@lifeaitools/rdc-skills 0.35.23 → 0.35.25

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "rdc",
-  "version": "0.35.23",
+  "version": "0.35.25",
   "description": "RDC typed-agent dispatch skill suite for Claude Code — plan, build, review, overnight unattended builds with work-item tracking and TDD enforcement.",
   "author": {
     "name": "LIFEAI",

package/.github/workflows/self-test.yml

@@ -1,29 +1,29 @@
-name: self-test
-
-on:
-  push:
-    branches: ['**']
-    tags: ['v*']
-  pull_request:
-
-jobs:
-  tier1:
-    name: Tier 1 static lint
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
-      - name: Setup Node 20
-        uses: actions/setup-node@v4
-        with:
-          node-version: '20'
-
-      - name: Run rdc:self-test --strict (Tier 1)
-        # Tier 2 requires headless claude + clauth daemon — not available in standard CI.
-        # Tier 1 catches the backtick-drift + frontmatter class of bugs.
-        run: node scripts/self-test.mjs --strict
-
+name: self-test
+
+on:
+  push:
+    branches: ['**']
+    tags: ['v*']
+  pull_request:
+
+jobs:
+  tier1:
+    name: Tier 1 static lint
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Node 20
+        uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+
+      - name: Run rdc:self-test --strict (Tier 1)
+        # Tier 2 requires headless claude + clauth daemon — not available in standard CI.
+        # Tier 1 catches the backtick-drift + frontmatter class of bugs.
+        run: node scripts/self-test.mjs --strict
+
   tier2-gate:
     name: Tier 2 behavioral (tag releases only)
     runs-on: ubuntu-latest
@@ -32,14 +32,14 @@ jobs:
     env:
       ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
     steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
-      - name: Setup Node 20
-        uses: actions/setup-node@v4
-        with:
-          node-version: '20'
-
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Node 20
+        uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+
       - name: Install claude CLI
         if: env.ANTHROPIC_API_KEY != ''
         run: npm install -g @anthropic-ai/claude-code

