Changelog

All notable changes to Liminal AI on the main branch.

Current stage: beta (v0.1.2 tip of main, 2026-06-18). The workspace package may read 0.1.0, but beta, RC, and a v0.1.0 public preview have not been declared as product releases yet.

Format: v0.0.x entries keyed to the last GitHub push in each slice. Dates are real push dates.

Surface	Where
Source of truth	changelog/releases.json — run npm run changelog:gen
Technical changelog (this page)	Generated from JSON — full Shipped bullets
Marketing changelog	vireondynamics.com/liminal/changelog
Short index	CHANGELOG.md

After editing JSON: npm run changelog:gen, commit, push, then in vireondynamics-website: npm run docs-portal:sync and npm run docs-portal:deploy.

---

v0.1.2 — 2026-06-18

Current beta tip. Multi-account mail, inbox watcher, file attachments, integrations hub polish, managed inference fixes.

Shipped

• Multi-account mail — mail_search_inboxes scans every connected mailbox; account_hint on Gmail/Outlook compose and send; session mail_account_context persists the chosen mailbox across turns (rule R-EMAIL-ACCOUNT)
• Connected mail only — mail_oauth_filter ignores unconnected integrations and Entra #EXT# guest admin mailboxes; list_connectors shows Connected mailboxes
• Email privacy — email_redaction masks SSN, cards, phones, and addresses in summaries; mail_search_inboxes defaults redact_sensitive: true (rule R-EMAIL-PRIVACY)
• Inbox watcher — Background Gmail/Outlook polling with label support; automated receipt workflow hooks; capability gates for mail tools
• File attachments — Drag, drop, or paste any file into web and desktop composer; chat_attachments persists to .agent_artifacts/uploads/ with limits (12 total: 4 images, 8 files)
• Integrations hub — Web IntegrationCard + per-account list; desktop provider rows, connection status, revoke account; integrations_snapshot and revoke_integration_account tools
• Managed inference fix — Settings managed mode no longer flips back to BYOK on Pro; sidecar skips BYOK auto-patch when user chose managed; regional Bedrock profile ids in desktop model picker
• Email compose — Expanded recipient guard, compose context/resume, Gmail compose guard tests; capability gates and long-horizon eval updates
• Docs — Google Workspace, Microsoft 365, Connectors

---

v0.1.1 — 2026-06-15

Desktop auto-update — harness hot-swap and app restart from GitHub Releases.

Shipped

• Desktop auto-update — Settings → About & updates; hub banner when a newer v*-desktop release exists
• Harness hot-swap — downloads liminald-runtime-v{version}.zip, stops sidecar, swaps liminald/, reconnects
• App update — downloads full platform archive, verifies SHA256, relaunches via bundled updater scripts
• Release manifest — liminal-desktop-manifest-v{version}.json on each desktop GitHub Release
• CLI — liminal update --check, --json, --harness-only for portable installs; git path unchanged
• Docs — Install (desktop auto-update)

---

v0.1.0 — 2026-06-15

First public preview release — stability milestone with CI, eval framework, latency optimization.

Shipped

• CI workflow — .github/workflows/ci.yml runs build, typecheck, and test on every push to main
• Eval sandbox framework — packages/eval/fixtures/ sandbox scenarios; npm run eval:sandbox for reproducible agent tests
• Inference latency mode — AGENT_LATENCY_MODE patches for performance-optimized inference paths
• Tool circuit — tool_circuit.ts for parallel tool execution optimization
• Kimchi constants — Extracted AGENT_KIMCHI_* constants for Cast AI provider tuning
• Desktop managed inference catalog — Flutter UI for managed inference model selection
• Sidecar multi-chat — chat_registry.ts for multi-session support
• Inference path validation — scripts/validate-inference-path.ts for diagnosing routing issues
• Release checklist — docs/RELEASE-CHECKLIST.md documents the v0.1.0 release process
• Docs — Running evals, Sandbox capability lab, Inference path validation

---

v0.0.26 — 2026-06-10

Live compose dock, email HTML preview, Azure ARM, Kimchi provider, OSC 133 shell.

