Merge pull request 'chore(ci): persistent pnpm store + Playwright caches (and Dockerfile BuildKit cache)' (#27) from gsd/quick-260618-tg2-ci-dep-cache into main

Reviewed-on: #27

.gitea/workflows/ci.yml

@@ -46,12 +46,12 @@ jobs:
       - name: Enable pnpm
         run: corepack enable pnpm
 
-      # actions/cache@v4 is intentionally omitted — probe (D-PROBE-04) showed it
+      # actions/cache@v4 intentionally omitted (D-PROBE-04). Installs now target the
-      # times out on this runner (socket hang-up between runner container and job
+      # host-mounted pnpm store at /pnpm-store (see act_runner config.yaml container.options).
-      # container cache server). pnpm install without cache takes ~30s; acceptable.
+      # Without the host mount the flag still works — pnpm creates an ephemeral store there.
 
       - name: Install dependencies
-        run: pnpm install --frozen-lockfile
+        run: pnpm install --frozen-lockfile --store-dir /pnpm-store --prefer-offline
 
       - name: Lint
         run: pnpm lint
@@ -104,10 +104,11 @@ jobs:
       - name: Enable pnpm
         run: corepack enable pnpm
 
-      # actions/cache@v4 intentionally omitted — same reasoning as fast-checks job (D-PROBE-04).
+      # actions/cache@v4 intentionally omitted (D-PROBE-04). Installs now target the
+      # host-mounted pnpm store at /pnpm-store (see act_runner config.yaml container.options).
 
       - name: Install dependencies
-        run: pnpm install --frozen-lockfile
+        run: pnpm install --frozen-lockfile --store-dir /pnpm-store --prefer-offline
 
       # Pitfall 11: service container healthy != MariaDB accepting connections.
       # No mysql CLI in the runner image (D-PROBE-03); poll via the already-installed
@@ -181,6 +182,9 @@ jobs:
       DB_USER: familysync
       DB_PASSWORD: testpass
       DB_NAME: familysync
+      # Persist Playwright browser binaries across runs via host-mounted /ms-playwright.
+      # Without the host mount CI still works — binaries are downloaded to the ephemeral dir.
+      PLAYWRIGHT_BROWSERS_PATH: /ms-playwright
     steps:
       - uses: actions/checkout@v4
 
@@ -191,10 +195,11 @@ jobs:
       - name: Enable pnpm
         run: corepack enable pnpm
 
-      # actions/cache@v4 intentionally omitted — same reasoning as fast-checks job (D-PROBE-04).
+      # actions/cache@v4 intentionally omitted (D-PROBE-04). Installs now target the
+      # host-mounted pnpm store at /pnpm-store (see act_runner config.yaml container.options).
 
       - name: Install dependencies
-        run: pnpm install --frozen-lockfile
+        run: pnpm install --frozen-lockfile --store-dir /pnpm-store --prefer-offline
 
       # Pitfall 11: service container healthy != MariaDB accepting connections.
       # No mysql CLI in the runner image (D-PROBE-03); poll via the mysql2 driver
@@ -270,7 +275,9 @@ jobs:
       # Install Playwright browsers with system deps BEFORE starting the API, so the long
       # browser download does not run during the API's lifetime.
       # Must run from apps/pwa/ where @playwright/test is installed (D-PROBE-05 confirmed exit 0).
-      # Do NOT cache browser binaries — Playwright explicitly recommends against it in CI.
+      # PLAYWRIGHT_BROWSERS_PATH=/ms-playwright (job-level env) persists binaries across runs via
+      # the host-mounted dir. The --with-deps apt step cannot be cached; baking a runner image
+      # with browsers preinstalled would also drop the --with-deps apt step (future optimization).
       - name: Install Playwright browsers
         run: npx playwright install --with-deps webkit chromium
         working-directory: apps/pwa
@@ -462,7 +469,8 @@ jobs:
             --exit-code 1
 
       # ── pnpm audit + outdated (code-change PRs only, D-12) ───────────────────
-      # actions/cache@v4 intentionally omitted — same reasoning as fast-checks job (D-PROBE-04).
+      # actions/cache@v4 intentionally omitted (D-PROBE-04). Installs now target the
+      # host-mounted pnpm store at /pnpm-store (see act_runner config.yaml container.options).
 
