stabilizePromptCache re-attaches Claude Code's <total_tokens> footer as a user message; routed models treat every tool result as a new user turn and re-plan #6646

Client or integration

Claude Code

Area

Proxy and routing

Summary

With claudeCode.stabilizePromptCache: true, the Claude Messages → Responses translation peels Claude Code's <total_tokens>N tokens left</total_tokens> footer off the system prompt (good for the cache) and re-attaches it as a trailing role: "user" input message on every request (src/claude/inbound.ts, the stabilized.dynamicNotice branch). The routed model reads that as a new user turn. After every tool result it starts a fresh turn: re-announces the task ("Starting on …"), re-runs orientation commands (git status, re-reads the issue), and never gets to the edit it was about to make.

Observed with Claude Code subagents routed to meta-muse/muse-spark-1.3-contributor through a combo, at low, medium and high effort: 19–302 plan restarts per agent, for hours, with the model answering normally on every turn. Short read-only tasks still finish (a fresh start just picks the next file), which hid the cause for a while.

Expected: the peeled footer stays out of the conversation, so the routed model sees only the user turns Claude Code actually sent.

Cause. stabilizeClaudeInstructionsForPromptCache returns the peeled footer as dynamicNotice; translateAnthropicRequest then does

input.push({ type: "message", role: "user", content: [{ type: "input_text", text: stabilized.dynamicNotice }] });

so the conversation ends with a user message the user never wrote. PR #4052 introduced this; its review thread raised "models that take the token XML as user speech" as a concern and it was not measured.

Fix that works here. Drop the peeled notice instead of re-attaching it. Nothing on the translated route acts on Claude Code's token bookkeeping (compaction is client-side), and the cache stays stable because the footer still leaves instructions. Patched locally this way: the reproduction below answers 1, prompt-cache hits are unchanged, and the same long agent runs that looped for hours now progress.

If the notice must be kept, a role: "developer" item would at least not read as the user speaking, but I have not tested whether every Responses-compatible provider accepts that role.

Reproduction

One request, no agent needed. POST /v1/messages to a translated model with:

• system: [{"type":"text","text":"You are a test harness. Answer only what is asked."}, {"type":"text","text":"<total_tokens>14900000 tokens left</total_tokens>"}]
• one user message: How many user messages are in this conversation, counting this one? Answer with the number, then quote each user message verbatim on its own line.

Version

2.76.0 (the same code is present in 2.78.0-preview.20261005)

Operating system

macOS 26.6.2 (arm64)

Provider and model

meta-muse / muse-spark-1.3-contributor (via combo); not provider-specific, the extra user message is added before routing

Logs or error output

No response

Screenshots and supporting files

No response

Redacted configuration

{ "claudeCode": { "stabilizePromptCache": true } }

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.