oh-my-openagent

Desktop memory sync now starts as soon as the engine boots, without waiting for a session. An engine started by omo host ensure loads extensions only when a session opens, while the desktop opens none until the engine has sent the memory-sync bootstrap handshake, so the two waited on each other. The compiled omo and the npm launcher now start that one claim before the engine loads whenever an RPC engine's launch declares a memory-sync channel; the first session reuses it. A claim that fails still reports memory sync as unavailable and never falls back to CLI credentials. (#12702) (#12704)

Internal (tests only): macOS trust acquisition reads /etc/ssl/cert.pem through the already-injectable bundle reader, and a test now proves that a malformed default file yields the structured openssl-default-file-malformed drop entry (file, cause, block index, narrowing). (#12633)

Internal (tests only): the OpenSSL default-file importer tests now pin that any high byte on a BEGIN line, not only a BOM, drops the whole file. (#12632)

Internal (release tooling): the omo-native trust mutation harness counts a mutant as killed only when a named test fails by an assertion; a mutant that only throws a runtime error or crashes on load is now reported as crashed and fails the run. (#12631)

Fixed

• Memory sync no longer loses an edit when two devices change the same line of a memory file at the same time. Devices on this version keep both edits as whole lines, and every device converges. Before, the edits were merged character by character into text that matched neither (#12279).
  • Between devices on this version, an edit does not disturb a line nobody edited. The exception is a very large rewrite: when one device adds and removes more than about 2,000 lines of a note in one save, that save is merged character by character, as before this fix, so a concurrent edit to the same line can still splice.
  • When the note's body or the new text does not end with a newline, the edit is still merged character by character, as before, so two concurrent edits to that last line can still splice.
  • With a device that still runs an older version, its old character merge can still go wrong in these ways:
    • Adjacent lines can fuse. The older device rewrites the line right after a line this version edited. Depending on the two devices' internal ids, the two edits can fuse into one line followed by an empty line. Measured: from head, one, two, tail, this version writes ONE and the older device writes TWO, and the result can be head, TWOONE, an empty line, tail. Two older devices give head, ONE, TWO, tail.
    • Leftover text moves. The older device edits a line this version deletes. Its leftover text joins a neighbouring line.
    • Its own diff touches other lines. The older device's character merge crosses into a line its user did not edit. This also happens between two older devices.
  • Two older devices keep their old character merge until they update. (#12597)
• Memory sync: the push phase of a large memory set no longer needs memory in proportion to the whole set. Every push step used to re-read and re-write one journal file holding every pending item's data, several times per batch. A device pushing a few thousand memory files therefore peaked at 2 GiB or more (#12286). The journal now keeps a small index of pending items with their data in separate files, and a push reads only the batch it is sending. Capturing the set before the push still holds each captured file's text until the capture ends. A journal written by an older

The macOS keychain helper's lookup paths and tests are correct on Windows CI. The macOS install locations (the OmO.app runtime and helper copies) are built with POSIX separators on every host, and the #11262 keychain-helper tests use platform-correct paths, so dev's Windows test legs pass again. (#12690, part of #11262)

Fixed: the first device listing no longer waits about 10 seconds for each signed-out or otherwise unsupported device; the bunshin SDK answers such a device's tool list as empty at once instead of trying to describe it. (#12680)

Internal: the vendored bunshin SDK imports its own pinned typebox through omo-senpi's #bunshin-ts-dep/ imports instead of the bare package name, so senpi's extension registry can no longer hand it a different typebox; no user-visible change. (#12676) (#12680)

Thread ops on another device spend fewer relay round trips: the session's attach is kept across ops, and the session's model catalog is held on it, so a model switch to an exact provider/id that catalog lists is one set_model. A set that may have run is still reported as idempotency_uncertain; a refused one re-reads the catalog and answers as a fresh catalog would. Status and transcript reads (get_state, get_messages) still go to the device every time, because the engine's in-memory state changes without a record on the attach. (#12672)

Starting omo no longer reinstalls and restarts a running device service that is only slow to answer: on macOS and Linux a handshake timeout is retried once and then reported as degraded, instead of being treated as a stopped service. (#12665)

A message sent to a thread on another device whose engine answers started without a turn id (senpi 2026.11.2) is now confirmed with the running turn that get_state reports, instead of idempotency_uncertain. (#12660)

Fixed: closing a thread on another device answers closed instead of uncertain (reply_lost); the close was applied there, but its answer was dropped when the other device ended the session's attachment. (#12655)

Fixed: a thread created on another device (thread_create with a device, or a fork there) is named by the id that device lists it by, so reading, messaging, interrupting, switching the model of and closing it work instead of answering session_not_found. (#12651)

Internal: knip ignores the closed-device-login preload fixture (a --preload test file, nothing imports it); no behavior change. (#12648)

A delegated child whose own model chain has no connected provider now fails with one message that names every model in that chain and the exact fix, instead of an opaque "no available model". The chain is listed in order with the providers each model needs, the refusal says which providers are not connected, and it names the key the app's Settings editor writes (categories.<name>.model for a category, agents.<name>.model for an agent), so the parent session can act on the first line instead of retrying a dead binding. The same refusal now covers a lane pinned to a model this machine's registry does not serve, naming the pin and the Settings key. Nothing is substituted and no child is spawned (#12634). (#12636)

Internal (tests only): in the omo-native trust-stage TLS probe, a channe

What changed

omo verifies the device service it downloads against the real release key set. The device service's release key set is now pinned to the minted Ed25519 keys, operational first and recovery second, in packages/device-service-core/src/release/release-key.json (a byte copy of the service's own release-key.json). A release signed by either key installs; one signed by neither, or with a malformed signature, is refused before anything is written. Until now the pin was a placeholder that refused every release. (#12437)

