workstation: per-user playwright browser MCP for all users, reproducible from git

Viktor asked that the playwright browser MCP be available for every devvm user
in every directory, with each user running their own server and multiple
concurrent sessions per user.

Before this, playwright was hand-set-up per user (~/.config/systemd/user/
playwright-mcp.service on 8931/8932/8933) and only wizard was actually wired —
emo's and anca's servers ran but their ~/.claude.json had no playwright entry,
so their Claude never connected. None of it was reproducible from git (units,
refresh script, and the Vault snapshot token lived only in user homes), so a
devvm rebuild would silently lose it.

This makes it reproducible and fixes the unwired users:

- roster_engine.py: sticky per-user PLAYWRIGHT_PORT (PLAYWRIGHT_BASE_PORT=8931,
  allocated for every roster user incl. the admin), emitted in the derive JSON.
- scripts/workstation/playwright/: system-level TEMPLATE units
  (playwright-mcp@.service + playwright-snapshot-refresh@.{service,timer},
  User=%i — system manager, so no systemd --user / linger) + the refresh script.
  @playwright/mcp pinned to 0.0.76 (avoids the @latest silent-fleet-roll
  footgun, same rationale as T3_PIN).
- setup-devvm.sh: install the templates + script (9e); stage the chrome-service
  snapshot bearer token from Vault to a root file (8c) — the hourly root
  reconcile has no Vault token, mirrors the Claude OAuth staging in 8a.
- t3-provision-users.sh: install_playwright() (ALL tiers incl. admin) writes
  PLAYWRIGHT_PORT, seeds the token if-absent, wires the user-scope ~/.claude.json
  by running `claude mcp add` AS the user (clobber-proof + if-absent, so it fixes
  existing/new/admin without rewriting a populated config), and enable --now's the
  instances (idempotent, never restarts a running server). Also hardened the
  section-1 *.env scan to skip the new playwright-*.env files (no T3_PORT -> grep
  no-match would abort under set -e -o pipefail).
- Docs: chrome-service-snapshot runbook (new Provisioning section + system-unit
  commands), multi-tenancy.md, and the 2026-06-07 plan Task 2.3.

Supersedes the hand-made per-user --user units (one-time idle-gated migration to
follow on the live host).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

scripts/workstation/playwright/playwright-mcp@.service

@@ -0,0 +1,35 @@
+[Unit]
+# Per-user isolated playwright-mcp HTTP server — the browser MCP each user's
+# Claude Code sessions connect to (user-scope `.claude.json` entry "playwright"
+# -> http://localhost:<PLAYWRIGHT_PORT>/mcp). System-level TEMPLATE unit (one
+# committed file, one instance per OS user: playwright-mcp@<user>.service), so
+# it is reproducible from git and root-manageable WITHOUT systemd --user / linger.
+# Installed to /etc/systemd/system by setup-devvm.sh; enabled per-user by
+# t3-provision-users.sh. Supersedes the hand-made ~/.config/systemd/user units.
+Description=Per-user isolated playwright-mcp HTTP server (%i)
+After=network-online.target playwright-snapshot-refresh@%i.service
+Wants=network-online.target playwright-snapshot-refresh@%i.service
+
+[Service]
+Type=simple
+User=%i
+# PLAYWRIGHT_PORT is written per-user by t3-provision-users.sh from roster_engine
+# (PLAYWRIGHT_BASE_PORT, sticky allocation). Required (no `-`): a missing port
+# file should fail loudly rather than start npx with an empty --port.
+EnvironmentFile=/etc/t3-serve/playwright-%i.env
+Restart=on-failure
+RestartSec=5
+# --isolated: each MCP HTTP connection (= each Claude Code session) gets a fresh
+#   ephemeral BrowserContext, so a single user's concurrent sessions never share
+#   tabs. --storage-state seeds each context from the hourly cookie snapshot
+#   harvested from in-cluster chrome-service (warm logged-in state).
+# Version PINNED (see the T3_PIN rationale in setup-devvm.sh): @latest re-resolves
+#   on every restart, so an upstream breaking release would silently roll the
+#   whole fleet. Bump deliberately in git. %h is NOT used (it resolves to /root
+#   in a system unit even with User=); the home path is spelled out as /home/%i.
+ExecStart=/usr/bin/npx -y @playwright/mcp@0.0.76 --port ${PLAYWRIGHT_PORT} --host localhost --headless --browser chrome --isolated --storage-state /home/%i/.cache/playwright-shared-storage-state.json
+StandardOutput=journal
+StandardError=journal
+
+[Install]
+WantedBy=multi-user.target