Shipped

• Compose dock — compose_preview + streaming tool_delta open the right pane immediately on write_file/edit_file/Gmail/Outlook compose; content streams live and the panel persists through turn_end
• Email HTML preview — Extracts <body> fragments and sanitizes for HtmlWidget; fixes raw-HTML monospace preview during streaming
• Azure ARM — Subscription, resource-group, and location tools; per-resource api-version inference; clearer OAuth/ARM errors; guide in docs/guides/azure.md
• Kimchi provider — Cast AI rate-limit parsing, transient 400/429 retries, min request interval, throttle-aware circuit breaker; AGENT_KIMCHI_* in Settings
• OSC 133 shell — Prompt markers at PTY launch (shell_integration_launch.ts); agent_shell_session + Osc133Tracker for deterministic command-done events without prompt regexes
• Provider backends — Kimchi backend routing in provider_backends.ts; web/desktop Settings presets updated
• Desktop WebView — desktop_webview_lifecycle.dart pauses off-screen WebView2 instances to cut GPU/memory churn
• Docs — Azure, Google Workspace

---

v0.0.25 — 2026-06-09

Desktop design system, Sheets auto-fit, integration arg reliability, research reply fix.

Shipped

• Desktop design system — Shared tokens, primitives (Button, Card, Badge), patterns (Sheet, EmptyState, SearchField), persona shells; wired through hub, home, chat panes, composer, and settings
• Vireon hub — Status strip, quick actions, mission control, searchable chat list, integration chips, two-column layout, pull-to-refresh
• Chat UI — Distinct user/assistant/tool/reasoning lanes (chat_message_frame, reasoning_block, tool_category); polished composer, drawer, message and tool activity tiles
• Tool arg pipeline — prepareToolArgsForValidation normalizes Google MCP, Google REST, and Linear args before schema validation; fixes Sheets requests/data JSON-string coercion and MCP HTTP 400 from hallucinated fields
• Sheets auto-fit — AGENT_SHEETS_AUTO_FIT (default on) resizes columns after value writes; empty spacer columns pinned narrow; sheets_rest_auto_fit for manual refit
• User reply finalize — maybeRunUserReplyFinalizePass runs one tool-free completion when a turn would end with tools/vault only; streams assistant text without fake user nudges (AGENT_USER_REPLY_FINALIZE)
• Linear REST — Expanded issue/project tools, linear_resolve helpers, and forgiving arg aliases
• Docs — Google Workspace + Linear guides updated for REST arg shapes and Sheets layout protocol
• Docs — Google Workspace, Linear, Desktop design system

---

v0.0.24 — 2026-06-09

Hosted OAuth connectors and desktop integrations hub.

Shipped

• Hosted OAuth bridge — /connect/<provider> begin + callback routes broker OAuth; desktop opens the flow and receives tokens on localhost callback
• Seven connectors — Google Workspace, Microsoft 365, GitHub, Slack, Linear, Notion, and Xero through the hosted bridge
• Hosted token refresh — Long-lived connections refresh access tokens server-side; secrets stay on Vireon
• Custom GitHub website login — vireondynamics.com sign-in via shared OAuth callback
• Desktop integrations hub — Connect, disconnect, and inspect providers from Liminal Desktop
• Lazy connector tools — Connector families stay off until a provider is linked
• Docs — Connectors

---

v0.0.23 — 2026-06-07

Microsoft 365, styled email dock, lazy tools, Google REST hybrid.

Shipped

• Microsoft 365 — Entra OAuth broker, connect_provider({ provider: "microsoft_365" }), MCP sidecar, Graph REST for Outlook send/draft, Teams, Planner, To Do, OneNote, OneDrive, Excel, SharePoint, and unified search; guide in docs/guides/microsoft-365.md
• Google Workspace hybrid — REST tools for Docs (docs_rest_*), Sheets, Slides, Calendar alongside existing Gmail MCP; styled HTML drafts via gmail_create_draft / gmail_send_message REST
• GitHub connector — connect_provider({ provider: "github" }) with token/OAuth and MCP tool attach
• Lazy tool loading — AGENT_TOOL_LAZY=1 default; connectors family activates on demand; persisted MCP connections no longer auto-register 300+ tools at boot (AGENT_INTEGRATION_AUTO_ACTIVATE=1 opt-in)
• OAuth mail routing — Gmail preferred over Entra #EXT# guest admin mailboxes; list_connectors shows primary mail provider
• Email composition — email_style_infer sidecar, R-EMAIL-COPY / R-PRODUCT-TRUTH protocol rules, MIME mojibake repair + em-dash humanization, placeholder URL guard on send, Liminal product facts baked into system prompt
• Desktop email dock — Live HTML/plain preview while gmail_* / outlook_* tools stream; preview sanitization matches send path
• Desktop shell — Integrations screen, Vireon hub, orchestrator panel, chat pane with file + email compose rails
• Sidecar orchestrator — Worker handoff protocol, integrations snapshot API, orchestration status on wire
• Web integrations — Settings Integrations accordion for Google, Microsoft 365, and GitHub connect/disconnect
• Docs — Microsoft 365, Google Workspace, Tool families

