v0.12.0-real-engine - Make everything real (W1-W4)

[v0.12.0-real-engine] - 2026-06-16 - Make everything real (W1-W4)

The "make everything real" workstream: tool stubs are gone and the combat layer runs on real rules. Full design in the planning dir spec 2026-06-09-dmai-make-everything-real-design.md.

Added

• W1 combat engine: real initiative is rolled and sorted server-side; the frontend renders the backend initiative order.
• W1 turn-gating: token dragging is restricted to the active PC token, the movement budget is consumed on drag, and dead tokens (hp<=0) cannot be moved.
• W1 conditions gate movement and actions; the ActionBar is shown only on the player's turn; active-turn highlight plus token-move animation.
• W1 damage resistance/immunity/vulnerability applied in apply_damage (migration 0006_token_resistances).
• W2 query_rules performs real SRD retrieval through embeddings (SrdRetriever).
• W2 set_scene persists to a new scenes table (migration 0005_scenes) and the current scene is injected into model context.
• W2 quick_save/load captures and restores real combat plus scene state end-to-end.
• W2 real spell resolution: compendium lookup, damage, resistances, and saves.
• W3 end-to-end video generation: sidecar /video/generate SSE, generate_video tool, VideoGenerated event, and <video> rendering in chat.
• generate_image split into generate_map (top-down, routed to the VTT) and generate_illustration (cinematic, routed to a chat tool-call card).
• VTT zoom/pan/fit-to-view and a layers popover (grid / scene image / tokens).
• Local text-only models Qwen3-0.6B and Qwen3-8B via ISQ.
• VRAM auto-swap around image generation: stops the local LLM, frees the card for SDXL, and restarts the LLM on the same port.

Changed

• W4 honest pending indicator replaces the fake cycling-digit progress.
• W4 agentTurnInFlight disables the model switch while a turn is in flight.
• W4 danger Button variant for destructive confirmations.
• W4 journal renders markdown as sanitized rich HTML.
• Migrated the local LLM sidecar from the deprecated mistralrs-server to mistralrs-cli serve (clean structured tool_calls), with a parser fallback for tool-calls that leak into delta.content.

Known limitations

• 7 biome warnings remain (non-null assertions in tests, exhaustive-deps, descending-specificity); they do not block the gates.
• GGUF loading crashes after load on non-TTY Windows (upstream mistralrs); the default Gemma/ISQ path works.
• Rendering a generated illustration directly onto the VTT map is deferred (the image is generated and delivered over SSE).

v0.11.0-m11

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server,
  plus src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.10.0-m10

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server,
  plus src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.9.0-m9

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.8.0-m8

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.7.2-m7

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.7.1-m7-5: M7.5-DM real GPU loaders close

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.7.0-m7

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.6.0-m6

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.

v0.6.0-m5: ci(release): drop darwin-x86_64 from first-GA matrix

Changelog

All notable changes to this project are tracked here. The format follows
Keep a Changelog and the project
uses Semantic Versioning.

Tags before v1.0.0 are pre-release milestones; the public surface and
on-disk schema may change between them. See docs/RELEASE.md for the
release pipeline and docs/RELEASE_CHECKLIST.md for the per-release
manual smoke list.

[Unreleased] - M5 polish

Added

• Persistent SessionSlice (src/state/session.ts) with lazily-minted
  (activeCampaignId, activeSessionId) UUID pair, persisted via the
  split-storage adapter (settings.json).
• useSession hook that ensures a session exists on app mount and
  rehydrates the chat from GET /sessions/{id}/messages so reopening a
  campaign restores chat history (incl. images sent in prior sessions).
• Frontend Stronghold-encrypted secrets vault via
  tauri-plugin-stronghold 2.x and @tauri-apps/plugin-stronghold.
  Provider configs and the Replicate API key now sit in
  dmai-vault.hold under the OS-native app data dir; argon2 salt at
  app_local_data_dir()/salt.txt is lazily created. One-shot migration
  drains the pre-M5 plaintext secrets.json into the vault on first
  read (closes M4 carry-over E.2 frontend half).
• Backend StrongholdSecretsRepo via iota_stronghold 2.x (closes M4
  carry-over E.2 backend half). dmai-server now opens an encrypted
  Stronghold snapshot on startup (path: DMAI_VAULT_PATH env override
  or sibling to the SQLite db file) and uses it as the SecretsRepo.
  Survives a sidecar restart without the frontend re-delivering keys
  via /settings POST.

