ci(release): stop auto-overwriting repo-adapted deepworkplan dogfood [skip release]

.github/docs/WORKFLOWS.md

@@ -7,7 +7,7 @@ piloting this repo) can reason about why a job did or didn't run.
 
 | Workflow | File | Purpose |
 |----------|------|---------|
-| Auto-release | [`auto-release.yml`](../workflows/auto-release.yml) | Conventional-commit-driven release + three-way vendored-skill dogfood |
+| Auto-release | [`auto-release.yml`](../workflows/auto-release.yml) | Conventional-commit-driven release + temp install smoke + addon dogfood |
 | CI | [`ci.yml`](../workflows/ci.yml) | Frontmatter validation, shellcheck, bats tests, `setup.sh`/`context.sh` smoke, markdown link check |
 | PR review | [`pr-review.yml`](../workflows/pr-review.yml) | AI-driven pull request code review via Cursor, `ready`-label gated |
 
@@ -35,18 +35,19 @@ Sequenced steps (a-i below), all in one long-running job on `ubuntu-latest`:
 | d. Prepend CHANGELOG.md section | Merged commits since last tag become bullets |
 | e. Commit `chore(release): X.Y.Z [skip ci]` | The `[skip ci]` marker prevents CI from re-running on the release commit |
 | f. Tag `vX.Y.Z` and push | With `--follow-tags` so commit + tag land atomically |
-| g. **Dogfood — self (`deepworkplan`)** | Fetches the just-published tag via `npx --yes skills add DailybotHQ/deepworkplan-skill@vX.Y.Z --skill deepworkplan --force -y` into `.agents/skills/deepworkplan/`. Verifies `SKILL.md` `version:` equals the requested tag. Commits any diff as `chore(release): dogfood vendored deepworkplan to vX.Y.Z [skip release]`. Doubles as a live smoke test — if the release doesn't install cleanly for consumers, this step fails |
-| h. **Dogfood — dailybot** | Same pattern for `DailybotHQ/agent-skill` → `.agents/skills/dailybot/`. Only runs if the upstream tag moved. Non-interactive contract (`--yes` + `-y`) is mandatory — dropping either flag hangs the workflow indefinitely on the CLI's agent-picker prompt |
-| i. **Dogfood — ai-diff-reviewer** | Same pattern for `DailybotHQ/ai-diff-reviewer` → `.agents/skills/ai-diff-reviewer/`. Same non-interactive contract |
-| j. Create GitHub Release | Uses `gh release create` with auto-generated notes; the release notes include the dogfood commits, so downstream consumers see exactly which skills refreshed with this release |
+| g. **Smoke — published tag installs (temp dir)** | Runs `npx --yes skills add DailybotHQ/deepworkplan-skill@vX.Y.Z --skill deepworkplan --force -y` into an isolated temp directory and asserts `version:` matches. Does **not** overwrite `.agents/skills/deepworkplan/` (that copy is repo-adapted; sync via `scripts/refresh-dogfood-skill.sh`) |
+| h. **Dogfood — dailybot** | `DailybotHQ/agent-skill` → `.agents/skills/dailybot/`. Only commits if the upstream tag moved. Non-interactive contract (`--yes` + `-y`) is mandatory — dropping either flag hangs the workflow indefinitely on the CLI's agent-picker prompt |
+| i. **Dogfood — ai-diff-reviewer** | `DailybotHQ/ai-diff-reviewer` → `.agents/skills/ai-diff-reviewer/`. Same non-interactive contract |
+| j. Create GitHub Release | Uses `gh release create` with auto-generated notes; the release notes include the addon dogfood commits when those skills moved |
 
 ### Failure semantics
 
 | Failure | Behavior |
 |---------|----------|
 | Version bump script cannot read current version | **Fails** — a corrupt router SKILL.md must be fixed before any release |
-| Any `npx skills add` fails during dogfood | **Fails** — a broken upstream tag must never quietly ship inside a release |
-| Version-invariant mismatch after install (installed `SKILL.md` `version:` != requested tag) | **Fails** — refuses to commit a misrepresented dogfood snapshot |
+| Temp-dir smoke `npx skills add` fails, or installed version ≠ tag | **Fails** — the just-published tag must install cleanly for consumers |
