fix(deepseek-web): SSE thinking/search routing and session lifecycle

[ovehbe]

Summary

• Fix DeepSeek web SSE parsing for thinking models using Chat2API-style currentPath routing so reasoning stays in reasoning_content instead of leaking into chat content
• Fix web search streaming: keep reading past FINISHED, collect search_results, convert [citation:N] tags, and append citation links at stream end
• Create a fresh DeepSeek chat session per request and delete it after completion to avoid stale-session failures when chats are removed from the DeepSeek web UI
• Expand unit tests for thinking separation, search citations, and session lifecycle

Related Issues

• Related to DeepSeek web provider follow-up fixes after v3.8.2 merge regressions

Validation

• node --import tsx/esm --test tests/unit/deepseek-web.test.ts (29 tests passing)
• npm run lint
• npm run test:unit
• npm run test:coverage

Tests Added Or Updated

• tests/unit/deepseek-web.test.ts

Coverage Notes

• Changes are covered by updated DeepSeek web executor unit tests, including thinking routing, search citation formatting, and session create/delete behavior

Reviewer Notes

• Session cleanup is always-on (fresh session per request, delete after stream ends) — no dashboard toggle included
• Native tool calling parity with Chat2API is intentionally out of scope for this PR and can be a follow-up

[kilo-code-bot]

Kilo Code Review could not run — your account is out of credits.

Add credits or switch to a free model to enable reviews on this change.

[gemini-code-assist]

Code Review

This pull request refactors the DeepSeek Web executor to support reasoning content, search citations, and improved session management by creating fresh sessions with explicit cleanup. Review feedback identifies a regression in the prompt builder that omits conversation history, logic errors in fragment handling that prevent correct path updates for empty content, and a bug where citation index filtering could exclude valid zero-indexed results.

[gemini-code-assist]

The new messagesToPrompt implementation introduces a regression by dropping the conversation history. It only extracts the system prompts and the last user message, ignoring all intermediate assistant responses and previous user messages. This breaks multi-turn chat functionality. Since a fresh session is created for every request (line 716), the full conversation context must be reconstructed in the prompt string to maintain continuity.

function messagesToPrompt(messages: Array<{ role: string; content: string }>): string {
  const extractText = (content: unknown): string => {
    if (Array.isArray(content)) {
      return (content as any[])
        .filter((item: any) => item.type === "text")
        .map((item: any) => item.text)
        .join("\n");
    }
    return String(content || "");
  };

  if (messages.length === 0) return "";

  return messages
    .map((m) => {
      const role = m.role.toLowerCase();
      const text = extractText(m.content).trim();
      if (!text) return "";
      if (role === "system") return `[System]: ${text}`;
      if (role === "assistant") return `[Assistant]: ${text}`;
      return text;
    })
    .filter(Boolean)
    .join("\n\n")
    .replace(/!\[.*?\]\(.*?\)/g, "");
}

[gemini-code-assist]

The handleFragment function contains redundant logic and a bug. When setPathFromType is false (default for initial fragments), the function returns early if content is empty (line 251), skipping the path update logic (lines 252-256). This prevents currentPath from being set correctly if the type-defining fragment has no content. The logic can be simplified to always update the path via applyFragmentType and only send content if present.

      const handleFragment = (frag: any, _setPathFromType = false) => {
        applyFragmentType(frag);
        if (typeof frag?.content === "string" && frag.content.length > 0) {
          sendByPath(frag.content);
        }
      };

[gemini-code-assist]

This handleFragment implementation in collectSSEContent shares the same bug and redundancy as the one in transformSSE. It should be simplified to ensure the path is updated even when the fragment content is empty.

  const handleFragment = (frag: any, _setPathFromType = false) => {
    applyFragmentType(frag);
    if (typeof frag?.content === "string" && frag.content.length > 0) {
      appendByPath(frag.content);
    }
  };

[gemini-code-assist]

The filter r.cite_index will exclude citations with an index of 0 because 0 is falsy in JavaScript. While citation indices are typically 1-based, it is safer to explicitly check for the existence of the property or its type to avoid missing valid data.

Suggested change

	.filter((r) => r.cite_index)
	.filter((r) => typeof r.cite_index === "number")

[diegosouzapw]

🎉 Thank you @ovehbe! Your comprehensive DeepSeek Web SSE overhaul has been merged into release/v3.8.3. The thinking/search routing, search result footnotes, and session lifecycle improvements are a major step forward! Conflicts with the release branch were resolved on your branch before merging. 🚀