apps/server/src/orchestration/workflowScriptQuery.ts (1)

25-27: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Record the provider decision for the hardcoded Claude scripts root.

scriptsRoot() returns ~/.claude/projects. Workflow scripts produced by other adapters resolve outside that root and fail with outside-root. This PR also adds Codex workflow observability, so the Agents "{} script" affordance will not work for Codex runs.

State the intended behavior for Codex, Cursor, Grok, and OpenCode. If only Claude persists workflow scripts today, document that in the module header so the limitation is explicit.

As per coding guidelines: "Provider-shaped features require a deliberate decision for each provider adapter: Codex, Claude, Cursor, Grok, and OpenCode."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/server/src/orchestration/workflowScriptQuery.ts` around lines 25 - 27,
Document the provider decision in the module header for the hardcoded
scriptsRoot() path: explicitly state the intended workflow-script behavior for
Codex, Claude, Cursor, Grok, and OpenCode, including that only Claude currently
persists scripts if that is the supported behavior. Ensure the documentation
makes clear which providers support the Agents “script” affordance and avoids
implying that all adapters use ~/.claude/projects.

Source: Coding guidelines

apps/server/src/orchestration/workflowScriptQuery.test.ts (1)

10-15: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Do not write test fixtures into the real home directory.

Lines 10-13 create files inside the developer's or CI runner's actual ~/.claude/projects, which is the same directory the Claude harness uses. Line 14 uses a fixed name in os.tmpdir(). Two concurrent runs collide, and afterAll deletes state under the real home directory.

Make the scripts root injectable in workflowScriptQuery.ts (for example an optional root parameter that defaults to scriptsRoot()), then point the test at a mkdtempSync directory. This also removes the need for the afterAll cleanup of real user state.

Also applies to: 23-26

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/server/src/orchestration/workflowScriptQuery.test.ts` around lines 10 -
15, The test currently writes fixtures into the real home directory and uses a
colliding fixed temporary filename. Make the scripts root injectable in the
relevant workflow script query API, defaulting to scriptsRoot(), then have the
test create an isolated mkdtempSync directory and place both fixtures under it.
Update calls and cleanup to use only this temporary directory, removing afterAll
cleanup of real user state.

apps/web/src/rightPanelStore.ts (1)

52-52: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Consider keeping the storage version at 7.

The change adds one surface kind. migratePersistedRightPanelState already tolerates the previous shape, because it only rewrites file and terminal surfaces and passes other kinds through. Raising the version to 8 discards every user's open right-panel tabs on upgrade without a correctness benefit.

If you keep version 8, that is a deliberate reset. If not, revert to 7.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/web/src/rightPanelStore.ts` at line 52, Revert
RIGHT_PANEL_STORAGE_VERSION to 7 because migratePersistedRightPanelState already
preserves the newly added surface kind and no reset is required.

apps/web/src/components/chat/MessagesTimeline.logic.ts (1)

531-533: 🎯 Functional Correctness | 🔵 Trivial | 💤 Low value

Compute onlyToolEntries from the hidden entries.

The toggle describes the hidden rows. visibleGroupedEntries includes the pinned spawn entries, which are never hidden. A spawn entry that is not tool-like flips onlyToolEntries to false even when every hidden row is a tool call, which changes the toggle label.

♻️ Proposed change

-              onlyToolEntries: visibleGroupedEntries.every((entry) =>
-                workLogEntryIsToolLike(entry),
-              ),
+              onlyToolEntries: hiddenEntries.every((entry) => workLogEntryIsToolLike(entry)),

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/web/src/components/chat/MessagesTimeline.logic.ts` around lines 531 -
533, The onlyToolEntries calculation currently uses visibleGroupedEntries, which
includes pinned spawn entries; update the logic around onlyToolEntries to
evaluate the hidden entries collection instead, while preserving the existing
workLogEntryIsToolLike predicate and toggle behavior.

apps/web/src/session-logic.ts (2)

652-672: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Move the quiet-timeline doc block onto deriveWorkLogEntries.

The comment at Lines 652-661 documents the work-log filtering contract of deriveWorkLogEntries. It now sits directly above isAgentTaskStartedActivity, which carries its own doc comment on Line 662. A reader (and any doc tooling) attaches the block to the wrong function.

Move the block to Line 712, immediately above deriveWorkLogEntries.

As per coding guidelines: "Use comments mainly to describe how a function is used; avoid annotating every line of behavior, and move comments when code moves."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/web/src/session-logic.ts` around lines 652 - 672, Move the
quiet-timeline guarantee doc block from above isAgentTaskStartedActivity to
immediately above deriveWorkLogEntries, preserving its text unchanged; keep the
separate isAgentTaskStartedActivity documentation directly above that function.

Source: Coding guidelines

---

882-895: 🗄️ Data Integrity & Integration | 🔵 Trivial | ⚡ Quick win

Carry parentAgentId into the derived entry and key by it.

toDerivedWorkLogEntry stores taskId and agentSpawn.workflowId, but not parentAgentId. If parentAgentId exists without a legacy :wf: encoding, agentSpawnGroupKey falls through to direct:<turnId> and splits the same workflow members into separate chat CTA rows from what the Agents panel groups under one coordinator.

