fix(agent-status): stop a child's work re-dating and un-reading its idle parent by brennanb2025 · Pull Request #21399 · stablyai/orca

brennanb2025 · 2026-09-18T06:44:17Z

	Files	Added	Deleted	Net
Test	9	$\color{#1a7f37}{\Huge{\mathbf{+}}}$338	$\color{#cf222e}{\Huge{\mathbf{−}}}$94	$\color{#1a7f37}{\Huge{\mathbf{+}}}$244
Prod	15	$\color{#1a7f37}{\Huge{\mathbf{+}}}$882	$\color{#cf222e}{\Huge{\mathbf{−}}}$46	$\color{#1a7f37}{\Huge{\mathbf{+}}}$836

ELI5

When you ask an agent to do something, it sometimes hands part of the job to a helper agent and
finishes its own turn. In the sidebar, the agent that had finished kept jumping to the top of the
list and saying it had just done something — every single time the helper did anything. Its blue
"unread" dot came back each time too, so a row you had already read went unread over and over for
work you never saw.

Now the finished agent's row shows when it actually finished, stays read once you have read it,
and simply shows as still busy for as long as its helper is still working.

What Changed

Before. A session that spawns a helper agent writes both agents' output into one timeline. The
row's "last activity" time was read off that whole timeline, so the helper's output counted as the
parent's. A user watching the sidebar saw the parent row stamped "now" while it sat idle, and the
helper row — the thing actually working — stamped "14m" beside it. The idle row looked live and the
live row looked stale. The same timestamp is what Orca compares against to decide whether you have
read a row, so the finished row also went unread again on every helper frame, and the notification
it produces changed identity with it.

After. The finished row shows the time the session's own agent finished, and holds it. It stays
read once read. While the helper is still working, the parent row reads as busy rather than
finished — so nothing goes quiet just because the work moved to a child.

The mechanism.

The clock stops counting other agents' rows. The journal reducer advanced lastActivityAt for
every row appended. It now advances only for rows the session's own agent produced, using the
producer marker and the shared isRootAgentJournalItem predicate that already ship on this
branch. Two consumers read that clock, and both are the session's status summary.
Subagent roster rows are marked as what they are. Claude and Codex both write child-only roster
rows from the parent's code. Those rows are rewritten on child transitions, so leaving either
provider unmarked still moved the parent clock. Both providers now stamp roster revisions as
child-produced, and Claude preserves that attribution when removing its roster row. Nothing is
hidden by that: the transcript deliberately renders every producer's rows, and the sidebar's
child list comes from the live roster the host publishes, not from this row.
Status, and only status, rolls up. The host's status feed now publishes working for a session
whose own agent has settled but whose subagent is still running. The recency clock, the prompt,
the running-tool line and the last assistant message all stay strictly the parent's own. The
rollup is deliberately narrow: only agent-kind children in a working state (a backgrounded shell
is not a subagent), attention still outranks it, and a session with no turn at all is not made
listable by a child.
The renderer stops holding a second copy of a rule. StructuredAgentSessionStatusBridge passed
its own stateStartedAt when writing the row. The store's entry builder already computes one —
same state keeps the stamp, a state change restamps it — and that rule is identical to the
canonical host-side writer's. The bridge's version added a done special case that restamped
every settled row on every publication, which is what let the acknowledgement clock move. The
property is gone; the builder's rule applies. The bridge itself stays, and so do its publication
filters.

Why

The failure was not a formatting bug. One timeline carries two agents' rows, and every "what is this
agent doing right now" answer was a scan over that timeline with nothing on a row saying who wrote
it. Attribution at the producer — already landed for the prose and tool-line readers on this branch
— is what removes the premise, rather than adding a check in front of it.