scripts/workstation/playwright/playwright-snapshot-refresh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+# Refresh the local cookie+localStorage snapshot served by chrome-service.
+#
+# Run per-user by the hourly playwright-snapshot-refresh@<user>.timer systemd
+# unit (as that user, so $HOME resolves to the user's home). Per-session Claude
+# Code MCP contexts (`@playwright/mcp --isolated --storage-state=…`) read this
+# file on each connection — fresh state is visible to NEW sessions, existing
+# ones keep what they were seeded with.
+#
+# Token: cached at ~/.config/playwright/token. Seeded per-user (if-absent) by
+#   t3-provision-users.sh from the root-staged /etc/t3-serve/chrome-service-token
+#   (which setup-devvm.sh writes from Vault `secret/chrome-service`
+#   api_bearer_token). Rotate by re-staging + re-copying; the snapshot endpoint
+#   reloads the token via Reloader, local caches must be refreshed.
+set -euo pipefail
+
+URL="${PLAYWRIGHT_SNAPSHOT_URL:-https://chrome.viktorbarzin.me/api/snapshot}"
+TOKEN_FILE="${PLAYWRIGHT_SNAPSHOT_TOKEN:-$HOME/.config/playwright/token}"
+DEST="${PLAYWRIGHT_SNAPSHOT_PATH:-$HOME/.cache/playwright-shared-storage-state.json}"
+
+if [ ! -r "$TOKEN_FILE" ]; then
+  echo "ERROR: token file $TOKEN_FILE missing or unreadable" >&2
+  exit 1
+fi
+
+mkdir -p "$(dirname "$DEST")"
+TMP="$DEST.new.$$"
+trap 'rm -f "$TMP"' EXIT
+
+TOKEN="$(cat "$TOKEN_FILE")"
+
+HTTP_CODE=$(curl -sS \
+  -H "Authorization: Bearer $TOKEN" \
+  -o "$TMP" \
+  -w '%{http_code}' \
+  --max-time 30 \
+  "$URL")
+
+if [ "$HTTP_CODE" != "200" ]; then
+  echo "ERROR: HTTP $HTTP_CODE from $URL" >&2
+  cat "$TMP" >&2
+  exit 1
+fi
+
+# Sanity: response must be valid JSON with at least the cookies/origins keys.
+python3 - "$TMP" <<'PY' || { echo "ERROR: response is not a valid storageState JSON" >&2; exit 1; }
+import json, sys
+with open(sys.argv[1]) as f:
+    data = json.load(f)
+if "cookies" not in data or "origins" not in data:
+    raise SystemExit("missing required keys")
+PY
+
+mv -f "$TMP" "$DEST"
+trap - EXIT
+chmod 600 "$DEST"
+echo "snapshot refreshed: $DEST ($(stat -c %s "$DEST") bytes)"

