Bedrock: "Cache point cannot be inserted after reasoning block" — non-retryable 400, distinct from #45620

[thinkryoda]

Description

Hit a 400 ValidationException from Bedrock's Converse API that bricks the session, in the same family as #45620 but a different specific mechanism.

statusCode: 400
isRetryable: false
x-amzn-errortype: ValidationException
message: "Cache point cannot be inserted after reasoning block. Please remove the invalid cache point and try again."

Endpoint: https://bedrock-runtime.ap-south-1.amazonaws.com/model/global.anthropic.claude-opus-5-5/converse-stream
Occurred: 2026-09-30T04:08:13Z (per Bedrock's own response date header)
Model: global.anthropic.claude-opus-5-5, reasoning variant max
Session: a long-running session (280 messages) using extended thinking throughout.

How this differs from #45620

#45620 (fixed by #45621 / #45769, merged 2026-08-27) was about unsigned reasoning blocks left behind by a stream that died mid-thinking — the fix filters out reasoning content the SDK can't replay before caching. This error is different: it's Bedrock's API rejecting the structural placement of a cache checkpoint immediately after a reasoning block, independent of whether that reasoning block's signature is valid. The existing fix filters unreplayable reasoning; it does not appear to prevent opencode's caching logic from placing a cache point directly after a (validly signed) reasoning block when that's the last content in the message.

Impact

This was the last message in the session when it occurred — no successful turn has gone through since. It's unclear whether the invalid cache point is purely session-state held by the running client (in which case killing the client and reattaching to the same session would clear it) or baked into the session's stored message data and replayed on every subsequent request (as #45620 originally described: "every request in the session is a non-retryable 400"). Haven't yet confirmed which, since retrying risked losing session continuity without being able to recover it.

Environment

• opencode v1.18.32 (latest as of 2026-09-21)
• Provider: amazon-bedrock, region ap-south-1
• Long session (280 messages), heavy extended-thinking usage ("max" variant)

Request for maintainers

1. Does opencode's cache-point placement logic (applyCaching) need to also skip placing a cache point directly after a reasoning block, not just filter unsigned/unreplayable reasoning content?
2. If a session does hit this, is there a supported way to recover it (e.g. reverting the last message/checkpoint) without losing the rest of the session, given the revert field that exists on stored sessions? Couldn't find documented syntax for this.

[github-actions]

---

This issue might be a duplicate of existing issues. Please check:

• #49859: Reports the same Bedrock ValidationException caused by placing a cache point after a reasoning block.
• #36517: Earlier report of the same Bedrock cache-point placement error.
• #31687: Earlier report of cache points being placed after reasoning blocks on Amazon Bedrock.

[thinkryoda]

Found the exact root cause via opencode export + inspecting the session's stored messages directly.

The assistant message immediately before the poisoned turn has "finish": "length" — it hit the 32,000-token output cap mid-thought. Its reasoning part has no signature field at all (checked directly, not just empty — the key is absent). That's the same missing-signature defect #45620 described, just triggered by hitting the output-length cap instead of a dead/aborted stream. The existing fix (#45621/#45769) filters unreplayable reasoning from a stream that died, but doesn't appear to also catch reasoning left unsigned because the turn was truncated by finish: length.

Confirmed this really does permanently poison the session (not just a stale client connection) — restarting the opencode client and reattaching to the same session reproduced the identical 400 immediately, because the same unsigned reasoning block is resent as part of history every time. Recovered by manually stripping that one reasoning part from the exported session JSON and re-importing — happy to share the repro session if useful, redacted.

[rekram1-node]

https://opencode.ai/v2/docs

Checked on V2, including the output-limit case described in the follow-up comment: thinking without a signature is converted to ordinary text when replayed. Signed/redacted thinking is preserved without placing a cache checkpoint directly after it.

The reported invalid request wasn’t reproduced. A separate cache-picker improvement remains, but it concerns missed caching opportunities rather than this 400.

Closing for V2. Please upgrade and reopen if you still encounter this error, including your V2 version and reproduction details.