“tool_use ids without tool_result” error

What you see

OpenClaw (or the Claude API) may reject a request with:

LLM request rejected: messages.N: `tool_use` ids were found without `tool_result` blocks
immediately after: process..., subagents.... Each `tool_use` block must have a corresponding
`tool_result` block in the next...

Cause

The Claude/Anthropic API requires that every assistant message containing tool_use blocks is immediately followed by a tool-role message that has a tool_result block for each of those tool_use ids. If that ordering is broken, the API rejects the request.

Common ways this happens:

Context-window trimming – When the conversation is trimmed to fit the context window, the code may keep an assistant message that includes tool_use (e.g. from process or subagents tools) but drop the following message(s) that contained the tool_results.
Session replay – Loading from session JSONL where an assistant turn with tool calls was persisted but the corresponding tool-result turn was not (e.g. crash or truncation).
Message building bugs – The code that builds the messages array for the API sometimes omits or reorders tool-result messages.

The fix belongs in the place that builds or trims the messages array before sending to the API (typically OpenClaw core), not in this plugin.

Fix options

1. Sanitize before send (recommended)

Run the provided sanitizer over messages immediately before calling the Claude API. It appends synthetic tool_result blocks for any orphan tool_use ids so the request is valid.

From OpenClaw (or any code that has the messages array):

import { sanitizeMessagesForClaude } from "openclaw-hybrid-memory";

const safeMessages = sanitizeMessagesForClaude(messages);
// send safeMessages to the API

The sanitizer inserts placeholder tool results ("[Output omitted or truncated.]") only for ids that are missing a tool_result in the next message. It does not remove or reorder existing messages.

2. Trim at safe boundaries

If you trim the conversation for context length, never cut in the middle of a turn. When trimming from the start (or dropping middle messages), ensure you never leave an assistant message that has tool_use blocks without the immediately following tool-role message(s). Either:

Remove the assistant message as well (trim one more message), or
Use sanitizeMessagesForClaude() after trimming so any orphan tool_use gets a synthetic tool_result.

Exported API

sanitizeMessagesForClaude(messages) – Returns a new array if it had to insert or extend tool messages; otherwise returns the same array. Message objects are not mutated; only new elements or new wrapper objects are added.
MessageLike – Type for message-like objects (role, optional content as string or array of blocks) for use in TypeScript.

Defined in extensions/memory-hybrid/utils/sanitize-messages.ts and re-exported from the plugin entrypoint.