chore(models): refresh stale model defaults (NodeRoom)

.env.example

@@ -59,10 +59,13 @@ AGENT_QUALITY_CANDIDATE_TIMEOUT_MS=45000
 OPENROUTER_FREE_MODEL_CACHE_MS=600000
 OPENROUTER_REQUIRE_NO_TRAINING=1
 OPENROUTER_FREE_ALLOW_FILE_EGRESS=0
-# Provider-neutral NodeAgent uses OpenRouter's governed free ladder first, then
-# this direct Gemini route when free routes are unavailable. Receipts retain the
-# catalog price because an account's free-tier entitlement cannot be asserted.
-NODEAGENT_FREE_AUTO_GOOGLE_MODEL=gemini-3.5-flash
+# Provider-neutral NodeAgent uses OpenRouter's governed free ladder only. This
+# direct Gemini route is an opt-in paid leg: it is taken only when the caller sets
+# allowPaidFallback AND passes a spend meter (adapter.ts), and no production caller
+# does either, so by default a free-route outage is an error, not a paid request.
+# Receipts retain the catalog price because an account's free-tier entitlement
+# cannot be asserted.
+NODEAGENT_FREE_AUTO_GOOGLE_MODEL=gemini-3.8-flash
 AGENT_MODEL=gpt-5.4-nano
 # Optional voice provider overrides. Browser TTS is the free client default;
 # Convex `/voice/*` endpoints are authenticated hosted fallbacks.
@@ -81,7 +84,7 @@ VOICE_CLOUDFLARE_TTS_MODEL=@cf/myshell-ai/melotts
 VOICE_CLOUDFLARE_TTS_LANG=en
 GEMINI_API_KEY=
 VOICE_GEMINI_TTS_FREE_TIER_CONFIRMED=0
-VOICE_GEMINI_TTS_MODEL=gemini-3.1-flash-tts-preview
+VOICE_GEMINI_TTS_MODEL=gemini-3.8-flash-tts
 VOICE_GEMINI_TTS_VOICE=Kore
 VOICE_GEMINI_TTS_INSTRUCTIONS=Read clearly and concisely for a collaborative work room.
 GOOGLE_CLOUD_TTS_API_KEY=
@@ -113,7 +116,7 @@ SEO_QA_GOOGLE_TARGET_PHRASE=NodeRoom collaborative AI room
 GOOGLE_SEARCH_CONSOLE_SITE_URL=https://noderoom.live/
 GOOGLE_SEARCH_CONSOLE_ACCESS_TOKEN=
 GOOGLE_SEARCH_CONSOLE_SERVICE_ACCOUNT_JSON=
-GEMINI_SEO_VIDEO_JUDGE_MODEL=gemini-3.5-flash
+GEMINI_SEO_VIDEO_JUDGE_MODEL=gemini-3.6-flash
 NEBIUS_SEO_JUDGE_MODEL=nebius/MiniMaxAI/MiniMax-M2.5
 NEBIUS_SMOKE_MODEL=nebius/MiniMaxAI/MiniMax-M2.5
 NEBIUS_ENABLE_VISION_JUDGE=0
@@ -162,8 +165,8 @@ FREE_JOB_GOAL=
 FREE_JOB_TIMEOUT_MS=900000
 
 # Optional server-side parser model overrides.
-PROVIDER_PARSER_GEMINI_MODEL=gemini-3.5-flash
+PROVIDER_PARSER_GEMINI_MODEL=gemini-3.8-flash,gemini-3.6-flash
 PROVIDER_PARSER_OPENAI_MODEL=gpt-5.4-mini
-PROVIDER_PARSER_ANTHROPIC_MODEL=claude-haiku-4-5
-PROVIDER_PARSER_OPENROUTER_MODEL=z-ai/glm-5.2,moonshotai/kimi-k2.7-code,cohere/north-mini-code:free,nvidia/nemotron-3-ultra-550b-a55b,nex-agi/nex-n2-pro
+PROVIDER_PARSER_ANTHROPIC_MODEL=claude-haiku-5-5,claude-sonnet-5-5
+PROVIDER_PARSER_OPENROUTER_MODEL=z-ai/glm-5.2,moonshotai/kimi-k2.7-code,cohere/north-mini-code:free,nvidia/nemotron-3-ultra-550b-a55b
 PROVIDER_PARSER_ALLOW_FILE_EGRESS=0

