Vc Publish

vc-publish

Output style: Follow process/development-protocols/communication-standards.md — answer-first, plain language, no unexplained jargon, TL;DR on long responses.

Push harness improvements from the current development repo to the remote kit repository (vibecode-pro-max-kit). This is the maintainer counterpart to vc-update.

• vc-update = user pulls latest harness INTO their project FROM the remote
• vc-publish = maintainer pushes improvements FROM the development repo TO the remote kit repo

Prerequisites

• Local checkout of the kit repo (git clone git@github.com:withkynam/vibecode-pro-max-kit.git)
• .vc-publish-config file in the current repo root (see Configuration below)
• Git push access to the remote kit repo

Configuration

Create .vc-publish-config in the repo root:

jsonCopy code
{"kitRepoPath": "/path/to/vibecode-pro-max-kit"}

If this file is missing, ask the user for the kit repo checkout path and offer to create it.

Workflow

Step 1: Load Configuration

1. Read .vc-publish-config from the current repo root.
2. If missing, ask the user for the kit repo local checkout path.
3. Verify the path exists and contains vc-manifest.json.
4. Verify the kit repo worktree is clean (git -C <kitRepoPath> status --porcelain). If dirty, warn and ask whether to proceed or abort.

Step 2: Read Manifest

5. Read vc-manifest.json from the kit repo checkout.
6. Extract the current version.
7. Before computing a version bump, check if the current kit version already matches the intended target version (e.g. 3.0.0). If the kit version already equals the target: skip the bump step entirely and proceed directly to Step 3 with a tag-as-is note — do not increment the version. Record in the publish summary that the version was unchanged.

Catalog-regen (pre-publish): Before resolving files in Steps 3–4, regenerate the skills catalog in the dev repo:

bashCopy code
node .claude/skills/vc-audit-context/scripts/generate-skills-catalog.mjs --write

This ensures process/context/generated-skills-catalog.json is current before it is copied into the kit repo.

Step 3: Resolve Kit File Set

7. Run the resolver against the kit repo to get the kit file list:

   bashCopy code
   node <kitRepoPath>/resolve-manifest.mjs --root <kitRepoPath> --json
   

   Extract files (kit managed files) and kitOnly (kit-exclusive files).

   Note: resolve-manifest.mjs reads vc-manifest.json from its --root directory and also scans files from that same root. vc-manifest.json is NOT installed into dev/user projects by install.sh, so the resolver must always be pointed at the kit repo checkout (which does have it). There is no separate dev-repo resolver call — the dev-side file comparison happens inside compute-sync-plan.mjs in Step 4.

Step 4: Compute Diff

9. Computation via compute-sync-plan.mjs: Use the shared computation core to produce the diff between the dev repo's managed files and the kit repo's current managed files.

   Direction note: compute-sync-plan.mjs loads vc-manifest.json from --kit-root and runs the resolver with --root <kit-root>. Since vc-manifest.json lives in the kit repo (not the dev repo), --kit-root must always be the kit repo checkout. --root is the dev repo (the project being compared). This is the same direction as a normal install — vc-publish uses it to see what a fresh install FROM dev INTO the kit would change.

   bashCopy code
   # --root = dev repo (the "project" being compared against the kit source)
   # --kit-root = kit repo (where vc-manifest.json lives; source of truth for file lists)
   # --resolver overrides the resolver path because compute-sync-plan
   # would otherwise look for resolve-manifest.mjs inside --kit-root,
   # which IS the kit repo here, so --resolver is optional but explicit for clarity.
   node <kitRepoPath>/compute-sync-plan.mjs \
     --root <devRepoRoot> \
     --kit-root <kitRepoPath> \
     --resolver <kitRepoPath>/resolve-manifest.mjs \
     --json
   

   Parse the JSON output: { toAdd, toModify, toDelete, toPreserve, staleWarnings }.

   • toAdd — files to copy from dev to kit (present in dev, not yet in kit).
   • toModify — files to overwrite in kit (tracked in both, content differs).
   • toDelete — files to remove from kit (no longer in dev managed set).
   • toPreserve — files to leave untouched (merge/copyIfMissing survivors, unchanged files).
   • staleWarnings — paths that failed the namespace guard — print to user; do NOT delete.

   The ownedPaths for the publish direction are the dev repo's resolved ownedPaths. CLAUDE.md and AGENTS.md are always in the merge category — they require special stripping regardless of diff status (see Step 7).