package/MANIFEST.md

@@ -1,224 +1,224 @@
-# RDC Skills Library — Manifest
-
-**Created:** 2026-04-13  
-**Last updated:** 2026-05-01  
-**Source:** Regen Root project `.claude/skills/user/`  
-**Files:** 31 skill files, 10 agent guides, command mirrors, and support scripts  
-**Size:** ~480K  
-
-## Generic Conversion Summary
-
-All skills have been converted from project-specific to generic/portable:
-
-### Path Replacements
-- `C:/Dev/regen-root/` → `{PROJECT_ROOT}/`
-- Hardcoded app names (rdc-marketing-engine, brand-studio, prt, etc.) → generic examples
-- Specific Supabase project IDs → generic table references
-- Developer-specific paths → variables
-
-### TDD Enhancements
-- Added mandatory post-wave test gate to rdc:build
-- Updated rdc:review with test coverage delta reporting
-- Added test-first requirement to all agent skills
-- TDD enforcement: red → implement → green cycle
-
-### Preserved Elements
-- All logic and procedures (100%)
-- All safety rules
-- All RPC patterns
-- Branch strategy (development → production)
-- Work item lifecycle
-- Escalation protocols
-- Unattended mode patterns
-- Status block format
-
-## Files Created
-
-```
-/c/Dev/rdc-skills/
-├── README.md
-├── MANIFEST.md (this file)
-└── skills/
-    ├── build/SKILL.md        [+TDD enforcement, +verify gate]
-    ├── plan/SKILL.md
-    ├── preplan/SKILL.md
-    ├── review/SKILL.md       [+test coverage delta, +verify gate]
-    ├── overnight/SKILL.md
-    ├── fixit/SKILL.md
-    ├── status/SKILL.md
-    ├── report/SKILL.md
-    ├── co-develop/SKILL.md   [peer-aware Claude/Codex co-development]
-    ├── collab/SKILL.md       [claude.ai ↔ Claude Code relay]
-    ├── handoff/SKILL.md
-    ├── prototype/SKILL.md
-    ├── workitems/SKILL.md
-    ├── design/SKILL.md       [RDC/Studio design, Palette Library, Rampa CLI]
-    ├── help/SKILL.md
-    ├── regen-media/SKILL.md [primary image generation + Regen Media upload]
-    ├── deploy/SKILL.md       [Coolify deployment ops]
-    ├── release/SKILL.md      [atomic version release]
-    ├── self-test/SKILL.md    [static lint + tier2 smoke tests]
-    └── watch/SKILL.md        [session log + browser viewer]
-└── guides/agents/     (agent-only playbooks — dispatched by rdc:build, not user-invocable)
-    ├── frontend.md
-    ├── backend.md
-    ├── cs2.md
-    ├── data.md
-    ├── design.md
-    ├── infrastructure.md
-    ├── content.md
-    ├── viz.md
-    ├── setup.md
-    └── verify.md
-```
-
-**Note (v0.7.9):** Skills migrated from flat `rdc-<name>.md` files to `<name>/SKILL.md` subdirectories. Plugin namespace is `rdc`, so composed skill names are `rdc:<name>` — matching the frontmatter. This eliminates the duplicate `rdc:rdc-*` registration that occurred when directories were named `rdc-<name>/`.
-
-## Skill Categories
-
-### Orchestration (8)
-- rdc:build — dispatch typed agents in waves
-- rdc:plan — architecture + epic/tasks
-- rdc:preplan — research before planning
-- rdc:review — quality gate
-- rdc:overnight — unattended build supervisor
-- rdc:fixit — quick-fix bypass
-- rdc:status — project dashboard
-- rdc:report — nightly report
-
-### DevOps (3)
-- rdc:deploy — Coolify deployment ops (deploy, new, diagnose, audit)
-- rdc:release — atomic version release with git tag + changelog
-- rdc:watch — session log + browser viewer for Claude activity
-
-### Quality (1)
-- rdc:self-test — static lint (tier1) + smoke tests (tier2)
-
-### Collaboration (1)
-- rdc:co-develop — peer-aware Claude/Codex co-development over clauth
-
-### Agent Guides (10, in `guides/agents/`) — [agent-only — dispatched by rdc:build, not user-invocable]
-- frontend — React, UI, Tailwind
-- backend — API, database, auth
-- data — migrations, schema
-- design — brand, tokens
-- infrastructure — deployment, CI/CD
-- content — copy, messaging
-- cs2 — paradigm-level work
-- viz — visualizations
-- setup — project scan + .rdc/config.json generation
-- verify — evidence-before-claims verification gate
-
-### Utility (1)
-- rdc:collab — bidirectional claude.ai ↔ Claude Code relay via file transport
-- rdc:channel-formatter — channel-native formatting and content repurposing,
-  including article-to-post, social-pack, campaign-pack, exec-pack, and
-  launch-pack outputs
-- rdc:regen-media — primary image generation and Regen Media upload workflow,
-  defaulting to local keyless Codex `gpt-image-2` via built-in `image_gen`
-
-### Bridge (3)
-- rdc:handoff — planning → work items
-- rdc:prototype — JSX prototypes
-- rdc:workitems — work item CRUD
-
-### Design (1)
-- rdc:design — Studio-aware design workflow with token, Palette Library, and Rampa CLI context
-
-### Reference (1)
-- rdc:help — skill index
-
-## Usage Instructions
-
-1. **Clone or reference** the `/c/Dev/rdc-skills/` directory
-2. **Update path variables** in all skills for your project
-3. **Create project guides** (agent-bootstrap.md, frontend.md, etc.)
-4. **Set up database** (work_items RPC, prototype_registry, design_context)
-5. **Invoke skills** via your agent framework
-   - Interactive: `/rdc:build <epic-id>`
-   - Unattended: `/rdc:overnight`
-   - Quality: `/rdc:review --unattended`
-   - Content repurposing: `/rdc:channel-formatter social-pack article.md`
