[Bug] Repeated automatic model switch from Fable 5 to Opus 5 mid-session

Bug Description

Repeated automatic model switch from Fable 5 to Opus 5 mid-session, with the message
"Fable 5's safeguards flagged this message. This sometimes happens with safe, normal
conversations. Switched to Opus 5." The session then silently continues on Opus 5.
This happened multiple times within a single session.

Workload — ordinary document and static-site work in a local git repository with
no remote: editing Markdown files, generating HTML with a small Python + pandoc
script, generating a .docx, and running two research subagents that read public web
pages. No code execution against third parties, no security tooling, no credentials,
no dual-use subject matter.

Expected — the session stays on the model I selected; or, if a safeguard must
intervene, I am told what triggered it and how to return to my chosen model.

---

There appear to be two separate problems.

1. False positive. The classifier fires on ordinary document work. In case it
helps narrow down the data shapes involved:

• A research subagent returned a single long result containing roughly two dozen real

individuals' names alongside their affiliations, personal site URLs and quoted text,
all gathered from public pages. The flag appeared immediately after one such result
arrived, so aggregation of many named individuals in one message seems like the
strongest candidate.

• A file in the repository containing a list of proper nouns written as regex

patterns, used to exclude those strings from generated output. Out of context this
may resemble evasion tooling; its function is the opposite — it blocks strings
rather than hiding them.

• Ordinary personal contact details (name, email address) appearing in files being

edited.

2. No documented way to restore the current session. Running /model prints
"Set model to Fable 5 and saved as your default for new sessions" — it changes the
default for FUTURE sessions but does not appear to move the CURRENT session back.
After a flag, the session stays on Opus 5 while the settings file correctly reads
claude-fable-5. This makes it look like the switch is happening repeatedly even
when it may have happened once, and there is no obvious way to undo it without
starting a new session.

Suggestions

• Surface which category or content triggered the safeguard, even coarsely.
• Let /model (or a dedicated command) move the CURRENT session back to the selected

model after an automatic switch, and say so in the message.

• Clarify in the switch message whether the change is permanent for the session.

Config (verified)

• ~/.claude/settings.json has "model": "claude-fable-5[1m]"
• No fallback model configured, no model-related environment variable set

Environment Info

• Platform: darwin
• Terminal: Termius
• Version: 2.1.220
• Feedback ID: c51cd725-5d14-460a-9f1a-c366302a88de

Errors

[]