Changelog

[Unreleased]

Memory system — recall redesign (Complementary Learning Systems model).

Changed

• Capture is split by "wake/sleep". A per-session, every-N-turns subagent does encode only (recent exchange → episodic). A new built-in, visible dream mission owns consolidate + evict globally on a daily cron plus a turn-seam catch-up (covers a machine that was off overnight). The dream mission is per-run stoppable and deletable as a supported opt-out (degrades automatic curation, never loses data); each run leaves an audit record.
• Write-time quality bar — the encoder writes only what a future task benefits from (episodic is recall-visible immediately), instead of liberal capture.
• Write routing by salience — explicit "remember…"/standing instructions go to the durable store immediately via Memory_write; incidental signal goes through episodic→consolidate.
• Reconcile contract — memory is reconciled against existing memory on any reactivation (write, recall, dream). Near-duplicates are deduped mechanically; a genuine contradiction is resolved with the user (asked, with dates) by appending the corrected row, never silently overwritten or merged into storage; the no-user dream run defers contradictions to a later recall. The store stays CRUD-only — reconciliation is append + read-time + explicit user delete (no structural "replaces" link). Live-agent guidance ships in the memory nudge; the full contract is in doc/memory-spec.md §2.

Added

• agent.dream_catchup_hours (default 24) — catch-up threshold for the built-in dream consolidation mission.

v0.10.0 (2026-04-29)

Proxy rooms, user isolation, mission as first-class subsystem, memory system redesign, and LAN access fixes.

Added

Proxy rooms (Phase 6a + 6b) — share AI model access with others through invite-only or public rooms.

• Private/public rooms with shared models, allowed tools, and allowed skills configured in ~/.linggen/room_config.toml.
• Two consumer modes — browser-based (linggen.dev/app) and linggen-server (outbound WebRTC client connects via relay signaling).
• Proxy provider — ProviderClient::Proxy routes inference over the WebRTC inference data channel and streams StreamChunks back to the consumer.
• Inference data channel — separated from the control channel; list_models + inference handlers run there, filtered by shared_models.
• Settings → Sharing tab — room management, shared model checkboxes, allowed tools/skills, member list.
• Persistent token budget — ~/.linggen/token_usage.json store with room-level (token_budget_room_daily) and per-consumer (token_budget_consumer_daily) daily limits. Auto-resets at midnight UTC, flushes every 30s, survives reconnects.
• Token usage UI — Room tab reads usage and budgets from local store; bar updates in real time.
• Auto-refresh model list on proxy connect/disconnect via StateUpdated event.
• Room chat — bidirectional chat over the inference channel with sender_id echo prevention. User profile (user_id, user_name, avatar_url) backfilled from relay on startup; falls back to instance_name when user_name is missing.
• Disconnect on disable — toggling a room off broadcasts RoomDisabled, kicks consumer peers, syncs status to the linggen.dev DB. UI shows amber "Disabled" badge and hides Connect.
• Model selector labels show the room name — proxy:gpt-5.4 (My Room) instead of generic owner attribution.

User isolation

• UserContext replaces ConsumerContext — every peer carries user_id + permission level. room_name flows through signaling → UserContext → page_state.
• Unified page_state filtered per-user (sessions, models, skills, busy_sessions) — no more separate consumer page state.
• SessionMeta.user_id tracks ownership; ChatRequest accepts user_id, injected by peer.rs.
• ConsumerFilter drops events for other users' sessions on outbound delivery.
• user_id persisted in remote.toml (returned by linggen.dev registration).
• Hidden messages — [HIDDEN] content is filtered from all session-state APIs (workspace, skill, missions) before sanitization, and from ChatPanel rendering. HTML comments (<!-- ... -->) are stripped from rendered markdown. New send_hidden action in the skill bridge for system-level prompts.

Mission as first-class subsystem — missions are no longer a "mission skill"; they are a sibling subsystem with skill-shaped markdown.

