perf+qa: alwaysLoad engine tools + finish-tool for honest gave_up (validated on .app)

.mcp.json

@@ -3,6 +3,7 @@
     "clawdnd-engine": {
       "type": "stdio",
       "command": "uv",
+      "alwaysLoad": true,
       "args": ["run", "--directory", "${CLAUDE_PLUGIN_ROOT}/servers/engine", "server.py"]
     },
     "clawdnd-rules": {

gaveupfix-vet.log

@@ -0,0 +1,61 @@
+[uipt-app] run=gaveupfix-vet world=baldurs-gate persona=veteran beats=40 budget=$8.00 part=AB player_agent=claude
+[uipt-app] build_sha=f1c796c version=v1.0.3-138-gf1c796c repo=/private/tmp/wos-gate
+[uipt-app] === PART A: native-transition gate (re-verifies #356) ===
+[A] pkill WorldOSApp + THIS checkout's stale viewers, rm -rf /private/tmp/wos-gate/dist/WorldOS.app, fresh build…
+[A] WOS_APP_NO_GLOBAL_KILL=1 — preserving other WorldOSApp processes.
+[A] launcher viewer ready on 8766; can_act(before)={"can_act":false,"is_live_view":false,"live":false}
+[A] before.png deferred to orchestrator
+[A] raising WorldOS to front + CGEvent-clicking the RESUME → PLAY CTA (with retries)…
+[A] window X=97.0 Y=36.0 W=1823.0 H=956.0; front='WorldOS'; click=(1701,240)
+[A] click posted.
+[A] polling for a minted live session (new run dir + can_act:true on a new port; deadline 420s / 140 polls)…
+[A] MINTED: run=play-20260602114235 port=8767 can_act=true (after 7 polls)
+[A] after.png deferred to orchestrator
+[A] tearing down minted backend (run=play-20260602114235) — cold-open was enough.
+[A] RESULT: PASS — #356 banner minted a live DM (can_act:true). surface={"can_act":true,"is_live_view":true,"live":true,"campaignId":null,"enabledActions":[]}
+<stdin>:23: DeprecationWarning: datetime.datetime.utcnow() is deprecated and scheduled for removal in a future version. Use timezone-aware objects to represent datetimes in UTC: datetime.datetime.now(datetime.UTC).
+[A] wrote /private/tmp/wos-gate/qa/ui_playtest_runs/gaveupfix-vet/native/transition.json
+[uipt-app] === PART B: persona loop on the .app-faithful backend ===
+[uipt-app] [B] Playwright not installed at /private/tmp/wos-gate/qa/playwright. Run: (cd qa/playwright && npm install && npx playwright install chromium)
+<stdin>:25: DeprecationWarning: datetime.datetime.utcnow() is deprecated and scheduled for removal in a future version. Use timezone-aware objects to represent datetimes in UTC: datetime.datetime.now(datetime.UTC).
+[uipt-app] === DONE. dir=/private/tmp/wos-gate/qa/ui_playtest_runs/gaveupfix-vet ===
+[uipt-app] part A (#356 gate): PASS   part B (persona loop): no_playwright
+[uipt-app] spend: DM ~$0 + player ~$0 = ~$0.0000 (budget $8.00)
+----- run.json -----
+{
+  "run": "gaveupfix-vet",
+  "world": "baldurs-gate",
+  "persona": "veteran",
+  "beats_cap": 40,
+  "budget_usd": 8.0,
+  "build_sha": "f1c796c",
+  "version": "v1.0.3-138-gf1c796c",
+  "part": "AB",
+  "part_a": {
+    "gate": "native_transition_356",
+    "result": "PASS",
+    "original_result": "PASS",
+    "failure_bucket": null,
+    "failure_detail": null,
+    "minted_run_dir": "play-20260602114235",
+    "minted_port": 8767,
+    "kept_backend_alive": false,
+    "first_turn_ready": false
+  },
+  "part_b": {
+    "persona_loop": "no_playwright",
+    "score_pass": false,
+    "provider": "claude",
+    "player_agent": "claude",
+    "original_result": "no_playwright",
+    "failure_bucket": "no_provider",
+    "failure_detail": "Playwright palette dependency is missing"
+  },
+  "spend_usd": {
+    "dm_and_companions": 0.0,
+    "player_agent": 0.0,
+    "total": 0.0
+  },
+  "surface": "BUILT dist/WorldOS.app (part A) + claude provider/claude player backend (part B)",
+  "at": "2026-06-02T11:42:56Z"
+}PLAYTEST_EXIT=1

qa/playwright/node_modules

@@ -0,0 +1 @@
+/Users/lume/ClawDnD-val/qa/playwright/node_modules

qa/playwright/palette_server.js

@@ -637,6 +637,29 @@ server.registerTool(
   }
 );
 
