Repository navigation
Fixes from running a real multi-agent app through AgentFox - #77
Merged
Merged
Conversation
Ran openai/openai-cs-agents-demo (5 agents, handoffs, 11 tools, its own guardrail agents) on GPT models through the gateway with the OpenAI Agents adapter, then configured it from the dashboard as a customer would. Gateway - The OpenAI/Anthropic proxy rebuilt requests and dropped tools, tool_choice, response_format and parallel_tool_calls, and always sent temperature: client fields now pass through to a provider speaking the same protocol. Streaming forwards and records tool-call deltas. - The proxy held the request transaction (agent row lock) across the model call; preflight now commits first. The non-streaming route ran the model call on the event loop; it now runs in a worker thread. - System, developer and earlier assistant messages were scanned as user input, so a guardrail prompt about jailbreaks blocked every request at High sensitivity. App-authored roles are skipped unless marked untrusted. - A model-provider outage no longer stops /v1/guard checks. - Granting a tool registers it (risk guessed from its name, editable), so Allow no longer leaves every call held by tool.not_declared. - An identical held call reuses its pending approval; the Agents adapter redeems an approved call when the agent retries it. Dashboard - Risk is editable per tool on the agent's Access tab. - Approval limits offer the agent's own tools as one-click answers. - Approval cards show the approval id the agent quoted. - Protect wizard starts protections at their inherited level and leaves model-backed ones off when the model is not installed. Co-Authored-By: Claude <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This branch was successfully deployed
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
1 participant
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.
Ran openai/openai-cs-agents-demo (5 agents with handoffs, 11 tools incl. cancel/compensation, its own guardrail agents) live on GPT models through the gateway, governed by the OpenAI Agents adapter, and configured it from the dashboard like a customer.
Bugs it found (fixed here)
tools,tool_choice,response_format,parallel_tool_callswere never forwarded, andtemperaturewas always sent. Tool-using agents could not work through the gateway. Now passed through per protocol; streaming forwards tool-call deltas.lock_timeout) concurrent requests for the same agent failed when the model took longer than 5 s. Preflight now commits before the call; the non-streaming route no longer blocks the event loop.X-AgentFox-Trust.tool.not_declared. Granting now registers the tool with a guessed risk, editable on the Access tab.Dashboard
Editable tool risk; approval limits offer the agent's tools as answers; approval cards show the id; Protect wizard start levels.
Tests
Full backend 3462 passed; dashboard 67; ruff/format/import-linter clean; wheels rebuilt.
🤖 Generated with Claude Code