• Frontmatter matches SKILL.md — nested permission { mode, paths, warning }, allowed-tools, allow-skills, requires, optional entry script.
• Three modes — agent (default; create session + run agent loop), app (open entry URL in browser), script (run entry as shell command). App/script modes skip the agent loop entirely.
• Entry script — runs before the agent loop, captures stdout/stderr to a per-run output dir, passes MISSION_* env vars. ~ in cwd is expanded before spawning, fixing the ENOENT that blocked Bash calls.
• Tool scope — allowed-tools + allow-skills drive mission_allowed_tools and consumer_allowed_skills. * means any skill; an empty allow-skills removes Skill from the allowlist; a concrete list gates the Skill tool even when allowed-tools is empty.
• Mission body injected into the system prompt via active_mission (mirrors active_skill), not as a 3 KB user message. The user turn is a short kickoff. get_system_prompt_api restores active_mission so copy-prompt returns the real prompt for mission sessions.
• In-memory mission cache — loaded once at startup, refreshed on create/update/delete and after skill install. Scheduler reads from cache (zero disk I/O per tick).
• MissionEditor rewritten — policy (4-way), permission mode + paths + warning, allowed-tools, allow-skills, requires, entry, description. MissionNav simplified; mission sessions live in the main session list under the Mission tab.
• API — POST/PUT /api/missions accepts new fields and legacy aliases (rewritten on next save). GET /api/missions/{id}/runs/{run_id}/output returns entry-stage stdout/stderr.
• Bundled "dream" mission rewritten in the new format with a real entry script.

Memory system redesign

• Skill install field — runs a script on install, wired into all install paths (marketplace, built-in, init). Replaces SkillMission / create_mission_for_skill — missions are now asset files copied by install scripts, not engine-managed.
• Memory frontmatter is fixed — templates are the source of truth; the model only edits the body.
• Edit added to the memory tool allowlist (alongside Read/Write).
• memory_descriptions_block_empty bootstrap template for first-run when no memory files exist. Old memory_block / global_memory_block templates removed.
• $SKILL_DIR now resolves correctly in skill content injection (prompt.rs + delegation.rs).

UI architecture

• App entry restructured — App.tsx / main.tsx replaced by apps/{Main,Consumer,Embed}App.tsx and entries/main.tsx, one bundle per surface.
• Event dispatcher split — per-kind handlers under eventHandlers/, table-driven dispatcher. Canonical EVENT_KINDS list; UiEvent.kind is now an exhaustive union.
• Store refactor — projectStore → sessionStore, agentStore → serverStore; new userStore and interactionStore.
• Session list reordered to User / Mission / Skill / All with per-tab counts; defaults to User. Trigger now sends a proper JSON body.
• Chat actions — Copy Chat / Copy System Prompt / Clear Chat moved out of HeaderBar into the expandable session bar inside ChatPanel, so every chat surface (main, skill, embed, consumer) has them.
• System prompt export — frontend sends session_id so skill-bound sessions export the active SKILL.md body; backend falls back to session cwd when project_root is empty; native tool schemas included.
• Skills sorted by usage — localStorage click count + last-used timestamp.
• HeaderBar shows the room name (clickable → Settings → Sharing). Leave button navigates to linggen.dev/app.
• Subagent return messages render as their own chat bubbles in live-stream sessions, matching the persisted messages.jsonl replay path.

LAN / WebRTC

• WebRTC binds to 0.0.0.0 and advertises the real LAN IP for ICE — fixes failures when the browser connects via LAN IP instead of localhost.
• --host passed to daemon — was ignored by the daemon child process before. Startup message now shows the actual host.
• Auth login uses the real host — was hardcoded to localhost, blocking login from LAN IPs.
• Skip browser open on headless Linux — when neither DISPLAY nor WAYLAND_DISPLAY is set (SSH sessions).
• ChatGPT OAuth tokens read from disk per-request — login/refresh takes effect immediately without restart. ModelManager rebuilt after OAuth completes.

