Sentinel Coder One Studio

Your models. Your keys. Your machine. An autonomous AI coding agent for VS Code that is strictly bring-your-own — API keys or local models, no bundled provider, no vendor lock-in, no telemetry surprises.

⚡ Works out of the box with any ONE of: a free local Ornith 9B (Ollama, tools + reasoning on a 12 GB GPU) · Kimi K3 frontier (1M context via your Moonshot key) · or any of 340+ models across OpenAI, Anthropic, Google, Groq, OpenRouter, DeepSeek, Mistral, Azure, Hugging Face, Moonshot, and custom OpenAI-compatible servers. If your selected model is unavailable, the agent automatically falls back to the first connected one.

Product links:

• Qubitpage platform: https://qubitpage.com
• Sentinel Coder repository: https://github.com/qubitpage/sentinel-coder-vscode
• Marketplace source of truth: docs/MARKETPLACE_CANONICAL_RECORD.md
• Multi-provider model guide: docs/MULTI_PROVIDER_MODEL_ARTICLE.md
• Remote workspace tooling: docs/REMOTE_WORKSPACE_TOOLS.md — run agent commands through the active VS Code Remote (SSH/WSL/tunnel) host; never asks for SSH private keys.
• Multi-session terminal pool: agent commands run in named persistent terminal sessions with resource guardrails — sentinelCoder.terminalMaxSessions caps parallel sessions, sentinelCoder.terminalMinFreeMemoryMb blocks new sessions under memory pressure, sentinelCoder.terminalIdleCleanupSeconds reclaims idle ones.

Current Release: 3.27.4 — Native MCP Tool Dispatch, Anti-Stall Agent Loop, TOON Token Compression

The 3.27 line is a reliability and cost release focused on one goal: the agent finishes the task instead of narrating it. Every fix below is verified in the compiled build (npx tsc -p ./ exit 0).

• NEW — native-loop MCP tool dispatch: MCP tools are now callable directly inside the native tool_calls agent loop via the mcp__<server>__<tool> bridge (src/mcpClient.ts). Filesystem, memory, search, Postgres and your own VS Code MCP-config servers are dispatched without falling back to the text protocol, so MCP-heavy turns stay on the fast native path.
• NEW — mandatory self-critique guard: when a non-trivial turn (one that used tools or touched files) tries to finish on prose alone, the agent loop requires the selfCritique tool to return PASS before completing. This eliminates the "explained the discovery but never applied the fix" failure mode.
• NEW — zero-tool loop guard: if the model replies with prose (no tool calls) twice in a row, the loop forces a tool call instead of letting the turn stall mid-task. Combined with the self-critique guard, mid-session stalls are structurally prevented.
• NEW — TOON token compression + dynamic tool loading: the context budgeter compresses tool/history payloads and loads only the tools a turn needs, cutting per-turn token spend.
• NEW — adaptive Moonshot tier-aware rate limiting: an AIMD pacing governor with Retry-After propagation is wired into both HTTP loops, ending the Kimi 429 mid-session stalls.
• Accurate token accounting: the token estimator now counts the tool_calls payload itself, so context-budget decisions reflect real request size; an undefined-tool strictness guard rejects malformed calls instead of silently dropping them.
• Agentic profiles: ships the Kimi Frontier Swarm (K3 orchestrator + K2.7 code workers + K2.6 vision) and the cost-optimized Kimi Council (all-Moonshot team tuned for cache-hit pricing), alongside the local-free and hybrid profiles.
• Regression coverage: tests/agent-loop-toolcall-regression.cjs (24 assertions) and tests/token-opt.test.cjs guard the native tool loop, MCP bridge, and token optimizer against regressions; tests/frontier-benchmark.cjs compares the orchestrator against configured frontier models on identical agentic tasks.