-
-## Channel Formatter Content Packs
-
-`rdc:channel-formatter` can strictly format copy for one channel or repurpose a
-long source into a channel-native output set.
-
-Examples:
-
-```bash
-rdc:channel-formatter linkedin article.md
-rdc:channel-formatter twitter-thread article.md
-rdc:channel-formatter social-pack article.md
-rdc:channel-formatter campaign-pack report.md
-```
-
-Use `rdc:convert` for actual Office/Markdown file conversion, `rdc:brochure`
-for HTML/folder/URL PDF rendering, and `rdc:brochurify` for Brochurify
-orchestration. Channel formatter writes and repurposes text; it does not produce
-binary Word or PDF artifacts.
-
-## Key Differences from Source
-
-### Added: TDD Enforcement
-
-**rdc:build (Step 8):**
-```bash
-# Post-wave test gate (mandatory)
-cd packages/<name> && npx vitest run 2>&1 | tail -20
-```
-All tests must pass before next wave.
-
-**All agent skills:**
-```
-TDD REQUIREMENT: Write tests FIRST for new functions/modules.
-Run: npx vitest run packages/<name>
-```
-
-**rdc:review (Step 3):**
-```bash
-# Check test coverage delta
-git diff origin/main...HEAD -- packages/*/src/ | grep -c "^+"
-git diff origin/main...HEAD -- packages/*/test* | grep -c "^+"
-```
-Flags any package with implementation added but 0 new tests.
-
-**rdc:review report format:**
-```
-| Package | Tests | Pass/Fail | New Tests Added | Issues |
-```
-
-### Preserved: All Logic
-
-- Work item RPC patterns (never raw SQL)
-- Agent type classification
-- Wave-based parallelization
-- Escalation protocols
-- Unattended mode patterns
-- Status block formatting
-- Safety rules (100%)
-- Branch strategy
-
-## Deployment
-
-These skills are ready to use in any project that has:
-
-1. A monorepo with `packages/`, `apps/`, or `sites/` structure
-2. A database (Supabase, PostgreSQL, etc.) with work_items RPC functions
-3. Capability to run agents (Claude Code, subagents, or CLI)
-4. Git-based version control with at least 2 branches (dev + prod)
-5. A credential management system (clauth daemon or equivalent)
-
-## Questions?
-
-- **Specific project integration:** Refer to your project's infrastructure documentation
-- **Skill logic:** All skills are pure logic extracted from Regen Root; no project-specific magic
-- **TDD requirements:** Built into every skill; red → implement → green is mandatory
-- **Database schema:** You define; skills use generic `work_items` table and RPC patterns
-
----
-
-**Manifest created:** 2026-04-13  
-**Source project:** Regen Root (github.com/LIFEAI/regen-root)  
-**Genericization:** Complete — all hardcoded references removed, all logic preserved
+# RDC Skills Library — Manifest
+
+**Created:** 2026-04-13  
+**Last updated:** 2026-05-01  
+**Source:** Regen Root project `.claude/skills/user/`  
+**Files:** 31 skill files, 10 agent guides, command mirrors, and support scripts  
+**Size:** ~480K  
+
+## Generic Conversion Summary
+
+All skills have been converted from project-specific to generic/portable:
+
+### Path Replacements
+- `C:/Dev/regen-root/` → `{PROJECT_ROOT}/`
+- Hardcoded app names (rdc-marketing-engine, brand-studio, prt, etc.) → generic examples
+- Specific Supabase project IDs → generic table references
+- Developer-specific paths → variables
+
+### TDD Enhancements
+- Added mandatory post-wave test gate to rdc:build
+- Updated rdc:review with test coverage delta reporting
+- Added test-first requirement to all agent skills
+- TDD enforcement: red → implement → green cycle
+
+### Preserved Elements
+- All logic and procedures (100%)
+- All safety rules
+- All RPC patterns
+- Branch strategy (development → production)
+- Work item lifecycle
+- Escalation protocols
+- Unattended mode patterns
+- Status block format
+
+## Files Created
+
+```
+/c/Dev/rdc-skills/
+├── README.md
+├── MANIFEST.md (this file)
+└── skills/
+    ├── build/SKILL.md        [+TDD enforcement, +verify gate]
+    ├── plan/SKILL.md
+    ├── preplan/SKILL.md
+    ├── review/SKILL.md       [+test coverage delta, +verify gate]
+    ├── overnight/SKILL.md
+    ├── fixit/SKILL.md
+    ├── status/SKILL.md
+    ├── report/SKILL.md
+    ├── co-develop/SKILL.md   [peer-aware Claude/Codex co-development]
+    ├── collab/SKILL.md       [claude.ai ↔ Claude Code relay]
+    ├── handoff/SKILL.md
+    ├── prototype/SKILL.md
+    ├── workitems/SKILL.md
+    ├── design/SKILL.md       [RDC/Studio design, Palette Library, Rampa CLI]
+    ├── help/SKILL.md
+    ├── regen-media/SKILL.md [primary image generation + Regen Media upload]
+    ├── deploy/SKILL.md       [Coolify deployment ops]
+    ├── release/SKILL.md      [atomic version release]
+    ├── self-test/SKILL.md    [static lint + tier2 smoke tests]
+    └── watch/SKILL.md        [session log + browser viewer]
+└── guides/agents/     (agent-only playbooks — dispatched by rdc:build, not user-invocable)
+    ├── frontend.md
+    ├── backend.md
+    ├── cs2.md
+    ├── data.md
+    ├── design.md
+    ├── infrastructure.md
+    ├── content.md
+    ├── viz.md
+    ├── setup.md
+    └── verify.md
+```
+
+**Note (v0.7.9):** Skills migrated from flat `rdc-<name>.md` files to `<name>/SKILL.md` subdirectories. Plugin namespace is `rdc`, so composed skill names are `rdc:<name>` — matching the frontmatter. This eliminates the duplicate `rdc:rdc-*` registration that occurred when directories were named `rdc-<name>/`.
