Release validate

Release Validation Skill

Purpose

Pre-release quality gate that validates the devops-ai-skill pack is ready to publish. Covers four areas that existing structure tests (test-structure.sh) do not fully address:

1. Package release checks — version sync, changelog, npm pack content, git state, badge accuracy
2. Cross-platform link validation — all 4 platforms correctly reference skills, prompts, and agents
3. Skill & pipeline parity — every skill/pipeline on disk is registered on every platform
4. Setup script smoke testing — dry-run validation of setup scripts in a sandboxed temp directory

This skill is the only one that validates the skill pack itself (meta-validation). It bridges the gap between test-structure.sh (static file checks) and release.sh (publish). Think of it as a pre-flight checklist that catches semantic issues — dead cross-references, missing registrations, npm packaging gaps, or setup script regressions — that static checks cannot detect.

Activation

This skill activates when the user:

• Asks to validate a release before publishing
• Runs a pre-release check or readiness assessment
• Wants to verify cross-platform consistency after adding a skill or pipeline
• Asks to smoke-test the setup scripts
• Asks to check if the pack is ready for pnpm release

Step 0: Discover Repository Layout

Do NOT assume hardcoded counts or file lists. Discover the pack structure at runtime.

0a: Discover All Skills

Build the canonical skill list dynamically:

find skills/ -mindepth 1 -maxdepth 1 -type d | sort

Store the count as <skill-count> and the list as <skill-names> for all subsequent steps.

0b: Discover All Pipelines

Build the canonical pipeline list:

# Horus pipelines
ls prompts/horus/*.md 2>/dev/null
# Zeus pipelines
ls prompts/zeus/*.md 2>/dev/null
# Shared prompts
ls prompts/shared/*.md 2>/dev/null

Store counts and lists for parity checks.

0c: Discover Platform Entry Files

Verify each platform's entry point exists:

Platform	Entry File	Config Files
Claude Code	CLAUDE.md	.claude/agents/*.md, .claude-plugin/*.json
Codex	AGENTS.md	.codex/config.toml
Gemini	GEMINI.md	.gemini/agents/*.md, .gemini/extensions/devops/gemini-extension.json, .gemini/commands/devops/**/*.toml
Antigravity	.agents/rules/devops.md	.agents/skills/*/SKILL.md

0d: Discover Version Files

Locate all files that must contain the version string:

• VERSION (source of truth)
• package.json → .version
• .claude-plugin/plugin.json → .version
• .claude-plugin/marketplace.json → .metadata.version AND .plugins[0].version
• .gemini/extensions/devops/gemini-extension.json → .version

0e: Read Current Version

VERSION=$(cat VERSION | tr -d '[:space:]')

Phase 1: Package Release Checks

1a: Version Consistency

Run the existing version consistency check and interpret results:

bash scripts/version-consistency.sh

Verify all 5 files from Step 0d contain the same version. If version-consistency.sh is missing, perform the checks inline by reading each file.

Marketplace.json has TWO version fields — both metadata.version and plugins[0].version must match. This is easy to miss because version-consistency.sh only checks metadata.version.

1b: Changelog Validation

Check that CHANGELOG.md contains:

• An entry header matching ## [$VERSION] (required by release.sh)
• A date in the entry (format: YYYY-MM-DD)
• At least one bullet point describing changes under ### Added, ### Changed, ### Fixed, or similar
• The entry is NOT the placeholder (describe changes here) left by version-bump.sh

1c: npm Pack Dry-Run

Simulate what npm publish would include:

pnpm pack --dry-run 2>&1

If pnpm is unavailable, fall back:

npm pack --dry-run 2>&1

Validate the tarball contents against the canonical lists from Step 0:

Required file categories (validate each dynamically, not by hardcoded count):

• Every skills/<name>/SKILL.md from Step 0a is present
• Every prompts/horus/*.md, prompts/zeus/*.md, prompts/shared/*.md from Step 0b is present
• Agent definitions: .claude/agents/horus.md, .claude/agents/zeus.md
• Plugin config: .claude-plugin/plugin.json, .claude-plugin/marketplace.json
• Setup scripts: scripts/setup.sh, scripts/setup/*.sh
• Core docs: docs/PROJECT.md, VERSION

Exclusion check — the tarball must NOT contain:

• .env, .env.* files
• node_modules/
• .DS_Store
• tests/ directory (test files should not ship to users)
• .git/ directory
• Any file larger than 1MB (binary bloat check)

Cross-reference with package.json files field — if a required category is missing from the files array, report it as a packaging gap.

