[BUG] clearOldReasoning mutates thinking blocks in place, then stripClearedReasoning sentinelizes them — both survive to the API request #162

Summary

clearOldReasoning mutates thinking and text properties of reasoning parts directly on the message object (setting them to the string "[cleared]"). stripClearedReasoning then converts those mutated parts to sentinels ({ type: "text", text: "" }). If either step touches a block that was returned by the most recent API response, Anthropic rejects the subsequent request with:

400: messages.1.content.80: `thinking` or `redacted_thinking` blocks in the latest
assistant message cannot be modified. These blocks must remain as they were in the
original response.

Root cause

// dist/index.js ~line 184913
function clearOldReasoning(messages, reasoningByMessage, messageTagNumbers, clearReasoningAge) {
  const maxTag = findMaxTag(messageTagNumbers);
  const ageCutoff = maxTag - clearReasoningAge;   // can be ≤ 0 on short sessions
  for (const message of messages) {
    const msgTag = messageTagNumbers.get(message) ?? 0;
    if (msgTag === 0 || msgTag > ageCutoff) continue; // skips untagged + recent
    for (const tp of parts) {
      tp.thinking = '[cleared]';   // ← in-place mutation of the original part object
      tp.text     = '[cleared]';
    }
  }
}

// dist/index.js ~line 184947
function stripClearedReasoning(messages) {
  // converts any part whose thinking/text === '[cleared]' to makeSentinel()
  // → { type: 'text', text: '' }
  message.parts[i] = makeSentinel(part);
}

The ageCutoff guard is designed to protect recent messages, but two edge-cases defeat it:

1. Small clearReasoningAge relative to maxTag — if clearReasoningAge is, say, 2 and maxTag is 3, then ageCutoff = 1. The message with tag 2 (second-most-recent) gets its thinking cleared, then sentinelized. If that message is the last assistant turn, Anthropic rejects it.
2. Tag assignment races on slash-command execution — when a slash command fires, the transform pipeline may run before the current assistant turn has been assigned a tag (msgTag === 0). Untagged messages are skipped correctly today, but downstream mutations to already-tagged messages from the same turn boundary are not protected.

Additionally, the in-place mutation of the part object (rather than producing a new object) means any other reference to the same part object also sees the corrupted value.

Reproduction

1. Use an Anthropic model with extended thinking.
2. Have a conversation long enough that maxTag > clearReasoningAge (depends on configured value).
3. Trigger a slash command so the transform pipeline runs.
4. The second-to-last (or earlier) assistant message's thinking blocks are cleared and sentinelized.
5. If any of those blocks are referenced as part of the "latest assistant message" by Anthropic's validator, the API returns 400.

Expected behaviour

• clearOldReasoning must never mutate parts from the last assistant message, regardless of tag number — add an explicit last-assistant-message guard mirroring the fix proposed in issue [BUG] stripReasoningFromMergedAssistants mutates thinking/redacted_thinking blocks in the last assistant message, causing Anthropic 400 #161.
• The in-place mutation (tp.thinking = '[cleared]') should be replaced with a new object or a sentinel substitution at the message.parts[i] level, not a property mutation on the shared part reference.
• stripClearedReasoning should be a no-op if invoked on the last assistant message.

Severity

P1 — triggers on longer sessions where tag numbers exceed clearReasoningAge, or on any session where the tag-assignment race occurs.