@awebai/oats 0.68.0 → 0.69.1
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.
- package/bin/oats.mjs +102 -22
- package/docs/capabilities.md +49 -5
- package/docs/configuration.md +68 -2
- package/docs/desktop-cli-api.md +355 -12
- package/docs/desktop-instance-start.md +88 -3
- package/docs/desktop.md +147 -13
- package/docs/execution-targets.md +514 -62
- package/docs/harness-trust.md +3 -1
- package/docs/implementation.md +28 -11
- package/docs/integrations.md +1 -1
- package/docs/oats-local.schema.json +2 -2
- package/docs/official-catalog.md +4 -3
- package/docs/packages.md +7 -7
- package/docs/release-notes/v0.68.1.md +102 -0
- package/docs/release-notes/v0.69.0.md +658 -0
- package/docs/release-notes/v0.69.1.md +25 -0
- package/docs/servers.md +44 -5
- package/docs/souls-and-instances.md +11 -7
- package/docs/tui.md +114 -31
- package/docs/upgrading.md +14 -5
- package/docs/workspaces.md +1 -1
- package/lib/core.mjs +345 -89
- package/lib/file-pieces.mjs +49 -0
- package/lib/instance-events.mjs +25 -4
- package/lib/launch-prompt-fixtures/README.md +62 -0
- package/lib/launch-prompt-fixtures/resume-confirmation.json +32 -0
- package/lib/launch-prompt-fixtures.mjs +49 -0
- package/lib/launch-prompt-resume.mjs +55 -0
- package/lib/login-environment.mjs +27 -9
- package/lib/schedule.mjs +3 -3
- package/lib/servers.mjs +117 -23
- package/lib/session-end.mjs +82 -0
- package/lib/session-input.mjs +43 -9
- package/lib/session-resume-contract.mjs +32 -0
- package/lib/session-resume.mjs +310 -0
- package/lib/session-seen.mjs +163 -0
- package/lib/session-viewer.mjs +147 -44
- package/lib/soul-composition.mjs +2 -1
- package/lib/start-env.mjs +110 -0
- package/lib/tmux-server-scope.mjs +139 -14
- package/lib/workspace.mjs +38 -4
- package/package-catalog.json +10 -4
- package/package.json +1 -1
- package/packages/client/cli-adapter.mjs +9 -0
- package/packages/client/cli-locator.mjs +3 -2
- package/packages/client/deployment-data.mjs +7 -0
- package/packages/client/deployment-panel.mjs +3 -0
- package/packages/client/file-output-exec.mjs +98 -0
- package/packages/client/instance-display-name.mjs +18 -0
- package/packages/client/launch-prompt-outcome.mjs +25 -13
- package/packages/client/own-codes.mjs +2 -0
- package/packages/client/session-attach.mjs +35 -5
- package/packages/client/session-resume-cli.mjs +64 -0
- package/packages/client/session-resume-trigger.mjs +167 -0
- package/packages/client/session-resume.mjs +332 -0
- package/packages/client/tmux-status.mjs +30 -7
- package/packages/client/waiting-on-you.mjs +11 -4
- package/packages/client/workspace-step.mjs +17 -5
- package/packages/record/lib/last-conversation.mjs +123 -0
- package/packages/record/lib/session-roots.mjs +10 -1
- package/packages/tui/README.md +1 -1
- package/packages/tui/lib/client.mjs +2 -1
- package/packages/tui/lib/main.mjs +5 -3
- package/packages/tui/lib/model.mjs +57 -15
- package/packages/tui/lib/store.mjs +8 -7
- package/packages/tui/lib/term/text.mjs +12 -9
- package/packages/tui/lib/views/capabilities.mjs +6 -6
- package/packages/tui/lib/views/frame.mjs +72 -7
- package/packages/tui/lib/views/help.mjs +9 -4
- package/packages/tui/lib/views/instances.mjs +46 -29
- package/packages/tui/lib/views/souls.mjs +9 -8
- package/skills/oats-getting-started/SKILL.md +117 -4
|
@@ -0,0 +1,658 @@
|
|
|
1
|
+
# OATS 0.69.0
|
|
2
|
+
|
|
3
|
+
Unreleased
|
|
4
|
+
|
|
5
|
+
## Upgrade
|
|
6
|
+
|
|
7
|
+
- **Install the Desktop first and the CLI second,** as
|
|
8
|
+
[Upgrading](../upgrading.md) says: a 0.68 Desktop refuses a 0.69.0 CLI. The
|
|
9
|
+
0.69.0 Desktop accepts OATS CLIs `>=0.26.0 <0.70.0`.
|
|
10
|
+
|
|
11
|
+
- **Package moves:** the official catalog pins oats.aweb 1.32.0,
|
|
12
|
+
oats.framework 1.15.1 and oats.engineering 1.15.0 (Provider updates, below).
|
|
13
|
+
oats.aweb 1.32.0 requires OATS >=0.68.0: every host that shares a
|
|
14
|
+
workspace runs OATS 0.68.0 or later before any of them syncs to it. A
|
|
15
|
+
deployment then takes them with `oats sync`; new spawns use them, and a
|
|
16
|
+
stopped home moves to them with `oats instance upgrade`.
|
|
17
|
+
|
|
18
|
+
- A broken machine file (unreadable, not YAML, empty, another top-level key,
|
|
19
|
+
an invalid entry) refuses every launch that reads launch configurations,
|
|
20
|
+
with `E_LAUNCH_CONFIG_INVALID` naming the file. A missing file changes
|
|
21
|
+
nothing.
|
|
22
|
+
|
|
23
|
+
- An instance last started before this version has no record of its launch
|
|
24
|
+
environment. Its first resume after the upgrade goes ahead with one warning:
|
|
25
|
+
"`<instance>`'s launch environment could not be verified this once (it was
|
|
26
|
+
last started before OATS recorded it); its next start records it". From
|
|
27
|
+
then on it is checked (#1559).
|
|
28
|
+
|
|
29
|
+
- **Run `oats status` once before you restart.** After upgrading, open OATS
|
|
30
|
+
Desktop or `oats tui`, or run `oats status`, once before you restart this
|
|
31
|
+
computer, or nothing from before the upgrade comes back: the sweep only
|
|
32
|
+
resumes what OATS saw running.
|
|
33
|
+
- **A boot nobody looked at.** A whole boot with no OATS read leaves that boot
|
|
34
|
+
unknown: what ran in it and was lost there is not resumed.
|
|
35
|
+
- **One legacy gap.** An instance spawned before this release, never
|
|
36
|
+
restarted since, whose harness ends by itself after the upgrade, reads
|
|
37
|
+
`lost` and comes back once. Its next start records what is needed; stop it
|
|
38
|
+
with OATS to keep it stopped.
|
|
39
|
+
- **Claude's resume-cost dialog.** On some plans Claude asks, at the first
|
|
40
|
+
message of a resumed long conversation, whether to resume from a summary or
|
|
41
|
+
in full. OATS sends no key: the person chooses. The resume is reported
|
|
42
|
+
waiting (`resume-confirmation`), open the instance's terminal to answer.
|
|
43
|
+
The dialog is recognised from Claude Code's strings and documentation, not
|
|
44
|
+
from a captured screen.
