[RGB-loop]

Problem

The public contract lets a function result omit output and error or set them to null (contracts/agents-api/execution-tools.md). Core admits {"success":true,"output":""}, "output":[] and "error":"", and maps them to empty function result content. With the Claude Harness, those results fail the whole Turn with execution_failed.

The pinned native harness (@anthropic-ai/claude-agent-sdk 0.3.269) does not record a blank tool result as submitted:

submitted function result	native tool_result content
success, [] / "" / whitespace only	(mcp__functions__<name> completed with no output)
error, []	Unknown error
error, ""	Command failed with no output
error, "x" + "", or "boom\n"	joined text, trimmed (x, boom)

FunctionBridge.consume requires the native block to equal the submitted content, so it throws Native function response differs from submitted result. although the native harness did apply the result.

Change

Adapter code only; Core admission and the public contract are unchanged.

• A result with no image and no non-blank text is confirmed by a native text tool_result with the same is_error. The native harness owns how it renders a blank result.
• Error text is compared after the same join("\n").trim() the native formatter applies.
• Every other result keeps exact matching, and an is_error mismatch still fails.
• packages/claude-sdk-adapter/README.md (the bridge's owning doc) states both receipt rules.
• The OpenAI compatibility ledger (contracts/agents-api/index.md, with the zh/ mirror and source_hash) records the model-visible difference: Claude Code shows its own placeholder for a blank result and trims error text.

Semantics and alternatives (CONTRIBUTING: raise semantics with a recommendation)

1. This PR: accept the native placeholder by outcome for blank results. This keeps receipt confirmation for every result that has content and does not depend on the SDK's wording.
2. Match the exact pinned placeholder strings. This is stricter, but it breaks on SDK pin upgrades and depends on the tool name. Not chosen.
3. Reject blank function results at Core admission for the Claude profile. This would narrow the public contract because of an engine limitation. Not chosen.

Validation

• packages/claude-sdk-adapter:
  • pnpm typecheck passes.
  • pnpm test passes (build plus all adapter tests).
  • The new function_bridge.test.mjs case covers six blank results (including a whitespace-only error) with the captured native strings. For each, a flipped is_error is rejected. It also covers trimmed error text (trailing empty part, trailing newline, leading space, leading empty part, empty part in the middle), and checks that non-blank success and error mismatches still throw.
  • On the baseline, the new test fails and the existing tests pass.
• make check-names passes. git diff --check is clean.
• The zh/index.md source_hash matches the new English ledger. Checked with an offline subset of website/tests/translations.test.mjs (source, hash, fences, list items, inline literals); the full website test was not run because its dependencies were not installed locally.
• Fixture end to end, not live:
  • Setup: the built adapter (dist/main.js) with its pinned native SDK binary (linux-x64), against a local fake Anthropic Messages API on 127.0.0.1, with a temporary HOME/cwd and a fake key.
  • Each case answers one function call with the case's result. 13 cases: 4 controls, 5 blank results, and 4 padded/mixed text results.
  • In the implementer's own runs, the blank and trimmed-error cases ended in execution_failed before this change; after it, all 13 emit function_applied and reach result.
• Independent review: a fresh subagent that received only the requirements, acceptance criteria, boundaries, repository path and base (no implementation summary) reviewed the complete diff against main and found no blocking issues, as AGENTS.md asks.
• Not covered: the native harness also truncates error text longer than about 10,000 characters, and replaces oversized success output with a preview. Exact matching still rejects both. That predates this change and is left for a separate fix.
• Live acceptance was not run. It lacks a prerequisite: we have no Anthropic provider account. So this change has not passed live Harness qualification. The fixture runs above check the native wire behavior with a synthetic model only.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}