Repository navigation
Releases
v5.1.75
Fixed
omo remote devicesshows this device offline while its device service is not connected, andomo remote sessions --devicenaming a device the account does not have reports it instead of listing nothing. (#13916)- Memory sync's own test suite no longer fails a later, unrelated test when one test times out while a sync cycle is still running: device teardown in the tests now stops the engine and waits for that cycle before removing its files, and the push-bounded test waits on the cycle instead of a fixed 30 s budget. (#13887, fixes #13680)
- The first memory sync cycle after signing in no longer fails with "the device's sign-in token could not be read or refreshed" when the system keychain briefly reads nothing. That one empty read is now checked again before sync treats the sign-in as gone; signing out still stops sync at once. (#13879, fixes #13739)
- Regenerating the computer-use prelude (
bun run generateinsenpi-desktop-prelude) now runs its generated method tables throughbun fmt, soprelude.js, itsindex.tsallowlist and the desktop protocol tier tables stay formatter-clean instead of reverting to unformatted output at every release regeneration. The published prelude JavaScript is the formatted copy again. (#13855)
Changed
-
A memory sync capture that receives files from other devices no longer scans every known file once per received file, so a large memory catching up on many incoming files stays fast. (#13862)
-
2ec6b5d ci(publish): operator-built darwin artifacts for the v5.1.75 release (workflows from #13928, merge 8f8f4d1)
-
89bbf71 Merge pull request #13887 from code-yeongyu/fix/memory-sync-push-bounded-test-13680
-
598ec23 Merge pull request #13862 from code-yeongyu/fix/memory-sync-capture-known-set
-
c3304b6 Merge pull request #13879 from code-yeongyu/fix/memory-sync-first-receive-token-13739
-
8342246 Merge pull request #13870 from code-yeongyu/docs/13855-pin-comment-wording
-
9ed8307 Merge origin/dev into fix/memory-sync-first-receive-token-13739
-
bc1abda Merge origin/dev into fix/memory-sync-push-bounded-test-13680
-
8c3f1a1 Merge origin/dev into fix/memory-sync-capture-known-set
-
821eafd Merge remote-tracking branch 'origin/dev' into docs/13855-pin-comment-wording
-
31e08cb Merge pull request #13916 from code-yeongyu/rt/13414-client-parity
-
984088b Merge origin/dev into fix/memory-sync-first-receive-token-13739
-
9bd2382 Merge pull request #13855 from code-yeongyu/fix/prelude-fmt-clean-regen
-
f865a7a Merge pull request #13918 from code-yeongyu/ci/platform-decision-body-comment
-
ebeaac0 Merge remote-tracking branch 'origin/dev' into fix/prelude-fmt-clean-regen
-
03e4b7e ci(pr-platform-decision): restore why the PR body is read from the event file (#13859)
-
f9f610d Merge pull request #13912 from code-yeongyu/ci/platform-decision-one-cap
-
add0ca2 docs(changes): fragment for remote client read-path parity (#13916)
-
bca0d62 Merge branch 'dev' into rt/13414-client-parity
-
72cecae fix(remote): the client's devices.list and sessions.list carry #13402's CLI behaviour (#13414)
-
02915ac Merge remote-tracking branch 'origin/dev' into fix/prelude-fmt-clean-regen
-
eac1e0a ci(pr-platform-decision): one file-cap number, owned by the classifier (#13859)
-
8694906 Merge origin/dev into fix/memory-sync-capture-known-set
-
d81b80a Merge remote-tracking branch 'origin/dev' into fix/prelude-fmt-clean-regen
-
2fc39fc Merge origin/dev into fix/memory-sync-first-receive-token-13739
-
00026f3 chore(changes): the fragment carries this PR's number (#13887)
-
223aefb test(memory-sync): teardown waits for a running cycle; push-bounded waits on the cycle
-
ea76605 fix(memory-sync): re-read a transiently empty credential once before retiring the capture
-
4e51d43 Merge remote-tracking branch 'origin/dev' into fix/prelude-fmt-clean-regen
-
7560233 docs(desktop-prelude): describe the whole-file bun fmt pass in the pin fixture comment
-
d15ec64 perf(memory-sync): look up known keys in a set when picking joining files
-
2a9d548 chore(fmt): drop the generate-assets.ts output ignores from .prettierignore
-
45ed149 Merge remote-tracking branch 'origin/dev' into fix/prelude-fmt-clean-regen
-
8c82979 fix(desktop-prelude): format the generator's outputs with bun fmt and regenerate them
bun add -g omo-aiv5.1.74
A secret input or grant approval that settled while the relay link was down is now dismissed on the terminal that raised it after omo remote attach reconnects, instead of staying open with an answer the session would refuse. (#13773)
The computer-use conformance preflight test cleans up its compiled stand-in engine reliably on Windows. The test's temporary directory is now removed with the shared retrying helper, because Windows briefly keeps a just-run executable locked and Bun ignores rmSync's retry options. QA tooling only; no runtime behaviour changes. (#13299, part of #13221)
Fixed
- The EQ.omo driver resolves a symlinked temp dir before it makes its sandboxes, so a recording on macOS no longer carries
/private<sandbox>paths or an untrusted project. A new test fails when a platform's EQ.omo goldens were recorded with a senpi version more than a week away from the pinned one. (#13865) - CI: the desktop-bench comparison summary prints its regression thresholds (x1.3 and +0.5 ms) instead of
undefined. It now reads them fromcompareBench's result rather than destructuring nameseval-core.mjsdoes not export. (#13838, follow-up to #12876 review L1, #12567) omo remote bootstrapinstalls device service 0.1.39 instead of 0.1.27, which was older than the wake and delivery minimum and answered open/create to the device withunsupported. (#13803)- A wake toward a device service older than the wake minimum now names the service's version, the minimum and the fix, instead of a bare
unsupported. (#13803) omo remote deviceslists the device it runs on first (the account roster held only the other devices), and--jsonmarks each rowthis_device.omo remote sessions --devicewith a device that is not on the account now fails with exit 2 naming it, instead of printing an empty list. (#13458, part of #13402)thread_interruptalso stops the background jobs the interrupted turn started (a bash command it ran in the background or that auto-detached), so the thread no longer wakes a minute later when that job ends and replies as if it had never been interrupted. The result names what was stopped asbackground: { status: "stopped", stopped, failed }; an engine that predates the stop answersbackground: { status: "not_supported" }. A thread on this device and a thread on another device run the same interrupt code and answer the same shape. (#13444)- Computer use v3 (fake backend): the source guard that bans environment reads now also reports a grouped
use std::{...}import of the environment module when a//or/* */comment comes before the item inside the group. A comment can no longer hide an item, or split one at a comma or brace inside it. (#13363) - Desktop engine: the client retries a connection reset before the first reply byte only on Unix sockets, never on a Windows named pipe, where the daemon may already have run the request. (#12744)
Changed
- The
thread_readwait limit (600,000 ms) is now internal to the thread schema instead of an unused export; the limit itself is unchanged. No behavior change. (#13864) - Internal: the Windows backend documents when its posted key route still sends a key under a held Alt: for modifier-only chords, when releasing held keys, and while a failed Alt release is still recorded as held. A Windows test is also renamed to match the order of its checks. Nothing changes for users. (#13824, fixes #13821)
- Internal: the module doc of the X11 input's client-process reader now names the
/proc/<pid>/exeread added for the background-input check. Nothing changes for users. (#13820, fixes #13818) /remote enable,/remote disableand/remote controlin the TUI now answer in a sentence ("Remote sessions are on for this device.", "Remote control is off: ...") instead of the CLI's raw JSON. When/remote enablecannot finish, it shows the CLI's reason (for example "Could not download the OmO device service.") instead of an exit code, and a warning the CLI prints while succeeding (for example that the device is not linked) now shows after the result. (#13796)- The remote module's public types no longer depend on internal modules, so the importable remote SDK ships clean type declarations. No command or tool behaves differently. (#13782)
- The remote module gains its client factories, one per kind of caller (a person, including
omo remote; a session; a connector; the thread SDK's own principal; the device service's bootstrap entry), and its public entry. Its alias operations work when the caller hands in the account's alias registry, naming each session by the same durable id the thread tools use, and a device that does not answer is reported as offline rather than as having no such session. Nothing calls them yet, so no command or tool behaves differently. (#13750) - The remote module now declares every operation a surface needs (part 4a-2): opening, messaging, reading, following, waiting on, interrupting and closing a session on any device, its model and thinking level, attaching with pending approvals, aliases and the wake switch. The operations that have not landed yet answer "unsupported" as data, and nothing calls them yet, so no command or tool behaves differently. (#13747)
- Remote access now has its own module (part 4a-1 of the one remote module design): typed results, one code list, a deadline table and a clock seam, with the service, devices and session list members. Nothing uses it yet, so no command or tool behaves differently. (#13498)
- The device service client takes one injectable clock for its handshake, request, endpoint-command and attach waits, so each of those deadlines can be tested on fake time. Real timers stay the default and no wait changed. (#13473)