|
|
45
|
+
- **Codex starts fresh.** A Codex instance is not resumed until a Codex resume
|
|
46
|
+
is proven end to end on a signed-in host: it starts fresh on `TASK.md`
|
|
47
|
+
(`resume-unproven`).
|
|
48
|
+
- **No recovery copy for an exited harness.** The exit marker no longer counts
|
|
49
|
+
as changed home bytes at retire, so an instance whose harness ended needs no
|
|
50
|
+
recovery copy because of it. Retirement baselines keep their digest version.
|
|
51
|
+
|
|
52
|
+
## Added
|
|
53
|
+
|
|
54
|
+
- Sessions come back after a shutdown, a reboot or a crash (feature
|
|
55
|
+
`session-resume`, #1559). `oats session resume` starts every instance whose
|
|
56
|
+
session OATS saw alive and then lost, on every deployment in this
|
|
57
|
+
computer's list, each in its
|
|
58
|
+
own harness conversation (Claude `--resume`, pi `--session`) with the launch
|
|
59
|
+
it recorded and a one-line resume notice as its prompt. Running instances
|
|
60
|
+
are left alone; instances stopped with OATS, or whose harness ended by
|
|
61
|
+
itself, stay stopped: stopping with OATS is the only way to keep an instance
|
|
62
|
+
down. One sweep runs per machine (`E_RESUME_BUSY`); each loss is answered
|
|
63
|
+
once, so a refused resume is not tried again, after a reboot either; and a
|
|
64
|
+
home is started by sweeps at most 3 times in a boot (on Linux and macOS,
|
|
65
|
+
which have a boot to count in), then reported `keeps-getting-lost` and left
|
|
66
|
+
down, after a reboot too, until someone starts it.
|
|
67
|
+
**An agent killed outside OATS (closed from tmux, or killed by a signal)
|
|
68
|
+
comes back within 24 hours: only `oats instance stop` keeps it down.** The
|
|
69
|
+
clients run it when they open, in a later change. See
|
|
70
|
+
[Resume after a shutdown](../execution-targets.md#resume-after-a-shutdown).
|
|
71
|
+
- `oats session start --resume --home <abs>`: one home's resume. It refuses a
|
|
72
|
+
launch choice (`E_BAD_ARGS`) and a home that did not lose its session
|
|
73
|
+
(`E_RESUME_NOT_LOST`); with no conversation to continue it starts fresh on
|
|
74
|
+
`TASK.md` and says why (`resumed: false`). Its answer and the sweep's are in
|
|
75
|
+
the [Desktop CLI API](../desktop-cli-api.md#resume-a-lost-session---resume).
|
|
76
|
+
A warning the sweep's starts say of the computer rather than of their
|
|
77
|
+
instance (today only that harnesses get no scope of their own) is said once,
|
|
78
|
+
in the sweep's own `warnings`, not in each row.
|
|
79
|
+
- `sessionEnd` (`lost`, `stopped`, `exited`) on `oats status --json` rows of
|
|
80
|
+
launched instances that do not run.
|
|
81
|
+
- Only a loss OATS saw is resumed. The sweep brings back a lost instance when
|
|
82
|
+
OATS saw it running (in `oats status`, OATS Desktop or `oats tui`), no
|
|
83
|
+
earlier sweep has answered that loss, OATS first saw it stopped less than a
|
|
84
|
+
day ago, and, if the computer did not restart in between, it stopped less
|
|
85
|
+
than a day after OATS last saw it run. Any other lost instance is reported
|
|
86
|
+
`not-seen-recently` and left as it is; start it by hand to bring it back.
|
|
87
|
+
So an agent killed in the night comes back in the morning, and a laptop
|
|
88
|
+
closed for two weeks brings back what ran when it was closed.
|
|
89
|
+
- A local `oats status` now records what it saw of each instance, for the
|
|
90
|
+
sweep: in a private file under `~/.oats/session-resume/`, one per
|
|
91
|
+
deployment. It never changes what status prints or exits with. A status
|
|
92
|
+
run on another machine with `--server` records nothing on this one.
|
|
93
|
+
|
|
94
|
+
- **Desktop: lost sessions are resumed.** After a shutdown, a reboot or a
|
|
95
|
+
crash, opening Desktop asks OATS to resume every instance whose session was
|
|
96
|
+
lost, on every deployment on this computer. While Desktop is open, it asks
|
|
97
|
+
again when it sees a session that was lost, at most once every 30 seconds,
|
|
98
|
+
and not again for an instance OATS could not resume. None of this holds
|
|
99
|
+
anything up. When an instance was resumed, started fresh, is waiting for you
|
|
100
|
+
or could not be resumed, a notice says so and stays until you dismiss it.
|
|
101
|
+
Its Details name each instance and what happened to it, and say which
|
|
102
|
+
deployments were not checked. An instance that keeps losing its session is
|
|
103
|
+
named once, and OATS stops resuming it until you start it yourself. More
|
|
104
|
+
than three notices fold into one. Nothing lost means no notice. Quitting
|
|
105
|
+
Desktop while the sessions are being resumed does not cut that short
|
|
106
|
+
([desktop-instance-start.md](../desktop-instance-start.md#when-desktop-opens)).
|
|
107
|
+
|
|
108
|
+
- A resume starts only under the `{fromEnv: SRC}` values its instance's last
|
|
109
|
+
start ran with. A value that changed since refuses the resume before
|
|
110
|
+
anything is launched: `E_LAUNCH_ENV_CHANGED`, naming the variables, never
|
|
111
|
+
their values. `oats session resume` reports it as a refused start
|
|
112
|
+
(`refused`, `start-refused`), and `--plan` predicts it. Start the instance
|
|
113
|
+
yourself to use the new values: an ordinary start never compares. Every
|
|
114
|
+
start records its launch environment as a keyed one-way digest per variable,
|
|
115
|
+
in two kernel-private files in the home. A value, which may be a credential,
|
|
116
|
+
is never written (#1559). See
|
|
117
|
+
[The launch environment a resume runs under](../execution-targets.md#the-launch-environment-a-resume-runs-under).
|
|
118
|
+
|
|
119
|
+
- **`Shift+PageUp` scrolls back in `oats session attach`**
|
|
120
|
+
([#1476](https://github.com/awebai/oats/issues/1476)). Every viewer binds
|
|
121
|
+
it to tmux's copy mode, a page up, with or without `--detach-key`. Under an
|
|
122
|
+
agent that takes the mouse, as Claude Code does, the wheel goes to the agent,
|
|
123
|
+
and a viewer had no way to scroll back. In copy mode `PageUp` and `PageDown`
|
|
124
|
+
move and `q` leaves it. See
|
|
125
|
+
[Inspect, input and attach](../execution-targets.md#inspect-input-and-attach).
