Spike: can a conversation mix local and upstream turns?

Issue: #7 · Date: 2026-10-11, case by case 2026-10-12 · Status: complete

Question

What does the Anthropic API do with a conversation in which some assistant turns came from a local model?

Method

No API key is available on this machine, so the API could not be probed with hand-built requests as the issue planned. Instead a real Claude Code session was run through Janus with a rule that sends only small subagent turns local ("class": "subagent", "maxTokens": 16500). A subagent's first turn is under that size and its second is over, so the conversation starts on qwen/qwen3.8-27b and continues on Claude, with the local turn in its history.

Finding 1: unsigned thinking blocks are rejected

The second turn, sent upstream, was refused:

HTTP 400
{"type":"error","error":{"type":"invalid_request_error",
 "message":"messages.1.content.0: Invalid `signature` in `thinking` block"}}

The history Claude Code sent back contained the local turn as:

[{"type":"thinking","thinking":"The user is asking to count the lines in…","signature":""},
 {"type":"tool_use","id":"587714761","name":"Bash", …}]

LM Studio returns thinking blocks with no signature field. Claude Code stores the block and replays it with "signature": "". The API signs the thinking it produces and rejects a block whose signature is missing or does not verify.

Claude Code recovered by starting the subagent again from scratch, so the session finished, having paid for the failed attempt and a fresh conversation.

This affected shipped behaviour. Any rule that can send part of a conversation local and part upstream produced it: a maxTokens condition, or a local rule falling back upstream when the conversation outgrew the local model's context.

Decision

Before a request goes upstream, Janus removes from its history any thinking block with an empty or missing signature (backend.RepairForUpstream, ledger stage drop_unsigned_thinking).

• Blocks the API signed are never touched, including the ones Claude Code receives with empty text under display: "omitted"; their signature is not empty. A conversation that only ever went to Claude is forwarded byte for byte, as before.
• The repair is on the request, on the way up, and not on the local response on the way down. That covers histories that already hold local thinking, and avoids rewriting block indexes in a live stream.
• Requests going to the local model keep the local thinking.

Finding 2: with the repair, the mixed conversation is accepted

The same session was run again with the repair in place. One subagent conversation, two turns:

Turn	Answered by	What happened
1	qwen/qwen3.8-27b	Thinking, text, and a Bash tool call with id "..." as a bare number
2	Claude, after drop_unsigned_thinking	HTTP 200. Claude read the tool result and finished the task correctly.

So, for this case:

1. An assistant turn with a tool call and no thinking block is accepted mid-conversation, with adaptive thinking enabled on the request.
2. LM Studio's numeric tool-use ids are accepted in tool_use and the matching tool_result.
3. Nothing else about the local turn tripped validation.

This is one run of a two-turn conversation on one model. It shows the repair removes the failure that was observed; it does not show that no other shape of local turn can fail.

Finding 3: case by case

On 2026-10-12 each kind of mixed history was tried on its own. A throwaway pass-through proxy sat between headless Claude Code (2.1.296, claude -p --effort high, model claude-opus-5-5) and the API. Claude Code authenticated as usual with the machine's Pro subscription; the proxy forwarded the auth header untouched and never read or stored it. On each request that already held assistant turns, it made one change to the history, the kind a local model's turn would bring, and recorded the API's status, error, and usage.

The task was a short tool loop over a scratch repo: three file reads, one of them needing worked reasoning, so that the history held signed thinking blocks (display: "omitted": empty text and a real signature) once the model had thought.

Change to the history	API answer
None (control)	200
Signature field removed from the last turn's thinking block	400 messages.7.content.0.thinking.signature: Field required
Ten characters of a real signature altered	400 Invalid `signature` in `thinking` block
Thinking text replaced with a local model's, signature ""	400 Invalid `signature` in `thinking` block
Signed thinking removed from the last assistant turn	200
Signed thinking removed from every assistant turn	200
Last turn made local-shaped: text first, no thinking, tool ids call_…	200
Every tool id renamed to call_…, in tool_use and tool_result	200
thinking: {type: "disabled"}	400 "thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort"

What this adds to findings 1 and 2:

• The API checks every signature it is shown, and nothing else about origin. A thinking block must carry a signature the API issued, for that text. Tool ids in any format, a turn with no thinking, and text before a tool call all pass.
• Removing signed thinking is accepted, but not free. In the turn after the removal, Claude worked the reasoning out again: 840 and 875 output tokens with a fresh thinking block, against 117 with no thinking at the same step of the control. One run each, and adaptive thinking varies from run to run, so read this as a direction, not a size.
• Thinking off is not an option on this model. The "thinking on and off" half of the original question does not arise for claude-opus-5-5; every request must use adaptive thinking.
• Claude Code repairs a rejected history itself. All three times a signature was refused, Claude Code sent the request again without the thinking blocks in the failing turn, and that succeeded. So a mistake here costs a failed request rather than a failed turn, which also means it is easy to miss without the ledger.

Routing granularity

Granularity	Safe?	Why
Per request class	Yes	Nothing crosses between conversations.
Per whole subagent conversation	Yes	A subagent's conversation never mixes origins.
Local, then upstream within a conversation (escalation, fallback)	Yes, with drop_unsigned_thinking	The only history the API rejects is thinking it did not sign, and the repair removes exactly that.
Turn by turn in the main loop	Accepted by the API; not recommended	No longer an integrity question. Each local turn costs Claude its reasoning for the next turn (finding 3), and a local turn takes 10 to 50 seconds where Claude takes 2 to 8 (#8). Whether that ever pays is for shadow and replay data on real work to show.

FR-4.4 is met by the repair: no routing Janus can do produces a history the API rejects, as long as no local thinking block is ever sent upstream as a thinking block.

Still open

• What a later Claude turn does, in quality terms, with a history that holds a local model's tool calls. Acceptance is not the same as working well.
• A local turn that is only thinking, with no text or tool call. The repair replaces it with a placeholder text block; that path is tested against a fake upstream only.