Changed

• useAgentTurn and useChat now read campaign and session IDs from
  the persistent SessionSlice instead of hardcoded zero-UUID
  placeholders.
• streamChat invocations pass the active sessionId so the backend
  persists user and assistant rows across launches.

Tests

• 181 vitest (+5 from Stronghold + session work) and 210 cargo
  (+4 stronghold round-trip tests).

[v0.5.5-m4.5] - 2026-05-09 - Vision input + chat persistence

Added

• Multimodal user messages: MessagePart enum (Text / Image) on
  both backend and frontend; LlmProvider::supports_vision() capability
  flag (Anthropic / OpenAICompat / mistralrs report true).
• Composer attachments: paperclip button, paste-from-clipboard, and
  drag-drop on the chat panel; PNG/JPEG/WebP, ≤ 5 MB, ≤ 4 images per
  message; thumbnail strip and ImageLightboxModal for full-size view.
• Backend chat persistence: messages table (migration
  0003_m4_5_messages.sql), db::insert_message,
  db::list_messages_by_session, and GET /sessions/{id}/messages
  endpoint returning Vec<ChatMessage> JSON.
• /chat and /agent/turn accept an optional session_id and persist
  user, assistant, and tool-result rows when present.
• HF Hub real tree-API walker (/api/models/{repo}/tree/{rev}) for the
  diffusers folder downloader; injectable HfEndpoints for testing.
• E2E vision-flow Playwright spec + RELEASE_CHECKLIST M4.5 section.

Changed

• ChatMessage::User { content: String } -> ChatMessage::User { parts: Vec<MessagePart> }; ChatMessage::user_text(impl Into<String>)
  helper for the common text-only path.
• HttpMessage accepts both legacy {role, content} and new {role, parts} shapes via custom dual-shape Deserialize (size guards: 5 MB
  per image, 4 images per message - return HTTP 413 on overrun).

Tests

• 206 cargo + 167 vitest (cargo per-crate to avoid Tauri-bin
  PDB-overflow LNK1140 on Windows).

[v0.5.0-m4] - 2026-05-08 - Local Mode + cross-platform packaging

Added

• Local LLM via mistralrs-server Tauri sidecar binary (pinned
  MISTRALRS_VERSION=v0.8.0), ProviderConfig variant local-mistralrs
  hot-swappable from Settings.
• Local image generation via Python sidecar (FastAPI + diffusers
  SDXL-Turbo), GPU coordination mutex (auto-swap LLM unload before
  image-gen).
• Model manifest (Qwen3.5 0.8B/2B/4B/9B + SDXL-Turbo), download
  manager with resumable HTTP Range + sha256 verify, per-model state
  machine with cancel.
• LocalRuntime state machine, free-port discovery, health-probe with
  exponential backoff retry, RuntimeRegistry aggregating two
  runtimes.
• Tauri updater plugin + UpdateAvailableModal + useUpdater hook
  (pubkey placeholder until first release).
• SecretsRepo trait + InMemorySecretsRepo (Stronghold real impl
  deferred to M5).
• Local Mode UI: LocalModeSlice, useModelDownload SSE hook,
  useLocalRuntimeStatus 5 s poll, ModelDownloadCard,
  RuntimeStatusPill, LocalModeModal.
• Cargo features: default with-local-runtime, cloud-only gates
  Local Mode routes off via cfg-attribute.
• CI: build-bundle.yml matrix (4 targets, no signing) + release.yml
  with Win EV cert + macOS rcodesign + notarytool + Linux AppImage.
• RELEASE_CHECKLIST.md for manual smoke after each release.

Tests

• 179 cargo + 157 vitest.

