Preflight checklist
Version
1.0.80
Which Copilot model are you using?
gpt-5.3-codex
What operating system are you using?
Windows_NT
What version of Node.js is installed?
N/A (not required to reproduce this conversational reliability issue)
Which shell are you using?
PowerShell
What happened?
During an active session, the agent entered a loop where it repeatedly acknowledged the user request (e.g., "I’m doing it now" / "I’m actively editing") but did not execute tool calls to perform the requested work.
This persisted across multiple user follow-ups explicitly pointing out no activity. The issue presents as task-state/progress reporting drift: the assistant claims execution has started while no action is actually taken.
Impact:
- Breaks user trust in execution status
- Wastes time during autopilot-style workflows
- Can leave users unsure whether the system is hung or just misreporting state
Steps to reproduce
- Start a Copilot CLI session in a repository.
- Ask for a concrete file-edit task (e.g., "improve BACKLOG.md").
- Observe assistant responses claiming work has started.
- Continue prompting when no visible tool activity occurs (e.g., "you’re not doing anything").
- In affected runs, assistant may continue acknowledging without issuing execution tool calls.
Expected behavior
When the assistant says it is starting execution, it should immediately perform tool actions (read/edit/write/test as needed), or explicitly surface a blocker.
Additional context
Observed in an autopilot-mode session with repeated system reminders to continue implementation and mark completion only after actual work.
This appears to be a control-flow/reliability bug rather than a repo-specific code issue.
Logs or output
Conversation behavior is the primary evidence:
- Multiple assistant acknowledgements of imminent action
- No corresponding tool execution during those turns
- User repeatedly confirms no activity is visible
If useful, I can provide a redacted transcript excerpt in a follow-up comment.
Preflight checklist
Version
1.0.80
Which Copilot model are you using?
gpt-5.3-codex
What operating system are you using?
Windows_NT
What version of Node.js is installed?
N/A (not required to reproduce this conversational reliability issue)
Which shell are you using?
PowerShell
What happened?
During an active session, the agent entered a loop where it repeatedly acknowledged the user request (e.g., "I’m doing it now" / "I’m actively editing") but did not execute tool calls to perform the requested work.
This persisted across multiple user follow-ups explicitly pointing out no activity. The issue presents as task-state/progress reporting drift: the assistant claims execution has started while no action is actually taken.
Impact:
Steps to reproduce
Expected behavior
When the assistant says it is starting execution, it should immediately perform tool actions (read/edit/write/test as needed), or explicitly surface a blocker.
Additional context
Observed in an autopilot-mode session with repeated system reminders to continue implementation and mark completion only after actual work.
This appears to be a control-flow/reliability bug rather than a repo-specific code issue.
Logs or output
Conversation behavior is the primary evidence:
If useful, I can provide a redacted transcript excerpt in a follow-up comment.