1d: Git State Check

Verify the repository is in a clean state for release:

• No uncommitted changes: git status --porcelain should be empty
• Current branch is main
• Tag v$VERSION does not already exist
• Local main is up-to-date with origin/main:

  git fetch origin main --dry-run 2>&1
  git rev-list HEAD..origin/main --count
  

  If unpushed or unpulled commits exist, report the count and direction.

1e: Badge and Count Accuracy

README badges display counts (skills, pipelines, agents) as shield.io URLs. These must match reality.

# Extract badge counts from README.md
grep -oP 'SKILLS-\K[0-9]+' README.md
grep -oP 'PIPELINES-\K[0-9]+' README.md
grep -oP 'AGENTS-\K[0-9]+' README.md

Compare against canonical counts from Step 0:

• SKILLS badge must equal <skill-count> from Step 0a
• PIPELINES badge must equal total horus + zeus pipeline count from Step 0b
• AGENTS badge must equal 2

Check all 3 README files: README.md, docs/README.zh-TW.md, docs/README.zh-CN.md.

Also scan for hardcoded skill counts in prose text:

grep -n '[0-9] skills\|[0-9] 個 Skills\|[0-9] 个 Skills\|[0-9] skill symlinks' README.md docs/README.zh-TW.md docs/README.zh-CN.md docs/setup.md scripts/setup.sh

Every count found must match <skill-count>.

1f: Version-Tag History Alignment

For informational purposes, verify that previous versions in CHANGELOG.md have corresponding git tags:

grep -oP '## \[\K[0-9]+\.[0-9]+\.[0-9]+' CHANGELOG.md | while read v; do
  if git rev-parse "v$v" &>/dev/null; then
    echo "  ✓ v$v tag exists"
  else
    echo "  ⚠ v$v tag missing (changelog entry exists but no git tag)"
  fi
done

Report missing tags as WARN (not FAIL) — the user may have only recently adopted tagging.

Phase 2: Cross-Platform Link Validation

2a: Discover All Cross-References

Build a reference graph by scanning these entry-point files:

Platform	Entry File	References To
Claude Code	CLAUDE.md	docs/PROJECT.md, prompts/, .claude/agents/
Claude Plugin	.claude-plugin/plugin.json	skills, agents, commands
Codex	AGENTS.md	docs/PROJECT.md, prompts/, skills/
Gemini	GEMINI.md	docs/PROJECT.md, .gemini/commands/, .gemini/agents/
Antigravity	.agents/rules/devops.md	docs/PROJECT.md, prompts/, .agents/skills/

2b: Validate All Referenced Paths Exist

For each cross-reference found in Step 2a:

1. Extract file paths from markdown links: [text](path) and backtick-quoted inline paths
2. Extract paths from code blocks and configuration files (JSON path fields, TOML prompt fields)
3. Resolve relative paths from the referencing file's directory
4. Verify each referenced file or directory exists on disk

Path extraction patterns:

• Markdown: [text](relative/path.md) or [text](relative/path)
• Backtick: `prompts/horus/full-pipeline.md`
• JSON: "path": "../../../skills/repo-detect/SKILL.md"
• TOML prompt blocks: references to prompts/ paths inside prompt = """..."""

Report:

Cross-Platform Link Validation
===============================
Scanning: CLAUDE.md
  ✓ docs/PROJECT.md
  ✓ prompts/horus/full-pipeline.md
  ✗ prompts/horus/migrate.md (NOT FOUND)

Scanning: .claude-plugin/plugin.json
  ✓ skills/terraform-validate/SKILL.md
  ...