---

v0.0.22 — 2026-06-07

Serper search, agentcard, desktop browser dock; widget apps gated off.

Shipped

• Serper web search — AGENT_WEB_SEARCH_PROVIDER duckduckgo|serper with fallback; optional Serper fetch for web_fetch; dedicated weather_fetch path
• Agentcard tools — agentcard harness tools + CLI shim (AGENT_AGENTCARD); operator guide in docs/guides/agentcard.md
• Gmail compose guard — Preflight gmail_send for missing recipients/subject/body; expanded Google OAuth scope catalog
• Desktop browser dock — AGENT_BROWSER_EMBED streams browser_view snapshots to the Flutter shell; browser dock UI in chat
• Default persona skip — Bundled persona/default Liminal preset installs on bootstrap skip without LLM generation
• Liminal widget apps (gated off) — spawn_app/update_app stack in core/sidecar/tools/desktop ships disabled (AGENT_LIMINAL_APPS=0, desktopAppsEnabled=false) until widget stability improves
• Desktop voice hardening — Dictation TTS suspend/resume depth counter; incomplete portable liminald bundle detection in launcher
• Docs — Liminal apps, Agentcard, Web research

---

v0.0.21 — 2026-06-06

Plan progress UI, chat replay, desktop voice, persona UI v2.

Shipped

• Plan progress UI — One updatable Progress card per turn (web PlanProgressBlock, TUI PlanCard, desktop PlanProgressTile); step_index checkoffs no longer spawn empty Plan shells
• Chat session replay — chat_session_replay + active_chat_state hydrate transcripts from .agent_sessions JSONL on web, TUI, and sidecar multi-chat registry
• Desktop voice — Sidecar /audio HTTP for TTS/STT; Flutter dictation (VAD), composer mic, and speech_output playback; docs/guides/voice.md
• Persona UI v2 — persona_ui_copy, persona_ui_layout, persona_ui_invariants in core; desktop persona chrome tokens and bootstrap polish
• Provider hygiene — provider_token_scrub strips leaked API keys from model output; OpenRouter routing and managed BYOK fallback refinements
• Docs — Voice, Rich message rendering

---

v0.0.20 — 2026-06-06

Vault knowledge graph, desktop provider presets, managed-inference BYOK fallback.

Shipped

• Vault knowledge graph — vault_ingest_entities decomposes research into per-entity dossiers (Identity / Current / History / Relationships); vault_ingest auto-wikilinks; vault_recall GraphRAG neighborhood retrieval; vault_lint orphan repair; AGENT_VAULT_ENTITY_EXTRACT on by default for research auto-write
• R-VAULT-ENTITIES — Protocol cookbook for one canonical name per dossier, parallel vault_write batching, event+cast workflows; kind tags (person, org, …) on ingested entities
• Managed inference fallback — On HTTP 402 (credits exhausted), switch to OpenRouter BYOK on free openrouter/owl-alpha (AGENT_MANAGED_FREE_FALLBACK); preempt when wallet is already empty; honor explicit OpenRouter baseURL for BYOK presets
• Desktop provider — Settings provider-preset dropdown (Owl Alpha Stealth single-model, DeepSeek, etc.); save_provider allows model/baseURL-only saves; API key hydration from bundled liminald/repo/.env (provider_api_key)
• Web Settings — Provider presets sourced from @liminal/core; inference usage banner hints when managed credits are exhausted
• Docs — Memory and vault, Managed inference

---

v0.0.19 — 2026-06-05

Desktop for macOS and Linux, GitHub Release downloads, install-page links.

