Changelog

• Downloads are fast and resumable again — installers now stream from dl.tanvrit.com instead of Cloudflare's public development endpoint, which throttled one 260 MB installer from 13.9 MB/s to 125 KB/s inside a single minute; the new path measured 14.9–35.5 MB/s and supports resume, so an interrupted download picks up where it stopped
• The Cost dashboard does its own arithmetic correctly — $8.20 was printing as $8.1999, and a −$12.34 credit rendered as $0.00 with both the sign and the magnitude gone
• Invoices and billing history show the right amount in every currency — yen, won and the other zero-decimal currencies were 100× out, and the three-decimal dinars 10× out; rupee and dollar amounts were always correct, which is exactly why it went unnoticed
• Built against Tanvrit SDK 2.0.76, with every module this app resolves verified present in the artifact bucket rather than merely answering through the proxy

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• The Sign out button on the trial, workspace and profile screens is now actually reachable — on a shorter window it was rendering below the fold with no way to scroll to it
• Plan cards and the buttons beneath them no longer get clipped on a narrow window

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• You can always sign out — if your free trial ended, or you were partway through creating a workspace or completing your profile, those screens took over the whole window and left no way back to Account. Each of them now has a Sign out button, so you can switch accounts without quitting the app
• The Android build reports its real version again — it had been shipping as 1.5.21 no matter which release it was actually built from

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• The agent self-heals when a model can't use tools — if a local model describes a tool call instead of running it, Tanvrit AI now automatically retries the turn on a tool-capable model (when auto-routing is on and a capable model is available), so a coding task just works instead of silently stalling
• Every agent turn starts grounded — the in-app agent, sub-agents, the headless runner and connected MCP clients all open with a fused Context Pack (relevant code, key symbols + blast radius, design decisions, staged diff), so the model works with your codebase from the first message

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• Preview lives in the Workspace now — a Chat / Preview tab switcher in the right panel lets you preview the agent's web output right inside the IDE, instead of navigating to a separate screen
• Runs on the latest Tanvrit SDK — desktop sign-in now persists at the SDK layer as well (belt-and-suspenders with the 1.5.22 session fix), so your login survives restarts

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• Slimmer sidebar — the left rail now leads with your coding workspace and tucks the deeper Intelligence / Agents / Models / Connect / Observability sections behind one-click section headers, so you see the primary flow first instead of 30+ items; ⌘K still jumps to anything instantly

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• Stay signed in — fixed the desktop app asking you to log in on every launch; your session now persists across restarts, and a merely-expired access token silently refreshes instead of bouncing you back to the login screen
• Preview no longer crashes — fixed a native crash in the in-app browser engine where the Chromium framework was unpacked in the wrong place, aborting the whole app whenever you opened Preview
• The local agent uses its tools more reliably — read-file / search / web-search calls emitted by small local models are now parsed more robustly

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• Project-first Workspace — open a project from the Projects screen and the Workspace shows just that project (a VS Code-style single root) instead of every indexed project at once; 'All projects' is still there when you want everything
• The web app renders again — fixed a Compose/skiko version skew that showed a blank white screen at ai.tanvrit.com/app
• The in-app version now shows the real build (it was hardcoded to '1.3.17' on every build)
• Removed the redundant 'Add Project' menu item — use the + on the Projects screen

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch.

• Workspace explorer now scopes to the project you're in — a VS Code-style single-root tree — instead of dumping every indexed project into one list
• Carries all 1.5.19 fixes (no login-twice, MCP SSE) on Tanvrit SDK 2.0.20

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch. The VS Code extension stays at v1.4.0.

• No more logging in twice — email/password and social sign-in now persist your session on the first try, so a fresh launch keeps you signed in instead of bouncing back to the login screen
• Claude Code MCP now connects out of the box — the generated config uses the correct SSE transport (was emitting an http type the local SSE server rejected, showing as 'failed')
• Carries all 1.5.18 fixes (login persistence, sticky model pick, engine) built against Tanvrit SDK 2.0.20

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch. The VS Code extension stays at v1.4.0.

• Stay signed in — fixed the bug that bounced you to the email login screen on almost every launch; your session now survives transient network hiccups and only re-prompts on a real sign-out
• Your model pick sticks — choosing a specific model no longer silently reverts to the auto/budget model when you send a message
• The local inference engine now ships and starts cleanly, resolving the "Engine failed" state on first run
• Built against the latest Tanvrit SDK 2.0.19

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch. The VS Code extension stays at v1.4.0.