|
|
126
|
+
|
|
127
|
+
- **Desktop: scroll back in a terminal with the keyboard.** Shift+PgUp, or
|
|
128
|
+
Ctrl+B then `[`, scrolls back through an agent's output, as the wheel does,
|
|
129
|
+
in every terminal tab on this computer, and on another computer running
|
|
130
|
+
OATS 0.69.0 or later. Ctrl+B followed by any other key still reaches the
|
|
131
|
+
agent as both keys. Ctrl+B `[` needs tmux 3.3 or later on the computer the
|
|
132
|
+
agent runs on. The shortcuts window lists the three under Tabs
|
|
133
|
+
([desktop.md](../desktop.md#copy-from-a-terminal),
|
|
134
|
+
[desktop-keyboard.md](../../packages/desktop/docs/desktop-keyboard.md)).
|
|
135
|
+
|
|
136
|
+
- **Desktop: the shortcuts window's Most useful names how to scroll back in a
|
|
137
|
+
terminal** ([#1502](https://github.com/awebai/oats/issues/1502)): "Wheel
|
|
138
|
+
up, Shift+PgUp or Ctrl+B then `[`", as the Tabs group's row does. The Tabs
|
|
139
|
+
row's tooltip also says that on another computer the two keys need OATS
|
|
140
|
+
0.69.0 or later there. On Linux and Windows, when a shortcut that works in
|
|
141
|
+
a terminal is bound to Ctrl+B, both rows leave Ctrl+B `[` out and the
|
|
142
|
+
tooltip names the shortcut
|
|
143
|
+
([desktop-keyboard.md](../../packages/desktop/docs/desktop-keyboard.md)).
|
|
144
|
+
|
|
145
|
+
- **Desktop: an agent's question shows over its terminal.** While an instance
|
|
146
|
+
waits on you, each terminal tab that shows it has a row over the terminal.
|
|
147
|
+
The row says **Needs input** and how long it has waited, and under that the
|
|
148
|
+
agent's question, as plain text. When the agent left no text, the row gives
|
|
149
|
+
the reason in words, for example "Waiting for a tool approval". The question
|
|
150
|
+
wraps to at most three lines, never takes more than a quarter of the pane,
|
|
151
|
+
and ends with "…" when it is longer. The terminal resizes only when the wait
|
|
152
|
+
starts, ends or its text changes, never as the age ticks over. In a split,
|
|
153
|
+
each group has its own row
|
|
154
|
+
([desktop.md](../desktop.md#when-an-agent-waits-on-you)).
|
|
155
|
+
|
|
156
|
+
- **TUI: the cheat sheet names a terminal's scroll keys**
|
|
157
|
+
([#1502](https://github.com/awebai/oats/issues/1502)). While a terminal
|
|
158
|
+
has the focus, the cheat sheet and, where it has room, a zoomed
|
|
159
|
+
terminal's title row name the viewer's own keys: `Shift+PageUp or
|
|
160
|
+
Ctrl+B [`, then `scroll back` (`Ctrl+B [` alone where the row holds one
|
|
161
|
+
key). In the Alt family a zoomed terminal 120 columns wide or less has no
|
|
162
|
+
room for them: its chords keep the row.
|
|
163
|
+
- The keys are the ones the tab's host's kernel says its viewer takes:
|
|
164
|
+
`Shift+PageUp` where it lists `session-attach-prefix-scroll`, and
|
|
165
|
+
`Ctrl+B [` too where its `sessionViewer.prefixScroll` is true. That is
|
|
166
|
+
this machine's `oats version --json` for a local tab, and the roster
|
|
167
|
+
probe of its server for a remote one. A host without the feature gets
|
|
168
|
+
no scroll keys named. `Ctrl+B [` is never named when the leader is
|
|
169
|
+
`Ctrl+B` (`open.detachKey: "C-b"`), which takes that key first.
|
|
170
|
+
- Each tab's viewer runs without its own status line where the kernel
|
|
171
|
+
has `--no-status-line`: the cheat sheet says it instead.
|
|
172
|
+
- See [the TUI's tabs](../tui.md#tabs).
|
|
173
|
+
|
|
174
|
+
- **TUI: an open agent that waits on you says what it waits on, above its
|
|
175
|
+
terminal** ([#1502](https://github.com/awebai/oats/issues/1502)).
|
|
176
|
+
- While the shown tab's agent has a live claim, a row stands under its
|
|
177
|
+
tab bar: `Needs input`, how long it has waited, and the claim's
|
|
178
|
+
message.
|
|
179
|
+
- A claim with no message says what it waits on, for example
|
|
180
|
+
`waiting on a permission prompt`.
|
|
181
|
+
- The message wraps to at most 3 rows, and to a quarter of the group's
|
|
182
|
+
height at most.
|
|
183
|
+
- The agent's pane is resized only when the claim starts, clears or
|
|
184
|
+
changes its text. Before, the question showed only in the agent's
|
|
185
|
+
detail, which an open terminal covers. See
|
|
186
|
+
[Tabs](../tui.md#tabs).
|
|
187
|
+
|
|
188
|
+
- **`C-b [` scrolls back in `oats session attach`, and the viewer says how**
|
|
189
|
+
([#1502](https://github.com/awebai/oats/issues/1502)). Not every keyboard
|
|
190
|
+
has a PageUp.
|
|
191
|
+
- **`C-b [`:** a viewer without `--detach-key` takes it to enter tmux's copy
|
|
192
|
+
mode, as a tmux prefix does. It is the viewer's only `C-b` key. Any other
|
|
193
|
+
key after `C-b` still reaches the agent as both keys (`C-b n`, `C-b c`,
|
|
194
|
+
`C-b C-b`), with no timeout. It needs tmux 3.3 or later, whose
|
|
195
|
+
`send-keys` can re-send the key after `C-b`. On an older tmux the viewer
|
|
196
|
+
has no `C-b` key, and the line names `Shift+PgUp` alone.
|
|
197
|
+
- **The status line:** the viewer shows one line under the agent,
|
|
198
|
+
`Shift+PgUp or C-b [ to scroll back`. It takes the terminal's last row.
|
|
199
|
+
- **`--no-status-line`** (feature `session-attach-no-status-line`) drops
|
|
200
|
+
the line, for a caller that shows the keys itself.
|
|
201
|
+
- **`sessionViewer.prefixScroll`** (feature `session-attach-prefix-scroll`)
|
|
202
|
+
on `oats version --json` and `oats status --json`, relayed as each
|
|
203
|
+
roster group's `probe.sessionViewer`: whether that host's viewer takes
|
|
204
|
+
`C-b [`, read from its tmux afresh at each run. A caller that names the
|
|
205
|
+
keys itself reads it per host.
|
|
206
|
+
- **Copy mode is the agent pane's.** While one viewer scrolls back, every
|
|
207
|
+
other viewer of that agent shows the same frozen screen.
|
|
208
|
+
- **Each viewer reads key tables of its own**, named as its session and
|
|
209
|
+
removed with it. Before, every viewer without a detach key shared one
|
|
210
|
+
server-wide table, `oatsview-locked`.
|
|
211
|
+
- **A tmux limitation:** a `C-b` pressed in the instant another viewer of
|
|
212
|
+
the same tmux server opens is dropped. tmux resets every client's key
|
|
213
|
+
table when any session sets `key-table`. The key after it reaches the
|
|
214
|
+
agent alone.
|
|
215
|
+
- See
|
|
216
|
+
[Inspect, input and attach](../execution-targets.md#inspect-input-and-attach).
