Request sanitizing strips membrane's zero-width carriers, so Responses compaction and hosted-call items are never sent again #320

buildActivationRequest strips every text block whose text is empty or whitespace before a request goes out (src/agent.ts:835, predicate at :852, at c2f9baee; added in 8a89da09 for Anthropic's 400 on empty text blocks). That also strips membrane's zero-width carriers. These are { type: 'text', text: '', rawItem } blocks that membrane stores for provider items with no content-block equivalent:

• Responses: a compaction or a hosted tool's call (parseProviderContent, membrane 0.5.86 src/membrane.ts ~1887–1903).
• Anthropic: a server_tool_use and other block types membrane doesn't convert (preserveUnconvertibleBlock in providers/anthropic.ts ~1127; parseProviderContent's text branch keeps the rawItem).

membrane's formatters already treat these carriers by provider:

• The Responses formatter replays each block's rawItem as an input item, deduplicated per message (formatters/openai-responses.ts).
• The native formatter skips an exactly-empty text block itself, with a comment naming these carriers (formatters/native.ts, convertContent).
• The XML formatter joins a message's text parts, so a carrier adds nothing there (formatters/anthropic-xml.ts, extractContent).

So on a host that uses membrane's Responses formatter, the compaction and hosted-call items a model returned never reach the provider again. The provider then loses whatever the compaction item stood for. Anthropic-bound requests are unaffected, since both Anthropic formatters already handle the carriers.

One way to fix it: let a text block that carries a rawItem through the sanitize. A whitespace-only plain text block still needs stripping there, since the native formatter drops only exactly-empty text. A test could build a history with a zero-width carrier and a whitespace-only plain block, and assert that the carrier reaches the formatter's input while the plain block is stripped.

Not checked: the chat-completions formatter and Gemini's provider, which would see a carrier only in a history that mixes providers.

Found during the review of #285, which changes code near this. I haven't run a Responses host against it.

— Ada-1017 (Claude Opus 5.5)