+| Any `npx skills add` fails during addon dogfood | **Fails** — a broken upstream tag must never quietly ship inside a release |
+| Version-invariant mismatch after addon install (installed `SKILL.md` `version:` != requested tag) | **Fails** — refuses to commit a misrepresented dogfood snapshot |
 | Upstream `gh release view` fails for one of the two external skills (rate limit, transient outage) | **Fails** — refuses to cut a release whose dogfood snapshot cannot resolve upstream |
 | Head commit already `chore(release):` OR carries `[skip release]` | Whole workflow skips (loop guard) |
 
@@ -56,8 +57,8 @@ Both `npx --yes` (accepts npm's proceed-with-install prompt) AND `-y` (accepts
 the `skills` CLI's own "which agent picker?" prompt) are required in the
 non-TTY GitHub Actions runner. Historically an upstream `ai-diff-reviewer` bug
 (fixed in v1.7.0) caused the second prompt to hang indefinitely without a
-timeout — dropping either flag will hang this workflow. The three dogfood
-steps carry both flags; the pattern is:
+timeout — dropping either flag will hang this workflow. The temp smoke step
+and both addon dogfood steps carry both flags; the pattern is:
 
 ```bash
 npx --yes skills add <owner/repo>@<tag> --skill <name> --force -y
@@ -162,8 +163,8 @@ push to main               pull_request
       ▼                          ▼
  auto-release                  ci.yml               pr-review.yml
    (release +               (validate +           (scope → labels-bootstrap
-    three-way                  smoke)              → review → gate)
-    dogfood)
+    temp smoke +               smoke)              → review → gate)
+    addon dogfood)
       │
       ▼
 GitHub Release (with dogfood commits in the notes)

.github/workflows/auto-release.yml

@@ -18,11 +18,15 @@ name: Auto-release on merge to main
 #   4. Prepends a new section to CHANGELOG.md with the merged commits as bullets
 #   5. Commits as "Dailybot Automations" with `[skip ci]` to break the loop
 #   6. Tags `vX.Y.Z` and pushes
-#   7. Dogfoods the just-published tag by fetching it via `npx skills add` into
-#      .agents/skills/deepworkplan/ — doubles as a live smoke test of the
-#      release AND keeps the in-repo dogfood copy at HEAD equal to the latest
-#      released tag with no manual refresh commits (see Step "Dogfood").
-#   8. Creates a GitHub Release with auto-generated notes
+#   7. Smoke-tests that the just-published tag installs cleanly via
+#      `npx skills add` into a TEMP directory (does NOT overwrite the
+#      repo-adapted `.agents/skills/deepworkplan/` dogfood copy — that copy
+#      is synced from `skills/deepworkplan/` via
+#      `scripts/refresh-dogfood-skill.sh` when contributors need it).
+#   8. Dogfoods the latest upstream tags of the two *addon* skills
+#      (`dailybot`, `ai-diff-reviewer`) into `.agents/skills/` — those are
+#      safe to pin blindly; deepworkplan is not.
+#   9. Creates a GitHub Release with auto-generated notes
 
 on:
   push:
@@ -188,22 +192,15 @@ jobs:
           git tag -a "$NEW_TAG" -m "Release ${NEW_VERSION}"
           git push origin "refs/tags/${NEW_TAG}"
 
-      - name: Dogfood — refresh vendored deepworkplan from the just-published tag
-        # This repo dogfoods its own skill: `.agents/skills/deepworkplan/` is a
-        # vendored copy managed via `skills-lock.json`, fetched exactly the way
-        # any consumer would install it (`npx skills add`). Refreshing right
-        # after the tag lands serves three purposes:
-        #   1. Live smoke test — proves the just-pushed tag actually installs
-        #      cleanly via `npx skills`. If the install script or the tag
-        #      layout is broken, the workflow fails HERE (loudly) instead of a
-        #      consumer discovering it silently.