|
|
217
|
+
|
|
218
|
+
- This machine's launch configurations: `~/.oats/config/launch-configs.yaml`
|
|
219
|
+
(`$OATS_HOME_DIR/config/` when set) declares launch configurations for every
|
|
220
|
+
deployment on the machine. Its one key is `launch-configs:`, with the
|
|
221
|
+
deployment block's schema. A deployment entry with the same name replaces
|
|
222
|
+
the machine's whole and lists the machine file in `shadows`. A deployment's
|
|
223
|
+
`default: true` for a harness wins over the machine's. With no deployment
|
|
224
|
+
in reach, the machine's entries alone apply. A relative `executable`
|
|
225
|
+
resolves against `~/.oats/config`. The person or an installer writes the
|
|
226
|
+
file; `oats launch-config set|remove` keep writing only `oats-local.yaml`.
|
|
227
|
+
See [This machine's launch configurations](../configuration.md#this-machines-launch-configurations).
|
|
228
|
+
- `oats launch-config list` gives each entry's `source` and `shadows`, and its
|
|
229
|
+
text marks machine entries. `oats spawn --preview` names the machine file
|
|
230
|
+
when it supplied the configuration, and its `--json` carries
|
|
231
|
+
`launchConfigSource`, the declaring file's directory. `oats doctor` reports the file, and its
|
|
232
|
+
`--json` carries `machineLaunchConfigs` (`file`, `present`, `ok`,
|
|
233
|
+
`configurations`, `error`).
|
|
234
|
+
|
|
235
|
+
- **Desktop: a font picker.** **Font…** in the settings cog's menu, or
|
|
236
|
+
**Font: choose…** in the command palette (no default shortcut; give it one
|
|
237
|
+
in the shortcuts editor), chooses between
|
|
238
|
+
**Omarchy's font** and **System font**. The choice is kept on this computer
|
|
239
|
+
and applies at once: open terminals switch font in place and keep their
|
|
240
|
+
session. The terminal font size and zoom are unchanged (**Theme and
|
|
241
|
+
font**, under [Workspaces](../desktop.md#workspaces)).
|
|
242
|
+
|
|
243
|
+
- **`oats.machine` is an official package** at `v0.1.1`
|
|
244
|
+
([official catalog](../official-catalog.md)). It brings a steward for one
|
|
245
|
+
computer, `oats.machine/steward`, and a technician that handles one machine
|
|
246
|
+
problem at a time; both use the `oats.machine-care` capability, which the
|
|
247
|
+
catalog aliases to the package. A workspace pins it as
|
|
248
|
+
`oats.machine: v0.1.1`.
|
|
249
|
+
|
|
250
|
+
## Changed
|
|
251
|
+
|
|
252
|
+
- On Linux, on the unified cgroup hierarchy, each agent's harness now runs in
|
|
253
|
+
a systemd scope of its own, `oats-agent-<hash>-<random>.scope` in the user
|
|
254
|
+
manager's default slice, at every launch (spawn, session start and restart),
|
|
255
|
+
wherever the OATS tmux server runs: systemd-oomd, which ends a whole cgroup
|
|
256
|
+
under memory pressure, then ends one agent, never the server or the others.
|
|
257
|
+
The harness stays the pane's direct child, and nothing recorded changes (the
|
|
258
|
+
home's command, its launch recipe, receipts, `oats spawn --preview`). Before
|
|
259
|
+
it wraps, a launch checks with `systemd-run --user --scope -- true`, in the
|
|
260
|
+
pane's environment, that a scope can be created. When it cannot on a host
|
|
261
|
+
booted with systemd (no user manager or bus in the pane, no `systemd-run`),
|
|
262
|
+
the harness is launched as before, and the answer's `warnings` say `the agent's harness runs in the OATS
|
|
263
|
+
tmux server's cgroup, not a scope of its own: <reason>`, as does a
|
|
264
|
+
`launch-warning` event once the launch is recorded (a start refused after
|
|
265
|
+
the check leaves none). Off Linux, on a Linux host not booted with systemd
|
|
266
|
+
(a container's own init, another init system) and on the old or hybrid
|
|
267
|
+
hierarchy, nothing changes and nothing is said (#1559).
|
|
268
|
+
- The OATS tmux server is now started in a scope of its own from inside any
|
|
269
|
+
unit of the user manager, not only a service: a client launched from a
|
|
270
|
+
desktop session (OATS Desktop or `oats tui` in its `app-*.scope`), a
|
|
271
|
+
terminal's scope, an agent's scope. A login session's `session-N.scope` (an
|
|
272
|
+
ssh or tty login) starts the server as before. A server that already runs
|
|
273
|
+
keeps the cgroup it has (#1559). See
|
|
274
|
+
[Where sessions run](../execution-targets.md#where-sessions-run-the-servers-scope-and-each-agents).
|
|
275
|
+
|
|
276
|
+
- **Desktop: the command that opens Workspace › Souls is named for what it
|
|
277
|
+
does.** ⌘N / Ctrl+Shift+N, in the command palette and the shortcuts editor,
|
|
278
|
+
now reads **Souls: choose a soul to spawn (Workspace › Souls)**, where it
|
|
279
|
+
read **Spawn instance: choose a soul in Workspace**, so it is not mistaken
|
|
280
|
+
for the sidebar's **Spawn instance** button, which opens Quick Open. Its
|
|
281
|
+
keys are unchanged
|
|
282
|
+
([Keyboard shortcuts](../../packages/desktop/docs/desktop-keyboard.md)).
|
|
283
|
+
|
|
284
|
+
- **Desktop: on Omarchy, the font is Omarchy's by default.** On a computer
|
|
285
|
+
that runs Omarchy, the window and its terminals use the computer's
|
|
286
|
+
monospace font, the one `omarchy font set` changes (fontconfig's
|
|
287
|
+
`monospace`), and the terminal's default size is 13px (JetBrains Mono,
|
|
288
|
+
Omarchy's default face, at 13px; 15px elsewhere). In Omarchy's font,
|
|
289
|
+
instance names in the sidebar and on tabs are 12px, a little smaller than
|
|
290
|
+
System font's 12.5px. Everywhere else the
|
|
291
|
+
default stays **System font**, the previous faces; **Omarchy's font** can be
|
|
292
|
+
chosen there too (on Windows the system's monospace font; on macOS JetBrains
|
|
293
|
+
Mono, with the terminals keeping Inconsolata). A chosen font or terminal size
|
|
294
|
+
always wins. A terminal font family set with **Terminal: set font
|
|
295
|
+
family…** still wins. Where the computer's monospace font does not draw at
|
|
296
|
+
a fixed width (none installed, or a bitmap font), Omarchy's font shows
|
|
297
|
+
System font's faces and its row in the picker says so.
|
|
298
|
+
|
|
299
|
+
- **Desktop: the sidebar's rows sit closer to the window's left edge.** The
|
|
300
|
+
workspace and instance rows, their headings and the foot's OATS mark all
|
|
301
|
+
move 4px left and stay in one column. The right side is unchanged.