       - uses: actions/setup-node@v4
         if: needs.changes.outputs.code == 'true'
@@ -475,7 +483,7 @@ jobs:
 
       - name: Install dependencies
         if: needs.changes.outputs.code == 'true'
-        run: pnpm install --frozen-lockfile
+        run: pnpm install --frozen-lockfile --store-dir /pnpm-store --prefer-offline
 
       - name: Dependency audit (blocking on High+Critical)
         if: needs.changes.outputs.code == 'true'

.gitea/workflows/publish.yml

@@ -88,6 +88,11 @@ jobs:
       # Build from REPO ROOT (T-08-10): the Dockerfile copies the pnpm workspace manifest +
       # lockfile from the root context; building from apps/api/ would fail to find them.
       - name: Build production image
+        # DOCKER_BUILDKIT=1 is required: the Dockerfile uses `RUN --mount=type=cache`
+        # (BuildKit) to persist the pnpm store across builds. The legacy builder would
+        # fail on that syntax. BuildKit is default on Docker 23+, set explicitly for safety.
+        env:
+          DOCKER_BUILDKIT: '1'
         run: |
           set -euo pipefail
           docker build --target production \

.planning/STATE.md

@@ -257,6 +257,7 @@ Recent decisions affecting current work:
 | 260613-fp9 | `.gitea`/`.planning`-only pushes to main no longer trigger the Docker publish — added `paths-ignore: ['.gitea/**', '.planning/**']` under `on.push` in `.gitea/workflows/publish.yml` (skips only when EVERY changed file matches; mixed code+docs pushes still publish). `.dockerignore` already excludes `.planning` so the image is byte-identical. Done in isolated worktree (phase-10 agent held main tree). | 2026-06-13 | cd5a88c | | [260613-fp9-gitea-and-planning-pushes-should-not-tri](./quick/260613-fp9-gitea-and-planning-pushes-should-not-tri/) |
 | 260613-ndv | Isolate local apps/api integration tests to a dedicated `familysync_test` DB so test runs stop polluting the dev `familysync` DB. New CI-gated vitest globalSetup root-provisions (CREATE DATABASE + GRANT) + migrates + truncate-resets `familysync_test` each run; `vitest.config.ts` forces `DB_NAME=familysync_test` for local workers (no-op under CI, so CI's `familysync` service DB + db:migrate are untouched). Verified: dev `familysync` users stays 3 across a run, `familysync_test` resets (186→93, not doubled), 244/244 tests pass (flaky list_shares timeout gone), typecheck 0. Branch off main. | 2026-06-13 | 07d5161 | Verified | [260613-ndv-wire-apps-api-integration-tests-to-a-ded](./quick/260613-ndv-wire-apps-api-integration-tests-to-a-ded/) |
 | 260618-smr | Remove unused Redis service and all references — Redis confirmed unused at runtime (no ioredis/redis client import, no `REDIS_*` env, not a dependency in any package.json). Dropped the `redis` service from both compose files and cleaned all references in CLAUDE.md, README.md, and docs/* + e2e config. Kept the in-memory-vs-Redis design-rationale comments (D-12/D-18) in listEmitter/reminderScheduler/linkNonceStore/localAuth. `docker compose config` parses clean (0 redis); `format:check` green. Branch off main. | 2026-06-18 | 0b42666 | Verified | [260618-smr-remove-unused-redis-service-and-referenc](./quick/260618-smr-remove-unused-redis-service-and-referenc/) |
+| 260618-tg2 | Persistent CI dependency caches — point all 4 CI `pnpm install` steps at a host-mounted `/pnpm-store` (`--store-dir /pnpm-store --prefer-offline`) and persist Playwright browsers via `PLAYWRIGHT_BROWSERS_PATH=/ms-playwright` on the harness job; added BuildKit `--mount=type=cache` to all 3 Dockerfile install stages + `DOCKER_BUILDKIT=1` on the publish build. Avoids `actions/cache` (D-PROBE-04 timeout). In-repo only — requires act_runner `config.yaml` `container.options` host mounts (manual host change). Verdaccio deferred. Branch off main. | 2026-06-18 | 6e93e24 | Verified | [260618-tg2-persistent-ci-dependency-caches-pnpm-sto](./quick/260618-tg2-persistent-ci-dependency-caches-pnpm-sto/) |
 
