Settings referenceRéférence des réglages
Every option of the settings page, with its key, kind, choices and help - generated from the code.Chaque option de la page de réglages, avec sa clé, son type, ses choix et son aide - générée depuis le code.
Every option of the settings page (F9, the palette Settings, the web page ⚙) with its key in config.toml, generated from the settings tables of the sources. The default values and the comments of the file are in the configuration reference. Settings applied only at the next start are marked ↻.
Chaque option de la page de réglages (F9, la palette Settings, le ⚙ de la page web) avec sa clé dans config.toml, générée depuis les tables de réglages du code. Les valeurs par défaut et les commentaires du fichier sont dans la référence de configuration. Les réglages pris en compte seulement au prochain démarrage sont marqués ↻.
Les libellés et aides viennent du code, en anglais.
1. InterfaceInterface
| Setting | Type | What it does |
|---|---|---|
Front-endui.frontend ↻ | choice: auto | tty | gui | auto: terminal when in a terminal, else a window (--tty / --gui win); also what the wake phrase opens when claude-cluster is closed (auto: a window, a terminal on its failure) |
Terminal emulatorui.terminal | text | the terminal opened when one is needed (the wake phrase with Front-end tty, or a window that failed), with its option to run a command (kitty, alacritty -e...); empty: the first installed |
Matching of the listsui.matching | choice: c2dmp | raw | the completions (/command, @file, !command, :command, models...), the palette and the search of the settings: c2dmp (c2dmp-hsm of libutils: typos and misplaced letters found, the closest first) or raw (what starts with the text typed, then what contains it) |
!command given to the agentui.shell_context | choice: ask | always | never | a !command typed in the prompt: its command and its output given to the agent of the session with your next prompt (like Claude Code); ask: chosen after it ran (this time, always, not this time, never: always / never are written here, ask again from here) |
Several terminals / windows at onceui.shared_core ↻ | bool | claude-cluster opened again: attached to the one running, every terminal and window in sync (the core of the sessions in a process of its own, until the last one closes); off: opened again is refused |
Restart a crashed coreui.restart_core ↻ | bool | the core runs under a small watcher: crashed (a fault, an abort), it is started again with its sessions restored - never after a normal end, never over another core, at most 5 times in 10 minutes; off: no watcher (at the next start of the core) |
Quit with sessions runningui.quit_background | choice: ask | always | never | the last terminal / window closed while a session works (or has prompts queued): • ask: a popup - keep them running in background (the core goes on without interface, its web access closed, stopped by itself once every session is idle; claude-cluster opened again meanwhile shows them) / stop everything / back • always: kept without asking • never: stopped (the turns interrupted, the sessions kept for the next start) |
Window modeui.gui_mode | choice: gui | tty | gui: the window and its own interface; tty: the window shows the terminal interface (a real terminal inside it, every key of it; -g --tty-mode for one run) |
Interfaceui.mode | choice: classic | simple | classic: every view and every setting; simple: what needs you first, a card per session in plain words, the same prompt, the useful settings only (--simple / --ipad-kid for one run) |
Language of the simple modeui.language | choice: en | fr | the words of the simple interface (English, French) |
Layoutui.layout | choice: list | grid | graph | hub | tree | tabs | the view at the start, the same five everywhere (F2 goes through them): list (the sessions and the active one), grid (every session in a cell), graph (the global session in the middle, its categories / tags, then their sessions), hub (every session around the global one), tree (the sessions by project folder); tabs: the window only |
Styleui.style | choice | colors of the interface (own ones: [styles.<name>] in the file) |
Prompts coloured by originui.prompt_origin | bool | the prompts of a transcript by where they come from: you, mcp (the global session, replies, answers), task (tasks and schedules); with a label in the header. Off: all prompts as before |
Color of your promptsui.prompt_color_user | text | #rrggbb, empty: the accent of the style (as before) |
Color of the mcp promptsui.prompt_color_mcp | text | #rrggbb, empty: violet (#8a3fb5 light, #bb8cf2 dark) |
Color of the task promptsui.prompt_color_task | text | #rrggbb, empty: magenta (#b3267a light, #f08ccb dark); tasks and schedules |
Panels of a new sessionui.panels | list | skills, tokens, context, cost, git, diff, tools, files |
Previous sessions at startui.restore ↻ | choice: ask | always | never | ask at every start (default, like Firefox: r restore, n new, q quit), always restore, never (to the trash) |
Vim mode of the promptui.vim | bool | Esc: Normal mode (hjkl w b e 0 $ x dd dw cw yy p u v V...), i / a / o: Insert |
Group the session list by categoryui.group_by_category | bool | the sessions under the category the global session or you gave them (terminal, window, web page; the palette: Group by category) |
Paste collapsed above (lines)ui.paste_lines | int | a longer paste is one [Pasted text #N +L lines] in the prompt (0: never) |
Stash: prompts keptui.stash_max | int | the prompts put aside (Ctrl+S, the list: Ctrl+L); the prompts copied by themselves (a prompt replaced by another one) never are refused |
Stash: when it is fullui.stash_full | choice: refuse | drop_oldest | refuse: the prompt stays in the editor, nothing is lost; drop_oldest: the oldest stashed prompt makes room (with a notice) |
Stash: the prompts a rewind dropsui.stash_rewound | bool | Esc (the chooser of the points of the conversation): the prompts it removes are put in the stash too |
Paste collapsed above (characters)ui.paste_chars | int | a longer paste is one [Pasted text #N] in the prompt (0: never) |
Editor of the prompt (Ctrl+E)ui.editor | text | empty: nvim, vim, vi, $VISUAL, $EDITOR, emacs -nw, nano (the first installed) |
Image viewer (right click on a chip)ui.image_viewer | text | empty: xdg-open (the viewer of the desktop) |
Image preview (terminal)ui.image_preview | choice: blocks | chafa | blocks: the pixels as half blocks; chafa: its symbols |
Answers typed as they comeui.typewriter | bool | the text appears a few characters at a time, also when the agent sends it at once |
Typing speed (characters / s)ui.typewriter_speed | int | faster by itself when a long text is behind (never more than ~1.5 s late) |
Desktop notificationsui.notify | bool | all of them (off: none); their face, their buttons (Allow / Deny, the options of a question, Open), one per session |
Start menuui.start_menu | bool | at each start of the terminal: open the claude-cluster that runs, an independent instance (its own sessions and global session), restore the previous run, start new, or quit (off: the previous run is asked only when there is one) |
Whenui.notify_when | choice: unfocused | always | unfocused: only when claude-cluster is not in front (its terminal / window not focused: its own popups are in front of you); always: also when it is |
The global session answersui.notify_global | bool | ^_^ the end of its answers (their first words) |
A session finishesui.notify_done | bool | >‿< the end of a turn of a sub-session (its tokens) |
A permission waitsui.notify_permission | bool | •_• a tool waits for your yes (Allow / Deny from the notification); o_o: the safety check |
A question of an agentui.notify_question | bool | •_• its options as buttons when there is one question |
An error, a crashui.notify_error | bool | o_o a turn failed, a session stopped |
A limit close or reachedui.notify_limits | bool | o_o 80 % / 95 % of the 5 h / 7 d windows; -_- reached (until its reset) |
Web devicesui.notify_web | bool | •_• a device asks to connect (Allow / Allow and trust / Deny), ^_^ a device connected |
The task queue doneui.notify_tasks | bool | >‿< one summary when every task is over |
Each task doneui.notify_task_each | bool | a notification at the end of each task (task_add notify: true asks it for one) |
Remindersui.notify_reminders | bool | •_• schedule add --notify, at its time |
Something slow is readyui.notify_ready | bool | ^_^ claude-cluster started (its global session), the speech model loaded, the map of this computer written: come back when it is there |
Android appui.notify_app | bool | the same notifications published to the Android app (the kinds above apply; shown when the app is not in front, each phone chooses its own kinds in its This device page; off: none sent) |
With a soundui.notify_sound | bool | a short sound of the desktop theme, one per kind (done, question, error, web) |
Map: sync proposed after (days)ui.machine_days | int | the map of this computer older: a sync proposed at start (0: never) |
| Réglage | Type | Rôle |
|---|---|---|
Front-endui.frontend ↻ | choice: auto | tty | gui | auto: terminal when in a terminal, else a window (--tty / --gui win); also what the wake phrase opens when claude-cluster is closed (auto: a window, a terminal on its failure) |
Terminal emulatorui.terminal | text | the terminal opened when one is needed (the wake phrase with Front-end tty, or a window that failed), with its option to run a command (kitty, alacritty -e...); empty: the first installed |
Matching of the listsui.matching | choice: c2dmp | raw | the completions (/command, @file, !command, :command, models...), the palette and the search of the settings: c2dmp (c2dmp-hsm of libutils: typos and misplaced letters found, the closest first) or raw (what starts with the text typed, then what contains it) |
!command given to the agentui.shell_context | choice: ask | always | never | a !command typed in the prompt: its command and its output given to the agent of the session with your next prompt (like Claude Code); ask: chosen after it ran (this time, always, not this time, never: always / never are written here, ask again from here) |