Previous Release: 3.24.18 Reliability — No More Mid-Task Stops, Kimi Council, Foundry IQ Off By Default

3.24.18 fixes the "stops in the middle of the task / very short answers" class of bugs and completes the Kimi lineup:

• FIXED — empty-output turn endings: reasoning models (Kimi K3, o-series) could burn the whole token budget on internal thinking and return empty content; the agent loop treated that as "done" and silently ended mid-task — even on Autopilot. The loop now detects empty output and demands the continuation/final answer (up to 2 retries) before ever giving up.
• FIXED — frozen-looking thinking: Moonshot/DeepSeek-style reasoning_content stream deltas are now surfaced live as thinking, so K3's long reasoning phase shows progress instead of a dead chat.
• FIXED — Moonshot 400 errors: the ENTIRE current Kimi catalog is fixed-temperature; the extension now always sends temperature: 1 for Moonshot (verified live 2026-07-26).
• Foundry IQ is OFF by default (the contributed setting default was true and silently overrode the code default). It is strictly optional grounding — local RAG and session memory work standalone.
• Live Moonshot catalog: defaults now match the real /v1/models lineup — kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6, moonshot-v1-auto (stale kimi-latest/k2-0905 removed).
• New profile — Kimi Council: best all-Moonshot team — K3 plans/reviews/finalizes, K2.7 Code drafts implementation, K2.7 Code Highspeed does fast boilerplate — tuned for cache-hit pricing ($0.30/M vs $3/M) and minimal K3 output tokens ($15/M).

Previous Release: 3.24.16 Model Options, Universal Provider Fallback, Studio Media Hub

3.24.16 adds Copilot-style per-model controls and removes the last provider requirement:

• Model Options panel next to the model selector: choose reasoning/thinking effort (Auto/Off/Low/Medium/High) and context size (Auto/8K→1M) per model, persisted across sessions. Effort maps to native provider controls (Ollama think, OpenAI/Azure reasoning_effort, Anthropic thinking budget, Gemini thinking budget); context size caps that model's request window and token budget — lower = cheaper/faster.
• Universal fallback: if no local model exists (or the provider is disconnected), the chat automatically uses the first connected model from ANY enabled provider — the extension never requires a specific provider, local or API, to work.
• Media Studio hub moved to the Studio tab (image · video · audio · transcript generator cards with a send-to-chat bridge); the chat block is now a slim collapsed dropdown.

Previous Release: 3.24.15 BYO-Only Providers, Ornith 9B Standard, Kimi K3 Frontier, Browser Tools

3.24.15 makes the provider model strictly bring-your-own, fixes the local runtime crash, activates Kimi K3, and adds the Sentinel Browser Tools foundation:

• BYO-only providers: every hardcoded personal/provider endpoint is removed from the default catalog. Azure OpenAI/Foundry and Sora are now optional empty templates you point at YOUR resource; the extension never requires Azure/Foundry or any specific provider to work.
• Ornith 9B standard local model: the selector offers exactly one known-good local model by default (Ornith 9B, tools + reasoning, 40K context), it is re-selected as the active model on every VS Code start, and it is warmed into GPU memory at startup. Set sentinelCoder.ollamaShowAllModels: true to list every installed Ollama model.
• Kimi K3 frontier via Moonshot API: 1,048,576-token context, native tool calling + reasoning + vision, ranked tier-0 frontier in Auto routing. Selector surfaces official pricing: $0.30/M input (cache hit) · $3.00/M input (cache miss) · $15.00/M output. The extension automatically sends the fixed temperature: 1 the K3 API requires.
• New built-in agentic profiles: Hybrid: Ornith 9B local + Kimi K3 frontier (free local worker drafts/implements/runs tools, K3 plans/reviews/finalizes) and Kimi K3 Frontier Production (provider-independent successor to the Azure GPT-5.5 triad, cache-pricing and NHP-1 optimized).
• FIXED Cannot find package 'node-llama-cpp': the NeuroPack (.npk) native runtime is strictly optional — it activates only when a models folder is configured AND the optional dependency is installed; otherwise local chat cleanly uses Ollama with no errors.
• Adds isolated browser-tool actions for live web navigation, click, type, hover, drag, screenshots, console logs, and bounded scripted flows through a trusted local/enterprise sidecar.
• Adds enterprise controls: disabled by default, explicit enable command, allow/block domain policy, sensitive permission approval gates, non-persistent agent tabs by default, and private sessions that are never persisted.
• Removes vendor-specific assistant branding from the Marketplace-facing copy and replaces it with Sentinel/Qubitpage-owned wording.