Changed

• Tool restriction model — single source of truth via EngineConfig.effective_tool_restrictions() (cascading mission ∩ consumer intersections); EngineConfig.is_tool_allowed() is the unified check used by both prompt building and the execution gate. Two separate gates (mission + consumer) collapsed into one in tool_exec.rs. consumer_allowed_tools normalized to HashSet.
• ModelInfo.provided_by field added so the UI can attribute proxy models.
• Increased proxy answer poll timeout — 30s → 60s to tolerate slow relay delivery before the owner picks up the offer.
• Filter messages input dropped — the browser's Cmd+F is sufficient.

Removed

• AgentTree component (dead code).
• /api/projects endpoints and projects from page_state. MissionPage derives working folders from sessions; project-related store methods deleted.
• ProjectInfo type and related project store methods.
• The "mission" skill — missions depend on skills via requires:, not on a gated skill. Scheduler and API no longer check for it.

Fixed

• Permission mode UI not updating after approval — page_state handler referenced an undefined _permissionSuppressedUntil, so Date.now() >= undefined always evaluated false and blocked all updates. The UI showed stale "read" mode even after the user approved a switch to "edit".
• Auto-cleanup proxy connection on disconnect — consumer auto-removes the stale connection, Room tab shows "Connect" instead of stuck "Connected".
• Room chat panel hidden when the owner had room enabled but no user_name yet.
• list_models race — wait for the inference channel to open before sending; double-clickable Connect button now shows a spinner and is guarded.
• Skills installed flag stale — Library card kept showing "Install" for freshly-installed built-in skills until the 10-min cache TTL expired. Cache only the GitHub-derived metadata (dir_name/name/description); recompute the installed flag from the filesystem on every call.
• File watcher removed — crashed on permission-denied paths; wasn't used.

v0.9.3 (2026-04-08)

Server-pushed PageState, TUI removal, auto-scroll rework, and UX polish.

Added

• Server-pushed PageState — server aggregates projects, sessions, models, skills, agents, missions, and permissions into a single message pushed over the WebRTC control channel at 0.5 Hz with a dirty-flag mechanism. Replaces the HTTP polling storm that fired on every agent run.
• set_view_context message — frontend tells the server which session/project is active, scoping PageState pushes to relevant data only.
• busy_sessions in PageState — remote clients now see session busy status without needing per-session event channels.
• Dismiss button on queued messages banner — manually clear the queue when it gets stuck (e.g. QueueUpdated event missed).
• Markdown links open in new tab — target="_blank" on rendered links so clicking doesn't navigate away and text selection is easier.

Changed

• Auto-scroll rework — replaced distance-threshold detection with scroll-direction detection. Added distanceFromBottom > 150 guard so layout reflows during streaming no longer falsely detach auto-scroll. Consolidated duplicate scroll tracking from ChatPanel into the single useAutoScroll hook.
• Removed HTTP polling — initial load fetches for projects, sessions, models, skills, agents, and config are all replaced by PageState delivery on WebRTC connect. Only Ollama status and session tokens remain as HTTP fetches.
• Removed 5 dead API endpoints — agent-children, agent-context, missions/:id GET, missions/:id/sessions, builtin-skills/install-all.
• SessionModeSelector simplified — reads mode and zone from store (pushed by PageState) instead of fetching /api/sessions/permission on every render.
• Extracted non-reactive agent tracking — agentTracker.ts singleton replaces 15+ direct Zustand store mutations in eventDispatcher.
• Memoized skill suggestions in ChatInput (was rebuilt twice per render).
• Tokens/sec display wired up — recordTokenSample + recomputeTokenRate were never called; now functional.

Removed

• TUI — terminal UI (ratatui) and --tui flag removed. Linggen is now Web UI only. ling starts the daemon and opens the browser; ling --web runs the server in foreground.
• SSE transport — server-sent events transport removed. All real-time communication uses WebRTC data channels.