Several terminals / windows at onceui.shared_core ↻ | bool | claude-cluster opened again: attached to the one running, every terminal and window in sync (the core of the sessions in a process of its own, until the last one closes); off: opened again is refused |
Restart a crashed coreui.restart_core ↻ | bool | the core runs under a small watcher: crashed (a fault, an abort), it is started again with its sessions restored - never after a normal end, never over another core, at most 5 times in 10 minutes; off: no watcher (at the next start of the core) |
Quit with sessions runningui.quit_background | choice: ask | always | never | the last terminal / window closed while a session works (or has prompts queued): • ask: a popup - keep them running in background (the core goes on without interface, its web access closed, stopped by itself once every session is idle; claude-cluster opened again meanwhile shows them) / stop everything / back • always: kept without asking • never: stopped (the turns interrupted, the sessions kept for the next start) |
Window modeui.gui_mode | choice: gui | tty | gui: the window and its own interface; tty: the window shows the terminal interface (a real terminal inside it, every key of it; -g --tty-mode for one run) |
Interfaceui.mode | choice: classic | simple | classic: every view and every setting; simple: what needs you first, a card per session in plain words, the same prompt, the useful settings only (--simple / --ipad-kid for one run) |
Language of the simple modeui.language | choice: en | fr | the words of the simple interface (English, French) |
Layoutui.layout | choice: list | grid | graph | hub | tree | tabs | the view at the start, the same five everywhere (F2 goes through them): list (the sessions and the active one), grid (every session in a cell), graph (the global session in the middle, its categories / tags, then their sessions), hub (every session around the global one), tree (the sessions by project folder); tabs: the window only |
Styleui.style | choice | colors of the interface (own ones: [styles.<name>] in the file) |
Prompts coloured by originui.prompt_origin | bool | the prompts of a transcript by where they come from: you, mcp (the global session, replies, answers), task (tasks and schedules); with a label in the header. Off: all prompts as before |
Color of your promptsui.prompt_color_user | text | #rrggbb, empty: the accent of the style (as before) |
Color of the mcp promptsui.prompt_color_mcp | text | #rrggbb, empty: violet (#8a3fb5 light, #bb8cf2 dark) |
Color of the task promptsui.prompt_color_task | text | #rrggbb, empty: magenta (#b3267a light, #f08ccb dark); tasks and schedules |
Panels of a new sessionui.panels | list | skills, tokens, context, cost, git, diff, tools, files |
Previous sessions at startui.restore ↻ | choice: ask | always | never | ask at every start (default, like Firefox: r restore, n new, q quit), always restore, never (to the trash) |
Vim mode of the promptui.vim | bool | Esc: Normal mode (hjkl w b e 0 $ x dd dw cw yy p u v V...), i / a / o: Insert |
Group the session list by categoryui.group_by_category | bool | the sessions under the category the global session or you gave them (terminal, window, web page; the palette: Group by category) |
Paste collapsed above (lines)ui.paste_lines | int | a longer paste is one [Pasted text #N +L lines] in the prompt (0: never) |
Stash: prompts keptui.stash_max | int | the prompts put aside (Ctrl+S, the list: Ctrl+L); the prompts copied by themselves (a prompt replaced by another one) never are refused |
Stash: when it is fullui.stash_full | choice: refuse | drop_oldest | refuse: the prompt stays in the editor, nothing is lost; drop_oldest: the oldest stashed prompt makes room (with a notice) |
Stash: the prompts a rewind dropsui.stash_rewound | bool | Esc (the chooser of the points of the conversation): the prompts it removes are put in the stash too |
Paste collapsed above (characters)ui.paste_chars | int | a longer paste is one [Pasted text #N] in the prompt (0: never) |
Editor of the prompt (Ctrl+E)ui.editor | text | empty: nvim, vim, vi, $VISUAL, $EDITOR, emacs -nw, nano (the first installed) |
Image viewer (right click on a chip)ui.image_viewer | text | empty: xdg-open (the viewer of the desktop) |
Image preview (terminal)ui.image_preview | choice: blocks | chafa | blocks: the pixels as half blocks; chafa: its symbols |
Answers typed as they comeui.typewriter | bool | the text appears a few characters at a time, also when the agent sends it at once |
Typing speed (characters / s)ui.typewriter_speed | int | faster by itself when a long text is behind (never more than ~1.5 s late) |
Desktop notificationsui.notify | bool | all of them (off: none); their face, their buttons (Allow / Deny, the options of a question, Open), one per session |
Start menuui.start_menu | bool | at each start of the terminal: open the claude-cluster that runs, an independent instance (its own sessions and global session), restore the previous run, start new, or quit (off: the previous run is asked only when there is one) |
Whenui.notify_when | choice: unfocused | always | unfocused: only when claude-cluster is not in front (its terminal / window not focused: its own popups are in front of you); always: also when it is |
The global session answersui.notify_global | bool | ^_^ the end of its answers (their first words) |
A session finishesui.notify_done | bool | >‿< the end of a turn of a sub-session (its tokens) |
A permission waitsui.notify_permission | bool | •_• a tool waits for your yes (Allow / Deny from the notification); o_o: the safety check |
A question of an agentui.notify_question | bool | •_• its options as buttons when there is one question |
An error, a crashui.notify_error | bool | o_o a turn failed, a session stopped |
A limit close or reachedui.notify_limits | bool | o_o 80 % / 95 % of the 5 h / 7 d windows; -_- reached (until its reset) |
Web devicesui.notify_web | bool | •_• a device asks to connect (Allow / Allow and trust / Deny), ^_^ a device connected |
The task queue doneui.notify_tasks | bool | >‿< one summary when every task is over |
Each task doneui.notify_task_each | bool | a notification at the end of each task (task_add notify: true asks it for one) |
Remindersui.notify_reminders | bool | •_• schedule add --notify, at its time |
Something slow is readyui.notify_ready | bool | ^_^ claude-cluster started (its global session), the speech model loaded, the map of this computer written: come back when it is there |
Android appui.notify_app | bool | the same notifications published to the Android app (the kinds above apply; shown when the app is not in front, each phone chooses its own kinds in its This device page; off: none sent) |
With a soundui.notify_sound | bool | a short sound of the desktop theme, one per kind (done, question, error, web) |
Map: sync proposed after (days)ui.machine_days | int | the map of this computer older: a sync proposed at start (0: never) |
2. AgentAgent
| Setting | Type | What it does |
|---|---|---|
Provider of every sessionagent.provider ↻ | text | the global session, every sub-session (existing, new, imported conversations): <provider>[/<model>] - claude, ollama/<model>, openai/<model>, opencode/<p>/<m>, codex; the global session needs a claude-driver one (claude, ollama, anthropic-api...) |
Permission modeagent.permission_mode | choice | mode of the new sessions (empty: the default of the agent) |
Thinking effortagent.effort | choice | of the new sessions (empty: the default of the agent; per session: /effort, the palette) |
bypassPermissions possibleagent.allow_bypass ↻ | bool | claude launched with --allow-dangerously-skip-permissions (needed to switch to bypass) |
Orchestration modeagent.orchestration | choice: compact | hybrid | ultra | how readily the global session splits work into sub-sessions (told to it in its system prompt, taken at its next start): compact - the fewest tokens, medium tasks go in ONE session (one per project), it asks you before splitting; hybrid - several sub-sessions when it pays, sub-globals rarely; ultra - a sub-session (and sub-globals) for each different task |
Workflows of the modelagent.workflow ↻ | bool | the multi-agent workflows of the model itself, when it has them (Claude Code: its Workflow tool, ultracode: dozens of agents, a lot of tokens): on - allowed and asked, it uses them on a big task; off - never (the tool taken away from it). For the sessions started from now |
Work methodagent.work_method | choice: free | list-review | one-task | plan-first | custom | how the work is organised: injected, short, in the system prompt of the sub-sessions and of the global session, which splits the requests accordingly; taken by each session at its next start |
Work method (free text)agent.work_method_text | text | your own words, added to the preset (alone with custom): short, 400 characters at most |
Detect the work methodagent.work_method_detect | bool | the global session may PROPOSE a method from how you phrase your requests (a small sample of your prompts, no model): a popup you answer Allow / Deny - never applied alone |
Detection: every N sessionsagent.work_method_every | int | sessions started by the global session between two checks (0: only when asked) |
Sub-sessions: see their tasksagent.sub_read | bool | a sub-session lists its own tasks and schedules (MCP task_list, schedule_list: the ones it made; never the others). Off: it has no such tool |
Sub-sessions: queue tasksagent.sub_tasks | bool | a sub-session queues tasks for itself or another session (MCP task_add, task_cancel, task_act on its own tasks) |
Sub-sessions: create schedulesagent.sub_schedules | bool | a sub-session schedules prompts, notifications or tasks (MCP schedule_add, schedule_edit): off by default - a repeated schedule runs by itself, long after |
Sub-sessions: tasks that start sessionsagent.sub_spawn | bool | a task or a schedule of a sub-session may start sessions of its own (each costs tokens): off by default |
Global session can pinagent.global_pin | bool | the global session may pin a message or a display block at the bottom of the view of any session (MCP session_pin, session_unpin); you can always pin by hand (right click, Ctrl+K: Pin / Unpin) |
