[GH-ISSUE #1076] [Build] Generate derived catalogs and skill metadata without feature-branch snapshot commits #393

New issue

Closed

opened 2026-10-15 18:24:42 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-15 18:24:42 +00:00

Owner

Originally created by @morluto on GitHub (Oct 8, 2026).
Original GitHub issue: https://github.com/morluto/rea/issues/1076

Problem

Parallel PRs repeatedly conflict in docs/product-catalog.json, docs/verification/managed-conformance-manifest.json, and generated metadata inside skills/reverse-engineer-anything/SKILL.md. PRs #947, #936, and #934 demonstrate the cycle: unrelated contract changes update repository-wide digests, and .github/workflows/regen-generated.yml commits regenerated snapshots back onto every feature branch. Merging main and regenerating only postpones the next conflict.

Proposed fix

Keep authored contracts and skill instructions in Git; generate derived catalogs, portable conformance projections, and catalog-dependent skill metadata for the exact revision being validated or packaged.
Separate the authored skill from its generated packaged representation, preserving runtime catalog identity and installed-skill diagnostics.
Replace feature-branch auto-commit automation with read-only generation and validation.
Update documentation, website, packaging, and release consumers so they never rely on a stale snapshot or a future bot commit.
Keep real verification evidence distinct from deterministic generated descriptions, and retain the manifest commitment to the exact skill bytes it describes.

Acceptance criteria

Independent contract changes do not require commits to shared catalog digests.
A clean checkout can generate and validate documentation and portable conformance outputs.
Packaged skill metadata matches the catalog shipped in the same package, while generation leaves authored skill instructions unchanged.
CI and release preparation validate the actual generated representation, including drift/error propagation.
No workflow pushes generated-only changes onto feature branches.

PR #1075 reduces schema-only digest churn while preserving runtime identity. This issue covers build-output ownership and the auto-push workflow more broadly; implementation should preserve that work rather than duplicate its doctor changes.

Originally created by @morluto on GitHub (Oct 8, 2026). Original GitHub issue: https://github.com/morluto/rea/issues/1076 ## Problem Parallel PRs repeatedly conflict in `docs/product-catalog.json`, `docs/verification/managed-conformance-manifest.json`, and generated metadata inside `skills/reverse-engineer-anything/SKILL.md`. PRs #947, #936, and #934 demonstrate the cycle: unrelated contract changes update repository-wide digests, and `.github/workflows/regen-generated.yml` commits regenerated snapshots back onto every feature branch. Merging main and regenerating only postpones the next conflict. ## Proposed fix - Keep authored contracts and skill instructions in Git; generate derived catalogs, portable conformance projections, and catalog-dependent skill metadata for the exact revision being validated or packaged. - Separate the authored skill from its generated packaged representation, preserving runtime catalog identity and installed-skill diagnostics. - Replace feature-branch auto-commit automation with read-only generation and validation. - Update documentation, website, packaging, and release consumers so they never rely on a stale snapshot or a future bot commit. - Keep real verification evidence distinct from deterministic generated descriptions, and retain the manifest commitment to the exact skill bytes it describes. ## Acceptance criteria 1. Independent contract changes do not require commits to shared catalog digests. 2. A clean checkout can generate and validate documentation and portable conformance outputs. 3. Packaged skill metadata matches the catalog shipped in the same package, while generation leaves authored skill instructions unchanged. 4. CI and release preparation validate the actual generated representation, including drift/error propagation. 5. No workflow pushes generated-only changes onto feature branches. ## Related work PR #1075 reduces schema-only digest churn while preserving runtime identity. This issue covers build-output ownership and the auto-push workflow more broadly; implementation should preserve that work rather than duplicate its doctor changes.

bsakaguchi

2026-10-15 18:24:42 +00:00