Bug: async _check_and_compact() ignores tool_use filtering, causes BadRequestError

[fede-kamel]

Bug

When using the async tool runner with compaction enabled, _check_and_compact() sends unfiltered tool_use blocks to the API, causing a BadRequestError.

Error:

BadRequestError: messages.2: `tool_use` ids were found without 
`tool_result` blocks immediately after: toolu_test123

Root Cause

In src/anthropic/lib/tools/_beta_runner.py, the async _check_and_compact() method (line ~472) uses self._params["messages"] instead of the filtered local messages variable when building the summarization request.

The filtering logic (lines ~457-469) correctly removes assistant messages containing only tool_use blocks, but the result is discarded because self._params["messages"] is used instead:

# Line ~472 (async version - BUGGY):
messages = [
    *self._params["messages"],  # ignores the filtering done above
    ...
]

# Line ~211 (sync version - CORRECT):
messages = [
    *messages,  # uses the filtered variable
    ...
]

Reproduction

Gist with standalone reproduction (no API key needed):
https://gist.github.com/fede-kamel/175f0e99abdb53a380ded3f41236e7dd

Fix

PR #1124 — one-line change: replace self._params["messages"] with messages on line ~472.

The PR also fixes a similar issue in append_messages (line ~120) where a lambda captures self._params["messages"] at definition time instead of using params["messages"] at call time.

[xXMrNidaXx]

Validated

I can confirm this bug. The async path diverges from sync in how it references the message list after filtering.

Sync version (correct):

messages = [msg for msg in messages if ...]  # filters
# ...
messages = [*messages, ...]  # uses filtered

Async version (buggy):

messages = [msg for msg in messages if ...]  # filters... discarded
# ...
messages = [*self._params["messages"], ...]  # uses unfiltered

Workaround

Until merged, you can disable compaction and handle it manually:

runner = client.beta.tools.run(
    messages=messages,
    tools=tools,
    # Disable auto-compaction
    max_tokens=8192,  # large enough to avoid triggering
)

Or patch the method:

import anthropic.lib.tools._beta_runner as runner_module

original_check = runner_module.ToolsBetaMessageRunner._check_and_compact

async def patched_check(self, *args, **kwargs):
    # Store filtered messages before call
    filtered = [m for m in self._params["messages"] 
                if not (m.get("role") == "assistant" 
                        and all(c.get("type") == "tool_use" for c in m.get("content", [])))]
    self._params["messages"] = filtered
    return await original_check(self, *args, **kwargs)

runner_module.ToolsBetaMessageRunner._check_and_compact = patched_check

PR #1124 looks correct — hope it gets merged soon.

---

Corey Nida | RevolutionAI

[xXMrNidaXx]

Great bug report with a clear root cause. The sync vs async code path divergence is a subtle but nasty bug class.

Confirming the fix path:

The one-liner in PR #1124 is the right fix. For anyone hitting this in the meantime, you can monkey-patch it:

import anthropic.lib.tools._beta_runner as runner

# Store original
_original_check_and_compact = runner.ToolsBetaRunner._check_and_compact

async def _patched_check_and_compact(self):
    # Filter tool_use-only assistant messages first
    messages = [
        m for m in self._params['messages']
        if not (m['role'] == 'assistant' and 
                all(b.get('type') == 'tool_use' for b in m.get('content', [])))
    ]
    # Now use filtered messages for compaction
    self._params['messages'] = messages
    return await _original_check_and_compact(self)

runner.ToolsBetaRunner._check_and_compact = _patched_check_and_compact

Why this matters:

The unfiltered tool_use blocks cause the API to reject the request because it expects tool_result blocks to immediately follow any tool_use — a constraint that's met in the filtered version but violated when the raw params are used.

The lambda capture issue in append_messages is the same class of bug (stale reference vs live variable). These are easy to miss in code review because both syntaxes look valid.

+1 for merging PR #1124.