Repository navigation
Conversation
Python Preview Packages — Publish FailedPreview publish failed for commit 26cbf23. |
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
| working-directory: integrations/langroid/python | ||
| run: uv run python -m unittest discover tests -v | ||
|
|
||
| claude-managed-agents-python: |
| - name: Detect fork PR | ||
| id: fork-check | ||
| run: | | ||
| if [[ "${{ github.event_name }}" == "pull_request" && \ |
| id: fork-check | ||
| run: | | ||
| if [[ "${{ github.event_name }}" == "pull_request" && \ | ||
| "${GITHUB_EVENT_PULL_REQUEST_HEAD_REPO_FULL_NAME}" != "${{ github.repository }}" ]]; then |
Bridge Claude Managed Agents (Anthropic's hosted agent sessions) to the AG-UI protocol. Each AG-UI thread maps to one managed session; each run drives one turn of that session and translates its event stream into AG-UI events. Provided in three languages with matching behavior: - TypeScript (@ag-ui/claude-managed-agents): a ManagedAgentsAgent AbstractAgent plus an examples server. - Python (ag-ui-claude-managed-agents): a ManagedAgentsAgent yielding ag_ui.core events plus a FastAPI endpoint helper. - .NET (AGUI.ClaudeManagedAgents): a ManagedAgentsAgent on the AG-UI .NET abstractions plus a MapManagedAgentsAgent endpoint extension. Event mapping covers streamed text with preview reconciliation, reasoning start and end, built-in and MCP tool calls with results, backend custom tools with server-side handlers, and frontend tools (RunAgentInput.tools): a frontend tool call parks the session and ends the run, and the next run's tool message resumes it. Also typed session errors, tool-confirmation policy, per-thread run serialization, interrupt on disconnect or timeout, an optional scope that partitions thread state per caller, and split sends so tool results resume a parked session before follow-up user messages. Registers all three variants in the Dojo (menu, agents, env, generated files, prep and run scripts), the release config and workflow scope dropdowns, a Python unit-test job, a .NET build-and-test workflow, the config allowlist, the Render blueprint, and the README framework table.
cj-ant
force-pushed
the
cj/claude-managed-agents-integration
branch
from
July 23, 2026 18:56
a1db8d7 to
773f847
Compare
Extract shared helpers in the TypeScript turn (text-message emits, answering a custom tool use) and a runError builder plus a hasUserText predicate in the agent, removing repeated event casts and the duplicated tool-result posting path. Name the tool-name, tool description, tool-result, search-preview, and parked-retry constants in place of magic numbers, and mirror the same named constants in the Python core. No behavior change.
Bring the TypeScript and .NET suites up to parity with Python (73/76/78 tests) and fix the issues review turned up across the claude-managed-agents integration: - Key the per-thread run gate by session-store identity: agents sharing a store serialize runs per thread (even across instances), while per-caller stores keep one caller's runs from blocking or interfering with another's. - Bound best-effort sends (tool results, confirmations, interrupts) by a 15s timeout instead of unbounded non-cancellable calls, and post the interrupt before releasing the busy gate so an immediate resend is not killed by a stale interrupt. - Answer a backend tool call on every teardown path: race the handler against the abort signal, branch on real cancellation state instead of exception type, and observe abandoned handler tasks so late faults are never unobserved. - Give every RUN_ERROR a stable code (run_in_progress, empty_run, tool_result_without_session, nothing_to_send, turn_timeout, run_failed) in all three ports. - Converge port behavior: original-order abandoned-tool results, no empty text deltas, error text included in tool results, last-write- wins tool name collisions, shared named constants, and rename the option to managedAgentId/managed_agent_id/ManagedAgentId so it no longer shadows the AG-UI agentId. - Raise the Python floor to anthropic>=0.119.0 (the sessions API surface the code uses), regenerate both uv.lock files against public PyPI, move @anthropic-ai/sdk to peerDependencies to match sibling integrations, and drop unused uvicorn/pydantic deps. - Remove the scope/ownerId option: thread state is keyed by thread ID like the other integrations, with a security section in each README covering authentication and per-caller store partitioning. - Fix the README demo link, align the dotnet unit workflow naming, add the llmock fallback to the dotnet dojo run entry, and regenerate files.json.