Previous Release: 3.24.14

3.24.14 migrates existing installs away from the retired api.qubitpage.com gateway:

• Existing sentinelCoder.cloud.gatewayUrl=https://api.qubitpage.com settings are automatically migrated to https://qubitpage.com during activation.
• Fixes getaddrinfo ENOTFOUND api.qubitpage.com for users who installed before the path-based qubitpage.com auth default.
• Keeps the visible Sign in / Account Sentinel Cloud buttons and the VS Code AuthenticationProvider PKCE browser login flow.

Previous Release: 3.24.13

3.24.13 adds a visible Sentinel Cloud login path in the chat UI and aligns the web platform auth default with qubitpage.com:

• Added top-bar Sign in and Account buttons that call the existing VS Code AuthenticationProvider VS Code AuthenticationProvider flow.
• Verified https://qubitpage.com/sentinel-auth/login returns the Authorize VS Code page.
• Verified POST https://qubitpage.com/sentinel-auth/start returns { "ok": true } for a PKCE start request.
• Corrected sentinelCoder.cloud.gatewayUrl default from the stale api.qubitpage.com host to https://qubitpage.com.

Previous Release: 3.24.12

3.24.12 fixes UI mojibake in the Sentinel chat shell and publishes the cleaned Marketplace overview:

• Fixed corrupted Media Studio / Project Studio separator text (image * video * audio * transcript, checkpoints/tasks/generated assets/file paths).
• Cleaned remaining mojibake from src/sidebarProvider.ts UI/system strings before compiling the packaged out files.
• Keeps the 3.24.11 canonical Marketplace record and deduplicated overview.

Previous Release: 3.24.11

3.24.11 consolidates the Marketplace description and local package metadata into one deduplicated canonical record:

• Removed the old stacked release-history overview from the Marketplace README.
• Reconciled the live Marketplace 3.24.8-era text with the local 3.24.10 source/manifest state.
• Scanned the extension manifest, source files, media webview files, docs, README, and changelog.
• Normalized provider/tool/setting/command terminology.
• Preserved all unique capabilities while correcting stale or duplicated wording.
• Added docs/MARKETPLACE_CANONICAL_RECORD.md as the master Marketplace/product record.

Production Local Agent

The default local path is ollama:ornith-9b-classic:latest.

• Free, local, private coding agent on a 12 GB GPU class workstation.
• Production runtime settings: num_ctx=40960, num_batch=512, num_gpu=999, keep_alive=60m.
• VS Code activation can auto-start ollama serve and warm the model.
• Windows login recovery is supported by scripts/install-ollama-autostart.ps1.
• Tested by tests/ornith-9b-production-e2e.cjs: service health, warmup, streaming, long-context recall, function calling, provider tool bridge, RAG fallback, multi-turn coherence, latency, and GPU/RAM snapshots.

Evidence summary: docs/MARKETPLACE_CANONICAL_RECORD.md

Native NeuroPack / NPK Runtime

Sentinel can run .npk models inside VS Code on the GPU without Ollama or an external server.

• Provider: npk / NeuroPack (native).
• Runtime: node-llama-cpp, Vulkan GPU backend, NPK2 payload extraction/cache, flash attention, context ladder, RAM/VRAM guards.
• NPK Q8 and Q4 Ornith 9B paths use the same agentic/tool/RAG loop as cloud/Ollama models through a text tool bridge.
• Local model compaction protects NPK from huge persisted chat sessions.
• Tested by tests/npk-production-e2e.cjs: real huge-session compaction, Q8/Q4 tool engagement, and Q8 coding/reasoning/tool-call behavior.