Fixed

• Auto-scroll fighting — removed duplicate scroll tracker in ChatPanel that competed with useAutoScroll hook.
• Session mode selector race — after user switches mode (e.g. admin → read), a 3-second suppress window prevents the next PageState push from overwriting the optimistic UI update.
• Page flash during streaming — floatingUserMsg effect was re-subscribing on every token.
• Duplicate React key in SubagentTreeView.
• Subagent state leak — agentTracker.reset() called on session switch.
• Plan message overwrite — mutateLast guard prevents fast-path from overwriting plan messages with streaming tokens.
• Skill session chat in remote mode — skill app chat iframe now routes through the relay connect page when accessed via linggen.dev (was loading the landing page instead of the compact chat).
• Skill session restore — reopening an existing skill session with no localStorage cache auto-triggers a fresh scan instead of showing an empty dashboard.

v0.9.1 (2026-03-31)

Simplified run system, daemon mode, ChatGPT default model, and bug fixes.

Added

• Background daemon mode — bare ling now spawns a background daemon and opens the Web UI in the browser. Terminal returns immediately. Use ling --tui for classic TUI mode.
• ChatGPT OAuth default — new installs default to GPT-5.4 via ChatGPT subscription. No API key or local model download needed.
• Unified working folder — all tools (Read, Write, Edit, Glob, Grep) resolve relative paths from the agent's cwd, not just Bash. When the agent cds into a git repo, the workspace root, CLAUDE.md, and permissions update automatically.
• User ! cd tracking — ! cd /path in the Web UI now persists cwd per session, same as agent Bash commands.
• UI follows cwd changes — selectedProjectRoot updates when the agent changes working folder.

Changed

• In-memory run store — agent run records are no longer persisted to {run_id}.json files on disk. Runs are tracked in memory only (for cancellation and status during execution).
• Removed run history UI — run picker dropdowns, context display, timeline, pin/unpin removed from ChatPanel and SubagentDrawer.
• Removed dead code — AgentsCard.tsx (never imported), timeline.ts, run context types (AgentRunSummary, AgentRunContextResponse, etc.).
• Simplified cancel response — POST /api/agent-cancel returns { status: "ok" } instead of { cancelled_run_ids: [...] }.
• Font size +1px — all UI font sizes bumped by 1px for mobile readability.
• Logo — shortened to "Linggen", links to linggen.dev.
• install.sh — removed --with-memory flag and ling-mem install block.

Fixed

• Plan reject buttons not disappearing — PlanUpdate events now carry session_id, so they're delivered via WebRTC data channels (was None, events were lost).
• Queued messages stuck after cancel — cancel_agent_run() now drains the queue for cancelled agents.
• Queued messages showing in chat — queued messages are no longer persisted to messages.jsonl at queue time. They're persisted when dequeued, preventing the sync-back from re-adding them.
• Session header not showing for user sessions — fetchSessions was resetting activeSessionId when the session wasn't in the project-filtered list. Now checks allSessions before resetting.
• Working folder in non-git dirs — check_working_folder_change() now uses the cwd as workspace root when no git repo is found (was falling back to ~).
• macOS /tmp symlink — cwd is canonicalized before use as workspace root (resolves /tmp → /private/tmp).

v0.9.0 (2026-03-30)

Working folder model, per-session engines, WebRTC-first transport, and UX improvements.

Added

• Working folder model — sessions start in HOME mode and auto-detect projects when the agent cds into a git repo. CLAUDE.md, permissions, and git context load dynamically on project entry. Configurable home_path in settings.
• Per-session agent engines — each session gets its own engine instance. No more lock contention between sessions — game-table and regular chat run truly in parallel.
• WebRTC-first transport — Web UI always uses WebRTC (local and remote). Per-session data channels provide natural isolation. SSE retained for TUI only.
• WebRTC session_id enrichment — events are tagged with session_id before routing to data channels, preventing cross-session event leaks.
• ChatGPT token expiry UX — inline re-login button when ChatGPT OAuth expires. After re-login, session engines are cleared so the fresh token is used immediately.
• Working folder changed event — WorkingFolderChanged server event emitted when the agent cds. UI header updates reactively.
• home_path config — configurable default working folder for new sessions (defaults to ~).
• Git root detection — find_git_root() walks up from cwd looking for .git/. Skips home directory dotfiles repos.

