feat(runtime): support Kimi OpenAI protocol

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.

[Nyvo-io]

Summary

• let the existing Kimi Coding Plan connection select Anthropic Messages or OpenAI Chat per model
• add Kimi OpenAI transport handling for reasoning dialect lifetime, explicit empty reasoning, tool replay, and usage capture
• preserve provider-owned reasoning metadata across RuntimeEvent and missing-ledger StoredMessage recovery
• keep unsupported or unmarked private thinking out of provider replay without dropping paired tool calls and results
• preserve recovered assistant step identity so reasoning, text, and tool calls reconstruct as one provider message

Context

This is the flat Runtime protocol PR requested in the maintainer review on #1451. Request-trace evidence and the benchmark harness are intentionally excluded and will be submitted separately.

Refs #1269.

Verification

• Biome format: 1,084 files passed
• Biome lint: 2,126 files passed
• all workspace typechecks passed
• git diff --check passed
• TDD regression reproduced the split recovered step before the fix and passed after it
• Kimi/Anthropic product protocol suite: 11 passed
• focused backend/backfill suites: 170 passed
• full npm test: all workspaces passed
• independent headless runner: 1,328 passed, 2 skipped
• local review-agent full base-diff review: no actionable correctness findings

No paid Kimi or Harbor benchmark was run; benchmark execution remains in the later benchmark PR.

[Astro-Han]

One P2 recovery issue is noted inline. The rest of the change uses the existing Kimi provider and Runtime replay path.

[Astro-Han]

Reviewed c38ecf7. When ledger events are missing, the replay plan now uses providerEventId ?? storedMessageId as the assistant step ID.

The regression test follows the full path from backfill through backend reconstruction to the final OpenAI-compatible request. It covers assistant text, empty reasoning, the tool call, and its result.

All required checks pass. No findings. Approving.