Shipped

• Liminal Desktop (all platforms) — Release builds for Windows x64, macOS arm64, and Linux x64; npm run desktop:build|package:macos|linux plus existing Windows scripts; portable liminald/repo with relative paths in bundle.json
• GitHub Releases — Tag v*-desktop runs unified CI (Windows + macOS + Linux), uploads zip/tar.gz artifacts and SHA256 checksums
• Install page — vireondynamics.com/liminal/get-started#desktop-app links direct downloads; product hero CTA points at the desktop app
• CI fixes — tsc builds without proprietary @liminal/enterprise on public clones; desktop bundle treats robocopy exit codes 0–7 as success
• Docs — Rich message rendering, Install (docs)

---

v0.0.18 — 2026-06-04

Liminal Desktop (Windows), provider-routing recovery, reasoning-stall breaker, Gmail API tools.

Shipped

• Provider routing recovery — Dynamic provider.ignore list is capped and cleared when OpenRouter returns “all providers ignored” (404), so long ReAct runs do not stall mid-task after repeated upstream 429s
• Reasoning-stall breaker — Consecutive reasoning-only rounds (think/reason/plan/…) trigger a hard nudge, then temporarily drop reasoning tools so the model must call an action tool or finish
• Liminal Desktop (Windows alpha) — Flutter app (apps/liminal_desktop) talks to bundled liminald over @liminal/protocol WebSocket: multi-chat, Vireon sign-in, harness settings, compact tool cards, image attachments, GFM markdown + HTML embeds
• Sidecar — @liminal/sidecar serves session bridge, /media for workspace images, and send_message attachments; npm run desktop:build:windows bundles Release + liminald
• Gmail — Native google_gmail_* tools (read/search) with proper body decoding; Google MCP probe documents separate Gmail API enablement
• Rich message rendering — Spec in docs/concepts/rich-message-rendering.md; desktop parity with web for alerts, code highlight, video embeds, streaming HTML fences
• Docs — Rich message rendering, Managed inference, Google Workspace

---

v0.0.17 — 2026-06-02

Team shared memory, Pro cloud sync APIs, NotesFacade in CE, and org-scoped control-plane storage.

Shipped

• Team shared memory — team.shared_memory entitlement drives org-scoped note replication keyed by org_id + workspace_fingerprint. Workspace and global notes sync across teammates; chat-scoped notes stay local-only (API + ranker enforced). Enterprise Edition pulls before round 0 and pushes after turn_end when AGENT_TEAM_MEMORY_SYNC=1
• Control plane — GET/PUT /api/team/memory/notes, status, optional SSE bus (/api/team/bus/*), Pro user notes/vault sync (/api/pro/cloud_sync/*), session history upload, org tables + Team checkout creates org owner, invite flow with seat limits
• CE contracts — NotesFacade, StoredNote org fields (userId, orgId, revision), runWithOrgContext, harness hooks onTurnStartMemorySync / onTurnEndMemorySync / onRecallMerge, AGENT_TEAM_EMBED_ON_RECALL to avoid full-store embedding on recall
• Web Settings — Team memory status + org id on /api/vireon/account and in the Settings modal
• Dev — AGENT_CONTROL_PLANE_URL for local EE sync against :3002; changelog now generated from changelog/releases.json (npm run changelog:gen)
• Docs — Team memory, Pro & Enterprise, Enterprise Edition

---

v0.0.16 — 2026-06-02

Managed inference for Pro, Vireon sign-in, dynamic OpenRouter routing, EE auto-install, security hardening.