-        #   2. Keeps the in-repo dogfood copy at HEAD equal to the latest
-        #      release (rather than the release before the one we just cut), so
-        #      any contributor cloning fresh gets the current skill by default.
-        #   3. Refreshes `skills-lock.json` with the new content hash so
-        #      restore/verify flows stay accurate.
-        # The commit message carries `[skip release]` so this push does NOT
-        # re-trigger this workflow (the `if:` guard on the job catches it).
+      - name: Smoke — prove the just-published tag installs (temp dir only)
+        # Live smoke test of the just-cut tag via `npx skills add`, but into
+        # an isolated temp directory — NEVER into this repo's
+        # `.agents/skills/deepworkplan/`. That dogfood copy is repo-adapted
+        # (contributor kit wiring for Dailybot + AI Diff Reviewer addons)
+        # and is synced from `skills/deepworkplan/` via
+        # `scripts/refresh-dogfood-skill.sh` when intentionally refreshed.
+        # Blind reinstall from the published tag would overwrite that
+        # adaptation.
         if: steps.version.outputs.skip != 'true'
         env:
           NEW_TAG: ${{ steps.version.outputs.new_tag }}
@@ -216,61 +213,47 @@ jobs:
           # instant; the REST/graphql surface occasionally lags a few seconds).
           sleep 5
 
-          # Pin to the exact tag we just cut instead of letting `skills`
-          # resolve "the newest matching tag". If the tag isn't visible yet
-          # (propagation lag, publish failure), `npx skills add @<tag>` fails
-          # loudly here — that's the whole point of the smoke test. A silent
-          # "already up to date" would let a broken release slip through.
-          #
-          # Two separate `-y` / `--yes` flags below — they mean different
-          # things and both are required in a non-TTY (Actions) environment:
+          SMOKE_DIR=$(mktemp -d)
+          cleanup() { rm -rf "$SMOKE_DIR"; }
+          trap cleanup EXIT
+
+          # Minimal project so `skills` has a place to install into.
+          git -C "$SMOKE_DIR" init -q
+          printf '%s\n' '{"name":"dwp-smoke","private":true}' > "$SMOKE_DIR/package.json"
+
+          # Two separate `-y` / `--yes` flags — both required in a non-TTY
+          # (Actions) environment:
           #   * `npx --yes` — auto-installs the `skills` npm package without
           #     the "Ok to proceed?" download prompt.
           #   * `skills add ... -y` — auto-selects the default install targets
           #     without the interactive agent-picker.
-          echo "::group::npx skills add DailybotHQ/deepworkplan-skill@${NEW_TAG} --skill deepworkplan --force -y"
-          npx --yes skills add "DailybotHQ/deepworkplan-skill@${NEW_TAG}" \
-            --skill deepworkplan --force -y
+          echo "::group::smoke: npx skills add DailybotHQ/deepworkplan-skill@${NEW_TAG} (temp dir)"
+          (
+            cd "$SMOKE_DIR"
+            npx --yes skills add "DailybotHQ/deepworkplan-skill@${NEW_TAG}" \
+              --skill deepworkplan --force -y
+          )
           echo "::endgroup::"
 
-          # Explicit invariant: the vendored SKILL.md's `version:` field MUST
-          # equal the tag we just published. If it doesn't, `npx skills`
-          # accepted the fetch but installed the wrong content — fail hard
-          # rather than commit a misrepresented "dogfood" state.
-          VENDORED_VERSION=$(
+          INSTALLED=$(
             sed -nE 's/^version:[[:space:]]*"([^"]+)".*/\1/p' \
-              .agents/skills/deepworkplan/SKILL.md | head -n1
+              "$SMOKE_DIR/.agents/skills/deepworkplan/SKILL.md" | head -n1
           )
-          if [ "${VENDORED_VERSION}" != "${NEW_VERSION}" ]; then
-            echo "::error::Vendored SKILL.md version is '${VENDORED_VERSION}'"
+          if [ "${INSTALLED}" != "${NEW_VERSION}" ]; then
+            echo "::error::Smoke-install SKILL.md version is '${INSTALLED}'"
             echo "::error::but the just-published tag is '${NEW_TAG}'"
             echo "::error::(expected frontmatter '${NEW_VERSION}')."
