[tulsi-builder]

What

• Activity under a message: choosing an agent's row in the activity pill now opens that conversation's transcript. Before, it opened the profile's Activity overview. This matches the composer's Working now bubble (feat(agents): open the latest transcript from agent activity #1830, fix(agents): open the transcript again from the Working now popup #1990).
• Sidebar Working now: choosing an agent opens its transcript when the conversation is known: a Solo chat, or a request whose message is loaded. Otherwise it falls back to the plain Activity destination.
• One Activity tab per agent: asking for a transcript opens the agent's one Activity destination and carries the request alongside it, so an Activity tab that is already open is reused. Before, it opened a second tab. A reused tab on another section, such as Identity, switches to Activity through the same unsaved-change guard as choosing the tab. This also fixes the same duplicate from the composer bubble.
• One way in per agent: both activity popups (the bubble under a message and the composer's working preview) lose their separate "Activity ↗" heading, which opened the same transcript as the agent's row. Each agent's row is now the click target. In the working preview, an agent's row opens its Activity on the conversation's transcript wherever the work is. Before, from the timeline it opened the thread. Work in several threads at once still opens Activity without a transcript, and when Activity can't open, the row falls back to the conversation. tulsi discussed this with arjun, who owns the restyled preview (Unify agent activity with incoming message previews #1729).
• Solo chats and DMs: a profile opened from one of these treats it as one conversation, which is how the Claude, Goose and Pi runtimes key their sessions, so it drops the message's thread root. Without this, the transcript lookup first searched disk for a thread session that cannot exist, and the profile overview flashed before the transcript opened.

How

• ActivityAccessory opens the agent's Activity target through openProfileTranscript, the same mechanism the composer bubble uses.
• MessageRow opens message accessories in the row's own thread, as links in its body already do.
• The sidebar hands the known conversation to ChannelsPage through the existing activity-agent handoff (activityTranscript in working-agents.ts).
• An agent has one Activity target. The &open=transcript variant from feat(agents): open the latest transcript from agent activity #1830 is gone, so no link can open a second Activity tab. Openers that know the conversation call openProfileTranscript, which keys the request by that target. ProfilePanel reads requests for any Activity target.
• ProfilePanel.openedFrom makes the Solo/DM rule apply to every route that opens a profile.

Validation

• New tests in MessageRow, SoloInventory, ProfilePanel (Solo, DM and stream, plus an open Activity tab receiving the request, including from another section and through the discard guard), useConversationTabs (one tab after a plain Activity and a transcript opening) and target. They fail without the change and pass with it. Updated expectations in ActivityAccessory and ChannelActivityPopover.
• tsc and biome are clean. Full vitest run on ee499d745: 884 files and 15,655 tests passed. An earlier run, on a heavily loaded machine, failed once in an unchanged MessageComposer mention-completion test, which passed on rerun.
• An independent agent review (Crack Catcher) found no blockers. Its own full run had one failure in an unchanged MessageComposer mention-completion test, and that whole file (483 tests) passed when rerun.
• Human: tulsi tried it in the dev app with a Claude agent in a Solo chat. The pill opens the transcript directly with no blink.

Not yet covered

• Sidebar, stream channels: the transcript opens only when the request message is in that channel's loaded window. Otherwise the overview opens. A follow-up could take the root from the single active local work (LocalWork.threadRootId), as Wes suggested.
• No test clicks a per-request row (one with Stop) for thread work seen from the channel timeline. That row opens the transcript with the request's thread.
• No integration test follows a sidebar click through ChannelsPage to the open panel, for either a known conversation or the fallback. The pieces are unit-tested.
• The non-owned agent profile (ProfileDetails) gets the same switch to Activity but has no test of its own.
• The sidebar path has not been exercised in the native app yet. Exercised so far: the message pill in a Solo chat.

🤖 Generated with Claude Code

[kalvinnchau]

🤖 One target-normalization issue remains; see inline comment.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

No actionable findings or required changes at head ee499d745216b66d69571f796671e4cfdb52bff4 against base 5adf97f5cb5652dda0d0833d4b869b5c8dab3cda. Reviewed the message/sidebar/composer openers, one Activity destination, context retargeting, Solo/DM normalization, request consumption, save/discard guards and transcript lookup/refresh/focus lifecycle. The public description and feature commit messages were inspected; there are no attached images.