.tours/01-primary-user-flow.tour

@@ -27,14 +27,14 @@
     },
     {
       "file": "src/ui/Chat.tsx",
-      "line": 1878,
+      "line": 1879,
       "pattern": "const send = \\(raw\\?: string",
       "title": "The primary user action",
       "description": "`send` is what the Send button calls. Two separate things happen and it matters that they are separate:\n\n1. The message is posted to the room, with a client-generated `clientMsgId` — that id is how a retry updates the same message instead of double-posting it.\n2. Only then is the text inspected to see whether it was addressed to the assistant."
     },
     {
       "file": "src/ui/Chat.tsx",
-      "line": 1166,
+      "line": 1167,
       "pattern": "function parsePublicNodeAgentRequest",
       "title": "Where text becomes intent",
       "description": "The assistant that can edit the shared sheet is addressable ONLY as a leading `@nodeagent …` directive. Mentioning it mid-sentence does not summon it.\n\nThis decides intent. It does not decide permission — that happens on the server, in the next step but one."
@@ -48,7 +48,7 @@
     },
     {
       "file": "convex/agentJobs.ts",
-      "line": 2782,
+      "line": 2799,
       "pattern": "export const startPublicAsk = mutation\\(",
       "title": "Where input becomes trusted (live tier)",
       "description": "This is the real trust boundary. Convex validators reject a malformed payload before a job row exists, and `requireActorProof` checks that this browser is actually a member of this room.\n\nThe client-side parse two steps back exists to decide what the user meant. This decides what they are allowed to do."

.tours/03-debug-and-recovery.tour

@@ -27,21 +27,21 @@
     },
     {
       "file": "src/ui/Chat.tsx",
-      "line": 269,
+      "line": 270,
       "pattern": "const MAX_FAILED_SENDS = 50;",
       "title": "Failure 3 — a message fails to post, and the list is bounded",
       "description": "`MAX_FAILED_SENDS = 50`, FIFO. Every in-memory collection on an agent path needs a cap and an eviction rule; agents loop, and an unbounded list fills memory in minutes rather than months."
     },
     {
       "file": "src/ui/Chat.tsx",
-      "line": 1943,
+      "line": 1944,
       "pattern": "const retrySend = \\(cid: string, text: string\\)",
       "title": "Retry that cannot double-post",
       "description": "`retrySend` re-sends with the SAME `clientMsgId`. That id is what makes the retry idempotent at the store, so a user hammering Retry on a flaky connection does not produce five copies of their message."
     },
     {
       "file": "src/ui/Chat.tsx",
-      "line": 935,
+      "line": 936,
       "pattern": "function buildAgentFailureNotice\\(",
       "title": "Failure 4 — the agent fails",
       "description": "`buildAgentFailureNotice` turns a provider error into a sentence a non-engineer can act on, carrying which model selection was used, what was asked, and which job it was. A stack trace on this surface is a defect, not a debug aid — this is a surface where a user decides whether to trust the assistant."

CHANGELOG/README.md

@@ -23,6 +23,7 @@ entries have not been invented or backfilled.
 ## Quality assurance
 
 - [Benchmark timeout budgets](qa/benchmark-timeout-budgets.md)
+- [Model default refresh](qa/model-refresh-noderoom.md)
 - [NodeKit note reference proof](qa/nodekit-note-reference-proof.md)
 
 See [TEMPLATE.md](TEMPLATE.md) before adding or editing a lane.