2c: Validate Skill Registration Parity

Ensure every skill from Step 0a is registered on every platform:

1. Canonical skill list: <skill-names> from Step 0a
2. Gemini extension: Parse .gemini/extensions/devops/gemini-extension.json → skills[] array. Each entry should have a name matching a skill directory and a path that resolves to an existing SKILL.md
3. Codex: Parse AGENTS.md for $<skill-name> references
4. Antigravity: Check .agents/skills/ — note that Antigravity registers agents-as-skills (horus/zeus) separately from shared skills (symlinked by setup). Only validate that the setup script would create symlinks for all skills.
5. Claude Plugin: Skills are auto-discovered from skills/ at runtime — verify the skills/ directory is in package.json files array
6. docs/PROJECT.md: Check skill tables list all skills

Report per-platform registration matrix:

Skill Registration Parity
===========================
                          Claude  Codex  Gemini  Antigravity  PROJECT.md
terraform-validate          ✓       ✓      ✓         ✓           ✓
terraform-security          ✓       ✓      ✓         ✓           ✓
release-validate            ✓       ✓      ✗         ✓           ✓
                                          ↑ missing from gemini-extension.json

2d: Validate Pipeline Registration Parity

Ensure all pipelines are documented consistently across platforms:

1. Canonical pipeline list: Horus pipelines from prompts/horus/*.md and Zeus pipelines from prompts/zeus/*.md (from Step 0b)
2. CLAUDE.md: Check command table lists all pipelines with correct *command names
3. docs/PROJECT.md: Check pipeline tables match
4. Gemini TOML commands: Check .gemini/commands/devops/pipelines/ has a TOML file for each pipeline. Naming convention: horus-<name>.toml, zeus-<name>.toml (with full-pipeline.md → horus-full.toml as a special case)
5. Antigravity workflows: Check that scripts/setup/setup-antigravity.sh references all expected workflow names
6. AGENTS.md (Codex): Check command table lists all *command entries
7. GEMINI.md: Check command table lists all *command entries

TOML count validation:

toml_count=$(find .gemini/commands/devops -name "*.toml" | wc -l)
expected_count=$((2 + horus_pipeline_count + zeus_pipeline_count + shared_prompt_count_that_have_tomls))

The 2 accounts for agents/horus.toml and agents/zeus.toml.

2e: Validate Gemini Extension JSON Integrity

Parse .gemini/extensions/devops/gemini-extension.json and verify:

• All agents[].definition paths resolve to existing files
• All skills[].path values resolve to existing SKILL.md files
• skills[] count matches <skill-count> from Step 0a
• Version matches VERSION file (already checked in Phase 1a, but verify here too)

Phase 3: Setup Script Smoke Testing

3a: Create Sandboxed Test Environment

TEST_DIR=$(mktemp -d)
trap 'rm -rf "$TEST_DIR"' EXIT  # Always clean up, even on failure

# Simulate a target repository with the skill pack installed
mkdir -p "$TEST_DIR/target-repo"
cp -r . "$TEST_DIR/target-repo/devops-ai-skill"
cd "$TEST_DIR/target-repo"

The trap ensures cleanup happens even if a step fails partway through.

3b: Test setup.sh for Each Platform

Run each platform setup script in the sandboxed directory. Use name-based assertions — verify each specific expected symlink exists, not just a count.

Claude Code (--claude):

bash devops-ai-skill/scripts/setup.sh --claude

Verify:

• .claude/agents/horus.md and .claude/agents/zeus.md are symlinks pointing to valid targets
• Every skill from <skill-names> has a symlink at .claude/skills/<skill-name>/ that resolves to a directory containing SKILL.md
• No [fail] in output
• Exit code is 0

Codex (--codex):

bash devops-ai-skill/scripts/setup.sh --codex

Verify:

• Every skill from <skill-names> has a symlink at .codex/skills/<skill-name>/
• AGENTS.md exists (or is symlinked)
• No [fail] in output
• Exit code is 0

Gemini (--gemini):

bash devops-ai-skill/scripts/setup.sh --gemini

Verify:

• .gemini/commands/devops/ directory has TOML command files
• .gemini/agents/ has agent definition files
• No [fail] in output
• Exit code is 0

Antigravity (--antigravity):

bash devops-ai-skill/scripts/setup.sh --antigravity

Verify:

• Every skill from <skill-names> has a symlink at .agents/skills/<skill-name>/
• Workflow symlinks exist in .agents/workflows/ for all expected pipeline names
• .agents/rules/devops.md exists
• No [fail] in output
• Exit code is 0

3c: Test Idempotency

Run each setup command a second time and verify:

• All items show [skip] (not [link] or [fail])
• Exit code is still 0
• No duplicate symlinks or files created
• The count of [skip] messages matches the expected total (skills + workflows where applicable)

3d: Test Uninstall

bash devops-ai-skill/scripts/setup.sh --uninstall

Verify:

• All symlinks created in 3b are removed
• No [fail] in output
• Original target repo files (if any) are preserved
• Directories created by setup (.claude/skills/, .codex/skills/, etc.) may remain empty — this is expected

3e: Cleanup

Handled automatically by the trap in Step 3a. If running manually, ensure:

rm -rf "$TEST_DIR"

Phase 4: Skill Fixture Suite Runs (v2.0.0)

Iterate every tests/*/run-fixtures.sh and aggregate PASS/FAIL counts per suite. Delegates to scripts/run_all_fixtures.sh:

bash skills/release-validate/scripts/run_all_fixtures.sh > /tmp/fixtures.json

Output JSON shape: {suites: [...], totalPass, totalFail, totalSuites, verdict}.

Gate: any suite with verdict: FAIL → overall verdict FAIL → release blocked. Empty list (no fixture runners present) → verdict: NONE (not a failure, but flagged in the report).

Phase 5: Shell Portability Static Checks (v2.0.0)

Static lint of every .sh file under skills/ and scripts/ for cross-OS compatibility issues. Delegates to scripts/check_shell_portability.sh:

bash skills/release-validate/scripts/check_shell_portability.sh . > /tmp/portability.json

Rules:

#	Rule	Severity	Why
1	#!/usr/bin/env bash shebang	WARN	/bin/bash and /bin/sh paths vary across distros
2	No declare -A	ERROR	macOS bash 3.2 lacks associative arrays — caught by us in v1.13.1
3	No mapfile / readarray	ERROR	Bash 4+ only
4	No sed -i without backup suffix	WARN	BSD vs GNU divergence; portable form is sed -i.bak ...
5	No readlink with -f	WARN	BSD lacks -f; use cd && pwd pattern

Gate: any ERROR → release blocked. WARN-only → release allowed but flagged.

Phase 6: Cross-Repo-Style Coverage (v1.15.0)

For each "Kustomize-touching skill" listed in references/repo-style-matrix.md, verify fixtures exist for each declared repo style (kustomize-argocd, helm-only, mixed). Delegates to scripts/check_repo_style_coverage.sh:

bash skills/release-validate/scripts/check_repo_style_coverage.sh --repo-root .

Output JSON: {scanned, matrix: [...], missing: ["skill:style"], coveragePct, verdict}.

Convention: any pre-v1.15.0 fixture under tests/<skill>/fixtures/ counts as kustomize-argocd-style (matches the current reality). Other styles must have a directory whose name contains the style keyword (e.g. tests/<skill>/fixtures/helm-only-basic/).

Gate: WARN-only. Coverage gaps are surfaced but do not block release — they're follow-up work, not regressions. To upgrade a row to FAIL, edit the matrix and add a new column with required (no code change needed).

Phase 7: Cross-AI-Tool Registration Parity (v1.15.0)

For every Zeus command (file under prompts/zeus/), verify it is registered across all 4 AI-tool surfaces. Delegates to scripts/check_ai_tool_parity.sh:

bash skills/release-validate/scripts/check_ai_tool_parity.sh --repo-root . --all

Checks per command:

#	Where	Required
1	prompts/zeus/<cmd>.md exists	yes
2	.gemini/commands/devops/pipelines/zeus-<cmd>.toml exists	yes
3	CLAUDE.md mentions *<cmd>	yes
4	AGENTS.md mentions *<cmd>	yes
5	GEMINI.md mentions *<cmd>	yes
6	docs/PROJECT.md mentions *<cmd>	yes

Gate: FAIL on any gap. Every Zeus command must surface identically across Claude Code, OpenAI Codex CLI, Google Gemini CLI, and Google Antigravity. Strengthens Phase 2 (cross-platform link validation), which only checks file references, not command registration.

Phase 8: Release Artifact Generation (v2.0.0)

Combine Phases 4 + 5 + 6 + 7 outputs into a single Markdown artifact at docs/reports/release-validate/<version>/RELEASE-CHECK.md. Suitable verbatim as the body of gh release create --notes-file ... or as the npm publish README excerpt.

