Repository navigation
fix: every app shows the same inbox for the same machines - #92
Conversation
Machine owners were a per-device client setting, so two apps credited the same pre-inbox threads to different founders. The Settled count added the servers' unfiltered older-settled totals to filtered loaded rows, so it matched no inbox and differed from the servers' history. Each machine now reports its owner (`ownerFounder` server setting, advertised by `machineOwner`), and clients resolve attribution with it, falling back to their own map only for servers too old to report one. `olderSettled` counts carry groups by primary inbox and sender, using the same sender rule as clients, so the Settled total applies the list filter to threads no app has loaded. Counts from older servers show as "N+" under a partial filter. Model: claude-opus-5-5 Task: fix Agent: claude-code Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Note 🤖 Claude Opus 5.5 responding on behalf of nullStack65 Bot review round 1 (Claude Opus 5.5, independent) @ e01b0ab Pinned: head Verdict: blocking, on one cheap item (#1, the repo's eval gate for changed MCP schemas). The design holds up for the main paths. Server-reported owners, plus groups that cancel each loaded older row once on each side, do give identical membership and totals across apps for the cases the new test covers. Findings #2 to #8 are non-blocking. Findings, most severe first
Checked and fine
Model-facing text changes
What I ranAt the PR head, with the repo's
Not run: web browser tests ( |
… machine owners Agents can no longer read or set the machine owner: `ownerFounder` is gone from `t3_environment_read` and `t3_environment_preferences_update`, so Settings is the only writer and no model-facing text changes. The server now resolves the sender of agent work without an owner from its parent chain and stamps it on the shell as `parentFounder`. Clients use the stamp, so a fork of an idle subagent is credited the same on an app that never holds the subagent. Older settled counts group by the stamp too, so the `limit: 0` recount reads only the requested projects' older rows, and parent chains are read only for ownerless agent rows. Snapshots group counts once, only when they omit older settled threads, and the shell index decodes inbox fields only for settled listed rows. The recount key also includes the routing of held parents a sender comes from. A lower-bound "N+" Settled header never reads below the rows on screen. Settings offers "Set on machine" when an upgraded machine has no owner but this device's old map chose one. Model: claude-opus-5-5 Task: fix Agent: claude-code Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Note 🤖 Claude Opus 5.5 responding on behalf of nullStack65 Repair round 1 @ d6cfb2f One commit on top of
Docs: Commands (Node 26.13.1, repo
Left open (judgment):
|
The one-inbox change moved the sidebar's inbox matching into the shared client-runtime filter (matchesListedInboxes over matchesPrimaryInbox). The guard looked for matchesInboxFilter in the web file, so it failed although the logic was intact. It now requires the shared matcher in the web file, guards the client-runtime filter file, and also requires matchesInboxFilter in contracts. Model: claude-opus-5-5 Task: fix Agent: claude-code Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Guard follow-up @ 4b3906eCI at e01b0ab failed in "Guard indispensable AssemblyAI voice implementation":
Ran the guard script locally at 4b3906e: |
|
Fleet build artifacts from
|
Two apps connected to the same machines could show different inboxes and counts:
founderByEnvironmentIdlived in client settings, so pre-inbox threads (and every thread on 0.1.9 servers) went to whichever founder that device's map named. Jeff's desktop showed dozens of threads in Kameron's inbox while Kameron's laptop showed 3.olderSettledper-project totals to the loaded rows that passed the filter, and subtracted only filtered loaded rows. So with an inbox or From filter on, the number didn't correspond to that inbox or to the servers' history.Fix
Machine owner is a fact the machine reports.
ownerFounder: Founder | null. It's a normalsettings:writekey, local and never shared-synced.machineOwner: truein their server config. That flag is needed because older servers decode the field as null without meaning it.resolveMachineOwners(client-runtime) uses the server's value, including null, whenever the server reports one. The legacy local map is used only for servers without the flag.One sender rule, resolved on the server.
resolveThreadSenders. Its result is symbolic: a founder,"machine", or null, plus byAgent.parentFounder. Clients use the stamp, so a sender never depends on which parents a client holds, for example a fork of an idle subagent that opted-in clients never receive. Clients walk parents only for older servers without the stamp."machine"through the merged owner map.resolveThreadAttributionsstays pure and takes that map.Settled counts are server totals under the same filter.
olderSettled.counts[]gains an optionalgroupsbreakdown: primary inbox, agent-created, sender founder, byAgent, count. It's built on all three server paths (full-shell snapshot, index-built snapshot, and thelimit: 0recount) with one function,countOlderSettledThreads, which reads only the older rows and their stamps.settledShelfTotal(client-runtime) runs each group through the same list filter as the loaded rows. Older rows a client holds in the shelf count once on each side, so the total is the same whichever pages are loaded.Why groups instead of "N+" everywhere: the founder asked for the full settled history. The breakdown is small: a handful of groups per project, computed from rows the server already reads. It's also the only way to give exact per-inbox and per-sender totals for rows no client has loaded.
Why no mesh index: routing and settled state already live on each thread's own server, and every app reads them from there. The two divergences above were both client-side derivations: a local owner map, and unfiltered totals mixed with filtered rows. Once the machine reports its owner and the server resolves senders and breaks down its own counts, every app computes the same answer from the same server data.
Mobile: mobile has no inbox or From filter and no attribution. Its Settled counts are unfiltered and already use the server totals, so nothing changes there beyond a test fixture.
Agent tools: unchanged. The MCP tool list and
agentContextBudget.jsonare identical to the base.Tests
packages/client-runtime/src/state/inboxConvergence.test.ts(new, 10 tests). Two simulated apps get the same server data: four machines (two report owners, one reports a cleared owner, one 0.1.9 server reports nothing), threads with and without inbox fields, agent threads whose founder comes from a parent, forks of an idle subagent that one app never holds, and a pinned older thread. The apps differ in their legacy owner maps and in which older pages they loaded. For seven filters both apps give identical membership, attribution and Settled totals, equal to a ground-truth app that holds every thread. Totals-only servers give an "N+" that is at least the rows on screen and at most the truth.olderSettledThreads.test.ts: the recount key changes when a held parent (or grandparent) a sender comes from is rerouted, and a lower-bound total never reads below the loaded rows.ShellStream.test.tsgroups by stamped parent founders.ProjectionStore.test.tschecks the stamp on snapshot and single-shell reads (including the subagent-fork chain), that the recount groups through it, that the recount query is scoped to the requested projects, and that a parent chain is read only when an ownerless agent row needs it.serverSettings.test.tsadds an owner round trip: set, persist, client view, clear, scope =settings:write.http.test.tsandws.test.tsare updated for groups.Commands run (
vp test run, Node 26):inboxConvergence,olderSettledThreads,threadAttribution: 28 passed.ShellStream,ProjectionStore,serverSettings,http,ws,ThreadSearch, environmenthandlers,AgentContextBudget: 229 passed.threadListFilter.logicandscopedSettings: 36 passed.SidebarandolderSettledShelf: 10 passed.useThreadListV2: 1 passed.vp linton the touched files reports one existing warning inProjectionStore.tsthat this PR doesn't touch.No real-client pass was done.
Migration
Set each machine's owner once, from any app connected to it with settings access (Settings → Threads → Machine owners), or click Set on machine where a device's earlier choice is shown:
Until a machine runs this version and has an owner set, its pre-inbox threads show as Unknown on every app. That's the same everywhere, rather than different per device. Machines still on 0.1.9 keep using each device's own map, so give those maps the same entries or upgrade them.
Model: claude-opus-5-5
Task: fix
🤖 Generated with Claude Code