• Platform release train — built against the latest Tanvrit SDK 2.0.11, with a coordinated end-to-end deploy of the web portal, desktop installers, and the server on a single version
• Corrected the in-app SDK version display (was stale)

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch. The VS Code extension stays at v1.4.0.

• Platform-wide zero-defect pass — ~84 fixed defects across providers, the agent runtime, chat, search, the editor, the indexer, dashboards, the database, auth, the MCP server, the 19-language parsers, and telemetry; each fix shipped with a test
• AI Mesh — route a chat to a peer/fleet node for free inference, with a sticky per-conversation Private mode (trusted-only routing) and opt-in node serving on desktop
• Privacy hardening — the opt-in telemetry guard is now fail-closed and catches paths, IPs, tokens, and code, upholding the local-first promise
• Editor — search and symbol results now jump to the exact matched line
• Simpler, honest plans — free is local models + AI Mesh + bring-your-own API key; Pro (₹1,000/mo) adds AI Mesh priority

Native installers for macOS (Apple Silicon), Windows, and Linux. Unsigned builds — your OS may warn on first launch. The VS Code extension stays at v1.4.0 and pairs with desktop 1.5.x.

• Persistent conversation memory — every chat is saved to a local SQLite store and survives restarts; long conversations no longer 'forget' what you asked 10+ turns ago
• Semantic recall (RAG) — relevant earlier turns are embedded and re-injected into context automatically, plus an explicit recall_conversation tool (in-app and over MCP); recall is scoped to your workspace so it never bleeds across businesses
• Universal embeddings — a hosted embedding fallback brings semantic search + recall to the web and mobile builds, not just desktop with a local engine
• Pluggable model catalog — browse and pull GGUF models straight from HuggingFace (license shown up front, with a quantization pick), on top of the curated set; more sources are rolling out
• First fully cross-platform release — native installers for macOS, Windows, and Linux plus Android and iOS builds, all cut from one pipeline

The Wasm portal at /app/ runs in any browser; engine-dependent features still degrade gracefully there. The VS Code extension stays at v1.4.0 and pairs with desktop 1.5.x.

• Four named Engineering Agents — Code Reviewer, Test Writer, Migration, and Docs (each 1.0.0): documented products with per-agent tool scopes enforced by the runtime, not just requested in the prompt
• Agent Marketplace — browse and install curated agents and Domain Packs in-app; payloads are validated, built-in ids can't be shadowed by a download, and every install is provenance-stamped
• Domain Packs — manifest-driven packs bind a business domain to the agent runtime; opt-in per pack, and mutating tools are approval-gated exactly like write_file
• Commerce pack v1 — 12 governed commerce_* tools mirroring the /api/ai/v1 gateway routes (catalog, orders, products, inventory, customers, analytics), authenticated with a scoped X-API-Key stored in your OS keychain
• Store Operations Agent 1.0.0 — works the order queue and proposes confirm / decline / inventory actions; gateway-only scope, every money-affecting call summarized first and gated on your approval
• Agent scoping, verification, and attribution across the runtime — write jails (the Docs agent can physically only write Markdown), exec denied by default, requireCleanExit rolls back unverified runs
• Headless agents + compute scheduling now pair with the named agents — tanvrit-ai-agent CLI and one-click scheduled agent jobs on Tanvrit Compute
• Run retention — agent run history is pruned to a configurable cap (default 500) and spilled transcripts are cleaned up with their rows
• Security: agent endpoints on the Local API now require the bearer token — extension 1.3.x agent runs get 401 against this build; 1.4.0+ needs no change

Server side: the marketplace curation flow and seats/entitlements are live; hosted inference on the gateway is still rolling out. A self-host bundle ships in-repo as a preview (evaluation mode; license enforcement is roadmap). The VS Code extension stays at v1.4.0 and pairs fully with desktop 1.5.0.