|
|
302
|
+
|
|
303
|
+
- **Desktop: switching back to a workspace puts the keyboard back where you
|
|
304
|
+
left it.** The terminal you were typing in has the focus again, ready to
|
|
305
|
+
type, with no click; so does the instance's row in the list of instances
|
|
306
|
+
(once the list is read), the tab, or the page's control. Picking a workspace
|
|
307
|
+
from the list of workspaces, a move (Alt+↑ / ↓, ⌥↑ / ↓, the leader's
|
|
308
|
+
↑ / ↓) and a window falling back after another window took its workspace
|
|
309
|
+
over all do it. A move made from the list of workspaces leaves the focus on
|
|
310
|
+
the list. A tab that closed meanwhile gives the focus to the tab active in
|
|
311
|
+
its place, or to the page
|
|
312
|
+
([desktop.md](../desktop.md#one-window-per-workspace)).
|
|
313
|
+
|
|
314
|
+
- **Desktop: the workspace moves wrap around.** Alt+↓ (⌥↓ on macOS, the
|
|
315
|
+
leader's ↓) on the last workspace of the list shows the first, and Alt+↑
|
|
316
|
+
on the first shows the last, instead of saying "No workspace below this
|
|
317
|
+
one." / "No workspace above this one." With only one workspace to show,
|
|
318
|
+
nothing moves and one line says "No other workspace to show." The TUI's
|
|
319
|
+
moves wrap too
|
|
320
|
+
([desktop.md](../desktop.md#workspaces)).
|
|
321
|
+
|
|
322
|
+
- **`oats-getting-started` sets up a machine's steward with a global
|
|
323
|
+
address** when its prompt hands it a machine found-list (a found-list with
|
|
324
|
+
`machine.platform` set, as oats-omarchy's installer writes):
|
|
325
|
+
- Step 1 is the machine's own round. It asks once, with a default for
|
|
326
|
+
every value: the workspace and the namespace (both the machine's
|
|
327
|
+
hostname, normalised), the steward's name (`steward`), the deployment
|
|
328
|
+
directory, and the harness when the found-list names none. Enter
|
|
329
|
+
accepts them all. No question comes after that round: a step that would
|
|
330
|
+
ask tells the person what happened and stops.
|
|
331
|
+
- The person's harness is recorded as the steward's launch preference
|
|
332
|
+
(`oats soul launch set`), so harness trust, readiness and the spawn all
|
|
333
|
+
use it.
|
|
334
|
+
- The new workspace is the machine's own. It pins `oats.machine`, and its
|
|
335
|
+
setup soul is `oats.machine/steward`; no `<workspace>-oats-setup-expert`
|
|
336
|
+
soul is written.
|
|
337
|
+
- Once messaging is set up, the agent creates the steward's resident with
|
|
338
|
+
`oats aweb resident create <name> --team-label <label>`: first with
|
|
339
|
+
`--plan`, then for real. It runs with no `AWEB_API_KEY`, using the team
|
|
340
|
+
root that signup made. The agent then spawns the steward with the grant-seat
|
|
341
|
+
options the provider prints. The steward's address is contacts-only:
|
|
342
|
+
`<namespace>/<name>`.
|
|
343
|
+
- With oats.aweb below 1.33.0, aw below 1.36.48, a machine that joined an
|
|
344
|
+
existing workspace by invite, or a plan that does not take the root route,
|
|
345
|
+
the steward gets a local team identity instead. The person is told so in
|
|
346
|
+
one line, with no question asked. oats.aweb 1.33.0 is not released yet,
|
|
347
|
+
so for now every steward gets the local team identity.
|
|
348
|
+
|
|
349
|
+
- **TUI: the workspace moves wrap around, as the Desktop's do.** The
|
|
350
|
+
leader then `Down` (or `Alt+Down` in the Alt family and on macOS) on the
|
|
351
|
+
last workspace of the list of workspaces opens the
|
|
352
|
+
first, and the leader then `Up` on the first opens the last, instead of
|
|
353
|
+
saying "No workspace below this one." / "No workspace above this one."
|
|
354
|
+
With only one workspace to open, nothing moves and one line says "No
|
|
355
|
+
other workspace to show." The list's own `Up` and `Down`, while it has
|
|
356
|
+
the focus, still stop at its ends
|
|
357
|
+
([tui.md](../tui.md#the-list-of-workspaces)).
|
|
358
|
+
|
|
359
|
+
- **Desktop: the sidebar's instance rows have no ⋯ button.** Their actions
|
|
360
|
+
(Open in split, Open pull request…, Knowledge & capabilities…, Start…,
|
|
361
|
+
Restart with…, Stop…, Retire instance…, and Forget… for an instance its
|
|
362
|
+
server no longer lists) are in the row's context menu (right-click,
|
|
363
|
+
Shift+F10 or the ContextMenu key), as a workspace row's are. A right-click
|
|
364
|
+
never opens the row's terminal. While the menu can't be used (the roster is
|
|
365
|
+
out of date, one of its actions is still running, or the row can't be used
|
|
366
|
+
from this computer), it opens nothing and says why in the line at the
|
|
367
|
+
bottom of the main area
|
|
368
|
+
([desktop.md](../desktop.md#the-sidebars-instance-tree),
|
|
369
|
+
[desktop-keyboard.md](../../packages/desktop/docs/desktop-keyboard.md)).
|
|
370
|
+
|
|
371
|
+
- OATS Desktop runs with the desktop session's environment, not your shell's.
|
|
372
|
+
A `fromEnv` variable exported only in a shell profile is not set there, so
|
|
373
|
+
the Desktop's sweep refuses that instance (`E_LAUNCH_ENV_MISSING`) while
|
|
374
|
+
`oats tui` run from that shell resumes it. Put the variable in the session's
|
|
375
|
+
environment, or use a literal in the launch configuration (#1559).
|
|
376
|
+
- A command you typed into an instance's window is not resumed. A resume runs
|
|
377
|
+
the instance's OATS launch, with the launch configuration, model and yolo its
|
|
378
|
+
last start chose. Put a custom command in a launch configuration (#1559).
|
|
379
|
+
|
|
380
|
+
- A start opens a fresh conversation on `TASK.md`, **except a resume**, which
|
|
381
|
+
continues the instance's own last conversation. This replaces the rule that
|
|
382
|
+
every start opened a new one.
|
|
383
|
+
- The window a spawn makes now records how its harness ended, as a start's
|
|
384
|
+
window does, and a spawn records its launch's `startId`; `oats instance
|
|
385
|
+
stop` records the start it ended, and records a stop of an instance that
|
|
386
|
+
already reads `lost` too, signalling nothing. These files are
|
|
387
|
+
kernel-private: read `sessionEnd`. Stop an instance with OATS to keep it stopped: a tmux window
|
|
388
|
+
killed by hand reads as lost and comes back.
|
|
389
|
+
|
|
390
|
+
## Fixed
|
|
391
|
+
|
|
392
|
+
- **An instance whose window has another pane is no longer unreachable**
|
|
393
|
+
([#1569](https://github.com/awebai/oats/issues/1569)). A split window, such
|
|
394
|
+
as oats.setup's no-echo prompt opens beside the agent, made status read the
|
|
395
|
+
instance `unreachable` and session inspect, input, wakes, attach, stop and
|
|
396
|
+
restart refuse it. Status and every session command now act on the agent's
|
|
397
|
+
pane alone, and the other panes never make the instance unknown. See
|
|
398
|
+
[which pane is the agent's](../execution-targets.md#which-pane-is-the-agents).