Step 5: Print Diff Summary

10. Print a summary table:

Copy code
vc-publish diff: current repo -> kit repo (v2.1.0)
================================================

FILES:
  [modified]  .claude/agents/vc-execute-agent.md  (+8 -3)
  [modified]  .claude/hooks/lib/scout-checker.cjs  (+2 -1)
  [new]       .claude/skills/vc-new-skill/SKILL.md
  [merge]     CLAUDE.md (needs content review)
  [merge]     AGENTS.md (needs content review)
  [unchanged] .claude/settings.json
  ... (350 more unchanged)

Total changes: 4 files modified, 1 new, 0 removed

Step 6: STOP -- Confirm Publish

11. STOP and ask the user:
    • Confirm they want to publish these changes.
    • Specify version bump type: patch, minor, or major.
    • Optionally provide release notes (1–3 sentences for the GitHub Release body). Leave blank to auto-generate from the diff summary (e.g. "4 modified, 1 new, 0 removed.").
    • Or abort.

Version bump semantics:

• Patch (2.1.0 -> 2.1.1): hook fixes, skill doc updates, minor agent prompt tweaks
• Minor (2.1.0 -> 2.2.0): new skills, new agents, new development protocols
• Major (2.1.0 -> 3.0.0): CLAUDE.md structure changes, manifest schema changes, breaking workflow changes

Step 7: Apply Changes

12. On confirm:

    • Copy all modified and new managed files from current repo to kit repo checkout.

    • For each removed file: delete it from the kit repo checkout.

    • CLAUDE.md and AGENTS.md stripping: Do NOT copy the current repo's project-specific versions directly. Instead:

      1. Read the current repo's CLAUDE.md/AGENTS.md.
      2. Read the kit repo's existing harness-only version as base.
      3. Apply only methodology/structural changes from the dev repo to the kit's harness-only version.
      4. Strip all project-specific content:
         • Technology stack details (frameworks, databases, versions)
         • Feature list / "Current features" entries
         • Project-specific context groups
         • Hardcoded package manager (replace with generic)
         • MCP server instructions (project-specific config)
         • Project-specific routing rules
         • Absolute paths (/Users/...)
         • Product name references (the project's product name and repo/directory name)
      5. Verify the result is harness-only methodology with no project leaks.

    • Manifest reconciliation (vc-manifest.json) — the manifest is NOT a normally-copied managed file (it is dev-only on one side and kit-resolved on the other), so its fields do NOT auto-sync. Handle it explicitly:

      1. version: bump per the chosen bump type. Kit-authoritative.
      2. legacyDeletions (the deprecation ledger): dev is authoritative and always the superset. Whenever dev removes a skill/dir/file from the harness it appends the path here so downstream projects clean it up on their next vc-update. Set kit.legacyDeletions = dev.legacyDeletions (preserve dev order). This field is a literal path array, NOT a glob — it genuinely changes every time the harness deprecates something, so it MUST be reconciled at every publish. (Historically this was skipped, which silently stranded the kit with stale deletions — never skip it.)
      3. All OTHER non-version fields (include, exclude, strip, merge, copyIfMissing, symlinks, kitOnly): these legitimately diverge — the kit carries packaging-only rules (e.g. excluding **/*.test.mjs and __tests__/** so tests are not shipped to users; kitOnly tooling like compute-sync-plan.mjs). Do NOT blindly overwrite them — a blanket dev→kit copy would strip the kit's test-excludes and ship test files. Instead run the field-level drift report below and reconcile any unexpected drift by hand.

      Drift-report command (run during Step 4 summary AND here before writing):

      bashCopy code
      node -e '
      const d=require("<devRepoPath>/vc-manifest.json");
      const k=require("<kitRepoPath>/vc-manifest.json");
      let drift=0;
      for(const key of new Set([...Object.keys(d),...Object.keys(k)])){
        if(key==="version") continue;
        if(JSON.stringify(d[key])!==JSON.stringify(k[key])){
          drift++;
          console.log("DRIFT field:",key);
          console.log("  dev:",JSON.stringify(d[key]));
          console.log("  kit:",JSON.stringify(k[key]));
        }
      }
      console.log(drift?("\n"+drift+" manifest field(s) drift — legacyDeletions auto-syncs dev→kit; reconcile the rest consciously."):"manifest in sync (besides version)");
      '
      

      legacyDeletions appearing in the drift report is EXPECTED and is auto-resolved (dev→kit). Any OTHER field in the report is a conscious decision: confirm the kit value is the intended packaging rule, or update dev/kit so they converge. Never let a drift go unexamined.

    • Create symlinks if missing (.agents/skills -> ../.claude/skills).