 ## Deferred Items
 

.planning/quick/260618-tg2-persistent-ci-dependency-caches-pnpm-sto/260618-tg2-PLAN.md

@@ -0,0 +1,99 @@
+---
+quick_id: 260618-tg2
+slug: persistent-ci-dependency-caches-pnpm-sto
+description: Persistent CI dependency caches (pnpm store + Playwright browsers)
+type: quick
+created: 2026-06-19
+files_modified:
+  - .gitea/workflows/ci.yml
+  - docs/DEVELOPMENT.md
+  - apps/api/Dockerfile
+  - .gitea/workflows/publish.yml
+---
+
+# Quick Task 260618-tg2: Persistent CI dependency caches (pnpm store + Playwright)
+
+## Why
+
+The Gitea runner re-downloads all deps every run: 4 jobs each run `pnpm install --frozen-lockfile`
+cold (lines 54/110/197/478), and the harness job re-downloads Playwright browser binaries every
+run (line 275). The runner is long-lived Docker-on-Unraid, so persisting these via host bind-mounts
+(`/pnpm-store`, `/ms-playwright`, wired in the act_runner `config.yaml` `container.options` — a
+separate manual host change) eliminates the repeat downloads. This avoids `actions/cache@v4`, which
+the Phase-8 runner probe found times out on this runner (D-PROBE-04).
+
+**Scope this task: the two CI caches only.** Verdaccio (registry mirror) and the Dockerfile
+BuildKit cache mount are explicitly OUT of scope for now.
+
+## Tasks
+
+### Task 1 — pnpm store: point all CI installs at the persistent store
+In `.gitea/workflows/ci.yml`, change each of the four install steps:
+```
+run: pnpm install --frozen-lockfile
+```
+→
+```
+run: pnpm install --frozen-lockfile --store-dir /pnpm-store --prefer-offline
+```
+Lines 54 (fast-checks), 110 (api), 197 (harness), 478 (security). `--store-dir` and
+`--prefer-offline` are valid pnpm 11.5.1 install flags (verify with `pnpm install --help`).
+Do NOT add `store-dir` to a repo `.npmrc` — local dev has no `/pnpm-store`.
+Update the stale comment near line 51 (the "no cache backend / ~30s acceptable" note) to reflect
+that installs now use the persistent host-mounted store.
+- verify: `grep -c -- '--store-dir /pnpm-store --prefer-offline' .gitea/workflows/ci.yml` → 4
+- done: all four installs use the persistent store; YAML still valid.
+
+### Task 2 — Playwright: persist browser binaries on the harness job
+Add a job-level `env:` to the `harness:` job so every step (browser install + test run) resolves
+the same persistent path:
+```yaml
+  harness:
+    runs-on: ubuntu-latest
+    env:
+      PLAYWRIGHT_BROWSERS_PATH: /ms-playwright
+```
+(If the harness job already has a job-level `env:` map, add the key to it rather than duplicating.)
+Leave `npx playwright install --with-deps webkit chromium` (line 275) as-is — the binary download
+is now cached by version; the `--with-deps` apt step can't persist (add a one-line comment noting
+"baking a runner image with browsers preinstalled would also drop the --with-deps apt step" as a
+future optimization).
+- verify: `PLAYWRIGHT_BROWSERS_PATH: /ms-playwright` present under the harness job; the test-run
+  step (PLAYWRIGHT_BASE_URL ~line 340) inherits it.
+- done: Playwright browsers persist across runs.
+
+### Task 3 — Document the host-mount dependency
+Add a short subsection to `docs/DEVELOPMENT.md` (CI/runner area) noting:
+- CI now uses persistent caches at container paths `/pnpm-store` and `/ms-playwright`.
+- These require the act_runner `config.yaml` `container.options` to bind-mount host dirs to those
+  paths (host change, not in this repo).
+- Without the mounts CI still works — it just falls back to uncached (writes to an ephemeral dir).
+- done: the host-side requirement is discoverable from the repo.
+
+## Constraints
+- Must pass local gates before each commit: `format:check` (prettier), eslint, YAML validity
+  (yq or actionlint if available), typecheck (no TS touched, but run if cheap).
+- Job names and the required-check contexts (CI / fast-checks, CI / api, CI / harness,
+  CI / security, CI / gate) MUST stay identical so branch protection still matches. Do not rename
+  jobs or restructure the job graph.
+- Atomic commits (Task 1, Task 2, Task 3 may be one or separate commits — keep changes coherent).
+
+### Task 4 — Dockerfile BuildKit pnpm-store cache (added mid-task per user request)
+`apps/api/Dockerfile`: add `# syntax=docker/dockerfile:1` (line 1) and a BuildKit cache mount to
+all three pnpm install stages (builder, pwa-builder, production):
+`RUN --mount=type=cache,target=/pnpm-store,id=pnpm-store,sharing=locked pnpm install ... --store-dir /pnpm-store`.
+`sharing=locked` because builder + pwa-builder run in parallel and would otherwise race the store.
+`.gitea/workflows/publish.yml`: set `DOCKER_BUILDKIT: '1'` on the "Build production image" step so
+the legacy builder can't break on the `--mount` syntax (BuildKit is default on Docker 23+; explicit
+for safety).
+- done: image build reuses a persistent BuildKit pnpm-store cache across builds.
+
+## OUT OF SCOPE (do not touch)
+- Verdaccio / any `.npmrc` registry change (deferred — user will set up later).
+
+## must_haves
+- truth: "All four ci.yml pnpm installs use --store-dir /pnpm-store --prefer-offline"
+- truth: "The harness job sets PLAYWRIGHT_BROWSERS_PATH=/ms-playwright"
+- truth: "ci.yml remains valid YAML with unchanged job names / required-check contexts"
+- truth: "docs note the act_runner config.yaml host-mount requirement"
+- artifacts: [.gitea/workflows/ci.yml, docs/DEVELOPMENT.md]

.planning/quick/260618-tg2-persistent-ci-dependency-caches-pnpm-sto/260618-tg2-SUMMARY.md

@@ -0,0 +1,121 @@
+---
+quick_id: 260618-tg2
+slug: persistent-ci-dependency-caches-pnpm-sto
+phase: "20"
+plan: tg2
+status: complete
+completed: 2026-06-18
+tags: [ci, caching, pnpm, playwright]
+key-files:
+  modified:
+    - .gitea/workflows/ci.yml
+    - docs/DEVELOPMENT.md
+    - apps/api/Dockerfile
+    - .gitea/workflows/publish.yml
+decisions:
+  - All four CI pnpm installs now target /pnpm-store via --store-dir --prefer-offline flags
+  - PLAYWRIGHT_BROWSERS_PATH added at harness job level (not per-step) so both install and run steps share the same path
+  - D-PROBE-04 comments updated to reflect the new store strategy rather than "no cache"
+---
+
+# Quick Task 260618-tg2: Persistent CI dependency caches (pnpm store + Playwright) Summary
+
+**One-liner:** Point all four CI pnpm installs at `/pnpm-store` and harness Playwright at `/ms-playwright` via host-mounted directories on the act_runner.
+
+## What Was Done
+
+### Task 1 — pnpm store (ci.yml, 4 install lines)
+
+Changed all four `pnpm install --frozen-lockfile` lines to
+`pnpm install --frozen-lockfile --store-dir /pnpm-store --prefer-offline`:
+
+- Line ~54: `fast-checks` job
+- Line ~110: `api` job
+- Line ~197: `harness` job
+- Line ~485: `security` job (conditional)
+
+Both flags confirmed valid against `pnpm 11.5.1 install --help` before use.
+
+Updated the stale D-PROBE-04 comment in each job from "no cache backend / ~30s acceptable"
+to reflect that installs now target the host-mounted store.
+
+### Task 2 — Playwright browsers (harness job env)
+
+Added `PLAYWRIGHT_BROWSERS_PATH: /ms-playwright` to the existing job-level `env:` block on
+the `harness:` job (alongside the DB_* creds). This means both the `Install Playwright browsers`
+step and the `Run harness` step inherit the same path, so cached binaries are found at install
+time and used at test time.
+
+Added a comment on the Playwright install step noting the future optimization: baking a runner
+image with browsers preinstalled would also eliminate the `--with-deps` apt step.
+
+### Task 3 — Host-mount documentation (docs/DEVELOPMENT.md)
+
+Added a "CI dependency caches" subsection under the CI Pipeline Overview. Documents:
+- The two container paths (`/pnpm-store`, `/ms-playwright`) with a reference table
+- That the act\_runner `config.yaml` `container.options` bind-mount is a **host-side** change
+- That CI still works without the mounts (ephemeral fallback — just no caching)
+
+## Verification
+
+```
+grep -c -- '--store-dir /pnpm-store --prefer-offline' .gitea/workflows/ci.yml
+→ 4
+