- Computer use v3: on Windows, one user's two logon sessions take turns at the computer instead of acting at once: the interaction lock is now the operating system's file lock on
%LOCALAPPDATA%\OmO\ComputerControl\run\interaction.lock, in a directory only that user and SYSTEM can open, on a local NTFS or ReFS volume. A session without a runtime directory cannot act on a real machine; it can still read. (#13367) - The Windows live tier's ledger job now fails when a fresh measurement's verdict differs from the committed
crates/omo-computer-backend-win32/probes/hosted-windows-live.tsv, naming each leg that moved, so the committed ledger only changes on purpose. (#13274, part of #13221)
Added
- Computer use v3: the session crate's side-effect funnel lands. Every mutating call runs the gate, writes a pre-input marker before the first OS call on every transport, performs the call, reads back and writes one audit event per side effect; a crashed call is never resent and reads
delivery_uncertain. Audit, lock, permits, batches and artifact retention come with it, and a typed secret is audited by length only. Nothing user-visible changes yet; the existing desktop engine is untouched. (#13367) - Computer use v3: the safety crate
omo-computer-safetylands. It holds the fixed gate order (each gate refuses before any OS call with nothing sent), the supervisor whose stop commit releases every hold within 1 s or exits, the stop latch and persisted stop marker under the per-uid state root, thelive/deaf/nonestop paths with the host-relay-only policy, the per-OS protected tables and dangerous chords, and the self set. The walking skeleton'sclickandtypeTextnow pass through the gate. Nothing user-visible changes yet; the existing desktop engine is untouched. (#13353) - Computer use v3: the walking skeleton lands.
omo-computer-engineruns on the fake backend with--stdio(NDJSON JSON-RPC),--schema,--selftestand--version; the session answerssession.open,capabilitiesand close plus the thinclickandtypeTextverbs; refusals go through the closedCOMPUTER_*table.script/qa/computer/replay.tsreplays conformance cases against the engine binary and prints the JSON pointer of the first mismatch. Nothing user-visible changes yet; the existing desktop engine is untouched. (#13352) - Spike SP-R probe and ledger for the computer-use relay pipe: ...
v5.1.73
Remote session views stop receiving each streamed delta twice. open_session for a remote create, open, or fork now goes on the endpoint dialer (a fresh connection per call, closed once it settles), instead of sharing the long-lived connection the later session.attach rides on; opening on the shared connection had attached the host's delegated engine connection to the session as well, so the engine wrote every delta twice. A connection a caller lends is now named borrowed, so the CLI's one shared connection survives the open. (#13761, adopting #13075 by @surlymochan)
The first session on a newly enabled device now starts memory sync. Sync used to start before the memory repository existed, fail to write its space binding, log "memory sync failed to start" and never retry; it now waits for the repository and starts on its own once the first memory write creates it. (#13493, fixes #13486) (#13627)
Memory sync drops a replaced desktop account cleanly. When the desktop moved to another account without announcing a retirement, the engine opened the new account but left the old one half alive: it kept listening on the desktop channel and its open sync sessions did not end. The old account now ends as soon as the engine sees it is no longer current, which also ends its sessions. In a session Worker, an operation started just before the desktop retired its account can no longer hand back a handle from the old account. (#13334, follow-up to #13327 and #13329) (#13627)
OmO is ready to load on the tool side of a split session: its first-paint startup gate is declared for the agent-loop side, so the tool side never registers those handlers, and on the tool side deferred startup work runs right away instead of waiting for a paint that never comes; local sessions register exactly as before. (#12329) (#13627)
Computer use is ready for the tool side of a split session: foreground control is revoked, and the engine session ended, when the run of the session holding it ends or that session closes, through that session's own lifecycle instead of loop events; local sessions behave exactly as before. (#12328) (#13627)
Adds the executable proof that computer use returns one screenshot per tool call, under today's capture caps and unchanged across the tool side of a split session; no behavior change. (#12337) (#13627)
Computer use is ready for the tool side of a split session: there, foreground control follows the device's own exec setting (refused until exec is enabled on that device, then granted to any session without a per-session confirm); local sessions keep today's confirm exactly. (#12332) (#13627)
Computer use is ready for the tool side of a split session: sessions sharing one device take its foreground one at a time, and only the holder can release it; local sessions behave exactly as before. (#12339) (#13627)
Computer use is ready for the tool side of a split session: a missing OS permission (Screen Recording, Accessibility) raised there is reported to the session that owns the call, once per permission, without a session marker on that side; local sessions behave exactly as before. (#12450) (#13627)
Computer use v3: the result contract refuses a receipt under an unverified delivery and types refusal reasons. An ActionReport that carries a consumption receipt under not_delivered, delivered_unverified or unknown/unknown (#13147) is now refused, both when it is built and when it is parsed from the wire. A backend refusal's reason is the closed Reason type, so a mistyped reason is a compile error. (#13137)
- CI: a computer-use bench regression job compares the desktop bench against a base and fails on a p50 regression above 30%; it runs nightly against the newest release tag and on manual dispatch, not on pull requests. (#12876)
A question or approval a remote session asks now reaches every omo remote attach on it, not only the one that sent the prompt. The terminal UI shows the request and sends your answer back by its request id, as it already did for prompts you sent yourself, and dismisses it when another device answers it first. A secret input or a grant approval is still shown only on the terminal whose prompt raised it. A request already waiting when you attach, or raised while the relay link was down, is shown too, and one that was settled during a relay gap is dismissed after the reconnect (#13474). (#12750)
Changed
- The ulw-plan plan template and reference, ulw-execute's decomposition step, and the Codex edition's ulw-loop reference and plan agent no longer say "target 5-8 per wave; fewer than 3 means under-splitting". A wave is sized to the tasks that are independent of each other. (#13851)
- The mass-ulw skill's listing description now says a graph is sized to the request with one run per phase, matching the rewritten skill body; it no longer says each phase's stages are defined up front in one run. (#13848)
- ulw-loop no longer tells HEAVY goals to target 5-8 tasks per wave or treats fewer than 3 as under-splitting; a wave is sized to the tasks that are independent of each other. Mass research gives lead nodes the same handoff file and source tagging as harvest nodes, and both ulw-research editions make slice reducers conditional like the mass-ulw planning reference does. (#13847)
- mass-ulw now sizes the workflow graph to the request: a small graph by default, a verification node only when the final check is too long to run inline, per-item chains instead of a review node between waves, and mass research that starts from the angles the topic has instead of a 60-angle floor. The skill and its planning reference were rewritten and are about a quarter shorter. (#13834)
- Internal: the desktop service's fake engine and test harness are now the
./testingpackage export ofsenpi-desktop-service, and the computer-use tests inomo-senpiandsenpi-desktop-toolimport them from there instead of a relative path into its test tree. Nothing changes for users. (fixes #13614) (#13712) - Memory sync: a test now covers a remote head that arrives on the stream while a sync cycle is running and is not applied by it, so the engine must run exactly one more cycle to pick it up; a lost wakeup there now fails the test. No behavior change. (#13656, fixes #11778)
- Remote sessions are now on by default. The moment
omo logincompletes it switches them on for this device and says so in one line ("Remote is on for this device: other signed-in devices can reach it. Turn off: omo remote disable"), and every omo start keeps the device set up.omo remote disableturns them off and remembers that, so neither a lateromo loginnor an omo start turns them on again;omo remote enableturns them back on (#13594). (#13602) (#13627) - device-service-core's local-API tests fail when omo starts sending a request type that has no reply row in the LA21 decoding test or that the conformance test owner still lists as unserved (#13491). (#13491)
omo remote, the thread tools and the memory sync use the device service's own local-API client, vendored byte for byte from omo-relay-daemon (main after #1115, plus the read-side tolerance of omo-relay-daemon#1130), instead of a hand-edited copy; a guard test keeps that client and the connection factories behind the device-service layer (#13475). (#13475)
Fixed
- A fanout that fails as a whole (an unknown or ambiguous target, invalid options, a duplicate target under
dedup: "reject") and a target that runs pasttimeoutMsget the same typed error, and a failedresultOfnames the request instead of treating its op id as a device. (#13813) - The shipped device account, device service, memory sync, wake, thread client and bunshin bundles, and the omo-ai launcher's sign-in module (
omo login/ `log...
