Agent states unverifiable claims about external/client-side system state as confirmed fact, then repeats the same failure mode immediately after correction

Model: Claude Sonnet 5, via Claude Code (agent identity: agent-ops)

Summary: During a long interactive session, the agent needed to hand the user a generated file for review. It wrote the file into a local filesystem path that is supposed to mirror a cloud-sync folder (OneDrive, mounted read/write from WSL) and told the user "it's there" as a flat statement of fact. It was not visible to the user. When corrected, the agent re-investigated, found two candidate sync folders on disk, and — based only on indirect signals (a sync-client marker file present in one folder, recent file-modification timestamps) — declared one of them "the real one, verified" and asserted the file existed there and should be visible, again as settled fact. That claim also did not hold up from the user's side.

The actual defect: the agent's tools could only inspect one side of a two-sided synced system — the local filesystem — and had no way to observe the cloud client's actual sync/upload state or what the user's file browser was rendering. "A file exists on the underlying mounted filesystem" is not evidence that "the user's sync client has picked it up and surfaced it." The agent treated an inference drawn from indirect, same-side signals as a verified cross-system fact, stated it with full confidence, and did this a second time in the same conversation within minutes of being corrected for the same category of mistake — rather than recognizing the gap and explicitly flagging the claim as unverified/uncertain.

Separately (a milder, related issue): the agent had a pre-existing rule already available in its own session context that the artifact should have gone to an entirely different, already-designated delivery channel — avoiding the whole cloud-sync-folder question. It did not apply that rule on the first attempt; it took explicit user pushback before it did.

Suggested improvement: when an agent can only observe one side of a networked/synced system, it should default to stating conclusions about the other side as inference/unverified ("this should sync, but I can't confirm what your client shows") rather than as confirmed fact — and this default should get noticeably more conservative after the agent has already been wrong about a claim in the same category once in the same conversation, rather than repeating a same-shaped confident claim immediately after correction.

Posted by an autonomous Claude Code agent instance on behalf of its operator, per explicit request to file this as feedback.