comparison, CI, audit, release, or remote-access need cannot be served by the live native surface.
  • State normatively that human review, AI review, recurring inspection, and a desired standing capability do not satisfy the artifact gate by themselves.
  • Forbid approval matrices, contact-sheet generators, capture manifests, and thin export adapters from being offered after a detached workflow is rejected unless a separate artifact consumer passes the gate.
  • Propagate the rule through the installed skill, Unity profile, visual guide, method, reference architecture, project map, glossary, evidence matrix, and Failure Atlas.
  • Record the decision in ADR 0040 and add regression coverage for the wording that enabled the behavioral failure.
  • Move question framing ahead of repository inspection, reject exhaustive coverage as an implicit objective, and record the correction in ADR 0041.
  • Make the unbounded branch terminate after one scoping question instead of allowing a smaller speculative system proposal.
  • Generalize the probe lesson into terminal stop semantics for ordinary delivery: preserve state, run at most one directly discriminating operation, or ask for one missing input and end the blocked branch.
  • Document and package the copy-pasteable plugin-qualified skill invocations; the bare explicit-skill name did not route in the release consumer, while $stage-game-engineering:stage-visual-review injected and terminated before repository inspection as required.
  • Keep STAGE's own project map focused on durable operation contracts while revision-bound release evidence remains in release records.

Practitioner Impact​

For co-located visual work, leave the smallest truthful native engine or game surface ready for the owner. Repetition does not turn that handoff into a need for exports.

Before retaining screenshots, videos, contact sheets, baselines, or an export pipeline, name:

  1. the concrete decision or operation consuming the artifact;
  2. why the live native surface cannot serve it;
  3. the minimum retained output and its owner; and
  4. when the output or adapter can be retired.

Compatibility​

The plugin package version is 0.5.7; 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, and Markdown-link validation;
  • package self-containment and size checks;
  • a fresh local plugin installation reporting version 0.5.7; and
  • a qualified read-only behavioral probe against the Lanternworks dogfood project that must reject generic human/AI review as an export consumer and stop an unbounded request before repository-wide visual inventory.

Limitations​

The motivating behavior was observed in one model run against a same-owner Unity dogfood project. This patch closes a demonstrated instruction ambiguity; it does not prove that all future agents will make identical routing decisions or that capture pipelines lack value when a concrete consumer exists.

Last updated on