fix(kimi): omit explicit empty assistant history on native Responses

[yuanyuanlove]

Summary

• Fix the native Kimi Code Responses request compiler for explicit content-empty assistant history. The observed rejection was Invalid request: the message at position 585 with role 'assistant' must not be empty on a native Responses → Responses request, so a Chat-only fix would miss the affected wire.
• Only final outgoing input for the exact api.kimi.com/coding/v1/responses host/path omits assistant messages whose content is explicitly "", [], or a list of empty text parts. Do not trim whitespace, invent answer text, rewrite the client's saved history, add retries, or change other Responses destinations. Keep refusal/multimodal/unknown content, nonempty annotations and extension fields, tools and reasoning carriers. Empty messages with same-input stored-item references are retained to avoid creating dangling references.
• Close the corresponding translated Chat guard's empty-string gap without dropping serialized reasoning or tool calls; nonempty text and whitespace remain unchanged. Tool-call/result ordering and raw request immutability have regression coverage.
• Separate from native reasoning/cipher replay and Sol hosted-search metadata projection. No provider default, protocol preference, model routing, credential, dependency, diagnostic, deployment or retry changes are included.
• The source item at position 585 was not captured; regressions use synthetic empty-string/empty-array/empty-text cases and an offline strict upstream model, not a published user conversation. This PR fixes those explicit empty shapes; it does not assert that the original long live conversation or every Kimi compatibility problem has been recovered.

Verification

• Fetched and merged upstream dev commit 9b6f4b2632492b3df637dccc009d6a01ed47c85d without conflicts (no rebase). Incoming upstream commits did not change these runtime sites. The PR diff is nine files; no unrelated runtime overlay is included.

• Test-first development of the original empty shapes: 7 pass / 7 fail before implementation, then 18 pass / 0 fail. The added referenced-item regression independently failed before its guard: 19 pass / 1 fail; after the fix all 20 synthetic behavior cases pass, including explicit empty shapes, the synthetic 585th-message rejection, references, multimodal/refusal/unknown content, whitespace, native destination exclusions, tool/reasoning carriers and immutable caller input.

• Final merged tree, pinned Bun 1.4.0: bun test --isolate --parallel=4 --timeout=15000 --changed=origin/dev ./tests/responses ./tests/adapters/openai: 4,133 pass / 1 skip / 0 fail, 165 files and 226,456 assertions. This is bounded import-connected adapter/Responses coverage, not the full repository suite. Test timeout is a runner setting, not a production timeout change. An earlier development batch at a 5-second test budget had timeout/handoff failures; its failed file passed independently on both candidate and unchanged baseline. Those old failures are recorded, not asserted to be baseline source defects; both this final batch and the preceding merged batch at a 15-second budget completed without failures.

• bun test --isolate tests/adapters/openai/openai-empty-assistant-history.test.ts tests/test-layout.test.ts tests/test-layout-tooling.test.ts tests/ci-workflows/file-size-ratchet.test.ts tests/ci-workflows/structure-ssot.test.ts: 99 pass / 0 fail, 826 assertions. Subsequently only the new test's index placement moved next to existing OpenAI entries to avoid shared hunks with the separate Sol PR; the three index/layout/ratchet files passed again (28 pass / 0 fail, 643 assertions). Runtime and behavior tests were unchanged by that placement edit.

• Final bun x --no-install tsc --noEmit, bun run structure:check, bun run privacy:scan: passed. English/Chinese docs build passed with existing compatible dependencies and isolated output/cache: 569 pages and 79,288 internal links checked; no install or lockfile change.

• git diff --check origin/dev and git diff --cached --check origin/dev: passed for the proposed PR. As in upstream itself, comparing the merge to the old HEAD reports an unrelated existing blank line at EOF in tests/providers/github-copilot/github-copilot-auto-runtime.test.ts; that file is unchanged relative to the base and not cleaned up in this PR.

• Independent read-only review raised the reference-preservation case, now fixed and covered; no implementation blocker remained. The runtime boundary is reviewed on the merged tree, not assumed from a pre-merge test result.

• Full repository suite, live Kimi long-history acceptance and cross-platform validation remain outstanding. The PR is submitted as draft, not as a full-CI/review-readiness attestation. No real request body, ciphertext, token or private capture is committed.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. (No credential/auth/workflow/dependency changes; synthetic fixtures only.)

Review readiness checklist

This PR stays in draft until every box below is ticked. Tick all four boxes once the requirements are met:

• 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.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

✅ Deterministic PR hygiene checks passed.

[github-actions]

⏳ DRAFT

• review readiness checklist open (0/4 boxes ticked).

What to do

• Tick all four boxes in the PR description once you're done (currently 0/4).

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.

0/4 boxes ticked.

This PR stays in draft until every box above is ticked.

[Ingwannu]

Source review of b49d4f6b5824a63998f1b82ee4cadb17c2a9f79d against trusted dev 9b6f4b2632492b3df637dccc009d6a01ed47c85d (nine files).

The Responses filter is narrow: exact Kimi Coding host/path, explicit empty message content, an allowlist of harmless message/text metadata, and same-input reference preservation. Unknown fields/parts, nonempty annotations, whitespace, refusals and independent reasoning/tool carriers are not treated as empty. It returns a detached input only when needed and leaves caller history unchanged. In translated Chat the check occurs before flushing pending calls; it drops an empty joined answer only when neither tool calls nor serialized reasoning remain. That matters for the tool-result pairing and reasoning-only cases.

I read the complete 20-case fixture and the production/doc/layout changes. No concrete source-level blocker found in this scope. The position-585 case is correctly labeled synthetic: it demonstrates the stated rejection shape with a strict offline fixture, not reproduction of the reporter's original conversation or a real Kimi response.

Independent execution is not completed. I prepared an exact-head ubuntu-owned worktree and a fresh private-home, CPU75% / memoryHigh1280MiB / max1536MiB / swap0 / tasks64 / finite-deadline unit. Its admission guard waited 60 seconds for foreign Cargo/rustc work and exited 75 before Bun started. This is neither a failing regression nor a passing test receipt; I did not overlap the foreign build or raise limits. The 20-case and broader receipts in the description remain yours, not mine.

Please retain Draft until readiness/required review and current-head CI complete. Cross-platform 38113761756, Catalog 38113761789 and React Doctor 38113761780 are action_required; target and hygiene success alone are not runtime validation. No live Kimi/provider/native-client call, scan, workflow authorization, sponsorship or contributor-branch modification was performed. This review is COMMENTED, not APPROVED.

[lidge-jun]

Carried into #6947 with credit to you (Co-authored-by trailers in the commit and description). The Responses filter for api.kimi.com/coding/v1/responses is unchanged from this PR. The Chat guard is narrower: independent review found that dropping every empty answer on all openai-chat providers would put two user turns side by side on strict-alternation templates such as Mistral's. #6947 drops an empty turn only while tool calls await their results, which keeps your call/result pairing fix, and keeps a standalone blank answer as content: "". The maintainer will close this PR once #6947 merges. Thank you for the precise fixtures.

[lidge-jun]

Carried into #6947 and merged to dev as c3a4cdf, with Co-authored-by credit. The Chat-side guard is narrower than in this PR: dropping every empty assistant turn would break strict user/assistant alternation, so a standalone empty answer keeps its slot. Thank you.