VERSION=$(cat VERSION) \
  FIXTURES_JSON=/tmp/fixtures.json \
  PORTABILITY_JSON=/tmp/portability.json \
  REPO_STYLE_JSON=/tmp/repo-style.json \
  AI_TOOL_PARITY_JSON=/tmp/ai-parity.json \
  OUTPUT="docs/reports/release-validate/${VERSION}/RELEASE-CHECK.md" \
  bash skills/release-validate/scripts/render_release_artifact.sh

(REPO_STYLE_JSON and AI_TOOL_PARITY_JSON are optional — when absent, the artifact renders Phase 6 + 7 verdicts as SKIPPED, matching the v1.14.0 caller contract for backwards compatibility.)

The artifact contains:

• Overall verdict (PASS / WARN / FAIL — worst of all phases)
• Per-phase verdict table (4 rows: Phases 4–7)
• Per-suite breakdown (Phase 4)
• Per-issue table (Phase 5)
• Phase 6 + 7 detail sections (coverage metrics and registration gaps)
• "How to interpret" + "Next steps" sections

The renderer also prints a one-line summary to stdout for CI logs:

release-validate v1.15.0: PASS (fixtures=OK, portability=OK, repoStyle=OK, parity=OK)

Gate: if overall verdict is FAIL, the renderer exits 1 — pnpm release should refuse to proceed.

Top-Level Orchestrator (v1.15.0)

scripts/release_check.sh at the repo root runs every phase in sequence and renders the artifact. Used by both operators (manual pre-release sanity check) and CI (.github/workflows/release.yml).

pnpm release:check    # or: bash scripts/release_check.sh

The orchestrator:

1. Creates docs/reports/release-validate/<version>/
2. Runs Phase 4 → fixtures.json
3. Runs Phase 5 → portability.json
4. Runs Phase 6 → repo-style.json
5. Runs Phase 7 → ai-parity.json
6. Invokes the renderer → RELEASE-CHECK.md
7. Exits 1 if overall verdict is FAIL, else 0

CI uploads docs/reports/release-validate/ as a workflow artifact named release-check-<tag> so the rendered Markdown is downloadable from the Actions run page.

Output Format

Present results as a release readiness report following the standard format from prompts/shared/report-format.md:

Release Validation Report — v1.6.0
====================================

Phase 1: Package Release Checks
  ✓ Version consistency (5/5 files match: 1.6.0)
  ✓ CHANGELOG.md has entry for 1.6.0 (dated 2026-04-07)
  ✓ npm pack includes all 9 skills, 18 prompts, 2 agents
  ✓ Git: clean tree, on main, tag v1.6.0 available
  ✓ Badge accuracy: SKILLS-9, PIPELINES-14, AGENTS-2 (all correct)
  ⚠ Version-tag history: v1.3.0 has no git tag (WARN)

Phase 2: Cross-Platform Links
  ✓ CLAUDE.md: 12/12 links valid
  ✓ AGENTS.md: 8/8 links valid
  ✓ GEMINI.md: 10/10 links valid
  ✓ Antigravity: 27/27 links valid
  ✓ Skill parity: 9/9 skills registered on all platforms
  ✓ Pipeline parity: 14/14 pipelines registered on all platforms
  ✓ Gemini extension: 9 skills, 2 agents, all paths valid

Phase 3: Setup Script Smoke Tests
  ✓ --claude: 11 items linked (2 agents + 9 skills), idempotent, uninstall clean
  ✓ --codex: 9 items linked (9 skills), idempotent, uninstall clean
  ✓ --gemini: agents + commands linked, idempotent, uninstall clean
  ✓ --antigravity: 27 items linked (9 skills + 18 workflows), idempotent, uninstall clean

====================================
Result: PASS — 0 errors, 1 warning
Ready for: pnpm release

When issues are found, append an actionable fix list:

====================================
Result: 3 issues found — fix before release
  1. [Phase 1a] marketplace.json plugins[0].version is 1.5.0 (expected 1.6.0)
     Fix: pnpm version:bump 1.6.0
  2. [Phase 2c] release-validate missing from gemini-extension.json
     Fix: Add skill entry to .gemini/extensions/devops/gemini-extension.json
  3. [Phase 3b] --gemini setup exit code 1
     Fix: Check .gemini/agents/ source directory exists

