-
School 0.3.4 Stable
released this
2026-11-13 02:17:51 +00:00 | 0 commits to main since this releaseSchool 0.3.4: fewer requests on Qwen models, a prompt that caches
The app runs the check after a lone edit (Qwen family)
Qwen models tend to send an edit alone and spend their next request only on running the check. When that happens the
app now runs the chat's own check right after the edit and returns its result together with the edit's.- Only a command the model itself ran earlier in the chat (its latest test or build, or its latest run of the changed
script) or one the goal names. The app never invents a command, and refuses installs, network access, deletes,
moves and redirects. - The transcript shows it as an ordinary command step marked as started by the app; the model reads
"Ran by the app after your edit". - Measured on Qwen 4 (12 runs each way): 131 requests down to 108, 4,738 s down to 3,529 s, input tokens 6.0M down
to 4.1M; the model never ran the check a second time. - On for the Qwen family (the local Flash Next included), off for other models;
check_after_editin a model's
profile switches it.
A stable start of every request
What is the same for every chat now comes first in the instructions, byte for byte; what belongs to one chat (a
sub-agent's path, the list of a project's other chats) moved into the chat's first message; the tool list a chat
starts with no longer changes when an MCP server comes up late (its tools are found with tool search). Sub-agents,
project chats and chats started while servers are still starting now reuse the provider's cache and the local
engines' prefix.Downloads
- Only a command the model itself ran earlier in the chat (its latest test or build, or its latest run of the changed
-
School 0.3.3 Stable
released this
2026-11-13 00:20:09 +00:00 | 0 commits to main since this releaseSchool 0.3.3: the finish check asks only for what is owed
0.3.2 made goal runs on large-window models slower on simple tasks: after writing a report, an answer file or notes,
the first "done" was refused with "nothing ran and passed after your last change", and the model had to invent a
check. Measured on small tasks: 67 s and 5 requests with 0.3.1, 93 s and 9 with 0.3.2.- A passing run is now owed only for code or build files changed in the turn, and only when something runnable
exists (a test or build the chat already ran, one the goal names, the folder's own tests, or the changed script). - Reports, answer files, notes, documentation and data owe nothing.
- A check that finished in the background counts.
- When the check does refuse, it names the one run that is owed in one line.
- A model that changed code and ran nothing is still stopped once, as before.
Measured after the fix on the same tasks: the check refused 0 of 15 runs (7 of 15 with 0.3.2), all passed, and the
times are back at or below 0.3.1 on three of five tasks.Downloads
- A passing run is now owed only for code or build files changed in the turn, and only when something runnable
-
School 0.3.2 Stable
released this
2026-11-12 21:31:03 +00:00 | 0 commits to main since this releaseSchool 0.3.2: a harness that fits each model
Model profiles
The app now adapts to the model a chat runs on, from what the router reports for it (window, effort levels, default
effort). Small-window models work exactly as before; large-window models get more room.- More goes back to a large model: command output, file reads and search results are cut later on a 500K or 1M
model than on a 256K one. - Fewer tools up front for the Qwen family (at most 20; the rest are found with tool search), as its makers advise.
- Effort: a new chat starts at the model's own default effort (the setting "Default"); every level in the menu is
a different thing. When a model is stuck, the next step runs one level higher and drops back after progress. - Replies that are only thinking (cut before any answer or tool call) are dropped and asked for again one level
lower, instead of ending the turn. - Qwen working rules: a short prompt section for the Qwen family (batch independent calls, send an edit together
with the command that checks it, read whole files on a large window). Other models' prompts are unchanged. - A second edit format (
edit/write_file) exists and can be chosen per model inconfig.toml
([profiles."<model>"]); measured against patches it did not win, so patches stay the default. - Details and the measurements:
docs/MODEL-PROFILES.md.
Smaller things
- The remote machines are called Devices (Settings > Workforce); the old "CC hosts" label is gone.
- The v2 core's output limit leaves room for the model's thinking.
Downloads
- More goes back to a large model: command output, file reads and search results are cut later on a 500K or 1M
-
School 0.3.1 Stable
released this
2026-11-12 06:25:28 +00:00 | 0 commits to main since this releaseSchool 0.3.1: helper sessions that last, and a live view of them
A helper session is never back at zero
- The helper is a task from its first question and keeps one session per chat and agent, with its own folder.
- If the agent can no longer open that session (it was cleared, the agent was updated, the folder is gone), School
starts a new one and gives it the whole earlier exchange from its record, then the question. The record marks
that exchange as "session restored from this record". - After a restart of the app, sessions that were working show as idle and can be asked again.
See a helper work
Helpers can take minutes. While one works, the right panel shows a live card for it, fed by the agent's own stream:
- its phase (starting, connected, thinking, writing) and a clock;
- the latest text as it arrives, sliding line after line; while the agent only thinks, a running count of its
thought tokens; - a flow strip of the last half minute and the loop mark, which runs faster the more data comes in;
- "quiet for N s · still working" when nothing has arrived for a while, so a slow answer is not mistaken for a
stuck one.
Click the card to open the session: status strip, the steps it has passed, the question asked and the answer so
far, streaming, above the full record of earlier exchanges. The session view is now as tall as the window.Codex sends its answer in one piece, so its card shows phase and clock until the answer lands.
Fixes
- A goal with help showed two Goal cards in the chat.
- The session record was only a few lines tall.