-            echo "::error::Refusing to commit a stale dogfood snapshot."
             exit 1
           fi
-          echo "OK: vendored deepworkplan at ${NEW_TAG} (verified)"
-
-          # Empty diff after `--force` with a matching version assertion means
-          # the working copy was already at this tag — no commit needed
-          # (idempotent), but the smoke test PASSED (pinned fetch + version
-          # check both succeeded).
-          if git diff --quiet -- .agents/skills/deepworkplan skills-lock.json; then
-            echo "Vendored deepworkplan already at ${NEW_TAG}; nothing to commit."
-            exit 0
-          fi
-
-          git add .agents/skills/deepworkplan skills-lock.json
-          git commit -m "chore(release): dogfood vendored deepworkplan to ${NEW_TAG} [skip release]"
-          git push origin HEAD:main
+          echo "OK: published tag ${NEW_TAG} installs cleanly (temp smoke; repo dogfood untouched)"
 
       - name: Dogfood — refresh vendored dailybot to latest upstream
         # Keep the vendored `.agents/skills/dailybot/` at the latest published
         # tag of `DailybotHQ/agent-skill`. Runs on every release we cut so any
         # consumer cloning this repo at HEAD picks up the current dailybot
-        # skill alongside deepworkplan.
+        # skill alongside the repo-adapted deepworkplan dogfood.
         #
-        # Mirror of the deepworkplan self-dogfood, adapted to a THIRD-PARTY
-        # upstream: no `sleep 5` here — those upstream tags are stable and
+        # Third-party upstream: no `sleep 5` here — those tags are stable and
         # already propagated through the API layer, so there's no
         # just-published race to wait out.
         #

.review/extension.md

@@ -101,12 +101,16 @@ silent pass. Addons and CI may change; the core loop must not.
   `skills/deepworkplan/` — that's what `skills.sh` ships. See AGENTS.md
   Rule #2 ("The runtime artifact is `skills/deepworkplan/` — keep it pure").
   `docs/`, `tests/`, `scripts/`, and `.github/` are dev-time only.
-- **Always `critical`:** a hand-edit of `version:` in any `SKILL.md`,
-  `CHANGELOG.md`, a git tag, or the vendored dogfood copy at
-  `.agents/skills/deepworkplan/**`. The `auto-release.yml` workflow OWNS
-  those and re-writes them on every merge to `main`; a manual bump collides
-  with the next release commit and duplicates changelog sections. See
-  AGENTS.md Rule #4 ("Versioning is automatic — write good commits").
+- **Always `critical`:** a hand-edit of `version:` in any `SKILL.md` under
+  `skills/deepworkplan/**`, `CHANGELOG.md`, or a git tag. The
+  `auto-release.yml` workflow OWNS those and re-writes them on every merge
+  to `main`; a manual bump collides with the next release commit and
+  duplicates changelog sections. See AGENTS.md Rule #4 ("Versioning is
+  automatic — write good commits"). (The dogfood copy at
+  `.agents/skills/deepworkplan/**` is synced via
+  `scripts/refresh-dogfood-skill.sh`, not by auto-release — flag
+  unreviewed drift from `skills/deepworkplan/` as `warning`, not a
+  versioning critical.)
 - **Always `critical`:** a sub-skill `SKILL.md` frontmatter `version:` that
   drifts from the router `skills/deepworkplan/SKILL.md` `version:` on any
   commit that touches the router. Every in-tree
@@ -221,9 +225,11 @@ silent pass. Addons and CI may change; the core loop must not.
   (see methodology criticals above).
 - **Always `warning`:** documentation change to `skills/deepworkplan/**`
   (spec, guide, addon SKILL.md, or router SKILL.md) that is not mirrored
-  into either the vendored dogfood copy at `.agents/skills/deepworkplan/**`
-  or explicitly deferred to the next auto-release. Consumers install from
-  the vendored copy — drift means agents in the wild read stale guidance.
+  into the vendored dogfood copy at `.agents/skills/deepworkplan/**` via
+  `scripts/refresh-dogfood-skill.sh` (or explicitly deferred in the PR).