+grep -n 'PLAYWRIGHT_BROWSERS_PATH' .gitea/workflows/ci.yml
+→ 187: PLAYWRIGHT_BROWSERS_PATH: /ms-playwright  (job-level env)
+→ 278: # PLAYWRIGHT_BROWSERS_PATH=/ms-playwright... (comment)
+
+python3 -c 'import yaml,sys; yaml.safe_load(open(".gitea/workflows/ci.yml")); print("YAML valid")'
+→ YAML valid
+
+pnpm format:check → All matched files use Prettier code style!
+pnpm md:lint     → Summary: 0 error(s)
+```
+
+### Task 4 — Dockerfile BuildKit pnpm-store cache (added mid-task by user request)
+
+Added by the orchestrator after the initial 3 tasks, when the user asked to include the Dockerfile:
+
+- `apps/api/Dockerfile`: added `# syntax=docker/dockerfile:1` (line 1) and a BuildKit cache mount
+  (`RUN --mount=type=cache,target=/pnpm-store,id=pnpm-store,sharing=locked ... --store-dir /pnpm-store`)
+  to all three pnpm install stages (builder, pwa-builder, production). `sharing=locked` because
+  builder + pwa-builder run in parallel and would otherwise race the shared store.
+- `.gitea/workflows/publish.yml`: set `DOCKER_BUILDKIT: '1'` on the "Build production image" step —
+  the publish path uses plain `docker build` (not buildx), and the legacy builder would fail on the
+  `--mount` syntax. BuildKit is default on Docker 23+; set explicitly for safety.
+- Verified: `format:check` clean (Dockerfile is outside prettier's scope), publish.yml valid YAML.
+
+## Commits
+
+| Hash | Message |
+| --- | --- |
+| `80b2038` | chore(20): persistent CI caches — pnpm store + Playwright browsers |
+| `f83d423` | docs(20): document CI persistent cache host-mount dependency |
+| `6e93e24` | chore(260618-tg2): BuildKit pnpm-store cache mount in Dockerfile build |
+
+## Deviations from Plan
+
+Dockerfile cache (Task 4) was added mid-task at the user's request after the initial 3-task plan
+(it had been explicitly deferred/out-of-scope). The publish workflow's `DOCKER_BUILDKIT=1` was a
+required companion change so the `--mount` syntax doesn't break the legacy builder.
+
+- `--store-dir` flag form matches plan exactly (plan said verify against pnpm; verified: valid)
+- harness job already had a job-level `env:` map; `PLAYWRIGHT_BROWSERS_PATH` was added to it as instructed
+- stale comment text updated as instructed
+
+## Known Stubs
+
+None.
+
+## Threat Flags
+
+None — YAML-only and doc-only changes; no new network endpoints, auth paths, or trust boundaries introduced.
+
+## Self-Check: PASSED
+
+- `.gitea/workflows/ci.yml` — modified and committed at 80b2038
+- `docs/DEVELOPMENT.md` — modified and committed at f83d423
+- YAML validity confirmed by python3 yaml.safe_load
+- 4 install lines confirmed by grep -c
+- PLAYWRIGHT_BROWSERS_PATH confirmed at job-level env line 187