Step 8: Leak Detection

13. Verify no project-specific content leaked into the kit repo. This is a resolved-set, two-check gate that scans the full shipped TEXT surface, not just CLAUDE.md/AGENTS.md.

    Resolve the shipped set via the kit's resolver, then restrict to TEXT surfaces:

    bashCopy code
    node <kitRepoPath>/resolve-manifest.mjs --root <kitRepoPath> --json
    

    Take the resolved files and keep only TEXT surfaces:

    • .claude/skills/** matching *.md, *.cjs, *.mjs, *.py, *.js, *.json
    • .claude/agents/** matching *.md
    • .codex/**
    • process/development-protocols/**
    • plus CLAUDE.md, AGENTS.md

    Exclude binaries and **/node_modules/**.

    Check (a) -- product-name grep over the resolved text set. Scan for the product names in the grep below ONLY. tRPC/Prisma are DROPPED from this skill-prose scan to avoid false positives in legitimate generic test guidance; the hosted-database product name is KEPT (see the pattern):

    bashCopy code
    grep -rIin "flowser\|CloakBrowser\|OpenClaw\|Supabase" <resolved-text-files>
    

    Allowlist the Bucket-4 lines that MUST keep the literal to function (otherwise the gate flags itself): author: flowser frontmatter; the isFlowserActivePlanPath identifier; this skill's OWN scrub-grep pattern lines below; the new validator's own pattern strings; and the one internal plan-generation validation comment in session-init.cjs.

    Check (b) -- non-portable context-path grep: any concrete backticked process/context/... file reference in the resolved text set, MINUS the shipped/seeded survivors, is a dangling-link leak → FAIL with file:line. Survivors (allowed): process/context/all-context.md, process/context/tests/all-tests.md. Portable directory refs (e.g. process/context/tests/) and the process/context/... placeholder are fine.

    Keep the existing narrow CLAUDE.md/AGENTS.md grep (this stays as-is on just those two files; tRPC/Prisma plus the hosted-database product name all REMAIN here, as shown in the pattern below):

    bashCopy code
    # Must return empty -- any matches indicate leaked content
    grep -ri "flowser\|tRPC\|Prisma\|Supabase\|CloakBrowser\|OpenClaw" CLAUDE.md AGENTS.md
    
    # Must return empty -- no absolute paths
    grep -r "/Users/" .
    

    Check (c) -- README badge counts: Verify the kit README.md badge counts match actual agent and skill counts:

    bashCopy code
    actual_agents=$(ls <kitRepoPath>/.claude/agents/*.md | wc -l | tr -d ' ')
    actual_skills=$(ls -d <kitRepoPath>/.claude/skills/vc-*/ | wc -l | tr -d ' ')
    readme_agents=$(grep -oE '[0-9]+-Agents' <kitRepoPath>/README.md | grep -oE '[0-9]+')
    readme_skills=$(grep -oE '[0-9]+-Skills' <kitRepoPath>/README.md | grep -oE '[0-9]+')
    echo "Agents: actual=$actual_agents badge=$readme_agents"
    echo "Skills: actual=$actual_skills badge=$readme_skills"
    [ "$actual_agents" = "$readme_agents" ] && [ "$actual_skills" = "$readme_skills" ] && echo "PASS" || echo "FAIL: badge counts mismatch"
    

    If FAIL: update README.md badges to match actual counts before committing.

    NOTE: the brand grep matches product names ONLY. It does NOT match .ck.json/.ckignore -- those Phase-2 legacy-fallback literals are intentional and must NOT be flagged. Do not add ck/ckignore to any leak grep.

    The standing validate-kit-portability.mjs validator (run by vc-audit-vc) mirrors checks (a) and (b) for between-release drift; this Step-8 gate is the publish-time enforcement.

14. If leak detection fails:

    • Print the offending lines (file:line).
    • Revert the changes in the kit repo (git -C <kitRepoPath> checkout .).
    • STOP and report the leak. Do NOT commit or push.

Step 9a: Commit and Tag

15. In the kit repo. If the version was already at the target (skip-bump path from Step 2): use tag-as-is and commit with the existing version number. Otherwise: bump the version in vc-manifest.json and commit with the new version.

bashCopy code
cd <kitRepoPath>
git add -A
git commit -m "Release vX.Y.Z"
git tag vX.Y.Z

Step 9b: STOP — Explicit Push Approval (separate from Step 6 confirm)

Leak detection passed. The commit and tag are ready locally. Before running git push, you MUST stop and get explicit user approval:

"Leak detection passed. The commit is ready locally (git log --oneline -1 shows the new commit). Type 'push' to publish to remote, or 'abort' to keep the commit local."