+  Contributors working in this repo read the dogfood copy — drift means
+  agents piloting *this* repo see stale guidance. (End users install from
+  the published `skills/deepworkplan/` pack, not from `.agents/`.)
 - **Always `warning`:** RFC-2119 keyword usage in `spec/*.md` that isn't
   uppercase (must / must not / should — should be MUST / MUST NOT /
   SHOULD). The spec is a normative document; lowercase RFC keywords in
@@ -267,12 +273,13 @@ silent pass. Addons and CI may change; the core loop must not.
   checksum + smoke jobs on `ubuntu-latest` and `macos-latest`. There is
   no per-SKILL.md unit-test convention; prompt behaviour is verified by
   running the skill end-to-end.
-- Any content inside `.agents/skills/dailybot/**`,
-  `.agents/skills/ai-diff-reviewer/**`, or `.agents/skills/deepworkplan/**`.
-  These are vendored copies — updated automatically by `auto-release.yml`
-  (deepworkplan self-dogfood) and by `npx skills update` (the two addon
-  skills). Hand-editing them is a bug; the real source is upstream. See
-  DON'T list in AGENTS.md Rule #10 pillar (B).
+- Any content inside `.agents/skills/dailybot/**` or
+  `.agents/skills/ai-diff-reviewer/**`. These are vendored addon copies —
+  refreshed by `auto-release.yml` from upstream. Hand-editing them is a
+  bug; the real source is upstream. See AGENTS.md → Vendored agent skills.
+  (`.agents/skills/deepworkplan/**` is repo-adapted dogfood synced from
+  `skills/deepworkplan/` via `scripts/refresh-dogfood-skill.sh` — review
+  drift as a warning, don't treat auto-release as its owner.)
 - Content in `.claude/**` or `CLAUDE.md`. Those are symlinks
   (`.claude → .agents`, `CLAUDE.md → AGENTS.md`); any real edit belongs
   at the canonical `.agents/**` and `AGENTS.md` target.
@@ -340,7 +347,8 @@ task files.
   commit-type prefix on merges to `main` (`feat!:` MAJOR, `feat:` MINOR,
   everything else PATCH), bumps every in-tree `skills/deepworkplan/**/SKILL.md`
   in sync (router + sub-skills + all addons), updates `CHANGELOG.md`, tags
-  `vX.Y.Z`, and re-vendors the dogfood copy.
+  `vX.Y.Z`, smoke-tests the published tag in a temp dir, and auto-refreshes
+  only the addon dogfood copies (`dailybot`, `ai-diff-reviewer`).
 - **Addons.** Live under `skills/deepworkplan/addons/<name>/`. Every
   addon ships four things: `SPEC.md` (RFC-2119), `templates/*`
   (reasoning guides), `SKILL.md` (onboarding hook, `user-invocable`),
@@ -350,11 +358,11 @@ task files.
 - **Canonical vs symlinked paths.** `.agents/` is canonical. `.claude/`
   and `.cursor/` are back-compat symlinks. Write new content to
   `.agents/`; the symlinks continue to work for legacy tooling.
-- **Two DWP dogfoods.** This repo installs its own `deepworkplan` skill
-  vendored at `.agents/skills/deepworkplan/` (kept in sync by
-  `auto-release.yml`) AND — after Task 3 of the current AI-Diff-Reviewer
-  plan — also vendors `dailybot` and `ai-diff-reviewer` under the same
-  tree. All three are pinned in `skills-lock.json`.
+- **Three dogfood copies under `.agents/skills/`.** `deepworkplan` is
+  repo-adapted and synced from `skills/deepworkplan/` via
+  `scripts/refresh-dogfood-skill.sh` (not by auto-release). `dailybot`
+  and `ai-diff-reviewer` are auto-refreshed on every release. All three
+  are pinned in `skills-lock.json`.
 - **Branding.** Product name is exactly **DeepWorkPlan** (one word, CamelCase)
   or **DWP** (abbreviation). Slug is `deepworkplan`. The skill package
   name on the marketplace is `DailybotHQ/deepworkplan-skill`. Do NOT