fix(agent): deliver skills-like fallback replies when streaming

[camera-2018]

Fixes #10024

Motivation

With skills_like and streaming enabled, a tool-schema re-query can return a plain assistant reply instead of another tool call. The re-query uses non-streaming text_chat(), but this fallback only emits llm_result. The streaming bridge consumes streaming_delta, so the final reply is saved to history without reaching the user.

Modifications

• In the no-tool re-query fallback only, emit the reply as streaming_delta when streaming is enabled, while retaining llm_result for non-streaming and stream-to-general consumers.
• Append the new streaming events after the existing _complete_with_assistant_response() call, so they use the post-hook reply. Do not move the existing llm_result emissions or hook invocation.
• Preserve reasoning as a separate chain so the bridge's reasoning visibility setting still applies.
• Leave the generic stream bridge unchanged: forwarding every llm_result there would duplicate ordinary streamed replies.

No new settings, dependencies, platform-specific changes, or changes to cron notification policy.

Scope: only the missing streaming events in #10024. The existing non-streaming/stream-to-general llm_result-before-hook behavior is preserved. The separate hook-ordering issue #9788 is left to #9789 / #9835; this PR no longer changes that ordering. The production diff only adds 20 lines in the fallback branch.

• This is NOT a breaking change.

Verification Steps and Test Results

Based on upstream master at 6fb50cae9.

Added regressions exercise the real ToolLoopAgentRunner, its schema re-query, and run_agent() with a deterministic fake provider/event. Coverage includes:

• Plain completion_text and result_chain fallback replies.
• Streaming with reasoning hidden/shown, non-streaming, and stream-to-general delivery.
• Exactly-once final reply and intermediate text delivery.
• Original complete-result events remaining before hooks, and only the newly added streaming deltas being emitted after hooks.
• Retained complete-result events and final history content.
• Ordinary streaming replies remaining single-delivery.
• Existing successful re-query/tool execution coverage extended to streaming mode.

Before the fix, the new regression matrix fails (including missing final replies on the streaming cases); the ordinary streaming control passes. After the fix:

python -m pytest tests/test_tool_loop_agent_runner.py tests/test_astr_agent_run_util.py \
  tests/unit/test_astr_main_agent.py tests/unit/test_astr_agent_tool_exec.py \
  tests/unit/test_agent_runner_config.py tests/agent -q

317 passed, 1 warning

The warning is the existing Python audioop deprecation warning.

ruff format .       # Ruff 0.15.22: 504 files left unchanged
ruff check .        # All checks passed
git diff --check    # clean

Tests are local and deterministic; no real provider calls or platform messages were sent, and the production deployment was not modified. This is not a live Feishu end-to-end test.

Checklist

• Changes are tested, with verification commands and results above.
• No new dependencies are introduced.
• No malicious code is introduced.

Summary by Sourcery

Ensure assistant replies produced by skills-like tool-schema re-queries reach streaming consumers without duplication.

Bug Fixes:

• Deliver fallback assistant replies from skills-like streaming re-queries to users exactly once.
• Preserve reasoning visibility and ensure agent-done hooks run before fallback stream content is emitted.

Enhancements:

• Extend regression coverage across streaming, reasoning, result-chain, and stream-to-general delivery paths while confirming ordinary streaming replies are not duplicated.

Tests:

• Add deterministic integration-style coverage for the real tool-loop runner, streaming bridge, hook ordering, response persistence, and final history content.

[sourcery-ai]

Hey - I've reviewed your changes and they look great!

Sourcery assessment

Approved.

---

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

• X
• Mastodon
• LinkedIn
• Facebook

[Copilot]

🟢 Approval recommended

The change is tightly scoped, aligns with the stated streaming bridge semantics, and is covered by targeted regression tests that exercise the real runner/bridge behavior.

Pull request overview

This PR fixes a streaming delivery gap in the skills_like tool-schema re-query fallback path: when the re-query returns a plain assistant reply (no tool calls), the final answer previously only emitted llm_result and could be persisted to history without being delivered to streaming consumers.

Changes:

• Emit the skills_like no-tool re-query fallback reply as streaming_delta (when streaming is enabled) while still emitting llm_result for non-streaming / stream-to-general consumers.
• Ensure _complete_with_assistant_response() (and thus agent-done hooks) runs before emitting the fallback content.
• Add regression tests covering streaming vs non-streaming, reasoning visibility, stream-to-general behavior, exactly-once delivery, and “normal streaming reply not duplicated” controls.

File summaries

File	Description
astrbot/core/agent/runners/tool_loop_agent_runner.py	Moves fallback completion before emission and emits fallback content via streaming_delta in streaming mode (plus llm_result for non-streaming consumers).
tests/test_tool_loop_agent_runner.py	Adds/extends tests to validate fallback delivery and hook ordering across streaming modes and to ensure ordinary streaming replies are not duplicated.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.