fix(agent): run response hooks before send in skills_like no-tool fallback

[ryeeda]

Fixes #9788

Modifications / 改动点

• This is NOT a breaking change. / 这不是一个破坏性变更。
• 修复非流式 skills_like 工具重查询无工具时的 hook/发送顺序。
• 在 fallback 发送链构造前调用 _complete_with_assistant_response()，确保 OnLLMResponseEvent 与 OnAgentDoneEvent 能先审查或改写最终响应。
• 发送链直接读取 hook 后的 LLMResponse，不复用 hook 前构造的消息组件。
• 保持正常工具调用、普通无工具终局、异常路径和 streaming/live 路径不变。
• 新增端到端契约测试 tests/test_skills_like_hook_order.py（fake provider 驱动真实 runner、scheduler 与 respond stage）。
• 没有新增依赖。

Why / 原因

旧顺序是 yield -> downstream send -> response hooks。这使所有依赖 response hook 做内容安全、脱敏、改写或审计的插件在该 fallback 分支失去发送前保障，也可能造成已发送文本与最终 history 不一致。

新顺序是 response hooks -> build chain from finalized response -> yield -> downstream send -> after-message hook -> history save。与正常无工具终局路径（先 _complete_with_assistant_response 再 yield）保持一致。

Verification Steps / 验证步骤

新增端到端契约测试以 fake provider 驱动真实 runner、scheduler 和 respond stage，验证：

1. skills_like 无工具重查询时，response hook 改写后的文本才被发送；
2. after-message hook 观察到的文本与发送文本一致；
3. history 尾项与发送文本逐字相等；
4. 普通无工具终局、正常工具终局、多轮工具和 provider 异常路径不回归。

验证基于最新 master（19d00fb）：

• 未打补丁：核心用例失败，复现 issue（1 failed, 4 passed）；
• 打补丁后：5 passed；
• 相关回归 tests/test_astr_agent_run_util.py + tests/agent/：99 passed。

Screenshots or Test Results / 运行截图或测试结果

tests/test_skills_like_hook_order.py: 5 passed, 1 warning in 3.70s
tests/test_astr_agent_run_util.py + tests/agent/: 99 passed, 1 warning in 3.53s

warning 为既有 audioop 弃用提示，与本次修改无关。

Non-goals / 不在本 PR 范围

本 PR 不调整 streaming/live 发送语义。流式响应的逐块发送与终局 hook 改写之间需要单独的 API/生命周期设计。

Related / 相关链接

• Fixes [Bug] skills_like 无工具重查询退化分支在 response hook 前发送响应 #9788
• 相关前史：skills_like 模式下用户收到的消息与 conversation history 中存储的消息不一致 #8238 / fix: skills_like 模式下保留原始 completion_text，修复上下文分裂问题 #8240（re-query 主路径的文本-history 分裂修复，未覆盖 fallback 分支）、feat(agent-runner): add tool_choice parameter to fix empty tool calls response in "skills-like" tool call mode #7101（skills_like re-query 机制引入）。
• 代码邻近：与 open PR fix(agent): preserve tool metadata after skills-like requery #9220 修改同一函数相邻区块，改动不冲突，但合并顺序靠后时可能需要 rebase。

Checklist / 检查清单

• 👀 更改已通过端到端契约和相关回归测试，并提供了验证步骤及结果。
• 🤓 未引入新的依赖库。
• 😮 更改不包含恶意代码。
• 本 PR 不添加新功能，因此“新增功能已先讨论”项不适用（对应 bug 已先在 issue 中报告）。

Summary by Sourcery

Run response hooks before finalizing and sending skills-like fallback responses while preserving correct error handling and history consistency.

Bug Fixes:

• Ensure response hooks run before sending the non-streaming skills-like no-tool fallback response, keeping sent content consistent with after-send processing and persisted history.
• Route error responses from skills-like re-queries through error handling instead of treating them as assistant content.

Tests:

• Add end-to-end lifecycle contract tests covering hook ordering, response consistency, normal tool flows, multi-step tool flows, and provider errors.

[sourcery-ai]

Hey - I've found 1 issue

Prompt for AI Agents

Please address the comments from this code review:

## Individual Comments

### Comment 1
<location path="astrbot/core/agent/runners/tool_loop_agent_runner.py" line_range="941" />
<code_context>
                     )
+                    # Finalize before yielding so response hooks can review or rewrite
+                    # the fallback text before downstream stages send it.
+                    await self._complete_with_assistant_response(llm_resp)
                     if llm_resp.reasoning_content:
                         yield AgentResponse(
</code_context>
<issue_to_address>
**issue (bug_risk):** When the skills-like re-query returns an `LLMResponse` with `role == "err"` and no tool calls, this newly added call treats the error as a successful assistant response: it transitions the runner to `DONE`, appends the error text to assistant history, and invokes response-completion hooks. The earlier error branch at line 883 is bypassed because the re-query occurs after that branch, so provider errors on this path no longer follow the normal error handling and history contract.

**Triggers:** When the skills-like re-query provider returns an error response rather than raising an exception.

**Suggested fix:** Check `llm_resp.role == "err"` before finalizing the fallback, and route it through the existing error-state/error-response handling instead of `_complete_with_assistant_response`.
</issue_to_address>

Sourcery assessment

Needs a human reviewer. 1 finding to address first, and if the hook-order change is wrong, a fallback response could be rewritten incorrectly and sent to the user, with the same incorrect text persisted in conversation history. Reverting prevents future occurrences, but it cannot retract messages already sent; the affected history is bounded and repairable.

Blocking findings: astrbot/core/agent/runners/tool_loop_agent_runner.py:941

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}

[ryeeda]

@sourcery-ai review

[sourcery-ai]

Hey - I've found 1 issue

Prompt for AI Agents

Please address the comments from this code review:

## Individual Comments

### Comment 1
<location path="astrbot/core/agent/runners/tool_loop_agent_runner.py" line_range="958-960" />
<code_context>
+                            ),
+                        )
+                        return
+                    # Finalize before yielding so response hooks can review or rewrite
+                    # the fallback text before downstream stages send it.
+                    await self._complete_with_assistant_response(llm_resp)
                     if llm_resp.reasoning_content:
                         yield AgentResponse(
</code_context>
<issue_to_address>
**issue (broader_impact):** When an `OnLLMResponseEvent` hook rewrites `llm_resp` without also mutating `run_context.messages`, `_complete_with_assistant_response()` saves the original completion to history before invoking the hook, while the fallback sends the rewritten completion afterward. The delivered text and persisted assistant history therefore diverge in this newly changed fallback path.

**Triggers:** When a response hook performs content rewriting, redaction, or safety filtering by changing only the `LLMResponse`.

**Suggested fix:** Run the response hooks before constructing the assistant history message, or synchronize the appended history message from the finalized `LLMResponse` after the hooks complete.
</issue_to_address>

Sourcery assessment

Needs a human reviewer. 1 finding to address first, and the fallback response is now finalized before it is yielded, allowing response and completion hooks to rewrite the text that is sent and saved to conversation history. If that ordering is wrong, a bad assistant message can be delivered externally and persisted; reverting the change cannot unsend it or automatically remove the saved history.

Blocking findings: astrbot/core/agent/runners/tool_loop_agent_runner.py:960

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}