fix(server): classify ACP Task tool calls as collab agent tool calls - #8443
fix(server): classify ACP Task tool calls as collab agent tool calls#8443NoahLinckeScout wants to merge 4 commits into
Conversation
A thread on an ACP-backed provider that launched subagents via its Task tool showed zero subagents in the UI. The launches were never dropped: over ACP they arrive as ordinary background tool calls titled "Task: ..." with `rawInput._toolName === "task"` and no agent identity on the wire, so `canonicalItemTypeFromAcpToolKind` filed them under its `dynamic_tool_call` default. Every consumer that makes delegated work visible keys off `collab_agent_tool_call`, so the rows sat in the timeline as anonymous tool noise. The other adapters already classify the same work as `collab_agent_tool_call` (`ClaudeAdapter.classifyToolItemType`, `OpenCodeAdapter`, `CodexAdapter`), so the shared ACP helper was the outlier. Recognize the Task spellings there so the ACP providers get the same classification: `rawInput._toolName === "task"` (case-insensitive) or a title matching `/^task:/i`. All other kinds are unchanged. Identity limits remain: ACP carries no model, agent type, or subagent prompt for these calls, so per-subagent attribution is still not possible. The row now at least reports that it was agent work.
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
github-actions
Bot
added
vouch:unvouched
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c5af99eea9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
ApprovabilityVerdict: Approved at Macroscope's review found this PR approvable — This is a localized classification and UI-rendering fix for existing ACP Task events, with unrelated tool handling preserved and targeted regression tests added. It introduces no schema, deployment, security, or workflow changes. You can add or adjust custom eligibility rules. Learn more. |
…c tools Review on pingdotgg#8443 noted that reclassified ACP Task launches were still invisible on mobile: the work-log icon derivation filed beside under the same hammer, the only presentation branch for either type. Right — the server change gave web bot styling and agent grouping, and mobile still showed anonymous tool noise. The icon now reads the collab type (or a task id) as agent work, matching how the web timeline classifies the same rows as . Other tool kinds keep their existing icons. Verified: apps/mobile suite 112 files passed, 765 tests passed. Server ACP classification tests still pass. Typecheck exit 0. The new test fails against the pre-fix hammer grouping.
…c tools Review on pingdotgg#8443 noted that reclassified ACP Task launches were still invisible on mobile: the work-log icon derivation filed collab_agent_tool_call beside dynamic_tool_call under the same hammer, the only presentation branch for either type. Right — the server change gave web bot styling and agent grouping, and mobile still showed anonymous tool noise. The icon now reads the collab type (or a task id) as agent work, matching how the web timeline classifies the same rows as agent-tool. Other tool kinds keep their existing icons. Verified: apps/mobile suite 112 files passed, 765 tests passed. Server ACP classification tests still pass. Typecheck exit 0. The new test fails against the pre-fix hammer grouping.
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 6b64534. Configure here.
Cursor Bugbot on the previous commit: the collab-agent branch ran before the error-tone branch, so a failed task.completed or terminal task.updated with a taskId rendered as agent instead of alert. Mobile has one icon where web overlays failure after selection, so nested-agent failures lost their only destructive signal. The error tone now wins above the agent branch. Added the failing-then-passing test: a failed collab_agent_tool_call with a taskId keeps the alert icon. Verified: apps/mobile suite 112 files passed, 766 tests passed.
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
1 participant
What Changed
canonicalItemTypeFromAcpToolKindinAcpCoreRuntimeEvents.ts(the helper shared by the Cursor and Grok adapters) now recognises Task tool calls and classifies themcollab_agent_tool_callinstead of falling through to itsdynamic_tool_calldefault.A Task call is recognised by
rawInput._toolName === "task"(case-insensitive) or a title matching/^task:/i. Every other kind is unchanged. One helper plus its single call site, and tests.Why
A thread on an ACP-backed provider launched several subagents through its
Tasktool. All of them ran to completion. The timeline showed them as anonymous tool rows, indistinguishable from ordinary tool noise.The launches are never dropped. Over ACP a Task launch arrives as a plain background tool call:
session/update→tool_call/tool_call_update,kind: "other"title: "Task: ..."rawInput: { _toolName: "task" },rawOutput: { isBackground: true }canonicalItemTypeFromAcpToolKindkeys only offkind, and has no Task branch, sokind: "other"lands in thedynamic_tool_calldefault.This is an inconsistency between adapters, not a new policy. Every other adapter already classifies the same work as
collab_agent_tool_call:ClaudeAdapter.classifyToolItemType—normalized === "task" || normalized.includes("agent") || ...OpenCodeAdapter—normalized.includes("task") || normalized.includes("agent") || ...CodexAdapter—type.includes("collab")The shared ACP helper was the only one without it, so Cursor and Grok were the odd ones out. This brings them in line.
