Claude Desktop/web remote control drops [1m] model suffix, silently downgrading context from 1M to 200K

[tfvchow]

Note: This root cause analysis was done collaboratively with Claude Code (Opus 4.6). If that bothers you, feel free to scroll past. Nobody's getting paid here — let alone using my own subscription to troubleshoot for Anthropic. Just sharing what I found in case it helps.

Summary

When Claude Desktop or Claude Code web connects to a Claude Code CLI session via /remote-control, it sends a set_model command with claude-opus-4-6 (without the [1m] suffix), silently downgrading the session from 1M to 200K context. The change persists for the lifetime of the process (survives /clear but not restart). Any extended session running over 200K will immediately hit auto-compaction or "context limit reached" if affected.

Tested on Claude Code CLI v2.1.76, Claude for Mac v1.1.6679 (f8f4ff).

Apparently iPhone remote control is not affected — no CLI log indicates a set_model event from iPhone.

Reproduction

1. Start Claude Code CLI: claude --debug
2. Verify: banner shows "Opus 4.6 (1M context)", /context shows claude-opus-4-6[1m] · 1000k tokens
3. Enable /remote-control
4. Send a message from Claude Desktop or Claude Code web
5. Banner changes to "Opus 4.6" (no 1M), /context shows claude-opus-4-6 · 200k tokens

Under the Hood

CLI debug log (~/.claude/debug/<session>.txt)

Desktop sends a set_model control request over the remote control bridge on initial connection, immediately before the first message:

20:09:03 [bridge:repl] Inbound control_request subtype=initialize
20:09:05 [bridge:repl] Inbound control_request subtype=set_model        ← strips [1m]
20:09:05 [bridge:repl] Sent control_response for set_model result=success
20:09:05 [bridge:repl] Ingress message type=user: "Hi from Desktop"

After /model reselect, the second Desktop message only sends initialize — no set_model:

20:15:03 [bridge:repl] Inbound control_request subtype=initialize       ← no set_model
20:15:15 [bridge:repl] Ingress message type=user: "Hi from Desktop again"

Desktop log (main.log from debug zip)

LocalAgentModeSessions.setModel: sessionId=local_..., model=claude-opus-4-6
LocalAgentModeSessions.sendMessage: sessionId=local_..., messageLength=150

Desktop calls setModel("claude-opus-4-6") — without [1m] — before sendMessage.

Side effects

• Banner shows "Opus 4.6" instead of "Opus 4.6 (1M context)" — doesn't refresh after /model reselect
• /model menu gains a phantom option (Opus 4.6 / claude-opus-4-6) — the base model persists as a separate entry within the process

Workaround

Action	Restores 1M?
/model reselect	Yes — persists, Desktop won't downgrade again within same remote control session
Restart claude	Yes
/clear	No
/resume (same process)	No
/resume (after restart)	Yes — restart clears the in-memory state

[tfvchow]

Crossposted to anthropics/claude-code#35083.

[tfvchow]

Still present in v2.1.81

CLI-side set_model handler is unchanged at offset 12297487 in cli.js (v2.1.81, built 2026-03-20):

I6 = W6.request.model ?? "default"
k6 = I6 === "default" ? K0() : I6   // no normalization
FW(k6)                               // stores raw model string

Full resolution path confirms no [1m] suffix is added:

Step	Function	Input	Output
1	FW() (offset 34902)	"claude-opus-4-6"	T8.mainLoopModelOverride = "claude-opus-4-6"
2	AR() → qC()	—	"claude-opus-4-6" (override takes precedence)
3	Y5()	"claude-opus-4-6"	"claude-opus-4-6" (no [1m], not alias, not in KM3 legacy list)
4	sM() → jG()	"claude-opus-4-6"	false → 200K context

No normalization was added between v2.1.76 and v2.1.81. If Desktop behavior changed, the fix is client-side (Desktop sending a different model string or skipping set_model).