scripts/workstation/playwright/playwright-snapshot-refresh@.service

@@ -0,0 +1,22 @@
+[Unit]
+# Per-user oneshot that pulls the warm cookie+localStorage snapshot from
+# in-cluster chrome-service into ~/.cache/playwright-shared-storage-state.json,
+# which playwright-mcp@%i seeds every new session from. System-level TEMPLATE
+# (one instance per user); runs the shared /usr/local/bin script as the user.
+Description=Refresh %i's playwright storage-state snapshot from chrome-service
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=oneshot
+User=%i
+# Runs as %i, so the script's $HOME-relative paths (token, cache dest) resolve to
+# the user's home. $HOME/$USER are set by systemd because User= is set.
+ExecStart=/usr/local/bin/playwright-snapshot-refresh
+StandardOutput=journal
+StandardError=journal
+# Don't hang if chrome-service is unreachable — the timer retries next hour.
+TimeoutStartSec=60
+
+[Install]
+WantedBy=multi-user.target

scripts/workstation/playwright/playwright-snapshot-refresh@.timer

@@ -0,0 +1,16 @@
+[Unit]
+Description=Hourly refresh of %i's playwright storage-state snapshot from chrome-service
+After=network-online.target
+
+[Timer]
+# 5 minutes after the in-cluster snapshot-harvester CronJob (runs at :23 every
+# hour) so the file we pull is the freshest one. Also once shortly after boot so
+# a freshly-booted box doesn't wait until the next :28 to populate the cache.
+OnCalendar=*-*-* *:28:00
+OnBootSec=2min
+Persistent=true
+RandomizedDelaySec=30
+Unit=playwright-snapshot-refresh@%i.service
+
+[Install]
+WantedBy=timers.target

scripts/workstation/roster_engine.py

@@ -21,6 +21,12 @@ from typing import Iterable
 import yaml
 
 BASE_PORT = 3773
+# Per-user playwright-mcp HTTP port (the browser MCP each user's Claude sessions
+# connect to). Distinct range from T3_PORT, allocated for EVERY roster user incl.
+# the admin (wizard is listed). Sticky from existing, so the live in-session
+# assignments (wizard 8931, emo 8932, ancamilea 8933) are preserved across
+# reconciles once seeded; a fresh box allocates from 8931 in sorted order.
+PLAYWRIGHT_BASE_PORT = 8931
 VALID_TIERS = ("admin", "power-user", "namespace-owner")
 # single    - ~/code IS the locked infra clone (the original non-admin layout)
 # workspace - ~/code is a plain directory of per-project clones; the locked
@@ -82,6 +88,7 @@ class DesiredState:
     ttyd_user_map: str
     dispatch: dict[str, dict]
     ports: dict[str, int]
+    playwright_ports: dict[str, int] = field(default_factory=dict)
 
 
 @dataclass(frozen=True)
@@ -203,13 +210,18 @@ def has_blocking_errors(issues: list[ValidationIssue]) -> bool:
 # --------------------------------------------------------------------------
 
 
-def _allocate_ports(roster: Roster, existing_ports: dict[str, int]) -> dict[str, int]:
+def _allocate_ports(
+    roster: Roster, existing_ports: dict[str, int], base: int = BASE_PORT
+) -> dict[str, int]:
+    """Sticky port allocation: keep every roster user's existing port, then assign
+    each new user the next free port from `base`. Used for both T3_PORT (base 3773)
+    and the per-user playwright-mcp port (base 8932)."""
     ports = {u: existing_ports[u] for u in roster.users if u in existing_ports}
     used = set(ports.values())
     for os_user in sorted(roster.users):
         if os_user in ports:
             continue
-        candidate = BASE_PORT
+        candidate = base
         while candidate in used:
             candidate += 1
         ports[os_user] = candidate
@@ -224,9 +236,14 @@ _TTYD_MAP_HEADER = (
 
