Skip to main content

STAGE 0.5.13

Released: 2026-07-23

STAGE 0.5.13 bounds valid project inspection by decision sufficiency. Agents continue gathering evidence only while another action can plausibly change the engineering decision, evidence route, authority boundary, diagnosis, or material risk. The patch keeps difficult diagnoses and broad audits expandable; it does not impose a universal action, file, token, or time quota.

Why This Patch Exists​

The installed 0.5.12 delivery skill correctly routed a bounded Lanternworks visual diagnosis through the production game and identified a plausible readability risk. It nevertheless used 72 action calls and more than one million aggregate input tokens before reaching a conclusion that a focused search exposed much earlier.

More inspection is not automatically more rigorous. After the smallest complete plan or next discriminating check is already supported, adjacent file reads and repeated corroboration consume owner attention without changing the decision. A fixed budget would create the opposite failure by prematurely truncating genuinely uncertain diagnoses.

Changes​

  • Define decision-bearing uncertainty as the reason to continue inspecting a project.
  • Stop orientation when remaining uncertainty cannot materially change ownership, implementation, verification, authority, diagnosis, risk, or the next discriminating check.
  • Broaden again when evidence conflicts, ownership or authority remains unresolved, blast radius is unclear, a production consumer is missing, verification fails, or broad coverage is the requested outcome.
  • Add a failure-atlas entry for corroboration after the decision is already supported.
  • Add explicit evaluation measures for decision relevance and post-sufficiency investigation.
  • Add a target-first pass and explicit sufficiency checkpoint for bounded named questions.
  • Make diagnose-only named questions an early terminal fast path rather than letting them fall through the full delivery workflow as an inspection checklist.
  • End source inspection when unavailable live observation or human judgment is the next truthful check.
  • Clarify that ADR 0008 retains only the human-authority lesson; its detached capture-plan, adapter, baseline, and receipt machinery is historical and superseded.

Practitioner Impact​

Ordinary delivery still begins in the real project and follows the strongest available information scent. The agent should not inventory neighboring systems merely to feel certain. It should be able to state which unresolved question the next inspection action can change.

This is a stopping rule, not permission for shallow work. A hard bug may need instrumentation, reproduction, minimization, and repeated hypothesis tests. An architecture or release audit may legitimately cover the whole declared scope. The amount of inspection follows the decision, not a preset quota.

Compatibility​

The plugin package version is 0.5.13; 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 both skill-package validators;
  • repository-local project-map, case-study, schema, and Markdown-link validation;
  • focused workflow-policy and packaging regression tests; and
  • a fresh installed-plugin behavior probe recorded after candidate installation.

The first installed candidate preserved the correct production owner, readability risk, and native check, but remained operationally disproportionate: 57 action calls and 1,239,429 aggregate input tokens. Its progress messages said that only live judgment remained, then continued through general settings, UI, editor-lock, and corroborating test reads. STAGE records this as a failed efficiency outcome and uses it to motivate the target-first pass and explicit sufficiency checkpoint.

A second installed candidate reached the same sound result in 52 action calls and 659,631 aggregate input tokens, but it still read the project map first, checked process state, omitted the checkpoint, and kept inspecting after live observation was the remaining discriminator. STAGE does not claim the reduced aggregate input was caused by the prompt change. The second semantic failure motivated the front-loaded terminal diagnosis branch. A final candidate probe then used 29 action calls and 405,961 aggregate input tokens. It searched the exact subject first, skipped process-state inspection, preserved the target, and stopped at the native Game-view check with the same sound diagnosis. It still expressed sufficiency informally and read additional composition context. The release therefore claims a clearer method contract and a material observed routing improvement, not strict behavioral conformance, causal token savings, or optimal inspection depth.

Limitations​

The correction originates in one model-bound Lanternworks probe, one same-revision retrospective, and related research on information foraging, bounded rationality, long-context retrieval, and human oversight of coding agents. It does not reveal the model's internal stopping point, prove that a particular later action was redundant, or establish a universal optimal amount of inspection. The behavioral probe remains evidence about one package, prompt, model, toolchain, target revision, and environment.

Last updated on