Shipped

• Managed inference (Pro) — Chat, embeddings, vision, and audio sidecars can route through Vireon's metered OpenAI-compatible proxy (AGENT_INFERENCE_MODE: auto | managed | byok; default auto). Pro sign-in prefers managed routing with included monthly credits and a clear top-up path when budget is exhausted. Community Edition stays bring-your-own-key with no account
• Vireon sign-in — liminal login (loopback browser flow) or Settings → Sign in stores the license under ~/.liminal/; no .env API-key paste for Pro managed mode
• Dynamic OpenRouter routing — Price-sorted provider selection (AGENT_PROVIDER_STRATEGY), sticky session_id for KV-cache affinity (AGENT_OPENROUTER_SESSIONS), and session-epoch rotation on upstream 429s (adaptive strategy) so rate limits do not stall long runs
• Enterprise Edition auto-install — Pro+ login downloads the proprietary EE bundle to ~/.liminal/enterprise/ (checksum-verified); implementations stay out of the public CE repo while gating cloud/team entitlements
• Security hardening — Control-plane rate limiting and auth hygiene, web API token gate bound to loopback, tighter guards on run_shell and execute_code
• Inference UX — Web inference usage/credits banner; steadier sign-in polling for managed sessions
• Dictation & ASR defaults — Default AGENT_TRANSCRIBE_MODEL is nvidia/parakeet-tdt-0.6b-v3 (English + EU); AGENT_DICTATION_WEB_SPEECH=0 keeps browser Web Speech off so dictation transcribes server-side only (no Chrome→Google path). Web /api/config exposes the toggle
• Docs — Managed inference, Accounts & licensing, Pro & Enterprise, Voice I/O

---

v0.0.15 — 2026-05-31

Open-core licensing, Vireon control plane, Enterprise Edition package, dynamic workflows.

Shipped

• Open-core entitlements — Offline-verifiable license tokens (Ed25519; harness holds public key only). Verified licenses cache to ~/.liminal/license.json with bounded offline grace. Community Edition needs no license — full harness on your machine
• License tiers — community · pro · team · enterprise (each inherits below). Tiers gate cloud/sync/governance features by entitlement key, not core agent capability
• Vireon control plane — @liminal/control-plane: Stripe subscriptions + webhooks, Supabase entitlements, signed license issuance (deploy beside vireondynamics-website)
• Enterprise Edition package — Proprietary Pro/Team/Enterprise implementations in a separate package; CE repo stays fully functional without EE installed
• Dynamic workflows — plan_workflow → run_workflow execute a declarative multi-phase spec: bounded concurrency waves per phase, per-agent results in an out-of-context store (.agent_workflows/), distilled phase summaries back to parent context only. Root-only, approval-gated (AGENT_WORKFLOWS). Intent classifier can auto-activate the workflow tool family under lazy loading
• Docs — Dynamic workflows, Pro & Enterprise, Enterprise Edition

---

v0.0.14 — 2026-05-30

LLM memory curator, output-effort dial, voice I/O, retrieval reranker, deeper prompt caching.

Shipped

• Memory curator — curate_memory (dry-run by default) sends a metadata-annotated slice of the note store to the model, which returns prune / merge / re-confidence ops; a deterministic safety-rail veto then protects user:/identity:/pref: keys, high-access, and too-young notes. Deletion is reversible: forget and the curator soft-delete the full note to notes.archive.json (AGENT_MEMORY_ARCHIVE) before removing; restore_memory recovers it. Tunables: AGENT_CURATOR_TIMEOUT_MS, AGENT_CURATOR_MAX_TOKENS, AGENT_CURATOR_PROTECT_GLOBAL (off — global is the default scope, not a durability signal)
• Output effort — AGENT_EFFORT (low | medium | high | xhigh) — a deliverable-thoroughness dial injected as a system-prompt directive (rule R-EFFORT), exposed as a Settings dropdown. A separate axis from reasoning: it scales completion budget and output completeness/coverage, not internal think()/reason() depth
• Voice I/O — Text-to-speech (speak, Kokoro / OpenAI voices, per-turn budget + near-duplicate suppression) and live microphone dictation in the web UI; voice-mode tool gating. See Voice I/O
• Retrieval — Second-stage recall reranker over first-stage BM25 + embedding hits; fast-model JSON response cache to cut repeat sidecar calls
• Prompt cache — Rolling cache_control breakpoint over conversation history (AGENT_PROMPT_CACHE_ROLLING) plus volatile-tail context ordering (AGENT_CTX_VOLATILE_TAIL), so the cache extends across accumulated tool-result history, not just the static prefix
• File tools — find_files (path/name glob) and delete_file (approval-gated); a catalog reachability test guards against tools unreachable under lazy loading
• Chat UX — Markdown renderer for assistant output, session screenshot export, persona shell refactor (Hud / Minimal / Studio / Terminal / Composer)
• Web — Greeting fires on reload without blocking, SSE reconnect robustness; OpenRouter session_id on chat completions for request grouping
• Reliability — More reliable long write_file paths and HTML chunk guards

