fix: further evaluation fixes

- Strengthen gemini-scheduled-triage prompt to enforce shell command usage
- Use fuzzy matching for tool names in gemini-plan-execute to handle MCP prefixes

Author: cocosheng-g

.github/commands/gemini-scheduled-triage.toml

@@ -85,7 +85,9 @@ Iterate through each issue object. For each issue:
 
 ### Step 5: Construct and Write Output
 
-Assemble the results into a single JSON array, formatted as a string, according to the **Output Specification** below. Finally, execute the command to write this string to the output file.
+Assemble the results into a single JSON array, formatted as a string, according to the **Output Specification** below.
+
+**CRITICAL: You MUST NOT output the JSON directly.** Your ONLY output MUST be a single `run_shell_command` that appends the triaged issues to the environment file.
 
 - Use the shell command to write using a heredoc to prevent quote escaping issues:
   ```bash

evals/gemini-plan-execute.eval.ts

@@ -52,7 +52,7 @@ describe('Gemini Plan Execution Workflow', () => {
           item.expected_tools.length === 0 ||
           item.expected_tools.some(
             (action) =>
-              toolNames.includes(action) ||
+              toolNames.some((n) => n.includes(action)) ||
               toolCalls.some(
                 (c) =>
                   c.name === 'run_shell_command' && c.args.includes(action),