omo-ai 5.1.78 → 5.1.81

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (72) hide show
  1. package/bin/darwin/omo-keychain +0 -0
  2. package/bin/lib/doctor.js +6 -17
  3. package/bin/lib/engine-commands.js +2 -0
  4. package/bin/lib/engine-environment.js +3 -0
  5. package/bin/lib/path-omo.js +53 -0
  6. package/bin/lib/provider-map.json +1 -1
  7. package/bin/lib/reflection-lock.generated.js +7 -7
  8. package/bin/lib/self-update.js +18 -2
  9. package/bin/lib/service-auth.generated.js +45 -43
  10. package/bin/lib/thread-output.js +9 -1
  11. package/bin/lib/thread.js +23 -3
  12. package/bin/lib/versioned-install.js +173 -0
  13. package/bin/lib/versioned-leftovers.js +58 -0
  14. package/package.json +2 -2
  15. package/plugin/CHANGELOG.md +140 -0
  16. package/plugin/README.md +2 -1
  17. package/plugin/extensions/gateway-store-worker.mjs +2 -2
  18. package/plugin/extensions/memory-run-supervisor.mjs +1 -1
  19. package/plugin/extensions/omo-claim-receipt.js +2 -2
  20. package/plugin/extensions/omo-computer-run-worker.mjs +1 -1
  21. package/plugin/extensions/omo-computer-use.js +1 -1
  22. package/plugin/extensions/omo-desktop-toolkit.js +1 -1
  23. package/plugin/extensions/omo-device-account.js +7 -7
  24. package/plugin/extensions/omo-device-service.js +6 -6
  25. package/plugin/extensions/omo-feedback.js +2 -2
  26. package/plugin/extensions/omo-init-deep-advisor.js +2 -2
  27. package/plugin/extensions/omo-member.js +2 -2
  28. package/plugin/extensions/omo-memory-commands.js +2 -2
  29. package/plugin/extensions/omo-memory-doctor.js +3 -3
  30. package/plugin/extensions/omo-memory-memfs.js +1 -1
  31. package/plugin/extensions/omo-memory-palace.js +2 -2
  32. package/plugin/extensions/omo-memory-sync.js +1 -1
  33. package/plugin/extensions/omo-onboarding-scan.js +2 -2
  34. package/plugin/extensions/omo-recall-store.js +1 -1
  35. package/plugin/extensions/omo-session-history.js +1 -1
  36. package/plugin/extensions/omo-side-panel.js +2 -2
  37. package/plugin/extensions/omo-task.js +2 -2
  38. package/plugin/extensions/omo-telemetry-otlp.js +2 -2
  39. package/plugin/extensions/omo-telemetry-posthog.js +1 -1
  40. package/plugin/extensions/omo-telemetry-recovery.js +1 -1
  41. package/plugin/extensions/omo-thread-client.js +3 -3
  42. package/plugin/extensions/omo-wake.js +6 -6
  43. package/plugin/extensions/omo.js +3 -3
  44. package/plugin/package.json +3 -2
  45. package/plugin/runtime/agent-toolkit-sdk/sdk.js +2 -2
  46. package/plugin/runtime/ast-grep-mcp/manifest.json +1 -1
  47. package/plugin/runtime/bunshin-ts/sdk.js +2 -2
  48. package/plugin/runtime/category-coverage/index.js +4 -4
  49. package/plugin/runtime/omo-sdk/sdk.js +65 -0
  50. package/plugin/runtime/rollback-migrate.js +1 -1
  51. package/plugin/runtime/task-config/index.js +1 -1
  52. package/plugin/runtime/thread-sdk/sdk.js +3 -3
  53. package/plugin/scripts/install.mjs +3 -1
  54. package/plugin/skills/browser/runtime/omowright/manifest.json +1 -1
  55. package/plugin/skills/omo-relay/SKILL.md +38 -0
  56. package/plugin/skills/omo-relay/references/errors.md +16 -0
  57. package/plugin/skills/omo-relay/references/machines.md +53 -0
  58. package/plugin/skills/omo-relay/references/parallel.md +51 -0
  59. package/plugin/skills/omo-relay/scripts/sdk/bind-eval.ts +153 -0
  60. package/plugin/skills/omo-relay/scripts/sdk/index.ts +5 -0
  61. package/plugin/skills/omo-relay/scripts/sdk/machines.ts +123 -0
  62. package/plugin/skills/omo-relay/scripts/sdk/outcome.ts +77 -0
  63. package/plugin/skills/omo-relay/scripts/sdk/ports.ts +78 -0
  64. package/plugin/skills/omo-relay/scripts/sdk/threads.ts +89 -0
  65. package/plugin/skills/omo-relay/scripts/sdk/tsconfig.json +19 -0
  66. package/plugin/skills/programming/SKILL.md +1 -0
  67. package/plugin/skills/ultrawork/SKILL.md +0 -22
  68. package/plugin/skills/ultrawork/references/gpt.md +67 -0
  69. package/plugin/skills/ulw-execute/SKILL.md +15 -5
  70. package/plugin/telemetry-schema.json +2 -1
  71. package/plugin/telemetry-schema.omo-cloud.json +2 -1
  72. package/plugin/skills/ultrawork/references/astra-variant.md +0 -66
@@ -41,7 +41,15 @@ const RENDER = {
41
41
  ],
42
42
  read: (result) => [
43
43
  ...result.items.map((item) => `[${item.role}] ${item.content}`),
44
- ...(result.truncated ? [`(truncated; continue with --cursor ${result.next_cursor})`] : []),
44
+ ...(result.truncated ? [`(truncated; older messages: --cursor ${result.next_cursor})`] : []),
45
+ ],
46
+ close: (result) => [
47
+ result.closed
48
+ ? `closed ${result.thread}${result.graceful === undefined ? "" : ` after turn ${result.graceful.turn} (${result.graceful.waited_ms} ms)`}`
49
+ : `${result.thread} was not running`,
50
+ ...(result.dropped_deliveries.length === 0
51
+ ? []
52
+ : [`returned to their senders: ${result.dropped_deliveries.join(", ")}`]),
45
53
  ],
46
54
  bind: (result) => [`bound ${bindingLine(result.binding)}${result.deduplicated ? " (already bound)" : ""}`],
47
55
  unbind: (result) => [`unbound ${bindingLine(result.binding)}${result.already_closed ? " (already closed)" : ""}`],
package/bin/lib/thread.js CHANGED
@@ -16,13 +16,14 @@ export { THREAD_EXIT } from "./thread-args.js"
16
16
  */
17
17
 
18
18
  const USAGE = [
19
- "usage: omo thread <list|send|read|bind|unbind|rebind|bindings|report|answer|outbox|ack> [options] [--json]",
19
+ "usage: omo thread <list|send|read|close|bind|unbind|rebind|bindings|report|answer|outbox|ack> [options] [--json]",
20
20
  "",
21
21
  " list [--all-scope] [--state running|idle|inactive|archived[,...]] (repeatable; default: all but archived)",
22
- " send <target> <text> [--mode auto|steer|follow_up] [--expected-turn <n>] [--idempotency-key <k>]",
22
+ " send <target> <text> [--mode auto|steer|follow_up] [--expected-turn <n>] [--idempotency-key <k>] [--collapse-key <k>]",
23
23
  " send --binding <id> [<target>] <text> [--idempotency-key <event-id>] [--mode auto|follow_up]",
24
24
  " [--author-id <platform-user-id> --author-name <display> [--author-user-id <id>]]",
25
25
  " read <target> [--limit <items>] [--max-bytes <n>] [--cursor <c>]",
26
+ " close <target> [--force | --graceful [--wait-ms <n>]] [--idempotency-key <k>]",
26
27
  " bind <session> --platform <p> --account <id> --chat <id> [--thread <id>] [--root-message <id>]",
27
28
  " [--progress-message <id>] [--direction in|out|both] [--inbound-mode auto|follow_up]",
28
29
  " [--events <kind,...>] [--policy <id>] [--ttl <seconds>|none] [--idempotency-key <k>]",
@@ -35,6 +36,9 @@ const USAGE = [
35
36
  " outbox <binding-id> [--after <cursor>] [--limit <n>] [--ack]",
36
37
  " ack <binding-id> <cursor> [--provider-message-id <id>]",
37
38
  "",
39
+ " <target>/<session> takes a thread id, session id, alias, exact name, unique id prefix,",
40
+ " or <device>/<thread> (device: id, here, name or unique device id prefix)",
41
+ "",
38
42
  " --all-scope resolve sessions in every workspace, not only this directory's",
39
43
  " --json print the result as one JSON value",
40
44
  ].join("\n")
@@ -54,7 +58,7 @@ const COMMANDS = {
54
58
  },
55
59
  },
56
60
  send: {
57
- values: ["--mode", "--expected-turn", "--binding", "--idempotency-key", ...AUTHOR_FLAGS],
61
+ values: ["--mode", "--expected-turn", "--binding", "--idempotency-key", "--collapse-key", ...AUTHOR_FLAGS],
58
62
  booleans: [],
59
63
  arity: [1, 2],
60
64
  call: (sdk, p, o, scope) =>
@@ -65,6 +69,7 @@ const COMMANDS = {
65
69
  expected_turn_id: integerOption(o, "--expected-turn"),
66
70
  binding_id: o["--binding"],
67
71
  idempotency_key: o["--idempotency-key"],
72
+ collapse_key: o["--collapse-key"],
68
73
  author: authorInput(o),
69
74
  }),
70
75
  },
@@ -78,6 +83,21 @@ const COMMANDS = {
78
83
  integerOption(o, "--limit"),
79
84
  ),
80
85
  },
86
+ close: {
87
+ values: ["--wait-ms", "--idempotency-key"],
88
+ booleans: ["--force", "--graceful"],
89
+ arity: [1, 1],
90
+ // omo#14313: the thread_close tool's body; --graceful lets the turn end and hands unread messages back.
91
+ call: (sdk, p, o, scope, flags) =>
92
+ sdk.close({
93
+ ...scope,
94
+ thread: p[0],
95
+ ...(flags.has("--force") ? { force: true } : {}),
96
+ ...(flags.has("--graceful") ? { graceful: true } : {}),
97
+ wait_ms: integerOption(o, "--wait-ms"),
98
+ idempotency_key: o["--idempotency-key"],
99
+ }),
100
+ },
81
101
  bind: {
82
102
  values: [
83
103
  "--platform",
@@ -0,0 +1,173 @@
1
+ import {
2
+ existsSync,
3
+ lstatSync,
4
+ mkdirSync,
5
+ readFileSync,
6
+ readlinkSync,
7
+ realpathSync,
8
+ renameSync,
9
+ rmSync,
10
+ symlinkSync,
11
+ writeFileSync,
12
+ } from "node:fs"
13
+ import { homedir } from "node:os"
14
+ import { basename, dirname, isAbsolute, join } from "node:path"
15
+ import { findBunBinary } from "./bun-runtime.js"
16
+ import { findOmoOnPath, readLauncherVersion, realpathOrSelf, reportsVersion } from "./path-omo.js"
17
+ import { liveForeignStages, pruneLeftovers } from "./versioned-leftovers.js"
18
+
19
+ // A side-by-side install (#12022): `<versions>/<version>/` holds a package.json pinning omo-ai, the bun
20
+ // install of it, and optionally an `omo` wrapper naming that dir; `<versions>/current` links the live
21
+ // one and PATH reaches omo through `current`. A version dir is live only once it has `.complete`.
22
+ const CURRENT = "current"
23
+ const WRAPPER = "omo"
24
+ const COMPLETE = ".complete"
25
+
26
+ /**
27
+ * @typedef {{ versionsDir: string, versionDir: string, current: string, manifest: Record<string, unknown> }} VersionedInstall
28
+ * @param {string} root the running omo-ai package root
29
+ * @param {NodeJS.Platform} platform
30
+ * @returns {VersionedInstall | undefined}
31
+ */
32
+ export function findVersionedInstall(root, platform) {
33
+ if (platform === "win32") return undefined
34
+ try {
35
+ const packageDir = realpathSync(root)
36
+ if (basename(packageDir) !== "omo-ai" || basename(dirname(packageDir)) !== "node_modules") return undefined
37
+ const versionDir = dirname(dirname(packageDir))
38
+ const versionsDir = dirname(versionDir)
39
+ const current = join(versionsDir, CURRENT)
40
+ if (!lstatSync(current).isSymbolicLink() || dirname(realpathSync(current)) !== versionsDir) return undefined
41
+ const manifest = JSON.parse(readFileSync(join(versionDir, "package.json"), "utf8"))
42
+ if (typeof manifest?.dependencies?.["omo-ai"] !== "string") return undefined
43
+ return { versionsDir, versionDir, current, manifest }
44
+ } catch {
45
+ return undefined
46
+ }
47
+ }
48
+
49
+ /** @param {VersionedInstall} install @param {string} target */
50
+ export function versionedUpdateTarget(install, target) {
51
+ return {
52
+ manager: "bun",
53
+ command: `bun install in ${join(install.versionsDir, target)}, then ${install.current} -> ${target}`,
54
+ argv: [],
55
+ }
56
+ }
57
+
58
+ function launcherIn(dir) {
59
+ return existsSync(join(dir, WRAPPER)) ? WRAPPER : join("node_modules", ".bin", "omo")
60
+ }
61
+
62
+ function setAside(versionsDir, dir, target, now) {
63
+ if (existsSync(dir)) renameSync(dir, join(versionsDir, `.failed-${target}-${now()}`))
64
+ }
65
+
66
+ function wrapperFor(dir, bun) {
67
+ const entry = join(dir, "node_modules", "omo-ai", "bin", "omo.js")
68
+ return [
69
+ "#!/bin/sh",
70
+ "# omo versioned launcher: runs this version's own entry so live sessions never share files with a newer one",
71
+ `if [ "$OMO_RUNTIME" != node ] && [ -x '${bun}' ]; then exec '${bun}' '${entry}' "$@"; fi`,
72
+ `exec '${entry}' "$@"`,
73
+ "",
74
+ ].join("\n")
75
+ }
76
+
77
+ async function installInto(input, finalDir, bun) {
78
+ const { install, target, run, env } = input
79
+ const stage = join(install.versionsDir, `.stage-${target}-${process.pid}`)
80
+ rmSync(stage, { recursive: true, force: true })
81
+ mkdirSync(stage)
82
+ const dependencies = { ...install.manifest.dependencies, "omo-ai": target }
83
+ writeFileSync(join(stage, "package.json"), `${JSON.stringify({ ...install.manifest, dependencies }, null, 2)}\n`)
84
+ let result
85
+ try {
86
+ result = await run(bun, ["install"], { cwd: stage, stdio: "inherit", windowsHide: true, env })
87
+ } catch {
88
+ result = { status: 1, signal: null }
89
+ }
90
+ if (result.signal || (result.status ?? 1) !== 0) {
91
+ rmSync(stage, { recursive: true, force: true })
92
+ return result.status || 1
93
+ }
94
+ if (existsSync(join(install.versionDir, WRAPPER))) {
95
+ writeFileSync(join(stage, WRAPPER), wrapperFor(finalDir, bun), { mode: 0o755 })
96
+ }
97
+ renameSync(stage, finalDir)
98
+ return 0
99
+ }
100
+
101
+ /**
102
+ * Installs omo-ai@<target> beside the running version the way the layout was built, proves the new dir
103
+ * runs, then swaps `current` with one rename. `current` never names a dir that failed either step, and
104
+ * the previous version dir stays for rollback. Exit 0 only when the omo on PATH reports the target.
105
+ * @param {{ install: VersionedInstall, before: string, target: string, run: Function, env: NodeJS.ProcessEnv,
106
+ * platform: NodeJS.Platform, log: (line: string) => void, error: (line: string) => void, now?: () => number }} input
107
+ * @returns {Promise<number>}
108
+ */
109
+ export async function updateVersionedInstall(input) {
110
+ try {
111
+ return await updateInto(input)
112
+ } finally {
113
+ pruneLeftovers(input.install.versionsDir)
114
+ }
115
+ }
116
+
117
+ async function updateInto(input) {
118
+ const { install, before, target, env, platform, log, error } = input
119
+ const now = input.now ?? Date.now
120
+ const { versionsDir, current } = install
121
+ const finalDir = join(versionsDir, target)
122
+ const unchanged = `${current} still points at ${before}`
123
+ const bun = findBunBinary({ env, platform, homedir: () => env.HOME ?? homedir() })
124
+ if (bun === undefined) {
125
+ error(`omo: ${versionsDir} is a bun-managed install, but no bun was found; ${unchanged}`)
126
+ return 1
127
+ }
128
+
129
+ const busy = liveForeignStages(versionsDir, target)
130
+ if (busy.length > 0) {
131
+ error(`omo: another installer is updating ${finalDir} (${busy[0]}); retry when it finishes, ${unchanged}`)
132
+ return 1
133
+ }
134
+ // A target dir another writer finished (with or without `.complete`) is adopted when it runs the target.
135
+ const runsTarget = async () =>
136
+ reportsVersion(await readLauncherVersion(join(finalDir, launcherIn(finalDir)), env), target)
137
+ if (!existsSync(finalDir) || !(await runsTarget())) {
138
+ setAside(versionsDir, finalDir, target, now)
139
+ const status = await installInto(input, finalDir, bun)
140
+ if (status !== 0) {
141
+ error(`omo: bun install of omo-ai@${target} into ${finalDir} failed; ${unchanged}`)
142
+ return status
143
+ }
144
+ const fresh = await readLauncherVersion(join(finalDir, launcherIn(finalDir)), env)
145
+ if (!reportsVersion(fresh, target)) {
146
+ setAside(versionsDir, finalDir, target, now)
147
+ error(`omo: the omo installed in ${finalDir} reports "${fresh}", not ${target}; set it aside, ${unchanged}`)
148
+ return 1
149
+ }
150
+ }
151
+ if (!existsSync(join(finalDir, COMPLETE)))
152
+ writeFileSync(join(finalDir, COMPLETE), `${new Date(now()).toISOString()}\n`)
153
+
154
+ const next = `${current}.tmp-${process.pid}`
155
+ rmSync(next, { force: true })
156
+ symlinkSync(isAbsolute(readlinkSync(current)) ? finalDir : target, next)
157
+ renameSync(next, current)
158
+
159
+ const updated = join(current, launcherIn(finalDir))
160
+ const onPath = findOmoOnPath(env)
161
+ const seen = await readLauncherVersion(onPath ?? updated, env)
162
+ if (reportsVersion(seen, target)) {
163
+ log(`omo ${before} -> ${target} (${finalDir}; ${install.versionDir} is kept for rollback)`)
164
+ return 0
165
+ }
166
+ error(`omo: updated ${versionsDir}: ${current} -> ${target} (${install.versionDir} is kept for rollback)`)
167
+ error(
168
+ onPath === undefined
169
+ ? `omo: but ${updated} reports "${seen}"`
170
+ : `omo: but PATH still resolves the old omo: ${onPath} (${realpathOrSelf(onPath)}) reports "${seen}"; the updated omo is ${updated}: put it first on PATH or remove ${onPath}`,
171
+ )
172
+ return 1
173
+ }
@@ -0,0 +1,58 @@
1
+ import { readdirSync, rmSync } from "node:fs"
2
+ import { join } from "node:path"
3
+
4
+ // Every writer of a versions dir names its work `.stage-<version>-<pid>` and a rejected version
5
+ // `.failed-<version>-<epoch ms>`, so the pid says whether a stage is still being written.
6
+ const KEEP_FAILED = 2
7
+ const TRAILING_NUMBER = /-(\d+)$/
8
+
9
+ function trailingNumber(name) {
10
+ const match = TRAILING_NUMBER.exec(name)
11
+ return match === null ? undefined : Number(match[1])
12
+ }
13
+
14
+ function pidAlive(pid) {
15
+ try {
16
+ process.kill(pid, 0)
17
+ return true
18
+ } catch (error) {
19
+ return error instanceof Error && "code" in error && error.code === "EPERM"
20
+ }
21
+ }
22
+
23
+ /**
24
+ * Stage dirs another live process is filling for exactly this version.
25
+ * @param {string} versionsDir
26
+ * @param {string} target
27
+ * @returns {string[]}
28
+ */
29
+ export function liveForeignStages(versionsDir, target) {
30
+ const prefix = `.stage-${target}-`
31
+ return readdirSync(versionsDir)
32
+ .filter((name) => name.startsWith(prefix) && /^\d+$/.test(name.slice(prefix.length)))
33
+ .filter((name) => {
34
+ const pid = Number(name.slice(prefix.length))
35
+ return pid !== process.pid && pidAlive(pid)
36
+ })
37
+ .map((name) => join(versionsDir, name))
38
+ }
39
+
40
+ /**
41
+ * Keeps the newest {@link KEEP_FAILED} `.failed-*` dirs and removes stages whose writer is gone; a live
42
+ * stage (this process's included) is never touched.
43
+ * @param {string} versionsDir
44
+ */
45
+ export function pruneLeftovers(versionsDir) {
46
+ const names = readdirSync(versionsDir)
47
+ const failed = names
48
+ .filter((name) => name.startsWith(".failed-"))
49
+ .sort((a, b) => (trailingNumber(a) ?? 0) - (trailingNumber(b) ?? 0))
50
+ const deadStages = names.filter((name) => {
51
+ if (!name.startsWith(".stage-")) return false
52
+ const pid = trailingNumber(name)
53
+ return pid !== undefined && pid !== process.pid && !pidAlive(pid)
54
+ })
55
+ for (const name of [...failed.slice(0, -KEEP_FAILED), ...deadStages]) {
56
+ rmSync(join(versionsDir, name), { recursive: true, force: true })
57
+ }
58
+ }
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "omo-ai",
3
- "version": "5.1.78",
3
+ "version": "5.1.81",
4
4
  "type": "module",
5
5
  "description": "OmO Native - the standalone omo command with the OMO extension built in: bun add -g omo-ai",
6
6
  "bin": {
@@ -17,7 +17,7 @@
17
17
  "postinstall": "node bin/senpi-patch.mjs"
18
18
  },
19
19
  "dependencies": {
20
- "@code-yeongyu/senpi": "2026.11.18-2"
20
+ "@code-yeongyu/senpi": "2026.11.19-5"
21
21
  },
22
22
  "repository": {
23
23
  "type": "git",
@@ -7,6 +7,146 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
7
7
 
8
8
  ## [Unreleased]
9
9
 
10
+ ## [5.1.81] - 2026-11-21
11
+
12
+ ### Changed
13
+
14
+ - The `programming` skill's language gate now routes rendered web work (`.html`, `.css`, `.jsx`, `.tsx`, `.vue`, `.svelte`, `.astro`, or any task that ships a UI) to the `frontend` skill before the first line of UI code. It tells the agent to take the look from the frontend skill's references when there is no design direction (no named brand, reference, `DESIGN.md`, or existing styled UI to match), and to finish only after `visual-qa` passes on the rendered surface. ([#14604](https://github.com/code-yeongyu/oh-my-openagent/pull/14604), fixes [#14602](https://github.com/code-yeongyu/oh-my-openagent/issues/14602))
15
+
16
+ ### Fixed
17
+
18
+ - Memory sync: a sign-out now retires the trust pin of the account it signs out of, so `omo account show --account <id>` no longer answers `joined` once that sign-in's keys are gone. Plain `omo account forget` (the desktop's sign-out of a CLI-item session) used to retire nothing. It now reads the stored sign-in first to find its account. A stored sign-in that cannot be read is still deleted and its pin is left; one warning line is printed only after the delete succeeds. A forget that fails on a locked store reports only that error (#12308, #12384). ([#13642](https://github.com/code-yeongyu/oh-my-openagent/pull/13642)) ([#14533](https://github.com/code-yeongyu/oh-my-openagent/pull/14533))
19
+ - Memory sync: some sign-ins earned their pin before `omo sync` recorded their account. On a refused refresh, plain `forget`, `forget --device` and `adopt --replace`, omo now asks the service for that account, as `omo logout` already did, and retires the pin when the service names it. It asks only while a live pin exists and the access token has not expired; on a refused refresh, the token may still be valid in the 30 s before expiry, when the refresh runs. An expired token is never sent. A pin whose account stays unknown is kept, and omo prints one line naming `omo account retire-sync --account <id>` to clear it, because `omo account show` keeps answering `joined` for it (#12383, #13662). ([#13642](https://github.com/code-yeongyu/oh-my-openagent/pull/13642)) ([#14533](https://github.com/code-yeongyu/oh-my-openagent/pull/14533))
20
+
21
+ ### Added
22
+
23
+ - `omo sync status` (and `--json`) shows this device's memory sync state: the signed-in account, whether this device has joined and whether its sync switch is on, the account's devices with their sync membership and last-seen time, pending join requests, and whether an engine is running on this machine. It writes no sync state, prints no key material, and each section reports its own failure instead of failing the command. ([#14437](https://github.com/code-yeongyu/oh-my-openagent/pull/14437), fixes [#14435](https://github.com/code-yeongyu/oh-my-openagent/issues/14435)) ([#14533](https://github.com/code-yeongyu/oh-my-openagent/pull/14533))
24
+
25
+ ## [5.1.80] - 2026-11-21
26
+
27
+ ### Fixed
28
+
29
+ - Thread tools: closing a session's gateway store no longer fails a store call that was still running. The close used to reach the store between the call's open and its request, so the call answered `the gateway store is not open`; it now waits for running calls to finish first. ([#14536](https://github.com/code-yeongyu/oh-my-openagent/pull/14536), fixes [#14535](https://github.com/code-yeongyu/oh-my-openagent/issues/14535)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
30
+ - A `thread_send` (or any send path) to a thread whose device left the account no longer resolves the row's stale device id onto a different device that shares its id prefix or carries it as a name: listing-row device ids now resolve as exact ids, answering `not_found` like an unknown device. ([#14521](https://github.com/code-yeongyu/oh-my-openagent/pull/14521)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
31
+ - An unnamed thread on another device is named after the folder its `path` names (the same folder its cwd and refusal candidates use), not the `project` field, so `thread_list`, the row's cwd and its refusal candidate cannot disagree. ([#14521](https://github.com/code-yeongyu/oh-my-openagent/pull/14521)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
32
+ - A thread that is not running no longer stays `queue_full` for good. Messages waiting for it now expire after 24 hours even while it is down, and they stop counting against its 128-message backlog, so later sends are accepted again. A refused send now says the target is not running (`target_running`) and when its oldest message frees a slot (`frees_at`). ([#14469](https://github.com/code-yeongyu/oh-my-openagent/pull/14469), fixes [#14466](https://github.com/code-yeongyu/oh-my-openagent/issues/14466)) ([#14518](https://github.com/code-yeongyu/oh-my-openagent/pull/14518)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
33
+ - A session that stops taking its thread messages no longer cuts itself off from every sender: steers have their own reserve (16 messages / 256 KiB), so they are admitted even when 128 undelivered follow-ups fill the target's backlog, and `queue_full` now says how many messages are held, how many the session took without writing them, how old the oldest is, and why they are not draining. ([#14431](https://github.com/code-yeongyu/oh-my-openagent/pull/14431), fixes [#14428](https://github.com/code-yeongyu/oh-my-openagent/issues/14428)) ([#14518](https://github.com/code-yeongyu/oh-my-openagent/pull/14518)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
34
+ - Advice that forbids ultrawork no longer starts it. "avoid ulw", "no ultrawork", "skip ulw-loop", "without ultrawork" and "don't ulw" are read as advice, the same as "do not use ulw", so pasted guidance and rule lists don't arm the mode. This re-lands #14392, which release train #14415 reverted before it reached dev. ([#14514](https://github.com/code-yeongyu/oh-my-openagent/pull/14514), fixes [#14389](https://github.com/code-yeongyu/oh-my-openagent/issues/14389)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
35
+ - OmO Native now reads `keyword_detector` from `omo.jsonc`. `"keyword_detector": { "disabled_keywords": ["ultrawork"] }` in a project's `.omo/omo.jsonc` or in `~/.omo/omo.jsonc` turns the `ulw`/`ultrawork` keyword off for every launch, and the key is no longer reported as unknown. The `--omo-senpi-ultrawork-disabled` flag still works for a single launch. ([#14514](https://github.com/code-yeongyu/oh-my-openagent/pull/14514), fixes [#14389](https://github.com/code-yeongyu/oh-my-openagent/issues/14389)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
36
+ - The DAG e2e ordering and missing-skill test cases journal to the in-memory store, and that store builds its paths with the file store's own path builder, so host disk latency no longer times them out. ([#14502](https://github.com/code-yeongyu/oh-my-openagent/pull/14502)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
37
+ - When reopening a thread, name devices whose lookup timed out separately from devices that did not answer, and suggest retrying the lookup. ([#14487](https://github.com/code-yeongyu/oh-my-openagent/pull/14487)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
38
+ - Switch live threads with a thinking suffix such as `:max`. If the target catalog has no models from the requested provider, `thread_set_model` now names the provider and lists the providers available there. ([#14482](https://github.com/code-yeongyu/oh-my-openagent/pull/14482)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
39
+ - A Mac first signed in through the desktop app before device names became machine names keeps its `OmO Desktop on <host>` label no longer: the next sign-in refresh renames the device to its machine name, so `<host>` resolves in thread tools and the device mesh. A name you chose is never changed. ([#14480](https://github.com/code-yeongyu/oh-my-openagent/pull/14480)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
40
+ - The Linux keyring probe reports a missing Secret Service only when gdbus says so: no session bus socket and no way to start one, or nothing owns `org.freedesktop.secrets`. Every other gdbus failure (a bus that refuses this process, a refused connection, a timeout, a Secret Service with no default collection) keeps asking libsecret. ([#13640](https://github.com/code-yeongyu/oh-my-openagent/pull/13640)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
41
+ - Thread tools answering about a session on another device that publishes no thread id (`thread_send`, `thread_set_model`, `thread_read`, and their refusals) now name it by its `<device id>/<listing key>` value, the one `thread_list` shows, with a link that opens it, instead of an empty thread. ([#14440](https://github.com/code-yeongyu/oh-my-openagent/pull/14440)) ([#14517](https://github.com/code-yeongyu/oh-my-openagent/pull/14517))
42
+ - `omo remote` no longer lists other devices as choices when the device you named is offline. ([#14440](https://github.com/code-yeongyu/oh-my-openagent/pull/14440)) ([#14517](https://github.com/code-yeongyu/oh-my-openagent/pull/14517))
43
+ - The DAG scheduler test cases now journal to the in-memory store, so host disk latency no longer times out a different case on each run of `scheduler.test.ts` on a loaded machine. ([#14495](https://github.com/code-yeongyu/oh-my-openagent/pull/14495))
44
+ - The DAG scheduler infrastructure-retry test cases now journal to an in-memory store instead of the file-backed store, removing host disk I/O latency as a source of flaky, load-dependent failures. ([#11624](https://github.com/code-yeongyu/oh-my-openagent/pull/11624)) ([#14495](https://github.com/code-yeongyu/oh-my-openagent/pull/14495))
45
+ - `omo --session <id>` attaches to a session the multi-session host (the desktop app) holds again, once the engine includes the matching senpi change. It exited 1 with "--extension cannot be applied to a session another process already runs" because the launcher's own plugin flag looked like one the user had given. Every omo launcher now names that plugin in the brand profile it hands the engine. ([#14467](https://github.com/code-yeongyu/oh-my-openagent/pull/14467))
46
+ - `gpt-6.2-astra-ultrafast` and `gpt-6-astra-ultrafast` are recognized as Astra models, so a DAG run on them gets the Astra verification directives instead of the default ones. ([#14460](https://github.com/code-yeongyu/oh-my-openagent/pull/14460))
47
+ - A thread on another device reports its working folder as `cwd`, not just the folder's name. `thread_list` with `all_scope`, a `thread_create` reopen whose engine answers without a cwd, and a `thread_handoff` or `thread_send` by name all now read the device's listed path, and fall back to the folder name only when the device could not name the folder. Before, they printed the bare folder name, which read like a relative path. ([#14458](https://github.com/code-yeongyu/oh-my-openagent/pull/14458), fixes [#14429](https://github.com/code-yeongyu/oh-my-openagent/issues/14429))
48
+ - Refresh the shipped Senpi plugin bundles and their Linux and macOS verification baselines for the existing GPT-6.2 category rules. The category selection rules themselves are unchanged. ([#14447](https://github.com/code-yeongyu/oh-my-openagent/pull/14447))
49
+ - `omo sync enable` now explains whether first-device recovery phrase setup is disabled in the CLI or the service has not confirmed support. Neither case implies a Mac with Touch ID is required. ([#14425](https://github.com/code-yeongyu/oh-my-openagent/pull/14425))
50
+ - A thread closed on request is no longer opened by a delivery that another path refused while the close settled it: a row whose `target_closed` refusal lost that race is not counted as queued. A wake that finds a closed session's deliveries all settled now reports their drained marks to the device service at once, instead of at a later carry end or the next start. ([#14402](https://github.com/code-yeongyu/oh-my-openagent/pull/14402), fixes [#14383](https://github.com/code-yeongyu/oh-my-openagent/issues/14383)) ([#14415](https://github.com/code-yeongyu/oh-my-openagent/pull/14415))
51
+ - `thread_create` on another device no longer answers `ok` with a thread nothing can address when that device's route does not reach the session its engine opened (for example the Desktop app's host, which the device service does not route): the name was not applied and the first prompt never ran. The create now ends that session where it was opened and answers `host_unavailable` with `details: { device, reason, created }`, so nothing is left behind (`created: false`). ([#14397](https://github.com/code-yeongyu/oh-my-openagent/pull/14397), fixes [#14394](https://github.com/code-yeongyu/oh-my-openagent/issues/14394)) ([#14415](https://github.com/code-yeongyu/oh-my-openagent/pull/14415))
52
+ - Waking many sessions at once no longer leaves messages to a busy session waiting behind the memory component: a session's repeated settle work now replaces itself instead of piling up, and an unchanged session journal is no longer re-read and rewritten on every settle. (#14351) ([#14379](https://github.com/code-yeongyu/oh-my-openagent/pull/14379)) ([#14415](https://github.com/code-yeongyu/oh-my-openagent/pull/14415))
53
+ - Reopening a closed thread with `thread_create({ thread })`, on this device or another, now names the folder the session file records (`cwd`) beside its file. The engine therefore resumes it without probing the file's path on its loop. On a loaded host that probe missed its 500 ms step deadline and refused the reopen with `open_failed: ... a filesystem step did not answer in time`. ([#14393](https://github.com/code-yeongyu/oh-my-openagent/pull/14393), fixes [#14296](https://github.com/code-yeongyu/oh-my-openagent/issues/14296)) ([#14404](https://github.com/code-yeongyu/oh-my-openagent/pull/14404))
54
+ - A `task` call or workflow node that names a builtin agent (`explore`, `librarian`, `plan-consultant`, `plan-reviewer` or a reviewer agent) together with `model` is now refused with `agent_with_model` instead of silently replacing that agent's designated model chain. Change a builtin agent's model with `agents.<name>.model` or `agents.<name>.models` in omo.json; custom agents still accept a call-site model. ([#14378](https://github.com/code-yeongyu/oh-my-openagent/pull/14378)) ([#14404](https://github.com/code-yeongyu/oh-my-openagent/pull/14404))
55
+ - A remote session member or `attach` no longer sends an unresolved session part to its device: only a thread id, a routing id `rpc-N` or a 22-character listing key passes through unlisted (a session opened a moment ago); any other text no listing resolves answers `not_found` with `details.kind: "session"`. ([#14327](https://github.com/code-yeongyu/oh-my-openagent/pull/14327)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
56
+ - A remote miss on an online device that did not answer the listing round answers `host_unavailable` with `details.reason: "listing_incomplete"` naming it, in `sessions.resolve`, the session members and `aliases.resolve` / `set` / `clear`; `device_offline` is only for a device the roster says is offline. A device-qualified miss is never blamed on another device's silence. ([#14327](https://github.com/code-yeongyu/oh-my-openagent/pull/14327)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
57
+ - `bind`, `rebind`, `bindings --session`, `report` and a bound send resolve on this device first; a session only another device holds answers `unsupported` with `details.reason: "session_on_other_device"`, where it answered a bare `not_found`. ([#14322](https://github.com/code-yeongyu/oh-my-openagent/pull/14322)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
58
+ - A thread address that misses while a device did not answer the listing round answers `host_unavailable` with `details.reason: "listing_incomplete"` naming that device, never `not_found`. ([#14318](https://github.com/code-yeongyu/oh-my-openagent/pull/14318)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
59
+ - An offline device resolves by id, name or id prefix alike; only a caller that needs it online (thread_create, the control tools) gets `device_offline`, by every form. ([#14318](https://github.com/code-yeongyu/oh-my-openagent/pull/14318)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
60
+ - `thread_list` and `omo thread list --json` say when one of this device's engines did not answer the listing. The result carries `listing: "incomplete"` and `unanswered_endpoints`, and this device's row reads `incomplete`. Threads held by that engine keep `error_note: "live_unresponsive"` instead of reading as ordinary closed threads. ([#14309](https://github.com/code-yeongyu/oh-my-openagent/pull/14309)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
61
+ - `thread_list` no longer reports a device as `listed` when the listing round did not hear back from it in full (the relay's page cap, a timeout, an undelivered request). Such a device is now `incomplete`. This device's row carries `dropped`, and the result carries `inactive_dropped`: the number of inactive threads the listing left out. ([#14304](https://github.com/code-yeongyu/oh-my-openagent/pull/14304)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
62
+ - `thread_read` shows the error text of an assistant message that ended with a provider error, instead of `[]`. A thread stuck in rate-limit retries no longer reads as a run of empty replies. ([#14283](https://github.com/code-yeongyu/oh-my-openagent/pull/14283)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
63
+ - `thread_list` rows of a running thread whose engine is sleeping between provider retries carry `waiting_on: { kind: "retry", reason: "rate_limit" | "retry", until, attempt, max_attempts }`, so the thread reads as waiting rather than working. This needs an engine that publishes the wait (senpi#7025); older engines list no wait. ([#14283](https://github.com/code-yeongyu/oh-my-openagent/pull/14283)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
64
+ - Following or attaching a session on this same device when its engine is not running, or when the engine does not accept the connection, now says that this device's engine is the problem and to start or restart omo on this device, instead of reporting that the device service did not answer. ([#13975](https://github.com/code-yeongyu/oh-my-openagent/pull/13975)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
65
+ - Reopening a closed thread keeps its name, and a name given at reopen is applied (checked for conflicts when the thread reopens on another device), where a reopened thread lost its name. ([#14282](https://github.com/code-yeongyu/oh-my-openagent/pull/14282), fixes [#14274](https://github.com/code-yeongyu/oh-my-openagent/issues/14274)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
66
+ - A closed thread's session file is found on any engine its device lists when that device's primary engine predates `find_session`, so reopening or reading it no longer fails there; the lookup names why it failed when no engine has the file. ([#14276](https://github.com/code-yeongyu/oh-my-openagent/pull/14276), fixes [#14271](https://github.com/code-yeongyu/oh-my-openagent/issues/14271)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
67
+ - The Kibitzer `tool_result` hook starts recall delivery without waiting for it, so a steer or a ledger write that never settles no longer holds the PostToolUse hook and the rest of the turn. `whenIdle()` still waits for that delivery. ([#13077](https://github.com/code-yeongyu/oh-my-openagent/pull/13077)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
68
+ - A thread message that starts a turn on an idle session now reports `effective_mode: "start"` in the tool result and `effective=start` in the delivery header the receiving session reads. It used to report `follow_up`, which looked as if the message had been queued. The lanes are now `start`, `steer` and `follow_up`, on this device and on another device alike: the result names the lane the receiving session's answer proves, so a message that queued is never reported as a start, and a session woken while it is held or waiting on an approval reports `follow_up`. An explicit `follow_up` to another device no longer fails when the state read before it is lost. The gateway store moves to schema v15 so it can record `start`; an older omo build refuses a v15 store with the existing `gateway_schema_too_new` error. ([#14339](https://github.com/code-yeongyu/oh-my-openagent/pull/14339), part of [#14333](https://github.com/code-yeongyu/oh-my-openagent/issues/14333)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
69
+ - The one-line installer (`curl -fsSL https://get.omo.dev/install.sh | bash`) now installs in a shell whose only user is root, such as proot-distro Ubuntu on Android or a container, instead of stopping with "refusing to run as root". It still refuses `sudo`/`doas` from a normal user, and its message now gives the pipeline-form override `curl -fsSL https://get.omo.dev/install.sh | OMO_INSTALL_ALLOW_SUDO=1 bash`. ([#14346](https://github.com/code-yeongyu/oh-my-openagent/pull/14346))
70
+ - `omo doctor` no longer counts a running one-shot engine command on a compiled install (`omo session-of-pid`, `omo schedule`, `omo models discover`) as an engine session it could report or reap. It now reads those commands from the same list omo hands to the engine. ([#14337](https://github.com/code-yeongyu/oh-my-openagent/pull/14337), fixes [#14336](https://github.com/code-yeongyu/oh-my-openagent/issues/14336))
71
+ - A `thread_send` to a busy thread on another device no longer fails with `unsupported` and "Update omo on that device". When that device refuses the steer because its turn just changed (`turn_conflict`, `not_steerable`), a default-mode send now delivers once as a follow-up (`effective_mode: "follow_up"`), and an explicit steer answers `turn_conflict` or `not_steerable`. ([#14326](https://github.com/code-yeongyu/oh-my-openagent/pull/14326)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
72
+ - `monitor({ type: "rpc", target })` accepts the addresses `thread_send` accepts (a `thread_list` thread value, alias, unique name or id prefix) and `<device name>/<thread>` as well as `<device id>/<thread>`; a target that resolves to nothing is refused at registration with a typed reason (`not_found`, `device_offline`, `ambiguous_target`) instead of starting a watch that can only time out (#14294). ([#14307](https://github.com/code-yeongyu/oh-my-openagent/pull/14307)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
73
+ - On the Claude subscription path, the `task` tool description no longer loses its spawn rules (batch-wide `run_in_background`, model only for `subagent_type`, English prompts, the plan-reviewer one-shot rule) to Claude Code's 4096-char MCP description cut: the rules now come before the agent and category catalogue (#14291). ([#14292](https://github.com/code-yeongyu/oh-my-openagent/pull/14292)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
74
+ - `thread_send` to a session held by a person's Stop, an interrupted run or its parent's run hold no longer answers `idempotency_uncertain`. It answers `queued` with `waiting_on: { kind: "hold" }`: the message is kept and runs once the hold is released, so do not send it again. ([#14287](https://github.com/code-yeongyu/oh-my-openagent/pull/14287)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
75
+ - Thread tools: a reply to a session on another device no longer fails as "not listed on device" while that session is live. The sender's device listing is read once more before any refusal, and each refusal names its reason (`host_unavailable` with `listing_incomplete`, `remote_unavailable`, `wake_unavailable` or `lookup_unavailable`); `not_found` is answered only when the device answered that it lists no such session and its probe found none. ([#14285](https://github.com/code-yeongyu/oh-my-openagent/pull/14285), fixes [#14280](https://github.com/code-yeongyu/oh-my-openagent/issues/14280)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
76
+ - `omo update` now updates the install your PATH actually runs when omo lives in a side-by-side versions layout (`~/.omo/omo-versions/<version>` with `current` pointing at the live one). It installs the new version beside the old one with bun, checks it runs, switches `current` in one step and keeps the previous version for rollback. It no longer runs `npm i -g` into a copy PATH never uses and then reports a false "update failed". If another omo earlier on PATH still answers with the old version, `omo update` says which install it updated and which path PATH still resolves, and exits non-zero. ([#14256](https://github.com/code-yeongyu/oh-my-openagent/pull/14256)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
77
+ - A thread_create whose first prompt is not delivered because this omo build is older than the agent dir's gateway store now answers dev's typed `gateway_schema_too_new` refusal (naming both schemas and the update command, with the new thread's id in details) instead of `internal_error`; the message says the thread was created and how to send its first prompt after updating (#13927). ([#14228](https://github.com/code-yeongyu/oh-my-openagent/pull/14228)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
78
+ - A `thread_*` call naming a session id that no answering device holds, while another device stays silent, now says which devices answered "not held" (this device first) and that the silent one may hold it, with a new `answered` field beside `devices`, instead of asking you only to wait for the silent device. ([#14279](https://github.com/code-yeongyu/oh-my-openagent/pull/14279)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
79
+ - `omo thread` and `omo daemon` in the compiled omo binary load the plugin's thread client again. They printed "the plugin's thread client could not be loaded ... (Cannot find package 'typebox')" and could reach only this device's threads, because the thread client imported TypeBox by name and the binary runtime ships no packages. The thread client now carries TypeBox inside itself. ([#14265](https://github.com/code-yeongyu/oh-my-openagent/pull/14265)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
80
+ - Paging a remote `thread_read` with a cursor from an older omo answers `cursor_stale` (re-read without a cursor) instead of `cursor_invalid`, and `omo thread read` says its cursor pages older messages. ([#14264](https://github.com/code-yeongyu/oh-my-openagent/pull/14264)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
81
+ - A `thread_send` reply no longer waits up to 35 s on an unanswered device-list read; that read ends with the call's own budget and the send answers `host_unavailable`. ([#14263](https://github.com/code-yeongyu/oh-my-openagent/pull/14263)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
82
+ - A `thread_send` to a thread on another device that the listing still shows running, but whose host has parked it, now asks that device again and wakes the session with the message, instead of failing. ([#14262](https://github.com/code-yeongyu/oh-my-openagent/pull/14262)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
83
+ - A blank `thread_create` prompt (only spaces or newlines) is refused with `invalid_arguments` and opens nothing; it used to open an unnamed, empty thread and send it a blank first message. ([#14261](https://github.com/code-yeongyu/oh-my-openagent/pull/14261)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
84
+ - `omo session-of-pid <pid> --json` reaches the engine and answers its one JSON line, on the compiled binary and on npm and bun installs. Before, omo put its plugin option in front of the command, so the engine never saw it and answered `Unknown option: --json`. A test now checks that every command the engine's entry dispatches reaches it unchanged. ([#14255](https://github.com/code-yeongyu/oh-my-openagent/pull/14255), fixes [#14253](https://github.com/code-yeongyu/oh-my-openagent/issues/14253))
85
+
86
+ ### Added
87
+
88
+ - `thread_send` `collapse_key` (`omo thread send --collapse-key`, SDK `send`): a repeated message replaces the sender's earlier undelivered one under the same key, so a notifier that re-sends an alert every few minutes leaves at most one undelivered message. Sessions on this device only. `thread_list` rows show each thread's undelivered `backlog`. ([#14431](https://github.com/code-yeongyu/oh-my-openagent/pull/14431)) ([#14518](https://github.com/code-yeongyu/oh-my-openagent/pull/14518)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
89
+ - The omo SDK signs a machine in by device code and renames a device: `account.signIn({ onUserCode })` shows the code once through the caller's callback, waits for approval and answers the new device, or a typed refusal such as `device_limit` (with the live devices; nothing is removed) or `machine_exists`. `devices.rename({ id, name })` renames one device. `omo account sign-in` drives the same function and prints JSON lines only, so an enroll call on a new machine can read the code before the approval it waits for. ([#14513](https://github.com/code-yeongyu/oh-my-openagent/pull/14513)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
90
+ - A session that ends 10 runs in a row with an error now raises an operator alert. Configure the targets under `operator_alerts.error_streak` in your user config (`~/.omo/omo.jsonc`; a project's `.omo/omo.json` cannot set them): `threads` (thread addresses that receive it through `thread_send`, sent as the failing session) and `webhook_url_env` (the name of an environment variable holding a JSON webhook URL; the body has a one-line `text`, so a Slack incoming webhook accepts it as is). The alert carries only metadata: the session, the device, how many runs failed in each error class (`rate_limit`, `auth`, `overloaded`, `network`, `unknown`), the first and last failure time, and the provider and model. It never carries prompt text, tool output or message content. While the streak lasts, a "still failing" alert repeats at most every `realert_minutes` (default 60). The first normal stop sends one "recovered" notice. Aborted runs never alert. Sending never holds the session. ([#14448](https://github.com/code-yeongyu/oh-my-openagent/pull/14448), fixes [#14430](https://github.com/code-yeongyu/oh-my-openagent/issues/14430))
91
+ - A shipped omo SDK for agent sessions and scripts: in a JS eval cell, `const { createOmoSdk } = await import(`${env("OMO_SDK_ROOT")}/sdk.js`)` gives `account.whoami()`, `devices.list()`, `deviceService.status()` and `config.status()`. Each call answers `{ kind: "ok", value }` or a typed `{ kind: "error", error: { code, message, next_action } }` and never returns a token or key. ([#14374](https://github.com/code-yeongyu/oh-my-openagent/pull/14374)) ([#14415](https://github.com/code-yeongyu/oh-my-openagent/pull/14415))
92
+ - OmO Native ships an `omo-relay` skill for working on the account's other devices from a js eval cell: running commands and jobs, moving and searching files, and running one operation on every machine, on chosen machines or on several threads at once, with one outcome per target and the error codes each one can return. ([#14381](https://github.com/code-yeongyu/oh-my-openagent/pull/14381)) ([#14404](https://github.com/code-yeongyu/oh-my-openagent/pull/14404))
93
+ - `ulw-execute --light` runs a plan with less ceremony: one worktree for the whole plan, no per-step receipts, verifier rounds or evidence files, and one verification against the plan's acceptance criteria at the end. Without the flag nothing changes. ([#14395](https://github.com/code-yeongyu/oh-my-openagent/pull/14395))
94
+ - Thread tools: `thread_send`, `thread_handoff` and `thread_create` take every address form the shared address module resolves, including `<device>/<thread>` (device by name, id or id prefix; thread by id, name, key or id prefix), `here/<thread>` and `<this device>/<thread>`. A device-qualified thread id reads only that device, and a listing row that publishes no thread id is addressed by its listing key. ([#14328](https://github.com/code-yeongyu/oh-my-openagent/pull/14328), part of [#14305](https://github.com/code-yeongyu/oh-my-openagent/issues/14305)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
95
+ - The remote client (thread SDK `remote`) accepts a device by its name (any Unicode form, case or inner spacing), `here`, or a unique id prefix of at least 6 characters in `sessions.resolve`, every session member and `attach`; the device part goes through the shared address module's resolver. ([#14327](https://github.com/code-yeongyu/oh-my-openagent/pull/14327)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
96
+ - The thread control tools (`thread_read`, `thread_close`, `thread_interrupt`, `thread_set_model`, `thread_set_reasoning`, `thread_rename`, `thread_archive`) accept `<device>/<thread>` and `here/<thread>`. The device part is its id, its name or a unique id prefix of at least 6 characters; the thread part is a thread id, name, alias, listing key or an 8-hex id prefix among that device's threads. One address module (`@oh-my-opencode/thread-address-core` plus `thread/address/`) now owns the grammar, the device resolver and the refusals. ([#14318](https://github.com/code-yeongyu/oh-my-openagent/pull/14318)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
97
+ - `bun run qa:rpc-relay-parity` measures the engine's RPC methods on a pair of signed-in devices: the same record locally and through the relay, compared and classified `same`, `differs` or `missing`. The local-versus-relay parity table (docs/reference/rpc-relay-parity.md) now records a measured verdict for 113 of the engine's 128 methods: the 15 `connection_local` methods are excluded by design, and the 36 rows marked `unsafe-to-probe` were probed only in a harmless form, with each row giving the answer both legs returned. ([#13895](https://github.com/code-yeongyu/oh-my-openagent/pull/13895)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
98
+
99
+ ### Changed
100
+
101
+ - The device and fleet keyword hint in OmO Native now points the agent at the `omo-relay` skill, instead of repeating the fleet client's import and shell details. It also fires when the user names their other machines or devices ("run this on my other machine"). The two warnings that keep a JS cell from exiting the kernel stay in the hint. ([#14416](https://github.com/code-yeongyu/oh-my-openagent/pull/14416)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
102
+ - On Linux with no keyring (no Secret Service on the session bus, or a Bun without `Bun.secrets`), `omo login` and `omo account` keep the shared sign-in in a private file, `~/.omo/account-state/omo-cli-<id>.json` (directory 0700, file 0600), instead of failing with `store-unavailable`. A locked keyring or a keyring whose store fails still reports its own error and never falls back to the file. When a keyring runs later, the sign-in moves into it only when the keyring holds no sign-in or an older copy of the same one, and the file is deleted only after the keyring returns exactly what was written. A keyring that holds another sign-in keeps it, the file is left as it is, and one line says so. A sign-out removes only the sign-in in use. On Linux, the errors no longer say "Credentials are never saved to a file."; macOS and Windows are unchanged. ([#13640](https://github.com/code-yeongyu/oh-my-openagent/pull/13640)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
103
+ - `omo account show` and `omo account adopt` answer an additional `storage` field, `"keyring"` or `"file"`, naming where the shared sign-in is kept. ([#13640](https://github.com/code-yeongyu/oh-my-openagent/pull/13640)) ([#14537](https://github.com/code-yeongyu/oh-my-openagent/pull/14537))
104
+ - Internal: the generated bundles checked in on `dev` match their sources again. Several had gone stale after merges that changed sources without regenerating them, which made every PR's bundle check against `dev` fail. No behavior change. ([#14501](https://github.com/code-yeongyu/oh-my-openagent/pull/14501), fixes [#14468](https://github.com/code-yeongyu/oh-my-openagent/issues/14468))
105
+ - DeepSeek V4.1 Flash now receives the GPT ultrawork prompt on every id shape (`deepseek-flash`, `deepseek-v4.1-flash`, `deepseek-v4p1-flash`, `deepseek-v4-1-flash`, with a provider prefix or a tier, snapshot or mode suffix). Its system prompt is a copy of the GPT-6.2 one, and the default ultrawork prompt gave it opposite answers on reporting, delegation, retries and stopping. V4 Flash, its 0731 snapshot and V4 Pro keep the default prompt. ([#14462](https://github.com/code-yeongyu/oh-my-openagent/pull/14462))
106
+ - Ultrawork now has two complete prompts, a default one and one for GPT models, each written as a whole file. GPT models (GPT-6.2 Astra and Sol on every speed tier, including `-ultrafast`, and the GPT-6 and GPT-5 families) receive the GPT prompt: it keeps the registered goal, the notepad, real-surface evidence, cleanup, commits and the review gate, and leaves reading, reporting cadence, retries and todo upkeep to the model's system prompt, so the two no longer give opposite instructions. Every other model receives the default prompt, whose text is unchanged. ([#14460](https://github.com/code-yeongyu/oh-my-openagent/pull/14460), fixes [#14441](https://github.com/code-yeongyu/oh-my-openagent/issues/14441))
107
+ - The built-in deep categories now require one exact model each: GPT-6.2 Astra at max for `ultrabrain`, GPT-6.2 Astra at high for `deep-high`, and plain GPT-6.2 Sol at medium for `deep-low`. Providers can still supply the same model, but an older model or a Fast sibling no longer activates these categories. Explicit model overrides still work. The Codex installer and bundled Astra agents now default to GPT-6.2 Astra. ([#14427](https://github.com/code-yeongyu/oh-my-openagent/pull/14427))
108
+ - The bundled senpi engine is now 2026.11.19-5, which resumes a host that was killed twice instead of leaving its sessions stranded. ([#14417](https://github.com/code-yeongyu/oh-my-openagent/pull/14417))
109
+ - `thread_list` rows of a thread held on this device carry `engine_version`, the version of the engine its host runs (the engine's own version, else an older host's package version), read from the `get_protocol_info` answer the listing connection already opens with. Another device's rows carry none yet. ([#14409](https://github.com/code-yeongyu/oh-my-openagent/pull/14409)) ([#14415](https://github.com/code-yeongyu/oh-my-openagent/pull/14415))
110
+ - `omo.dev/download` (and every `/<locale>/download`) now sends you to the stable desktop download page, https://get.omo.dev/desktop/stable/, which links the beta page; `/download?channel=beta` still opens https://get.omo.dev/desktop/beta/. The home page's desktop button now reads "Download the desktop app" in every language and opens the same stable page. ([#14401](https://github.com/code-yeongyu/oh-my-openagent/pull/14401))
111
+ - OmO now runs on senpi 2026.11.19-3. A running thread that is sleeping between provider retries now lists `waiting_on` (senpi#7027), and a message from another session to an idle terminal starts its turn even while a draft sits in the composer, leaving the draft untouched (senpi#7045). ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
112
+ - The thread send default is now one shared constant, `DEFAULT_THREAD_DELIVERY` (`auto`: steer a running turn, start an idle one). `thread_send`, `thread_create`'s first prompt, `thread_handoff`, the thread SDK, `omo thread send`, `omo remote send|open` and the gateway engine all read it instead of keeping their own copy. The `delivery` schema now declares `default: "auto"`, and its description says that `follow_up` is only for a message that must wait for the receiver's current turn. ([#14334](https://github.com/code-yeongyu/oh-my-openagent/pull/14334), part of [#14333](https://github.com/code-yeongyu/oh-my-openagent/issues/14333)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
113
+ - `omo remote send|model|interrupt|close|attach` resolve their target through the shared address module: a thread id, a listing key, an alias, a name or an 8+ hex id prefix, alone or behind `<device>/`, where the device is its id, `here`, its name or a 6+ character id prefix. Before, only a thread id and `<device>/<session>` were accepted. An ambiguous name or prefix is refused with its candidates and nothing runs, a device that missed the listing round answers `host_unavailable` instead of a plain miss, and `model`, `interrupt`, `close` and `attach` refuse an offline device as `device_offline` by every form. ([#14331](https://github.com/code-yeongyu/oh-my-openagent/pull/14331), part of [#14305](https://github.com/code-yeongyu/oh-my-openagent/issues/14305)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
114
+ - `omo remote` session verbs on `<device>/<thread>` now read one filtered session listing of that device before the command, so a mistyped thread is a typed `not_found` with nothing sent; on dev the text went to the device unread. A thread id or listing key that device does not list yet is asked of the device itself only when this device's service is 0.1.31 or later; an older service answers `host_unavailable` (`wake_unavailable`) and sends nothing, where dev sent it straight to the device. ([#14331](https://github.com/code-yeongyu/oh-my-openagent/pull/14331)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
115
+ - `omo remote` session verbs: a thread this device holds answers `not_found` with nothing sent, whatever form reached it, and an ambiguous device, name or prefix prints its candidates one per line. ([#14331](https://github.com/code-yeongyu/oh-my-openagent/pull/14331)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
116
+ - `omo remote attach` checks the relay before it attaches, as the other session verbs do: a device service that is off, signed out, not set up, waiting for sign-in, still connecting or on a plan without remote sessions now answers `host_unavailable` (or the plan refusal) naming its state, and nothing is attached; before, attach had no relay check. ([#14331](https://github.com/code-yeongyu/oh-my-openagent/pull/14331)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
117
+ - `omo remote sessions --device` names the candidates when the device it was given is ambiguous. ([#14331](https://github.com/code-yeongyu/oh-my-openagent/pull/14331)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
118
+ - `omo account devices revoke` takes a device by id, name or unique id prefix. A name or prefix two devices share is refused with both candidates (in `--json`, the `candidates` ids) and nothing is revoked; `here` and this device keep the "that is this device" refusal. ([#14331](https://github.com/code-yeongyu/oh-my-openagent/pull/14331)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
119
+ - Thread tools: a send to a name or id that no device lists, while a device missed this listing round, now answers `host_unavailable` (`listing_incomplete`, naming that device) instead of `not_found`. A reply to a sender whose delivery names no device reads the account listing once instead of twice. In `ambiguous_target` candidates, a thread on this device carries this device's id as `device` (and an `address` that resolves as written) instead of `"here"`. A thread named on an offline device that its last listing does not show answers `device_offline`, and nothing is sent to that device. ([#14328](https://github.com/code-yeongyu/oh-my-openagent/pull/14328)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
120
+ - A caller's `deadlineMs` now covers a whole remote `sessions.resolve` or `attach` call: the roster read, the listing and the attach (or a session member's call) share one deadline instead of each taking it in full. A spent budget answers the typed `timeout`, whose `details.deadline_ms` is what was left of it. ([#14327](https://github.com/code-yeongyu/oh-my-openagent/pull/14327)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
121
+ - The thread SDK and `omo thread` resolve every address (`read`, `send`, `bind`, `rebind`, `bindings --session`, `report`, `locate`) through the shared thread address module, so they accept the same forms as the thread tools, `<device>/<thread>` and `here/<thread>` included, with or without the account's devices. ([#14322](https://github.com/code-yeongyu/oh-my-openagent/pull/14322)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
122
+ - A `send` to a name only another workspace of this device holds answers `scope_denied` unless `--all-scope`, as `read` does, where it answered `not_found`. ([#14322](https://github.com/code-yeongyu/oh-my-openagent/pull/14322)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
123
+ - A chat connector that delivers a message through a closed binding, or through a binding that takes no inbound messages, is now told to deliver through the binding the gateway keeps for that chat. Before, it got the steps meant for a reporting session. ([#13994](https://github.com/code-yeongyu/oh-my-openagent/pull/13994)) ([#14356](https://github.com/code-yeongyu/oh-my-openagent/pull/14356))
124
+ - OmO now runs on senpi 2026.11.19-2. A message one session sends to another that is paused only because nobody is watching it (an unattended hold) now continues that session instead of waiting silently. ([#14287](https://github.com/code-yeongyu/oh-my-openagent/pull/14287)) ([#14332](https://github.com/code-yeongyu/oh-my-openagent/pull/14332))
125
+ - Test fixtures and one source comment no longer name internal machines or local paths; they use neutral names instead. No behavior change. ([#14281](https://github.com/code-yeongyu/oh-my-openagent/pull/14281)) ([#14284](https://github.com/code-yeongyu/oh-my-openagent/pull/14284))
126
+
127
+ ## [5.1.79] - 2026-11-19
128
+
129
+ ### Fixed
130
+
131
+ - The `/remote` close in the TUI follows `thread_close`'s close contract (#14234). A close refused because the thread has messages it has not read names `/remote close <session> --force` and lists those messages, or says the ids are unavailable on that device's engine. The picker then offers "Close anyway". A forced close lists the messages it dropped. The refusal's `details.reason` is the engine's own reason (`turn_active` or `work_pending`). ([#14246](https://github.com/code-yeongyu/oh-my-openagent/pull/14246), fixes [#14245](https://github.com/code-yeongyu/oh-my-openagent/issues/14245)) ([#14248](https://github.com/code-yeongyu/oh-my-openagent/pull/14248))
132
+ - Thread tools: a command for a thread whose session parked on another device after it was listed no longer fails `unsupported` with advice to update omo on that device. The device service answers such a command `route_unlisted` (its host no longer lists that session, so nothing was sent to it). `thread_close` now answers `closed: false, reason: "not_live"`, and the other thread tools answer `not_resumable`, as for a session that ended there. A create, or a reopen's session-file lookup, refused that way is `host_unavailable`, with a retry. ([#14237](https://github.com/code-yeongyu/oh-my-openagent/pull/14237), fixes [#14233](https://github.com/code-yeongyu/oh-my-openagent/issues/14233)) ([#14248](https://github.com/code-yeongyu/oh-my-openagent/pull/14248))
133
+ - The cold-wake session host unit (and every other unit or plist the cold-wake install writes) is now written 0644 under any umask. Under umask 077 it was written 0600, because a file's creation mode is masked by the umask; the mode is now set explicitly before the file is moved into place. ([#14249](https://github.com/code-yeongyu/oh-my-openagent/pull/14249), fixes [#14247](https://github.com/code-yeongyu/oh-my-openagent/issues/14247)) ([#14248](https://github.com/code-yeongyu/oh-my-openagent/pull/14248))
134
+ - A cold-wake unit or plist an earlier build already left 0600 is set back to 0644 the next time the install (or a refresh to a newer runtime) finds its content unchanged, with no reload or restart. A file this user cannot change the mode of (another owner's, or on a read-only filesystem) is left as it is and the install still succeeds. ([#14249](https://github.com/code-yeongyu/oh-my-openagent/pull/14249)) ([#14248](https://github.com/code-yeongyu/oh-my-openagent/pull/14248))
135
+
136
+ ### Changed
137
+
138
+ - **`thread_close` no longer drops messages a thread has not read unless you ask it to.** Before, it always interrupted the session and silently dropped any messages queued behind its run. Now:
139
+ - A busy thread with unread messages is refused `turn_active`. The refusal names those messages (`pending_deliveries`) and gives the `force` option.
140
+ - `thread_close({ thread, force: true })` closes it anyway.
141
+ - Every close result carries `dropped_deliveries`, the messages it handed back unread (an empty list when there were none).
142
+ - An idle thread still closes without `force`, even when an interrupted run left messages stranded (#14169); those are listed too.
143
+ - On a device whose engine cannot name its queued messages (older than code-yeongyu/senpi#7006), a default close never interrupts. It says the message ids are unavailable there, and `force: true` still closes it.
144
+ - `omo remote close` takes `--force` and reports the dropped messages as well. ([#14242](https://github.com/code-yeongyu/oh-my-openagent/pull/14242), fixes [#14234](https://github.com/code-yeongyu/oh-my-openagent/issues/14234)) ([#14248](https://github.com/code-yeongyu/oh-my-openagent/pull/14248))
145
+
146
+ ### Added
147
+
148
+ - The desktop can ask OmO which model each category and agent runs. A new read-only extension request, `omo.modelRouting.resolve`, answers the model every category (under its canonical name) and every task agent would spawn on, using the same resolvers a `task` call uses against the session's project config, model registry and sign-ins, and says whether that model came from your omo.json or from OmO's defaults. ([#14025](https://github.com/code-yeongyu/oh-my-openagent/pull/14025))
149
+
10
150
  ## [5.1.78] - 2026-11-18
11
151
 
12
152
  ### Changed
package/plugin/README.md CHANGED
@@ -4,7 +4,7 @@
4
4
 
5
5
  ## Shipped skills
6
6
 
7
- The plugin ships 27 skills by default:
7
+ The plugin ships 28 skills by default:
8
8
 
9
9
  - `ast-grep` — structural code search and rewrites
10
10
  - `browser` — drive a real browser from the js eval kernel: signed-in sites, forms, screenshots, web QA
@@ -20,6 +20,7 @@ The plugin ships 27 skills by default:
20
20
  - `init-deep` — deep project initialization and scaffold setup
21
21
  - `lsp-setup` — Language Server configuration and troubleshooting
22
22
  - `mass-ulw` — dependency-ordered child work through the native workflow tool
23
+ - `omo-relay` — work on the account's other machines and threads, one at a time or all at once
23
24
  - `onboarding` — guide a first-time omo user through setup and the core workflow
24
25
  - `programming` — type-strict, modern-stack programming guidance
25
26
  - `refactor` — safe refactoring and code quality improvement