+
+## Skill Categories
+
+### Orchestration (8)
+- rdc:build — dispatch typed agents in waves
+- rdc:plan — architecture + epic/tasks
+- rdc:preplan — research before planning
+- rdc:review — quality gate
+- rdc:overnight — unattended build supervisor
+- rdc:fixit — quick-fix bypass
+- rdc:status — project dashboard
+- rdc:report — nightly report
+
+### DevOps (3)
+- rdc:deploy — Coolify deployment ops (deploy, new, diagnose, audit)
+- rdc:release — atomic version release with git tag + changelog
+- rdc:watch — session log + browser viewer for Claude activity
+
+### Quality (1)
+- rdc:self-test — static lint (tier1) + smoke tests (tier2)
+
+### Collaboration (1)
+- rdc:co-develop — peer-aware Claude/Codex co-development over clauth
+
+### Agent Guides (10, in `guides/agents/`) — [agent-only — dispatched by rdc:build, not user-invocable]
+- frontend — React, UI, Tailwind
+- backend — API, database, auth
+- data — migrations, schema
+- design — brand, tokens
+- infrastructure — deployment, CI/CD
+- content — copy, messaging
+- cs2 — paradigm-level work
+- viz — visualizations
+- setup — project scan + .rdc/config.json generation
+- verify — evidence-before-claims verification gate
+
+### Utility (1)
+- rdc:collab — bidirectional claude.ai ↔ Claude Code relay via file transport
+- rdc:channel-formatter — channel-native formatting and content repurposing,
+  including article-to-post, social-pack, campaign-pack, exec-pack, and
+  launch-pack outputs
+- rdc:regen-media — primary image generation and Regen Media upload workflow,
+  defaulting to local keyless Codex `gpt-image-2` via built-in `image_gen`
+
+### Bridge (3)
+- rdc:handoff — planning → work items
+- rdc:prototype — JSX prototypes
+- rdc:workitems — work item CRUD
+
+### Design (1)
+- rdc:design — Studio-aware design workflow with token, Palette Library, and Rampa CLI context
+
+### Reference (1)
+- rdc:help — skill index
+
+## Usage Instructions
+
+1. **Clone or reference** the `/c/Dev/rdc-skills/` directory
+2. **Update path variables** in all skills for your project
+3. **Create project guides** (agent-bootstrap.md, frontend.md, etc.)
+4. **Set up database** (work_items RPC, prototype_registry, design_context)
+5. **Invoke skills** via your agent framework
+   - Interactive: `/rdc:build <epic-id>`
+   - Unattended: `/rdc:overnight`
+   - Quality: `/rdc:review --unattended`
+   - Content repurposing: `/rdc:channel-formatter social-pack article.md`
+
+## Channel Formatter Content Packs
+
+`rdc:channel-formatter` can strictly format copy for one channel or repurpose a
+long source into a channel-native output set.
+
+Examples:
+
+```bash
+rdc:channel-formatter linkedin article.md
+rdc:channel-formatter twitter-thread article.md
+rdc:channel-formatter social-pack article.md
+rdc:channel-formatter campaign-pack report.md
+```
+
+Use `rdc:convert` for actual Office/Markdown file conversion, `rdc:brochure`
+for HTML/folder/URL PDF rendering, and `rdc:brochurify` for Brochurify
+orchestration. Channel formatter writes and repurposes text; it does not produce
+binary Word or PDF artifacts.
+
+## Key Differences from Source
+
+### Added: TDD Enforcement
+
+**rdc:build (Step 8):**
+```bash
+# Post-wave test gate (mandatory)
+cd packages/<name> && npx vitest run 2>&1 | tail -20
+```
+All tests must pass before next wave.
+
+**All agent skills:**
+```
+TDD REQUIREMENT: Write tests FIRST for new functions/modules.
+Run: npx vitest run packages/<name>
+```
+
+**rdc:review (Step 3):**
+```bash
+# Check test coverage delta
+git diff origin/main...HEAD -- packages/*/src/ | grep -c "^+"
+git diff origin/main...HEAD -- packages/*/test* | grep -c "^+"
+```
+Flags any package with implementation added but 0 new tests.
+
+**rdc:review report format:**
+```
+| Package | Tests | Pass/Fail | New Tests Added | Issues |
+```
+
+### Preserved: All Logic
+
+- Work item RPC patterns (never raw SQL)
+- Agent type classification
+- Wave-based parallelization
+- Escalation protocols
+- Unattended mode patterns
+- Status block formatting
+- Safety rules (100%)
+- Branch strategy
+
+## Deployment
+
+These skills are ready to use in any project that has:
+
+1. A monorepo with `packages/`, `apps/`, or `sites/` structure
+2. A database (Supabase, PostgreSQL, etc.) with work_items RPC functions
+3. Capability to run agents (Claude Code, subagents, or CLI)
+4. Git-based version control with at least 2 branches (dev + prod)
+5. A credential management system (clauth daemon or equivalent)
+
+## Questions?
+
+- **Specific project integration:** Refer to your project's infrastructure documentation
+- **Skill logic:** All skills are pure logic extracted from Regen Root; no project-specific magic
+- **TDD requirements:** Built into every skill; red → implement → green is mandatory
+- **Database schema:** You define; skills use generic `work_items` table and RPC patterns
+
+---
+
+**Manifest created:** 2026-04-13  
+**Source project:** Regen Root (github.com/LIFEAI/regen-root)  
+**Genericization:** Complete — all hardcoded references removed, all logic preserved

