…n speed

Edit page (ourdream-style):
- remove tab system -> single scrolling page
- fixed W++ trait set (no format picker / dynamic attributes); reuse
  appearance/gender/age, add species/personality/mind/likes/dislikes/background
- read-only warning for non-standard W++ attributes (dropped on save)
- remove Voice + Advanced sections
- creator auto-set to user handle; characterVersion auto-increments

Image:
- gallery max 10, pick up to 4 for the LoRA; "Build LoRA" encodes via Z-Image
  and stores loraName on the character
- server-side asset->base64 resolution (SSRF-guarded, avoids CDN CORS)
- native generate_image tool: model can request an image mid-chat -> Z-Image
  with the character's LoRA, appended to the reply
- resolve the chat character for every image so the LoRA always applies
- size by source: chat 512, character/avatar 768

Progression:
- per-character speed (slow/normal/fast) -> XP per message 1/3/7

Other:
- summaries no longer charge credits or advance XP
- plain/inference prompts sent as user role (endpoint requires a user message)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Character traits:
- align fixed persona traits to the canonical Charluv W++ keys (species, mind,
  personality, age, job, description, sexuality, likes, loves, zodiac, hates,
  country, body, appearance) so real characters round-trip instead of being
  dropped as "extras"
- gender moves out of the persona into charluv metadata (char.gender); art style
  is derived from tags (anime/realistic)
- inject charluv metadata (gender, age range) into the prompt at chat time so the
  model still sees facets that aren't in the W++ persona

Gallery:
- re-seed the gallery from the saved character once it loads (was captured empty
  at mount, so images vanished on reopen / only appeared after adding one)
- readAssetBase64 reads directly from S3/R2 when storage is enabled, so the
  Build-LoRA encode can resolve gallery URLs regardless of ASSET_URL/CDN config

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

- Discover card click now opens a character profile/preview page
  (/discover/:id) with the gallery, age, description, facets and an editable
  name field instead of instantly matching
- "Match & Start Chat" clones the template with the chosen name, creates a chat
  directly and jumps to /chat/:id — the create-chat form step is bypassed
- server: POST /match/:id honors an optional name; GET /match/:id returns the
  template character for the profile page (registered after /discover)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

- Create wizard rebuilt as an ourdream-style image-card flow: gender/style/age,
  ethnicity + skin tone, hair + hair/eye color, body/breast/butt, personality,
  finish. Step-icon progress, two-tone hero, color swatches for skin/hair/eyes.
- Option previews load from /assets/wizard/...; appearance groups are gender +
  style aware: /assets/wizard/<style>/<gender>/<group>/<slug>.png. Missing
  assets degrade to labelled placeholders.
- Profile page: large two-tone "Name, Age" title (name live-editable).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

- srv/memory: server-side embeddings via @xenova/transformers (all-MiniLM-L6-v2,
  lazy in-process) and a Mongo-backed vector store with brute-force cosine
  retrieval, scoped per user+character (persists across chats)
- native `remember` tool: the model durably saves facts about the user/relationship;
  stored on the assistant turn via the existing tool-call meta channel
- RAG: recall top-k relevant memories from the recent turns and inject them into
  the {{memory}} prompt slot at generation time
- replaces memory books; backend swappable for pgvector/endpoint later

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Root cause: image persistence was gated on the SAVE_IMAGES flag (unset), so chat
images were uploaded to R2 but never attached to a chat message — they showed
live over WS then vanished from history on refresh (orphaning the R2 object).

- persist whenever the image belongs to a chat (has a chatId), regardless of
  SAVE_IMAGES; the flag still forces saving for non-chat images
- the native image tool now creates its OWN image message (parented to the reply)
  instead of appending onto the text turn — appending set adapter:'image' on the
  text message, which the renderer treats as an image filename and would blank
  the reply text

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…tics

- memory logs now carry only metadata (length/count/topScore); the actual text
  is gated behind MEMORY_DEBUG_LOG_TEXT=1 (dev only)
- openai adapter: debug-log when tools are attached and whether the model
  returned any tool_calls (finish_reason + names) to diagnose tool-calling

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…he field)

The single-page editor dropped the culture input, but the form guard still
required culture:'string', so getStrictForm rejected save/create with
"Request body is invalid: .culture is undefined".

- newCharGuard: culture -> optional
- getPayload: default culture from state / defaultCulture
- create wizard: set culture 'en-us' on the payload

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Native function-calling doesn't fire on the charluv finetune: the prompt
prefills the assistant turn (the "<name>:" steering), so the model generates
text mid-turn and can't emit tool_calls (logs showed finish_reason:stop,
toolCallCount:0 with tools correctly attached + qwen3_xml parser).

Switch to in-band markers the model writes inside its reply:
  <image>visual description</image>   -> fires Z-Image (char LoRA)
  <remember>fact</remember>          -> stores long-term memory
Server parses + strips them from the reply and feeds the existing downstream
(meta.imageTool / meta.rememberFacts). Markers are also stripped from streaming
partials so raw tags never show. Instructions are prepended to the system
message when the features are available.

Native OpenAI tools/tool_calls remain behind CHARLUV_NATIVE_TOOLS=1.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

The dream-date wizard had no way to set an image. Added a "Portrait (optional)"
section on the Finish step:
- Generate from choices: text-to-image via Z-Image using a prompt built from the
  selected style/ethnicity/gender/appearance (no LoRA yet)
- Upload: pick an image file
The chosen image is set as the new character's avatar in the create payload.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Reword the marker instructions from a meta "you have special abilities / the
system processes tags" into an in-world platform capability ("on Charluv, every
companion can share photos and remember things"). Characters are far more likely
to use the tags in-character; the previous framing was easy for a restrictive
persona to refuse.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

- Dockerfile: pnpm rebuild sharp — pnpm v10 skips dependency build scripts, so
  @xenova/transformers' sharp native binary was missing, breaking long-term
  memory ("Cannot find module sharp-linux-x64.node"). Image gen was fine (remote
  Z-Image); only the embedder failed.
- Marker instructions made more selective: at most one <image> per reply and only
  when fitting; <remember> only for genuinely important facts — the model was
  emitting both on nearly every turn. Also told not to narrate doing it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Both are superseded:
- memory books -> long-term memory (RAG) now fills {{memory}} server-side; the
  keyword-book injection is removed (it was leaking the old character_book, e.g.
  "Julia's cat...", into unrelated chats)
- attached scenario books -> relationship progression (the LEVEL token). Only the
  character's initial scenario (or the chat's overridden scenario) remains, plus
  the progression token + charluv meta.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

The native image tool created a detached image message that popped in with no
feedback. Now it behaves like normal in-chat image gen:
- server attaches the image to the reply (messageId + append) and no longer sets
  adapter:'image' on append (preserved adapter, so the reply text still renders)
- image-generation-started is broadcast to chat members with the messageId
- client shows a Spinner on the reply while generating (imagesGenerating state),
  renders the attached image from extras under the text, clears on message-retry

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

`pnpm rebuild sharp` silently no-ops under pnpm v10's build-script gate, so the
native binary was still missing. Run sharp's own install script directly in the
pnpm store path to fetch the prebuilt binary, unblocking the memory embedder.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

- One "Images" card: cover (avatar) + gallery in a single grid; every tile
  (incl. the avatar) is selectable for the LoRA; Generate / Upload / Build LoRA /
  Make-cover / Remove. Server: encode-lora allows the avatar ref; new
  POST /character/:id/cover sets the avatar from an existing image.
- Gallery "Generate" now composes a real prompt from the persona
  (appearance/body/species/age + name) instead of the empty appearance field.
- Sprite feature removed from the editor UI (schema field kept, no migration;
  original sprite data forwarded on save).
- LoRA re-encode reuses the character's existing loraName so it REPLACES the old
  LoRA instead of minting a new uuid each time.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…22 bookworm

- remove dead deps: all @parcel/* + parcel (Vite is the bundler), localtunnel +
  @types/localtunnel (dev tunnel); drop their npm scripts and the localtunnel
  usage in srv/start.ts, srv/bin.ts, srv/config.ts; drop .parcelrc from Docker
- add pnpm.overrides (33) pinning patched transitive deps (protobufjs, pbkdf2,
  sha.js, fast-xml-parser, shell-quote, undici, esbuild, rollup, qs, ws, ...) —
  clears all 6 criticals + most highs
- bump dompurify, exifreader, js-cookie, svgo, vite to patched
- Dockerfile: node:22.14.0-bullseye-slim -> node:22-bookworm-slim

Remaining 9 are justified (no patch / major-break / non-exploitable):
pdfjs-dist+webpack (v4 layout break), vite (dev-server only, v6 major),
uuid (v4-only usage, advisory is v3/5/6), showdown/elliptic (no patch).
discord.js kept — still imported (Discord rip is a later phase).

srv build:server exit 0; web tsc unchanged (790); frozen-lockfile OK.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…build)

The `>=7.29.6` override resolved to Babel 8.0.1, which is incompatible with
babel-preset-solid and crashed `vite build` in CI. Pin to `^7.29.6` (resolves
7.29.7) — still patched, builds clean (verified locally, 1m42s).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

The remaining 9 advisories cannot be fixed without breaking the build or removing
a kept feature, or are non-exploitable here. Recorded them in
pnpm.auditConfig.ignoreGhsas with per-advisory justification in
SECURITY-ACCEPTED.md, so `pnpm audit` now exits 0 and each is tracked.

- no upstream patch: elliptic, showdown
- non-exploitable + fix breaks CJS build: uuid (v4-only usage)
- dev-server/Windows-only, fix needs Vite 6 (breaks node-polyfills build): vite x3
- pdfjs-dist high + its webpack peer x2: v4 ESM/worker break entangled with
  pdfdataextract; exploit is self-inflicted (user's own PDF, text extraction)

Net: audit 154 -> 0 active (6 crit + 59/61 high removed via deps/overrides/bumps).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…rrides

The `path-to-regexp: >=0.1.13` override forced express@4.21.2 onto
path-to-regexp@8.x (incompatible API) -> server crashed at boot
("pathRegexp is not a function" registering routes). Also `lodash: >=4.18.0`
was a bogus range (latest is 4.17.21).

- replace blanket override with scoped `express>path-to-regexp: 0.1.13` (patched
  AND keeps the 0.1.x function API express needs; other consumers keep 8.x)
- remove the invalid lodash override

Verified: `pnpm audit` exit 0, build:server ok, and the server BOOTS
(MongoDB + "Server started") — the Docker build alone didn't catch the runtime
crash, so this was boot-tested.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Broaden the remember instruction (marker + native tool) so the model also saves
self-facts it states or invents in the moment — its pet's name, family, where it
lives, backstory — not just facts about the user. Keeps the character consistent
across chats (these get stored per user+character and recalled via RAG).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

… screen)

- My AI card now opens an owned-character profile (/mine/:id): big cover +
  gallery, click any image for fullscreen (settingStore.showImage), Chat + Edit +
  Back buttons. Reuses the Discover profile styling.
- startChat(char, navigate, {forceNew?}) helper: resume the most-recent chat or
  create one, then go straight to /chat/:id. Replaces per-character
  /chats/create/:id navigations across the character views (no more confusing
  create-chat screen).
- ChatList: per-chat inline rename (pencil -> editChat) — replaces the only
  feature the create-chat screen offered (naming the chat).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…My AI)

Wrap the Discover profile cover + gallery images in buttons that open the
fullscreen viewer (settingStore.showImage), matching the My AI profile.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

- Every character carries an imageSeed (loaded or freshly generated). The editor's
  image generation (cover + gallery) sends this seed so all generated images keep
  a consistent base look — good for building a coherent LoRA reference set.
- Chat image generation omits the seed entirely, so chat images vary (the LoRA
  provides identity). generateImage now uses opts.seed (editor-supplied) instead
  of reading character.imageSeed, so chat never locks it.
- Editor UI: a "Base look seed" row with the value + a Dice "Reroll" button and a
  warning that rerolling changes the base appearance (saved images unaffected).
- Threaded seed: editor -> generateAvatar -> characterStore.generateAvatar ->
  generateImageWithPrompt -> POST /character/image -> generateImage. Persisted via
  the character create/update mapping + FormData.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

… seed)

With the locked base-look seed, generating again with the same Appearance prompt
yields the same image. Added a note under the gallery actions telling users to
vary the Appearance prompt (pose/outfit/setting) between generations for varied
reference shots while keeping the core looks consistent.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Wire isPresent into the PriorityGate singleton so the gate can drop
requests from disconnected clients. Start/stop the presence-refresh
interval in startQueue(). Mark presence on socket open and user login;
clear on socket close, error, and user logout.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

…nect

Two paths charged credits before entering the gate. When ClientGoneError
is thrown (client gone before generation), image generation now refunds
IMAGE_COST (logged-in users only) and guidance refunds 10 credits.
IMAGE_COST is now exported from srv/image/index.ts and imported by the
image route instead of being redefined locally.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XXNb2Kcdayr1ZvUpy5a3nA

Fixes the PR format gate (prettier --check) that blocked merge to trunk.
Formatting only — no behaviour changes.

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

setAvatarFile(file) tripped the Setter's `U extends Function` overload branch
under CI's TypeScript. Use the functional-updater form. Surfaced now that the
prettier gate passes and CI reaches the typecheck step.

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

Align the test suite with the rewrite:
- resolve-scenario: scenario books are ignored (feature disabled)
- progression: unknown archetype id returns no steps (no fallback)
- prompt: remove keyword memory-book tests (feature removed)
- refresh snapshots for CHARLUV_SAFEGUARD_PROMPT and dropped memory entries

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