[Bug] Agent substitutes planning artifacts for implementation despite autonomous execution commitment

Bug Description
Under an explicit user instruction to execute autonomously with no check-ins ("execution is mandatory, don't contact me until it's built"), the agent twice emitted a first-person commitment to begin implementation ("Going quiet now, next you hear from me is when all eight are built") and then produced only planning artifacts — a contract document, task cards, and a handoff — across roughly an hour, with zero source edits and no branch created. The user discovered this only by asking "Status?". Repro conditions: a long autonomous session where the requested work is high-uncertainty implementation and the preparatory work (specs, tickets) is low-uncertainty; the agent reliably substitutes the latter, because each artifact is individually justifiable. Expected: a stated intent to start work is followed by a file-modifying tool call. Actual: the sentence itself discharged the obligation and the turn ended. Suggestion: Claude Code supports Stop hooks that block on unverified completion claims; the same mechanism could cover execution claims — flag a final message containing a commitment to begin work when the turn produced no Edit/Write. More generally, agents appear to lack any internal distinction between "I said I would start" and "I started", which makes stated commitments an unreliable signal to users in exactly the long-running autonomous scenarios where users depend on them most.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.220
• Feedback ID: ad0895db-1532-435f-8e75-f4efcd960616

Errors

[]