fix(sdk): strip leading 思考 block from LLM-generated title

[santhiprakash]

HUMAN:

Santhi Prakash — small correctness fix for the SDK title-leak bug class cross-referenced from #4541; see AGENT section for evidence.

---

AGENT:

Why

Some providers (notably Qwen3 behind Nebius) return reasoning inline
in content as 思考...思考 instead of the normalized
reasoning_content field. The transparent leak surfaced in Agent
Canvas 1.14.0 (software-agent-sdk#4541)
as the autogenerated conversation title containing the raw 思考
reasoning markup — the SDK ships the reasoning into the title, and Agent
Canvas just renders it.

This is the SDK-side root-cause fix: peeling the leading 思考 block in
the SDK means every consumer of Conversation.generate_title() and the
agent-server auto-title path benefits, not just Agent Canvas.

Summary

• generate_title_with_llm() now peels a leading 思考 block from the
  LLM title response before using it as the title.
• A response that is only a 思考 block (closed or unclosed) yields
  no LLM title and falls back to the existing user-message truncation.
• Mid-text literal 思考 occurrences in a legitimate title are
  preserved.

Issue Number

Fixes #4530

How to Test

uv run pytest tests/sdk/conversation/test_generate_title.py -q

Expected: 12 passed (4 new regression tests + 8 existing).

The 4 new tests cover:

1. test_generate_title_strips_leading_think_block — closed
   思考…思考 block is peeled, the real title after the block is returned.
2. test_generate_title_unclosed_think_block_falls_back — unterminated
   leading 思考… returns no LLM title; truncation fallback supplies the
   title.
3. test_generate_title_think_only_response_falls_back — response that
   is only a 思考 block yields no LLM title; truncation fallback
   supplies the title.
4. test_generate_title_preserves_mid_text_think_literal — a mid-text
   literal 思考 occurrence in a legitimate title is preserved.

I verified the regression coverage by reverting
generate_title_with_llm() to its pre-fix behavior and re-running the
new tests: 3/4 fail on the unfixed code (the literal-preservation test
correctly passes either way), proving the tests do catch the bug class.

$ uv run pytest tests/sdk/conversation/test_generate_title.py -q
collected 12 items
tests/sdk/conversation/test_generate_title.py ............               [100%]
============================== 12 passed in 0.38s ==============================

Type

• Bug fix
• Feature
• Refactor
• Breaking change
• Docs / chore

Notes

• The fix is intentionally scoped to the SDK-side title generation path;
  it does not touch the Agent Canvas frontend rendering issues (plaintext
  thoughts / duplicate thought block) covered separately in
  [Bug]: Weird <think> behavior with Qwen3-32B #4541.
• The peel is a leading-block conservative peel — only the first leading
  block is stripped, so a legitimate title that happens to contain the
  characters 思考 mid-text is preserved.
• The unterminated-leading case consumes the entire response rather than
  returning the raw reasoning string, so the LLM path surfaces no title
  for the truncation fallback to take over.

[all-hands-bot]

👋 This PR needs a couple of things fixed before OpenHands can review it:

• the PR description's HUMAN: section needs at least 20 characters describing what you tested, not just the template placeholder

Push an update once this is addressed and this check re-runs automatically.

This is an automated check - no AI was used to generate this comment.

[all-hands-bot]

🚦 CI is currently failing on this PR's latest commit.

Please fix the failing checks before OpenHands reviews it - this is re-checked automatically once you push a new commit. (A maintainer can also request @all-hands-bot as a reviewer to have it reviewed regardless of CI status.)

This is an automated check - no AI was used to generate this comment.

[github-actions]

📁 PR Artifacts Notice

This PR contains a .pr/ directory with temporary PR-specific documents. Because this is a fork PR, the directory will be automatically removed from main immediately after merge.

[santhiprakash]

Rebased onto current main (1de2e6d, which restructured generate_title_with_llm with the new on_error callback in #4535) — the conflict was only in the import block; the think-block peel now sits inside the updated signature alongside the on_error plumbing.

Verified locally on the rebased head:

• uv run python -m pytest tests/sdk/conversation/test_generate_title.py -q → 13 passed (incl. the new on_error test from fix(agent-server): propagate out-of-band run failures as ConversationErrorEvent (#16686) #4535 and the three 思考-block tests from this PR)
• uv run python -m pytest tests/agent_server/test_conversation_service.py -q → 109 passed
• ruff check / ruff format --check clean

Heads-up: every workflow run on this PR (including the original Aug 18 head) has been sitting in action_required — "awaiting workflow approval" — so CI has never actually executed. The bot's "CI failing" notice reflects that gate, not a test failure. Could a maintainer approve the workflow runs so the suite can run?

[neubig]

I think this is probably a fix on the litellm side. The "think" blocks should be parsed out into the reasoning_content response. I don't think this is something that we should be making heuristic changes to here.