v1.6.0

v1.6.0

Keep browser sessions and MCP connections ready between messages, with bounded idle cleanup and explicit ownership. Add browser-free Codex OAuth and integration with the real official C# SDK, making three explicit backend choices available. Each website generation still obtains fresh Sentinel answers; MCP recovery avoids repeating tool side effects.

MCP reliability

• Reuse owned HTTP and stdio connections within the same account, application user and chat. Endpoint, credentials and process settings form part of the connection identity; scoped server selections and tool allowlists are checked again on each message.
• Refresh tool discovery for every message and retry transient initialization or discovery failures once by default. An expired HTTP session is replaced before execution. Authentication failures and application-supplied clients are not automatically retried.
• Resume interrupted Streamable HTTP SSE responses with GET and Last-Event-ID when the server supports it. Never repeat the original tool-call POST; an outcome that cannot be confirmed retains the existing reconciliation requirement.
• Accept tool blocks wrapped in Markdown fences and request one hidden correction for malformed JSON or arguments before executing any tools in that batch. Approval, call limits and durable execution records remain in effect. Progress observer failures no longer change a confirmed tool outcome.

Browser and Sentinel

• Retain isolated account sessions for two minutes of inactivity by default, avoiding a fresh Chromium process and context on every turn. Reuse is configurable, session counts are bounded, changed credentials replace the context, and asynchronous disposal releases owned resources.
• Add WarmUpAsync() to load the authenticated composer ahead of a send, plus sanitized phase timings through BrowserSentinelOptions.TimingObserved. Warm-up does not submit a draft or request challenge answers.
• Preload the next composer after capture while the current HTTP generation runs, enabled by default. Reuse the pending navigation on the next send instead of loading it again. Preloading collects no fresh proofs; disabling it restores loading at send time.
• Keep the same browser open through repeated Cloudflare challenge pages until the configured timeout or cancellation. A repeated challenge response alone does not close or restart it; configured startup retries follow a full failed attempt.
• Keep generation interception active while a session is idle, including conversation URLs with query parameters. Fresh answers are captured only during an SDK handshake, and the browser-generated draft request is aborted before transmission.
• Preserve invisible Chromium support and automatic hardware-to-software rendering fallback. HTTP-only mode continues to accept an external authorized Sentinel provider. Actual challenge latency depends on the upstream service and sign-in state; a one-second live handshake is not guaranteed.

Optional Codex OAuth transport

• Add a separate browser-free transport adapted from evanzhoudev/openai-oauth, with upstream Apache-2.0 attribution retained in NuGet packages and downloadable bundles. The default web backend continues to bind actual ChatGPT website conversations.
• Add explicit Codex auth-file import, PKCE authorization/callback exchange, account validation and serialized refresh-token rotation. Credential writes are atomic, preserve unrelated file fields and use owner-only permissions on Unix. The SDK does not automatically inspect or modify a shared Codex login.
• Support the existing official .NET Responses, Chat Completions and model clients over OAuth, including SSE, native function-call results, one-shot tool arguments and non-streaming response collection. Truncated or failed streams do not become successful completions, and generation requests are never automatically replayed.
• Support direct OAuth image generation/editing through the existing ImageClient, with base64 output, up to five reference images, bounded uploads and explicit unsupported-option errors.
• Discover account-specific models and cache their capabilities, including reasoning/verbosity defaults and Responses Lite request normalization. Keep local continuation and item-reference history isolated by account, user and thread, with count and memory limits. This stateless backend replays history and does not create website chats.

Native official SDK and backend selection

• Add ChatGPTWebSdk.Native, which references the real signed OpenAI 2.14.0 NuGet library and supplies explicit web and Codex client factories. Existing native constructors retain their platform defaults; official, web and OAuth clients can coexist without globally changing transport.
• Add ChatGPTOfficial to both package alternatives for API-key clients with the native endpoint, authentication, retry options and platform features. Backend selection never silently retries a generation or moves a website conversation ID to another service.
• Keep the integration and drop-in packages separate to avoid duplicate OpenAI assemblies. Publish the Native NuGet alongside the existing packages and verify its three methods against the original SDK.

Lifecycle, packaging and validation

• Add asynchronous runtime/client disposal and preserve application ownership of supplied providers and MCP clients. Update the proxy and diagnostic commands to release the resources they create.
• Require native browser bundles to pass an offline test in real Chromium: repeated challenge pages, warm-up, two fresh handshakes, one reused session and interception of generation requests. The diagnostic must load the exact packaged SDK assemblies; prepare/verify receipts include the core, browser and compatibility DLL hashes.
• Add 155 core/drop-in SDK cases, 18 native SDK integration cases and a native packaging gate, bringing the automated gate to 1,361 core/drop-in SDK, 18 native SDK, 167 automation and 12 extension tests. Packages and browser extensions share version 1.6.0; the OpenAI compatibility assembly keeps version 2.14.0.0.

See verification details for the measured local results and live-service limits. The release workflow adds all linked commits in a collapsed categorized appendix.

All 36 commits

Features

• aeb856a feat(resources): add bounded exclusive async leases
• def7774 feat(mcp): configure discovery recovery and SSE resumption
• 8d87b2c feat(mcp): reuse scoped connections and repair tool requests
• b6c0940 feat(runtime): add warmup and asynchronous ownership cleanup
• f1f2f9c feat(oauth): add explicit sessions and serialized refresh
• 868d3d7 feat(codex): bound isolated response continuation history
• 5c839c3 feat(codex): adapt Responses and Chat over direct OAuth HTTP
• bfbf1bd feat(images): support Codex generation and reference edits
• 089feb0 feat(codex): expose explicit typed OAuth clients
• d11cc24 feat(official): add explicit platform API clients
• eae7140 feat(native): integrate the real official OpenAI package
• dfbf13c feat(cli): verify real browser reuse without remote generation

Tests

• a1e1a0c test(resources): cover lease contention and cleanup
• f478316 test(mcp): verify HTTP and stdio session recovery
• e06ae60 test(mcp): cover tool protocol repairs and uncertain calls
• f9948cb test(sentinel): verify fresh proofs and challenge waiting
• b852ab3 test(runtime): verify owned and external resource lifetimes
• 2b7f956 test(oauth): verify callbacks refresh and private storage
• f223813 test(codex): cover streaming history and model capabilities
• 93d4c53 test(images): verify OAuth image input and output contracts
• b2d54ff test(official): verify platform and web clients coexist
• d06ab3d test(native): validate all three backends and public signatures
• cdcc17e test(ci): enforce complete native build and release gates

Implementation

• 4ad86cd perf(sentinel): reuse sessions and preload the next composer

Fixes

• da6e3f9 fix(proxy): dispose retained browser and MCP resources

Security

• 078ef99 feat(security): add state-checked PKCE authentication

Documentation and maintenance

• e9c9581 chore(licenses): retain OAuth source attribution and notices
• 00f5b70 docs(mcp): explain connection reuse and recovery controls
• a0cfe5e docs(auth): document OAuth setup and warm browser sessions
• 70daacf docs(native): explain official SDK migration and backend selection
• e414225 docs(readme): introduce three backends and faster warm sessions
• 4827547 docs(verification): record package and native browser acceptance
• ff02769 chore(release): 1.6.0
• d9dca71 docs(native): explain release package installation

Build and CI

• c8f5c37 build(packaging): ship native integration and verify reused browsers
• aef4c8e ci(release): require native package and integration tests