|
|
399
|
+
- **The mark.** OATS sets the tmux pane option `@oats-agent`, its value the
|
|
400
|
+
instance name, on the agent's pane whenever it creates or respawns it: a
|
|
401
|
+
spawn's window, a start's or restart's new window, and a restart in place.
|
|
402
|
+
A mark that cannot be set does not fail the launch. It is internal, not an
|
|
403
|
+
interface.
|
|
404
|
+
- **The rule.** In a window with one pane, that pane is the agent's. With
|
|
405
|
+
several, the pane marked with the instance's name; when none is, the
|
|
406
|
+
unmarked pane with the lowest pane id (the window's own pane comes before
|
|
407
|
+
any split of it). A pane marked with another instance's name is never the
|
|
408
|
+
agent's. A window launched before this release has no mark and resolves
|
|
409
|
+
to its lowest pane id until the next start or restart marks it.
|
|
410
|
+
- Wakes and `oats session input` paste into the agent's pane even when the
|
|
411
|
+
person's focus is in another pane. Stop signals only the processes under
|
|
412
|
+
the agent's pane, and a restart in place respawns that pane alone. No JSON
|
|
413
|
+
field or state name changed.
|
|
414
|
+
- **The viewer** (`oats session attach`, the Desktop and the TUI) still
|
|
415
|
+
shows the whole window, every pane. In a window with more than one pane it
|
|
416
|
+
selects the agent's pane before its client attaches, so typing goes to
|
|
417
|
+
the agent. The viewer has no key or click that moves to another pane.
|
|
418
|
+
- **When OATS cannot tell** (two panes marked with the instance's name, or
|
|
419
|
+
several panes and none unmarked or marked with it), status reads
|
|
420
|
+
`running: null`, `runtimeState: "unreachable"`, and the sentence that
|
|
421
|
+
names the panes and how to close them is the `runtimeError`; session
|
|
422
|
+
inspect, input, attach, stop and restart refuse with the same sentence.
|
|
423
|
+
- **oats.setup route 2 guidance.** The no-echo prompt's instructions no
|
|
424
|
+
longer say that the window is refused while the prompt is open: wakes and
|
|
425
|
+
input keep reaching the agent, and a viewer opened after the prompt
|
|
426
|
+
selects the agent's pane, so the agent closes the prompt and opens it
|
|
427
|
+
again while the person watches. It
|
|
428
|
+
reaches deployments with the next oats.framework release.
|
|
429
|
+
- **Known limit.** A development-channel Claude started or restarted in
|
|
430
|
+
place on a split window comes back with its launch prompt `blocked`,
|
|
431
|
+
Claude waiting at its development-channel prompt: the confirmation still
|
|
432
|
+
requires the agent's window to have one pane. Close the other pane before
|
|
433
|
+
the restart
|
|
434
|
+
([the development-channel confirmation](../configuration.md#the-development-channel-confirmation)).
|
|
435
|
+
|
|
436
|
+
- **Desktop and TUI: an agent whose window has a second pane can be opened
|
|
437
|
+
again** ([#1569](https://github.com/awebai/oats/issues/1569)). A window
|
|
438
|
+
with more than one pane read as unknown, "Instance window has multiple
|
|
439
|
+
panes; check its terminal before starting", and its row had no Open. So
|
|
440
|
+
you couldn't reach the extra pane to close it. OATS's own setup splits one
|
|
441
|
+
off for a no-echo prompt, such as an API key or an invite token.
|
|
442
|
+
- Desktop's and TUI's own terminal check now read the agent's pane as
|
|
443
|
+
the kernel does. That's the pane OATS marks as the agent's; for a
|
|
444
|
+
window from before the mark, the window's own first pane. The other
|
|
445
|
+
panes never make the agent unknown. When it can't tell (two panes marked
|
|
446
|
+
as the agent's, or none that can be), the row stays unknown and says
|
|
447
|
+
which panes, with the kernel's sentence.
|
|
448
|
+
- Desktop: a terminal opened on such a window starts on the agent's pane,
|
|
449
|
+
wherever the focus was, so typing goes to the agent and never into a
|
|
450
|
+
forgotten prompt. In the TUI, and in Desktop for another machine, the
|
|
451
|
+
kernel's viewer does the same. A click doesn't select a pane in these
|
|
452
|
+
terminals, so a terminal opened after the split can't reach the prompt:
|
|
453
|
+
the agent closes the prompt and opens it again while you watch.
|
|
454
|
+
- Desktop: a pull request's **Send N threads to <instance>** pastes into
|
|
455
|
+
the agent's pane, as the kernel names it, never the window's active pane. When the kernel can't
|
|
456
|
+
tell which pane is the agent's, nothing is pasted.
|
|
457
|
+
|
|
458
|
+
- **Desktop: a waiting instance's row says why, in words.** A message of
|
|
459
|
+
only spaces showed as a blank line; it now reads as no message, so the
|
|
460
|
+
row, the tab's tooltip and the sidebar's card give the reason in words.
|
|
461
|
+
An instance waiting for a reason Desktop doesn't know read "Needs input"
|
|
462
|
+
twice in its terminal's row, its tab's tooltip and an unavailable sidebar
|
|
463
|
+
row's tooltip; they now say it once
|
|
464
|
+
([desktop.md](../desktop.md#when-an-agent-waits-on-you)).
|
|
465
|
+
|
|
466
|
+
- **A question to the person is no longer hidden by a later claim with no
|
|
467
|
+
text** ([#1502](https://github.com/awebai/oats/issues/1502)). On a new
|
|
468
|
+
user's machine the steward's yes/no question sat unseen for 2.5 days: a
|
|
469
|
+
later `oats.core` permission claim with `message: null` replaced it in
|
|
470
|
+
`waitingOnYou`.
|
|
471
|
+
- `waitingOnYou` in `oats status`, `oats session inspect` and
|
|
472
|
+
`oats instance events` now shows the newest claim that has a message. A
|
|
473
|
+
text-less claim only fills in the reason when the shown claim has none.
|
|
474
|
+
With no message anywhere, the newest claim is shown, as before.
|
|
475
|
+
- **oats.core 2.8.0: Claude Code's waiting claims carry the prompt's
|
|
476
|
+
text.** An `AskUserQuestion` gives the question's own words. A permission
|
|
477
|
+
prompt gives Claude's "Claude needs your permission to use <tool>". See
|
|
478
|
+
[the Claude Code emitter](../capabilities.md). It reaches deployments
|
|
479
|
+
with the next oats.framework release.