[v0.4.0-m3] - 2026-05-08 - LLM agent loop + RAG + journal + NPC

Added

• AgentOrchestrator N-round tool-call loop with a 15-tool surface
  (combat: roll_dice, apply_damage, start_combat, end_combat,
  add_token, update_token, remove_token; narrative: set_scene,
  cast_spell, remember_npc, recall_npc, journal_append,
  quick_save, generate_image, query_rules).
• SRD content (20 spells / 15 monsters / 30 rules / 4 classes) +
  fastembed BGE-small-en (default switched to MultilingualE5Small for
  RU+EN coverage) + cosine retriever; RAG injection in
  context_builder.
• Journal: journal_entries table, JournalViewer overlay (full-screen
  modal, EN/RU, parchment polish deferred to M5).
• NPC memory: npc_memory table, NpcMemoryGrid (responsive 320x420
  dossier cards, name/role search, disposition chip filters).
• ImageProvider trait + ReplicateProvider (90 s deadline + 10 s
  per-request timeouts + status-checked polling) + cache-key
  (FNV-1a 64-bit hex of scene_id + sorted(npc_ids) + style_preset);
  LocalSdxlSidecarProvider stub for M4.
• POST /agent/turn SSE endpoint wrapping AgentOrchestrator.
• DM system prompts (EN+RU reference templates).
• Frontend ToolCallCard with settle animation (cycle digits at 100 ms
  while pending, snap to JSON + 600 ms gold flash on settle),
  ToolLogSlice, streamAgentTurn + useAgentTurn hook,
  ToolInspectorDrawer (480 px right slide-in, copy-as-cURL).
• Settings Model tab (system prompt + temperature 0..2 step 0.1 +
  Replicate API key).

Tests

• 37 cargo + 140 vitest.

[v0.3.0-m2] - 2026-04-30 - D&D 5e combat resolver + VTT overlay

Added

• 4-phase turn (initiative + action economy + conditions + damage +
  healing + saves+checks + attack rolls) via CombatResolver.
• ToolCallValidator dispatch table for the seven M2 combat tools.
• sqlx 0.8 + uuid 1 + chrono 0.4 SQLite persistence: campaigns,
  sessions, snapshots (with v2 forward-compat parent_save_id /
  branch_id columns), combat_encounters, combat_tokens.
• Three combat HTTP endpoints (/combat/start, /action,
  /combat/end) streaming six SSE-event variants.
• PixiJS CombatToken HTML overlay (HP-by-percent bar / AC chip /
  status ring / active-turn pulse), AoeTemplate cone/sphere/line/cube
  SVGs from school-of-magic palette, CombatOverlay container (280 ms
  cross-fade entry), InitiativeTracker slide-in, ActionBar.
• Chat refinements: TypingIndicator three Cinzel diamonds 1.4 s loop,
  useStickyScroll 100 px threshold, drop-cap on finalised assistant
  bubbles.
• Global prefers-reduced-motion override.

Tests

• 76 cargo + 116 vitest.

[v0.2.0-m1.5] - 2026-04-29 - Multi-provider settings + UI hardening

Added

• OpenAICompatProvider for BYO base_url + model + api_key (LM Studio,
  Ollama /v1/, llama.cpp server, vLLM, mistral.rs, OpenRouter, Groq,
  DeepSeek, Together, Fireworks).
• Hot-swap atomic via RwLock<Arc<dyn LlmProvider>> in AppState,
  driven by POST /settings.
• Self-hosted Inter / Cinzel / JetBrains Mono in public/fonts/.
• C1-C6 design tokens folded into src/styles/theme.css.

Changed

• Toolchain: npm + eslint + prettier + @vitejs/plugin-react-swc ->
  bun + biome + @vitejs/plugin-react v6.
• valibot (replaces zod) for runtime schemas.
• Zustand persist middleware with split-storage adapter
  (secrets.json for provider configs / API keys; settings.json for
  non-sensitive prefs).
• All inline styles migrated to CSS Modules.
• --color-accent-soft -> --color-accent-tint (translucent rename),
  freeing the canonical name for the hover-gold semantic.

[v0.1.0-m1] - 2026-04-28 - Tauri shell + chat skeleton

Added

• Tauri v2 shell (Rust workspace: app-domain, app-llm, app-server
  • src-tauri/).
• axum HTTP server as embedded backend, Tauri sidecar launch.
• LlmProvider trait + AnthropicProvider (genai 0.6.0-beta.18 with
  native Anthropic prompt caching).
• Streaming chat skeleton via SSE (/chat endpoint).
• React 19 + TypeScript + Vite + Zustand + react-i18next frontend
  scaffolding.