OpenAI-compatible server reliability improvements via HTTP/1.1 keep-alive config and error handling#404
Open
ahundt wants to merge 4 commits into
Open
Conversation
Previous behavior: Local OpenAI-compatible servers experienced intermittent connection failures and streaming issues due to missing HTTP/1.1 keep-alive configuration and inadequate timeout handling. What changed: - packages/core/src/core/openaiContentGenerator.ts: Add HTTP agent configuration with proper keep-alive settings, implement retry logic with exponential backoff, add constants for timeout and connection pooling - packages/cli/src/config/config.ts: Consolidate OpenAI configuration functions, add setupOpenAIFromCliArgs and getEffectiveAuthType for proper config hierarchy (CLI > env > settings) - packages/cli/src/validateNonInterActiveAuth.ts: Consolidate debug logging into multi-line statements, use getEffectiveAuthType for consistent auth resolution - packages/cli/src/gemini.tsx: Fix import order, add error logging in debug mode, update imports after config consolidation - packages/core/src/core/contentGenerator.ts: Add getBackendName function for human-readable API names in error messages - packages/core/src/core/geminiChat.ts: Add getConfig() method to expose configuration for error reporting - packages/core/src/core/turn.ts: Use getBackendName for better error messages - packages/cli/src/nonInteractiveCli.ts: Apply OpenAI CLI arguments before auth initialization - docs/cli/openai-auth.md: Document local server configuration options Why: Local OpenAI-compatible servers often have different connection characteristics than cloud services, requiring specific HTTP agent configuration to maintain stable connections. The changes ensure reliable streaming and prevent premature connection closures. Testable: Run with local OpenAI-compatible server using OPENAI_BASE_URL=http://localhost:1234/v1 and verify stable streaming without connection drops.
wenshao
requested changes
Apr 19, 2026
|
|
||
| // Use config values if provided, otherwise use sensible defaults | ||
| const timeout = contentGeneratorConfig?.timeout || 600000; // 10 minutes default | ||
| const maxRetries = contentGeneratorConfig?.maxRetries ?? 2; // 2 retries default |
Collaborator
There was a problem hiding this comment.
[Critical] fetchOptions is computed in the constructor before Agent is loaded. On the first OpenAIContentGenerator instance Agent is still undefined, so this client is created without the custom dispatcher and never picks up the new keep-alive / timeout settings that this PR is trying to add. Please load undici before new OpenAI(...), or recreate the client after Agent becomes available.
— gpt-5.4 via Qwen Code /review
|
|
||
| if (!effectiveAuthType) { | ||
| return; | ||
| } |
Collaborator
There was a problem hiding this comment.
[Critical] initializeAuth() now swallows every auth initialization failure here. That changes the sandbox startup path too, because the surrounding try/catch no longer sees the error and cannot fail fast before launching. Broken or expired credentials will now slip past startup and fail later in a much harder-to-diagnose place. Please rethrow after logging, or return an explicit failure result that callers must handle.
— gpt-5.4 via Qwen Code /review
| if (process.env.GEMINI_API_KEY) { | ||
| return AuthType.USE_GEMINI; | ||
| } | ||
| if (process.env.GOOGLE_GENAI_USE_VERTEXAI === 'true') { |
Collaborator
There was a problem hiding this comment.
[Critical] This new shared auth resolver no longer considers QWEN_OAUTH_TOKEN, even though other code paths still treat that env var as a valid signal for Qwen auth. As soon as validateNonInteractiveAuth() switches to this helper, existing token-based Qwen setups stop resolving an auth type in non-interactive/shared flows. Please preserve the previous env-based auth matrix here before making this the source of truth.
— gpt-5.4 via Qwen Code /review
DragonnZhang
pushed a commit
that referenced
this pull request
Apr 30, 2026
# v0.7.5 — Network Proxy, Webhooks & Community Fixes --- ## Features - **Network proxy support** — configure HTTP/HTTPS proxies with bypass rules directly from App Settings. The proxy engine routes traffic through `undici` ProxyAgent instances, respects `NO_PROXY` rules, and configures both Node and Electron browser sessions. (46049d28, 0ed01c48, 81d1b1c3, 0da3f265) - **Webhook actions for automations** — automations can now fire HTTP webhooks with configurable auth, form payloads, response capture, replay, and persistent retry with exponential backoff. (993c75ce, 588aab6c) - **Gemini 3.1 Flash Lite** — added to Google AI Studio preferred defaults. Thanks to [@naishyadav](https://github.com/naishyadav) for the suggestion in [#357](craft-ai-agents/craft-agents-oss#357). (e36e9a22) - **Dismiss working directory history items** — hover-visible X button on each recent directory entry to remove it from history. Thanks to [@jjjrmy](https://github.com/jjjrmy) ([#346](craft-ai-agents/craft-agents-oss#346)) and [@jonzhan](https://github.com/jonzhan) ([#391](craft-ai-agents/craft-agents-oss#391)) for requesting this. (3bd55d6d) ## Improvements - **History truncation** — consolidated all history field truncation to a single `HISTORY_FIELD_MAX_LENGTH` constant instead of scattered hardcoded limits (8689a1d4) - **Craft source docs** — added Collections section to guide with emphasis on the nested title + properties item format (0837afa0) ## Bug Fixes - **MiniMax CN authentication** — removed incorrect `minimax-cn → minimax` alias, added lightweight direct HTTP test for Pi providers, and stripped MiniMax-prefix for CN API compatibility. Thanks to [@Kathie-yu](https://github.com/Kathie-yu) ([#396](craft-ai-agents/craft-agents-oss#396)) and [@RimuruW](https://github.com/RimuruW) ([#386](craft-ai-agents/craft-agents-oss#386)) for reporting. Fixes [#396](craft-ai-agents/craft-agents-oss#396). (612c0e7e) - **Inline code in messages** — text between inline badges (sources, skills, files) now renders through the Markdown component, restoring inline code, bold, italic, and links. Thanks to [@linusrogge](https://github.com/linusrogge) for reporting [#378](craft-ai-agents/craft-agents-oss#378). Fixes [#378](craft-ai-agents/craft-agents-oss#378). (e7f88a38) - **Zod/JSON Schema passthrough** — default Zod object schemas to `.passthrough()` to match JSON Schema semantics where `additionalProperties` defaults to `true`. Also preserved `additionalProperties` in MCP proxy tool schema round-trip. Fixes tools with loosely-typed schemas silently losing fields. (cf4b6ac1, 42c173cf) - **Self-signed TLS certificates** — accept self-signed certificates for the configured remote server origin, fixing `ERR_CERT_AUTHORITY_INVALID` on `wss://` connections (7148ebec) - **File attachment in thin client mode** — paperclip button now uses browser-native `FileReader` API instead of server-side `fs.readFile()`, fixing silent failures when client and server filesystems differ (680cd197) - **URL linkification** — strip trailing markdown characters (`**`, etc.) from linkified URLs that were producing broken links (24385e78, 65b8f350) - **Automation action badges** — sidebar now shows correct "Prompt" / "Webhook" badges based on actual action types instead of hardcoding "Prompt" (dc422573) - **Packaged server path fallback** — added `dist/resources` fallback for builds where `extraResources` output layout differs (edc5f61f) - **`$CRAFT_EVENT_DATA` missing labels** — all automation event payloads now include the session's current `labels` array, so webhooks and scripts can access label data. Fixes [#406](craft-ai-agents/craft-agents-oss#406). (0f98f090) - **Multi-select non-adjacent sessions** — Cmd-click now always toggles selection (standard OS behavior); opening in a new panel moves to Cmd-Shift-click. Fixes [#404](craft-ai-agents/craft-agents-oss#404). (d4d7aff1) - **@ mention autocomplete with spaces** — spaces are now allowed in file mention queries (e.g. `@app availability.md`). The menu auto-closes Slack-style when a space produces no matches. Thanks to [@alexzadeh](https://github.com/alexzadeh) for reporting [#398](craft-ai-agents/craft-agents-oss#398). Fixes [#398](craft-ai-agents/craft-agents-oss#398). (1d063177) ---
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR improves the reliability of Qwen Code when using local OpenAI-compatible servers by implementing proper HTTP/1.1 keep-alive configuration and robust error handling.
Problem
Local OpenAI-compatible servers (running on localhost) were experiencing:
Solution
1. HTTP Agent Configuration
2. Retry Logic
undicimodule is not available3. Configuration Improvements
getBackendName()function to display actual backend name ("OpenAI API", "Qwen API", "Vertex AI", etc.) instead of undefined in error messagesChanges
config.ts, enforced CLI > env > settings.json precedenceopenai-auth.mdwith environment variable examplesTesting
To test with a local OpenAI-compatible server:
Compatibility
Fixes issues with local model inference reliability when using self-hosted models.