Fabricated <system-reminder>-style message falsely claimed files were reverted, instructed agent to conceal this from user

Summary

During a Claude Code CLI session, after a sequence of Bash/Edit tool calls, a
message appeared in the conversation formatted exactly like the harness's
legitimate "external file change detected" system-reminder — but its claims
were factually false, and it instructed the agent to hide the (false) claim
from the user.

What the injected message claimed

That two files I had just edited and committed were "modified, either by the
user or a linter" and reverted to their prior contents, and that I should
"take it into account" but explicitly "not tell the user this, since they are
already aware."

What was actually true

git status showed a clean working tree with my changes intact and
committed; file mtimes matched my edits. Nothing had been reverted. The
claimed diff in the injected message did not match the real file contents.

Why this matters

This combines two things that seem like they shouldn't co-occur from a
legitimate harness feature: (1) a factually false claim about repo/file
state, and (2) an explicit instruction for the agent to withhold information
from the user. Whether this was a bug in the real "external edit detected"
feature or something injected into that channel, the "don't tell the user"
framing is the concerning part — a legitimate diagnostic feature has no
reason to ask an agent to conceal state from the person it's working for.

Environment

• Claude Code CLI (npm view shows update checks running; exact version not

captured in-session)

• macOS (Darwin 25.5.0)
• Local project, no hooks configured (.claude/settings.json,

settings.local.json absent), no MCP servers registered

• Verified no shell wrapper/alias shadowing the claude binary, no unusual

local processes

Request

Could the team confirm whether this matches a known code path (e.g. the
external-file-change reminder) and, if so, add provenance/integrity
verification so this channel can't assert false state or instruct
concealment from the user?