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>

docs/architecture/multi-tenancy.md

@@ -547,6 +547,8 @@ Separate from the in-cluster namespace-owner model above, the **devvm** (`10.0.1
 
 **Claude Code runtime — native, per-user (2026-06-15):** `claude` is the **native** install (`~/.local/bin/claude` → `~/.local/share/claude/versions/<v>`, self-updating; `installMethod: native`) — NOT npm-global or npx. It is the runtime for both the ttyd launcher and each `t3-serve` instance. `setup-devvm.sh` installs node ONLY for the `t3` CLI (not claude); per-user native claude is provisioned by the reconcile's `install_user_claude_native` (covers terminal + t3, idempotent, skip-if-present) and self-bootstrapped by `start-claude.sh` on first launch — both via the official `https://claude.ai/install.sh`. The legacy machine-wide `npm install -g @anthropic-ai/claude-code` bootstrap and the launcher's `npx` fallback were removed; existing users had already auto-migrated to native, and the npm-global dir was empty. **PATH (`~/.local/bin`, where the native binary lives):** ensured three ways — `/etc/profile.d/10-local-bin.sh` for login shells (machine-wide, fresh-user-safe), `start-claude.sh` itself (the launcher runs in tmux's non-login env that skips the user's shell rc), and `t3-serve@.service` (`Environment=PATH=…:/home/%i/.local/bin`).
 
+**Per-user browser MCP — playwright, reproducible from git (2026-06-16):** every user (incl. the admin) gets their OWN isolated `@playwright/mcp` server so their concurrent Claude sessions don't fight over tabs (`--isolated` → a fresh browser context per MCP connection), wired into Claude in **every directory** via a user-scope `~/.claude.json` entry (`playwright → http://localhost:<PLAYWRIGHT_PORT>/mcp`). Mechanism: **system-level template units** `playwright-mcp@<user>.service` + `playwright-snapshot-refresh@<user>.{service,timer}` (`User=%i`, sourced from `scripts/workstation/playwright/`, installed by `setup-devvm.sh` §9e — system manager, so NO systemd --user / linger). `roster_engine.py` allocates a sticky per-user `PLAYWRIGHT_PORT` (`PLAYWRIGHT_BASE_PORT=8931`); the reconcile's `install_playwright()` writes it, seeds the chrome-service snapshot token if-absent (staged from Vault `secret/chrome-service` to `/etc/t3-serve/chrome-service-token` by `setup-devvm.sh` §8c, since the hourly root reconcile has no Vault token), wires `~/.claude.json` by running `claude mcp add --scope user` 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). The `@playwright/mcp` version is **pinned** in the unit (the `@latest`-silently-rolls-the-fleet footgun — see `T3_PIN`). Replaced the earlier hand-made `~/.config/systemd/user/playwright-*` units (one-time idle-gated migration; pre-migration emo/anca had servers running but never wired into their `.claude.json`). Cookie-warming pipeline + ops: `../runbooks/chrome-service-snapshot.md`.
+
 **Infra access:** non-admins get their own **writable, git-crypt-LOCKED** clone of the (public) infra repo — code/docs plaintext, secret files (`*.tfvars`, `secrets/**`) stay ciphertext. Its location depends on the per-user `code_layout` in `roster.yaml`: `single` (default) puts the clone AT `~/code`; `workspace` makes `~/code` a plain directory of per-project clones — the infra clone at `~/code/infra` plus each roster `repos` entry cloned from Forgejo `viktor/<name>` **as the user** (their PAT authenticates, so private repos work; clone failures WARN and retry next hour). Flipping a user to `workspace` auto-migrates their existing `~/code` clone to `~/code/infra` (local branches/dirty state survive; running processes follow the moved inode). ancamilea = workspace + `tripit` since 2026-06-10. The provisioner clones infra anonymously from the public GitHub mirror; **contribute access is wired per-user on top** (see below). The apply boundary still holds (`scripts/tg apply` needs an admin Vault token + cluster RBAC), but **pushing `master` is NOT inert** — the Forgejo→Woodpecker webhook fires `.woodpecker/default.yml` (`event: push, branch: master`, `require_approval: forks` only), which terragrunt-applies changed stacks. `master` is **branch-protected on Forgejo** (force-push disabled for everyone — history is append-only; push + merge whitelists = `viktor` + explicitly granted users, deploy keys allowed). **Allow-then-audit (Viktor, 2026-06-10):** `ebarzin` (emo) is on the whitelist and pushes straight to `master` — no PR gate. The tracking burden moves to: (a) **commit messages that record what + why** (the agent instructions in AGENTS.md and the managed claudeMd require the body to paraphrase the user's request), (b) the **`notify-nonadmin-push` Slack audit step** in `.woodpecker/default.yml` — every master push by a non-admin author is posted to Slack (admin pushes are not), and (c) non-admins **never use `[ci skip]`** so every change fires the pipeline (and thus the audit feed). Users NOT on the whitelist fall back to `<user>/<topic>` branches + PRs. **Clones stay fresh automatically** (2026-06-10): the hourly `t3-provision-users` reconcile runs `refresh_user_clone` over every managed clone — the infra clone and any workspace repos (fetch all remotes + fast-forward `master`, ONLY when on master with a clean tree and an upstream — dirty trees and local commits are left alone with a WARN) — and also `wire_forgejo_remote`, which idempotently adds the documented `forgejo` remote + `forgejo/master` upstream to infra clones that predate that contract. `start-claude.sh` does the same freshen at session launch (10s fetch cap per repo so an offline remote never stalls the session; workspace layouts freshen each repo under `~/code`).
 