• Agent run history — every run is persisted with tokens, cost, edit count, and source (UI / API / VS Code / Compute), browsable in the new Agent Runs screen
• Hybrid code search — FTS5 full-text fused with real semantic embeddings (local engine or Ollama nomic-embed-text); the new semantic_search MCP tool brings the set to 32
• Team Knowledge — opt-in sync of curated decisions, rationales, and annotations across your team; default OFF, every payload runs through the privacy redactor, and the raw code index never leaves your machine
• Headless agents — the tanvrit-ai-agent CLI runs plan-only by default; --auto-approve stages, verifies, and rolls back on failure; one-click scheduled agent jobs on Tanvrit Compute, cron included
• Pending-edits HTTP API — GET /v1/edits plus apply/reject endpoints power the VS Code 1.4.0 Pending Edits review through the same single-writer staging path as the UI
• Org policy enforcement — organization model/provider allowlists, local-only mode that locks routing, and a team usage dashboard
• /v1/knowledge/search on the Local API powers the VS Code knowledge view
• iOS Keychain entries hardened to AfterFirstUnlock accessibility, with lazy migration of legacy plaintext values

Desktop and the VS Code extension both reached v1.4.0 and pair fully. Cloud side at release time: the managed gateway and the team-knowledge backend were still rolling out — both are live as of 1.5.0.

• VS Code Copilot v2 — the extension graduates from a sidebar into a full copilot surface, live now on the Marketplace as v1.4.0
• Agent runs panel — start and watch streaming agent runs without leaving the editor, with live tool-call output
• Pending Edits review — every staged hunk shows up in the editor for apply / reject; the extension never writes files itself, the desktop app applies what you approve
• Editor context sync — your active file and selection follow you into agent runs automatically
• Agent run history — browse past runs with status from the sidebar
• Knowledge view feature-detects the connected desktop version and only lights up when the backing surface exists
• Chat panel fixes — message ordering, scroll retention, and clean reconnect after the desktop app restarts

VS Code extension-only release — desktop binaries remain at v1.3.1 and pair with this extension today. Platform program: the six-pillar work (AI Router gateway, Agent Runtime staged edits + run attribution, Team Knowledge sync, Enterprise controls, hybrid search) lands with desktop 1.4; the new extension surfaces light up automatically when it ships.

• Windows: local model downloads fixed — the models folder was being created beneath the index.db file on backslash paths; path derivation is now separator-agnostic
• Windows: Editor file tree renders real folders (paths were arriving as one flat list of \Client\… labels); re-index existing projects to pick this up
• Preview: the one-time browser-runtime download now survives navigation, shows live progress, and failures get a Retry banner — no more permanently blank pane
• Database screen: table tabs and wide tables now scroll horizontally instead of cropping at narrow window widths
• App window/taskbar icon fixed on Windows; Linux .deb ships from this release onward
• Web: on-device AI (wllama) module loading fixed; in-browser local inference works again

Patch release for the Windows QA sprint — recommended upgrade for all Windows users on 1.3.0.

• All-new premium interface — dense, keyboard-first design with a hairline surface system, 13sp type scale, and a single indigo→cyan brand moment per screen
• ⌘K / Ctrl+K command palette — fuzzy search over every screen and app action (copy MCP URL/token, re-index, switch appearance), with persisted recents
• Appearance is now persisted — light mode first by default; Dark and System one tap away in Account; code highlighting follows the active theme
• Global server-status surfaces — service dots in the sidebar footer plus an incident banner with one-click Retry when the MCP or Local API fails
• Architecture Explorer now computes a real dependency graph — imports are resolved into file-to-file edges during indexing (re-index to populate)
• Native macOS folder picker for Import Project (Finder sheet, replacing the Swing dialog); web app boot fixes and on-device model loading on the web target
• MCP connections — Setup Guide config snippets now include the Authorization header and type: "http"; session attribution names clients by User-Agent

• Silent first-launch auto-install — fresh installs get a working local default model with no manual download step
• Qwen 2.5 Coder 3B Q4_K_M flipped to isAutoInstallOnFirstRun in LocalModelCatalog (the existing chat default)
• Dashboard pill shows download progress; banner flips to "Tanvrit ready" on completion
• Auto-install only fires when no chat model is on disk AND no cloud provider API key is saved — cloud-only users are never interrupted
• Same DMG / R2 / Pages pipeline as v1.2.2, no schema changes

• Tanvrit Local is now the default chat provider on every target — Web, Desktop, Android (and iOS via scripts/enable-ios-llama.sh)
• Fixed: Tanvrit Local missing from the Agent provider picker on Wasm / Android / iOS — now registered FIRST in the common aiModule registry, not just the desktop override
• Fixed: Wasm provider self-disabled by the LocalModelManager install gate — new LlamaEngine.requiresLocalFile lets streaming engines bypass the install check and load the catalog proxy URL directly
• availableChatModels() replaces installedChatModels() to reflect the new installed-vs-streamed dual mode
• Project-wide FQN cleanup — fully-qualified class names replaced with import statements across composeApp/src/

