Product / area
GitHub Copilot cloud agent in GitHub Desktop app
Version
GitHub Enterprise Cloud (latest)
Severity / impact
High user impact. The agent loops on low-value actions before completing a simple request, making sessions unreliable and inefficient.
Summary
In a fresh ad-hoc session with no project and no custom instructions, the cloud agent repeatedly renames the session to the same title and repeatedly re-reads the same mcp-config.json file before and after running copilot mcp list. This appears to happen before any repository or project custom instructions could influence behavior.
Environment
Fresh ad-hoc session
No project attached
No custom instructions active
Prompt: can you check the MCP status
Steps to reproduce
Open the desktop app.
Start a new ad-hoc session.
Do not attach a project.
Do not add or load any custom instructions.
Enter: can you check the MCP status
Observe the tool/activity timeline.
Actual result The agent:
Renames the session to MCP Status Check
Views mcp-config.json (293 lines)
Renames the session again to the same name
Views the same file again
Repeats the rename + file-read pattern multiple times
Eventually runs copilot mcp list
Then renames again and re-views the same file
Expected result The agent should:
avoid repeated renaming to the same session title
avoid repeatedly reading the same unchanged file across iterations
inspect the relevant MCP configuration once, run the needed MCP status command, and return a result without looping
Observed behavior / analysis
The rename loop begins on the first turn, before project-level custom instructions could reasonably be applied.
The repeated file-read behavior is not specific to one repository or document set; it appears to reflect the agent not trusting or reusing prior-turn context.
This suggests a cloud-agent or desktop-app harness issue rather than a repository instruction issue.
Business impact
Wastes agent execution on repeated low-value actions
Makes simple tasks much slower than expected
Reduces trust in the desktop app agent workflow
Creates risk for broader use cases like repeated automated runs if the same behavior occurs there
Repro frequency Consistent in fresh sessions
Evidence
Attach screenshot showing repeated rename events and repeated reads of the same mcp-config.json
If available, include copied activity/session log showing the exact sequence
Notes Repository instruction files may still help later task behavior, but they do not appear to prevent this first-turn loop.
Product / area
GitHub Copilot cloud agent in GitHub Desktop app
Version
GitHub Enterprise Cloud (latest)
Severity / impact
High user impact. The agent loops on low-value actions before completing a simple request, making sessions unreliable and inefficient.
Summary
In a fresh ad-hoc session with no project and no custom instructions, the cloud agent repeatedly renames the session to the same title and repeatedly re-reads the same mcp-config.json file before and after running copilot mcp list. This appears to happen before any repository or project custom instructions could influence behavior.
Environment
Fresh ad-hoc session
No project attached
No custom instructions active
Prompt: can you check the MCP status
Steps to reproduce
Open the desktop app.
Start a new ad-hoc session.
Do not attach a project.
Do not add or load any custom instructions.
Enter: can you check the MCP status
Observe the tool/activity timeline.
Actual result The agent:
Renames the session to MCP Status Check
Views mcp-config.json (293 lines)
Renames the session again to the same name
Views the same file again
Repeats the rename + file-read pattern multiple times
Eventually runs copilot mcp list
Then renames again and re-views the same file
Expected result The agent should:
avoid repeated renaming to the same session title
avoid repeatedly reading the same unchanged file across iterations
inspect the relevant MCP configuration once, run the needed MCP status command, and return a result without looping
Observed behavior / analysis
The rename loop begins on the first turn, before project-level custom instructions could reasonably be applied.
The repeated file-read behavior is not specific to one repository or document set; it appears to reflect the agent not trusting or reusing prior-turn context.
This suggests a cloud-agent or desktop-app harness issue rather than a repository instruction issue.
Business impact
Wastes agent execution on repeated low-value actions
Makes simple tasks much slower than expected
Reduces trust in the desktop app agent workflow
Creates risk for broader use cases like repeated automated runs if the same behavior occurs there
Repro frequency Consistent in fresh sessions
Evidence
Attach screenshot showing repeated rename events and repeated reads of the same mcp-config.json
If available, include copied activity/session log showing the exact sequence
Notes Repository instruction files may still help later task behavior, but they do not appear to prevent this first-turn loop.