Global session can edit the taxonomyagent.taxonomy_edit | bool | the global session may set the parent of a category / tag, rename, delete and merge them (MCP taxonomy), which moves the sessions in the graph and the grouped list; you can always edit it by hand (graph, palette, claude-cluster taxonomy) |
Sub-sessions can pinagent.sub_pin | bool | a sub-session may pin an item on its own view (a task list, a form you keep seeing); never on another session |
Sub-sessions: sessions started at mostagent.sub_spawn_cap | int | the sessions of its tasks / schedules a sub-session may have waiting or open at once (past it a new one is refused); the global session has no cap |
Extra arguments (claude)agent.extra_args ↻ | list | added to every claude session, comma separated |
| Réglage | Type | Rôle |
|---|---|---|
Provider of every sessionagent.provider ↻ | text | the global session, every sub-session (existing, new, imported conversations): <provider>[/<model>] - claude, ollama/<model>, openai/<model>, opencode/<p>/<m>, codex; the global session needs a claude-driver one (claude, ollama, anthropic-api...) |
Permission modeagent.permission_mode | choice | mode of the new sessions (empty: the default of the agent) |
Thinking effortagent.effort | choice | of the new sessions (empty: the default of the agent; per session: /effort, the palette) |
bypassPermissions possibleagent.allow_bypass ↻ | bool | claude launched with --allow-dangerously-skip-permissions (needed to switch to bypass) |
Orchestration modeagent.orchestration | choice: compact | hybrid | ultra | how readily the global session splits work into sub-sessions (told to it in its system prompt, taken at its next start): compact - the fewest tokens, medium tasks go in ONE session (one per project), it asks you before splitting; hybrid - several sub-sessions when it pays, sub-globals rarely; ultra - a sub-session (and sub-globals) for each different task |
Workflows of the modelagent.workflow ↻ | bool | the multi-agent workflows of the model itself, when it has them (Claude Code: its Workflow tool, ultracode: dozens of agents, a lot of tokens): on - allowed and asked, it uses them on a big task; off - never (the tool taken away from it). For the sessions started from now |
Work methodagent.work_method | choice: free | list-review | one-task | plan-first | custom | how the work is organised: injected, short, in the system prompt of the sub-sessions and of the global session, which splits the requests accordingly; taken by each session at its next start |
Work method (free text)agent.work_method_text | text | your own words, added to the preset (alone with custom): short, 400 characters at most |
Detect the work methodagent.work_method_detect | bool | the global session may PROPOSE a method from how you phrase your requests (a small sample of your prompts, no model): a popup you answer Allow / Deny - never applied alone |
Detection: every N sessionsagent.work_method_every | int | sessions started by the global session between two checks (0: only when asked) |
Sub-sessions: see their tasksagent.sub_read | bool | a sub-session lists its own tasks and schedules (MCP task_list, schedule_list: the ones it made; never the others). Off: it has no such tool |
Sub-sessions: queue tasksagent.sub_tasks | bool | a sub-session queues tasks for itself or another session (MCP task_add, task_cancel, task_act on its own tasks) |
Sub-sessions: create schedulesagent.sub_schedules | bool | a sub-session schedules prompts, notifications or tasks (MCP schedule_add, schedule_edit): off by default - a repeated schedule runs by itself, long after |
Sub-sessions: tasks that start sessionsagent.sub_spawn | bool | a task or a schedule of a sub-session may start sessions of its own (each costs tokens): off by default |
Global session can pinagent.global_pin | bool | the global session may pin a message or a display block at the bottom of the view of any session (MCP session_pin, session_unpin); you can always pin by hand (right click, Ctrl+K: Pin / Unpin) |
Global session can edit the taxonomyagent.taxonomy_edit | bool | the global session may set the parent of a category / tag, rename, delete and merge them (MCP taxonomy), which moves the sessions in the graph and the grouped list; you can always edit it by hand (graph, palette, claude-cluster taxonomy) |
Sub-sessions can pinagent.sub_pin | bool | a sub-session may pin an item on its own view (a task list, a form you keep seeing); never on another session |
Sub-sessions: sessions started at mostagent.sub_spawn_cap | int | the sessions of its tasks / schedules a sub-session may have waiting or open at once (past it a new one is refused); the global session has no cap |
Extra arguments (claude)agent.extra_args ↻ | list | added to every claude session, comma separated |
3. ServiceService
| Setting | Type | What it does |
|---|---|---|
Keep the computer awakeservice.keep_awake | choice: off | working | always | the computer does not go to sleep by itself (idle suspend; a locked screen never stops claude-cluster, only a suspend does): off - never; working - while a session works on a turn; always - while claude-cluster runs. The lid and the suspend menu still work (systemd-inhibit, idle only) |
| Réglage | Type | Rôle |
|---|---|---|
Keep the computer awakeservice.keep_awake | choice: off | working | always | the computer does not go to sleep by itself (idle suspend; a locked screen never stops claude-cluster, only a suspend does): off - never; working - while a session works on a turn; always - while claude-cluster runs. The lid and the suspend menu still work (systemd-inhibit, idle only) |
4. Global sessionSession globale
| Setting | Type | What it does |
|---|---|---|
Enabledglobal.enabled ↻ | bool | the session that manages the others (MCP) |
Modelglobal.model ↻ | text | empty: the default model of the provider ([agent] provider) |
Permission modeglobal.permission_mode ↻ | choice | applied at every start (bypassPermissions: it never asks) |
Thinking effortglobal.effort ↻ | choice | applied at every start (empty: its own, /effort, else the one of [agent]) |
Context at mostglobal.context ↻ | choice | its conversation compacted once it reaches this size (Claude Code: CLAUDE_CODE_AUTO_COMPACT_WINDOW): a smaller context costs less at every turn and a wake-up after its cache expired reloads less (not 1M tokens); auto: the window of the model. 100k to 1M |
Folderglobal.cwd ↻ | text | working folder of the global session |
System prompt fileglobal.prompt ↻ | text | its own system prompt, added to CLAUDE.md |
Remote Control nameglobal.remote_name | text | name of the Remote Control session |
System prompt (edit)global.edit_prompt | action | opens it in the editor ([ui] editor, else nvim, vim, vi, $VISUAL, $EDITOR...); saved: the global session restarted on the same conversation |
Map of this computer (sync)global.machine_sync | action | scans again what the global session knows of this computer (asks first): OS, user, shell aliases, tools, git repositories |
| Réglage | Type | Rôle |
|---|---|---|
Enabledglobal.enabled ↻ | bool | the session that manages the others (MCP) |
Modelglobal.model ↻ | text | empty: the default model of the provider ([agent] provider) |
Permission modeglobal.permission_mode ↻ | choice | applied at every start (bypassPermissions: it never asks) |
Thinking effortglobal.effort ↻ | choice | applied at every start (empty: its own, /effort, else the one of [agent]) |
Context at mostglobal.context ↻ | choice | its conversation compacted once it reaches this size (Claude Code: CLAUDE_CODE_AUTO_COMPACT_WINDOW): a smaller context costs less at every turn and a wake-up after its cache expired reloads less (not 1M tokens); auto: the window of the model. 100k to 1M |
Folderglobal.cwd ↻ | text | working folder of the global session |
System prompt fileglobal.prompt ↻ | text | its own system prompt, added to CLAUDE.md |
Remote Control nameglobal.remote_name | text | name of the Remote Control session |
System prompt (edit)global.edit_prompt | action | opens it in the editor ([ui] editor, else nvim, vim, vi, $VISUAL, $EDITOR...); saved: the global session restarted on the same conversation |
Map of this computer (sync)global.machine_sync | action | scans again what the global session knows of this computer (asks first): OS, user, shell aliases, tools, git repositories |
5. SessionsSessions
| Setting | Type | What it does |
|---|---|---|
Short descriptionssessions.describe | choice: model | prompt | off | a few words under the name of each session (the trash too), to find it again: model: the title Claude Code made, else written by a model (Model of the descriptions; its own process, nothing kept) after its 1st, 3rd, 9th... turn; prompt: its first prompt; off: none written (the ones given kept) |
Model of the descriptionssessions.describe_model | text | <provider>/<model> writing them, any provider (claude, codex / OpenAI, opencode, qwen, a local one: ollama/llama3.2...); empty: the provider of each session (haiku for Anthropic) |
Trash kept (days)sessions.trash_days | int | a closed session stays restorable this long |
Parallel sessionssessions.max_parallel | int | sub-sessions working at the same time: past it, the tasks (task_add, the groups of the global session) wait in the task queue and are sent as soon as a place is free |
Queued promptssessions.queue_given | choice: end | step | a prompt typed while the turn runs (Claude Code sessions): • end (default): held here until the end of the turn - editable, cancelled (Esc, ↑) or sent now (F8: the turn interrupted) until then • step: given to claude after 1.5 s (Esc edits it before), taken at its next step - between two tool calls - like the queue of Claude Code; given, it can no longer be taken back |
Prompt history (prompts)sessions.history_limit | int | the prompts kept for ↑ / Ctrl+R (history.jsonl): the oldest dropped beyond it; 0: all |
Transcript size (MB)sessions.transcript_limit_mb | int | each conversation kept here (logs/<session>.jsonl, the pictures and display blocks in it): its oldest entries dropped beyond this size (the conversation of the agent itself untouched); 0: no limit |