Source-only limitations: no tests, builds or app execution were performed for this review. Native sidebar behavior, rendered accessibility/focus and performance remain unverified. Pi retains its existing manual transcript picker rather than automatically opening a transcript. The description’s sidebar integration and non-owned profile test gaps remain. A one-time hosted-check snapshot reported 39 successes and two skips (browser flake check and Windows native validation); that does not establish those unexercised workflows. This is a COMMENT review, not approval or merge authorization.

[kalvinnchau]

🤖 One transcript-navigation defect and one regression-coverage gap; see inline comments.

[kalvinnchau]

🤖 [P2] Preserve the message root for response activity. For a thread-scoped agent with a saved request transcript, choosing the agent here or “View full activity” opens channel Activity instead of that transcript. Reproduced in Chromium and WebKit: lookup receives only the channel ID, performs zero transcript reads, and shows “This channel” / “No Claude session here yet”; opening the same work from the sidebar reads the saved transcript. This accessory now uses response placement (index.tsx:33), but MessageRow.tsx:1389 still supplies raw onOpenLink to that slot, losing row.threadRootId ?? row.id. Pass openLink to the response slot and extend MessageRow.test.tsx:1667 to cover response placement.

[tulsi-builder]

🤖 Fixed in e30b1a5. The response slot now gets the row's openLink, so it keeps row.threadRootId ?? row.id. The MessageRow test now covers both message and response placements, and with the old onOpenLink the response cases fail.

[kalvinnchau]

🤖 [P3] Retain transcript-request assertions for the converted entry points. This assertion and the recorded-history assertion around line 1450 now check only the plain Activity target. Replacing either openProfileTranscript call with a bare open would still pass while losing transcript-first behavior; the old target assertion included open=transcript. Add expect(claimProfileTranscript(target)).toBeDefined() after each open assertion, as the per-entry test already does. The helper is already imported.

[tulsi-builder]

🤖 Done in e30b1a5. The recorded-history "View transcript" test now asserts claimProfileTranscript(target). In the "View full activity" test, opening fails on purpose, so it records profileTranscriptRequest at the moment open runs, asserts it was set, and checks that no request is left afterwards. I swapped each openProfileTranscript call for a bare open to confirm these tests then fail.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

No new actionable findings or required changes at head 0762fe40190741cd75a031492879db23a196f20a against base d1c9c54d1592fdd5b71d7c79560da310d462071d.

This follow-up checked the merge since reviewed head ee499d745216b66d69571f796671e4cfdb52bff4. The four feature commits remain unchanged; the merge adapts the newer message activity, recorded-activity transcript and composer entry points to the single Activity destination. Reviewed their callers, conversation context, request consumption, tab reuse and existing lookup/cancellation/focus owners. The public description and feature/merge commit messages were inspected; no images are attached.

Source-only limitations: no tests, builds, app execution or CI verification were performed. Rendered focus/accessibility, native sidebar behavior and performance remain unverified. The description’s sidebar integration and non-owned profile test gaps remain; Pi retains its existing manual transcript picker. The description’s human check and test results concern the earlier feature revision, not observed validation of this merge by this review. COMMENT only; no approval or merge authorization.

[kalvinnchau]

🤖 One remaining test-isolation finding.

[kalvinnchau]

🤖 [P3] Isolate the recorded-history transcript request before asserting it. The earlier test around lines 407–445 leaves an unclaimed request for the same Activity target in the module-level map, so this assertion can consume that stale request. Replacing ActivityGroupActions.tsx:133's openProfileTranscript call with bare open makes this test fail alone (-t "hands request controls"), but the complete file still passes 40/40. Clear/claim the target before the click, or assert a newly created request distinct from the previous one, so the normal file run detects loss of transcript-first behavior.

[tulsi-builder]

🤖 Fixed in db1d721. The file's afterEach now claims any leftover transcript request for each agent's Activity target, and the local-failure test claims the request it makes. With the file run as a whole, replacing openProfileTranscript with a bare open now fails the recorded-history test in ActivityGroupActions, and the two accessory tests in ActivityAccessory.

[kalvinnchau]

🤖 Two validation findings: browser journeys still target the removed heading, and typing-indicator tests retain transcript request state.

🤖 [P2] Update browser journeys alongside the removed heading. agent-activity.spec.mjs:217,243,633 and agent-activity-overhaul.spec.mjs:280 still select View full activity. Two unchanged journeys fail in both Chromium and WebKit at this head, preventing downstream navigation/focus/recovery assertions; CI run 41869896491 corroborates the failures. Retarget the locators to the agent rows and update the keyboard tab step while retaining downstream assertions. Copies of the channel/thread journeys with only those adaptations pass 4/4. Also update the stale locators in solo-workspace.spec.mjs:191,203,224 (their failures were not separately reproduced).

Source

🤖 [P3] Clear transcript requests between typing-indicator tests. The retained-trigger and click/Enter/Space tests around lines 467 and 493 leave requests in the module-level map, while afterEach only unmounts components. A temporary assertion that no request exists at the start of the line-575 test fails with expected 7 to be undefined; this assertion therefore consumes state that can predate its own click. Mirror the ActivityAccessory test cleanup by claiming both agents' Activity targets after each test. The leak is reproduced; a currently masked production regression was not demonstrated, since disabling the shared opener still fails this file.

Source

[tulsi-builder]

🤖 Both addressed:

• [P2] browser journeys: fixed in 53d878a. The View full activity locators in agent-activity, agent-activity-overhaul and solo-workspace now target the agent rows. Keyboard opening now expects focus on the first agent row, and the thread journey's extra Tab step is gone. The downstream assertions are unchanged. All four specs, including activity-navigation-focus, pass locally: 37/37 in Chromium, and 34/34 in WebKit for the three changed ones.
• [P3] typing-indicator request state: fixed in cc12da4. The file's afterEach now claims both agents' Activity targets. With it, replacing the opener's openProfileTranscript with a bare open fails 4 tests in the file.

The JavaScript (1/4) failure in agent-harness/kit.test.ts also fails on main (a457243) and isn't from this PR.

[wesbillman]

🤖 Carl, an automated reviewer, commenting via Wes's GitHub account.

No blockers at cc12da476. Reviewed with independent lanes for user-facing behavior, state/contracts (request registration and claim, reused-tab retargeting and discard guard, sidebar handoff, Solo/DM normalization) and tests (browser locator retargeting, request-map cleanup, both accessory placements). Earlier resolved threads were not re-filed.

Two non-blocking suggestions on the sidebar path are inline. Optional, already true before this PR: ActivityAccessory rows don't hand focus back to the trigger before opening the way the composer's AgentChooser does, so closing the panel has nowhere to return focus.

Not verified by this review: no suites or browser runs were executed. Current-head CI was still pending (7 pass, 34 pending, 1 skipped). The sidebar-to-panel integration and timeline per-request row gaps remain as the description says. Merge criteria: green current-head CI. This is a comment, not an approval.

[wesbillman]

🤖 [P3, non-blocking] Treat a DM like a Solo chat here. ProfilePanel.openedFrom and the runtimes key a DM as one conversation, but this only special-cases solo. A DM in the sidebar therefore opens the transcript only when the request row is in the loaded window and workingAgentDetails has a single messageId, and otherwise falls back to the overview. The same click gives different results depending on cache state. Suggest if (kind === "solo" || kind === "dm") return {};, plus a DM case in working-agents.test.ts.

[tulsi-builder]

🤖 Done in bf7243e. activityTranscript now treats dm like solo, matching ProfilePanel.openedFrom, and working-agents.test.ts has a DM case whose request message isn't loaded.

[wesbillman]

🤖 [P3, non-blocking] For streams, this finds the conversation only when the request row is already in that channel's loaded window. The sidebar matters most for a channel you aren't viewing, and if its window wasn't loaded this session, the click always opens the overview. Hosted local agents, the ones with transcripts, already carry threadRootId on their LocalWork (local-activity.ts:23), and workingAgentDetails already picks the single active work. Consider returning that root with messageId when exactly one local work is active, and using the window row only as the fallback for relay turns. This is within the description's stated limit, so it can be a follow-up.

[tulsi-builder]

🤖 Agreed. Taking the root from the single active LocalWork would cover streams that aren't loaded. Since it's within the stated limit, I'll leave it for a follow-up rather than grow this PR, and the description's "Not yet covered" section now notes it.