test(xai): cover encrypted Responses tool continuation - #7975

Summary

• Follow up on the already merged fix in fix(openai-responses): drop the include entry xAI rejects #7911 (392b786bcfee4078caa575e3fd0dee5b68eaf290); this PR does not duplicate its implementation. The fix is present on dev but absent from the installed 2.94.0 release where the failure was observed.
• Add eight regression cases to the existing xAI Responses include tests. Exercise Grok 4.7/xhigh with encrypted reasoning and paired function-call replay on the public, US regional and Grok CLI OAuth destinations. Verify duplicate-selector removal, retained supported selectors/tools/state, caller immutability, absent/empty/non-array input, and unchanged OpenAI/unknown/lookalike/nonstandard-port destinations.
• Make the test helper use the actual fixture model, assert its synchronous request-builder contract, and use the typed key auth-mode value in API-key fixtures. No runtime, routing, authentication, dependency or documentation behavior changes.

Verification

Validated against dev base 3ea89653f5727aea41948ddbd5e833f74afe3e36, using Bun 1.4.4 and frozen root dependencies.

• Sensitivity check: temporarily disable only the existing stripXaiUnsupportedInclude call in the isolated checkout, then run bun test tests/providers/xai/xai-responses-adjacency.test.ts -t 'preserves Grok 4.7 encrypted tool continuation': 3 failed, each retaining output.encrypted_content. Restore the production source byte-for-byte; its SHA-256 remains 1d7c6e51ec03092782e8c7a880da566bb0c0b32151b5a906b18b25b7fbd44a40. No runtime source change is in this PR.
• bun test tests/providers/xai/xai-responses-adjacency.test.ts tests/providers/xai/xai-transport.test.ts tests/responses/responses-xai-request-compat.test.ts --timeout 30000: 106 passed, 0 failed, 385 assertions.
• bun run test:changed: 20 passed, 0 failed, 104 assertions; selected the one changed test file against origin/dev.
• bun run typecheck: passed, 2,243 files.
• bun run structure:check, bun run privacy:scan, git diff --check: passed.
• Full-suite scope exception: this is a one-file test follow-up to an existing runtime fix. Focused adapter/transport coverage plus changed-mode execution were used instead of all 2,810 test files on the shared machine. Full and cross-platform suites were not run locally; no manual workflow dispatch was made. Repository CI remains separate evidence.

Bounded native Codex/Grok evidence

This evidence comes from an earlier, owner-authorized native Codex run on the affected installed OCX 2.94.0 runtime, after applying a minimal local backport of the same exact-entry xAI filter. It is not a fresh inference run against this PR head or the full current dev tree. The local backport retains an empty array when the rejected selector is the only entry; upstream #7911 omits that field, and the upstream tests retain its omission contract.

Before repair, the native hard worker's first physical request failed before any tool call:

request: ocx-399bbed9700070f41ee17167aa6326d6
served target: xai/grok-4.7, xhigh, openai-responses, Grok OAuth
HTTP 400: Argument not supported: "output.encrypted_content" in "include" field.
physical sends: 1; tool calls: 0

After the local repair and a managed service restart, one fresh native lazycodex-worker-high used the unchanged combo/omo-quota-deep/xhigh role, fork_context:false, with no model or effort override. Native pre/post spawn admission and attachment were recorded; the worker read a synthetic owned fixture, wrote its exact nonce and doubled numeric value, and continued to terminal output after tool results.

Total 3 physical sends, 2 real tool calls, 13.828 seconds. The parent independently checked the saved result and collected actual child request IDs, served targets, efforts and physical attempts. The completion receipt was accepted, the child closed, and unused allowance retired. Bounds were one worker, six requests, four tools and 120 seconds. All 128 captured routing/role/certificate/controller fingerprints remained unchanged.

Retained local native completion receipt SHA-256: 06cc447c175818e440a75a5819fab5cde0d8c3ed37e9e4cab4940679ea11b2e5. Applied local adapter SHA-256: 174dd7090783dbbba774cf7d706a5f6f98465b7ba77a3402cae28ecc648514da.

Accounting limit: the external watchdog saw three logical responses while the history index still had zero indexed physical sends; the collector subsequently verified all three actual physical attempts. This proves bounded native request/tool continuation, not live physical-send hard-cap enforcement, provider outages, load/concurrency behavior or broad model quality. No credentials, account identifiers, private source or raw conversation transcripts are included here.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed: no user-facing behavior changes; fix(openai-responses): drop the include entry xAI rejects #7911 already documents the runtime fix.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. Only synthetic test fixtures and a sanitized native evidence summary are included.

Current review and CI disposition

CodeRabbit completed its review of 2745922 and raised one valid test-coverage finding: the helper copied the top-level request before passing it to the adapter, so caller assertions could miss top-level mutations. Fixed in 9aac551a9d0bfc40c7b9d468bd8925f0e64378f4: reuse model-bearing input as the actual _rawBody, preserve default-model fallback, and assert the same request object's immutability inside the helper. Thread reply and validation; the thread is resolved.

An isolated adapter-side top-level mutation caused all three Grok continuation cases to fail at the new assertion. Restoring production source byte-for-byte made all three pass. No runtime source change is committed. Final changed-head validation: focused 106 passed / 385 assertions, changed-mode 20 passed / 104 assertions, typecheck 2,243 files, structure/privacy/diff checks passed. The documented full-suite exception remains unchanged. Earlier native Grok evidence above remains explicitly scoped to its original local-backport run; no new inference was performed.

The fork CI approval requirement remains external. The prior exact-head CI run was action_required with zero jobs; new-head CI must be read separately and is not inferred from that older run. No workflow approval, manual dispatch, CI success, merge or release is claimed. CodeRabbit may run incrementally on the new commit; the resolved finding does not claim completion of a future review.

Review readiness checklist

• Required local validation passed; commands, results, and any full-suite exception are documented.
• I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).
• I resolved all correct Codex and CodeRabbit findings.
• My PR is ready for review.

Summary by CodeRabbit

• Tests
  • Expanded coverage for xAI request handling across supported destinations, including how include options and other request details are preserved.
  • Added checks that requests to non-xAI destinations retain their existing include behavior, including on lookalike hosts and non-default ports.
  • Verified that missing, empty, or non-array include values do not gain an invented option, and that request inputs remain unchanged.