Non-CLI clients (Android + claude.ai web) prepend a <system-reminder>Message sent at …</system-reminder> block to message bodies, breaking slash-command parsing (regression)

Summary

The Claude Code mobile (Android) client prepends a <system-reminder>Message sent at <ISO-8601 timestamp>.</system-reminder> block, plus a trailing newline, to the body of every message it sends.

Slash commands only fire when the message starts with /. The injected prefix pushes /compact, /context, /bug, etc. off the start of the line, so they are parsed as plain text and sent to the model as an ordinary prompt instead of executing.

Regression

This is a regression, not longstanding behavior. Slash commands — /compact in particular — previously worked from the Android client and were used routinely for a long time; they broke recently (observed 2026-06-13). The same commands continue to work from the macOS CLI on the same session, so a recent Android client (or backend) change appears to have introduced the prefix. It is not the case that slash commands "never worked" on Android.

Reproduction

1. Open a Claude Code session and send a slash command (e.g. /compact or /context) from the Android client.

• Observed: the command does not execute; the model responds as if to a prose prompt.
• Expected: the command executes.

1. Send the same slash command from the macOS CLI on the same session — it executes normally.
2. View the session transcript from the CLI: every Android-originated message is prefixed with the raw <system-reminder>Message sent at …</system-reminder> block.

Persists across an Android app restart. Normal (non-slash) prompts are unaffected — they simply carry the extra prefix.

Impact

1. No slash commands are usable from Android — and there is no slash-command picker/dropdown on Android, so there is no fallback.
2. Context grows unbounded — /compact cannot be invoked from Android, so a mobile-only session runs until auto-compact.
3. /exit is not a usable workaround on mobile — after exiting, the session cannot be resumed from the Android client.
4. Internal framing leaks into the user-visible message body — the <system-reminder> tag is harness-internal and should not appear in user message content.

Expected behavior

• Message timestamp metadata should be carried out-of-band, not prepended into the message body.
• Slash-command detection should be robust to (or run before) any such injected prefix.

Environment

Claude Code — Android mobile client vs. macOS CLI, same session.

Onset timeline (verified from the session transcript)

This bug began at a specific, identifiable moment in a long-running multi-day single session. Checking the model on every assistant turn and the raw body of every user message:

• 2026-06-09 → 2026-06-13 09:20 UTC — no prefix at all (~4 days). This window includes an earlier stint on Opus 4.8 (2026-06-09 20:08 → 2026-06-10 10:42) and a long stint on Fable 5 (06-10 10:42 → 06-13 09:20). Slash commands from mobile worked throughout.
• 2026-06-13 ~09:20 UTC — forced migration off Fable 5 onto Opus 4.8. Fable 5 (and Mythos 5) were emergency-disabled by Anthropic on 2026-06-12 to comply with a US export-control directive, so the session fell back to Opus (/model was used to select it) — i.e. an unplanned, rushed shutdown, not a planned feature rollout.
• 2026-06-13 10:06:06 UTC — the FIRST message carrying the <system-reminder>Message sent at …</system-reminder> prefix. Its inner timestamp matches the message's delivery timestamp, confirming it is client/transport-injected (not a quoted example). Every mobile-originated message since carries it; CLI-originated messages never do.

Conclusions from the timeline:

• Not the model. The earlier Opus 4.8 period (06-09/06-10) had no prefix, so Opus itself is not the cause.
• Not a client app update. Onset is mid-session with no app reinstall, and it survives an Android app restart — which points to a server-side / backend change rather than a client release.
• Coincides with the emergency Fable 5 / Mythos 5 shutdown. The prefix first appears right as that rushed 06-12/06-13 disablement forced the session off Fable. Whether the prefix was introduced by that deployment or a separate concurrent change, the emergency Fable/Mythos shutdown deploy is the natural first place to look. Suggested triage window: backend/transport changes around 2026-06-13 ~09:20–10:06 UTC.

Note for triage

While drafting this report, the harness appears to strip complete <system-reminder>…</system-reminder> elements out of user-typed message bodies (apparently anti-injection sanitization): a typed/pasted example of the offending prefix is removed before it reaches the model, while the client-injected prefix itself is not stripped. This makes the bug awkward to report by hand, so this issue body was filed via tooling to keep the literal tags intact.