Alternatives considered:

Just delete the bridge's stateStartedAt. One line, and it does fix the reported screen. But
the session's published updatedAt stays contaminated, so worktree ps, mobile and the canonical
host row still report the parent as freshly active, and the host still re-broadcasts the whole
summary to every subscriber once per child frame. It hides the symptom from one reader.
Publish the turn record's own completion time as a second timestamp and stamp the row from
that. Exact, and cheap, because turns are already root-only. Rejected: it adds a second recency
concept beside updatedAt instead of making updatedAt correct, and attribution yields the same
answer for free — the parent's last own row is its turn record.
Leave the parent quiet while its child works. Correct about the clock and wrong about the
user: a parent whose only outstanding work is a child is not finished. Hence the status rollup,
which is the one thing a child is allowed to say about the row it hangs under.

Disclosure, because it is the honest shape of this change rather than a footnote: Orca keeps two
writers of one status row apart with publication filters; this change does not remove that
structure. A renderer component still writes structured pane keys, which the store's own guidance
says a reader should not do. What this change removes is the renderer's policy — after it, the
structured path has one stateStartedAt rule instead of two, and they agree. Consolidating the
writers is a separate, already-planned change, and it gets easier for this one having landed, not
harder.

One further limit worth stating: attribution here is a convention, not a type. A future append site
that forgets the marker re-opens the hole, and the mitigation is the producer-boundary tests below
rather than something the compiler can enforce.

Linked Issue

Fixes #

Visual Proof

Validated in the exact PR worktree with an isolated background Electron session. The first image
shows the parent row still Working while its subagent is active; the second shows the same row
settled to Done after the child reports back.

Testing

pnpm tc clean. pnpm exec oxlint clean on every changed file.
pnpm run check:code-quality:changed passes — 0 new findings across 37 changed files.

Suites run green: src/main/native-chat, src/main/claude, src/main/codex, src/main/runtime,
src/shared, src/renderer/src/store, src/renderer/src/attention,
src/renderer/src/components/sidebar, src/renderer/src/components/native-chat, mobile.

Electron validation launched this branch with ORCA_BACKGROUND_LAUNCH=1, attached over CDP, and
verified the app identity before interaction. A real Claude session spawned a real subagent from
the UI. While the child worked, the visible parent row stayed Working, the backing entry exposed
the child as working, and its stateStartedAt held at 1789973127195 while later child activity
advanced updatedAt to 1789973140619. The row then visibly settled to Done after the child
completed. The renderer console had no errors.

Every new test was proven red by removing the corresponding production change at the final head and
green with it restored:

Test	Removed to prove it red
`journal-reducer.test.ts` — the session's recency clock (own rows advance it, a child's do not, on both the item and lifecycle-batch paths)	the producer check around `lastActivityAt`
`structured-agent-session-status-feed-subagents.test.ts` — an idle session's clock and publication count hold still across five child rows	the same
`structured-agent-session-status-feed-subagents.test.ts` — a session with a working subagent reads `working`, claims no tool of its own, and ignores a backgrounded shell or a settled child	the status rollup
`claude-subagent-roster.test.ts` — every roster row is child-produced, first write and each revision	the roster row's marker
`codex-subagent-roster.test.ts` — Codex roster rows are child-produced	the Codex roster marker
`structured-agent-session-event-sink.test.ts` — child attribution survives a tombstone append	tombstone attribution forwarding
`claude-structured-journal-translation-subagents.test.ts` — the roster row is stamped	the same
`StructuredAgentSessionStatusBridge.test.tsx` — a settled row holds its completion stamp, its attention timestamp, and its read state as the host clock advances	the bridge's `stateStartedAt`
`StructuredAgentSessionStatusBridge.test.tsx` — a restored completion is stamped with host journal time and the bridge sends no `stateStartedAt`	the same

Two harness details worth a reviewer's eye, because both would have produced a test that passes
either way:

The roster test's fake sink discarded the options argument entirely, so every append read as root
and the attribution assertion would have passed against the unfixed code. It now captures options.
The feed test first used the wall clock, and five appends inside one millisecond share a ts, so
Math.max moved nothing and the test passed under ablation on a re-run. It now drives an explicit
advancing journal clock and fails under ablation on every run.