• Real engines for the two targets v1.2.0 stubbed with NoopLlamaEngine
• Web (WASM) real WasmLlamaEngine via the wllama npm package (@wllama/wllama 3.4.1) over the existing R2 proxy — bridges wllama's onNewToken callback into a Kotlin Flow
• PlatformModule.wasmJs.kt binds the real engine and registers TanvritLocalLlmProvider; ~4 GB browser-tab memory cap means only smaller catalog entries are practical
• iOS opt-in script scripts/enable-ios-llama.sh — builds llama.cpp static libs for iosArm64 (Metal) + iosSimulatorArm64 (CPU) and flips the DI binding from Noop to IosLlamaEngine
• Shared TanvritLocalLlmProvider + LlamaEngine abstraction untouched — same ChatML prompt rendering everywhere

• KMP LlamaEngine lit up across every target with the same de.kherud:llama JNI jar
• Android real-engine binding (AndroidLlamaEngine) — 1:1 mirror of DesktopLlamaEngine; libjllama.so extracts into the app cache on first use (smaller Qwen 0.5B / 1.5B entries practical on phones)
• iOS + Web ship NoopLlamaEngine with a documented path forward (cinterop against llama.cpp for iOS, wllama WASM for Web); TanvritLocalLlmProvider self-disables so the UI keeps compiling
• ServerHealthRegistry.updateLocalLlm adds localLlm: ServerState to the health snapshot so the Dashboard reflects the in-process LLM lifecycle
• Same commonMain code path everywhere — provider, ChatML rendering, and per-family stop sequences written once and reused

• MCP `/mcp` endpoint reachable — SSE handshake at http://localhost:19281/ (SDK 0.8.3 mounts at routing root)
• Login wall at launch removed — cold start lands directly on Dashboard
• Tiny Stack auto-install works — placeholder-trust SHA path, real catalog URLs
• Settings → AI Providers → Ollama: one-tap install + start daemon card
• AgentScreen empty-state CTAs (Install Tiny Stack / Install Ollama) + send-with-no-model surface
• AuthUiConfig restricted to email-OTP (phone-OTP hidden until SMS gateway)
• release-* workflows on self-hosted Tanvrit Mac runners — bypasses Actions cap

• 90-day plan shipped end-to-end — A1–A4 / B1–B4 / C1–C3 / D1–D3 / E1–E3 (17 themes)
• Plan/Execute mode with explicit Approve / Reject buttons
• Sub-agents framework (Task tool)
• Refactor preview with undo stack
• Context budget UI with pin/drop and ratio header
• Symbol-aware retrieval (get_symbol_context MCP tool)
• Knowledge-graph schema + Learn-from-PR ingestion
• Tanvrit Tiny Stack — Qwen2.5-Coder 3B chat + 1.5B FIM + BGE-small + jina-reranker-tiny + ModernBERT (~3.65 GB total)
• AutoRouter with rationale-pill surfacing
• Privacy redaction (13 secret rules + opt-in PII)
• Universal MCP bus + handoff URI + MCP 2026 RC scaffolding
• Annotation layer (multiplayer V1)
• Hooks system (Claude Code parity)

• Initial public release
• 31 MCP tools live on localhost:19281
• 19 language parsers (ANTLR-based)
• Ollama + Anthropic / OpenAI / Gemini / DeepSeek / Groq / Mistral providers
• OpenAI-compatible local API on localhost:19280
• Wasm portal at /app/ (engine-dependent features degraded)
• VS Code extension v1.0.0
• DPDPA-aligned privacy posture, telemetry off by default

• Final beta — RC for v1.0
• ConversationSummarizer landed for long agent sessions
• Bearer token rotation in Settings
• Self-hosted release runner pipeline stabilised

• AgentLoop with budget guards and stopping conditions
• Qwen2.5-VL vision integration with Tesseract OCR fallback
• Dependency graph + impact analysis tools

• MCP HTTP transport (alongside stdio)
• Authorization: Bearer auth for HTTP transport
• Conversation memory tools (store / search / update / entities)

• Wasm portal first cut (read-only)
• Telemetry framework with PII guard (off by default)
• Settings UI for AI provider keys

• Git intelligence: git_blame, git_history
• find_dead_code and find_complex_files
• Cyclomatic complexity metrics

• First closed beta
• Indexer + parser pipeline for 19 languages
• Initial MCP tool set (12 tools)
• Desktop builds for macOS, Windows, Linux