+server.registerTool(
+  "finish",
+  {
+    description:
+      "End the playtest because you have PLAYED ENOUGH to fairly judge the experience and you are " +
+      "NOT blocked (use give_up ONLY when you are genuinely stuck). Give your honest overall " +
+      "satisfaction (1-10) and a 1-2 sentence closing verdict. This is the normal, satisfied way to end.",
+    inputSchema: {
+      satisfaction: z.number().int().min(1).max(10).describe("Your honest overall satisfaction, 1-10."),
+      verdict: z.string().describe("A 1-2 sentence closing verdict."),
+    },
+  },
+  async ({ satisfaction, verdict }) => {
+    const pg = await ensurePage().catch(() => null);
+    const sc = pg ? await snap(pg, "finish") : "";
+    logAction("finish", { satisfaction, verdict: String(verdict).slice(0, 200), screenshot: sc });
+    try {
+      fs.writeFileSync(STATUS, JSON.stringify({ ended: true, reason: "finish", satisfaction, detail: String(verdict).slice(0, 500), at: nowIso() }));
+    } catch (_e) {}
+    return textResult({ ok: true, ended: true, note: "Run ended — thanks for playing." });
+  }
+);
+
 // Graceful browser teardown on exit.
 async function shutdown() {
   try {

qa/run_duo.sh

@@ -55,7 +55,7 @@ mkdir -p "$T" "$STATE_DIR"; rm -rf "$STATE_DIR/campaigns" 2>/dev/null
 DM_CFG="$STATE_DIR/dm.mcp.json"; PLAYER_CFG="$STATE_DIR/player.mcp.json"
 MOVES="$STATE_DIR/player_moves.jsonl"; : > "$MOVES"  # the player's structured moves (It.1)
 python3 - "$ROOT/qa/qa.mcp.example.json" "$STATE_DIR" "$DM_CFG" "$ROOT" <<'PY'
-import json, sys
+import json, sys, os
 cfg_path, state, out, root = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4]
 cfg = json.load(open(cfg_path))
 # RE-ROOT every MCP server's `--directory` at THIS repo ($ROOT) so the DM engine,
@@ -81,6 +81,11 @@ for name, srv in cfg.get("mcpServers", {}).items():
         args[i + 1] = f"{root}/servers/{pkg}"
     if name == "clawdnd-engine":
         srv.setdefault("env", {})["CLAWDND_STATE_DIR"] = state
+        # Parity with scripts/play.sh: pin the engine tools (un-defer) so the DM stops burning
+        # ~2 ToolSearch round-trips/beat re-discovering them. Set CLAWDND_ENGINE_ALWAYSLOAD=0 for
+        # the deferred baseline (the latency A/B arm).
+        if os.environ.get("CLAWDND_ENGINE_ALWAYSLOAD", "1") == "1":
+            srv["alwaysLoad"] = True
 json.dump(cfg, open(out, "w"))
 PY
 # The player gets ONLY the constrained move facade (clawdnd-player): it acts through

qa/ui_playtest_app.sh

@@ -660,7 +660,7 @@ PY
   cat > "$player_prompt" <<EOF
 $persona_brief
 
-You have a budget of about $BEATS actions for this whole session. Spend them trying to start and play the story, reporting friction as you go. When you have either (a) genuinely gotten stuck after reporting it, or (b) played a few real turns and seen enough to judge the experience, you may stop — give a final 1-2 sentence verdict and, if you got stuck, call give_up. Start now.
+You have a budget of about $BEATS actions for this whole session. Spend them trying to start and play the story, reporting friction as you go. Waiting for the DM to narrate is FREE and does not use your budget — a rich beat can take a few minutes, and the spinner / streaming text is PROGRESS, not a hang, so keep waiting rather than abandoning a beat. When you have played a few real turns and seen enough to judge the experience, call finish with your honest satisfaction (1-10) and a 1-2 sentence verdict. Call give_up ONLY if you are genuinely BLOCKED — a dead control, an error, or the DM truly stalled with no narration — NOT because a beat is slow. Either way, also end your final message with a line exactly like: Satisfaction: N/10. Start now.
 EOF
   log "[B] player agent starting (agent=$PLAYER_AGENT persona=$PERSONA, ~$BEATS actions, budget \$$player_budget)…"
   case "$PLAYER_AGENT" in
@@ -705,7 +705,7 @@ EOF
         -c "mcp_servers.clawdnd-uiplayer.env_vars=[\"CLAWDND_UIPT_URL\",\"CLAWDND_UIPT_RUNDIR\",\"CLAWDND_UIPT_CHANNEL\",\"CLAWDND_UIPT_PERSONA\"]" \
         -c "mcp_servers.clawdnd-uiplayer.required=true" \
         -c "mcp_servers.clawdnd-uiplayer.default_tools_approval_mode=\"approve\"" \
-        -c "mcp_servers.clawdnd-uiplayer.enabled_tools=[\"screenshot\",\"a11y_tree\",\"click\",\"type\",\"key\",\"wait\",\"report_bug\",\"give_up\"]" \
+        -c "mcp_servers.clawdnd-uiplayer.enabled_tools=[\"screenshot\",\"a11y_tree\",\"click\",\"type\",\"key\",\"wait\",\"report_bug\",\"give_up\",\"finish\"]" \
         - < "$player_prompt" > "$player_out" 2>> "$PLAYERDIR/player.err"
       ;;
   esac