Changed

• Flat session storage — all sessions stored in ~/.linggen/sessions/ (flat directory). No more per-project/mission/skill session directories. Session metadata tracks cwd, project, project_name, mission_id.
• Simplified chat creation — clicking + immediately creates a session. Removed project picker dialog.
• Removed project management UI — no more workspace section, project cards, or manual project add/remove in sidebar. Projects are auto-discovered from git repos.
• Skill search ordering — community skills from skills.sh and ClawHub are interleaved by relevance instead of sorted by install count.
• ClawHub ZIP install — handles root-level SKILL.md (no subdirectory) in ClawHub ZIP archives.
• Ollama status polling — only polls when Ollama models are configured, eliminating 404 spam.
• Auto-scroll — any upward scroll stops auto-scroll (was 10% threshold). Resumes within 20px of bottom.
• IME composition — Enter key during Chinese/Japanese input composition no longer triggers send.
• Models card scroll — auto-scrolls to default (starred) model when the model list loads.
• Session list — session rows use <div> instead of nested <button> (fixes React DOM nesting warning).
• Skill reload — installing/uninstalling skills clears session engines so new skills are available on next message.
• install.sh — post-install output now shows ling init as the first step.

Fixed

• Session isolation — WebRTC events no longer leak between sessions. Added session_id enrichment in WebRTC peer handler (was missing, only SSE had it).
• emit_outcome_event — plan/outcome events now carry session_id (was hardcoded None).
• Compact mode race — skill app iframe now explicitly fetches workspace state after setting isSkillSession, preventing stale API calls.
• Session engine memory leak — remove_session_engine called on all session deletion paths.
• TUI session creation — get_session_meta check uses Ok(Some(_)) instead of is_ok() (was always true).
• UiEvent.kind type — added 'working_folder' to TypeScript union type.

Removed

• ~/.linggen/projects/ session directories — sessions no longer stored per-project.
• session_root on EngineConfig — removed; all persistence goes through global sessions.
• ProjectContext.sessions — removed; all session access through AgentManager.global_sessions.
• ProjectStore::session_store() — removed dead code.
• missions_sessions_dir() / skill_sessions_dir() — removed from paths.rs.
• NewChatDialog component — removed project picker dialog from UI.

v0.8.0 (2026-03-25)

Remote access, mobile UI, Google login, and infrastructure improvements.

Added

• Remote access — access your linggen from any device. Run ling login to link to your linggen.dev account, then connect from any browser at linggen.dev/app. Peer-to-peer connection — no VPN or port forwarding needed.
• ling login / ling logout / ling status — CLI commands for managing remote access. Fully automatic browser-based OAuth flow with token exchange; no manual steps needed.
• ling auth login — ChatGPT subscription auth. Auto-detects headless/SSH environments and falls back to device code flow (removed --device flag).
• Google login — sign in to linggen.dev with Google or GitHub. Email-based account matching across providers.
• Signaling relay — lightweight relay on linggen.dev handles connection setup. Nonce-based offer/answer exchange via stateless HTTP.
• Mobile UI — responsive layout auto-detected on narrow viewports (or via ?mode=mobile). Full-bleed chat, larger touch targets, iOS safe area support. Right-side drawer for models and skills.
• Gzip chunked transfer — large responses (skill files, API data) are gzip-compressed and sent as base64 chunks over data channels. Handles SCTP backpressure correctly.
• Skills open in-app — web launcher skills now open in an in-page iframe panel instead of a new browser tab. Works in both local and remote mode.
• Session project names for missions — mission sessions now show their project name in the session header, matching the behavior of user sessions.

Changed

• ling login non-interactive — uses hostname automatically, no instance name prompt.
• Heartbeat interval — increased from 30s to 5 minutes to reduce relay load. Online threshold set to 10 minutes.
• Online status via D1 — instance online/offline status is now determined by updated_at timestamp in D1 database instead of KV TTL keys. Eliminates KV write quota consumption from heartbeats.
• JWT sessions — linggen.dev authentication switched from KV-stored sessions to signed JWT cookies (HMAC-SHA256). Eliminates KV reads on every authenticated request.
• Settings page mobile layout — scrollable tab strip, responsive model card grid, reduced padding on small screens.
• Header compact mode — shorter title ("Linggen" on mobile), status dot without text label, sparkles button for info drawer.
• Session delete on mobile — trash button always visible on touch devices (was hover-only).
• InfoPanel component — extracted models + skills cards into shared component used by desktop sidebar and mobile drawer.