 
 def derive_desired_state(
-    roster: Roster, existing_ports: dict[str, int]
+    roster: Roster,
+    existing_ports: dict[str, int],
+    existing_playwright_ports: dict[str, int] | None = None,
 ) -> DesiredState:
     ports = _allocate_ports(roster, existing_ports)
+    playwright_ports = _allocate_ports(
+        roster, existing_playwright_ports or {}, base=PLAYWRIGHT_BASE_PORT
+    )
     ordered = sorted(roster.users.values(), key=lambda u: ports[u.os_user])
     ttyd_lines = [f"{u.authentik_user}={u.os_user}" for u in ordered]
     ttyd_user_map = _TTYD_MAP_HEADER + "\n".join(ttyd_lines) + "\n"
@@ -246,7 +263,7 @@ def derive_desired_state(
         )
         for u in roster.users.values()
     }
-    return DesiredState(accounts, ttyd_user_map, dispatch, ports)
+    return DesiredState(accounts, ttyd_user_map, dispatch, ports, playwright_ports)
 
 
 def groups_to_add(desired: Iterable[str], current: Iterable[str]) -> list[str]:
@@ -303,6 +320,7 @@ def _desired_state_to_dict(ds: DesiredState) -> dict:
         "ttyd_user_map": ds.ttyd_user_map,
         "dispatch": ds.dispatch,
         "ports": ds.ports,
+        "playwright_ports": ds.playwright_ports,
     }
 
 
@@ -318,7 +336,11 @@ def _main(argv: list[str]) -> int:
     pv.add_argument("--k8s-users-json", required=True, help="JSON map {k8s_user: tier}")
     pd = sub.add_parser("derive", help="emit desired state as JSON")
     pd.add_argument("--roster", required=True)
-    pd.add_argument("--ports-json", required=True, help="JSON map {os_user: port}")
+    pd.add_argument("--ports-json", required=True, help="JSON map {os_user: T3_PORT}")
+    pd.add_argument(
+        "--playwright-ports-json",
+        help="JSON map {os_user: PLAYWRIGHT_PORT} (optional; sticky allocation)",
+    )
     args = parser.parse_args(argv)
 
     roster = load_roster_file(args.roster)
@@ -329,7 +351,12 @@ def _main(argv: list[str]) -> int:
             print(f"{issue.severity.upper()}: {issue.message}", file=sys.stderr)
         return 1 if has_blocking_errors(issues) else 0
     with open(args.ports_json, encoding="utf-8") as fh:
-        desired = derive_desired_state(roster, json.load(fh))
+        existing_ports = json.load(fh)
+    existing_playwright_ports = {}
+    if args.playwright_ports_json:
+        with open(args.playwright_ports_json, encoding="utf-8") as fh:
+            existing_playwright_ports = json.load(fh)
+    desired = derive_desired_state(roster, existing_ports, existing_playwright_ports)
     json.dump(_desired_state_to_dict(desired), sys.stdout, indent=2, sort_keys=True)
     sys.stdout.write("\n")
     return 0

scripts/workstation/setup-devvm.sh

@@ -140,6 +140,16 @@ if command -v vault >/dev/null; then
   else
     log "WARN: secret/workstation codex_shared_auth_json absent -> shared Codex auth not staged"
   fi
+  # 8c) chrome-service snapshot bearer token -> root file the provisioner copies
+  #     per-user (if-absent) to ~/.config/playwright/token, which the per-user
+  #     playwright-snapshot-refresh reads. One token for all users (single shared
+  #     warm profile, by design). 0600: the snapshot it fetches holds cookies.
+  if cs_tok="$(vault kv get -field=api_bearer_token secret/chrome-service 2>/dev/null)"; then
+    install -m 0600 /dev/stdin /etc/t3-serve/chrome-service-token <<<"$cs_tok"
+    log "staged /etc/t3-serve/chrome-service-token (playwright snapshot auth)"
+  else
+    log "WARN: secret/chrome-service api_bearer_token absent -> playwright snapshot refresh will 401"
+  fi
 fi
 