v5.1.72
A thread send now reaches a session that is not running, on this device or on another of your devices, and starts it. It needs senpi 2026.11.15-2 or later (omo pins it), and device service 0.1.31 or later on both devices for a wake from another device. A reply to a sender on another device reads only that device's sessions instead of the whole account listing; a send to a live session that is paused or holds a Stop answers queued with waiting_on: { kind: "hold", device }; and waiting_on.kind in a thread_send result can be approval, hold or wake. (#13599)
Changed
omo remote enablenow sets the device up in one command afteromo login. It installs the OmO device service when none runs, registers this omo's agent dir with it, turns remote sessions on, and installs and starts the always-on session host unit. If you signed in before the device service was installed, it also links that sign-in to the service, soomo loginthenomo remote enableleaves the device connected. Running it again changes nothing that is already in place (#13594). (#13602) (#13606)omo remote statusalso shows this device's own units as the service manager reports them: the session host unit (running, enabled, pinned host), the device service's unit, whether the cold-wake entry is in place, and the agent dir omo registers. With--jsonthese are alocalobject beside the status fields (#13594). (#13602) (#13606)- Thread gateway: waking a session that is not running is on (#13037). It runs on senpi 2026.11.15-2 or later (omo now pins it), the release whose session release drops a woken session's retained record. A message from another of your devices wakes a session here once both devices run device service 0.1.31 or later. The local omo daemon (
omo daemon run, or the device service's cold-wake unit) registers this device's agent dir with the device service and takes deliveries as soon as it starts, with no session open. (#13599) - Thread gateway: the operator daemon declares the delivery consumer as soon as it loads omo, with no session open (#13037 cold wake).
omo daemon runsends the daemon it ensured senpi'swarm, so a daemon the device service starts on demand can be handed deliveries before any session runs. A warm that is refused or unsupported prints one line and the run still exits 0. (#13596) (#13599) - Thread gateway: the session wake loads from its own sidecar (
extensions/omo-wake.jsbehind#omo-wake-runtime): the carrier, the delivery consumer and the device-service client never enter the mainomo.jsbundle, which carries only lazy shells; the ensured host's idle-exit setting is read inside the sidecar. (#13456) (#13599) - Thread gateway: waking a session that is not running needs no setting (#13037). A
thread_sendfrom this machine to a session here that is not running now starts it by default, and the answer isqueuedwithwaiting_on: { kind: "wake" }.session_wake.same_devicestays in the configuration schema only as an explicit opt-out (false), read for the target session's own project. The wake starts on senpi 2026.11.15-2 or later (senpi#6449), which omo now pins. (#13456) (#13599) - Thread gateway: when no operator host is running, a session that is not running still starts (#13037). Before it opens the session, the carrier asks the engine's own
host ensure --policy never, from omo's daemon launch spec, on the socketthread_createuses. A running host is reused whatever its build; one is started only when nothing answers. The ensured host runs withomo daemon's own host environment: the agent directory is named (OMO_AGENT_DIR) and the configuredtask.host_idle_exit_msrides along (SENPI_RPC_HOST_IDLE_EXIT_MS) instead of the launch spec's default, and from a source tree the launch spec resolves to the plugin's tracked copy so the CLI is always handed a file that exists. (#13456) (#13599) - Thread gateway: a delivery from another device is no longer gated on the device service's wake switch being told on. While this engine's delivery consumer is served (from its serving acknowledgement, before its startup pending sweep, so a remote row that sweep wakes is never told the gate is off), the carrier opens the session. The consumer serves on senpi 2026.11.15-2 or later with device service 0.1.31 or later, and it no longer reads the wake the service tells: the wake switch it tracked (
wakeAllowed, thedelivery.wakeevent) had no production reader left. (#13456) (#13599) - The delivery consumer (design 5.4) runs beside the carrier. The engine minimum for a woken session,
MIN_WAKE_ENGINE_VERSION(A8.1, senpi 2026.11.15-2, the release that drops a released session's retained record), gates the carrier as well. A device service older than 0.1.31 does not carry the delivery protocol, so on it the consumer serves nothing and does not ask again. (#13455) (#13599) - The gateway store's schema moves to version 12: three nullable columns on deliveries and a unique index on the new ledger id. The first process on this release migrates the store; any process still on the previous build then reports the store's "schema too new" error on its thread tools and inbox drain until it restarts. (#13208) (#13279) (#13599)
- Thread gateway: once a build carrying this store ships, every gateway store open runs the schema step to version 13 (one new table,
woken_sessions) and raises the store'suser_versionfor good. An older build, or a rolled-back one, that shares the same agent directory then refuses to open that store with its "schema too new" error until it is upgraded again. (#13212) (#13279) (#13599) - A reply to a message that came from another of your devices goes to the session on that device that sent it, whether you reply by its session id or by the name it gave. A session on another device, or on this one, that only carries the same id or name no longer receives the reply. (#13218, part of #13037) (#13599)
Fixed
- An omo engine on a device whose device service runs on a non-default
--state-dirnow registers with it:deviceServiceStateDirtakes a non-empty absoluteOMO_DEVICE_SERVICE_STATE_DIR, then thedevice-service.location.jsondiscovery file beside the default dir, and only then~/.omo/device-service(#13597). (#13598) (#13606) - A remote worktree remove refused for an empty reflog now says reflogs may be off in the repository, instead of claiming they are: an expired or deleted reflog looks the same (#13542 follow-up). (#13546) (#13606)
- Thread gateway: the session wake keeps working after the session that started it ends (#13037). The carrier and the delivery consumer are one per agent directory for the whole process, but they used the first session's gateway store, which that session's shutdown closed, so every later wake in the process failed. The running wake now opens its own store for the life of the process. (#13596) (#13599)
- Thread gateway: the carrier re-checks the sender's peer admission at the moment of every open of a row from another device (#13456 review follow-up), so the retry and lease-takeover paths - which the accept-time check of the consumer and the before-turn check of the drain bracketed but nothing else covered - start nothing when the answer is refused or unknown. A lease takeover now consults the same allowance a wake does, ending its record when nothing queued allows the open; the runtime answers the carrier's admission ask on the live delivery port, so the startup pending sweep's own wakes are admitted too. (#13511) (#13599)
- Thread gateway: the session wake loads from its own sidecar (
extensions/omo-wake.jsbehind#omo-wake-runtime): the carrier, the delivery consumer and the device-service client never enter the mainomo.jsbundle, which carries only lazy shells. ...
