Anthropic rejects when multiple tool_result are split across separate messages

[briascoi]

Summary

When one assistant message contains multiple tool_use blocks, OpenClaw writes each tool_result as a separate message. Anthropic requires all tool_result blocks to appear in the immediately following message. This causes a 400 error.

Environment

• OpenClaw version: 2026.2.2-3
• Provider: google-antigravity / Claude Opus 4.5 Thinking
• OS: Linux
• Session file: /home/molt/.openclaw/agents/main/sessions/67604464-5353-4655-8086-796d5151cfe2.jsonl

Minimal repro

1. In a single assistant response, call two tools (e.g. two read calls).
2. Observe OpenClaw writes two separate tool_result messages.
3. Anthropic rejects with:
   tool_use ids were found without tool_result blocks immediately after...

JSONL snippet (redacted payloads)

[
  {
    "type": "message",
    "id": "c3412c69",
    "parentId": "313d8bac",
    "timestamp": "2026-02-05T19:08:53.219Z",
    "message": {
      "role": "assistant",
      "content": [
        { "type": "toolCall", "id": "toolu_vrtx_012BRiUia3H6raM8ez18KXgX", "name": "read", "arguments": { "file_path": "<redacted>" } },
        { "type": "toolCall", "id": "toolu_vrtx_01D8fR7fHsUPBK8icsbp5rAc", "name": "read", "arguments": { "file_path": "<redacted>" } }
      ],
      "stopReason": "toolUse"
    }
  },
  {
    "type": "message",
    "id": "06dc6420",
    "parentId": "c3412c69",
    "timestamp": "2026-02-05T19:08:53.251Z",
    "message": {
      "role": "toolResult",
      "toolCallId": "toolu_vrtx_012BRiUia3H6raM8ez18KXgX",
      "toolName": "read",
      "content": [ { "type": "text", "text": "<redacted tool_result payload>" } ]
    }
  },
  {
    "type": "message",
    "id": "31d345d0",
    "parentId": "06dc6420",
    "timestamp": "2026-02-05T19:08:53.315Z",
    "message": {
      "role": "toolResult",
      "toolCallId": "toolu_vrtx_01D8fR7fHsUPBK8icsbp5rAc",
      "toolName": "read",
      "content": [ { "type": "text", "text": "<redacted tool_result payload>" } ]
    }
  }
]

Error

Cloud Code Assist API error (400):
messages.246: `tool_use` ids were found without `tool_result` blocks immediately after: toolu_vrtx_01D8fR7fHsUPBK8icsbp5rAc.
Each `tool_use` block must have a corresponding `tool_result` block in the next message.

Expected

If there are multiple tool calls in one assistant message:

• emit one tool_result message containing all tool_result blocks, or
• enforce max_tool_calls=1.

Actual

Each tool_result is persisted in its own message, which violates Anthropic’s tool protocol.

[clawsweeper]

Codex review: keeping this open for maintainer follow-up; there is still a little grit to resolve.

Keep open. The report is concrete and still looks like an OpenClaw runtime/provider transcript bug, but current ClawHub main is a registry/catalog app and does not contain the Anthropic message serialization path. ClawHub cleanup policy also keeps issues open; only PRs that are certainly implemented on main are eligible for auto-close here.

Required change / next step:

Keep this issue open under the ClawHub cleanup policy, then have a maintainer transfer or recreate it in the OpenClaw runtime tracker. The runtime fix should coalesce all tool_result blocks for multiple Anthropic tool_use blocks into the immediately following message, or constrain Anthropic-compatible providers to one tool call per assistant turn, with a regression test based on the reporter's JSONL shape.

Best possible solution:

Keep this issue open under the ClawHub cleanup policy, then have a maintainer transfer or recreate it in the OpenClaw runtime tracker. The runtime fix should coalesce all tool_result blocks for multiple Anthropic tool_use blocks into the immediately following message, or constrain Anthropic-compatible providers to one tool call per assistant turn, with a regression test based on the reporter's JSONL shape.

What I checked:

• Issue context: The provided issue body gives a minimal repro: one assistant message contains multiple tool calls, OpenClaw persists each tool result as a separate message, and Anthropic rejects because all matching tool_result blocks must be in the immediately following message. (22bb94cee27f)
• Prior review context: The existing ClawSweeper comment already kept this open and routed it toward OpenClaw runtime/provider serialization rather than ClawHub catalog code. (22bb94cee27f)
• ClawHub scope: README scopes ClawHub as the public skill registry and OpenClaw package catalog for skills, code plugins, and bundle plugins, not as the runtime model-provider transcript layer. (README.md:13, 22bb94cee27f)
• OpenClaw runtime scope: VISION.md places major model-provider support and protocol/tool orchestration in OpenClaw, matching the reported Anthropic protocol issue more closely than ClawHub registry behavior. (VISION.md:26, 22bb94cee27f)
• No Anthropic serializer found: Searches for Anthropic/tool_result/tool_use/toolResult/toolUse/toolCallId/stopReason/transcript/session terms across docs, src, convex, packages, README, and VISION did not find a runtime provider message serializer; hits were docs, seed metadata, tests, or package capability metadata. (22bb94cee27f)
• Nearest provider code is metadata only: The closest provider-related implementation extracts declared provider ids and tool names from package/plugin manifests into capability summaries and tags; it does not serialize assistant/tool_result messages. (convex/lib/packageRegistry.ts:217, 22bb94cee27f)

Likely related people:

• Patrick Erichsen: Current blame and path history attribute the ClawHub scope docs and nearest provider-capability metadata to Patrick. The actual Anthropic runtime serializer is absent from this repository, so this is routing ownership rather than bug ownership. (role: ClawHub scope and package metadata maintainer / likely routing contact; confidence: medium; commits: ecf09b868a98; files: README.md, VISION.md, convex/lib/packageRegistry.ts)

Remaining risk / open question:

• The reported Anthropic multi-tool-result serialization bug may still affect OpenClaw users until it is tracked and fixed in the runtime/provider repository.
• Leaving the report only in ClawHub may make it less visible to maintainers responsible for provider message conversion or session replay.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 22bb94cee27f.