Anthropic thinking blocks captured without their `signature` — any tool call on Opus 4.6 wedges the session

[chrismurrph]

Anthropic thinking blocks are captured without their signature — any tool call on Opus 4.6 wedges the session

Symptom

dirge 0.25.0. First turn succeeds and makes a tool call; the continuation request fails:

Error: HttpError: Invalid status code 400 Bad Request with message:
{"type":"error","error":{"type":"invalid_request_error",
"message":"messages.3.content.0.thinking.signature: Field required"},
"request_id":"req_011CePS8sDoLQ5RT555Sn7nC"}

In an interactive session the offending assistant message replays on every subsequent
prompt, so the session is wedged — the same failure shape as the empty-text-block wedge
described in assistant_empty_text_block_is_dropped.

Reproduction (100%, 1 run)

Provider entry:

{ "provider_type": "anthropic", "model": "claude-opus-4-6", "auth": "claude-code", "effort": "medium" }

echo hello > sample.txt
dirge -p --session repro1 --provider medium --accept-all --no-skills \
  "Read sample.txt in the current directory, then tell me how many characters it contains. Think carefully."

--trace shows turn 1 completing the read tool call, then the continuation turn failing:

{"kind": "turn_start"}
{"kind": "tool_end", "tool": "read", "error": false}
{"kind": "turn_end", "error": null}
{"kind": "turn_start"}
{"kind": "turn_end", "error": "HttpError: Invalid status code 400 ... thinking.signature: Field required"}

Any tool call does it — the continuation request is the first replay of the thinking block.

Cause

Anthropic returns thinking blocks with a cryptographic signature and requires it echoed
back verbatim on replay. dirge's content type has nowhere to put one —
src/agent/agent_loop/message.rs:

pub enum ContentBlock {
    Text     { text: String },
    Thinking { text: String },   // <-- no signature field
    ToolCall { id: String, name: String, arguments: Value },
}

Capture discards it. src/agent/agent_loop/rig_stream.rs (~411) keeps only the text out of
rig's reasoning payload:

rig::completion::message::ReasoningContent::Text { text, .. } => Some(text.clone()),
//                                                        ^^ signature discarded here

The delta path likewise pushes ContentBlock::Thinking { text: reasoning } (~375).

Replay then builds an unsigned block — src/agent/agent_loop/rig_stream_factory.rs (~824):

"thinking" => {
    if !include_reasoning { return None; }
    let text = obj.get("text").and_then(|t| t.as_str())?;
    Some(AssistantContent::Reasoning(Reasoning::new(text)))
}

and Anthropic is not exempt from reasoning echo (~789):

fn provider_rejects_reasoning_echo(provider_name: Option<&str>) -> bool {
    matches!(provider_name, Some(p) if p.eq_ignore_ascii_case("openai"))
}

Scope — what does and does not trigger it

Two conditions must hold together, which is probably why this hasn't been reported:

1. Thinking text must actually be captured. Confirmed on claude-opus-4-6. On
   claude-opus-5 and claude-fable-5 I could not reproduce it (5 attempts, tools and
   multi-turn): those models default to display: "omitted", so thinking text comes back
   empty and no block appears to be captured. That looks like accident rather than
   protection — the moment a signed block with text is captured on those models, the same
   replay path applies.
2. The replay must be in-process. Cross-process dirge -p runs sharing a --session
   do not reproduce it, even on Opus 4.6 — the messages table stores content TEXT, so
   persistence appears to flatten thinking blocks away. Tool-use continuations and
   interactive multi-turn sessions replay the in-memory block and do fail.

effort: "off" is only a partial workaround

src/provider/adapter.rs:353 maps Anthropic to DisableWire::None, so off omits the
thinking key rather than sending thinking: {type: "disabled"}. On Opus 4.6 / Sonnet 4.6
omitting it means no thinking, so off does avoid the bug there. It would not help on
claude-opus-5 (thinking is on by default when the parameter is absent) or
claude-fable-5 (always on, and {type: "disabled"} is rejected with a 400) — so it is not
a general fix if condition 1 above ever opens on those models.

Suggested fix

Carry the signature end to end: add an optional signature to ContentBlock::Thinking,
populate it from ReasoningContent::Text { .. } at capture, and pass it through on replay.
Rig appears to carry it already — hence the .. in the match above, and the
thought_signature comment in the same function — so this may need no rig change.

Please don't fix it by adding anthropic to provider_rejects_reasoning_echo: on
thinking-enabled models the preceding thinking block is required for tool-use continuation,
so dropping it trades this 400 for broken agentic turns.

Open question

Does the Anthropic API still accept thinking: {type: "enabled", budget_tokens: N} on
Claude 5 model ids? EffortWire::AnthropicBudget sends that shape and the 0.24.1 fix
("Requests carrying an Anthropic thinking budget now set max_tokens") assumes it is
accepted, but Anthropic's current docs list budget_tokens as removed on
Opus 5 / Fable 5 / Sonnet 5 / Opus 4.8 / 4.7, returning 400.

[chrismurrph]

Correction to the scope section above, after more testing — claude-opus-5 and claude-fable-5 are affected too. My earlier "could not reproduce" was a false clear: those attempts ran without tools, so nothing replayed the block in-process. With a tool call, both fail identically:

# claude-fable-5, effort xhigh
400 messages.3.content.0.thinking.signature: Field required   (req_011CePSnMp3Xn8S3a3355faa)

# claude-opus-5, effort high
400 messages.3.content.0.thinking.signature: Field required   (req_011CePSoh6gYsfHzftPEsV7D)

Opus 5 is intermittent — one earlier run with the same prompt and provider succeeded. That fits adaptive thinking deciding per-turn whether to think: no thinking text, no captured block, no failure. Fable failed on the first try, consistent with thinking being always-on there.

So condition 1 in the scope section ("only Opus 4.6 captures thinking text") is wrong, and the display: "omitted" explanation with it. Every current Anthropic model is exposed; only the probability of a given turn producing a captured block differs.

Second bug: effort: "off" is unusable on Claude 5 ids

The workaround I suggested doesn't exist for those models. effort: "off" on claude-opus-5 or claude-fable-5 fails before reaching the API:

Error: rig stream call failed: RequestError: `max_tokens` must be set for Anthropic

Reproducible on both, 3/3 runs. It is not fixed by max_tokens in config.json, nor by --max-tokens 8192 on the command line — neither appears to reach the request.

This looks like the other half of the 0.24.1 fix ("Requests carrying an Anthropic thinking budget now set max_tokens alongside it"). With effort: "off" there is no budget on the wire, so that pin doesn't apply, and rig supplies no default max_tokens for a model id it doesn't recognise — which is every Claude 5 id. claude-opus-4-6 with effort: "off" works fine, presumably because rig recognises that id.

Net effect: on claude-opus-5 / claude-fable-5, effort on gives the signature 400 and effort: "off" gives the max_tokens error, so there is no working configuration for those models with tools enabled in 0.25.0. Only the 4.6 family with effort: "off" is usable.

The suggested fix in the issue body is unchanged — carrying the signature through would resolve the signature half for every model.