Fixed

• SSRF bypass — URL-decode path before validation in WebRTC HTTP proxy (blocks %2e%2e traversal).
• JWT algorithm validation — verify alg: HS256 in token header before signature check.
• Free-tier instance limit — use COUNT(*) query instead of single-row check (prevents bypass via new instance IDs).
• Token panic — guard api_token length before slicing in ling status (no crash on corrupted config).
• Double reconnect — guard handleDisconnect against firing multiple times from concurrent ICE/connection state changes.
• Double connect — guard doConnect against concurrent calls (prevents RTCPeerConnection leak).
• Session channel leak — unsubscribeSession now called on session change in useTransport hook.
• Token lost on write error — browser response write in ling login callback no longer discards the received token if the browser closes early.
• Relay poll blocking — handle_remote_offer spawned in separate task so the offer poll loop stays responsive.
• Nonce URL encoding — relay signaling nonce is now URL-encoded in poll requests.
• Relay offer missing Content-Type — added Content-Type: application/sdp to relay offer POST.
• Logout CORS headers — logout response now includes CORS headers for cross-origin requests.
• Dead code — removed unused split_utf8_safe function.
• Stale comments — updated heartbeat interval comments from "30s" to "5 minutes".

v0.7.0 (2026-03-11)

Major release with native tool calling, mission system, TUI, permissions, and extensive UI improvements.

Added

• Native tool calling — models use structured function calling (OpenAI, Ollama) instead of JSON-in-text. Default for all providers; falls back gracefully for legacy models.
• TUI interface — full terminal UI via ratatui. Default mode runs TUI + embedded server; --web for web-only.
• Mission system — agents self-initiate work on cron schedules when a mission is active. Idle scheduler prompts agents between user messages.
• Plan mode — agents can enter plan mode (EnterPlanMode) for research and structured planning before making changes. Plans require user approval via ExitPlanMode.
• File-scoped permissions — AcceptEdits mode, deny rules, and per-project permission persistence.
• Credential storage — secure API key management via /api/credentials endpoint.
• Model auto-fallback — health tracking with automatic fallback to next model in the routing chain on errors or rate limits.
• AskUser bridge — agents can ask structured questions mid-run with options and multi-select.
• Web search & fetch — WebSearch (DuckDuckGo) and WebFetch tools for agents.
• Skills marketplace — search, install, and manage community skills from the web UI or CLI (ling skills add/remove/search).
• ling init command — scaffolds ~/.linggen/ directory tree, installs default agents, creates config, downloads skills.
• ling auth command — ChatGPT OAuth authentication (browser and device code flows).
• Session-scoped SSE — events are tagged with session ID; clients filter to their own session.
• Per-session working directory — cd in one session doesn't affect others.
• SSE reconnect handling — automatic state resync on reconnect with UI indicator.
• Context window management — adaptive compaction with importance-based message pruning.
• Prompt caching — stable system prompt prefix cached across iterations.

Changed

• Config file renamed from linggen.toml to linggen.runtime.toml.
• Prompt system refactored from hardcoded strings to TOML templates.
• Tool calls render as individual inline widgets (aligned with Claude Code style).
• ChatPanel.tsx refactored into focused modules under chat/ folder.
• tools.rs and app.rs split into module directories for maintainability.
• Default supports_tools changed to true even for unrecognized model IDs (prevents fallback to text-based JSON mode).

Fixed

• SSE session isolation — events no longer leak across sessions.
• Streamed text-only responses no longer disappear after generation.
• Ollama 500 error — use role "tool" for tool result messages in native mode.
• Agent context loss on long conversations.
• Glob pattern matching edge cases.
• Queued message display order (now chronological).
• Think tag stripping for models that emit <think> blocks.

v0.1.1 (2025-12-15)

Initial patch release.

v0.1.0 (2025-12-14)

Initial release — multi-agent engine, web UI, skills system, Ollama and OpenAI providers.