v5.1.71
The builtin architect category now defaults to Claude Fable 5.2 at thinking level xhigh instead of max: both chain rungs (the 5.2 lead and the 5.1 fallback, kept as the fallback at the same effort) carry xhigh, verified against the engine model catalog per the #9378 rule. Docs that spell the builtin default follow. Other categories, agent chains, and model profiles are unchanged (#13543). (#13545) (#13563)
Memory sync names the last unexplained "authority is unavailable" causes, and its log stays free of memory text. When closing a sync operation, releasing its claim, or the request handler itself failed, Settings still showed the bare "Memory sync authority is unavailable." with nothing in the log. These now carry a code (close_failed, release_failed, handler_failed) and log a warning. Logged errors name the error type, not its message; only a sync service error keeps its service code and status, because other messages can quote memory content. A device or pending-request read whose failure returns after the service refused a request is logged again. Carried by the next train top's bundle regeneration. (#13332, follow-up to #13157) (#13563)
A signed-out desktop no longer looks like a memory-sync error to the engine. When the desktop had no signed-in account, the engine's main thread treated the desktop's refusal as a broken channel and the sync runner reported an error; it now reads it as signed out, as session Workers already did, and picks up the account once you sign in. Carried by the next train top's bundle regeneration. (#13329, follow-up to #13325) (#13563)
Memory sync keeps working when you sign out and back in on the desktop. Signing out and signing in again (with the same or another account) while the engine runs used to leave Settings > Memory > Sync on "Memory sync authority is unavailable." until the engine restarted. The engine now ends only the old account when the desktop retires it, ends that account's open sync sessions once, and picks up the new account on the next status read. It needs a desktop that keeps the private channel open across the rollover; with an older desktop the behaviour is unchanged. Carried by the next train top's bundle regeneration. (#13327, fixes #13325) (#13563)
omo remote accepts -- to end its options. omo remote send <device>/<session> -- --text that starts with dashes now sends that text, and the same works for a model id given to omo remote model; before, any argument starting with -- was refused as an unknown option. Options before -- still apply. (#13496, fixes #13431)
The omo remote reference matches the command again. It lists open, send, model, interrupt and close with their flags, and describes attach as a view you can type into; it no longer says the attach view is observe-only. The Senpi plugin README lists all 27 skills it ships (it said 19). (#13492, fixes #13433)
omo remote --help and omo remote <command> --help now answer. They print the usage (or that command's entry with its flags) on stdout and exit 0, before any sign-in or device-service check, instead of failing as an unknown argument. omo --help lists remote among the commands. -h now also answers as a help flag after any verb, including in a value position (for example omo remote open dev --project --help prints help instead of taking --help as the project name). (#13489, fixes #13405)
omo-opencode tests no longer depend on the order test files run in. Under bun, module mocks leaked into later test files whenever files ran in sorted order (as bun test --shard runs them): six tmux-utils tests replaced modules with partial mocks, two modules that tests mock re-exported functions (the team-mode member resolver's dependencies and the legacy plugin toast's entry migrator), and two claude-code-hooks tests module-mocked shared modules. The tmux mocks are now complete, those two modules define their own forwarding functions so a mock of them stays scoped, and dispatch-hook.test.ts and execute-http-hook-security.test.ts use spies that are restored after each test. Tests, plus two forwarding-only source changes (features/team-mode/team-runtime/resolve-member-dependencies.ts and hooks/legacy-plugin-toast/plugin-entry-migrator.ts); no runtime behaviour changes. (#13333, fixes #13331)
Changed
- OmO now pins
@code-yeongyu/senpiand its packages at 2026.11.15. That senpi release brings the permission-prompt abort fix, the tool-result slash-command injection fix, sign-in-with-ChatGPT scope and input fixes, and the binary update that reads the GitHub release. (#13563) - Daily · Normal now walks Claude Opus 5.5 (medium), then Claude Fable 5.2 (medium), then Kimi K3.1 Pro (max), matching senpi's shipped Opus 5.5 fallback chain; GLM 5.3 is no longer a Daily · Normal rung, so a GLM-only setup gets no automatic Daily · Normal model (#13562). (#13562) (#13563)
- Daily · Heavy runs Claude Fable 5.2 at high (was xhigh); its Fable 5.1 rung stays at xhigh (#13562). (#13562) (#13563)
@oh-my-opencode/utilsno longer exportsmigrateConfigFilefrom its package root or from its migration entry point (@oh-my-opencode/utils/migration): the re-export inpackages/utils/src/migration.tsis removed, and the root entry re-exports that module. Nothing imported it through either entry. The function itself is unchanged. (#13544) (#13563)- Internal:
omo remotestatus and session commands now read the account without loading the code that signs in or refreshes, and reach the device service through the local API client directly. Nothing changes for users; this prepares a faster start for these commands (#13440). (#13477) (#13563) - The gateway store's schema moves to version 12: three nullable columns on deliveries and a unique index on the new ledger id. The first process on this release migrates the store; any process still on the previous build then reports the store's "schema too new" error on its thread tools and inbox drain until it restarts. (#13208) (#13279) (#13563)
- Thread gateway: once a build carrying this store ships, every gateway store open runs the schema step to version 13 (one new table,
woken_sessions) and raises the store'suser_versionfor good. An older build, or a rolled-back one, that shares the same agent directory then refuses to open that store with its "schema too new" error until it is upgraded again. (#13212) (#13279) (#13563) - The standalone
omobinary starts about 4x faster on Linux and 6x faster on macOS: release binaries now carry Bun's bytecode cache, so a launch no longer parses the whole bundled program first (omo --version880 ms -> 202 ms on Linux x64, 214 ms -> 33 ms on Apple silicon). Each binary is about 43 MB larger, and the per-binary size budget rises from 210 MiB to 260 MiB. (#13464)
Fixed
- The remote worktree remove guard's comment no longer claims a failed open's worktree always has exactly one reflog entry: a repo with
core.logAllRefUpdatesoff reports none even for a branch just made, so there the worktree is refusedworktree_left_behindand stays for its local owner; a test pins that refusal (#13450 follow-up). (#13542) (#13563) - The bundle marker digest no longer folds the running node's
builtinModulesinto the digested build settings: the builtin family is recorded as one stable token and only source-derived externals are digested, so a byte-identical bundle built under node 26 (which addsnode:ffi) no longer reads as stale. A RED test pins the stability and that a source-derived external still changes the digest (#13155). (#13541) (#13563) - A memory reflection or dream launcher now hears its supervisor's exit even when the supervisor dies while the "launched" receipt is still being written. The launcher used to subscribe to the exit only after that write, so an early supervisor death was never reported and the launcher waited for an outcome that could not arrive: until the run deadline (15 min...