Scope Flags

The user can run specific phases:

• --all (default): Run all 3 phases
• --package: Phase 1 only (version, changelog, npm, git, badges)
• --links: Phase 2 only (cross-platform reference and parity validation)
• --smoke: Phase 3 only (setup script smoke tests)
• --quick: Phase 1 + 2 only (skip slow smoke tests — recommended for iteration)

Auto-Fix Capabilities

This skill can auto-fix:

• Badge counts: Update shield.io URLs in README files to match actual counts
• Hardcoded skill counts: Update prose text ("9 skills", "9 個 Skills") across docs
• CHANGELOG placeholder: Warn that the placeholder text from version-bump.sh needs replacing

This skill CANNOT auto-fix (requires user judgment):

• Missing platform registrations (each platform has different formats)
• Broken cross-references (the fix depends on whether the target was renamed or deleted)
• Setup script failures (root cause must be diagnosed)
• Git state issues (uncommitted changes may be intentional)

Error Handling

Discovery Failures

• Missing VERSION file: Report "Cannot determine version — VERSION file not found" and stop. This is fatal — all other checks depend on knowing the version.
• Missing scripts: If version-consistency.sh or setup.sh is missing, report the missing script and skip that check. Continue with remaining checks.
• pnpm not installed: Fall back to npm pack --dry-run for package content check. If neither is available, skip Phase 1c and report as WARN.
• python3 not installed: version-consistency.sh uses python3 for JSON parsing. If unavailable, parse JSON manually using grep and jq (if available) or skip with WARN.

Phase 2 Failures

• Entry file missing: If a platform's entry file doesn't exist (e.g., GEMINI.md deleted), report as FAIL for that platform and continue with others.
• JSON parse error: If gemini-extension.json or plugin.json is malformed JSON, report the parse error and skip that file's checks.
• TOML parse error: If a Gemini command TOML is malformed, report the file and error but continue checking others.

Smoke Test Failures

• mktemp fails: Skip Phase 3, report that smoke testing requires a writable temp directory.
• setup.sh crashes mid-run: Capture both stdout and stderr, report the failure with the error output, continue with remaining platforms.
• Symlink target doesn't exist: Report which source file is missing — this usually indicates a file was deleted from skills/ or prompts/ but the setup script still tries to link it.
• Permission denied: Report and skip. May happen in CI environments with restricted temp directories.

Edge Cases

• New skill added but not registered: Phase 2c catches this — the parity matrix shows exactly which platforms are missing the registration.
• Pipeline renamed but old name in docs: Phase 2d catches stale references in command tables.
• Version already tagged: Phase 1d reports this — user must either bump version or delete the existing tag.
• marketplace.json dual-version mismatch: Phase 1a checks both metadata.version and plugins[0].version — they can diverge if only one is updated manually.
• CHANGELOG placeholder not replaced: Phase 1b checks for the literal text (describe changes here) left by version-bump.sh scaffolding.
• Git ahead/behind remote: Phase 1d reports both directions — unpushed commits and unpulled changes — with counts.

Dry-Run Support

This skill is read-only for Phases 1 and 2 — no files are modified.

Phase 3 creates a temporary directory via mktemp but does not modify the source repository. All symlinks and files are created inside the isolated sandbox. The trap EXIT ensures cleanup even on failure.

Rollback Strategy

• Phases 1 and 2: No rollback needed — these are pure read operations.
• Phase 3: The sandbox is in a temp directory and cleaned up automatically. If cleanup fails (e.g., process killed), the OS will clean /tmp eventually. No source files are touched.
• Auto-fix (badge/count updates): If auto-fix is applied, show the diff before writing. Suggest git checkout -- <file> to revert if the fix is wrong.

Dependencies

• scripts/version-consistency.sh — Reused for version checks (Phase 1a)
• scripts/setup.sh — Tested in Phase 3
• prompts/shared/report-format.md — Output format reference
• pnpm or npm — For pack dry-run (Phase 1c)
• git — For state checks (Phase 1d, 1f)
• python3 — Used by version-consistency.sh for JSON parsing
• mktemp — For sandboxed smoke tests (Phase 3)