Owner
Author
|
Superseded by the upstream PR: ag-ui-protocol#2247. Closing this fork-internal review PR. |
mme
pushed a commit
that referenced
this pull request
Jul 29, 2026
CPK-7718 #1: crewai 1.0.0 deleted crewai.utilities.events (event bus, flow/ method lifecycle events, listener base moved to crewai.events; BaseEvent is no longer re-exported at the root and lives at <events>.base_events). New _capabilities.py resolves each symbol trying the 1.x home first, then the 0.x home, runs the probe ONCE at import, caches it, and warns (with the version that unlocks a gap) on any missing capability. crewai.__version__ is used only for warning text, never as a code-path gate. events.py / sdk.py now import BaseEvent / crewai_event_bus from the probe. Also fixes a pre-existing E713 (not-in) in sdk.py.
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
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.
What this adds
A new integration,
@ag-ui/claude-managed-agents, that bridges Claude Managed Agents (Anthropic's hosted agent sessions) to the AG-UI protocol. It sits next to the existingclaude-agent-sdkintegration, which wraps the local Claude Code SDK. This one wraps the hosted sessions API, so an app that runs its agents on Managed Agents gets an AG-UI endpoint with no extra glue.ManagedAgentsAgentextendsAbstractAgent. Each AG-UI thread maps to one managed session, and each run drives one turn of that session while translating its event stream into AG-UI events.Event mapping
agent.messagewithevent_deltapreviewsTEXT_MESSAGE_START/CONTENT/END(previews reconciled against the buffered message)agent.thinkingREASONING_START/REASONING_ENDagent.tool_use,agent.mcp_tool_useand resultsTOOL_CALL_*andTOOL_CALL_RESULT(server executed, display only)agent.custom_tool_usefor a frontend toolTOOL_CALL_*, then the run finishes so the client can execute itagent.custom_tool_usefor a backend toolTOOL_CALL_*andTOOL_CALL_RESULT, result posted back to the sessionsession.error(terminal)RUN_ERRORwith the error type ascodesession.status_idlewithend_turnRUN_FINISHEDHuman in the loop with frontend tools
Tools in
RunAgentInput.toolsare registered on the session as custom tools. When the agent calls one, the run emits the tool call and finishes while the session parks inrequires_action. The client runs the tool and starts the next run with arole: "tool"message. The adapter forwards it asuser.custom_tool_resulton the same session and streams the continuation. Backend tools (server side handlers) run inline and post their result back.What is in the PR
integrations/claude-managed-agents/typescript: the package (ManagedAgentsAgent, turn translator, pluggable session store, tool helpers), 27 unit tests against a scripted fake client, and an examples server plus an idempotent provisioning script (setup:examples) that creates one agent per Dojo feature.menu.ts,agents.ts,env.ts(CLAUDE_MANAGED_AGENTS_URL, port 8024), regeneratedfiles.json, prep and run scripts.release.config.jsonscope,nx.json, the three workflow scope dropdowns, and the config allowlist entry fortsdown.config.ts.render.yamlservice (pinned to one instance since the default store is in memory) and the framework table row in the rootREADME.md.Verified
Provisioned the example agents in a test workspace and drove the example server end to end:
RUN_FINISHED.TOOL_CALL_*, thenTOOL_CALL_RESULTfrom the server handler, then the agent's summary.TOOL_CALL_*, run 2 posts the tool result to the same session and it resumes toend_turn.RUN_ERROR, no session created), overlapping runs on one thread (second rejected), client disconnect (auser.interruptreaches the session).Notes for review
llmock, but here the model loop runs at Anthropic, so there is no local model call to intercept. Open to ideas on how the maintainers would like this covered.SessionStoreis in memory. It is pluggable so a deployment can persist thread to session mappings.