v5.1.70
Plain omo login signs a computer in from the terminal while OmO accounts are closed. A computer with no OmO desktop app — a server over SSH, a fresh Linux box — runs omo login, gets a device code to approve on another device, and stores the credential in the OS credential store; the attach-before-mint refusal (#13096) answers a computer that already has its device before anything is asked of the service. The success line names the signed-in account (Signed in to OmO as <email>.), and omo whoami names it too; a failed account read still says how to see it. omo logout and omo whoami run while closed, so no message names a command that refuses. login --help describes the route (the command stays hidden from omo --help), every other closed login form refuses with a message that names the route, the closed hints name both routes, and the compiled entry classifies bare login, logout and whoami as one-shot launches with the same shared parse the command uses (login -- included). The damaged-item message names the terminal route too instead of a desktop-app-only dead end. Tests run the flow through the npm launcher (against a service-auth bundle built from this source into a temp copy of the launcher tree), the compiled entry and the toolkit CLI, with an isolated HOME and runtime dir, no inherited session bus, an injected store, and a refusal-text assertion on every zero-mint case. (#13281) (#13328)
The computer-use conformance preflight tests pass on Windows. The missing-driver message now prints its repo-relative path with forward slashes on every OS, so a run reads the same on every host, and the stdio-probe test builds its stand-in engine as a real executable for the host instead of a POSIX shell script. QA tooling only; no runtime behaviour changes. (#13295, part of #13221)
Computer use gets a declared-outcome conformance catalog for its QA. script/qa/computer/conformance/ adds a typed catalog in which every action, toolkit and mode cell on each OS declares that it delivers or refuses with an exact engine code, a report that fails a missing, duplicate, skipped or mis-coded cell or one that disturbed the person's session, an environment preflight that ends a broken host's run as ENVIRONMENT_ERROR, and twelve model-free reference tasks. QA tooling only; no runtime behaviour changes. (#13241, part of #13221)
The desktop session's actionability wait test no longer fails on a busy host. The battery's wait now runs on an injectable clock: the host's monotonic time in a session, as before, and the fake backend's own clock in the test. So the_battery_waits_for_an_element_that_appears_enabled_after_500_ms checks the 500 ms wait in fake time instead of wall-clock bounds that host load could break. Runtime behaviour does not change. (#13180, fixes #13178)
A fuzzy handoff no longer fails on an account with many ended sessions, and never goes to a different session than an exact name would. A fuzzy handoff first tries the thread's exact id, alias or name across every session, ended ones on other devices included, so a name that exactly matches one session always reaches that session. Otherwise local, live and ended sessions are compared together and the clearly best match wins wherever it lives; two equally good matches ask you to pick. Ended sessions that cannot match no longer count against the 500-candidate fuzzy cap, so a large history no longer blocks a unique match. (#13141, follows #13066)
A thread whose alias claim asks for a sign-in after it was created is still named by the alias. When the account token expires between thread_create checking an alias and claiming it, the thread it already created is now named by the alias on its device and the call answers ok with alias_scope: "local", instead of an error. thread_rename does the same when the account is gone by the claim. An alias used as a local name is now trimmed on every path. (#13179, follows #13112) (#13141)
omo login no longer registers a second device for a computer that already has one. Before any code, browser tab or new key, it checks the stored sign-in and the device the computer's OmO device service holds (parked included). A held, uncertain, locked or unreadable state refuses a new device and keeps the existing one, instead of adding an "omo CLI on " row beside the desktop's. When the service advertises machine identity, both sign-in flows send a hashed hardware id (SHA256("omo-machine-v1" + id); the raw id never leaves the computer), and a machine_exists refusal ends the sign-in without removing or taking over any device. (#13096) (#13141)
A thread alias no longer blocks thread creation when no account is signed in. thread_create with an alias now creates the thread when the alias registry cannot get an account token. A desktop launch is always in that state. The alias becomes the thread's local name unless a name is given, and the result says the alias is local only (alias_scope: "local"). thread_rename with an alias does the same. Every other alias error is unchanged. (#13112, fixes #13104)
Fixed: file reads through the bundled bunshin SDK over the relay work again (readText, readBytes and read no longer reject with "truncated must be a boolean for byte results"), an op sent with a minted capability token is no longer refused as malformed, and an unsupported clientName now fails at once with invalid_client_name naming the supported role. (#13108)
The machine-binding module no longer exports an unused helper. deviceProfileOf from #13095 is now module-private, so the knip dead-code check on dev is green again. No behavior change. (#13097)
One computer is one OmO device: omo login binds to the machine's device service instead of registering its own. A sign-in now advertises machine binding. When the service asks for the machine hash, omo asks this machine's device service (device.machine_binding) for it, and computes the same hash itself when no device service runs. The sign-in presents the device keys this machine already holds, so the service keeps the machine's existing device instead of adding a row. If the device service holds a device whose keys omo does not have (the desktop app's sign-in), omo refuses and points to the desktop app instead of minting a second device. The raw OS machine id never leaves the computer. A computer with no readable machine id signs in exactly as before. (#13095)
The programming skill opens with code ownership. Its first paragraph tells the model it owns the code: fit new behavior into the existing structure, leave no duplicate or dead code, and fix a defect in the code being changed at its root cause instead of excusing it as pre-existing or someone else's. It replaces the "lazy senior engineer" persona line. (#13091)
A monitor no longer loses a batch that arrives while the previous one is being sent. flushMonitor deleted the monitor's pending entry after awaiting a dispatch, so a batch queued during that await vanished without counting as dropped. The flush now deletes the entry only when it is still the one it drained. (#13089)
A foreground task waits for its child instead of failing at once. Every task with run_in_background: false ended with Session error about a second in and aborted the child, because opencode reports a running assistant turn with error: null and the poller treated the key itself as a failure. Only an error that is set ends the child now. (#13076)
The beta channel no longer falls behind the stable release. Every stable release now also moves omo-ai@beta on npm and the beta installer channel on get.omo.dev to that release, unless beta already holds a newer prerelease, so bun add -g omo-ai@beta and install.sh beta never install something older than latest. The beta move and its beta >= latest guard now also run when a release is re-run after it was already published, and after the move beta is re-read: a beta someone moved further forward is left alone, while a beta still older than latest or the newest prerelease fails the run with the exact npm dist-tag add omo-ai@<version> beta repair command. (#12987, #13148, #13149, #12990)
- Computer use: the omo
computertool now runs through the desktop SDK; omo sessions open with an explicit report mode (sandbox default unchanged). (#12840)
Fixed
- The hashline test project's lockfile now resolves zod 4.6.7, matching the
^4.6.6its manifest asks for since #13151, so that project no longer installs the zod 4.6.5 affected by GHSA-mrhx...
v5.1.69
OmO runs on senpi 2026.11.10-2. A client can read and edit the workspace's instruction files over RPC through read_project_file / write_project_file (senpi#6307), a built session can be parked and handed out through claim_session (senpi#6350), the Fable 5.2 prompt preset gains the codebase-ownership sentence (senpi#6360), generate_image is eval-only like bash (senpi#6272), eval cells reach every eval-only tool through tool.<name>(...) and the agent()/output() helpers (senpi#6276), and the EQ goldens carry the new prompt (senpi#6355). (#13054)
An agent that answered a message from another agent can keep answering it in the turns the engine starts on its own, with no prompt. The reply allowance a consumed delivery grants (a reply to that sender by id or stated name, approved from the delivery alone) now lasts for the work the delivery started: the consuming turn and every later turn the engine starts by itself - an open-todo reminder, a goal continuation, a monitor wake. A settle alone no longer ends it; it ends when a new message starts work after that answer: a person's message, or any delivery, whose own sender the new group then holds (a delivery that joins running work adds its sender). A reply in a covered turn continues the delivery's chain at the next hop, so the chain limit still closes it. (#13042, part of #12996) (#13044)
omo remote send now goes through the same delivery pipeline as every other send, and --steer reads the session once. A person's message to a session on another device enters createDeliveryPipeline with the relay address book's reference for the <device>/<session> address, so the CLI no longer builds its own transport call; the relay transport reads the running turn once and steers it, where the CLI used to read it a second time first. A --steer to a session that is running nothing still sends nothing. The unused second prompt implementation (promptRemote) and its argument check are removed. Which sender kinds each device admits is unchanged. (#13038)
A message to a session on this device that waits on a question now says so, as one to another device does. The send is still queued, and its delivery now carries waiting_on: { kind: "approval", device }. device is this install's account device, and is absent when the install has no device identity. (#13035)