Skills and prompts no longer point at skills, files or agents that do not exist, and CI now refuses any new one. The Native ultrawork directive told sessions to read the bun-1-4 skill, which Senpi renamed to bun; it now names bun. remove-deadcode asked for a typescript-programmer skill, which made OpenCode refuse the whole task call; it now loads programming. frontend and ulw-research no longer load an open-design skill that omo never shipped and instead name the nexu-io/open-design library. The ulw-research assembly worker loads imagegen and ulw-loop only when the session lists them, and frontend loads the browser skill where it asked for OpenCode's playwright. OpenCode now ships the hyperplan skill that its hyperplan mode tells the session to load; until now only this repository's own checkout had it. Seven programming references and one frontend cross-reference that pointed at files in the wrong directory, or at a file that never existed, now open the right file. A new check in the root test suite resolves every skill name, skill-relative path, agent and category in shipped skills and prompts for each edition that ships them, and names the closest real target when one is missing. (#12484)

When omo drops the OpenSSL default certificate file (/etc/ssl/cert.pem on macOS) from the engine's trust store, it now tells you why. omo doctor prints an INFO line naming the file, the cause, the certificate block where reading stopped, and whether the drop trusts fewer certificates than plain Bun would. omo trust ensure --json reports the same facts as defaultFileDrop. It is never a warning or a failure. (#12499, part of #10762)

2026-11-05 - Name the OpenSSL default-file drop (#12499, part of #10762)

Design revision 11, 4.4 item 6. readDefaultFileCandidates takes its narrowing flags from the exported DROP_CAUSE_NARROWING table. On a drop, acquireEngineTrust records one openssl-default-file-malformed entry carrying file, cause, blockIndex and narrowing instead of two plain codes, and returns it as defaultFileDrop. runTrustStage prints it in the trust ensure --json line. trust-diagnostics.ts accepts that structured entry and still accepts the plain code for an unreadable file; it refuses another file name, an unknown cause, a non-integer block index or a narrowing flag that does not match the cause. engine-trust-doctor.ts reads the adopted generation's diagnostics.1.jsonl and diagnostics.jsonl, takes the last import-exec or ensure record of that generation, and gives omo doctor one INFO line: the drop, or default-file: kept / unreadable. A missing or torn log yields no line and never fails doctor.

Native skills no longer point Senpi sessions at OpenCode-only agents, tools and commands. A community audit of the installed skills found 16 more references that exist only in OpenCode or nowhere: refactor consulted an oracle agent and called team_list and team_status, debugging ended a team with team_shutdown_request and team_approve_shutdown, frontend loaded skills through a skill tool and address

Runs on senpi 2026.11.3-3: Claude Opus and Fable get their full system prompt back (the pre-diet wording; every other model keeps the slimmer prompt), Anthropic Subscription no longer refuses a working Claude login as blocked when an older omo process is still running or when Claude Code can't read its login behind a locked keychain, a usage limit with an ISO reset time blocks that account until the reset instead of retrying every minute, a goal waits for the first limited account to come back instead of spending turns, and Shift+Enter inserts a newline in macOS Terminal.app. (#12311)

Runs on senpi 2026.11.3-3: the engine socket can now send compact message updates (compact_message_updates_v1) with event_resync recovery, so long streamed turns cost less to relay. It stays off unless SENPI_ENGINE_SOCKET_COMPACT_UPDATES=1 is set. (Part of code-yeongyu/senpi#4272) (#12215)

macOS computer use keeps the display awake while it operates, and says when the screen is locked or asleep. The engine now holds a PreventUserIdleDisplaySleep power assertion ("OmO computer use is operating the desktop" in pmset -g assertions) from the first request other than capabilities, control.state and stopPath.* until 60 seconds after the last one; session.close, a disconnect and a stop release it at once. capabilities.displayAsleep reports when every display of the session sleeps, omo doctor prints it on the display line, and a CaptureFailed, InputFailed or Timeout message ends with (the screen is locked) or (the display is asleep) when that is the likely cause. Other platforms are unchanged. (#11530, part of #11466)

Fixed

• In the desktop's first-run setup, "Understanding you" no longer stops with "No retained scan with this id" after About you brings anything in. Applying an onboarding scan now keeps it available to understanding until the scan expires; a scan is still applied only once. (#12509, fixes #12508)
• The engine now registers its agent dir with the device service even when the service's state directory appears right after the engine started (for example, when the device service is installed or started at the same moment). A freshly armed directory watch can miss an entry created in its first milliseconds, so the registrant never noticed the service and other devices listed no sessions. Each watch arm now re-checks the path once after it settles. (#12482)
• Your other signed-in devices now see this device's sessions after the OmO device service starts or restarts. Every omo engine, in the terminal and the ones the desktop app starts, keeps its own agent directory registered with the device service and registers it again when the service comes back or the device is signed in again; it never holds an engine open and stays off with "device_service": { "enabled": false }. (#12380, fixes #12379)
• A reflection sandbox probe that timed out once (for example on a busy machine) no longer turns the sandbox off until restart. Reflection checks again on its next run, so it runs sandboxed again once bwrap responds. A machine where bwrap definitively cannot create a sandbox is still checked only once (#12376). A reflection run that falls back to running unsandboxed because bwrap is unusable at that moment is now logged once as a degraded sandbox, and once more when the sandbox comes back, instead of pa