apps/api/Dockerfile

@@ -1,3 +1,4 @@
+# syntax=docker/dockerfile:1
 # Built from the REPO ROOT context (see docker-compose.yml: build.context: .)
 # so the pnpm workspace manifest + lockfile are available for a deterministic,
 # workspace-aware install. apps/api is one package in the pnpm workspace.
@@ -12,7 +13,8 @@ FROM base AS builder
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json ./
 COPY apps/api/package.json ./apps/api/
 COPY apps/pwa/package.json ./apps/pwa/
-RUN pnpm install --frozen-lockfile --filter @familysync/api...
+RUN --mount=type=cache,target=/pnpm-store,id=pnpm-store,sharing=locked \
+    pnpm install --frozen-lockfile --filter @familysync/api... --store-dir /pnpm-store
 COPY apps/api ./apps/api
 RUN pnpm --filter @familysync/api build
 
@@ -28,7 +30,8 @@ FROM base AS pwa-builder
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json ./
 COPY apps/api/package.json ./apps/api/
 COPY apps/pwa/package.json ./apps/pwa/
-RUN pnpm install --frozen-lockfile --filter @familysync/pwa...
+RUN --mount=type=cache,target=/pnpm-store,id=pnpm-store,sharing=locked \
+    pnpm install --frozen-lockfile --filter @familysync/pwa... --store-dir /pnpm-store
 COPY apps/pwa ./apps/pwa
 RUN pnpm --filter @familysync/pwa build
 
@@ -36,7 +39,8 @@ FROM base AS production
 COPY pnpm-workspace.yaml pnpm-lock.yaml package.json ./
 COPY apps/api/package.json ./apps/api/
 COPY apps/pwa/package.json ./apps/pwa/
-RUN pnpm install --frozen-lockfile --prod --filter @familysync/api...
+RUN --mount=type=cache,target=/pnpm-store,id=pnpm-store,sharing=locked \
+    pnpm install --frozen-lockfile --prod --filter @familysync/api... --store-dir /pnpm-store
 COPY --from=builder /app/apps/api/dist ./apps/api/dist
 WORKDIR /app/apps/api
 # Enforce production identity — engages the NODE_ENV=production hard guard

docs/DEVELOPMENT.md

@@ -205,6 +205,21 @@ Every PR to `main` runs through `.gitea/workflows/ci.yml`. A `changes` path-filt
 
 All five jobs must pass (or be legitimately skipped) before a PR can merge. See [docs/TESTING.md](TESTING.md) for test suite details.
 
+### CI dependency caches
+
+CI uses two persistent cache paths inside job containers:
+
+| Path             | Content                                                                 |
+| ---------------- | ----------------------------------------------------------------------- |
+| `/pnpm-store`    | pnpm content-addressable store (`--store-dir /pnpm-store`)              |
+| `/ms-playwright` | Playwright browser binaries (`PLAYWRIGHT_BROWSERS_PATH=/ms-playwright`) |
+
+These paths must be bind-mounted from host directories in the act_runner `config.yaml`
+`container.options` field — that is a **host-side change, not tracked in this repo**. Without the
+mounts, CI still works correctly — pnpm creates an ephemeral store at `/pnpm-store` inside the
+container and Playwright downloads browsers fresh each run. The mounts only eliminate repeat
+downloads across runs.
+
 ## Drizzle Migration Workflow
 
 Schema changes follow a strict two-step process. **`drizzle-kit push` is not available** — it has been removed from the scripts because it emits a false destructive diff (table truncation) on populated MariaDB databases.