Every message one session sends another now names the device it was sent from, on this device and across devices alike, and a message from another device counts as the cause of the receiving run's next send the same way a local one does. When no device id can be read (no device service, or it does not answer), the message still goes and records why it names no device. (#13034, part of #13004)
A message sent or handed off to a session on another device now reports its delivery_id, effective_mode and endpoint, as one sent within this device does. The answer came from the account's session listing and dropped those fields. It now passes the relay transport's own delivery facts through. The endpoint is null because the sender knows no endpoint on that device. (#13033)
A reply to an agent on another device now reaches it by the id in the message header or by its name, with no prompt on the same account. One resolver now finds a thread across this device's threads and the account's other devices: an exact durable id, then an account alias, then the name ladder judged over every entry (two devices sharing a name are an ambiguous_target naming both devices, never a silent pick). thread_send, thread_handoff, the control tools, thread_create (open or fork) and the omo thread CLI all use it. A reply to the sender of a delivery the run took in is approved from that delivery alone - by the sender's durable id or the name it stated - and must resolve to that sender, else it is refused as reply_target_mismatch without sending. The id shown for a thread on another device is its durable id; the relay listing carries it when the device publishes one. The omo remote verbs still dispatch by the id you type as the device's session key, so pass the id the device lists under until that path moves onto the shared resolver. (#13000)
A send to an agent on another device now reports the same facts as one on this device. thread_send and thread_handoff answer delivery_id and effective_mode for a target on another device too, so a caller reads the delivery's id and the lane it took the same way whichever device holds the thread; message_seq and endpoint stay the facts of a gateway row on this device, and device names another one. The reply allowance (a run answering the agent whose delivery it took in, by that agent's id or stated name, with no prompt) is proved identical for a same-device and a cross-device delivery, and one parity table now runs every send assertion against both legs, so a rule that reads the transport fails the build instead of reaching a user. (#13006, on top of #13000)
A task child no longer starts against a senpi engine that refused its process-wide redaction registry; the start fails with SenpiEngineRestartRequiredError, which says to restart omo. The engine refuses to load next to a registry slot it did not install, and that refusal lasts for the whole process, so a restart is the only remedy. (#12991)
omo now installs the signed device service 0.1.27 instead of 0.1.3. The device service pin had stayed at 0.1.3 across every device-service release since 0.1.3, and an installed omo only replaces a service older than its pin, so no machine received a newer one. (#12981)
The thread_* tools are now eval-only, like bash: when an eval tool is present they are no longer in the model's direct tool list and are called as tool.thread_*() inside an eval cell. Behaviour, permissions and results are unchanged, and without eval they stay ordinary direct tools. (#12907, closes #12903)
A long-running engine keeps reclaiming crashed one-shot memory runs. The memory component's cleanup of abandoned one-shot runs, which runs once per engine process outside session opens, now repeats once a day in an engine that stays up, such as the desktop's or a daemon host's, instead of waiting for a restart. (#12899)
Opening a session no longer re-sweeps the memory root. The memory component's cleanup of abandoned one-shot runs now runs once per engine process, shortly after the first open, instead of inside every session open. A session that ends before that cleanup starts hands it to the next open, so crashed runs are still reclaimed. (#12896)
Fixed
- The senpi barrel load-failure test imported the barrel by a pathname-derived
/D:/...specifier, which failed on Windows and turned dev red (#13068). It now imports the barrel by file URL. (#13068) - Computer use reports a click or typed text that went out as window events with its real input route, and judges whether input landed only for multi-event input such as drags, so a single click whose window then closes is reported delivered. (#13055)
- The actionability check no longer refuses
presson a Linux button that lists only another action such asclick, since the backend presses its first action. (#13055) - Computer use flags text typed into a secure field of a background window, and reports how much text was delivered when the user intervenes mid-send. (#13055)
- On macOS, key priming now bounds and classifies its focus reads, so a slow or failing focus read no longer stalls background typing into a multi-window app. (#13055)
- Computer use keeps holding the machine-wide interaction lock while it hands key focus back after background typing, so another session can no longer type in between. (#13055)
- A steer sent to a session on another device that was held behind that session's open draft, and whose turn ended before the draft was released, was answered
queuedand then lost without a word (#12927). The send now records that refusal: a retry of the same send under the sameidempotency_keyanswersturn_conflict, naming the delivery, instead ofqueued. (#12935)
Changed
- Computer use now lets the user keep typing and clicking while it works by default (
computer.intervention: "continue"); set it to"pause"to have input wait for the user and stop withUserIntervened. (#13055) - While computer use waits for an accessibi...
v5.1.68
Release binaries may be up to 210 MiB. The 5.1.67 publish stopped on Linux x64 because its binary measured 209,889,456 bytes, 174,256 bytes over the 200 MiB per-binary budget (5.1.66's was 209,664,176). The budget moves to 210 MiB, about 5% above that size, and still fails any release that goes past it; the 5 MiB headroom warning is unchanged. (#13052) (#13052)
- 8ed3326 Merge pull request #13052 from code-yeongyu/fix/binary-size-budget-210mib
- 5e5c60b docs(changes): add the 13052 release-note fragment
- 957de2b fix(release): raise the per-binary size budget to 210 MiB
- 42d5746 Merge pull request #13047 from code-yeongyu/release/v5.1.67-source-state
- 3b88a00 Merge pull request #13005 from code-yeongyu/fix/12989-endpoint-cache-at-resolve
- d26f4d7 Merge remote-tracking branch 'origin/dev' into fix/12989-endpoint-cache-at-resolve
- 245869b Merge pull request #12993 from code-yeongyu/fix/12989-credential-lock-torn-record
- 8ce72e7 Merge remote-tracking branch 'origin/dev' into fix/12989-credential-lock-torn-record
- bb340a0 Merge pull request #12815 from code-yeongyu/fix/12801-file-backed-ps-snapshot
- 7a0e5b3 chore(train/omo-2): regenerate bundles at the train top on dev f922d6a (#12944 #12949 #12815)
- 302b797 Merge #12815 (4f50baa) into train/omo-2 (re-cut without #12351)
- 4c32820 Merge #12949 (e6de1f0) into train/omo-2 (re-cut without #12351)
- a463e15 Merge #12944 (e91165d) into train/omo-2 (re-cut without #12351)
- 2215cc2 Merge origin/dev into fix/12989-credential-lock-torn-record
- d985fa7 Merge pull request #12979 from code-yeongyu/fix/12978-credential-lock-account-home
- a4bdc34 changes: fragment for #13005
- 60e17c0 fix(thread): a published-id send reuses only an answered enumeration (#12989)
- 24f897c fix(thread): share one endpoint enumeration and start its cache window when it answers (#12989)
- f04cdf1 Merge origin/dev into fix/12978-credential-lock-account-home
- fa05c27 Merge pull request #12975 from code-yeongyu/feat/12886-approval-wait-outcome
- 1acab30 changes: fragment for #12993
- 5515951 fix(service-auth): reclaim a torn credential lock record instead of holding it forever (#12989)
- 6874054 fix(service-auth): keep test CLI children on their own lock and run only System32 tools on Windows (#12978)
- e91165d Merge origin/dev into fix/914-tolerant-reply-decoders
- a777a72 Merge remote-tracking branch 'origin/dev' (#12953 merged standalone, 04ba4a7) into train/omo-1
- 04ba4a7 Merge pull request #12953 from code-yeongyu/feat/claude-fable-5-2
- a66d3d1 fix(knip): export-check device-service-core test/local-api/event-examples.ts
- 0d993cd Merge #12953 (dc68e0f) into train/omo-1
- 0bb53e3 Merge #12975 (3f95b95) into train/omo-1
- b221a27 Merge #12954 (ffa66d4) into train/omo-1
- 4b35d36 Merge #12966 (06241b8) into train/omo-1
- 05dc774 Merge #12909 (cd21814) into train/omo-1
- d14bda0 Merge #12938 (06ebd08) into train/omo-1
- cac87c0 fix(service-auth): one credential lock per OS user on Linux, macOS and Windows, with a homedir fallback (#12978)
- 06ebd08 Merge remote-tracking branch 'origin/dev' into fix/12853-follow-effective-mode
- 01cfed3 fix(thread): a permission prompt (pendingDialogs) parks the receiver too, and a failed post-admission read keeps the probe's parked state (review MEDIUM + LOW, #12975)
- 5755bb8 Merge remote-tracking branch 'origin/dev' (#12922) into feat/12886-approval-wait-outcome
- 86441b3 Merge remote-tracking branch 'origin/dev' into feat/claude-fable-5-2
- 4085650 chore(changes): fragment for #12979
- cd21814 Merge remote-tracking branch 'origin/dev' into fix/12908-canary-user-intervened
- 0763a35 fix(service-auth): take the credential lock in the OS account's home, not $HOME (#12978)
- 363d0e3 Merge commit '4543c85d49fab896634507a28fb3a47401521554' into fix/914-tolerant-reply-decoders
- d79bc6c Merge pull request #12922 from code-yeongyu/fix/omowright-attribution-290
- 06241b8 Merge remote-tracking branch 'origin/dev' into fix/qa-local-plugin-path
- c0931c9 Merge remote-tracking branch 'origin/dev' (v5.1.66 source state) into feat/12886-approval-wait-outcome
- 4543c85 Merge pull request #12973 from code-yeongyu/release/v5.1.66-source-state
- cab4d52 docs(changes): fragment for #12975
- 088ea31 feat(thread): a send to a peer parked on an approval is queued with waiting_on approval on that device, not idempotency_uncertain (part of #12886)
- e148829 Merge remote-tracking branch 'origin/dev' into fix/12886-peer-reply
- 9e0bce1 Merge remote-tracking branch 'origin/dev' into fix/omowright-attribution-290
- aa3312f Merge remote-tracking branch 'origin/dev' into fix/914-tolerant-reply-decoders
- 8c5cb0e Merge remote-tracking branch 'origin/dev' into fix/12853-follow-effective-mode
- b628bc8 docs(changes): name the #12966 fragment
- c995498 test(thread): type the expected reply request literally (#12954)
- dc85dad fix(thread): a reply keeps a scoped thread_send request with its ask auto-approved, only for an exact session id, and ends when the run settles (#12954 review)