 **Contribute access (per non-admin, manual — the anca/tripit PAT precedent):**
@@ -559,7 +561,7 @@ Separate from the in-cluster namespace-owner model above, the **devvm** (`10.0.1
 
 **Web-terminal session persistence (2026-06-10):** the tmux-based web terminal's named sessions (each running one Claude conversation) survive devvm reboots — `tmux-persist-save.timer` (5-min) snapshots every terminal user's sessions (name, cwd, conversation uuid from argv or the cwd-slug transcript dir) to `/var/lib/tmux-persist/<user>.tsv`, and `tmux-persist-restore.service` recreates missing sessions at boot with `claude --resume <uuid>` (per-session idempotent; also handles partial loss). The web terminal also exposes an **on-demand "Restore sessions" button** (terminal-lobby: `tmux-api` `POST /restore` → the validated root `tmux-restore-user` wrapper → `tmux-persist restore <user>`, a single-user mode of the same script): the boot-only restore service never fires when an **OOM kills a user's tmux server *without* a reboot** (the common case under multi-user memory pressure), so the button covers that gap. This is a **tmux/terminal-surface** feature, deliberately outside the t3 namespace: the t3 chat surface persists its own threads (`~/.t3` state, plus the daily `t3-backup-state` dump), and Claude conversations themselves were always durable (`~/.claude/projects/`) — what this adds is the volatile tmux wiring.
 
-**Status (2026-06-10):** built + verified on the live host — capacity (8 GiB swap), config inheritance, roster-driven provisioner, per-user locked clone, per-user OIDC kubeconfig + the `oidc-power-user-readonly` ClusterRole + emo's `k8s_users` entry (applied + impersonation-verified), the Authentik `T3 Users` edge gate, **the emo Phase-5 cutover (own clone + launcher repoint + `code-shared` removal, completed 2026-06-10) and emo's contribute access (`ebarzin` write collaborator + PAT + protected `master`)**, and **per-user `code_layout` with the ancamilea workspace cutover (infra → `~/code/infra`, `tripit` alongside, 2026-06-10)**. Per the live `/etc/skel` design, non-admin `~/.claude/{rules,skills}` symlinks into the admin base are **kept** (they ARE the shared-base delivery mechanism — the plan's step to remove them is obsolete). **Remaining (held / future):** the offboarding apply-side (Phase 7), per-user MCP/auth injection, and roster-reconciled `T3 Users` membership. See `../runbooks/offboard-user.md` for deprovisioning.
+**Status (2026-06-10):** built + verified on the live host — capacity (8 GiB swap), config inheritance, roster-driven provisioner, per-user locked clone, per-user OIDC kubeconfig + the `oidc-power-user-readonly` ClusterRole + emo's `k8s_users` entry (applied + impersonation-verified), the Authentik `T3 Users` edge gate, **the emo Phase-5 cutover (own clone + launcher repoint + `code-shared` removal, completed 2026-06-10) and emo's contribute access (`ebarzin` write collaborator + PAT + protected `master`)**, and **per-user `code_layout` with the ancamilea workspace cutover (infra → `~/code/infra`, `tripit` alongside, 2026-06-10)**. Per the live `/etc/skel` design, non-admin `~/.claude/{rules,skills}` symlinks into the admin base are **kept** (they ARE the shared-base delivery mechanism — the plan's step to remove them is obsolete). **Remaining (held / future):** the offboarding apply-side (Phase 7), the rest of per-user MCP/auth injection (`ha` + `claude_memory` + `.credentials.json` + beads Dolt cred — **per-user playwright browser MCP done 2026-06-16**, see above), and roster-reconciled `T3 Users` membership. See `../runbooks/offboard-user.md` for deprovisioning.
 
 ## Related