 # 9) service layer: install + enable the machine-wide systemd units (sources in
@@ -177,6 +187,16 @@ for u in t3-serve@.service \
          t3-dispatch.service; do
   install -m 0644 "$SCRIPTS/$u" "/etc/systemd/system/$u"
 done
+# 9e) per-user playwright-mcp browser MCP: system-level TEMPLATE units (one
+#     instance per OS user) + the snapshot-refresh script. Reproducible-from-git
+#     replacement for the hand-made ~/.config/systemd/user/playwright-* units
+#     (no systemd --user / linger needed). Enabled per-user by the provisioner;
+#     PLAYWRIGHT_PORT (roster_engine) + the chrome-service token (8c) feed them.
+install -m 0755 "$HERE/playwright/playwright-snapshot-refresh" /usr/local/bin/playwright-snapshot-refresh
+for u in playwright-mcp@.service playwright-snapshot-refresh@.service playwright-snapshot-refresh@.timer; do
+  install -m 0644 "$HERE/playwright/$u" "/etc/systemd/system/$u"
+done
+log "playwright: template units + snapshot-refresh script installed (per-user enable in provisioner)"
 systemctl daemon-reload
 systemctl enable --now t3-dispatch.service \
   t3-autoupdate.timer t3-backup-state.timer t3-provision-users.timer >/dev/null 2>&1 || \

scripts/workstation/test_roster_engine.py

@@ -296,6 +296,53 @@ def test_derive_is_deterministic():
     )
 
 
+# --------------------------------------------------------------------------
+# derive_desired_state: per-user playwright-mcp ports (reproducible browser MCP)
+# --------------------------------------------------------------------------
+
+# wizard (admin) IS a roster user, so playwright ports are allocated for every
+# user incl. the admin, from PLAYWRIGHT_BASE_PORT=8931. The live in-session
+# assignment is wizard 8931, emo 8932, ancamilea 8933.
+LIVE_PLAYWRIGHT_PORTS = {"wizard": 8931, "emo": 8932, "ancamilea": 8933}
+
+
+def test_derive_allocates_playwright_ports_for_all_users_incl_admin():
+    ds = eng.derive_desired_state(_roster(THREE), {})
+    # fresh box: sorted os_user order (ancamilea, emo, wizard) from 8931
+    assert ds.playwright_ports == {"ancamilea": 8931, "emo": 8932, "wizard": 8933}
+
+
+def test_derive_preserves_existing_sticky_playwright_ports():
+    # Seeded with the live assignment -> preserved exactly (nobody's port moves).
+    ds = eng.derive_desired_state(
+        _roster(THREE), {}, existing_playwright_ports=LIVE_PLAYWRIGHT_PORTS
+    )
+    assert ds.playwright_ports == LIVE_PLAYWRIGHT_PORTS
+
+
+def test_derive_allocates_next_free_playwright_port_for_new_user():
+    # Existing users sticky; a brand-new user gets the next free port from 8931.
+    ds = eng.derive_desired_state(
+        _roster(THREE), {}, existing_playwright_ports={"wizard": 8931, "emo": 8932}
+    )
+    assert ds.playwright_ports["wizard"] == 8931
+    assert ds.playwright_ports["emo"] == 8932
+    assert ds.playwright_ports["ancamilea"] == 8933  # next free, skipping 8931/8932
+
+
+def test_playwright_ports_are_disjoint_from_t3_ports():
+    ds = eng.derive_desired_state(_roster(THREE), LIVE_PORTS, LIVE_PLAYWRIGHT_PORTS)
+    assert set(ds.ports.values()).isdisjoint(ds.playwright_ports.values())
+
+
+def test_desired_state_dict_includes_playwright_ports():
+    # The JSON adapter is the contract the bash provisioner consumes via jq.
+    d = eng._desired_state_to_dict(
+        eng.derive_desired_state(_roster(THREE), {}, LIVE_PLAYWRIGHT_PORTS)
+    )
+    assert d["playwright_ports"] == LIVE_PLAYWRIGHT_PORTS
+
+
 # --------------------------------------------------------------------------
 # groups_to_add: the additive-only invariant (module #1)
 # --------------------------------------------------------------------------