123123213weqw opened a new pull request, #5819: URL: https://github.com/apache/rocketmq-dashboard/pull/5819
## Why `sendMessage` prepared the agent workspace at admission for every engine alike, and the prepare call has two visible side effects: it throws for a configuration the operator should hear about now (an unusable `workspace-dir`, an instance with no resolvable credential), and an empty result persists the degraded/unbound tools notice as the run's first event. The `http` engine has no agent CLI - `streamHttp` goes straight to the OpenAI-compatible client and `options()`, the only consumer of the preparation, is never built for it. So a plain completion turn could be killed by a RocketMQ-tools configuration it never touches, and an unbound conversation accumulated a "RocketMQ tools unavailable" warning on every chat turn - noise about tools this engine architecturally does not have (`streamHttp`'s own doc: no reasoning, no tools, no resumable session). ## What changed `sendMessage` skips `workspace.prepare` - and the degraded notice - when `AiRunExecutor.isHttpEngine(engine)` says the turn is a completion. The agent engines (`claude-code`, `qoder`) keep both behaviours unchanged. ## Validation Red test first: `anHttpTurnShouldNotPrepareAnAgentWorkspaceItCannotUseTest` fails on `upstream-studio` - a working http turn dies on the workspace exception before generating anything - and passes with the fix, which also asserts no `notice` row lands in a completion turn's timeline. Full `AiRunServiceTest` (24 tests) green. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
