Keep nested replies together in thread conversations

[morgmart]

What this does

Reply to a specific message inside a thread, then expand or collapse its conversation branch. The composer shows who you are replying to without discarding your draft. Sending opens the branch containing your reply. Sessions stay inline; unrelated legacy thread features are out of scope.

Behavior and implementation

• Nested replies retain both the thread root and immediate parent through folding, row identity, signing and publishing. Direct replies keep their existing format; retry preserves the signed event.
• Messaging owns the hierarchy and branch state, composed from Base UI Collapsible, existing design-system buttons and tokens. No new persistence store or native adapter.
• Closing a branch closes descendants; reopening reveals one level. The whole thread can collapse too. Deep links reveal and focus the target through collapsed ancestors. Missing loaded parents get an explanation, not a hidden reply. In-thread branch counts describe loaded replies, not all history. The timeline badge uses the relay’s whole-thread total, including nested replies.
• Root-keyed drafts survive changing the reply target. Main's composer editing, edit cancellation/native undo, channel-scoped names and removal of byline presence are preserved. The editor keeps the stable accessible name “Reply to thread”; its placeholder names the target.
• Straight branch connectors, sibling grouping and quiet continuation timestamps follow the approved legacy treatment. Indentation is capped responsively. Hover/focus controls reserve their natural height permanently, so they neither cover their own message nor shift its layout; coarse pointers retain visible collapse labels.

Validation

Current head: 8e44612b44e9510c9a0265c9e4760756a2d53d37, rebased onto main 30c3589b. Local, remote branch and PR heads match; working tree is clean. Hosted CI and reviewer re-review remain gates.

Main rebase (8e44612b)

• Rebased the existing branch onto current main, preserving original authorship and signoffs. Updated the remote with an exact lease against the previously verified head, not a blind force push.
• Resolved the author-byline overlap by retaining main's compact, non-focusable custom status inside the author wrapper while preserving continuation-row semantics and MessageTimestamp. Kept the approved stable reply composition, inert Settings retention, Todos/Bestie stacking, and incoming scrollend/drained-read browser synchronization.
• Explicitly restored narrow-panel behavior and coverage previously carried in merge commit 0640e581, which ordinary rebase does not replay. Compared the entire resulting tree with the expected merge of the old head and new main: only the deliberately resolved MessageRow conflict differs. Independent exact-head source review found no preservation blocker.
• On clean committed 8e44612b, complete layout, message-navigation, nested-replies, and user-status browser files passed 78/78, Chromium/WebKit, two workers, 3.0 minutes. Isolated output: test-results/main-rebase. No new or removed browser cases in this rebase. This is not the full browser suite; fixtures exercise production app/broker with modeled upstream, not deployed-community or packaged-native behavior.
• Mandatory push checks passed at the same head: 127 test files / 1,894 related tests, TypeScript, design-system types and all design guards. All 16 PR commits have Signed-off-by trailers. No hook bypass or relaxed assertions. New-head hosted CI and reviewer approval remain pending.

Earlier validation below belongs to its named pre-rebase snapshots.

Stable reply rows and side-panel stacking (f605f757)

• Keep the reply message subtree mounted through first-child arrival and last-child deletion. Conditional disclosure no longer remounts the message or resets spoiler/focus state. No repeated reveal or new focus owner. Independent bounded source review found no material defect.
• Limit retained-panel grid sharing to the actual Settings-containing wrapper and inert retained panels. Todos stacks with threads and linked panels in either opening order; Settings and Bestie keep their existing geometry. An initial overly narrow selector broke Settings/Bestie geometry and was corrected before commit, without relaxing assertions.
• Added two rendered component regressions and three browser cases per engine (six net browser cases added, none removed): two focus variants of first/last-child transitions and one side-panel stacking journey. Browser-only justification: actual DOM identity, native focus, spoiler state and panel geometry across production app/broker wiring.
• Fail-then-pass evidence: original leaf/branch composition failed all four engine/focus cases; repaired composition passed all four. Original panel selector failed both engines with an 882px stacking-gap error versus the unchanged <2px assertion; corrected full layout file passed 22/22.
• After main integration at 6d32dc05, complete layout, message-navigation, and nested-replies browser files passed 76/76, Chromium/WebKit, about 2.8 minutes, isolated output test-results/integrated-main-review. This is an affected-file run, not the full browser suite. Final f605f757 changes only an existing unit-test traversal helper; production/browser-test content is identical to the integrated run.
• The first push caught three existing shallow-traversal tests that did not inspect the reply render callback. The helper now invokes that callback without weakening assertions. Successful mandatory push hooks at f605f757: 122 test files / 1,856 related tests, TypeScript, and all design-system type/color/contrast/foundation/icon/adoption guards. No hook bypass. Hosted DCO passed; remaining hosted checks were pending at readback.
• Local production-app/broker fixtures use modeled upstream behavior; no claim of deployed-community publication or packaged-native validation. Both repaired GitHub threads have evidence replies and are resolved; reviewer approval remains outstanding.

Persistent-sidebar integration (0640e581)

• Resolved the edit-provenance test conflict by preserving both the required attachment source and direct/nested ancestry assertions. Independent source review covered the merged signing, folding, row identity, and resolved test.
• The combined code reproduced horizontal thread overflow in both engines at a 390px viewport: the persistent sidebar leaves a 188px history, while the old depth cap left deep rows only 20px wide. For thread containers <=320px, stop further indentation at depth 0 and use the existing compact trailing spacing token. Root rail, touch summaries, and per-row keyboard/pointer collapse controls remain. No sidebar redesign.
• Complete nested-replies.spec.mjs: 14/14 passed, Chromium/WebKit, two workers, 1.3 minutes. Includes new 390px crowded-branch case per engine (one net case added per engine, none removed), all 12 collapse/reopen/hit/focus cycles, touch no-overflow, deep-row readability, and threshold-adjacent widths. Browser-only justification: actual layout, hit testing, focus, and pointer media behavior. Existing desktop assertions remain; mobile actions are correctly asserted visible rather than hidden.
• Complete layout, message-actions, message-navigation, and thread-unread files: 70/70 passed, both engines, two workers, 2.0 minutes. Both browser runs used the resolved working tree based on 38f86ac2 plus main 119195ea, committed here without formatter changes. A preceding diagnostic WebKit run once failed the existing Up-to-edit focus step; subsequent full-file runs passed without a related code/test change. No claim that this intermittent failure is eliminated.
• Mandatory hooks at 0640e581: 119 Vitest files / 1,786 tests passed, TypeScript, design types/guards, formatting/lint. All 15 PR commits have signoffs; hosted DCO passed. Local policy-relay/browser evidence only, not packaged-native verification.

Crowded-thread CI test isolation (38f86ac2)

• Hosted WebKit shard 2 at 282abdde exceeded the single 120-second deadline in the crowded-controls test. Its trace showed 33/36 completed branch cycles, 239 clicks and 430 attribute reads; completed individual actions stayed under one second. All other automatic lanes passed; this shard had 149 other passing tests.
• Split the three existing viewport walkthroughs into independent fixture-backed tests. Every per-width interaction, hit-test, focus, hover, geometry and screenshot assertion remains; timeout unchanged. Two net additional cases per engine, no production changes. The separate nested-reply journey still exercises resizing within one session. Independent source review found no coverage loss in the tested contract.
• Full nested-replies.spec.mjs passed 12/12, Chromium/WebKit, two workers, in 1.0 minute on the pre-format working tree based on 282abdde committed here. Each isolated crowded case took 9.6–12.8 seconds locally. Earlier unsplit WebKit case took 47.8 seconds in the local two-file run; hosted exhaustion is not a directly comparable local performance measurement. Additional per-width fixture setup is intentional; hosted new-head timings remain pending.
• Mandatory hooks passed at 38f86ac2: TypeScript, 104 Vitest files / 1,587 tests, design types/guards and staged formatting/lint. New-head CI and reviewer re-review remain pending. GitHub readback at repair start exposed only the two original review comments, with no newer inline or issue comments.

Late ancestry focus and retained spoiler repair (282abdde)

• Preserve exact-target focus when late ancestry replaces its DOM node. Transfer only focus owned at detach during the synchronous ancestry update; deliberately moved focus, aborted navigation, inert settings and later collapse/reopen do not reclaim it. The one-shot navigation reveal stays unchanged.
• Retained settings panels now suppress subtree painting with opacity as well as visibility/inert, preventing revealed spoilers from overriding the hidden wrapper while retaining geometry.
• Added three browser cases per engine (two nested held-history focus cases and one spoiler-paint/restoration case), none removed. Browser-only justification: native focus on DOM reparenting and actual screenshot compositing cannot be proved by jsdom. Existing direct-reply identity coverage remains. Signed nested fixture supplies the parent only through held history; settings screenshots are compared against the same panel with forced subtree transparency.
• Before production fixes, target-focus and spoiler-paint assertions failed in both engines; moved-focus controls passed. After fixes, complete message-navigation and nested-replies files passed 44/44 in Chromium/WebKit, two workers, about 2.4 minutes, on the pre-format working tree based on 9d5fc9b8 committed here. Includes programmatic collapse/reopen without stale focus restore. Independent bounded source review found no actionable defect; reviewer did not run tests.
• Mandatory hooks passed on 282abdde: TypeScript, 104 Vitest files / 1,587 tests, design types/guards and staged formatting/lint. First push attempt encountered an existing MessageActionBar keyboard-focus test failure; an unchanged second full hook run passed. No test bypass, timeout change or assertion relaxation.
• Hosted CI and reviewer re-review at this head remain pending. This is local policy-relay evidence, not packaged-native or deployed-relay verification.

Timeline count repair (9d5fc9b8)

• Prefer validated relay descendant_count over direct-only reply_count for the existing presentation total. Valid zero wins; absent/invalid totals retain the validated direct-count fallback. No UI, layout, persistence, or relay changes.
• Confirmed both actual relay serializers supply these fields: block/buzz channel-window bridge and live summary emitter. Signed fixture coverage checks one direct reply with three total, live increases/decreases, stale and spoofed summaries, and zero. Render coverage checks visible and accessible badge text. Five assertions failed before the correction; all 140 tests in the three affected files passed afterward.
• Mandatory hooks passed on 9d5fc9b8: TypeScript, 104 Vitest files / 1,587 tests, design types/guards, formatting/lint and secret scan. Independent bounded source review found no actionable defect; reviewer did not run tests.
• No new browser cases: this is summary projection through existing UI, covered below the browser layer. No live reply/delete mutation or verification of Morgan’s exact channel was performed. New-head hosted CI remains pending; earlier browser passes below belong to their stated snapshots.

Earlier nesting and settings repair (e35f02b7)

• Mandatory pre-push TypeScript, 104 related Vitest files / 1,578 tests, design-system typecheck and all design guards passed on clean e35f02b7. This includes all 10 nesting component tests and the regression preserving collapse preference across temporary root loss. Mandatory commit hooks passed for the repair and merge. All PR commits have Signed-off-by trailers.
• 40/40 browser checks passed at e35f02b7, Chromium and WebKit, two workers: complete message-navigation and agent-activity files (about 1.4 minutes locally). These exercise exact reveal, retained target identity/focus, reading dwell, settings isolation, restored navigation, and main's new activity preview.
• 86/86 affected browser checks passed before formatting/commit/main integration, on the repair working tree based on 9437955b: complete channel-settings, completion-layout, layout, message-navigation, navigation-thread-history, nested-replies, thread-unread, messages, and typing files, both engines, two workers (about 3.5 minutes locally). This is not claimed as an 86-case run at the integrated head.
• Earlier complete nested-replies, navigation-thread-history, thread-unread, and message-edit files passed 20/20 in both engines on 6e0bb292 plus the test-only publication wait committed as 9437955b. The nested journey covers composer → broker signing → policy relay → delivered row, rejected send/retry with signed ancestry, held pending sends, selected-parent editing/cancel/native undo, collapse, deep links, missing parents, and responsive layouts in both themes.

Repair of the six previous hosted browser failures

• Replace obsolete “N replies loaded” assertions with rendered message-row counts; account for collapsed nested descendants in the typing fixture.
• Preserve the covered primary panel's layout while channel settings is open, using inert/visibility rather than removing its geometry. No new scroll bookkeeping or changes to message container queries. The companion retains its separate grid track.
• Keep the root reply branch mounted while surrounding context loads. A late root no longer replaces the exact reply's DOM node and loses focus. If a collapsed root temporarily disappears, available replies remain visible and its collapse preference returns with it.
• Failure-to-pass evidence: the held-context identity regression failed in all four engine/focus combinations before the stable-root repair and passed all four afterward. WebKit settings restoration failed with collapsed retained layout and passed after retaining geometry. A controlled-clock browser case proves covered content cannot take focus or publish reading progress, then resumes after settings closes.
• Browser case delta versus main: six net new cases, none dropped. Four nested-reply journeys cover production broker wiring, layout, pointer and focus behavior; the existing late-context test now covers both retained and deliberately moved focus; one new settings case checks real inert focus/reading behavior. Existing settings/companion layout assertions were strengthened at wide and narrow sizes. Root-loss/collapse state coverage stays in Vitest.
• Bounded independent source/diff review covered 9437955b..c33122c1, not runtime or the merge. Integrated-head inspection confirms the merge did not change the repaired panel layout or root-branch files; integrated runtime scope is the 40-case run above.

# Integrated-head browser run
bin/pnpm test:browser --project chromium --project webkit --no-deps --workers 2 \
  message-navigation.spec.mjs agent-activity.spec.mjs

# Mandatory integrated-head push validation
BUZZ_TEST_WORKERS=2 git push origin HEAD:refs/heads/morganm/nested-replies

All automatic hosted CI lanes passed at e35f02b7, including all four browser shards and CI required. That pass does not cover the subsequent count repair; new-head CI is a separate gate.

Scope and remaining checks

• Local policy-relay fixtures only; deployed-relay sending, packaged native behavior, real touch devices and a complete action-by-action keyboard walk were not exercised. Hosted CI completion and human approval remain gates.
• The approved permanent controls row makes dense threads roomier. Please eyeball the narrow desktop example before shipping; the independent crowded fixture was approximately 45% taller across 11 rows than the earlier overlapping layout.
• Nonblocking accessibility follow-up: action controls precede author/text in DOM order despite their visual placement. No claim of complete screen-reader/tab-order validation.
• Separate, out-of-scope finding recorded in Buzz: the main timeline's existing absolute hover tray can overlap message details at narrow widths. This PR fixes thread geometry, not the timeline tray.

Diff size at f605f757 against integrated main: tests +1,782/−127 across 19 files; implementation +1,014/−196 across 17 files; docs +17/−11 across 1 file. Tests are counted separately from implementation.

Current narrow-panel screenshots

Synthetic fixture content from the successful final browser run, not a deployed conversation. A 1024-wide desktop window yields a roughly 311-wide thread panel. Indentation stops growing to preserve reading space. Controls reserve space even while hidden.

Crowded sibling/continuation example, showing the intentionally reserved controls spacing:

[wesbillman]

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

Changes requested: two P2 regressions, detailed inline. Merge criteria: preserve the exact target’s focus when delayed ancestry changes its DOM placement (without reclaiming deliberately moved focus), and fully hide retained conversation contents while Channel Settings is displayed.

Reviewed head 9d5fc9b831a2db41defd28fd7eaa5125b55d6ea3 against base/merge-base 8842b3ac05862e069ab0adf2f30e11a3af084042. Integrated independent protocol/count and presentation reviews; no additional blocker found in those inspected paths. Approved spacing and controls-row decisions are unchanged.

Validation: source/contract review plus temporary, fixture-only Playwright probes against that head with no production edits. Both regressions reproduced in Chromium and WebKit; the two moved-focus controls passed. Existing delayed-context coverage uses direct replies, not a nested target whose parent arrives later. Broad suites were not duplicated; live relay mutations, packaged native behavior and physical touch were not exercised.

CI/mergeability, checked at review closeout: all automatic hosted lanes and CI required pass; GitHub reports mergeable. The optional Windows lane was skipped. These are separate from the reproduced defects above.

[wesbillman]

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

Source re-review clear at 38f86ac27d8414bfc236ed72859d43cac862a150 against 8842b3ac05862e069ab0adf2f30e11a3af084042: both prior P2 findings are source-fixed; no new actionable defect found. COMMENT only, not approval or merge readiness.

• Focus transfer is limited to focus owned at selected-row detach, bounded to the synchronous ancestry update, with moved-focus, aborted-navigation and inert guards. Retained wrapper opacity prevents revealed spoilers painting through Settings without discarding layout/state. The targeted regression sources include nested DOM replacement, deliberately moved focus, later collapse/reopen and spoiler restoration.
• Crowded checks retain all three widths and their interaction/geometry assertions. Isolating fixtures adds two cases per engine without changing the timeout or retries; the separate journey still exercises same-session resizing. Hosted runtime cost remains unmeasured.
• No CI workflow ran on this head. Only Semgrep, zizmor and DCO passed. The previous WebKit timeout belongs to 282abdde, not this tip. GitHub reports conflicts with main: resolve them and validate the resulting head, including Chromium/WebKit repairs and isolated crowded cases, before merging.

This review used source/contract analysis and existing hosted metadata only; no PR code or tests were executed. Live relay mutations, packaged native behavior, physical touch and a complete keyboard/screen-reader walk remain unverified. Approved controls spacing is unchanged and not a blocker.

[wesbillman]

Pinky, an automated reviewer, commenting on Wes's behalf. This review also includes Brain's thread review.

Two P2 regressions need fixing at 0640e5813702c9c3a874e66c7b1ef3c2a5865e62; details are inline. To merge:

1. A reply's message subtree stays stable when it goes from zero to one child, or from one to zero.
2. Outside Settings, a thread or linked panel stays stacked with a side channel panel instead of overlapping it.

Each fix needs regression coverage.

Checked and found no problems:

• Nested reply signing adds a root tag only when the parent isn't the root, which matches the relay's ancestry validation.
• The fold and row identity carry replyParentId.
• Timeline totals use the relay's descendant_count and fall back to a validated reply_count.
• The merge resolution of the traffic integration test keeps main's assertions.

Hosted CI is green at this head: 3,340 Vitest tests, all Chromium/WebKit journeys including nested-replies, browser measurements, and Rust. Windows native validation was skipped.

Both findings come from reading the source; neither was reproduced in a browser, and no local suites were run. This is a COMMENT review, not an approval.

[wesbillman]

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

Changes requested at 0640e5813702c9c3a874e66c7b1ef3c2a5865e62 against 119195ea331de33c8480bab180df0091ca8e9421: both existing P2 findings remain actionable. I independently traced the owning code and callers; details and regression expectations are already in the linked inline threads:

• Preserve the message subtree across zero↔one-child transitions. The keyed outer li and its memoized ref remain mounted while the inner row is replaced, so selectedBranchRef does not run; the completed one-shot reveal does not repair lost focus. Preserve local row state as well, rather than adding repeated focus grabs.
• Restore side-panel stacking outside Settings. Limit the shared grid cell to Settings retention; preserve the existing thread/link + side-panel arrangement rather than closing a user's thread as a workaround.

Exit criteria: fix both transitions and add regression coverage, retaining the prior late-ancestry focus, moved-focus, Settings paint/inertness and crowded-layout protections.

Exact-head hosted CI passed on the merge with the stated base: 3,340 Vitest tests and all Chromium/WebKit shards. Those passes do not cover these two missing transitions. Windows native validation was skipped. This was source-only review; neither finding was reproduced in a browser here, and packaged-native/deployed-relay behavior remains unverified. Approved permanent controls spacing is unchanged and not a blocker.

[wesbillman]

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

Re-review clear at f605f7571c2ab5afe95db32dee3a2f3e1b7d6441 against 6fa0e9a63690e7eab413f9e5c11be7a05a4f6c4f. Both prior P2 findings are repaired; no new actionable defect found in the changed paths. COMMENT only, not approval or merge readiness.

• Reply rows now retain the same ReplyBranch/MessageRow composition through first-child arrival and last-child deletion. Regression sources check node identity, revealed spoilers, owned focus and deliberately moved focus. The final helper-only commit restores shallow traversal of the render callback without removing assertions.
• Shared-grid overlap is restricted to Settings and inert retained panels. The new Todos journey checks thread/link stacking in both opening orders; existing Settings/Bestie geometry, late-ancestry focus, Settings spoiler-paint and nested-collapse protections remain intact. Approved controls spacing is unchanged.
• Hosted CI run 36078744083 passed, including 3,553 Vitest tests, Rust/Node integration, browser measurements and all Chromium/WebKit journey shards. The JS checkout log verifies merge c5beddd75d1c349bd30d824f3ef3f10253631638 of this head into the stated base. Windows native validation was skipped.

Validation was source-only on Blox plus existing CI evidence. This invocation executed no PR code/tests and did not exercise deployed-relay, packaged-native, physical-touch or full screen-reader journeys. Other jobs’ checkouts were not individually audited; normal human review/branch protections remain.

[wesbillman]

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

Source re-review clear at 8e44612b44e9510c9a0265c9e4760756a2d53d37 against 30c3589ba57ce926ac19c5526d8b87cde35d6e48. No actionable regression found in the rebase from the previously reviewed head. The stable reply-row composition, Settings/panel stacking fixes, exact-target focus behavior and narrow-panel rules are preserved; the author/status integration retains continuation timestamps.

Required CI is still red, so this is not merge-ready certification. Run 36087531104 exercised synthetic merge 4593bdd25d01781f5a0fc546e382e04fabe66382 of these exact pins. All four Chromium/WebKit journey shards passed, including the affected nesting/navigation/layout coverage. JavaScript finished with 3,647 passing tests and one failure in AgentModelPicker.test.tsx (the “Pi discovers extension providers…” case cannot find the accessible “Cancel model lookup” button). No causal regression from this PR was established; resolve that required gate before merge.

Source-only review on isolated Blox plus existing hosted CI evidence. No PR code executed or CI rerun by this review; Windows native validation was skipped and live/packaged-native acceptance is not established. This is a COMMENT, not approval.