Walkthrough

The embedded Python inside check_trajectory_acceptance in the E2E test script is updated so that source_calls is built by iterating over all assistant messages containing tool-call blocks, rather than extracting content only from the last such message.

Changes

Trajectory Acceptance Tool-Call Aggregation

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

Poem

🐇 One message was never enough for me,
I hopped through each assistant turn to see,
All tool calls collected, none left behind,
The trajectory checks now cover each find.
✨ A loop replaces a single peek — hooray!

❌ Failed checks (1 warning)

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

E2E Recommendation Advisor

Base: origin/main
Head: HEAD
Confidence: high

Required E2E

• None. No merge-blocking E2E is required because this is a tests-only change and does not affect NemoClaw runtime or user flows. Running the touched Kimi compatibility E2E is still useful to validate the updated test acceptance logic.

Optional E2E

• kimi-inference-compat-e2e (medium): Optional confidence check for the modified E2E script itself. This job runs test/e2e/test-kimi-inference-compat.sh and validates Kimi K2.6 safe exec splitting through OpenClaw trajectories, including the acceptance logic changed by this PR.

New E2E recommendations

• None.

Dispatch hint

• Workflow: .github/workflows/nightly-e2e.yaml
• jobs input: kimi-inference-compat-e2e

Vitest E2E Scenario Advisor

Base: origin/main
Head: HEAD
Confidence: high

Required Vitest E2E scenarios

• None. Only a legacy test/e2e shell test changed outside test/e2e-scenario/ and the Vitest scenario workflow; this does not affect Vitest scenario behavior or dispatchable Vitest scenario coverage.

Optional Vitest E2E scenarios

• None.

Relevant changed files

• None.

🛠️ Needs attention

• None.

🔎 Worth checking

• Source-of-truth review needed: Kimi trajectory acceptance verifier: The advisor marked localized patch analysis as needs_followup.
  • Recommendation: Identify the invalid state, source boundary, source-fix constraint, regression test, and removal condition before merging the localized behavior.
  • Evidence: The diff changes last assistant tool-call message inspection to aggregating all selected messages, but does not filter the flattened blocks to `type == "toolCall"`.
• Filter flattened source blocks to toolCall entries (test/e2e/test-kimi-inference-compat.sh:652): The new loop extends each selected assistant message's entire content array. Because messages are selected when they contain at least one toolCall, any valid assistant message that also includes text, reasoning, or another non-tool content block will add an entry whose command is None, making the strict source_commands == ["hostname", "date", "uptime"] assertion fail even though the actual tool calls are correct.
  • Recommendation: When flattening across assistant messages, append only blocks with block.get("type") == "toolCall" before deriving source_commands. A small fixture/helper test for mixed text-plus-tool-call content would make the intended acceptance rule explicit.
  • Evidence: The prior predicate already checks any(block.get("type") == "toolCall" ...), but the new code does source_calls.extend(message.get("content", [])) without applying that same block-type filter.

🌱 Nice ideas

• None.

• **Acceptance clause:** No deterministic linked issue clauses or linked issue comments were available in the validation context. — add test evidence or identify existing coverage. validation_context.linkedIssues is empty. The PR body says it is related to fix(e2e): stabilize nightly recovery coverage #5401, but no linked issue clauses/comments were provided by the deterministic context to map literally.
• **Acceptance clause:** Flatten tool-call content from all assistant tool-call messages in `test/e2e/test-kimi-inference-compat.sh` before checking the expected `hostname`, `date`, `uptime` order. — add test evidence or identify existing coverage. The diff now aggregates content from all assistant messages with tool calls, replacing last-message-only behavior. However, it flattens all content blocks from those messages rather than only tool-call blocks.
• **Kimi trajectory acceptance verifier** — No dedicated fixture or helper test was added for split multi-turn source commands or mixed content-plus-tool-call messages; coverage currently depends on running this E2E script against live/mock behavior.. The diff changes last assistant tool-call message inspection to aggregating all selected messages, but does not filter the flattened blocks to `type == "toolCall"`.