package/RELEASE.md

@@ -69,21 +69,63 @@ The public `/health` version must equal the released npm version and its skill
 count must match the packaged catalog. The local installer never restarts or
 deploys that endpoint.
 
-The Vultr host checks out this repository at `/srv/regen/rdc-skills` — it runs
-straight from the checkout (no `npm install`), so `/health`'s `git_sha` comes
-from `bin/rdc-skills-mcp.mjs`'s runtime `git rev-parse HEAD` fallback, NOT a
-stamped `git-sha.json` (that file is `.gitignore`d on purpose — it's a
-pack-time-only artifact for npm-installed copies; see the comment above its
-`.gitignore` entry). `install-systemd.sh` alone does not restart the running
-process. After pulling `master`:
+The Vultr host checks out this repository at `/srv/regen/rdc-skills` and runs the
+server straight from that checkout, with production dependencies installed there
+by `npm ci`. `/health`'s `git_sha` comes from `bin/rdc-skills-mcp.mjs`'s runtime
+`git rev-parse HEAD`, resolved once when the process starts — NOT a stamped
+`git-sha.json` (that file is `.gitignore`d on purpose — it's a pack-time-only
+artifact for npm-installed copies; the updater deletes one if it appears in the
+checkout). `version`, by contrast, is re-read from `package.json` on every
+request, so **only `git_sha` proves the process restarted onto a release.**
+
+**The host follows release tags by itself.** `rdc-skills-mcp-update.timer` runs
+the updater ([`deploy/update-from-tag.sh`](deploy/update-from-tag.sh), installed
+at `/usr/local/libexec/rdc-skills/` so no rollback can delete it) every five
+minutes. The service is only ever restarted onto a **release**: the newest plain
+`vX.Y.Z` tag merged into `master`, or the last release the updater proved serving.
+Untagged commits, pre-releases and tags off `master` are never served — a checkout
+pulled ahead to untagged `master` is moved back to the tag. The updater reinstalls
+dependencies and unit files only when they differ from the last install, restarts
+the MCP, and succeeds only when `/health` reports the release's version **and**
+commit. "Current" is decided by `/health`, never by where `HEAD` is — a run that was
+interrupted converges on the next tick.
+
+A release only moves forward: a tag that is not a descendant of the proved release
+(a higher number put on an older commit) is held off, so rolling back means tagging
+a new, higher release. A release that restarts but never answers is recorded as
+failed in `/var/lib/rdc-skills-update/failed`, the proved release is served again,
+and the failed one is not retried until a newer tag exists. An install or restart
+error says nothing about the release: the proved release is served again and the
+new one is retried next run. A registry that cannot be reached is not counted
+unless it stays unreachable for that release for a whole day; any other install or
+restart error, the third time against the same release, marks it failed. To retry a release marked failed without tagging again, delete its line
+from that file on the host. A host with no history adopts the commit it is already
+serving (if that commit is on `master`) as the proved release before touching
+anything, so the first move has somewhere to go back to. Tracked changes on the
+host, or a commit made on the host, are refused and the service is left as it is.
+So pushing the tag is the whole public MCP deploy — expect it live within about
+five minutes:
 
 ```bash
-git pull --ff-only origin master
-sudo ./deploy/install-systemd.sh        # re-verify/re-enable the unit config
-sudo systemctl restart rdc-skills-mcp.service   # actually load the new commit
-curl -fsS https://rdc-skills.regendevcorp.com/health   # confirm git_sha == new HEAD
+tag=v$(node -p "require('./package.json').version")
+curl -fsS https://rdc-skills.regendevcorp.com/health   # git_sha == the line below
+git rev-list -n 1 "$tag"
+# on the host: journalctl -u rdc-skills-mcp-update.service -n 20
+```
+
+`install-systemd.sh` is the one-time host bootstrap. It installs the units and the
+updater, runs the updater once (which restarts the MCP onto the checked-out release
+and proves it), then enables the timer:
+
+```bash
+cd /srv/regen/rdc-skills
+git fetch --tags origin && git merge --ff-only v<X.Y.Z>
+sudo ./deploy/install-systemd.sh
 ```
 