Three existing tests changed, which is worth saying plainly rather than leaving in the diff:

StructuredAgentSessionStatusBridge.test.tsx, "sorts restored completions by host time and
advances identical turns" — split. The restore half is a real fix from another change and is kept
verbatim. The half asserting that an idle row's stamp advances on an identical republication is
inverted, because it is the defect written down as intent: it contradicts the canonical
host-side writer, it contradicts the documented invariant in the timestamp's own module, and its
final assertion is a same-state ping re-triggering attention, which the acknowledgement module
states outright cannot be allowed to happen. What it was really protecting — two completions in a
row must not leave the row frozen on the first — now has its own test that drives the turn
directly.
claude-structured-journal-translation-subagents.test.ts — asserted the roster row stays root, on
the stated reasoning that stamping it would hide the subagent list from the parent. It does not:
the transcript projection is deliberately unscoped and the sidebar's children come from the live
roster. Inverted, with the reason recorded next to it.
StructuredAgentSessionStatusBridge.test.tsx, "accepts an authoritative older journal age after
a host upgrade reconnect" — the corrected clock still lands on updatedAt, which is the point of
the test. The completion stamp no longer moves with it, because the row was already done and
stayed done — which is exactly what the canonical writer does with its own copy of the same
session. One assertion updated.

Also refactored for the line budget rather than suppressing it: the status feed's test setup moved
into a shared -test-bed module used by both feed test files, and the old
-test-session helper folded into it. That removed a pre-existing double type assertion in the
process — the fake session map is now a real Map subclass.

Platforms: the change is platform-independent (journal reducer, host status projection, renderer
store write). Verified on macOS.

I manually tested these changes locally
Automated tests added/updated, or explained why not below

AI Disclosure

Claude Opus 4.5.

Review

Worth a second opinion on two judgement calls:

The rollup's state set. It fires for agent-kind children in working or monitoring, and
for a child whose state is absent (an older host's live task). waiting and blocked are
deliberately excluded: a child that wants something is not a parent that is busy, and carrying
them would be claiming the child's status rather than reporting outstanding work.
Where the rollup lives. It is in the host's status feed, so worktree ps, mobile and the
sidebar cannot disagree about one session. The alternative was the renderer, which would have
meant adding renderer policy in the same change that removes some.

Agent skill upstream boundary

Not applicable.

Notes

Backwards compatibility / remote wire. No new field, no new frame, no schema version bump. Two
published values change meaning, which the wire-compatibility guidance treats as a wire change
even with no codec movement: a session's updatedAt now derives from its own agent's rows only,
and its status can be working because of a child. Both are values every existing client
already renders, and the new meaning is the corrected one, so an older paired client reads it
without a capability gate and simply shows the fixed behaviour.
Journals written before this lands carry no attribution, so a replayed child row is still read
as the session's own. This is forward-only and self-heals on the session's next turn; nothing is
backfilled.
Not fixed here, deliberately. A subagent re-invoked after a pause keeps its first spawn stamp,
so it can render "40m" while its current invocation is seconds old. That is a separate latch in
the roster, it needs its own evidence, and the apparent version of it in the original report — an
idle parent reading "now" beside a child reading "14m" — is resolved by this change, because the
two now read coherently.
Performance. Strictly less work: an idle session with a running child no longer re-broadcasts
its whole status summary to every local and remote subscriber once per child frame.
Security, SSH and cross-platform: no execution, path, or transport surface touched.

