Summary

What problem does this PR solve?

• Transcript repair was treating any tool call with non-null arguments as complete, then stripping partialJson without checking whether the remaining block was a finalized OpenAI Responses call or an interrupted Anthropic artifact.

Why does this matter now?

• That leaves interrupted Anthropic artifacts replayable in repaired history, which can corrupt later tool-result repair and preserve bad session state.

What is the intended outcome?

• Drop interrupted Anthropic partialJson artifacts, keep finalized OpenAI Responses tool calls, and preserve retained sessions_spawn attachment content on kept blocks.

What is intentionally out of scope?

• This PR does not claim to fix the upstream Kimi transport ordering problem itself, and it does not change provider transport handling outside transcript repair.

What does success look like?

• Interrupted Anthropic artifacts are removed during transcript repair, finalized OpenAI Responses calls remain usable after partialJson stripping, and existing sessions_spawn attachment preservation still works.

What should reviewers focus on?

• The keep/drop split in repairToolCallInputs() and the matching regression coverage in src/agents/session-transcript-repair.test.ts.

Linked context

Which issue does this close?

• None claimed.

Which issues, PRs, or discussions are related?

• Related [Bug]: Kimi k2p5 (anthropic-messages): "Unexpected event order: message_start before message_stop" — chat unusable #57523
• Supersedes fix(agents): drop partialJson streaming artifacts from session history repair #61151

Was this requested by a maintainer or owner?

• No.

Real behavior proof (required for external PRs)

• Behavior or issue addressed: interrupted Anthropic tool-call artifacts that still have id, name, arguments: {}, and partialJson should be dropped, while finalized OpenAI Responses tool calls should keep valid arguments, lose partialJson, and preserve retained sessions_spawn attachment content.
• Real environment tested: local OpenClaw checkout on Windows 11 with Node 22.
• Exact steps or command run after this patch:
  1. Ran a temporary TypeScript proof script through pnpm exec tsx that imports sanitizeToolCallInputs() from src/agents/session-transcript-repair.ts.
  2. Fed that script one finalized OpenAI-style sessions_spawn block (call_spawn|fc_456) plus one interrupted Anthropic-style block (toolu_123 with arguments: {} and partialJson).
  3. Printed the repaired assistant content from the current PR head.
• Evidence after fix (screenshot, recording, terminal capture, console output, redacted runtime log, linked artifact, or copied live output):

{
  "droppedAnthropicArtifact": true,
  "assistantContent": [
    {
      "type": "toolCall",
      "id": "call_spawn|fc_456",
      "name": "sessions_spawn",
      "arguments": {
        "attachments": [
          {
            "content": "secret data",
            "name": "a.txt"
          }
        ]
      }
    }
  ],
  "sessionsSpawnAttachmentContent": "secret data"
}

• Observed result after fix: the interrupted Anthropic artifact was removed, the finalized OpenAI-style sessions_spawn block stayed, partialJson was removed from the retained block, and attachment content was preserved.
• What was not tested: a live remote-provider replay round-trip against Anthropic or OpenAI Responses in this environment.
• Proof limitations or environment constraints: no live provider credentials or remote replay harness were available in this environment, so proof is limited to direct transcript-repair execution plus focused regression tests.
• Before evidence (optional but encouraged): the previous branch behavior kept the initialized Anthropic shape after deleting partialJson, which is the blocker called out in prior review.

Tests and validation

Which commands did you run?

• node scripts/run-vitest.mjs src/agents/session-transcript-repair.test.ts --run
• node scripts/run-vitest.mjs src/agents/session-transcript-repair.attachments.test.ts --run
• pnpm check:changed

What regression coverage was added or updated?

• src/agents/session-transcript-repair.test.ts now locks in the keep/drop split between finalized OpenAI Responses tool calls and interrupted Anthropic partialJson artifacts.

What failed before this fix, if known?

• The initialized Anthropic id + name + arguments: {} + partialJson shape survived cleanup as a replayable tool call after partialJson removal.

If no test was added, why not?

• N/A

Risk checklist

Did user-visible behavior change? (Yes/No)

• Yes. Transcript repair now drops interrupted Anthropic partialJson artifacts instead of replaying them as valid tool calls.

Did config, environment, or migration behavior change? (Yes/No)

• No.

Did security, auth, secrets, network, or tool execution behavior change? (Yes/No)

• No.

What is the highest-risk area?

• Misclassifying a valid non-OpenAI tool call that still carries partialJson.

How is that risk mitigated?

• Current source only persists partialJson on Anthropic streaming artifacts and finalized OpenAI Responses blocks, and the updated regression tests pin the intended split and preserved sessions_spawn payload path.

Current review state

What is the next action?

• Wait for CI and maintainer review on the fresh replacement PR.

What is still waiting on author, maintainer, CI, or external proof?

• Waiting on GitHub CI and maintainer review. No additional author-side code change is planned unless CI or maintainer feedback finds a concrete regression.

Which bot or reviewer comments were addressed?

• Reworded the PR body to remove the over-claiming fully tested footer.
• Narrowed #57523 from a closing claim to related context only, because this PR repairs transcript fallout rather than claiming to fix the upstream Kimi transport ordering bug itself.

AI-assisted: drafted with Copilot.