Transcript in memory (MB)sessions.memory_mb | int | each conversation shown: its newest entries kept in memory within this size (1 to 1024); scrolled to its top, the older ones are read again from its log, up to its very start |
Transcripts in memory, all (MB)sessions.memory_total_mb | int | every conversation in memory together: past this size, the oldest entries of each are dropped from memory (still in their log, read again on demand) |
Transcript in memory (entries)sessions.memory_entries | int | also at most this number of entries in memory per conversation; 0: off (the size only, the default) |
Context at mostsessions.context ↻ | choice | the conversation of every other session compacted once it reaches this size, like [global] context (the global session may give a sub-session its own when it starts it: smaller for a small task); auto: the window of the model. 100k to 1M |
Keep warm (auto wake-up)sessions.keep_warm | choice: off | chosen | all | background | an idle session woken up a little before its prompt cache expires, so its context is never unloaded (the next prompt fast and cheap): a tiny turn (the agent answers ok) costing a cache read. • off: never (the default) • chosen: the sessions marked (/keepwarm, the palette: Keep this session warm) • all: every session • background: only a session waiting on its background (shells / agents running: a long compilation), so its context is not unloaded then loaded again when they end The sessions kept warm show ♨ |
Keep warm: cache life (min)sessions.keep_warm_ttl | int | how long the prompt cache of the agent lives: 5 (the default of Claude Code), 60 with a 1 h cache |
Keep warm: margin (s)sessions.keep_warm_margin | int | the wake-up this long before the cache expires (16 to 600 s: checked every 10 s; 45: room for a slow start) |
Keep warm: for (hours)sessions.keep_warm_hours | int | kept warm so long after its last real use (a prompt, a take-back, /keepwarm), then left to cool, its ♨ gone (0: never stops) |
Task queuesessions.task_queue | int | tasks waiting for a free place, at most (a new one past it is refused) |
Task result sizesessions.task_result_chars | int | the answer kept in a task (its result, chained into the prompt of the next one with {{tN.result}}): its first characters, at most |
Budget per session (USD)sessions.budget_usd | float | 0: none; reached: the session refuses new prompts |
Questions on the global sessionsessions.questions_to_global | bool | a question of the agent of a sub-session (AskUserQuestion) shown on the global session, its own one greyed while it waits; answered with the keys, the mouse or the voice (a number or an option, valide / suivant, retour, autre <answer>, annule); off: on its own session only |
Global session answers questionssessions.global_answers | bool | the global session is told the questions of the sub-sessions and may choose their answer (MCP session_allow answers) when it knows what you want; you can answer first |
Bypass / auto: global decidessessions.bypass_to_global | bool | a sub-session that never asks you (bypassPermissions, auto, dontAsk) still asks for some commands (the safety check of Claude Code: a removal it cannot resolve...): the global session decides them (it asks you when unsure); off: you answer them. In the other modes such a command is asked to you (auto mode: after the check stopped it) |
Global session can forksessions.global_fork | bool | the global session may copy a sub-session of Claude Code (MCP session_fork: its conversation so far, each one goes on apart) to give a new one its context; you can always (palette, fork) |
Global session can editsessions.global_edit | bool | the global session may rewrite the conversation of a sub-session of Claude Code (MCP session_edit: texts replaced, messages deleted, its end cut; a new conversation, the old one kept) |
Archive one-shot sessionssessions.archive_one_shot | bool | a session the global session started for ONE task (session_spawn one_shot) is closed to the trash (restorable) once its answer is relayed; kept when you typed in it, look at it or it waits for you. Off: the flag is ignored |
Global session can rewrite its own system promptsessions.rewrite_global_prompt | bool | the global session may rewrite its own system prompt (MCP config_write global_prompt): what it obeys at every start, it is started again on it. Off: it asks you (Ctrl+K: edit the system prompt) |
Global session can rewrite the system prompt of the sub-sessionssessions.rewrite_sessions_prompt | bool | the global session may rewrite the system prompt of the sub-sessions (MCP config_write sessions_prompt): each one takes it at its next start. Off: it asks you (Ctrl+K: edit the system prompt) |
System prompt filesessions.prompt | text | system prompt of every sub-session (claude driver), added to CLAUDE.md; empty file: none |
System prompt (edit)sessions.edit_sessions_prompt | action | opens it in the editor ([ui] editor, else nvim, vim, vi, $VISUAL, $EDITOR...); taken by each sub-session at its next start |
| Réglage | Type | Rôle |
|---|---|---|
Short descriptionssessions.describe | choice: model | prompt | off | a few words under the name of each session (the trash too), to find it again: model: the title Claude Code made, else written by a model (Model of the descriptions; its own process, nothing kept) after its 1st, 3rd, 9th... turn; prompt: its first prompt; off: none written (the ones given kept) |
Model of the descriptionssessions.describe_model | text | <provider>/<model> writing them, any provider (claude, codex / OpenAI, opencode, qwen, a local one: ollama/llama3.2...); empty: the provider of each session (haiku for Anthropic) |
Trash kept (days)sessions.trash_days | int | a closed session stays restorable this long |
Parallel sessionssessions.max_parallel | int | sub-sessions working at the same time: past it, the tasks (task_add, the groups of the global session) wait in the task queue and are sent as soon as a place is free |
Queued promptssessions.queue_given | choice: end | step | a prompt typed while the turn runs (Claude Code sessions): • end (default): held here until the end of the turn - editable, cancelled (Esc, ↑) or sent now (F8: the turn interrupted) until then • step: given to claude after 1.5 s (Esc edits it before), taken at its next step - between two tool calls - like the queue of Claude Code; given, it can no longer be taken back |
Prompt history (prompts)sessions.history_limit | int | the prompts kept for ↑ / Ctrl+R (history.jsonl): the oldest dropped beyond it; 0: all |
Transcript size (MB)sessions.transcript_limit_mb | int | each conversation kept here (logs/<session>.jsonl, the pictures and display blocks in it): its oldest entries dropped beyond this size (the conversation of the agent itself untouched); 0: no limit |
Transcript in memory (MB)sessions.memory_mb | int | each conversation shown: its newest entries kept in memory within this size (1 to 1024); scrolled to its top, the older ones are read again from its log, up to its very start |
Transcripts in memory, all (MB)sessions.memory_total_mb | int | every conversation in memory together: past this size, the oldest entries of each are dropped from memory (still in their log, read again on demand) |
Transcript in memory (entries)sessions.memory_entries | int | also at most this number of entries in memory per conversation; 0: off (the size only, the default) |
Context at mostsessions.context ↻ | choice | the conversation of every other session compacted once it reaches this size, like [global] context (the global session may give a sub-session its own when it starts it: smaller for a small task); auto: the window of the model. 100k to 1M |
Keep warm (auto wake-up)sessions.keep_warm | choice: off | chosen | all | background | an idle session woken up a little before its prompt cache expires, so its context is never unloaded (the next prompt fast and cheap): a tiny turn (the agent answers ok) costing a cache read. • off: never (the default) • chosen: the sessions marked (/keepwarm, the palette: Keep this session warm) • all: every session • background: only a session waiting on its background (shells / agents running: a long compilation), so its context is not unloaded then loaded again when they end The sessions kept warm show ♨ |
Keep warm: cache life (min)sessions.keep_warm_ttl | int | how long the prompt cache of the agent lives: 5 (the default of Claude Code), 60 with a 1 h cache |
Keep warm: margin (s)sessions.keep_warm_margin | int | the wake-up this long before the cache expires (16 to 600 s: checked every 10 s; 45: room for a slow start) |
Keep warm: for (hours)sessions.keep_warm_hours | int | kept warm so long after its last real use (a prompt, a take-back, /keepwarm), then left to cool, its ♨ gone (0: never stops) |
Task queuesessions.task_queue | int | tasks waiting for a free place, at most (a new one past it is refused) |
Task result sizesessions.task_result_chars | int | the answer kept in a task (its result, chained into the prompt of the next one with {{tN.result}}): its first characters, at most |
Budget per session (USD)sessions.budget_usd | float | 0: none; reached: the session refuses new prompts |
Questions on the global sessionsessions.questions_to_global | bool | a question of the agent of a sub-session (AskUserQuestion) shown on the global session, its own one greyed while it waits; answered with the keys, the mouse or the voice (a number or an option, valide / suivant, retour, autre <answer>, annule); off: on its own session only |
Global session answers questionssessions.global_answers | bool | the global session is told the questions of the sub-sessions and may choose their answer (MCP session_allow answers) when it knows what you want; you can answer first |
Bypass / auto: global decidessessions.bypass_to_global | bool | a sub-session that never asks you (bypassPermissions, auto, dontAsk) still asks for some commands (the safety check of Claude Code: a removal it cannot resolve...): the global session decides them (it asks you when unsure); off: you answer them. In the other modes such a command is asked to you (auto mode: after the check stopped it) |
Global session can forksessions.global_fork | bool | the global session may copy a sub-session of Claude Code (MCP session_fork: its conversation so far, each one goes on apart) to give a new one its context; you can always (palette, fork) |