---

v0.0.13 — 2026-05-28

Shipped

• Provider model presets — One-click OpenRouter packs in web Settings (main + fast slots, provider order pins); DeepSeek, Claude 4.7, GPT 5.5, Gemini 3.5, Qwen 3.6, MiMo, Llama 4, Kimi, GLM, and cross-vendor mixes; slugs verified against OpenRouter GET /models
• Settings completeness — Full harness env catalog in the web Settings modal with field metadata (aligned with HARNESS_ENV_DEFAULTS)
• Turn end — Simpler end-of-turn pipeline on send() (less redundant work)
• Repo — Removed experimental Android app scaffold and mobile-contract package

---

v0.0.12 — 2026-05-26

Research ledger, identity memory, FSL relicense, web chat chrome polish.

Shipped

• Research ledger — Per-send() tracking of web_search / web_fetch (queries, surfaced URLs, fetch outcomes, deduped URLs); compact context injection each round; harness tool research_state (summary / pending / fetched / failures views)
• Identity memory — Intent identityQuery / identityProvision; LLM-generated recall queries; on-disk scan of user: / identity: / pref: notes; auto-remember with scope: global for user:name; protocol reminder not to use OS username as display name
• Recall fix — recall_relevant accepts query or queries
• Web UI — Chat switcher embedded in each persona shell header row; portaled dropdown + click-outside dismiss
• PASTE (optional) — Predictive tool-call speculation (AGENT_PASTE_*, off by default)
• License — FSL-1.1-MIT (MIT Future) across the harness; compliance documentation expanded; fair-source framing on marketing/install
• Reliability — write_file integrity check fix for apostrophes inside comments

---

v0.0.11 — 2026-05-24

User-global storage, web multi-chat, audio transcription, and connector tools.

Shipped

• Storage layout — ~/.liminal/ user-global memory, persona, prefs, vault; per-chat folders for sessions/artifacts; AGENT_STORAGE_LAYOUT=legacy
• Web multi-chat — ChatManager, chat list/switch APIs, idle bridge eviction, SSE scoped to the active chat
• Audio — transcribe_audio, upload attachments, OpenRouter Whisper defaults, dictation env knobs, web mic UI
• API connectors — api_connect / api_disconnect / api_list (OpenAPI → dynamic tools, persisted connections)
• MCP — mcp_attach to wire MCP servers into the live registry
• Memory — memory_promote, memory_neighbors, consolidate_chat; federated rank scoring; exploratory-turn debias (AGENT_MEMORY_DEBIAS, optional AGENT_MEMORY_EXPLORATORY_AUTO_RECALL)
• Harness — AGENT_PROMPT_CACHE (provider cache breakpoints); intent inference heuristics + research finalize judge; vault index helper
• Docs — Roadmap

---

v0.0.10 — 2026-05-23

Task Worlds landed earlier on May 23 then was removed the same day in favor of execution state and compensation wiring.

Shipped

• liminal CLI (dev) — setup, doctor, web, tui, update, path
• Install scripts: scripts/install.sh, scripts/install.ps1
• Recipe library v2 — compounds on strategy shape, not generic recipe: notes
• Harness power pass — execution state, compensation playback, breakdown, tool DAG depends_on, contract→tool mapping
• Workflow evals + additional core baseline tests
• Web fetch — Wikipedia extracts, pagination, distill bypass, model max_chars floor
• Persona UI themes — distinct palettes, HudShell theme wiring
• Per-harness workspace root, streaming edit previews in web

---

v0.0.9 — 2026-05-22

Streaming writes, reasoning budget, Playwright browser + CAPTCHA, document engine, license file (later FSL), compensation ledger.

