[bingdongni]

Summary

Experiment comparisons can corrupt multibyte agent output when a UTF-8 character spans stream chunks, or parse an incomplete response when the process exits before its output streams close. Decode the streams incrementally and parse after they close.

What changed

• Preserve UTF-8 characters across stdout and stderr chunks.
• Wait for the child process output streams to close before parsing the JSON response.
• Cover split characters and output written after the child process exits with subprocess tests.

How to test

• Linux: node --test test/experimentAgentRunner.test.js test/experimentComparison.test.js (32 passed) and npm test (3,308 passed), CI job.
• Windows: the focused tests passed (31 passed, one inherited-pipe test skipped).
• node --check index.js passed. I did not run bare node index.js because it starts an evolution run.
• The template's full test selection had seven failures in proxyTracePlatformInstall.test.js and spawnReplacementProcess.test.js, both excluded by npm test. The same seven failures occur on the upstream base: platform templates are missing and some source assertions target obfuscated files.

Risk

Medium: an agent that leaves its output pipes open after exiting will now wait for the pipes to close or the existing per-arm timeout. The result shape and command invocation are unchanged.

Harness/evaluator governance

Upstream governance surface: src/experiment/agentRunner.js subprocess output handling for comparison arms.
Downstream EvoX impact: Experiment responses preserve complete UTF-8 output; the AgentResult shape and comparison contract are unchanged.
Rollout-local scope: Applies only when an experiment arm invokes its agent CLI; no live rollout is initiated by this change.
Promotion boundary: The experiment runner behavior changes by default only after maintainer review and merge.
Evaluator mismatch sets: Observation and evidence can now include previously corrupted or late output; action, repair, verification, and belief policies are unchanged.
Non-regression evidence: Real subprocess tests cover split stdout, split stderr, and delayed pipe closure; 3,308 project tests passed on Linux. The full template test selection has seven independently reproduced baseline failures as noted above.
Fix-severity review: medium
Owner approval: Experiment runner maintainer review is requested and remains pending.
Security boundary: The existing command, arguments, environment, host, network, and credential access are unchanged.
Rollback: Revert this commit to restore the previous output handling.
Live promotion: no
Autonomous evaluator self-editing: no

Self-check

• No new runtime dependencies added.
• Tests added for the new behavior; full template suite passes locally (node --test test/*.test.js). The seven baseline failures and CI results are described above.

Related

None.

[evomap-aiqa-bot]

Sentinel 已审核此变更，未发现需要报告的问题。

开发 agent 查询详情： 调用 MCP 工具 sentinel_pr_review_context，传入 owner_repo="evomap/evolver"、pr_number=644、head_sha="078408daf9f776d8b490dbac5cf2af17b3faf91c"，并按条目追加 finding_id 或 finding_key。

由 Sentinel 审核 · 目标提交 078408daf9f7