Checklist

This PR is small and focused
I explained what changed and why (ELI5, the user-facing before/after, the mechanism, and why over the alternatives)
Before/after screenshots or videos attached for UI changes, or N/A with reason
Self-reviewed for correctness, security, and performance
Cross-platform, SSH/remote, and path/shortcut impact considered (or N/A)
pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass (or CI will cover; local preferred)

…traints

… fact

… and plan the fix

…angs under A structured session's journal holds its own agent's rows and every subagent's, and the session's recency clock was a scan over all of them. So while a backgrounded child worked, the parent's row was restamped to "now" on every frame the child emitted: the settled row read as if the agent had just spoken, and the live child beside it read stale. The stamp it moved is `stateStartedAt`, which is also the acknowledgement clock, so the parent went unread again each time too. The producer marker this branch already lands is the fix. The journal's `lastActivityAt` now advances only for rows the session's own agent produced, and the Claude subagent roster row — written from the parent's context, with no `parent_tool_use_id` near it, yet holding nothing but children's state and rewritten on every child transition — is marked child-produced, because unmarked it kept moving the clock on its own. A parent whose only outstanding work is a child must still read busy, so the status feed rolls the children's status up, and only the status: a settled session with a working subagent publishes `working` while its clock, prompt, tool line and assistant prose stay strictly its own. `attention` outranks the rollup and a session with no turn stays unlisted. The renderer bridge stops passing `stateStartedAt`. The store's entry builder already held that rule, identical to the canonical host-side writer, and the bridge was overriding it with a second, divergent one whose `done` special case is what let a settled row's stamp move. The bridge stays, and so do the publication filters that keep one pane key to one writer; what leaves is write-time policy the renderer should never have owned. Three existing tests changed. One asserted that an idle row's stamp advances on an identical republication — that was the defect written down as intent, and it is now inverted, with the behaviour it was really protecting (two completions in a row) covered directly. One asserted the roster row stays root, on the reasoning that stamping it would hide the subagent list; it does not — the transcript renders every producer's rows and the sidebar's child list comes from the live roster. One asserted a corrected older journal clock also moves the completion stamp; the row was already `done` and stays `done`, which is what the canonical writer does with its own copy of the same session.

…parent-recency

pullfrog

Important

The root directory guard check is failing: this PR adds ATTRIBUTION-parent-recency.md at the repo root, which the guard rejects for any new root-level entry, and verify fails with it. Move the file under an existing tracked directory or drop it before merging.

Reviewed changes

Journal recency is now the session's own — applyJournalRow advances lastActivityAt only when isRootAgentJournalItem(row) (absence of the marker means root), so a child's rows no longer move the parent's clock.
Claude's roster row is stamped child-produced — the "Ran 3 agents" row is written from the parent's context but holds only children's state, and now carries producedBySubagent: true; the empty-group tombstone stays root.
Status rolls up, and only status — the host status feed publishes working for a session whose own agent settled while an agent-kind background task is still working.
The bridge stops writing stateStartedAt — the store entry builder's same-state-keeps / state-change-restamps rule applies instead of the bridge's done special case.
Shared helper + test-bed refactor — isWorkingSubagentBackgroundTask on the wire, and the feed test setup consolidated into structured-agent-session-status-feed-test-bed.ts.

ℹ️ The clock fix and the rollup only cover Claude today

The reducer scoping and the status rollup both key off producedBySubagent, which only the Claude translator stamps. Codex journals child-thread items into the parent session and writes its roster row with no marker, so an idle Codex parent can still be re-dated and re-marked unread by a child — the exact symptom this PR fixes for Claude. The PR body discloses this; flagging it so the scope of the bug being closed is explicit.

Technical details

# Codex child rows remain unattributed

## Affected sites
- `src/main/codex/codex-structured-journal-items.ts:259` — `items.handle(event)` runs for any `threadId`, appending child-thread items to the session journal with no `producedBySubagent`.
- `src/main/codex/codex-subagent-roster.ts:353-356` — the roster row is appended with `options` that carry no marker.
- `src/main/codex/` — `grep producedBySubagent` returns no matches.

## Required outcome
- Either accept the Claude-only scope explicitly where the bug is tracked, or stamp Codex child rows before the clock scoping can claim to cover Codex.

## Open questions for the human (optional)
- Does the issue this fixes reproduce on Codex? If so, is a follow-up planned?

ℹ️ Nitpicks

journal-reducer.test.ts: the new "a lifecycle batch a subagent did not produce, and holds on one it did" case builds a lifecycle-batch row with producedBySubagent: true, but no production builder can emit one — journalLifecycleBatchRowBuilder has no such option (journal-row-builders.ts:104-108) and the sink's lifecycle append drops it. The reducer handles it correctly; the case just guards a shape that cannot reach the reducer today.
structured-agent-session-status-feed-test-bed.ts: StatusFeedSessionsView extends Map but overrides only get and [Symbol.iterator], so size, values, and has still read the empty super-map. A future feed read through one of those would silently see zero sessions.

^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}

pullfrog · 2026-09-18T07:25:01Z

@@ -0,0 +1,658 @@
+# Defect A — an idle parent row is stamped "now" by its child's work


This new root-level file fails the root directory guard job, which blocks any entry not present at the PR base and not in REVIEWED_ROOT_ENTRIES; verify fails with it.

Technical details

# New root-level file fails the root directory guard ## Affected sites - `ATTRIBUTION-parent-recency.md:1` — new root-level file, not present at base `brennanb2025/attr-parent-label`. - `.github/scripts/check-root-directory-entries.mjs` — compares base and head trees, rejects every new root entry except `cloud`. ## Required outcome - No new root-level entry in the PR. ## Suggested approach (optional) - Move the diagnosis doc under an existing tracked directory (it would need an allow-list entry if placed in `docs/reference/`, since `docs/**` is gitignored), or drop it from the PR if it is a scratch planning artifact.

pullfrog · 2026-09-18T07:25:02Z

+export function isWorkingSubagentBackgroundTask(task: AgentSessionBackgroundTask): boolean {
+  return (
+    task.kind === 'agent' &&
+    (task.state === undefined || task.state === 'working' || task.state === 'monitoring')


This branch is the only reason a Codex subagent rolls up — Codex agent tasks are built with kind: 'agent' and never set state — yet every rollup test passes an explicit state, so a regression here would break Codex silently.

Technical details

# Untested branch: `task.state === undefined` is the Codex shape ## Affected sites - `src/shared/agent-session-background-task-wire.ts:64` — the absent-state arm. - `src/main/codex/codex-background-task-tracker.ts:63-67` — Codex agent tasks carry no `state`. - `structured-agent-session-status-feed-subagents.test.ts:72,99` — the only rollup tests, both with explicit states. ## Required outcome - A feed test that drives the rollup with the Codex task shape (`kind: 'agent'`, no `state`) so the absent-state arm is pinned.

pullfrog

ℹ️ No new changes since the prior review. The only new commit (d9269ee) merges base brennanb2025/attr-parent-label into the head, and the PR diff is byte-identical to the one reviewed at 5fdaf62. The findings from the previous review still stand unchanged: the root directory guard failure on ATTRIBUTION-parent-recency.md, the Claude-only scope of the attribution and rollup, and the untested state === undefined arm of isWorkingSubagentBackgroundTask.

^{｜ Fix it ➔ ｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}

pullfrog

ℹ️ No critical issues in this delta — one scope note below.

Reviewed changes

Codex roster rows are now child-produced — CodexSubagentRoster.write stamps producedBySubagent: true, so Codex roster churn no longer moves the parent's recency clock.
Claude's roster removal is child-produced too — the empty-group tombstone now carries the marker alongside the roster row it removes.
Tombstone attribution is threaded end to end — JournalTombstoneInput, journalTombstoneRowBuilder/buildJournalTombstoneRow, AgentSessionJournal.appendTombstone, both deferred-sink tombstone paths, and the resolved appendItem path forward producedBySubagent.
Tests added — a Codex roster-marker test and a sink test proving attribution survives a tombstone append; the Codex harness now captures the append options it previously discarded.
Docs updated — the PR body and its checked-in copy now describe both providers' roster stamping.

ℹ️ Codex roster rows are stamped, but Codex child-thread items still are not

The Codex change stops the roster row from moving the parent clock, but Codex journals child-thread items into the same session journal with no marker, so an idle Codex parent's updatedAt still advances on a child's prose and tool rows. The doc change in fef2626 removed the earlier disclosure that only the Claude translator attributes today and now presents coverage as provider-complete, which oversells what actually landed — either stamp Codex child-thread rows at the producer (the shared isRootAgentJournalItem predicate is already provider-neutral) or keep the caveat.

Technical details

# Codex child-thread rows remain unattributed

## Affected sites
- `src/main/codex/codex-structured-journal-translation.ts:259` — `items.handle(event)` runs inside the main `handle` for any `event.threadId`, appending child-thread items with no `producedBySubagent`.
- `src/main/codex/codex-structured-journal-translation.ts:288` — the generic-frame fallback (`genericFrames.appendUnhandled(..., event.threadId)`) does the same for any thread.
- `src/main/codex/codex-subagent-roster.ts:353-356` — the new marker covers only the roster row.
- `PR-BODY.md:33-36` — now claims both providers are stamped, with the previous Claude-only caveat removed.

## Required outcome
- Either mark Codex child-thread rows at their producer (the reducer's `isRootAgentJournalItem` scoping is shared and provider-neutral), or restore the removed disclosure that this change covers roster churn and Claude items, not Codex items.

## Open questions for the human (optional)
- Is the Codex child-item gap tracked as a follow-up? If so, the PR body can point at it instead of dropping the caveat.

^{｜ Fix it ➔ ｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}

…ent-recency

…tablyai/orca into brennanb2025/attr-parent-recency

pullfrog

ℹ️ No critical issues — the delta is a clean merge of origin/main; one reviewability note below.

Reviewed changes

Since the prior pullfrog review at fef2626, the authored delta is small; the bulk of this run's diff is a merge of origin/main.

Merged origin/main (259 commits) into the branch — the only manual conflict resolution is StructuredAgentSessionStatusBridge.tsx, which now imports getStructuredAgentSessionTabs from main's extracted structured-agent-session-tabs.ts module while keeping this PR's stateStartedAt removal. Every other file merged cleanly; the branch is a strict superset of origin/main (only 39 files differ from it).
Added a deferred-sink test — structured-agent-session-event-sink.test.ts now pins producedBySubagent riding through both tryAppendResolvedItem and tryAppendResolvedItemAndPublish.
Base reconciliation — 81995c8/4918c1a folded the moved base and the const/observedAt tweaks with no net content change to the reviewed files.

ℹ️ The merge folds 259 `origin/main` commits into this PR's diff

The declared base brennanb2025/attr-parent-label (1136af7) does not contain the origin/main commits this branch just merged. Because 1136af7 is an ancestor of the head, git diff base..head now includes every upstream commit pulled in, and GitHub reports 3,253 changed files — so "Files changed" no longer shows the PR's own ~24-file surface. The merged code is correct (main's changes are all present, and the only resolution is the bridge); this is purely a reviewability cost.

Technical details

# Stack base is behind the merged main

## Affected sites
- Branch topology: head `98d5244` parents `4918c1a` (PR) + `663d670` (`origin/main`); base `1136af7` is an ancestor of the head and predates the 259 merged `origin/main` commits.
- `git diff 1136af7 HEAD --name-only` spans ~3,000 files, almost all of them from `origin/main`; `git diff 663d670 HEAD --name-only` is only the 39-file authored surface.

## Required outcome
- The PR's "Files changed" shows only the PR's own work again.

## Suggested approach (optional)
- Merge `origin/main` into the base branch #21398 as well, so the shared base tip carries main and the stack's diff excludes it, or retarget this PR to `main` once #21398 lands (GitHub retargets automatically when the stack base branch is deleted).