Shipped

• Streaming large-file writes — .agent_write_staging/, manifest resume after truncation; write integrity nudges; legacy patch tools removed
• Compensation ledger for partial plan rollback; session tool-output index; outcome scoring; optional self-heal lint
• Intent classification and reasoning budget (effort + think-depth per turn); external think() + reason() surface; two-tier main/fast models; effort learning
• Playwright browser family + stealth; CAPTCHA solving (2captcha / CapSolver)
• Document engine — compose → lint → repair → PPTX / DOCX / PDF export with quality gate
• self_telemetry, rename_symbol, git_worktree; recipe library redesign; Settings catalog + provider presets in web UI
• MIT LICENSE file at this milestone (superseded by FSL-1.1-MIT in v0.0.12)

---

v0.0.8 — 2026-05-15

Harness defaults + web Settings, Obsidian vault discovery, persona bootstrap/themes, orchestration expansion.

Shipped

• HARNESS_ENV_DEFAULTS typed module (~190 keys); web Settings modal with field metadata and provider presets
• Obsidian vault auto-discovery from obsidian.json; web_fetch wall-clock budget + readability worker
• Removed web_research tool → web_search + parallel web_fetch
• Persona generator, bootstrap modal (TUI + web), PersonaUiTheme + web shell router, personality heartbeat (off by default)
• Expanded harness lifecycle hooks; orchestration catalog growth; VitePress docs restructure

---

v0.0.7 — 2026-05-12

World context, session JSONL, platform context, agenda/scheduler, auto-dream telemetry.

Shipped

• World context gatherers + harness rule recall at round 2; runtime prefs (.agent_runtime_prefs.json); platform context (OS, shell, git, ports, terminal snapshots)
• Agenda / scheduler / synthesis / independence tool family; web_fetch readability improvements
• Session JSONL — rollup / delta / trace modes under .agent_sessions/; provider pacing gate; auto-dream scanner telemetry in web UI
• TUI/web session metrics; runtime settings tools; experimental Android chat scaffold (later removed)

---

v0.0.6 — 2026-05-10

Unified file edits, research credibility tiers, dynamic tools, MCP client.

Shipped

• write_file + edit_file canonical surface; grep_file always-on
• Source credibility tiers and contradiction surfacing in research rules
• Dynamic tools persisted to disk; MCP client; progress events for long-running tools
• SSE reconnect hardening; provider retry lines visible in UI

---

v0.0.5 — 2026-05-09

Upgrade V — shared memory bus, tool DAG, semantic compression, goal/branch tools.

Shipped

• Shared memory bus; spaced-repetition decay in memory_rank
• TaskOrchestrator DAG with depends_on; semantic compression for context
• decompose_goal, branch_explore, verify_contract; plugin loader; adaptive protocol suffix

---

v0.0.4 — 2026-05-08

Vision sidecar, markets quotes, self-heal lint, document engine scaffold.

Shipped

• Vision sidecar (vision_analyze); markets_quote multi-source
• Self-heal lint loop (off by default); document engine scaffolding (doc_* family begins)
• Streaming resilience; long-horizon autonomy hooks; expanded file operations

---

v0.0.3 — 2026-05-05

Lazy tool families, hybrid recall, safety judge, workspace-root grounding.

Shipped

• Tool argument guard; vault path resolution; workspace-root grounding
• Safety judge optional pre-flight; hybrid recall (BM25 + embeddings)
• Lazy tool loading — list_tool_families / activate_tool_family
• README + architecture diagram fixes

---

v0.0.2 — 2026-05-03

Dynamic persona generation, child harness wiring.

Shipped

• LLM-generated personas with strength dial and modifiers
• onChildCreated child harness tool registration; resource lock caps on persona tooling

---

v0.0.1 — 2026-05-03

Monorepo scaffold, AgentHarness, tools, TUI, web, eval, Liminal rebrand.

Shipped

• Workspaces: packages/core, tools, tui, web, eval
• AgentHarness ReAct loop, approvals, context manager
• Initial tools: files, shell, git, memory, web, think / plan / reason
• Ink TUI + Express/React web with SSE; first eval scenarios

---

Planned (not started)

See the full Roadmap for what each milestone means.

Stage	Target	Status
Beta	Stability + defaults freeze candidate	Not started
RC	Ship checklist, docs freeze	Not started
v0.1.0	Public preview tag + install GA	Not started

---

Commits on main: github.com/traidy2222/liminal-ai/commits/main