Agent repeatedly acknowledges work without executing tool actions

Preflight checklist

• I searched existing issues to avoid creating a duplicate.
• I understand this repository is for GitHub Copilot CLI issues.
• I can reproduce this behavior.

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

1. Start a Copilot CLI session in a repository.
2. Ask for a concrete file-edit task (e.g., "improve BACKLOG.md").
3. Observe assistant responses claiming work has started.
4. Continue prompting when no visible tool activity occurs (e.g., "you’re not doing anything").
5. 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.

报告没有指出源文件、测试或实现入口点。首先在 Windows 上使用 PowerShell 和 gpt-5.3-codex 重现 Copilot CLI 1.0.80 中的行为，记录助手确认消息以及没有工具调用的情况。只有当执行以工具操作或明确的阻塞因素开始，而不是提出没有依据的进度声明时，才算完成。