Evidence summary: docs/MARKETPLACE_CANONICAL_RECORD.md

Core Capabilities

• Sentinel Browser Tools for live web-app verification: navigate, click, type, hover, drag, screenshots, console logs, bounded scripted flows, isolated agent sessions, non-persistent private sessions, explicit permission prompts, and enterprise allow/block domain controls.
• Qubitpage platform integration: Sentinel Cloud login, account/credits visibility, qubitpage.com gateway default, and roadmap alignment with the Qubitpage web IDE and AI Compute Pool / GPU pooling network.
• Agent, Ask, and Plan modes.
• Tool-using autonomous agent loop with file, editor, terminal, git, Docker, SSH, HTTP, browser, RAG, media, and security tools.
• Approval modes: default, bypass, autopilot.
• Plan tracking, checkpoints/undo, diagnostics verification, repeat-call guard, and forced finalization for stuck loops.
• Configurable Agentic Profiles with boss/worker/reviewer pools and cost policies.
• Intelligent orchestration router: direct, delegate, swarm, review, escalate.
• Native Inter-Agent Protocol (NHP-1) with compact handoffs and shared context ledger; measured about 48.6% fewer orchestration tokens.
• Orchestration trace panel with timing, token attribution, escalation, verification, synthesis, and savings events.
• Persistent sessions and live partial-output persistence across switches/reloads.
• Dynamic context injection: active file, open tabs, diagnostics, git state, model/provider metadata.
• Inline completions, AI code review, and PR summary commands.
• Media & Document Studio for Azure Sora 2 video generation, azure:MAI-Image-2e image generation, Azure gpt-image-2 images, speech, transcripts, screenshots/OCR, Office files, and generated assets.
• RAG plus Microsoft Foundry IQ grounding, with local fallback memory when the RAG server is unavailable.
• MCP servers: filesystem, memory, Brave Search, Postgres, and imported VS Code MCP configs.
• VS Code Web/remote bridge support for browser workspaces with a trusted remote tool host.
• Sentinel Cloud sign-in/account/credits plus bring-your-own-key provider support.

Providers

Sentinel supports local, hosted, and OpenAI-compatible providers:

• Local: Ollama, native NPK.
• Microsoft/Azure: Azure OpenAI / Azure AI Foundry, Azure Foundry Sora 2, Azure Speech, Foundry IQ.
• Cloud APIs: OpenAI, Anthropic, Google/Gemini, Groq, OpenRouter, DeepSeek, Mistral, Together AI, Vultr Inference, Hugging Face, Featherless, Moonshot/Kimi.
• Custom/self-hosted: custom OpenAI-compatible endpoints, QubGPU endpoint, experimental NeuroPack endpoint.

Tool Categories

The tool registry includes filesystem, editor, terminal, git, search, web, network, RAG, workspace/media, and security tools. Representative examples:

• Files: createFile, readFile, inspectFile, editFile, appendFile, deleteFile, listDirectory.
• Editor/workspace: getActiveFile, getSelection, insertText, getDiagnostics, getOpenTabs, getWorkspaceInfo.
• Terminal/remote: runCommand, remoteWorkspaceCommand, dockerCommand, sshCommand.
• Search/git: searchFiles, searchText, codebaseSearch, gitStatus, gitDiff, gitLog, gitCommit, gitPush.
• Web/RAG/MCP-adjacent: webSearch, httpRequest, queryRAG, ingestRAG.
• Media/documents: prepareGeneratedWorkspace, discoverMediaModels, generateImage, generateVideo, generateSpeech, transcribeAudio, captureScreenshot, ocrImage, analyzeImage, createOfficeDocument.
• Safety: firewallScan.

See docs/MARKETPLACE_CANONICAL_RECORD.md for the full deduplicated table with category and danger level.

