fix: add server_tool_use content blocks and tests for OpenAI native web search

[noanflaherty]

Summary

Fixes two gaps identified during plan review for managed-openai-native-web-search.md.

Gap 1: Missing test coverage for web_search_call event handling
Gap 2: No server_tool_use content blocks in ProviderResponse.content for conversation history parity with Anthropic

---

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f428c06d06

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Emit paired web-search result with server tool use

This now persists server_tool_use blocks for OpenAI web-search calls but never emits a matching web_search_tool_result block, which means every such message is treated as orphaned server-tool output by repairHistory and gets a synthetic web_search_tool_result_error injected on subsequent turns (via the pre-run repair path in conversation-agent-loop.ts). In practice, successful native web searches are rewritten as synthetic errors in runtime history, adding noisy repairs and polluting follow-up context; either emit a paired result block or avoid adding server_tool_use until a pair can be represented.

Useful? React with 👍 / 👎.

[devin-ai-integration]

Devin Review found 1 potential issue.

View 3 additional findings in Devin Review.

[devin-ai-integration]

🔴 Missing web_search_tool_result blocks causes history repair to inject synthetic error results on every subsequent turn

The new code injects server_tool_use content blocks into the response (lines 305-311) but never produces the paired web_search_tool_result content blocks that the rest of the codebase expects. The repairHistory function (assistant/src/daemon/history-repair.ts:71-113) runs before every provider API call (assistant/src/daemon/conversation-agent-loop.ts:1023) and explicitly checks that every server_tool_use in an assistant message has a matching web_search_tool_result in the same message. When the pairing is missing, it injects a synthetic error result with {type: "web_search_tool_result_error", error_code: "unavailable"}, logs a warning, and increments stats.missingToolResultsInserted. This means every successful native web search through OpenAI will be retroactively marked as failed in the conversation history on the very next turn. The Anthropic provider produces both block types (see assistant/src/providers/anthropic/client.ts:1386-1393), and the comment on line 302 even says the goal is to match the Anthropic shape — but the pairing is incomplete.

Expected fix pattern

After each server_tool_use block, also push a web_search_tool_result block:

for (const toolUseId of webSearchCallIds) {
  content.push({
    type: "server_tool_use",
    id: toolUseId,
    name: "web_search",
    input: {},
  });
  content.push({
    type: "web_search_tool_result",
    tool_use_id: toolUseId,
    content: [],
  });
}

Suggested change

	for (const toolUseId of webSearchCallIds) {
	content.push({
	type: "server_tool_use",
	id: toolUseId,
	name: "web_search",
	input: {},
	});
	}
	for (const toolUseId of webSearchCallIds) {
	content.push({
	type: "server_tool_use",
	id: toolUseId,
	name: "web_search",
	input: {},
	});
	content.push({
	type: "web_search_tool_result",
	tool_use_id: toolUseId,
	content: [],
	});
	}

---

Was this helpful? React with 👍 or 👎 to provide feedback.