- 615c2a2 Merge remote-tracking branch 'origin/dev' into fix/omowright-attribution-290
- d928b62 fix(memory): the dry run counts a completion whose run the same pass removes (#12949 review r2)
- 3bfc1eb test(memory): RED for the preview missing completions whose run the same pass removes (#12949 review r2)
- c10ef48 Merge remote-tracking branch 'origin/dev' into fix/914-tolerant-reply-decoders
- f99219c fix(memory): count every pruned store in the dry run and bound one log line to 32 KiB (#12949 review)
- 32e9ef3 test(memory): RED for the review findings on #12949 (preview undercount, oversize log line)
- d1fd0d1 docs(changes): name the fragment for #12954
- 622fedb docs(changes): add the Fable 5.2 changelog fragment (#12953)
- 0bee45b fix(thread): a peer-delivered run is told to answer with thread_send, and its reply to the sender asks no permission (#12886)
- 89d4074 feat(model-core): route the current Fable to Claude Fable 5.2 and retire the 5.5 launch gate
- 2a136d1 docs(changes): name the fragment for its PR (#12949)
- d966b3f docs(changes): fragment for the memory cleanup preview and log (#12932)
- 3e79ddb feat(memory): retention dry-run preview, removal log and last cleanup in omo doctor and /doctor (#12932)
- 97fe975 test(memory): RED for retention dry-run preview, removal log and last cleanup in doctor (#12932)
- cb1982f docs(changes): fragment for #12944
- 52c9a28 docs(changes): fragment for #12938
- 4f50baa Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- c2eb4bd fix(device-service-core): ignore unknown fields in replies the client reads
- a7c9fc4 fix(browser-skill): refuse a LICENSE-less omowright before digesting it
- 57fd35e fix(browser-skill): fold review findings
- 0a681b3 docs(changes): name the fragment for its PR
- 8cdaf95 docs(changes): fragment for the browser skill LICENSE and attribution fix
- ebc8ed9 fix(browser-skill): ship omowright's LICENSE beside the bundle and correct ATTRIBUTION.md
- e88915b Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- c1b70eb docs(changes): note the canary UserIntervened fix (#12909)
- 9d7d945 fix(desktop-macos): a user who interrupts the receipt canary gets UserIntervened, not BackgroundUnavailable (#12908)
- 2fc29d3 Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- 0f5c0ec Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- 462e8e4 Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- 0005c45 Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- ca8fde7 Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- fb08108 Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- 8795a27 Merge remote-tracking branch 'origin/dev' into fix/12801-file-backed-ps-snapshot
- 3e40b1b fix(memory-core): write the git-timeout process snapshot to a temp file, not a capped pipe
bun add -g omo-aiv5.1.66
omo thread says when a broken install leaves it on this device only. A thread client sidecar that cannot be loaded no longer makes omo thread read <alias> silently answer not_found: stderr names the sidecar, the error, and the reinstall step. The CLI sign-in's own warnings (a retired trust pin, a failed device-service forget) now reach stderr too. (#12936)
A message to a busy thread now reaches the turn it is running, by default. thread_send and thread_handoff with no delivery mode (and omo thread send without --mode) steer the recipient's running turn and start a turn when it is idle, on this device and on another device alike; a message queues behind the running turn only when the caller asks for follow_up, or when the turn cannot take a steer (a waiting question, a compaction, an engine that reports no turn epoch). A send that steers a turn ending at that very moment falls back to the lane the target has then instead of failing. (#12921, part of #12879)
OmO runs on senpi 2026.11.9-2. A tool that an eval cell awaits no longer leaves the running cell blank for a moment when it finishes: its streamed lines stay until the cell's own output or its result replaces them (senpi#6257). (#12920)
omo remote send now travels the same road as a session's thread_send to another device, and every reply from another device is explained by one mapping. The person at a terminal who sends to a session on another device goes through the relay transport (the same offline check, the same steer preflight against the running turn, a receipt in this device's gateway store) while keeping their own prompt record; what the user reads when a command on another device is refused (omo remote, the attach TUI, the thread tools and the SDK) comes from a single module, so a device that did not answer, a missing grant or an uncertain result reads the same everywhere. (#12911, part of #12879, fixes #12885)
OmO runs on senpi 2026.11.9. It brings senpi's admit_external_message (a message relayed from another device's session is admitted as an ordinary delivery) and makes a relayed delivery read exactly like a same-machine one, with the device kept in its metadata only. The one-channel relay transport (#12879) needs both. (#12904)
The Install button on omo.dev's header no longer presses against or slips past the pill's right edge. The floating nav bar gives its last control the same inset as the logo on the left, and below 1000 px the section links move into the menu so nothing is pushed outside the pill in any language. (#12893)
A thread an agent creates or forks with thread_create now starts in the creator's own approval mode on this device too, not the host's default. A session created on another device already did (#6189); the two legs now run one body: a name is checked against the device the thread lands on (a name that device already lists live is name_conflict there), opening a thread that is live on another device answers that thread without opening it again, and the created thread's facts (name, working directory, timestamps) come from the engine's reply on both devices. (#12891, part of #12879)
Stopping, closing, renaming, re-modelling, changing the reasoning level of, or reading a thread works the same way whether that thread runs on this machine or on another device of your account. thread_set_reasoning now reaches a thread on another device too, and a session with no file yet that another client still holds answers the same on both. (#12890, part of #12879)
A message an agent sends to a session on another device now arrives there exactly as a message between sessions on one device. The receiving session gets a delivery entry (never a user message) with the same sender line and authority notice a same-device delivery carries, deduplicated under the same key, and it wakes on receipt the same way; a reply is just another send back. Nothing in the delivered text names the device it crossed, and omo remote send from a terminal is unchanged. A send from omo thread send to a session on another device is refused as data (the engine admits a command-line delivery only from its own machine) and points at omo remote send. (#12889, part of #12879, fixes #12882, fixes #12877, fixes #12346, fixes #12865, fixes #12857)
Sending a message to another session now goes through one delivery pipeline whether that session is on this machine or on another device. The device difference is a transport module behind one port; nothing a user sees changes in this release. (#12887, part of #12879)
A message to a session on another device now reads exactly like one from a session on the same machine. The receiving agent sees which session sent it, not which device, and replies by that session as usual. (#12878)
A subagent whose model takes more than 90 s to start answering is no longer aborted by runtime fallback. With runtime_fallback enabled, the first-prompt watchdog cut every subagent that showed no progress within a fixed 90 s and re-prompted the fallback model, so reasoning-heavy models that legitimately think longer before their first token were killed and their prompt was sent twice. The watchdog now waits 300 s, the same stream-start bound the engines use, and runtime_fallback.first_prompt_timeout_seconds sets it (0 disables it). (#11893, fixes #11887)
Fixed
omo thread sendto a thread alias held on this device now reaches that thread when run from another workspace, the same wayomo thread readdoes, instead of answering that no visible thread matches. (#12860)- Memory shutdown drain: the session-shutdown wiring no longer claims senpi puts no bound on its handler. senpi bounds one
session_shutdownhandler with its own warn threshold, per-handler timeout and total quit budget (shipped defaults 2 s / 10 s / 2 s, each user-tunable and 0-disableable), and the 1500 ms drain budget is sized under those defaults and pinned by a test that reads them from the installed senpi, so a host release that lowers a default below the drain fails there. No host cap was added from OmO and nothing bypasses the host abort; a journal flush skipped by the budget still raises its error-level alarm. (#12795, refs #10336) - The browser skill's docs and examples teach BrowserSkill 0.4 shared control for the omowright it now ships (
^0.3.0): every gated page action carries the context of a freshsession.cooperate("sync"), and acooperation_sync_requiredrefusal waits for the person withsession.cooperate("wait"), then re-syncs and replans. They also note which calls are gated, that a gated action throws aTypeErroron an unknown option key (with the two calls that pass stray keys through instead), and thatpage.on("dialog")listeners receive aDialogthey can answer. The omowright this release ships includes running-browser discovery throughDevToolsActivePort,gotoon sparse and frameset pages, UTF-8 decoding across BrowserSkill IPC chunks and unique-only stale-ref recovery (thanks @ashmoonori-afk, code-yeongyu/omowright#115 to #118), and agent tabs in their own window (thanks @devxoul, code-yeongyu/omowright#120). (#12494) - The senpi-task cold-revive residency test is now bounded by the product's own liveness rules instead of a flat 20 s total (Part of #10993). A Windows child that blocks for a while during startup no longer fails it unless the product would also give up, and a real stall fails with the last stage it reached and the stall length. The test's child also reports senpi's startup phase timings, so the next block names its phase. (#11706)
Changed
- Skills:
ulw-researchand mass research inmass-ulwnow require every conclusion to rest on something a tool actually observed. Workers tag claimsOBSERVED:and dead endsPROBED:, anything known only from memory is sent back as a lead, and mass-research verify nodes re-open each observed source. Both research editions state the rule in their opening contract, and every mass-ulw node prompt carries it in its VERIFY item. ([#12...
