Please address the comments from this code review:

## Overall Comments
- In `step`, `_simple_print_message_role` now always logs `self.run_context.messages`, so the `[BefCompact]` and `[AftCompact]` logs will be identical; consider passing/using the provider-processed `_provider_messages` for the post-compaction log so the debug output actually reflects the transformation.
- The tool call filtering in `step` builds `valid_indices` from `zip(llm_resp.tools_call_name, llm_resp.tools_call_ids)` but then indexes into `llm_resp.tools_call_args` using those indices; if `tools_call_args` is shorter than `tools_call_name`/`tools_call_ids`, this will raise `IndexError`, so it would be safer to derive indices from zipping all three lists or clamp to the minimum length.
- The new `_should_fix_modalities_for_provider` only treats a non-empty `list` as a configured modalities value, whereas previously any truthy value (including non-lists) triggered sanitization; if there are providers with non-list `modalities` configs, this silently changes behavior, so consider either validating/normalizing `modalities` to a list or preserving the old truthiness semantics.

## Individual Comments

### Comment 1
<location path="astrbot/core/agent/runners/tool_loop_agent_runner.py" line_range="596-598" />
<code_context>
         log_context_sanitize_stats(stats)
         return sanitized_contexts

+    def _should_fix_modalities_for_provider(self) -> bool:
+        modalities = self.provider.provider_config.get("modalities", None)
+        return (
+            isinstance(modalities, list) and modalities
+        )  # Empty list is treated as unconfigured
</code_context>
<issue_to_address>
**suggestion (bug_risk):** The new modalities check only supports non-empty `list` values and will skip sanitization for other sequence types.

This narrows the definition of “configured” from any truthy value to only a non-empty `list`. If a provider config uses another sequence type (e.g. a tuple), sanitization will now silently stop applying. If that’s intended, consider normalizing `modalities` to a list earlier; otherwise, widen the check to accept other iterable sequence types (or coerce to `list`) so the previous behavior is preserved while still treating `None`/`[]` as unconfigured.

Suggested implementation:

```python
    def _should_fix_modalities_for_provider(self) -> bool:
        modalities = self.provider.provider_config.get("modalities", None)

        # Treat None and empty sequences (e.g. [], (), etc.) as "unconfigured".
        # For non-sequence types, preserve the previous "truthy means configured" behavior.
        if modalities is None:
            return False

        try:
            from collections.abc import Sequence  # local import to avoid module-level changes if not already present
        except ImportError:  # very defensive; Sequence should be available in supported Python versions
            return bool(modalities)

        if isinstance(modalities, Sequence) and not isinstance(modalities, (str, bytes)):
            return len(modalities) > 0

        return bool(modalities)

```

If `Sequence` is already imported at the module level (e.g. `from collections.abc import Sequence`), you should remove the local import inside `_should_fix_modalities_for_provider` and rely on the existing import instead to keep imports centralized and avoid redundant imports.
</issue_to_address>

### Comment 2
<location path="astrbot/core/agent/runners/tool_loop_agent_runner.py" line_range="905-914" />
<code_context>
                 parts = None
+
+            # 过滤掉无效的 tool calls，确保 assistant 消息不包含无 id/name 的条目
+            if llm_resp.tools_call_name and llm_resp.tools_call_ids:
+                valid_indices = [
+                    i for i, (name, tid) in enumerate(
+                        zip(llm_resp.tools_call_name, llm_resp.tools_call_ids)
+                    )
+                    if name and tid
+                ]
+                if len(valid_indices) < len(llm_resp.tools_call_name):
+                    llm_resp.tools_call_name = [llm_resp.tools_call_name[i] for i in valid_indices]
+                    llm_resp.tools_call_args = [llm_resp.tools_call_args[i] for i in valid_indices]
+                    llm_resp.tools_call_ids = [llm_resp.tools_call_ids[i] for i in valid_indices]
+
+            # 如果过滤后没有有效的 tool calls，跳过构建 tool_calls_result
</code_context>
<issue_to_address>
**issue (bug_risk):** Filtering tool calls by index assumes `tools_call_args` is at least as long as `tools_call_name`/`tools_call_ids`, which may raise `IndexError` on malformed provider output.

`valid_indices` comes from `zip(tools_call_name, tools_call_ids)`, but is then applied to `tools_call_args` without verifying that `tools_call_args` is at least as long. If a provider returns fewer args than names/ids, indexing `tools_call_args[i]` will raise `IndexError`. To avoid this, build the filtered lists in a single pass over `zip(tools_call_name, tools_call_args, tools_call_ids)` and append only fully valid triples, or use `zip_longest` with explicit handling of missing values.
</issue_to_address>