feat(agent): name the dropped tool when _init_prompt_tools swallows a conversion exception

[camerono]

What

When _convert_tool_to_prompt_message_tool raises inside _init_prompt_tools, the tool gets silently dropped from the LLM payload and the only trace is Failed to convert tool to prompt message tool — which doesn't name the tool. Operators trying to diagnose "why isn't my agent calling X anymore" have no log to grep for.

This PR adds the tool name to the log message and bumps the level from exception (ERROR) to warning, since plugin_daemon routes ERROR through a debug stream by default. No behaviour change beyond observability.

Why this matters

Burned a couple hours this week on the symptom. A custom API tool provider was recreated at some point — same name, new UUID in tool_api_providers. Every existing app's agent_mode.tools[*].provider_id still pointed at the prior UUID. App.deleted_tools (in langgenius/dify at models/model.py) joined on the stale IDs, marked the tools deleted, and _init_prompt_tools here dropped them. The LLM saw a partial tools list and either shortcut-synthesized or returned its own "unknown tool" rejection in-stream. Reading the wire-level capture made it obvious; the existing log alone wouldn't have.

The longer-term fix is probably either (a) a Dify-side migration that rewires stored provider_id references when a provider is recreated under the same name, or (b) refusing the recreate when dependent apps exist. Both are larger changes that belong in langgenius/dify. This PR is the small observability fix that would have shortened my session today.

Diff

-            except Exception:
-                # api tool may be deleted
-                logger.exception("Failed to convert tool to prompt message tool")
+            except Exception:
+                # api tool may be deleted (e.g. its provider was recreated and
+                # the app's stored provider_id no longer matches any row in
+                # tool_api_providers). Naming the tool lets operators grep for
+                # the drop in plugin_daemon logs instead of guessing why the
+                # model can't see it.
+                tool_name = getattr(
+                    getattr(tool, "identity", None), "name", "<unknown>"
+                )
+                logger.warning(
+                    "Dropping tool %r from prompt: conversion failed "
+                    "(provider may be deleted or its schema is stale)",
+                    tool_name,
+                    exc_info=True,
+                )
                 continue

Compatibility Check

• Backward compatible — log line change only, no API change
• Forward compatible
• No breaking change
• No version bump needed — observability-only fix

Available Checks

• just build — haven't run it locally; happy to iterate if CI flags anything

—

I'm grateful for this project's efforts and appreciate the chance to help evolve/improve it.

[gemini-code-assist]

Code Review

This pull request improves error logging in strategy.py when converting a tool to a prompt message tool fails. It safely extracts the tool's name and logs a detailed warning instead of a generic exception, making it easier to identify which tool failed to convert. There are no review comments to address.