v5.1.65
A message an agent sends to a session on another device now says who sent it. It used to arrive there as plain text, indistinguishable from that session's own user. It now carries the same sender line and authority notice as a message between sessions on one device (the sending session, its workspace and project, and the device it came from), and a leading / in it no longer runs a command or expands a prompt template on the receiving side. omo remote send from a terminal is unchanged. (#12858, part of #12852)
Memory retention no longer evaluates its completion-record guard twice per record. The completion sweep now reads the guard once and keeps or removes the record on that single answer, and the retention tests check their own preconditions instead of asserting them away. No change to what retention keeps or removes. (#12783, follow-up to #12532)
The standalone omo binary's omo doctor reports memory runtime size again, like the npm install. Each memory identity's runtime size against its cap and the total across identities are printed by both distributions, so binary/npm doctor parity holds (#12785). (#12783)
Old binary runtimes are pruned safely when two omo launches start at once. A launch no longer deletes a runtime another launch is still pruning, and a runtime that launch already removed is not reported as a failure. omo --version, --help and omo update never prune. (#12770, follow-up to #12627)
A task child whose engine cannot hold its designated model chain is no longer started. Before, when a process child's engine could not take its chain (an older engine, a refused set_retry_fallback, or no answer in time) or a task host did not advertise retry_fallback_profile, the child was still prompted and ran on the engine's own fallback settings, which can reach models outside its chain. The start now fails before the first prompt and names every designated model and the reason, and such a host is refused as incompatible, naming the same models. (#12759) (#12761)
A command on a closed session on another device now reports the same error as on this device. It answered session_not_found remotely and unknown_session locally; both now answer unknown_session. (#12755) (#12756)
omo hands its device keys to the device service from memory, with no credential store read for relay. A fresh sign-in (omo login, omo account adopt) deletes the signed-out marker and then hands its in-memory keys and account to a running device service. The memory sync session re-adopts from the credential it already holds when the service reports no adopted device. Every sign-out path (logout, forget, a session the service ended) tells the service to forget the device. The start-up bootstrap no longer reads the stored credential. device.adopt carries account only to a device service at 0.1.3 or newer, and the pinned device service is raised to 0.1.3 (design D PR 4). (#12735)
omo remote can open, send, switch model, interrupt and close sessions on your other devices. omo remote open <device> [--project P] [--model M] starts a session there and prints <device>/<session>; omo remote send <device>/<session> <text> prints the answer as it streams in and returns when the turn settles (--steer steers the running turn, --follow-up queues behind it, --no-wait returns once the session took the message); model, interrupt and close do what they say. Each verb is the same remote op the agent's thread tools use, and --json prints its typed result. A steer to a session on another device now reaches it when that device's engine reports its running turn, instead of always being refused. (#12709, part of #12320)
Creating or forking a thread on another device no longer waits for a separate session.locate round trip: the create asks the relay for the new session's id on the open_session reply and uses it. A relay daemon without that answer (an older host, or an older service on this device) is handled as before, with session.locate. (#12686)
omo update on the standalone binary now replaces the installed omo. It used to print a replace command for the copy under ~/.omo/binary-runtime/<version>/, which the next launch put back, so the update never took. A launch also removes old runtime versions there, keeping the previous one for a rollback and any version still running. (#12627, fixes #12626)
Each memory identity's runtime directory is capped at 2 GiB, and omo doctor and /doctor show how much it holds. After retention, the 12-hour memory maintenance pass evicts oldest-first until the identity fits memory.runtime.max_bytes (default 2 GiB, 0 for unbounded, per-agent overridable under memory.agents.<id>.runtime). It removes folded reflection runs first, then the transcripts of fully reflected sessions, idle sidecars, and finally old failed runs beyond the newest ten. Journals that were never or only partly reflected are kept by default; memory.runtime.discard_unreflected_journals: true lets the cap evict them too, and only while reflection is succeeding on the identity. Live sessions, in-flight runs and the run-id high water are never touched. omo doctor prints each identity's runtime size and the total, and warns when an identity reaches 80 % of its cap. /doctor adds a runtime-size check with the largest stores and what retention would still free. (#12538)
Memory no longer grows without bound on disk. The background memory maintenance pass now removes finished reflection runs, fully reflected transcripts, recall sidecars and legacy stores past configurable windows (memory.runtime.retention_days 7, completions_days 30), in time-boxed slices that resume where they stopped. Transcript input reflection never folded into memory is kept unless you turn on discard_unreflected_journals (then removed after abandoned_journal_days, 14, only while reflection works for that agent). Nothing a live session owns, no locked or unfinished work and no run id reflection still needs is touched. (#12532)
A reflection lock left behind by a run whose processes died (launcher killed before the run started, supervisor and child gone with no outcome, an unreadable run ledger, or a torn active.lock) is now released at session start, on the next turn, and by the new omo doctor --fix (--dry-run to preview, /doctor --fix in a session), instead of keeping every later reflection queued; omo doctor --fix also reclaims runtime locks whose owner is gone. (#12517)
Fixed
- Senpi: omo now names any senpi feature it relies on that the running senpi lacks in one startup warning, and opening a monitor on another device's session on a senpi without remote wake fails with
remote_wake_unavailableinstead of starting a watch that never wakes. A contract test checks every such feature against the pinned senpi. (#12862, fixes #12861) - The browser skill now depends on the published
omowrightpackage (^0.3.0, the first release that ships a builtdist/that loads under Node as well as Bun) instead of a git commit. Its runtime stager and loader accept omowright's TypeScript sources, the older JavaScript sources, or a compileddist/, in that order, so a future layout change cannot silently break the staged browser runtime (#12842, #12851, #12847). (#12847) - Memory facts extraction records facts again with senpi 2026.11.6-3: the facts child (also used by onboarding's understand step) keeps the approval mode its own settings give it and may run only its own record tools without asking, instead of being refused as a child whose parent is gone. (#12839)
- Concurrent memory account resolves on a desktop host share one open, so the app sees one
descriptorrequest instead of one per caller; a failed open is retried by the next resolve (carried from #12717, which this PR supersedes). (#12839) - Thread tools: an agent can now reach a session on another device of the same account by its name. A name that matches the title of one live session on another device sends there, and a name shared across devices (or with a thread on this device) is refused as ambiguous with every candidate and its device listed, never picked silently. Claiming or renaming a thread alias on a signed-in CLI device now reaches the account's alias registry instead of always an...