unbridled-41 opened a new issue, #4736:
URL: https://github.com/apache/rocketmq-dashboard/issues/4736
### Studio Version
branch: rocketmq-studio
git commit id: 7ce9a68215c4cd1c7025013d02135b43b6aa6ba1
### Problem
The manual tool playground keeps an in-flight execution alive across a tool
or scope change, so its response lands in the output pane of whatever tool is
selected afterwards. The pane never names the tool that produced the output, so
tool A's result is presented as tool B's.
Chain: `handleExecuteTool`
(`web/src/pages/ai/components/ToolPlaygroundModal.tsx:200-225`) awaits
`executeTool(selectedToolName, parsedInput, selectedClusterId)` and assigns the
response to `toolResult` with no generation check. Neither invalidation path
touches that call:
- `selectTool` (`:127-135`) clears the panel (`setToolResult(undefined)`)
but does not invalidate the running request, and the tool `Select` (`:267-281`)
stays enabled while `toolExecuting` is true (only the Run button goes into
`loading`, `:247-250`).
- `loadTools` (`:137-161`, reached by a cluster-scope change, `:191-198`)
clears the panel the same way.
- `handleClose` (`:227-233`) bumps only `toolLoadRequestRef` — the catalog
guard — so a response can also arrive after the modal was closed and reopened
(the modal has no `destroyOnHidden`, so its state survives).
Trigger (deterministic): open 工具 on the AI page, run any tool that takes a
few seconds (the RocketMQ tools do), then switch the tool in the selector
before it returns. The panel was cleared by the switch; when the response
arrives it is repopulated, the success toast fires, and the output is shown
under the newly selected tool's name and risk/permission tags (`:283-302`).
### Evidence
- `web/src/pages/ai/components/ToolPlaygroundModal.tsx:200-225` —
`setToolResult(await executeTool(...))` with no request-generation check; the
only guard is `if (!selectedToolName || toolExecuting) return` at the top,
which prevents a second *start*, not a late *landing*.
- `web/src/pages/ai/components/ToolPlaygroundModal.tsx:127-135` and
`:137-161` — the two paths that clear `toolResult` without invalidating the
execute.
- `web/src/pages/ai/components/ToolPlaygroundModal.tsx:227-233` —
`handleClose` bumps `toolLoadRequestRef` only.
- `web/src/pages/ai/components/ToolPlaygroundModal.tsx:318-342` — the output
pane is labelled with the generic `ai.tool.output` and renders
`formatToolResult(toolResult)`; nothing ties the text to the tool that produced
it.
- The same file documents and implements the analogous guard for the catalog
load ("Stale catalog guard", `:40-43`), so the pattern is established in this
component; the execute path is the one call that never got it.
- Regression test added with this report (fails on 7ce9a682): two tools in
the catalog, `executeTool` returns a deferred promise, run the first, switch to
the second in the selector, then resolve — `queryByTestId('tool-result')` must
be absent, but the stale output is rendered.
### Impact
An operator who uses the playground to confirm what a transcript's tool call
actually returns can read a different tool's output under the selected tool's
name and risk level. Because the pane is the "human-driven escape hatch" for
exactly that comparison, a misattributed result is worse than a missing one: it
looks like a successful manual reproduction of a call that was never made.
### Expected behavior
A response for a tool or cluster scope that is no longer the one on screen
is discarded — no result and no success toast — while the Run button's loading
state still clears. Output that does land must belong to the tool currently
selected, in every path that changes the selection or closes the modal.
### Related work
- #4609 (PR #4615) and #4612 (PR #4618) — a request's failure and a stale
response respectively, rendered as if they were the current state: same
"response outlives its context" family.
- #4585 (PR #4586) — a superseded comparison overwriting a newer result.
- #4606 (merged, 5b761df8) — introduced this component.
#
## PR
Fix: TBD.
--
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]