+Before the timer existed the checkout moved only when someone logged in and
+pulled, and the public catalog sat on 0.35.18 through six releases.
+
 Port `3110` is reserved for this public MCP ingress. Application fleet manifests
 must not claim it.
 

package/deploy/install-systemd.sh

@@ -1,9 +1,15 @@
 #!/usr/bin/env bash
+# One-time host bootstrap for the public MCP. From then on the timer keeps the
+# checkout on the newest release tag (deploy/update-from-tag.sh).
+#
+#   cd /srv/regen/rdc-skills
+#   git fetch --tags origin && git merge --ff-only v<X.Y.Z>
+#   sudo ./deploy/install-systemd.sh
 set -euo pipefail
 
 repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
-unit_source="$repo_root/deploy/systemd/rdc-skills-mcp.service"
-unit_target="/etc/systemd/system/rdc-skills-mcp.service"
+unit_dir="/etc/systemd/system"
+libexec_dir="/usr/local/libexec/rdc-skills"
 
 if [[ "$repo_root" != "/srv/regen/rdc-skills" ]]; then
   echo "rdc-skills must be checked out at /srv/regen/rdc-skills" >&2
@@ -14,9 +20,26 @@ cd "$repo_root"
 # The MCP imports declared runtime dependencies (including express).  A Git
 # checkout alone cannot satisfy those imports after a clean host or dependency
 # cleanup, so install the committed production graph before the unit is enabled.
-npm ci --omit=dev --no-audit --no-fund
+# No install scripts: nothing the server imports needs one, and this runs as root.
+npm ci --omit=dev --no-audit --no-fund --ignore-scripts
 
-install -m 0644 "$unit_source" "$unit_target"
+install -m 0644 deploy/systemd/rdc-skills-mcp.service "$unit_dir/rdc-skills-mcp.service"
+install -m 0644 deploy/systemd/rdc-skills-mcp-update.service "$unit_dir/rdc-skills-mcp-update.service"
+install -m 0644 deploy/systemd/rdc-skills-mcp-update.timer "$unit_dir/rdc-skills-mcp-update.timer"
+install -D -m 0755 deploy/update-from-tag.sh "$libexec_dir/update-from-tag.sh"
 systemctl daemon-reload
-systemctl enable --now rdc-skills-mcp.service
-systemctl --no-pager --full status rdc-skills-mcp.service
+systemctl enable rdc-skills-mcp.service
+
+# `enable --now` does not restart a process that is already running, so an
+# existing host would keep serving the old commit. The updater restarts into the
+# checked-out release and succeeds only when /health proves it; if that release
+# does not come up it returns to the commit that was serving. The timer is
+# enabled either way — the updater's own state decides what the next run does —
+# and the installer exits with the first run's status.
+first_run=0
+systemctl start rdc-skills-mcp-update.service || first_run=$?
+systemctl enable --now rdc-skills-mcp-update.timer
+systemctl --no-pager --full status rdc-skills-mcp.service || true
+systemctl --no-pager list-timers rdc-skills-mcp-update.timer
+journalctl -u rdc-skills-mcp-update.service -n 20 --no-pager || true
+exit "$first_run"