Downloads
-
School 0.3.0 Stable
released this
2026-11-11 22:48:52 +00:00 | 0 commits to main since this releaseSchool 0.3.0: Workforce
Workforce (Settings > Workforce)
The page that held the remote hosts is now Workforce: the outside coding agents of this computer and of your
other machines.- Every agent shows whether it is installed, its version and whether it is signed in.
- Install runs the maker's own install command. Sign in opens the agent's own sign-in in a terminal.
- Pick a remote device to see its agents the same way and install over SSH. Sessions opens the sessions screen.
- Listed: CC, Codex, Antigravity, Devin, Copilot, Cursor, opencode, Qwen Code.
- The remote devices and the CC app link are still on this page, under Devices.
Helper sessions: ask a frontier agent without telling it the project
A chat that is stuck can ask CC, Codex or Antigravity for help (
agent_help).- A dedicated session per chat and agent, kept and shown under Tasks in the right panel. Click it to read every
question, every answer and every message that was stopped. - The helper runs at its best model and highest effort, from an empty folder, with its tools off.
- The helper never learns what the project is. Every message is checked before it is sent. A message with a
project, file or folder name, a name from the code, a real path, a host, an address, an e-mail, a secret, code
copied from the project or your own wording is not sent; the model is told why and rewrites it. - If no useful answer comes, the model asks again another way, then another agent.
- Skills:
agent-help-claude,agent-help-codex,agent-help-antigravity. - Switch: Settings > Workforce > "Chats may ask outside agents for help" (on). "Private terms" adds words that
must never be sent.
Goal with help
/goal-with-help <objective>(orfable run --goal-with-help): a goal that asks a helper when it hits a wall
instead of stopping or repeating. When a turn is stopped for repeating the same call, it starts again by asking.
Skill:goal-with-help.Tasks for outside agents
Per agent, "May be given tasks" (off until you turn it on) lets a chat hand an agent a whole task in the chat's
folder (agent_delegate). Unlike help, the agent then sees the project. It runs in the background and reports
back; it is listed under Tasks.Sessions
- Right-click a session on the External sessions screen: Fork into a new chat brings its whole conversation
into a new chat and opens it.
Fixed
- Status lines and their tests follow the short provider names.
Downloads
-
School 0.2.10 Stable
released this
2026-11-11 11:18:54 +00:00 | 0 commits to main since this releaseSchool 0.2.10
- Model names before the sign-in. When the app has no model list from the gateway yet (before the sign-in, or while the gateway does not answer), it showed the model's raw id. It now shows the short name there too: the local model is School native FN.
Downloads
-
School 0.2.9 Stable
released this
2026-11-11 10:19:43 +00:00 | 0 commits to main since this releaseSchool
The app has a new name: School. Your chats, settings and sign-in stay as they are.
- New name and logo. The window, the sign-in page and the Windows shortcut now say School. The logo is an endless loop; in the app a light keeps running round it (it stands still when Motion is off).
- Short provider names. The model menu, the status line, Usage and Billing show providers as L, RT, LTH and QC.
- Short model names. A model's name starts with its kind (c, uc, uc-a) and then its initials, for example
c · CO 5.5. The local model is School native FN. A name you gave a model yourself is shown as you wrote it. - CC. The linked coding app is called CC everywhere.
- Windows. The old shortcut is replaced by "School" on update; the install folder does not move.
The names on screen are the only change: model ids in settings and scripts are the same as before.
Downloads
-
NIG 0.2.8 Stable
released this
2026-11-10 22:53:55 +00:00 | 0 commits to main since this releaseWhat changed
A goal no longer repeats work it already did. When a chat has a goal, NIG keeps the agent going with an automatic "continue" message after each turn. That message only repeated the goal and told the agent to check every requirement, so a goal with a one-time step, such as "set up this tool first", had the setup checked and done again on every round.
The continue message now lists what is already done: the commands that ran successfully, the files the agent changed, and the plan steps it finished. The list comes from NIG's own record of the agent's actions, not from what the model says it did. The agent is told not to redo those and to go on with the first thing that is still open.
Downloads
-
NIG 0.2.7 Stable
released this
2026-11-10 16:00:53 +00:00 | 0 commits to main since this releaseWhat changed
The chat follows the model again. While a model was working, the chat often stopped scrolling with the answer and showed "Jump to latest", even though you had not touched it. This happened most when a tool ran or a block of code or a sub-agent's card appeared in the middle of an answer. Now the chat stays at the newest output until you scroll up yourself. Scrolling up stops the following and shows "Jump to latest"; scrolling back to the bottom, clicking "Jump to latest" or sending a message starts it again.
Downloads
-
NIG 0.2.6 Stable
released this
2026-11-10 05:23:08 +00:00 | 0 commits to main since this releaseWhat changed
NIG uses the whole local fleet by default. The local model (Qwen3.8 Flash Next) is served by every local engine at once: the 4090 cards on all three servers and the DGX Sparks, about 96 seats in all. A chat can now run up to 24 sub-agents at the same time, where it was 6. NIG still waits when the router says every seat is busy, so a big fan-out only takes seats that are free. If you set your own number in Settings > Agent > Max sub-agents, it is kept.
The local model has its right name. It was still listed as "Qwen3.8-27B (local)", a model that is no longer used; it now reads "Qwen3.8 Flash Next (local)", also in existing installs that kept the old name.
Cloud models stay off unless you turn them on.
Downloads