Do NOT run git push or git push --tags until the user types 'push' (or a clear affirmative). This is a separate gate from the publish-confirm at Step 6 — even if the user approved publishing in Step 6, they must re-confirm before the actual remote push.

If the user says 'abort':

• The local commit and tag are preserved.
• Print: "Commit and tag preserved locally. Run git push origin main && git push --tags when ready."
• Stop.

Only on explicit 'push': proceed to Step 10 (git push origin main && git push --tags).

Step 10: Push

16. Push to remote (only after Step 9b approval):

bashCopy code
git push origin main && git push --tags

17. If push fails (e.g., rejected, auth error), report the error. The commit and tag are preserved locally for retry.

Step 11: Create GitHub Release

18. After a successful push, create a GitHub Release so watchers are notified and the release appears in the Releases tab:

    bashCopy code
    gh release create vX.Y.Z \
      --repo <remote-owner>/<remote-repo> \
      --title "vX.Y.Z: <first sentence of release notes>" \
      --notes "<full release notes>"
    

    • If the user provided release notes at Step 6, use them verbatim.
    • If left blank, auto-generate: "N modified, M new, P removed. See commit log for details."
    • The --title one-liner should be the first sentence of the notes (truncate at 72 chars if longer).
    • If gh is unavailable or the push to remote failed, skip this step and note it in the summary.

Step 12: Post-Publish Remote Verify

After a successful push, clone the kit from remote to a temp dir and verify the catalog works on a fresh install:

bashCopy code
TS=$(date +%s)
git clone <remote-kit-url> /tmp/vc-kit-verify-$TS
node /tmp/vc-kit-verify-$TS/.claude/skills/vc-context-discovery/scripts/discover-skills.mjs 2>&1
rm -rf /tmp/vc-kit-verify-$TS

Expected: exit 0 and expected skill count in output. If FAIL: note the error in the publish summary — the push succeeded but the remote install may have a catalog issue.

Step 13: Print Summary

19. Print publish summary:

Copy code
vc-publish complete
===================
Version:       v2.2.0 (was v2.1.0)
Files changed: 4
Remote:        git@github.com:withkynam/vibecode-pro-max-kit.git
Tag:           v2.2.0
Release:       https://github.com/<owner>/<repo>/releases/tag/v2.2.0

Key Changes from v1.0

• Glob arrays are stable; the deprecation ledger is not. The glob patterns in include/exclude/kitOnly are stable — adding a new skill or agent requires zero manifest edits (new files are auto-included by the globs). BUT legacyDeletions is a literal path array, not a glob: it grows every time the harness deprecates a skill/dir/file, and it MUST be reconciled dev→kit at every publish (see Step 7 Manifest reconciliation). Skipping it strands the kit with a stale deletion ledger so downstream vc-updates never clean up the newly deprecated dirs. So publish-time manifest edits are: (1) version bump, (2) legacyDeletions sync, (3) conscious reconciliation of any other field flagged by the drift report.
• Resolver-driven diffing. The kit repo is resolved via resolve-manifest.mjs (which requires vc-manifest.json in its --root). The dev-side file comparison is handled inside compute-sync-plan.mjs, which reads the manifest from --kit-root (always the kit checkout). Dev repos do not carry vc-manifest.json.
• No managed/managedDirs arrays to update. The old workflow of adding new files to these arrays is eliminated.

Rules

• ALWAYS run the full resolver diff (Steps 3-4) even when changes already exist in the kit repo. Direct kit edits (README, translations, community files) do not replace the dev→kit diff. Both change sources must be captured in the same publish.
• NEVER copy project-specific files: process/context/all-context.md (with real content), process/features/*, process/general-plans/* (with real plans)
• ALWAYS verify no project-specific content leaked before committing (Step 8)
• ALWAYS show the diff summary before publishing (Step 5-6)
• NEVER auto-push. git push must be preceded by a separate explicit 'push' confirmation (Step 9b), distinct from the publish-confirm at Step 6. This rule holds even when VC_KIT_SOURCE is a local path.
• CLAUDE.md and AGENTS.md require special handling -- never copy the development repo's project-specific versions directly
• Kit repo checkout path is stored in .vc-publish-config (add to .gitignore)
• The only manifest edit at publish time is the version bump -- glob patterns are stable

Reference

See references/vc-publish.md for the detailed algorithm, CLAUDE.md/AGENTS.md stripping rules, error handling, and example outputs.