Repro: start a thread on an ACP-backed provider (Cursor or Grok), ask it to delegate something via its Task tool, and watch the timeline. Before this change the launches render with the generic tool icon;
projection_thread_activitiesshowstool.completedrows withitemType: dynamic_tool_calland aTask: ...summary.Known limitation
ACP carries no model, agent type, or subagent prompt for these calls, so per-subagent attribution still is not possible for these providers. The row now at least reports that it was agent work.
UI Changes
No UI code is touched, but the reclassification does change how these rows render, via two existing call sites:
MessagesTimeline.tsx:2475— icon goes fromhammer(dynamic_tool_call) tobot(collab_agent_tool_call)MessagesTimeline.logic.ts:413— group summary kind goes fromdynamic-tooltoagent-toolI have not attached before/after screenshots: reproducing the row requires a live Cursor or Grok session that actually delegates through its Task tool, which I could not capture cleanly. The rendering delta above is the whole of the visible change, and it comes entirely from those two existing branches rather than from anything in this diff. Happy to add screenshots if you want them before reviewing.
Tests
Added to the existing
AcpCoreRuntimeEvents.test.ts: three Task spellings (both recognizers, plus one with no title) assertcollab_agent_tool_call, with negative controls (_toolName: "mcp__x"→dynamic_tool_call,kind: "search"→web_search) so the new branch cannot swallow unrelated calls.Verified red/green rather than assuming: with the call site reverted to
canonicalItemTypeFromAcpToolKind, the new test fails withexpected collab_agent_tool_call, received dynamic_tool_call; with the fix it passes.Verified locally:
vp test run src/provider/acp/AcpCoreRuntimeEvents.test.ts→ 4 passedapps/serversuite → 245 files passed, 2816 passed / 10 skipped (baseline on this commit's parent: 245 files, 2815 passed)vp run typecheck→ exit 0vp fmt --check→ clean;vp lintreports nothing new for the changed filesChecklist
Note
Low Risk
Narrow classification and presentation change for a specific tool shape; downstream impact is showing agent work correctly rather than altering auth or data paths.
Overview
ACP Task tool launches from Cursor/Grok were emitted as generic
dynamic_tool_callrows because classification only looked at toolkind. The server now detects Task calls (rawInput._toolNamecase-insensitively or atask:title) and emitscollab_agent_tool_call, aligning ACP with other adapters.On mobile, work-log icons follow that type:
collab_agent_tool_calland task rows with ataskIduse the agent icon instead of the hammer;dynamic_tool_callstays hammer. Error tone still maps to alert so failed delegated tasks stay visibly failed.Tests cover ACP classification (including negatives) and the mobile icon rules.
Reviewed by Cursor Bugbot for commit 24764ff. Bugbot is set up for automated code reviews on this repo. Configure here.
Note
Classify ACP Task tool calls as
collab_agent_tool_callin event mapper and mobile icon logicrawInput._toolName == "task"or title starting withtask:) now emititemTypecollab_agent_tool_callinstead of the kind-based classification in AcpCoreRuntimeEvents.ts.workEntryIconutil in threadActivity.ts now returns theagenticon forcollab_agent_tool_callentries and any entry carrying ataskId; entries with error tone still returnalert.collab_agent_tool_callentries that previously showed thehammericon now showagent;dynamic_tool_callstill maps tohammer.Macroscope summarized 24764ff.