Also update the :wf:-slot check only after copying the parent link so this only affects unencoded workflow members.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/web/src/session-logic.ts` around lines 882 - 895, Update
DerivedWorkLogEntry and toDerivedWorkLogEntry to preserve parentAgentId, then
update agentSpawnGroupKey to use the copied parentAgentId for workflow grouping
when no legacy :wf: taskId encoding exists. Keep the existing :wf: parsing
precedence unchanged, and apply the parent-link fallback before the coordinator
and direct turnId cases.

apps/server/src/provider/Layers/ClaudeAdapter.test.ts (1)

1453-1601: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add focused tests for task_updated and workflow member synthesis.

The two new tests cover interrupt fan-out and model/effort inheritance. Two behaviors with more branching remain untested in this file:

• task_updated: the CLAUDE_TASK_PATCH_STATUS mapping (killed → cancelled, paused → idle), removal from liveTaskIds on a terminal patch, and the endedAt conversion.
• emitWorkflowMemberProgress: the <coordinatorTaskId>:wf:<index> member id, per-index dedupe, and workflowAgentStatus.

Both are backend behavior changes, so the guidelines require focused tests. Do you want me to draft them in the same harness style?

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/server/src/provider/Layers/ClaudeAdapter.test.ts` around lines 1453 -
1601, Add focused harness-style tests in ClaudeAdapter.test.ts for task_updated
and workflow member synthesis. Cover killed→cancelled and paused→idle status
mapping, terminal patch removal from liveTaskIds, endedAt conversion, and
emitWorkflowMemberProgress behavior including <coordinatorTaskId>:wf:<index>
IDs, per-index deduplication, and workflowAgentStatus.

Source: Coding guidelines

apps/server/src/orchestration/ActivityPayloadProjection.test.ts (1)

5-15: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Set the activity kind per test so the task-lifecycle case is really exercised.

The helper hardcodes kind: "tool.completed". The second test is named for task lifecycle payloads, but it still passes a tool activity. If projectActivityPayload branches on kind, this test never reaches the task branch and the regression tripwire described in the file comment does not fire.

💚 Proposed change

-function activity(payload: Record<string, unknown>): OrchestrationThreadActivity {
+function activity(
+  payload: Record<string, unknown>,
+  kind = "tool.completed",
+): OrchestrationThreadActivity {
   return {
     id: "activity-1",
     tone: "tool",
-    kind: "tool.completed",
+    kind,
     summary: "Tool",
     payload,
     turnId: null,
     createdAt: "2026-08-01T10:00:00.000Z",
   } as unknown as OrchestrationThreadActivity;
 }

-    const source = activity({
-      taskId: "task-9",
+    const source = activity(
+      {
+        taskId: "task-9",
...
-      timelineBypass: true,
-    });
+        timelineBypass: true,
+      },
+      "task.started",
+    );

Also applies to: 47-62

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/server/src/orchestration/ActivityPayloadProjection.test.ts` around lines
5 - 15, The activity helper currently hardcodes the tool kind, preventing
task-lifecycle tests from exercising the task branch. Update activity and its
callers so each test supplies the appropriate activity kind, preserving
tool.completed for tool tests and using the task lifecycle kind in the second
test.

apps/server/src/provider/Layers/ClaudeAdapter.ts (1)

4358-4375: 🩺 Stability & Availability | 🔵 Trivial | ⚡ Quick win

Log stopTask failures instead of discarding them.

Every stopTask rejection is swallowed twice: catch: () => undefined and Effect.ignore. A provider that refuses to stop background tasks then produces no signal at all, and users see agents keep burning tokens after pressing Stop. Keep the best-effort semantics, but record the failure.

♻️ Proposed change to keep failures observable

-        yield* Effect.forEach(
-          liveIds,
-          (taskId) =>
-            Effect.tryPromise({
-              // Invoke through the query object: SDK methods rely on `this`.
-              try: () => context.query.stopTask!(taskId),
-              catch: () => undefined,
-            }).pipe(Effect.ignore),
-          { concurrency: 8, discard: true },
-        );
+        yield* Effect.forEach(
+          liveIds,
+          (taskId) =>
+            Effect.tryPromise({
+              // Invoke through the query object: SDK methods rely on `this`.
+              try: () => context.query.stopTask!(taskId),
+              catch: (cause) => cause,
+            }).pipe(
+              Effect.catchCause((cause) =>
+                Effect.logWarning("claude.task.stop-failed", { threadId, taskId, cause }),
+              ),
+            ),
+          { concurrency: 8, discard: true },
+        );

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/server/src/provider/Layers/ClaudeAdapter.ts` around lines 4358 - 4375,
Update the stop-task cleanup around context.query.stopTask to preserve
best-effort per-task execution while logging each rejection or thrown failure
with the taskId and error details. Replace the silent catch/ignore handling
without allowing one failed stopTask call to prevent remaining tasks from being
attempted or the subsequent interrupt from running.

apps/server/src/provider/Layers/CodexAdapter.ts (1)

760-766: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Record the intentional omission for Cursor, Grok, and OpenCode child task events.

This PR adds task.* child events for Codex and Claude only, while Cursor, Grok, and OpenCode do not emit collabAgent/...* child events. Capture that decision in the PR description or a code comment so the provider-level behavior is explicit.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/server/src/provider/Layers/CodexAdapter.ts` around lines 760 - 766,
Document in a nearby comment to mapToRuntimeEvents that Cursor, Grok, and
OpenCode intentionally do not map collabAgent/... child task events because
those providers do not emit them; clarify that task.* child events are currently
supported only for Codex and Claude.

Source: Coding guidelines

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes