Skip to content

fix: harden board-owned threads after QA (shell trim, routing, seat view, project removal) - #187

Merged
polymaniadeveloper-hub merged 8 commits into
masterfrom
feat/board-owned-threads-hardening
Nov 22, 2026
Merged

polymaniadeveloper-hub merged 8 commits into
masterfrom
feat/board-owned-threads-hardening

Conversation

@polymaniadeveloper-hub

@polymaniadeveloper-hub polymaniadeveloper-hub commented Nov 22, 2026 •

Copy link
Copy Markdown
Collaborator

Follow-up to #186 (merged as dbe1d21aaf). It hardens the board-owned threads spine against the findings of the independent QA pass, and adds the tests that could be written without a second machine or a running client. Thread creation, binding and the task.source stand-in swap are not in it (see Deferred).

Four commits, one per step. Branched from origin/master at dbe1d21aaf.

1. Shell trim and tag-filter routing (QA findings 1 and 2) 44e2e76859

Finding 1. Every un-archived thread carrying the marker went to every client in the shell snapshot, where the client only hid it.

  • A subscriber that reads thread families (omitIdleSubagents, the same opt-in the idle-subagent trim uses) now gets no marked thread in the socket or the HTTP snapshot, and none of the seats and side chats rooted at one. The store's index decides this, so the rows are never built.
  • The archived snapshot leaves marked threads out on the server, where before only new clients filtered them.
  • Kept on purpose: a marked thread that delegated a task to a linked environment. That task nests under it on the other machine, and without the row it would show up there as a top-level thread.
  • What is not fixed, stated plainly: deltas. A change to a marked thread still sends its row to every connected client (thread.updated). An open task view needs that, because the branch picker, git actions, terminals and the diff panel read the open thread's shell row. So within one session a client collects the rows of the board threads that changed, and loses them at its next snapshot. The snapshot is bounded; the session is bounded by the tasks that were active during it. Cutting the deltas too needs the server to know which threads a subscriber has open, which does not exist today.
  • A client that does not opt in (an old build) still gets every row, since it has no other way to them.

Finding 2. With a tag filter on, ThreadRouteView opened a thread the shell leaves out only as the child of a visible parent. A board thread has no parent, so with the trim above none would have opened while a filter was on.

  • A server-owned thread now opens by id on its lineage, which never changes, so the stale-detail worry behind the rule does not apply.
  • An archived one opens too (resolveThreadRouteRenderState). No archive list shows it, so its id is the only way in. Ordinary archived threads still close, as before.
  • I narrowed the instruction here: the brief said "a shell-absent thread opens by id regardless of tag filters". I did that for server-owned threads only. Doing it for every thread would let the address of an archived thread with a hidden tag open it, which is what the rule exists to stop.

Tests. The routing probe (ThreadRouteView.browser.test.tsx) is now a matrix: tag filter on and off, against a row that arrived as a delta, a row the shell left out, and an archived thread. Each half of the fix fails it on its own. The probe in #186 passed for a reason worth knowing: the fake kept the first test's shell rows and settings for the whole file, and a cached detail from an earlier case opened the thread before the route closed it. Each case now has its own thread, proves the filter is in force before it opens anything, and asserts the shell row state. ProjectionStore.test.ts reads the client-bound snapshot from the real store through loadSubscriberShellSnapshot; ShellStream.test.ts covers the family rule and the linked-environment exception.

2. Seat view stability (QA finding 3) 6d0a877525

A seat in a linked environment opens in the direct thread view or the read-only linked view, two components. The choice was made on every render, so the centre flipped when the links answered after the view had mounted, and twice more on a short reconnect. Each flip remounted the conversation.

  • A newly opened seat waits until the links have an answer before either view mounts. Meanwhile the centre shows the seat bar and "Loading the conversation...", not the task's own thread.
  • An open direct view stays through a dropped connection. It gives way to the linked view only after the connection has been down for ten seconds (SEAT_LINKED_VIEW_DELAY_MS). The number is a judgement, not a measurement.
  • A connection of the client's own still replaces the linked view at once, and a linked view is not given up for a link that only looks absent while the list reloads.
  • The rule is one pure function, steadySeatThread, in lib/taskConversation.ts, which stays the single reader.

Tests. Unit tests for the loading and reconnect cases. A browser test mounts the hook with the view and checks that the same composer element keeps its text and focus through a reconnect, moves to the linked view only when the wait ends, and that the direct view never mounts while links load. Removing either rule fails both.

3. Guards and cleanups 3178d156fb

  • Marker guard. boardThreadMarker.test.ts reads the source of the server, web, mobile, desktop, contracts, client-runtime and shared, and fails when anything but NativeIntake reads the marker's value, or when anything but the marker module knows the board-task: form. Every other line that names the field must match a listed presence check, declaration or pass-through.
  • Project removal. The owner ruled archive, never silent delete.
    • Count: the client no longer holds these rows (commit 1), so the shell snapshot now carries serverOwnedThreads, a per-project count. The web and mobile confirmation adds a line: "3 conversations started by a board or the engine are archived, not deleted."
    • Removal: ProjectService ends the work of a marked thread the way a deletion does (runs, requests, sessions, terminals) and archives it, with the seats and side chats under it. Attachments stay. An already archived one is left as it was. Everything else is deleted as before. planThreadDeletion gained a keep: "archived" mode so both paths share one teardown.
    • The wording says "a board or the engine" because the client reads presence only and cannot tell a board thread from an engine intake. Engine intake threads are archived on removal too, for the same reason.
    • The count is exact at the client's last shell snapshot. A task started or archived since can make it off by that change.
  • German catalog. "Phase nicht bekannt" is now "Phase unbekannt". Still machine-translated; no native speaker has read it.
  • QA finding 5 (the first-open key kept per task, and one more automatic open after upgrading): left alone, as the brief allowed.

4. Verification debt 2c8d0c5441

  • Mobile open by id. use-thread-selection.test.tsx takes the link the app itself writes for a thread, reads it with the Thread route's path, and checks the route's selection is that thread when the shell left its row out, when it is archived, and when its row came as a delta. Limit: mobile has no fake socket, so the probe stands in for the environment at the two reads the route makes. It does not render the thread screen.
  • A seat's environment going offline with nothing linked to it (the saved-but-offline path): the direct view stays, composer and text included, and no linked view appears however long it lasts.
  • An ordinary thread under a tag filter (ThreadRouteView.tagFilter.browser.test.tsx): a listed one opens, a hidden one and an archived one land on /.

Is the marker enough to resolve a task from a thread?

The question: can the pull request and issue linked-thread rows show "Board task: <title>" and open the task view from the marker alone, without waiting for the engine's boardThread field?

What the marker gives. A thread whose lineage.engineIntake starts with board-task: names its board id and task id. Lineage is written once by the server at creation (withCreationProvenance strips it from client commands) and never changes. So thread to task identity is available today, on the server and on any client that holds the thread.

What it does not give.

  • The title. It is an engine fact, and the spec keeps card fields to engine facts (section 4, rule 7). The row needs one engine read by task id. The thread's own title is not a substitute.
  • The task to thread direction. Nothing can find a task's thread from a marker, and nothing should: the engine's boardThread is the only source of which thread is the task's conversation. A task restarted after its thread was archived can leave two threads with the same marker, and the marker cannot say which is current.
  • Which engine. The marker has no environment or engine in it. The row has to be resolved in the environment the thread lives in.

Drift risks.

  1. The value is not parseable as written. Board and task ids are joined with : and neither is escaped, so board-task:a:b:c is ambiguous. The ids have to be pinned to a grammar without :, or escaped, before anything reads the value. The lead's review already asks this of the creation slice.
  2. The board id can go stale: the engine can move a task to another board, and the marker cannot be rewritten. Resolve by task id and treat the board id as a hint, if task ids are unique across boards in one engine. I did not check that.
  3. A deleted task leaves a marker pointing at nothing. The row needs a fallback.
  4. A resolver is a second reader of the value. The guard test in commit 3 will fail, on purpose. The resolver belongs in boardThreadMarker.ts, and the test's list is extended deliberately.
  5. Today's exclusions are presence checks, so they also hide engine intake threads from linked rows. Restoring board rows must select on the board-task: prefix, not lift the exclusion.

Verdict. The marker is enough to restore the linked rows without the engine link field: marker to task id, one engine read for the title, open the task view. It is not enough to say that the row's thread is the one the task view shows. Until the stand-in swap, the task view opens the thread the task was filed from, which is a different thread. That is a reason to do the restoration in the same pull request as creation, as ruled, and not before.

T3 pieces reused

  • The shell trims and their opt-in (omitIdleSubagents, shellSubscriberThreadIds, the active shell index): the marked-thread trim is one more filter in the same place. No new subscription option.
  • Thread families (getThreadFamily, useOpenThreadFamily): how an open board thread gets its shell row back. Unchanged.
  • ChatView's shell-less fallback (presentThreadShellFromProjection): why an archived board thread renders with no row at all.
  • OrchestrationV2OlderSettledCount and the readUnloadedThreadCount pattern: the per-project count and its readers in web and mobile.
  • planThreadDeletion: the teardown an archive on removal shares with a deletion.
  • planProjectRemoval in client-runtime: one confirmation for web and mobile.
  • subagentThreadTarget, resolveLinkedThreadOpen and LinkedThread: unchanged; the steadiness rule sits in front of them.
  • The fake environment and fakeTimers() of the web test harness. The fake gained tags, archived threads, a shell delta, and settings that reach the live connection.
  • Nothing new was built for folder, branch or model picking; none was needed.

Deferred, and why

  • Thread creation on task start, binding to the team lead, and the task.source stand-in swap. Blocked on the engine's task to thread link (boardThread, spec section 6a), which is not published.
  • Linked-row restoration ("Board task: <title>", opening the task view). Ruled to land with creation; see the note above.
  • Deltas for marked threads (the rest of finding 1). Needs per-subscriber interest on the shell stream.
  • A door to archived board threads. They open by id now, but nothing in the UI lists them, and one archived by a project removal belongs to a project that no longer exists.
  • Approvals and unread in a board thread (QA finding 4). An owner question, untouched.

Surfaces

Exercised by tests: the server's client-bound shell snapshot against the real store, the trim rules, the archived snapshot filter, project removal against the real event sink and projections, the marker guard, web routing in a real browser (board threads and ordinary ones), the seat view in a real browser, the removal confirmation text, the mobile route selection.

Covered by construction, not exercised: the HTTP shell route (it calls the same loader as the socket), the mobile removal confirmation (the shared plan is tested through web), the mobile and web count readers (readServerOwnedThreadCount), the wiring of useSteadySeat into useTaskConversation (the hook is tested with the view, the certainty inputs are not), desktop (it shares the web code).

Not verified

  • No pass in a real client, on any surface. I started no dev server, as instructed. There are no before and after images for the two visible changes: the seat's loading line and the new line in the removal confirmation.
  • Whether an archived board thread takes a message, and what the composer does there. The route opens it; I did not check sending.
  • Opening a thread that a project removal archived. Its project is gone, and I did not check what the thread view does without one.
  • The mobile probe does not render the thread screen and does not go through the app's navigation container.
  • The linked seat view with two real environments. The tests feed the connection and link states in.
  • The new confirmation line is translated by the model in ten web and ten mobile catalogs. No native speaker has read any of them. i18n:check passes.
  • ProjectService.ts and ShellStream.ts are imported by most of the server's tests. I ran the files that exercise the changed behaviour (project deletion and mutation, the project CLI and MCP handlers, thread deletion, the shell stream, the socket, the store), not every importer. I ran no repo-wide check.
  • Two tests fail on this machine and are the two feat(board): board-owned threads stay out of every thread list #186 recorded as failing on base: ProjectionStore.test.ts "reads a fixed SQL turn-item window for long histories and repeated clients", and ChatView.browser.test.tsx "a shortcut conditioned on rightPanelOpen sees the thread's open right panel", which passes alone. I did not re-run them on base myself.
  • Typecheck ran per package (server, web, mobile, contracts, client-runtime); lint and format on the changed files; knip in the commit hook.

Claude Fable 5.2, Claude Code.

🤖 Generated with Claude Code


Independent QA (round 2), head 2c8d0c5441

Tester: Claude Opus 5.5 (Claude Code), a different model from the builder. No product code changed, nothing pushed or merged. Worktree at the PR head, one dev server (vp run dev, seeded copy in the worktree's .t3), real web client in T3's Browser panel. Threads were made with the CLI on the free Replay provider. Three were then marked as board threads by setting lineage.engineIntake (board-task:qa-board:<id>) in the projection rows with the server stopped, and two carry the tag private. The marker went in by a direct projection write, not through NativeIntake, so the event log has no marker for them.

Machine: swap was nearly full the whole time (24 to 27 GB used; about 31% of memory free). The server logged 10 second event-loop stalls, so steps that depend on timing were not attempted.

Verdict: pass with findings, 0 blocking

Findings, ranked

1. Medium. A board thread that carries a hidden tag still lands on / under the tag filter. Item 2 of the PR says a server-owned thread opens by id whatever the tag filter. Opening one by id brings its family into the client shell (useOpenThreadFamily), so its row with its tags arrives, and hiddenByTagFilter in ThreadRouteView.tsx (about line 122) closes it. The server-owned exemption only covers offShellBlocked.
Repro: set hideTaggedThreads: true and hiddenThreadTags: ["private"]. Tag a marked, live thread private. Open /<env>/<threadId> by deep link or in-app push: the URL goes to /, then to a draft. Control cases under the same filter: a marked untagged live thread (otter) opens, a marked archived untagged one (walrus) opens, and an ordinary tagged or archived thread lands on / as intended. The routing matrix in the tests never puts a tag on a board thread. It is not blocking only if no board thread can carry a hidden tag. Otherwise exempt server-owned lineage from hiddenByTagFilter too, or drop server-owned rows from hiddenThreadKeysAtom.

2. Medium. The web composer sends into an archived board thread and a full turn runs; the CLI refuses the same send. Archived board threads now open by id, so the composer can be reached. On walrus (archived, project still there), Enter sent the message, the Replay turn ran, and the thread stayed archived and out of every list. That is agent work running where no list shows it. t3 thread send on the same thread answers "Thread ... is archived and cannot receive messages." (ThreadManagementService.sendToThread). The owner should pick one rule, either read-only or unarchive on send, and have both entry points follow it.

3. Low. On an archived board thread whose project was removed, send does nothing and says nothing. On otter after the project removal, the thread opened with its full history and "No project" in the status line. The submit button is enabled. Pressing Enter or clicking submit leaves the text in the composer, raises no toast or error, and no message reaches the server.

4. Low, read from the code only, from #186. listProjectThreads with includeSubagents: true (the MCP t3_thread_list path) drops marked roots but checks only the thread's own lineage. A seat or side chat rooted at a board thread would be listed, which goes against the shell rule this PR sets ("nor seats and side chats rooted at one"). I had no seat fixture to show it live. dbe1d21aaf has the same code.

5. Nit. Confirmation wording. "Those threads and any archived threads are deleted from T3 Code" sits right above the new archive line. A board thread that was already archived is kept, not deleted (deleteChildThread returns early), so the first line overstates. The new count covers active board threads only, so the two lines do not contradict each other, but the first line is now inexact.

Verified

  • A. Lists. Board threads (zebra live, walrus archived, otter in a second project) appear in none of these, while the control threads do: the sidebar (Settled shelf at 0, nothing to expand), sidebar search by title, the command palette's recent threads, palette search by title and by message text, and the archive list (heron listed, walrus not). The CLI thread list shows the controls and no board thread. A delta mid-session: after a fresh load, two turns were sent to otter through the CLI. Otter stayed out of the sidebar, the palette's recent threads and palette search. Opening a board thread by id does not add it to the sidebar either.
  • B. Open by id, filter off. Live zebra and archived walrus open by deep link and after a full reload. Back and forward across zebra, walrus and a control keep the right thread every time. Ordinary archived heron lands on /.
  • B. Filter on (after a server restart; the setting did not hot-reload from a hand edit of settings.json, which is probably expected). Walrus and otter open. Ordinary tagged ibis and ordinary archived heron land on /. Back and forward keep walrus. Zebra fails, see finding 1.
  • D. Project removal, run on qa-remove (2 control threads plus board thread otter). The confirmation said: Remove project "qa-remove" and delete its 2 threads? ... 1 conversation started by a board or the engine is archived, not deleted. That matches what happened: otter was archived (archived_at set, deleted_at null), both controls were deleted, and the project was deleted. Otter then opened by id with its history and no project in the breadcrumb. Sending into it: see findings 2 and 3. A capture of the dialog is saved on the test machine; it is not uploaded.
  • G. The two failures the builder reported both exist before this PR. On 9db6a1bf84, dbe1d21aaf and 2c8d0c5441:
    • ProjectionStore.test.ts "reads a fixed SQL turn-item window..." fails the same way on all three: TypeError: statement.stat is not a function. This Node 24.21 node:sqlite has no StatementSync.stat, so it comes from the machine, not the code.
    • ChatView.browser.test.tsx "a shortcut conditioned on rightPanelOpen..." fails in a full-file run on all three (1 failed, 17 passed, about 15 s timeout) and passes alone at the head.
  • The PR's own tests at the head all pass. Server: boardThreadMarker, ShellStream, ProjectService.deletion (31 tests); ProjectionStore and ThreadDeletion (63 of 64, the one failure above). Web unit: taskConversation, threadRoutes, ProjectSettingsPanel.logic (39). Web browser: ThreadRouteView, ThreadRouteView.tagFilter, TaskConversation (15). Mobile: use-thread-selection (4).

Not verified

  • C. Seat view in a real client. That covers the loading line, a reconnect of less than 10 seconds keeping the composer, more than 10 seconds moving to the linked view, and a direct connection returning it. It needs a second environment and a linked task, and I had neither. With the event loop stalling for 10 seconds, timing on this machine could not have been trusted anyway. Only the PR's browser test covers it.
  • A, partly. The thread link picker and the PR and issue linked-thread rows: the scratch repositories have no remote, PR or issue. The sidebar's search box matches titles only (a control could not be found by its message text either), so message-text search was checked in the palette only.
  • E. The live MCP tools t3_thread_list and t3_thread_search on the PR build. My MCP tools talk to the host environment, not the dev server. I checked the same service through the CLI thread list and read the SQL ThreadSearch uses (it excludes engineIntake). includeSubagents both ways was checked by reading the code only (finding 4).
  • F. Mobile and desktop in a real client. Not run, because of memory. Only the mobile unit test passed.
  • The HTTP shell route, the mobile removal confirmation, and translations, none of which a native speaker has read.
  • Whether a marker written by NativeIntake behaves exactly like my direct projection write.

Claude Opus 5.5, Claude Code.


Fix round after QA round 2, head 97b8011ba3

Builder: Claude Fable 5.2 (Claude Code). Four commits on top of 2c8d0c5441, one per fix. QA finding 5 is fix 4 here.

Finding 1, tag filter bc2b85ce9e. ThreadRouteView now exempts server-owned lineage from hiddenByTagFilter as well as from offShellBlocked, read from the shell row once the family brings it in and from the detail before that. The routing matrix gained three cases (a board thread carrying a hidden tag, filter on, for each shell state). The two live ones fail without the fix; the ordinary tagged thread test still lands on /.

Findings 2 and 3, one rule: an archived thread receives nothing 9c22c0d416.

  • Server: the orchestrator refuses a message.dispatch into an archived thread with OrchestratorThreadArchivedError, in the same place it refuses a message to a provider's own subagent. That is the one point the web socket, the CLI and the MCP tools share. The sentence is the CLI's ("Thread ... is archived and cannot receive messages."), now one function both errors use. Message intake treats it as not accepted, so claimed attachments are released.
  • Not refused, on purpose: the server's own follow-ups to a run (steer follow-up, usage-limit and restart continuations, notifications, idle compaction). They already settle quietly on an archived thread, and an existing test (runtimeLayer.test.ts, "guards a scheduled usage-limit continuation against archive") fails if they raise instead.
  • Web: the composer of an archived thread shows a notice above it ("This thread is archived and cannot receive messages.") and the send button is disabled with the same reason. This also covers the thread archived by a project removal, since it is archived. Reused: the composer banner stack (ComposerBannerStackItem) and sendDisabledReason; nothing new was built.
  • Tests: ThreadLaunchService.test.ts sends the web composer's command into an archived board thread and an archived ordinary thread (both refused, no run, no message) and into a live board thread (one run). ChatView.archivedSend.browser.test.tsx opens an archived board thread, checks the notice and the disabled send, presses Enter, and checks the text is kept and nothing was dispatched; a live board thread still sends.

Finding 4, MCP list fd32da6277. listProjectThreads with includeSubagents: true now drops any thread whose root is server-owned, the shell's family rule. Test: a seat under a marked thread is listed neither way; a seat under an ordinary thread still is with subagents included.

Finding 5, wording 97b8011ba3. The first line is unchanged. The board line now ends "Board and engine conversations already archived are kept.", which is the exception to "any archived threads are deleted" above it. One message changed, in the ten web and ten mobile catalogs, machine-translated like the web notice above; no native speaker has read either. i18n:check passes.

Verification

  • Typecheck: server, web, client-runtime, exit 0. Lint and format on the changed files; knip in the commit hook.
  • Server: all 85 test files that import Orchestrator.ts, ThreadManagementService.ts or ThreadMessageIntake.ts, two workers. 2440 passed, 3 failed. Two are the statement.stat is not a function failures of this machine's Node (Orchestrator.control-reads.test.ts), the same cause QA recorded. The third was mine and led to the narrowing above. After narrowing I re-ran 11 files (592 tests, all passing), not all 85 again: the change only refuses less.
  • Web: ThreadRouteView.browser, ThreadRouteView.tagFilter.browser, ChatView.archivedSend.browser, ProjectSettingsPanel.logic (21 passed). ChatView.browser.test.tsx: 17 of 18, the known rightPanelOpen failure.

Not verified

  • No pass in a real client, and no dev server started. No image of the notice.
  • Mobile shows no notice. The server refuses the send there too; what the mobile composer shows for that refusal was not checked.
  • An ordinary archived thread on web still lands on /, so its composer cannot be reached; the rule for it is tested on the server only.
  • A seat whose board root is archived while the seat is not: the list reads active threads, so the root is not in it and the seat would be listed. Project removal archives a root's seats with it; I did not check every other archive path.
  • Other browser tests that mount ChatView were not all run.

Independent QA (round 3, fix round), head 97b8011ba3, 2026-11-22

Tester: Claude Opus 5.5 (Claude Code), a different model from the builder. No product code changed, nothing pushed or merged. Same worktree and seeded data as round 2, moved to 97b8011ba3. One dev server, real web client in T3's Browser panel, run before any test batch. New fixtures for this round: project qa-remove-2 with a control thread, a live board thread (puffin) and a board thread archived through the web UI's own Archive action (gannet). The markers were written into the projection rows with the server stopped, as in round 2.

Machine: swap was nearly full again (about 25 to 26 GB of 26 to 28 GB used), and the client showed "froze for 5s/8s" toasts during the pass. Every step below completed; nothing was skipped for memory.

Verdict: pass with findings, 0 blocking

All five round 2 findings are fixed as described. What is left is small.

Findings, ranked

1. Low, from the code, not shown live. The new "already archived are kept" sentence only shows when the project also has a live board thread. ShellStream.shellTrimFacts counts active (unarchived) rows only, and planProjectRemoval adds the board line only when that count is above 0. So in a project whose board threads are all archived, the dialog says only "Those threads and any archived threads are deleted from T3 Code", while removal keeps those threads. Round 2 finding 5 in that one case. Repro by reading: a project with one control thread and one archived board thread and no live one gets keptCount 0, so no board line. Fix options: count archived server-owned threads too, or show the "kept" sentence whenever the project has any.

2. Low, disclosed by the builder, confirmed by reading. listProjectThreads with includeSubagents: true reads active threads only, so a seat whose board root is archived while the seat is not would still be listed. Project removal archives seats with their root; other archive paths were not checked.

3. Nit. On an archived board thread (walrus, otter) the browser tab title stays "T3 Code (Dev)". A live board thread gets "project / title". This is probably because there is no shell row. It may predate this PR; I did not check that.

Verified live

  • Fix 1, tag filter. This is the round 2 setup exactly: filter on (hideTaggedThreads: true, hiddenThreadTags: ["private"]), and zebra is a live board thread tagged private. Zebra now opens by deep link and full reload, with its history and composer, and stays out of the sidebar. Screenshot: shots/01-zebra-tagged-board-opens-filter-on.png. Controls under the same filter: ordinary tagged ibis goes to / and then to a draft. Ordinary archived heron goes to /. Ordinary listed alpha opens.
  • Fixes 2 and 3, archived threads receive nothing.
    • Walrus (archived board thread, project present). Opened by id. The notice "This thread is archived and cannot receive messages." sits above the composer. The submit button is disabled, and its accessible label is the same sentence. I typed text and pressed Enter, and the text stayed in the composer. I also forced form.requestSubmit() past the disabled button, and that sent nothing either. On the server, walrus still had 6 messages and 2 runs, and the event log gained no entry for it. Screenshot: shots/02-walrus-archived-board-notice.png.
    • Otter (archived by a project removal; its project no longer exists, "No project" in the status line). This used to be a silent no-op, and now it gets the same notice and a disabled send. Enter and a forced submit both kept the text. On the server, otter still had 9 messages and 3 runs, with no new event. Screenshot: shots/03-otter-archived-by-removal-notice.png. The composer also still held my round 2 draft text, which is the normal draft persistence.
    • Live board thread puffin still sends. Enter sent the message: messages went from 2 to 4, and a second run completed. Screenshot: shots/04-puffin-live-board-sends.png.
    • CLI. t3 thread send into walrus, otter and ordinary archived heron is refused each time with "Thread ... is archived and cannot receive messages." No message was stored.
    • Ordinary archived threads. Heron redirects to /, so its composer cannot be reached from the web, as expected. Nothing odd turned up.
  • Fix 5, removal dialog, on qa-remove-2 with a control, a live board thread (puffin) and an archived board thread (gannet). The dialog says: Remove project "qa-remove-2" and delete its 1 thread? ... Those threads and any archived threads are deleted from T3 Code. Provider sessions are kept. 1 conversation started by a board or the engine is archived, not deleted. Board and engine conversations already archived are kept. Screenshot: shots/05-removal-dialog.png. The database matches after Confirm:
    • The control is deleted.
    • Puffin is archived (archived_at set at removal, deleted_at null).
    • Gannet is kept with its original archived_at (05:35:09, before removal) and not deleted.
    • The project is deleted.
  • Fix 4, MCP list. Not proven live: I still could not build a live seat rooted at a board thread, because the Replay provider gives no seat. I accept the server test as evidence. ThreadManagementService.test.ts and ThreadLaunchService.test.ts pass (114 tests). I also ran a mutation probe in my local worktree, restored afterwards: with the root check removed, the new test fails (board-seat listed), so it catches the bug.

Regression, brief

  • The sidebar shows no board thread (zebra, walrus, otter, puffin, gannet) with the filter on or off. Opening one by id does not add it.
  • In the command palette, searching zebra, walrus, puffin, gannet and otter by title, and quokka, narwhal and capybara by message text, finds nothing. Alpha is found as the control. Recent threads lists no board thread.
  • With the filter off (hideTaggedThreads: false, server restarted, ibis back in the sidebar), zebra and walrus open by deep link and full reload. Walrus still shows the notice and the disabled send.
  • The removal count matches the database (above).

Not verified

  • The new Orchestrator refusal in a live run. The web client blocks the send before it leaves, and the CLI is refused earlier by ThreadManagementService's own check, so no live send reached the new guard. Only ThreadLaunchService.test.ts exercises it, and it passes. MCP t3_thread_send was not tried, because my MCP tools reach the host environment, not this dev server.
  • Fix 4 live (see above). Also finding 1's case live: a project with only archived board threads.
  • Mobile, including what its composer shows when the server refuses. Desktop. Translations: no native speaker has read them.
  • ChatView.archivedSend.browser.test.tsx and the other web browser tests were not re-run; the live pass covered the same behaviour.
  • Whether a marker written by NativeIntake behaves exactly like my direct projection write.

Screenshots are on the test machine under /private/tmp/claude-501/-Users-tevinmuparadzi-Projects-Operation-PaperClip/ba95e9e4-2f49-4262-8edf-602ebaa9d4ff/scratchpad/shots/, not uploaded.

Claude Opus 5.5, Claude Code.

Tevin2119 and others added 8 commits November 22, 2026 03:44
…l open by id

QA finding 1: every un-archived thread carrying the server-owned marker went to
every client in the shell snapshot, where clients only hid it. A subscriber
that reads thread families now gets no such thread, nor the seats and side
chats rooted at one, in the socket or HTTP snapshot. The archived snapshot
leaves them out too. A thread that delegated a task to a linked environment
stays, so that task still nests under it. Changes to a marked thread still
arrive as deltas, which keep an open one current.

QA finding 2: with a tag filter on, a thread the shell leaves out opened only
as the child of a visible parent, and a board thread has none. A server-owned
thread now opens by id on its lineage, which never changes, and an archived
one opens too: no archive list shows it, so its id is the only way in.

The routing probe covers tag filter on and off against a row that arrived as a
delta, a row the shell left out, and an archived thread.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…a reconnect

QA finding 3. A seat that runs in a linked environment opens in the direct
thread view or in the read-only linked view, two components. The choice was
made on every render from the connection and the link list, so the centre
flipped from one to the other when the links answered after the view was
mounted, and twice more on a short reconnect. Each flip remounted the
conversation and lost composer focus, text and scroll.

A newly opened seat now waits until the links have an answer before either
view mounts, and shows the seat bar with a loading line, not the task's own
thread. An open direct view stays through a dropped connection and gives way
to the linked view only after the connection has been down for ten seconds.
A connection of the client's own still replaces the linked view at once.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…guard the marker

Project removal. The removal confirmation counted listed threads only, so a
person was not told a project had board task conversations, and the server
then deleted them out of sight. The shell snapshot now counts the threads
carrying the server-owned marker per project, and the web and mobile
confirmation says how many are archived, not deleted. On removal the server
ends their work the way a deletion does and archives them, with the seats and
side chats under them; their attachments stay. An archived one is left as it
was. The count is exact at the last shell snapshot.

Marker guard. A contract test reads the source and fails when anything but
the engine intake reads the marker's value, or when anything but the marker
module knows the board-task form. Every other reader asks whether it is set.

German catalog: "Phase nicht bekannt" becomes "Phase unbekannt". The new
confirmation line is translated by the model in all ten catalogs of web and
mobile; no native speaker has read it.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…f the board work

Mobile: a link to a board task's thread, read with the Thread route's own
path, selects that thread when the shell left its row out, when the thread is
archived, and when its row came in as a delta. The probe stands in for the
environment at the two reads the route makes (the shell row and the thread's
detail); mobile has no fake socket to render a whole screen against.

Reverse states on web: a seat whose saved environment goes offline with no
link to it stays in the direct view with its composer, and ordinary threads
are unchanged under a tag filter: a listed one opens, a hidden one and an
archived one stay closed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The tag filter closed a board task's thread that carried a hidden tag: opening
it brings its row into the client, the row carries the tag, and only the
off-shell check exempted server-owned threads. The filter hides rows of lists,
and this thread is in none, so its lineage now opens it past the tag check too.
An ordinary tagged thread still lands on /.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The web composer ran a full turn in an archived board task's thread, and did
nothing at all, silently, in one whose project was removed. The CLI already
refused the same send. The orchestrator now refuses a message sent to an
archived thread from any entry point, with the CLI's sentence; the server's own
follow-ups to a run still settle quietly there, as before. The web composer of
an archived thread shows a notice and a disabled send, so the refusal is seen
before a message is written. The notice is translated by the model in the ten
web catalogs; no native speaker has read it.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…rver-owned work

With subagents included, the project thread list (the MCP t3_thread_list path)
dropped a board task's thread but still listed a seat rooted at it. It now
follows the shell's family rule: nothing rooted at server-owned work is listed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
"Any archived threads are deleted" sat above the line about board and engine
conversations being archived, and one of those already archived is kept, not
deleted. The second line now ends with that exception. Translated by the model
in the ten web and ten mobile catalogs; no native speaker has read it.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@polymaniadeveloper-hub
polymaniadeveloper-hub merged commit 360306b into master Nov 22, 2026
22 checks passed
polymaniadeveloper-hub pushed a commit that referenced this pull request Nov 22, 2026
Brings in the merged hardening round (#187).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants