Evidence-backed preparation and on-demand capabilities
This release packages the capability foundation and the completed review-preparation work for managed Claude Code and Codex installations.
Changes
- Nine specialized procedures with assessed on-demand selection. Task owners record need, scope, available prerequisites and budget; keywords alone do not grant eligibility.
- Immutable source and input evidence, exact artifact retrieval, freshness checks and isolated review runs.
- Focused specialist assignments and bounded follow-up requests with durable cumulative limits and explicit unresolved outcomes.
- One preparation operation for scope, packets, budget measurement and bounded waves, returning ready inputs or named blockers without trimming required proof or coverage.
- Safe reuse, interruption reconciliation, cancellation and concurrent preparation ownership.
- Proportional real-PR feature acceptance and separate comparative-benefit qualification.
- Updated README, wiki and generated reference pages; historical evaluation outcomes remain preserved.
Validation and limits
A fresh disposable checkout with Node 24.21.0 and locked dependencies passed 227 of 228 deterministic test commands initially. The sole failure was an obsolete capability-boundary wording assertion; the corrected focused context-packet check passes. Source smoke and strict bridge typecheck pass. Final documentation and artifact checks are recorded in the release announcement. Existing bounded owned-PR validation supplies functional evidence; no new model comparison is part of this release. All nine procedures retain evaluation benefit labels. General accuracy, latency and cost advantages, live client discovery, live providers, public deployments, additional operating systems and physical fabrication remain unqualified. The lean Codex plugin does not expose the managed capability entry point.
Updating
Update the source checkout, run the managed installer for your host and restart its session. Existing permissions and exclusions remain authoritative. The runtime archive excludes evaluation fixtures and completion study runners; GitHub's automatic source archives retain the repository's historical research material. No new redistribution-rights claim is made for upstream-derived research fixtures. GitHub Actions remains disabled.