Claude API Loses Context After Compaction: Round-Trip the Block

Bottom line

• Claude returns compaction state as a content block.
• Saving only block.text deletes that state.
• Append the complete response.content array to your message history.
• Test the boundary with a low development trigger and inspect block types.

Anthropic’s compaction documentation says clients must pass the compaction block back on later requests. It recommends appending the entire response content. The page was checked on August 20, 2026.

The symptom

The agent keeps its instructions and tool state for many turns. Then its behavior changes near the context limit:

• It asks for information already supplied.
• It repeats completed work.
• It forgets the current acceptance criteria.
• It stops referring to an earlier tool result that the summary should preserve.

The API call still succeeds. There is no context-window error.

The cause

Compaction replaces older history with a summary inside a compaction content block.

Many message loops keep only visible text:

# Wrong: discards compaction and other non-text blocks.
assistant_text = "".join(
    block.text for block in response.content
    if block.type == "text"
)
messages.append({"role": "assistant", "content": assistant_text})

That code works until the API returns a non-text block that carries state. The next request contains readable prose but no compaction summary.

The model did not forget the state. The client removed it.

The fix

Append the full content array:

messages.append({
    "role": "assistant",
    "content": response.content,
})

In TypeScript:

messages.push({
  role: 'assistant',
  content: response.content,
});

Do not serialize each block into a custom text format. Preserve the typed objects that the SDK returned.

Add a boundary test

Log only block types, not private content:

print([block.type for block in response.content])

When compaction occurs, the output includes compaction before later response blocks.

Add an assertion to your loop test:

returned_types = [block.type for block in response.content]

if "compaction" in returned_types:
    messages.append({"role": "assistant", "content": response.content})
    saved_types = [block.type for block in messages[-1]["content"]]
    assert "compaction" in saved_types

The test proves that your storage layer did not flatten the response.

Check every storage boundary

The in-memory append can be correct while persistence still breaks it.

Inspect these points:

1. The SDK response object.
2. The object written to your database or session file.
3. The object loaded for the next request.
4. The final messages payload.

At each point, the compaction block must remain typed and unchanged.

Do not patch the symptom with a larger context window

A larger window delays the boundary. It does not fix a message loop that deletes state.

Repair the round trip first. Then choose a compaction trigger that matches the workload.

Related context guides

• Why Your Claude Code Context Window Fills Up
• Restart Claude Code Sessions at 200K Tokens
• Your Prompt Cache Is Not Working