|
|
480
|
+
|
|
481
|
+
- **Add a machine (`oats server connect`) finds an OATS and npm that only the
|
|
482
|
+
host's login files put on PATH**
|
|
483
|
+
([#1517](https://github.com/awebai/oats/issues/1517)). Without `--path`, connect
|
|
484
|
+
asks the host user's own login shell (bash, zsh or fish, run `-l -i`) for its
|
|
485
|
+
PATH once. It records that PATH as the registration's `path`, and the oats it
|
|
486
|
+
finds there as `oatsPath`, so installing, onboarding, readiness and every later
|
|
487
|
+
`--server` command use the same environment. When the shell cannot be read,
|
|
488
|
+
the `ssh` step's detail says why and connect uses ssh's own PATH. When even the
|
|
489
|
+
login shell has no npm, the remedy names the PATH it searched and a command
|
|
490
|
+
that checks it on the host, and you act on it by checking again, in Desktop
|
|
491
|
+
too. An explicit `--path` is used as given. A machine registered before keeps
|
|
492
|
+
its `path` and `oatsPath`; connect fills in only the ones it lacks.
|
|
493
|
+
|
|
494
|
+
- **`oats instance upgrade` moves a home whose oats.framework changes**
|
|
495
|
+
([#1549](https://github.com/awebai/oats/issues/1549)). oats.core has spawn
|
|
496
|
+
and launch hooks but declared no `upgrade` hook, so the upgrade refused every
|
|
497
|
+
home that carries it ("oats.core does not declare an upgrade hook; respawn
|
|
498
|
+
the instance to change it").
|
|
499
|
+
- **oats.core 2.8.0 declares a non-required `upgrade` hook.** The hook that
|
|
500
|
+
runs is the destination's, so homes spawned on any earlier oats.core are
|
|
501
|
+
covered, including framework 1.4.1, 1.8.0 and 1.14.1 homes.
|
|
502
|
+
- `plan` adopts when the home records no oats.core metadata (no oats.core
|
|
503
|
+
ever wrote any), `{}`, or a falsy value (`null`, `false`, `0`, `""`),
|
|
504
|
+
which the kernel passes as `{}`. Any other recorded value is refused with
|
|
505
|
+
the remedy `respawn the instance: oats retire <instance>, then oats spawn`.
|
|
506
|
+
- `apply` carries the empty metadata unchanged. Neither phase reads the home
|
|
507
|
+
or calls the host. The next start's launch hook rewrites the Claude Code
|
|
508
|
+
waiting hooks for the moved module copy.
|
|
509
|
+
- **oats.core's kernel floor rises from 0.40.0 to 0.49.0.** A kernel before
|
|
510
|
+
0.49.0 drops a capability that declares a hook event it does not know, so
|
|
511
|
+
it would lose oats.core on every soul. From 0.49.0 the kernel ignores the
|
|
512
|
+
hook; only OATS 0.68.0 and later (`oats instance upgrade`) run it.
|
|
513
|
+
- Known effect: after the move, the record holds
|
|
514
|
+
`capabilityMeta["oats.core"]: {}`, as oats.aweb-tasks' adoption leaves it.
|
|
515
|
+
- See [oats.core across `oats instance upgrade`](../capabilities.md#oatscore-needs-input).
|
|
516
|
+
It reaches deployments with the next oats.framework release.
|
|
517
|
+
|
|
518
|
+
- **An instance with a file over 2 GiB in its work retires**
|
|
519
|
+
([#617](https://github.com/awebai/oats/issues/617)). The retire refused such
|
|
520
|
+
an instance before doing anything, with `E_WORK_INSPECTION_FAILED: ... File
|
|
521
|
+
size (...) is greater than 2 GiB`. A Git clone with a large pack in a
|
|
522
|
+
directory instance's `work/` was enough. The retire's digests now read
|
|
523
|
+
every file in pieces of 1 MiB, and so does the verification of its recovery
|
|
524
|
+
copy. A home's `instance.json`, whose stored digest leaves out the fields
|
|
525
|
+
the kernel writes after spawn, is read whole up to 16 MiB; a larger one is
|
|
526
|
+
hashed as it is. Otherwise a digest is the same as before for the same
|
|
527
|
+
bytes, so spawn baselines stay valid. A file that shrinks while the retire
|
|
528
|
+
reads it refuses the retire with `E_WORK_INSPECTION_FAILED`, `<path>
|
|
529
|
+
changed while it was read`. Retry once nothing writes to it. See
|
|
530
|
+
[souls-and-instances.md](../souls-and-instances.md).
|
|
531
|
+
|
|
532
|
+
- **Desktop on Linux: a right-click on a workspace in the sidebar's list opens
|
|
533
|
+
its menu (Open in new window) and keeps it open.** It used to close as soon
|
|
534
|
+
as the mouse button came up. Shift+F10 and the ContextMenu key were not
|
|
535
|
+
affected.
|
|
536
|
+
|
|
537
|
+
- **Desktop: Instance: name and colour… from the command palette keeps your
|
|
538
|
+
place when there is no instance to act on.** It says "No instance is focused
|
|
539
|
+
or selected." and the focus goes back where it was before the palette
|
|
540
|
+
opened; it used to be lost.
|
|
541
|
+
|
|
542
|
+
## Provider updates
|
|
543
|
+
|
|
544
|
+
- **oats.aweb 1.32.0: stopped seats upgrade in place, and joined teams mint
|
|
545
|
+
only from a member root.** The catalog and this workspace pin
|
|
546
|
+
`oats.aweb: v1.32.0`. It requires OATS >=0.68.0; the aw floor is unchanged.
|
|
547
|
+
- **`oats instance upgrade` moves a stopped seat** (local, global grant or
|
|
548
|
+
retained, including 1.17.1 metadata) to the locked package. Plan is
|
|
549
|
+
read-only. Apply keeps the identity and grant, and makes no host
|
|
550
|
+
registration or `aw` call. The next start re-registers the home and, where
|
|
551
|
+
configured, renews the grant before revoking the old one. That ends the
|
|
552
|
+
lost wakes after an upgrade or a restart: a grant seat's new grant home is
|
|
553
|
+
registered first, and a local Codex seat recorded under `delivery: channel`
|
|
554
|
+
registers `<home>/.aw` as a session home. `oats.aweb-tasks` adopts with it.
|
|
555
|
+
- **A team other than the default** is joined at spawn, and its roster read,
|
|
556
|
+
only from `roots[team]` or `root`, and only when `aw team list --json`
|
|
557
|
+
there lists the team. The cwd, `OATS_WORKSPACE` and the home's own `.aw`
|
|
558
|
+
never mint for it. A missing or non-member root refuses, naming the team,
|
|
559
|
+
the root tried and the remedy (`oats aweb setup --join <label>
|
|
560
|
+
--invite-stdin`, then a respawn). `oats aweb join --labels L --plan`
|
|
561
|
+
shows the root a join would use, without inviting.
|
|
562
|
+
- **Claude channel:** a start clears a cached aweb-channel failure, which
|
|
563
|
+
made every Claude started in the next 15 minutes skip the channel, and
|
|
564
|
+
readiness reads Claude's MCP log to prove a home registered on the channel
|
|
565
|
+
(`channel-not-connected`, with the restart remedy).
|
|
566
|
+
- **Upgrade:** every host sharing the workspace runs OATS >=0.68.0 first.
|
|
567
|
+
Then `oats sync` locks it; new spawns get it, or a stopped home moved with
|
|
568
|
+
`oats instance upgrade`.
|
|
569
|
+
|
|
570
|
+
- **oats.engineering 1.14.0: maintainers per human, and shorter instructions.**
|
|
571
|
+
The catalog and this workspace now pin 1.15.0, which carries it (below);
|
|
572
|
+
the floor is unchanged.
|
|
573
|
+
- **What a lane is.** A lane is a standing area of work: areas of the
|
|
574
|
+
repository, a kind of work, or one human's requests. Work that would be
|
|
575
|
+
finished when one goal lands is an effort in an existing lane. Any
|
|
576
|
+
maintainer opens a lane with nobody's sign-off.
|
|
577
|
+
- **One maintainer per human in each project lane.** The release lane stays
|
|
578
|
+
one maintainer for the whole project, with the final say on merges and
|
|
579
|
+
releases. A lane's maintainer owns its lane: it doesn't report to the
|
|
580
|
+
maintainer that spawned it. By default a lane merges its own PRs once the
|
|
581
|
+
release lane has signed off at that head.
|
|
582
|
+
- **The lane map shows who holds each seat and what each maintainer is
|
|
583
|
+
on.** Each maintainer keeps an entry of its own on it. Your messaging
|
|
584
|
+
layer's presence says which maintainers are running.
|
|
585
|
+
- **Experts find their maintainer.** An expert that no maintainer launched
|
|
586
|
+
finds its human's maintainer in its lane. If there is none, it spawns
|
|
587
|
+
one.
|
|
588
|
+
- **Shorter injects.** The maintainer, expert and developer injects open
|
|
589
|
+
with a short block of IMPORTANT, ALWAYS and NEVER rules, and run about
|
|
590
|
+
half as long. The skills are about a fifth shorter, and no rule was
|
|
591
|
+
dropped.
|
|
592
|
+
- **Upgrade:**
|
|
593
|
+
- `oats sync` locks it. New spawns get it, and a stopped home moves to it
|
|
594
|
+
with `oats instance upgrade`.
|
|
595
|
+
- Maintainers add their entry to the lane map at their next intake.
|
|
596
|
+
|
|
597
|
+
- **oats.engineering 1.15.0: maintainers trust what experts relay from their
|
|
598
|
+
human.** The catalog and this workspace pin `oats.engineering: v1.15.0`,
|
|
599
|
+
which carries 1.14.0 (above); the floor is unchanged.
|
|
600
|
+
- **A decision is trusted.** When an expert reports a decision it made
|
|
601
|
+
with its own human, the maintainer takes it as that human's decision
|
|
602
|
+
without asking them again. That covers an architecture, a design, a
|
|
603
|
+
scope or a product call, and the same holds for a peer maintainer
|
|
604
|
+
relaying its own human. The maintainer still gates it on soundness, the
|
|
605
|
+
roadmap, the product direction and the architecture.
|
|
606
|
+
- **A relay says "Decided with <human>: <decision>"** and where it is
|
|
607
|
+
recorded. Anything else is a proposal.
|
|
608
|
+
- **Authority is still confirmed with the human directly.** That covers a
|
|
609
|
+
merge without its second reader, a hold or freeze exception, a tag or
|
|
610
|
+
publication, a deploy, a destructive or outward-facing act, and marking
|
|
611
|
+
another maintainer `left`.
|
|
612
|
+
- **Upgrade:** `oats sync` locks it. New spawns get it, and a stopped home
|
|
613
|
+
moves to it with `oats instance upgrade`. Experts relay their human's
|
|
614
|
+
decisions in the "Decided with" form.
|
|
615
|
+
|
|
616
|
+
- **oats.framework 1.15.0: oats.core 2.8.0.** oats.core's floor
|
|
617
|
+
(`compatibility.oats`) rises from 0.40.0 to 0.49.0. The other three
|
|
618
|
+
capabilities are unchanged.
|
|
619
|
+
- **`oats instance upgrade` moves a home's oats.framework.** oats.core
|
|
620
|
+
declares a non-required `upgrade` hook
|
|
621
|
+
([#1549](https://github.com/awebai/oats/issues/1549)), so a
|
|
622
|
+
stopped home is upgraded in place instead of refused with "respawn". A
|
|
623
|
+
kernel before 0.49.0 would drop a capability that declares a hook event it
|
|
624
|
+
does not know, so oats.core now refuses such a kernel with the kernel's own
|
|
625
|
+
compatibility message. Only OATS 0.68.0 and later run the hook.
|
|
626
|
+
- **A waiting question carries its text**
|
|
627
|
+
([#1502](https://github.com/awebai/oats/issues/1502)). oats.core's Claude Code
|
|
628
|
+
emitter puts the prompt's own words on its needs-input claim: the
|
|
629
|
+
question, or Claude's "needs your permission to use <tool>". A claim with
|
|
630
|
+
text is shown over one without. `/oats-operate` says so, from oats.core
|
|
631
|
+
2.8.0.
|
|
632
|
+
- **Upgrade:** `oats sync` locks it once the workspace or the catalog pins
|
|
633
|
+
`oats.framework: v1.15.0`. New spawns get it; a stopped home moves with
|
|
634
|
+
`oats instance upgrade`, from OATS 0.68.0. Every host that shares the
|
|
635
|
+
workspace runs OATS 0.49.0 or later first.
|
|
636
|
+
|
|
637
|
+
- **oats.framework 1.15.1: oats.setup 2.9.2.** oats.setup's inject follows
|
|
638
|
+
the split-window fix (#1569). While an instance's no-echo prompt pane is
|
|
639
|
+
open, OATS 0.69.0 and later keep reaching the instance's own pane, so wakes
|
|
640
|
+
and `oats session input` still arrive, and status reads it as before. The
|
|
641
|
+
inject names that version and says what an earlier OATS does instead. The
|
|
642
|
+
floors are unchanged, and the other three capabilities are unchanged.
|
|
643
|
+
- **Upgrade:** `oats sync` locks it once the workspace or the catalog pins
|
|
644
|
+
`oats.framework: v1.15.1`. New spawns get it, and a stopped home moves with
|
|
645
|
+
`oats instance upgrade`, from OATS 0.68.0.
|
|
646
|
+
|
|
647
|
+
## Documentation
|
|
648
|
+
|
|
649
|
+
- **A test Desktop or TUI never restarts the operator's agents.** The live-check rule
|
|
650
|
+
(`electron-live-verification`, and the Desktop and kernel developer souls) now gives a test
|
|
651
|
+
Desktop or TUI a private `HOME`, `OATS_HOME_DIR` and `TMUX_TMPDIR`, and runs every command that
|
|
652
|
+
builds or reads its fixture with the same three. The session resume sweep a client runs when it
|
|
653
|
+
opens restarts the lost homes it may resume on this computer's deployments list, which lives
|
|
654
|
+
under `OATS_HOME_DIR`, and the list's first use imports `$HOME/Agents`. With all three private, a
|
|
655
|
+
test client's sweep sees only its own fixture: a home of the operator's that was lost is never
|
|
656
|
+
restarted on the test's private tmux server, out of the operator's sight, and the fixture never
|
|
657
|
+
joins the operator's list. The rule also opens only the fixture deployment, names the CLI by
|
|
658
|
+
absolute path and never runs the sweep, `--plan` included, under the operator's environment.
|