feat(openai-temporal): render hosted/server-side tool calls in TemporalStreamingModel

[declan-scale]

Summary

Stacked on #436. Renders Responses-API hosted/server-side tool calls (web_search, file_search, code_interpreter, image generation, server-side mcp, computer_call, local_shell_call) as ToolRequestContent + ToolResponseContent pairs in TemporalStreamingModel.

These tools execute inside the Responses API, so they never become function_call items and the SDK RunHooks (on_tool_start/on_tool_end) never fire for them. Previously they produced no UI event at all. Each completed hosted-tool output item is now surfaced as a request/response pair, mirroring how function_call tools already render.

This upstreams the hosted-tool half of golden_agent's vendored RichToolStreamingModel so that agent can stop vendoring (its "Phase 4"). The reasoning double-render fix the vendored model also carried already landed upstream (a5335983, fixed at the streaming-layer root cause), so only hosted-tool handling is ported here.

Changes

• New module helpers _HOSTED_TOOL_TYPES, _coerce_args, _hosted_tool_request, _hosted_tool_result.
• New _post_tool_message method (one-shot full message, no deltas).
• New elif item.type in _HOSTED_TOOL_TYPES branch in the ResponseOutputItemDoneEvent handler.
• Purely additive: existing function_call, reasoning, and text handling are unchanged.

Test plan

• 8 new unit tests (tests/test_hosted_tools.py) covering the extraction helpers.
• Existing openai_agents plugin suite (45 tests) stays green; ruff clean.
• Live model-agnostic turn exercising a hosted tool (e.g. web_search / server-side mcp) renders a request+response pair in the UI.

Note: base is declan-scale/init-templates-codex (#436 head) — review/merge #436 first.

🤖 Generated with Claude Code

Greptile Summary

This PR adds rendering of Responses-API hosted/server-side tool calls (web_search, file_search, code_interpreter, image_generation, mcp, computer_call, local_shell_call) as ToolRequestContent+ToolResponseContent pairs in TemporalStreamingModel. Previously, these tools executed silently inside the API with no UI event.

• New module-level helpers (_HOSTED_TOOL_TYPES, _coerce_args, _hosted_tool_request, _hosted_tool_result) extract call ID, display name, args, and a capped result string from completed hosted-tool SDK items.
• A new _post_tool_message one-shot method posts both halves of the request/response pair using the same streaming_task_message_context START→FULL pattern used by the existing function_call handler.
• An elif branch in the ResponseOutputItemDoneEvent handler triggers the new path without touching any existing function_call, reasoning, or text handling.

Confidence Score: 5/5

Purely additive change with no modifications to existing function_call, reasoning, or text paths; safe to merge.

All new code follows the established START→FULL streaming pattern used by the function_call handler. The elif branch is correctly scoped, helper functions are well-tested with real SDK types, and the hosted-tool items naturally flow through the existing final response processing via ResponseCompletedEvent. No changes to existing streaming paths.

No files require special attention.

Important Files Changed

Filename	Overview
src/agentex/lib/core/temporal/plugins/openai_agents/models/temporal_streaming_model.py	Adds ~100 lines of purely additive hosted-tool helpers and one new elif branch in the event loop; existing function_call, reasoning, and text paths are unchanged. Logic is consistent with the established streaming pattern (START→FULL via streaming_task_message_context). No correctness issues found.
src/agentex/lib/core/temporal/plugins/openai_agents/tests/test_hosted_tools.py	8 focused unit tests covering all extraction helpers; key SDK types (ResponseFunctionWebSearch, ResponseComputerToolCall, LocalShellCall, ImageGenerationCall) are now used as real types rather than SimpleNamespace stand-ins, closing the gap flagged in a prior review round.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant API as Responses API
    participant TSM as TemporalStreamingModel
    participant PM as _post_tool_message
    participant SC as StreamingTaskMessageContext
    participant UI as UI / Redis Stream

    API->>TSM: ResponseOutputItemDoneEvent (hosted tool)
    TSM->>TSM: _hosted_tool_request(item) → call_id, name, args
    TSM->>TSM: _hosted_tool_result(item) → result string

    TSM->>PM: _post_tool_message(ToolRequestContent)
    PM->>SC: async with streaming_task_message_context(initial_content)
    SC->>UI: START event (ToolRequestContent)
    PM->>SC: stream_update(StreamTaskMessageFull)
    SC->>UI: FULL event (ToolRequestContent)
    SC-->>PM: __aexit__ → close() early-returns (already closed)

    TSM->>PM: _post_tool_message(ToolResponseContent)
    PM->>SC: async with streaming_task_message_context(initial_content)
    SC->>UI: START event (ToolResponseContent)
    PM->>SC: stream_update(StreamTaskMessageFull)
    SC->>UI: FULL event (ToolResponseContent)
    SC-->>PM: __aexit__ → close() early-returns (already closed)

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant API as Responses API
    participant TSM as TemporalStreamingModel
    participant PM as _post_tool_message
    participant SC as StreamingTaskMessageContext
    participant UI as UI / Redis Stream

    API->>TSM: ResponseOutputItemDoneEvent (hosted tool)
    TSM->>TSM: _hosted_tool_request(item) → call_id, name, args
    TSM->>TSM: _hosted_tool_result(item) → result string

    TSM->>PM: _post_tool_message(ToolRequestContent)
    PM->>SC: async with streaming_task_message_context(initial_content)
    SC->>UI: START event (ToolRequestContent)
    PM->>SC: stream_update(StreamTaskMessageFull)
    SC->>UI: FULL event (ToolRequestContent)
    SC-->>PM: __aexit__ → close() early-returns (already closed)

    TSM->>PM: _post_tool_message(ToolResponseContent)
    PM->>SC: async with streaming_task_message_context(initial_content)
    SC->>UI: START event (ToolResponseContent)
    PM->>SC: stream_update(StreamTaskMessageFull)
    SC->>UI: FULL event (ToolResponseContent)
    SC-->>PM: __aexit__ → close() early-returns (already closed)

Loading

_{Reviews (4): Last reviewed commit: "fix(openai-temporal): render args for co..." | Re-trigger Greptile}