qa/ui_playtest_score.py

@@ -147,23 +147,31 @@ def is_image_404(n: dict) -> bool:
     gave_up = status.get("reason") == "give_up"
 
     # --- satisfaction --------------------------------------------------------
-    satisfaction = extract_satisfaction(verdict)
-    if satisfaction is None:
-        # Derive a rough satisfaction when the player didn't state one: start at 8,
-        # subtract for the friction we measured. (Informational, clearly derived.)
-        s = 8
-        if not completed_intro_flow:
-            s -= 3
-        if gave_up:
-            s -= 2
-        s -= min(3, by_sev.get("critical", 0) * 2 + by_sev.get("major", 0))
-        s -= min(2, dead_clicks)
-        if console_errors:
-            s -= 1
-        satisfaction = clamp10(s)
-        satisfaction_source = "derived"
-    else:
+    # A `finish` tool call records a structured 1-10 satisfaction in status.json — the most
+    # reliable self-report (a validated tool arg, impossible to mis-parse). Prefer it; then a
+    # "N/10" in the verdict text; then derive from measured friction.
+    status_sat = status.get("satisfaction")
+    if isinstance(status_sat, (int, float)) and not isinstance(status_sat, bool):
+        satisfaction = clamp10(int(status_sat))
         satisfaction_source = "self-reported"
+    else:
+        satisfaction = extract_satisfaction(verdict)
+        if satisfaction is None:
+            # Derive a rough satisfaction when the player didn't state one: start at 8,
+            # subtract for the friction we measured. (Informational, clearly derived.)
+            s = 8
+            if not completed_intro_flow:
+                s -= 3
+            if gave_up:
+                s -= 2
+            s -= min(3, by_sev.get("critical", 0) * 2 + by_sev.get("major", 0))
+            s -= min(2, dead_clicks)
+            if console_errors:
+                s -= 1
+            satisfaction = clamp10(s)
+            satisfaction_source = "derived"
+        else:
+            satisfaction_source = "self-reported"
 
     critical = by_sev.get("critical", 0)
     major = by_sev.get("major", 0)

scripts/play.sh

@@ -105,7 +105,7 @@ python3 - "$ROOT" "$STATE_DIR" "$DM_CFG" <<'PY'
 import json, sys
 root, state_dir, out = sys.argv[1], sys.argv[2], sys.argv[3]
 cfg = {"mcpServers": {
-    "clawdnd-engine": {"type": "stdio", "command": "uv",
+    "clawdnd-engine": {"type": "stdio", "command": "uv", "alwaysLoad": True,
         "args": ["run", "--directory", f"{root}/servers/engine", "server.py"],
         "env": {"CLAWDND_STATE_DIR": state_dir}},
     "clawdnd-rules": {"type": "stdio", "command": "uv",

scripts/play_party.sh

@@ -141,7 +141,7 @@ python3 - "$ROOT" "$STATE_DIR" "$DM_CFG" <<'PY'
 import json, sys
 root, state_dir, out = sys.argv[1], sys.argv[2], sys.argv[3]
 cfg = {"mcpServers": {
-    "clawdnd-engine": {"type": "stdio", "command": "uv",
+    "clawdnd-engine": {"type": "stdio", "command": "uv", "alwaysLoad": True,
         "args": ["run", "--directory", f"{root}/servers/engine", "server.py"],
         "env": {"CLAWDND_STATE_DIR": state_dir}},
     "clawdnd-rules": {"type": "stdio", "command": "uv",