Commands

• Sentinel Coder: Set Ollama Endpoint
• Sentinel Coder: Clear Chat History
• Sentinel Coder: Refresh Media & Document Studio
• Sentinel Coder: Open Media & Document Studio
• Sentinel Coder: Review Changes (AI Code Review)
• Sentinel Coder: Generate PR Summary
• Sentinel Coder: Toggle Inline Completions
• Sentinel Coder: Add NeuroPack (.npk) Model from Folder
• Sentinel Cloud sign-in/sign-out/account commands
• Atlas Voice Bridge status/send commands
• VS Code Web/Remote Tool Bridge status/help/configure/test commands

Settings Summary

Important setting groups:

• Local models: sentinelCoder.ollamaUrl, npkModelsDir, model, ollamaAutoStart, ollamaWarmModelOnStartup, ollamaKeepAlive, ollamaContextWindow, ollamaNumBatch, ollamaGpuLayers.
• Generation/context: maxTokens, contextBudgetTokens, temperature, defaultMode, approvalMode.
• Providers/secrets: providers, apiKeysFile, Azure OpenAI resource group/account.
• Inline/GitHub: inlineCompletions.*, github.reviewModel.
• Terminal safety: terminalMaxSessions, terminalMinFreeMemoryMb, terminalIdleCleanupSeconds.
• MCP: mcpServers.
• Dynamic context: dynamicContext.* plus legacy flat keys for compatibility.
• Web remote bridge: webRemoteBridgeEnabled, webRemoteBridgeUrl.
• Microsoft IQ / Foundry IQ: microsoftIq.*.

Quick Start

Local Ornith via Ollama

1. Install Ollama from https://ollama.com/.
2. Install/link ornith-9b-classic:latest from the Ornith 9B GGUF package: https://huggingface.co/qubitpage/ornith-9b-classic-gguf.
3. Open Sentinel Coder One Studio. The extension can start/warm Ollama automatically when configured.
4. Use Agent mode for tool-using local coding.

Native NPK

1. Convert or download an .npk model.
2. Run Sentinel Coder: Add NeuroPack (.npk) Model from Folder.
3. Select the NPK model in the picker.
4. Use Agent mode for tool/RAG/SSH/file workflows without Ollama.

Converter: https://github.com/msrusu87/npk-converter

Cloud Providers

1. Open the Sentinel settings/provider panel.
2. Add provider keys; secrets are stored in VS Code SecretStorage where applicable.
3. Select Azure/OpenRouter/Groq/OpenAI/etc. models from the live model picker.

Security and Privacy

• Bring your own keys; no provider keys are committed to source.
• Local models keep code on-device.
• Cloud providers receive prompt/tool context only when selected.
• Foundry IQ credentials are read from environment variables and preflighted for bearer/JWT freshness.
• Tool operations carry danger levels and are mediated by approval mode.
• Full terminal/SSH/Docker/MCP/local model tools require a trusted desktop or remote extension host.

Honest Limitations

• Ollama Ornith Q8 uses about 10 GB VRAM after warmup on RTX 3060; one large local model at a time is the safe rule.
• The default production local context is 40K even when model metadata advertises more, because VRAM/RAM safety matters.
• NPK Q8 runtime context is about 16K on this 12 GB GPU.
• Q4 NPK Windows/Vulkan explicit teardown can be unstable; production runtime guards unload and isolated tests avoid crashing the main VS Code host.
• Live vector RAG requires training/rag_server.py; otherwise local fallback memory is used.
• Pure browser VS Code workspaces need a trusted remote bridge for terminal/SSH/Docker/MCP/local-model tools.
• Hosted cloud agents are not included.
• Ornith 35B NPK-TQ on 12 GB is research-only; measured TQ2 quality is too lossy without new packed-code GPU kernels and layer eval.

Evidence and Reports

• Canonical Marketplace Record
• Verified Evidence
• Current Limitations and Honest Notes

Development

npm --prefix d:\QubitDev\vscode-ext run compile
npm --prefix d:\QubitDev\vscode-ext run package
node d:\QubitDev\vscode-ext\tests\ornith-9b-production-e2e.cjs
node d:\QubitDev\vscode-ext\tests\npk-production-e2e.cjs

For isolated extension-host testing without restarting the main VS Code window:

powershell -NoProfile -ExecutionPolicy Bypass -File d:\QubitDev\vscode-ext\scripts\run-extension-devhost.ps1 -Workspace d:\QubitDev