^{｜ Fix it ➔ ｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}

brennanb2025 · 2026-09-23T18:26:36Z

Superseded by #22520, which fixes the same symptom — a subagent's activity re-dating its idle parent session and marking it unread again — against the producer linkage now on main (#22299), instead of the reverted producedBySubagent flag this branch was built on. #22520 keeps this PR's mechanism, drops the reverted-flag lines and the added doc file, and also excludes the subagent roster row, which on its own was re-dating the Claude parent. Closing in favour of #22520.

brennanb2025 added 6 commits September 17, 2026 12:07

docs(attr-parent-recency): record the attribution defect and its cons…

0619bd6

…traints

docs(attr-parent-recency): add reference findings and the feasibility…

009523a

… fact

docs(attr-parent-recency): verify the chain, challenge the mechanism,…

9626cae

… and plan the fix

docs(attr-parent-recency): add the PR body for the parent-recency change

5fdaf62

Merge branch 'brennanb2025/attr-parent-label' into brennanb2025/attr-…

d9269ee

…parent-recency

pullfrog Bot reviewed Sep 18, 2026

View reviewed changes

brennanb2025 added 3 commits September 20, 2026 23:41

fix(native-chat): keep child roster churn off parent recency

29c01df

docs(attr-parent-recency): record provider-complete review fixes

fef2626

docs(attr-parent-recency): add Electron validation evidence

bd4f069

pullfrog Bot reviewed Sep 21, 2026

View reviewed changes

brennanb2025 added 3 commits September 21, 2026 00:05

chore: reconcile parent recency with current base

81995c8

fix(native-chat): preserve resolved append timestamps

4918c1a

Merge remote-tracking branch 'origin/main' into brennanb2025/attr-par…

98d5244

…ent-recency

brennanb2025 requested a review from Jinwoo-H as a code owner September 21, 2026 16:45

Merge branch 'brennanb2025/attr-parent-label' of https://github.com/s…

0375ab3

…tablyai/orca into brennanb2025/attr-parent-recency

pullfrog Bot reviewed Sep 21, 2026

View reviewed changes

brennanb2025 mentioned this pull request Sep 23, 2026

fix(native-chat): date a session by its own lifecycle, not its subagents' work #22520

Merged

3 tasks

brennanb2025 closed this Sep 23, 2026

Navigation Menu

fix(agent-status): stop a child's work re-dating and un-reading its idle parent - #21399

fix(agent-status): stop a child's work re-dating and un-reading its idle parent#21399
brennanb2025 wants to merge 13 commits into
brennanb2025/attr-parent-labelfrom
brennanb2025/attr-parent-recency

brennanb2025 commented Sep 18, 2026 •

edited by github-actions Bot

Loading

Uh oh!

pullfrog Bot left a comment

Uh oh!

pullfrog Bot Sep 18, 2026

Uh oh!

pullfrog Bot Sep 18, 2026

Uh oh!

pullfrog Bot left a comment

Uh oh!

pullfrog Bot left a comment

Uh oh!

pullfrog Bot left a comment

Uh oh!

brennanb2025 commented Sep 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		@@ -0,0 +1,658 @@
		# Defect A — an idle parent row is stamped "now" by its child's work

Conversation

brennanb2025 commented Sep 18, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

ELI5

What Changed

Why

Linked Issue

Visual Proof

Testing

AI Disclosure

Review

Agent skill upstream boundary

Notes

Checklist

Uh oh!

pullfrog Bot left a comment

Choose a reason for hiding this comment

ℹ️ The clock fix and the rollup only cover Claude today

ℹ️ Nitpicks

Uh oh!

pullfrog Bot Sep 18, 2026

Choose a reason for hiding this comment

Uh oh!

pullfrog Bot Sep 18, 2026

Choose a reason for hiding this comment

Uh oh!

pullfrog Bot left a comment

Choose a reason for hiding this comment

Uh oh!

pullfrog Bot left a comment

Choose a reason for hiding this comment

ℹ️ Codex roster rows are stamped, but Codex child-thread items still are not

Uh oh!

pullfrog Bot left a comment

Choose a reason for hiding this comment

ℹ️ The merge folds 259 origin/main commits into this PR's diff

Uh oh!

brennanb2025 commented Sep 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

brennanb2025 commented Sep 18, 2026 •

edited by github-actions Bot

Loading

ℹ️ The merge folds 259 `origin/main` commits into this PR's diff