Global session can editsessions.global_edit | bool | the global session may rewrite the conversation of a sub-session of Claude Code (MCP session_edit: texts replaced, messages deleted, its end cut; a new conversation, the old one kept) |
Archive one-shot sessionssessions.archive_one_shot | bool | a session the global session started for ONE task (session_spawn one_shot) is closed to the trash (restorable) once its answer is relayed; kept when you typed in it, look at it or it waits for you. Off: the flag is ignored |
Global session can rewrite its own system promptsessions.rewrite_global_prompt | bool | the global session may rewrite its own system prompt (MCP config_write global_prompt): what it obeys at every start, it is started again on it. Off: it asks you (Ctrl+K: edit the system prompt) |
Global session can rewrite the system prompt of the sub-sessionssessions.rewrite_sessions_prompt | bool | the global session may rewrite the system prompt of the sub-sessions (MCP config_write sessions_prompt): each one takes it at its next start. Off: it asks you (Ctrl+K: edit the system prompt) |
System prompt filesessions.prompt | text | system prompt of every sub-session (claude driver), added to CLAUDE.md; empty file: none |
System prompt (edit)sessions.edit_sessions_prompt | action | opens it in the editor ([ui] editor, else nvim, vim, vi, $VISUAL, $EDITOR...); taken by each sub-session at its next start |
6. VoiceVoix
| Setting | Type | What it does |
|---|---|---|
Enabledvoice.enabled ↻ | bool | voice of the global session (voice-listen / voice-say) |
Listeningvoice.mode ↻ | choice: push | auto | wake | push-to-talk key, always, or after the wake phrase |
Microphonevoice.source | text | PulseAudio / PipeWire source of the wake phrase heard while claude-cluster is closed (Services > Service), empty: the default one |
Wake phrasevoice.wake_word | text | wake mode, variants separated by commas (ok claude, okay claude) |
Armed (seconds)voice.wake_seconds | int | the phrase alone: the next sentence within this time is the command |
Spoken when armedvoice.wake_reply | text | ex: Oui ? (empty: silent) |
Only my voicevoice.only_me ↻ | bool | voice-listen --only-me (needs voice-enroll) |
Confirmation (seconds)voice.confirm_seconds | int | the transcription is shown this long before it is sent |
Spoken answersvoice.speak | bool | short summary of the answers of the global session |
Summary lengthvoice.summary_chars | int | characters of the spoken summary |
Speech enginevoice.say_engine | choice: piper | kokoro | elevenlabs | cartesia | text to speech of the answers (←→: what each one handles); an engine failing (no key, quota, model missing): Piper |
Voices of the enginevoice.say_voices | text | its voice per language: fr=ff_siwis, en=bf_emma (voice-tts --list <engine>); another engine for a language: es=kokoro:ef_dora; a language missing: the default voice of the engine |
French voice (Piper)voice.voice_fr | text | Piper voice of the French answers (engine piper, the fallback of the others) |
English voice (Piper)voice.voice_en | text | Piper voice of the English answers (engine piper, the fallback of the others) |
Listen commandvoice.listen_command ↻ | text | speech to text (one JSON per sentence) |
Say commandvoice.say_command | text | text to speech |
Speech recognitionvoice.listen_model | choice: | parakeet | whisper | kroko-fr | groq | deepgram | speech to text (←→: what each one handles): local models, or the free-tier APIs groq / deepgram (their key: claude-cluster auth login groq); an API failing: the local model |
Language spokenvoice.listen_language | text | for whisper / groq / deepgram: fr, en...; empty: found by the model (deepgram: multi, languages mixed) |
End of a sentence (seconds)voice.end_silence | float | the silence that ends a sentence (0: 1.0); a short pause goes on the same one |
Voice words filevoice.words | text | the cancel / send words said alone while a transcription waits; empty: voice-words.txt |
Voice words (edit)voice.edit_words | action | opens it in the editor: one word per line under [cancel] / [send] |
Vocabulary filevoice.vocabulary | text | proper names written back in what is heard (+ sessions, projects, your names); empty: vocabulary.txt of the config folder |
Vocabulary (edit)voice.edit_vocabulary | action | opens it in the editor: one name per line, "name = how it is heard, ..." |
| Réglage | Type | Rôle |
|---|---|---|
Enabledvoice.enabled ↻ | bool | voice of the global session (voice-listen / voice-say) |
Listeningvoice.mode ↻ | choice: push | auto | wake | push-to-talk key, always, or after the wake phrase |
Microphonevoice.source | text | PulseAudio / PipeWire source of the wake phrase heard while claude-cluster is closed (Services > Service), empty: the default one |
Wake phrasevoice.wake_word | text | wake mode, variants separated by commas (ok claude, okay claude) |
Armed (seconds)voice.wake_seconds | int | the phrase alone: the next sentence within this time is the command |
Spoken when armedvoice.wake_reply | text | ex: Oui ? (empty: silent) |
Only my voicevoice.only_me ↻ | bool | voice-listen --only-me (needs voice-enroll) |
Confirmation (seconds)voice.confirm_seconds | int | the transcription is shown this long before it is sent |
Spoken answersvoice.speak | bool | short summary of the answers of the global session |
Summary lengthvoice.summary_chars | int | characters of the spoken summary |
Speech enginevoice.say_engine | choice: piper | kokoro | elevenlabs | cartesia | text to speech of the answers (←→: what each one handles); an engine failing (no key, quota, model missing): Piper |
Voices of the enginevoice.say_voices | text | its voice per language: fr=ff_siwis, en=bf_emma (voice-tts --list <engine>); another engine for a language: es=kokoro:ef_dora; a language missing: the default voice of the engine |
French voice (Piper)voice.voice_fr | text | Piper voice of the French answers (engine piper, the fallback of the others) |
English voice (Piper)voice.voice_en | text | Piper voice of the English answers (engine piper, the fallback of the others) |
Listen commandvoice.listen_command ↻ | text | speech to text (one JSON per sentence) |
Say commandvoice.say_command | text | text to speech |
Speech recognitionvoice.listen_model | choice: | parakeet | whisper | kroko-fr | groq | deepgram | speech to text (←→: what each one handles): local models, or the free-tier APIs groq / deepgram (their key: claude-cluster auth login groq); an API failing: the local model |
Language spokenvoice.listen_language | text | for whisper / groq / deepgram: fr, en...; empty: found by the model (deepgram: multi, languages mixed) |
End of a sentence (seconds)voice.end_silence | float | the silence that ends a sentence (0: 1.0); a short pause goes on the same one |
Voice words filevoice.words | text | the cancel / send words said alone while a transcription waits; empty: voice-words.txt |
Voice words (edit)voice.edit_words | action | opens it in the editor: one word per line under [cancel] / [send] |
Vocabulary filevoice.vocabulary | text | proper names written back in what is heard (+ sessions, projects, your names); empty: vocabulary.txt of the config folder |
Vocabulary (edit)voice.edit_vocabulary | action | opens it in the editor: one name per line, "name = how it is heard, ..." |
7. UserUtilisateur
| Setting | Type | What it does |
|---|---|---|
Nicknameuser.nickname | text | how the global session calls you |
Contact emailuser.email | text | for commits, messages the global session writes for you |
Main editoruser.editor | text | nvim, vim, code... (Ctrl+E too when [ui] editor is empty) |
Language of the answersuser.language | choice: | fr | en | empty: the one of each message |
Notesuser.notes | text | anything the global session should know about you |
| Réglage | Type | Rôle |
|---|---|---|
Nicknameuser.nickname | text | how the global session calls you |
Contact emailuser.email | text | for commits, messages the global session writes for you |
Main editoruser.editor | text | nvim, vim, code... (Ctrl+E too when [ui] editor is empty) |
Language of the answersuser.language | choice: | fr | en | empty: the one of each message |
Notesuser.notes | text | anything the global session should know about you |
8. Web accessAccès web
| Setting | Type | What it does |
|---|---|---|
Open the web access at startweb.enabled | bool | a browser or the app of a phone / another computer drives claude-cluster. • who: the devices of this network (same LAN); beyond it: Beyond the LAN • how: the address and the 6-digit code shown at the top (a click copies the code), then your yes here (a popup) • several at once; trusted: no code next time; banned: refused (devices › at the top, the palette: Web devices) • opened / closed live: the palette, Web access |
Approve each new deviceweb.approve | bool | after its code, a new device waits for your yes here. • a popup: its name, its address, its browser; allow, allow and trust, deny • off: the code is enough on the LAN (from outside the LAN: always asked) |
Portweb.port | int | the address shown: https://<this computer on the LAN>:<port> |
Port outsideweb.remote_port | int | router: the port opened on the router (0: the same as here) |
Supervisor portweb.supervisor_port ↻ | int | the web access of the supervisor (claude-cluster --supervisor, w): every claude-cluster of this computer at a glance in the page and the app, its own code and devices |
Terminal page portweb.tty_port | int | the terminal interface of claude-cluster in a browser (its own port, the same code and approval): a real terminal of this computer, its keys, its mouse; the browser only (never the app); 0: off |
Terminal page port outsideweb.tty_remote_port | int | router: the port opened outside for the terminal page (0: the same) |
Found by the appweb.discovery | bool | the app searching this network finds this computer by itself (its setup page: Search this network; its address changed: found again). • while the web access is open: an answer on UDP 8789 to the devices of this network only • told: its name, its port, the fingerprint of its authority - never its code (still asked, then your yes) • off: the address typed in the app |
Beyond the LANweb.remote | choice: off | tailscale | router | the page reached away from home (the app goes there by itself when the LAN is not reached), encrypted end to end. tailscale (recommended: nothing open to the Internet) 1. install Tailscale here and on the phone, log in both with the same account 2. choose tailscale here: the address shown at the top becomes its tailnet one router (open to the Internet) 1. choose router: the port is opened on the router by UPnP (upnpc), else forward it yourself (TCP, Port outside → Port) 2. a changing public address: a dynamic DNS name in Name from outside Always from outside: each new device approved here, 5 wrong codes block its address 15 min, the bans apply |
Name from outsideweb.remote_host | text | router: a dynamic DNS name pointing at your router (duckdns, no-ip...). • empty: its public address • put in the certificate (the browser checks it) |
Shell from the pageweb.shell | bool | !command (and :r !command) typed in the web page or the app run on this computer. • in the folder of the session, like in the terminal • off (default): refused there - on, every device connected gets a shell of this computer, from outside the LAN too |
Settings from the pageweb.settings | list | the sections of the settings a device of the web page (the app) may change; empty (default): none, the settings read only there (This device stays its own). • sections: ui, notify, agent, commands, global, sessions, voice, user, service, web, mcp, services (the actions of the Services tab: start / stop, open / close, restart...) • all: every one (web itself: never this option from the page) |
HTTPSweb.https | bool | TLS with a certificate of its own, made once. • the conversation and the code encrypted, the microphone of the page allowed • the browser asks to accept it the first time (or: the palette, Web access: trust its certificate) • off: plain http, no microphone |
| Réglage | Type | Rôle |
|---|---|---|
Open the web access at startweb.enabled | bool | a browser or the app of a phone / another computer drives claude-cluster. • who: the devices of this network (same LAN); beyond it: Beyond the LAN • how: the address and the 6-digit code shown at the top (a click copies the code), then your yes here (a popup) • several at once; trusted: no code next time; banned: refused (devices › at the top, the palette: Web devices) • opened / closed live: the palette, Web access |
Approve each new deviceweb.approve | bool | after its code, a new device waits for your yes here. • a popup: its name, its address, its browser; allow, allow and trust, deny • off: the code is enough on the LAN (from outside the LAN: always asked) |
Portweb.port | int | the address shown: https://<this computer on the LAN>:<port> |
Port outsideweb.remote_port | int | router: the port opened on the router (0: the same as here) |
Supervisor portweb.supervisor_port ↻ | int | the web access of the supervisor (claude-cluster --supervisor, w): every claude-cluster of this computer at a glance in the page and the app, its own code and devices |
Terminal page portweb.tty_port | int | the terminal interface of claude-cluster in a browser (its own port, the same code and approval): a real terminal of this computer, its keys, its mouse; the browser only (never the app); 0: off |
Terminal page port outsideweb.tty_remote_port | int | router: the port opened outside for the terminal page (0: the same) |
Found by the appweb.discovery | bool | the app searching this network finds this computer by itself (its setup page: Search this network; its address changed: found again). • while the web access is open: an answer on UDP 8789 to the devices of this network only • told: its name, its port, the fingerprint of its authority - never its code (still asked, then your yes) • off: the address typed in the app |
Beyond the LANweb.remote | choice: off | tailscale | router | the page reached away from home (the app goes there by itself when the LAN is not reached), encrypted end to end. tailscale (recommended: nothing open to the Internet) 1. install Tailscale here and on the phone, log in both with the same account 2. choose tailscale here: the address shown at the top becomes its tailnet one router (open to the Internet) 1. choose router: the port is opened on the router by UPnP (upnpc), else forward it yourself (TCP, Port outside → Port) 2. a changing public address: a dynamic DNS name in Name from outside Always from outside: each new device approved here, 5 wrong codes block its address 15 min, the bans apply |
Name from outsideweb.remote_host | text | router: a dynamic DNS name pointing at your router (duckdns, no-ip...). • empty: its public address • put in the certificate (the browser checks it) |
Shell from the pageweb.shell | bool | !command (and :r !command) typed in the web page or the app run on this computer. • in the folder of the session, like in the terminal • off (default): refused there - on, every device connected gets a shell of this computer, from outside the LAN too |
Settings from the pageweb.settings | list | the sections of the settings a device of the web page (the app) may change; empty (default): none, the settings read only there (This device stays its own). • sections: ui, notify, agent, commands, global, sessions, voice, user, service, web, mcp, services (the actions of the Services tab: start / stop, open / close, restart...) • all: every one (web itself: never this option from the page) |
HTTPSweb.https | bool | TLS with a certificate of its own, made once. • the conversation and the code encrypted, the microphone of the page allowed • the browser asks to accept it the first time (or: the palette, Web access: trust its certificate) • off: plain http, no microphone |
9. AttachmentsPièces jointes
| Setting | Type | What it does |
|---|---|---|
Attachmentsattachments.enabled | bool | files, images and photos attached to a prompt (the paperclip, a drop, a paste; the app: the camera) |
Largest file (MB)attachments.max_mb | int | a bigger file is refused |
Files per promptattachments.max_files | int | at most this many attached to one prompt |
Kept per sessionattachments.keep | int | the oldest attachments of a session go past this many (0: all kept); all of them go when the session is deleted |
Downscale images above (px)attachments.max_pixels | int | a photo larger than this side is reduced before it is stored and sent (0: as it is); needs ImageMagick (magick) |
Strip the EXIF of imagesattachments.strip_exif | bool | photos re-encoded without their metadata (GPS position, camera, date); needs ImageMagick (magick) |
Stored inattachments.dir | text | the folder of the attachments, one folder per session (empty: <data>/attachments) |
Allow link modeattachments.link | bool | a file of this computer linked by its path, nothing copied (window, terminal, a page opened on this computer); a phone or another computer always uploads |
The global session may attachattachments.global | bool | it may attach a local file to a prompt it sends (MCP session_send attachments): the project folder of the session and the home of the user only (no hidden folder of the home, nothing of claude-cluster itself) |
Files per prompt of the global sessionattachments.global_max | int | at most this many |
| Réglage | Type | Rôle |
|---|---|---|
Attachmentsattachments.enabled | bool | files, images and photos attached to a prompt (the paperclip, a drop, a paste; the app: the camera) |
Largest file (MB)attachments.max_mb | int | a bigger file is refused |
Files per promptattachments.max_files | int | at most this many attached to one prompt |
Kept per sessionattachments.keep | int | the oldest attachments of a session go past this many (0: all kept); all of them go when the session is deleted |
Downscale images above (px)attachments.max_pixels | int | a photo larger than this side is reduced before it is stored and sent (0: as it is); needs ImageMagick (magick) |
Strip the EXIF of imagesattachments.strip_exif | bool | photos re-encoded without their metadata (GPS position, camera, date); needs ImageMagick (magick) |
Stored inattachments.dir | text | the folder of the attachments, one folder per session (empty: <data>/attachments) |
Allow link modeattachments.link | bool | a file of this computer linked by its path, nothing copied (window, terminal, a page opened on this computer); a phone or another computer always uploads |
The global session may attachattachments.global | bool | it may attach a local file to a prompt it sends (MCP session_send attachments): the project folder of the session and the home of the user only (no hidden folder of the home, nothing of claude-cluster itself) |
Files per prompt of the global sessionattachments.global_max | int | at most this many |
10. MCPMCP
| Setting | Type | What it does |
|---|---|---|
Custom MCP serversmcp.servers ↻ | list | more tools for the global session (MCP servers), comma separated. • a local program: name: command args (stdio; env VAR=value command... for its variables) e.g. files: npx -y @modelcontextprotocol/server-filesystem ~/notes • a site: name: https://url (HTTP) • its tools: mcp__<name>__*, on / off one by one in this MCP page Applied at the next start of the global session |
| Réglage | Type | Rôle |
|---|---|---|
Custom MCP serversmcp.servers ↻ | list | more tools for the global session (MCP servers), comma separated. • a local program: name: command args (stdio; env VAR=value command... for its variables) e.g. files: npx -y @modelcontextprotocol/server-filesystem ~/notes • a site: name: https://url (HTTP) • its tools: mcp__<name>__*, on / off one by one in this MCP page Applied at the next start of the global session |
11. UsageUsage
| Setting | Type | What it does |
|---|---|---|
Account windowsusage.account | bool | the 5 h / 7 d windows of the Claude account logged in here (the usage endpoint of claude.ai, the token of Claude Code: ~/.claude/.credentials.json, never shown), read again once too old (Account refresh) and after each turn; off: the ones the last turn told only |
Account refresh (s)usage.account_refresh | int | the account windows read again past this age (30 to 3600 s); one read shared by every instance of this user (~/.cache/claude-cluster/account-usage.json): an independent instance or a second core reads the one already made |
Claude Code transcriptsusage.transcripts | list | the folders of the transcripts of Claude Code (~/.claude/projects): every answer's usage (the API's own numbers: input, cache written / read, output, thinking) summed by day and model, every session of this computer - in claude-cluster or not; empty: the ledger of claude-cluster's own sessions only. The prices: [usage.prices] in config.toml (model = $ per million input tokens), else the ones learned from the cost Claude Code tells, else the public table |
| Réglage | Type | Rôle |
|---|---|---|
Account windowsusage.account | bool | the 5 h / 7 d windows of the Claude account logged in here (the usage endpoint of claude.ai, the token of Claude Code: ~/.claude/.credentials.json, never shown), read again once too old (Account refresh) and after each turn; off: the ones the last turn told only |
Account refresh (s)usage.account_refresh | int | the account windows read again past this age (30 to 3600 s); one read shared by every instance of this user (~/.cache/claude-cluster/account-usage.json): an independent instance or a second core reads the one already made |
Claude Code transcriptsusage.transcripts | list | the folders of the transcripts of Claude Code (~/.claude/projects): every answer's usage (the API's own numbers: input, cache written / read, output, thinking) summed by day and model, every session of this computer - in claude-cluster or not; empty: the ledger of claude-cluster's own sessions only. The prices: [usage.prices] in config.toml (model = $ per million input tokens), else the ones learned from the cost Claude Code tells, else the public table |
12. DebugDébogage
| Setting | Type | What it does |
|---|---|---|
Show everythingdebug.showcase_all | action | every popup below in a row (the sounds apart), on this interface: the display blocks in the showcase session, its states, a progress popup, a failure popup, a notification, a permission, the questions (QCM, Q&A, chat, review) |
Display blocksdebug.showcase_displays | action | markdown, callouts, code, table, charts (bar, line, pie), progress, metrics, heatmap, image, svg, diagram, diff, json, tree, terminal, checklist, steps, timeline, kanban, buttons, form, details, compare, math, audio, files - and how this platform draws each one |
Questions popupdebug.showcase_questions | action | the questions of an agent: a single choice, a multiple choice, a free answer (Other), the chat about them, back / next, the review before Submit / Cancel - answered here, sent to nobody |
Questions with previewsdebug.showcase_previews | action | a question whose options carry a preview beside the list: markdown, a coloured diff, a table, a chart, a mock-up and a picture - answered here, sent to nobody |
Update popupdebug.showcase_upgrade | action | the popup of a newer claude-cluster installed (Update / Ignore / Quit with the arrows and Enter): a pretend version, nothing is installed or stopped |
Permission popupdebug.showcase_permission | action | a permission asked by an agent (Allow / Always / Deny): nothing is run |
Session statesdebug.showcase_states | action | the showcase session working, waiting, in error, idle: its icon, the title of the terminal / window / tab, the face |
Progress popupdebug.showcase_progress | action | a long action: its steps, its log scrolling, then done |
Failure popupdebug.showcase_failure | action | an action that failed: what went wrong, OK / Retry |
Notificationsdebug.showcase_notify | action | a desktop / phone notification and the notices (information, error) |
Sounds: all of themdebug.sound_all | action | every sound of the notifications, one after the other, on this computer (whatever With a sound says): what you hear when something happens |
Sound: a turn donedebug.sound_message | action | message: a turn done, a message |
Sound: a job donedebug.sound_complete | action | complete: a heavy job done |
Sound: for youdebug.sound_message-new-instant | action | message-new-instant: a question, a permission, a reminder |
Sound: a problemdebug.sound_dialog-warning | action | dialog-warning: an error, a dangerous permission, a job failed |
Sound: readydebug.sound_service-login | action | service-login: claude-cluster ready, a device connected |
| Réglage | Type | Rôle |
|---|---|---|
Show everythingdebug.showcase_all | action | every popup below in a row (the sounds apart), on this interface: the display blocks in the showcase session, its states, a progress popup, a failure popup, a notification, a permission, the questions (QCM, Q&A, chat, review) |
Display blocksdebug.showcase_displays | action | markdown, callouts, code, table, charts (bar, line, pie), progress, metrics, heatmap, image, svg, diagram, diff, json, tree, terminal, checklist, steps, timeline, kanban, buttons, form, details, compare, math, audio, files - and how this platform draws each one |
Questions popupdebug.showcase_questions | action | the questions of an agent: a single choice, a multiple choice, a free answer (Other), the chat about them, back / next, the review before Submit / Cancel - answered here, sent to nobody |
Questions with previewsdebug.showcase_previews | action | a question whose options carry a preview beside the list: markdown, a coloured diff, a table, a chart, a mock-up and a picture - answered here, sent to nobody |
Update popupdebug.showcase_upgrade | action | the popup of a newer claude-cluster installed (Update / Ignore / Quit with the arrows and Enter): a pretend version, nothing is installed or stopped |
Permission popupdebug.showcase_permission | action | a permission asked by an agent (Allow / Always / Deny): nothing is run |
Session statesdebug.showcase_states | action | the showcase session working, waiting, in error, idle: its icon, the title of the terminal / window / tab, the face |
Progress popupdebug.showcase_progress | action | a long action: its steps, its log scrolling, then done |
Failure popupdebug.showcase_failure | action | an action that failed: what went wrong, OK / Retry |
Notificationsdebug.showcase_notify | action | a desktop / phone notification and the notices (information, error) |
Sounds: all of themdebug.sound_all | action | every sound of the notifications, one after the other, on this computer (whatever With a sound says): what you hear when something happens |
Sound: a turn donedebug.sound_message | action | message: a turn done, a message |
Sound: a job donedebug.sound_complete | action | complete: a heavy job done |
Sound: for youdebug.sound_message-new-instant | action | message-new-instant: a question, a permission, a reminder |
Sound: a problemdebug.sound_dialog-warning | action | dialog-warning: an error, a dangerous permission, a job failed |
Sound: readydebug.sound_service-login | action | service-login: claude-cluster ready, a device connected |
13. ServicesServices
| Setting | Type | What it does |
|---|---|---|
Refreshservices.refresh | choice: live | interval | manual | live: what changes shown at once (the voice heard, the web, the sessions), the slow parts (systemd, its journal, tailscale) read every Delay; interval: everything every Delay; manual: R reads it again (and after each action) |
Delay (s)services.delay | int | seconds between two reads (interval), the slow parts kept so long (live) |
Service (systemd)services.svc_service | action | the service of your session (systemd --user): it listens for the wake phrase while claude-cluster is closed and opens it when heard. • Setup: its unit written · Enable: started now and with your session · Disable: stopped, not started any more • Start / Stop / Restart: now only · Remove: its unit and its detector removed Its detector is paused while claude-cluster runs (the microphone is the voice's of the core): what is heard now is under Voice |
Web accessservices.svc_web | action | the web page and the app of the phone: open / close it now (Open the web access at start: at the start), a new code (the devices connected stay), the router asked again (Beyond the LAN) |
Coreservices.svc_core | action | the process that runs the sessions, shared by the terminals and windows (Several terminals / windows at once). • Restart: it stops, its terminals / windows start it again, the sessions resumed (their conversations kept) • Stop all: the core and every terminal / window closed |
Processes and versionservices.svc_processes | action | what claude-cluster runs now: its version (this interface and the core), since when, the changes of the last versions; the memory of the core, of each interface and of each session with what it started (its shells, its MCP servers), the tokens of this launch. • Stop the shells of a session (or every shell): the shells it runs in the background stopped, the session goes on |
Voiceservices.svc_voice | action | the voice of the core: listening or not, what it hears now (what is being said, the transcriptions waiting), what it misses |
Updateservices.svc_update | action | the version running and the one installed on disk (./setup.sh install claude-cluster): a newer one is offered in a popup (Update / Ignore / Quit); ignored, it comes back here. • Update now: the smooth update (the turns end first, the sessions kept) • Offer it again: its popup back (no longer ignored) • Check again: the binary on disk read again (also /update) |
MCP serversservices.svc_mcp | action | the MCP servers given to the global session (MCP > Custom MCP servers): Restart the global session to load them again |
Outside (router / tailscale)services.svc_outside | action | how the web access is reached from outside (Web > Beyond the LAN): ask the router again (UPnP), remove its redirection, log this computer in to tailscale |
| Réglage | Type | Rôle |
|---|---|---|
Refreshservices.refresh | choice: live | interval | manual | live: what changes shown at once (the voice heard, the web, the sessions), the slow parts (systemd, its journal, tailscale) read every Delay; interval: everything every Delay; manual: R reads it again (and after each action) |
Delay (s)services.delay | int | seconds between two reads (interval), the slow parts kept so long (live) |
Service (systemd)services.svc_service | action | the service of your session (systemd --user): it listens for the wake phrase while claude-cluster is closed and opens it when heard. • Setup: its unit written · Enable: started now and with your session · Disable: stopped, not started any more • Start / Stop / Restart: now only · Remove: its unit and its detector removed Its detector is paused while claude-cluster runs (the microphone is the voice's of the core): what is heard now is under Voice |
Web accessservices.svc_web | action | the web page and the app of the phone: open / close it now (Open the web access at start: at the start), a new code (the devices connected stay), the router asked again (Beyond the LAN) |
Coreservices.svc_core | action | the process that runs the sessions, shared by the terminals and windows (Several terminals / windows at once). • Restart: it stops, its terminals / windows start it again, the sessions resumed (their conversations kept) • Stop all: the core and every terminal / window closed |
Processes and versionservices.svc_processes | action | what claude-cluster runs now: its version (this interface and the core), since when, the changes of the last versions; the memory of the core, of each interface and of each session with what it started (its shells, its MCP servers), the tokens of this launch. • Stop the shells of a session (or every shell): the shells it runs in the background stopped, the session goes on |
Voiceservices.svc_voice | action | the voice of the core: listening or not, what it hears now (what is being said, the transcriptions waiting), what it misses |
Updateservices.svc_update | action | the version running and the one installed on disk (./setup.sh install claude-cluster): a newer one is offered in a popup (Update / Ignore / Quit); ignored, it comes back here. • Update now: the smooth update (the turns end first, the sessions kept) • Offer it again: its popup back (no longer ignored) • Check again: the binary on disk read again (also /update) |
MCP serversservices.svc_mcp | action | the MCP servers given to the global session (MCP > Custom MCP servers): Restart the global session to load them again |
Outside (router / tailscale)services.svc_outside | action | how the web access is reached from outside (Web > Beyond the LAN): ask the router again (UPnP), remove its redirection, log this computer in to tailscale |
14. PluginsPlugins
| Setting | Type | What it does |
|---|---|---|
+ Add a sourceplugins.plugsrc_add | action | a repository (GitHub owner/repo, an https URL), a folder of this computer (~/... or /...), a .zip (a path; dropped on the terminal: its path). • a Claude Code marketplace (.claude-plugin/marketplace.json): its plugins (claude plugin install) • else its skills (the folders with a SKILL.md): linked in ~/.claude/skills Code of these sources runs as you: only sources you trust |
Sync every (hours)plugins.sync_hours | int | the sources with auto sync pulled again every N hours (and at each start of the core); 0: at the start only. A source with auto sync off is pulled only by hand (Sync in its page) |
| Réglage | Type | Rôle |
|---|---|---|
+ Add a sourceplugins.plugsrc_add | action | a repository (GitHub owner/repo, an https URL), a folder of this computer (~/... or /...), a .zip (a path; dropped on the terminal: its path). • a Claude Code marketplace (.claude-plugin/marketplace.json): its plugins (claude plugin install) • else its skills (the folders with a SKILL.md): linked in ~/.claude/skills Code of these sources runs as you: only sources you trust |
Sync every (hours)plugins.sync_hours | int | the sources with auto sync pulled again every N hours (and at each start of the core); 0: at the start only. A source with auto sync off is pulled only by hand (Sync in its page) |
15. ProfilesProfiles
| Setting | Type | What it does |
|---|---|---|
Profilesprofiles.svc_profiles | action | a profile keeps every setting of now under the name you choose; Apply puts them back on every claude-cluster of this computer (they share config.toml and read it again when it changes), Undo brings back the settings before. Archive keeps one apart, Delete puts it in the trash for days (Restore), then it goes for good |
| Réglage | Type | Rôle |
|---|---|---|
Profilesprofiles.svc_profiles | action | a profile keeps every setting of now under the name you choose; Apply puts them back on every claude-cluster of this computer (they share config.toml and read it again when it changes), Undo brings back the settings before. Archive keeps one apart, Delete puts it in the trash for days (Restore), then it goes for good |
16. Accounts and limitsComptes et limites
| Setting | Type | What it does |
|---|---|---|
Limit per day (USD)limits.daily_usd | float | what all the sessions may spend since the local midnight; 0: none. Reached: the prompts are refused (Block) and you are notified; warned at the share of Warn at |
Limit per week (USD)limits.weekly_usd | float | the last 7 days; 0: none |
Limit per month (USD)limits.monthly_usd | float | since the 1st of the local month; 0: none |
Countedlimits.scope | choice: keys | all | what the limits count. • keys: only the turns billed through an API key (openrouter, anthropic-api...): a subscription costs nothing more per turn • all: every turn, the notional cost of a subscription too (the price of its tokens) |
Block at the limitlimits.block | bool | true: a limit reached refuses the new prompts (send --force passes); false: only warned |
Warn at (%)limits.warn_percent | int | warned (a notice and a notification) at this share of a limit; 0: only when it is reached |
| Réglage | Type | Rôle |
|---|---|---|
Limit per day (USD)limits.daily_usd | float | what all the sessions may spend since the local midnight; 0: none. Reached: the prompts are refused (Block) and you are notified; warned at the share of Warn at |
Limit per week (USD)limits.weekly_usd | float | the last 7 days; 0: none |
Limit per month (USD)limits.monthly_usd | float | since the 1st of the local month; 0: none |
Countedlimits.scope | choice: keys | all | what the limits count. • keys: only the turns billed through an API key (openrouter, anthropic-api...): a subscription costs nothing more per turn • all: every turn, the notional cost of a subscription too (the price of its tokens) |
Block at the limitlimits.block | bool | true: a limit reached refuses the new prompts (send --force passes); false: only warned |
Warn at (%)limits.warn_percent | int | warned (a notice and a notification) at this share of a limit; 0: only when it is reached |
17. Accounts and limitsComptes et limites
| Setting | Type | What it does |
|---|---|---|
spent todayaccounts.spent_day | action | the spending of today (the usage ledger, the scope below) |
spent this weekaccounts.spent_week | action | the last 7 days |
spent this monthaccounts.spent_month | action | since the 1st of the month |
| Réglage | Type | Rôle |
|---|---|---|
spent todayaccounts.spent_day | action | the spending of today (the usage ledger, the scope below) |
spent this weekaccounts.spent_week | action | the last 7 days |
spent this monthaccounts.spent_month | action | since the 1st of the month |
18. Model policyPolitique des modèles
| Setting | Type | What it does |
|---|---|---|
Default model of the sub-sessionspolicy.default_model | text | the model the sub-sessions start from (opus, sonnet, haiku, an id...); empty: none. The global session chooses the model of each sub-session by the difficulty of its task, from this one |
Model: how the default appliespolicy.model_mode | choice | free: no default | advise: the default is the start, deviate for a reason | force: nothing else is allowed (refused, with the reason) |
Default thinking of the sub-sessionspolicy.default_effort | choice | the thinking effort they start from (empty: the default of the agent) |
Thinking: how the default appliespolicy.effort_mode | choice | free | advise | force, like the model |
Allowed modelspolicy.allowed_models | list | the only models the global session may choose, comma separated (a name or a part: haiku, sonnet); empty: every one |
Allowed thinking effortspolicy.allowed_efforts | list | the only efforts it may choose, comma separated; empty: every one |
Ask before a costly choicepolicy.ask_costly | bool | true: the global session asks you (a popup: Allow / Deny, on every interface and a notification) before running a session on a costly model or thinking effort (below); false: it chooses alone inside the limits above |
Costly modelspolicy.costly_models | list | models that cost as much or more than these (a name or a part: opus, fable) |
Costly thinking effortspolicy.costly_efforts | list | efforts of this level or above (xhigh, max) |
| Réglage | Type | Rôle |
|---|---|---|
Default model of the sub-sessionspolicy.default_model | text | the model the sub-sessions start from (opus, sonnet, haiku, an id...); empty: none. The global session chooses the model of each sub-session by the difficulty of its task, from this one |
Model: how the default appliespolicy.model_mode | choice | free: no default | advise: the default is the start, deviate for a reason | force: nothing else is allowed (refused, with the reason) |
Default thinking of the sub-sessionspolicy.default_effort | choice | the thinking effort they start from (empty: the default of the agent) |
Thinking: how the default appliespolicy.effort_mode | choice | free | advise | force, like the model |
Allowed modelspolicy.allowed_models | list | the only models the global session may choose, comma separated (a name or a part: haiku, sonnet); empty: every one |
Allowed thinking effortspolicy.allowed_efforts | list | the only efforts it may choose, comma separated; empty: every one |
Ask before a costly choicepolicy.ask_costly | bool | true: the global session asks you (a popup: Allow / Deny, on every interface and a notification) before running a session on a costly model or thinking effort (below); false: it chooses alone inside the limits above |
Costly modelspolicy.costly_models | list | models that cost as much or more than these (a name or a part: opus, fable) |
Costly thinking effortspolicy.costly_efforts | list | efforts of this level or above (xhigh, max) |
19. OrchestratorOrchestrateur
| Setting | Type | What it does |
|---|---|---|
Profileorchestrator.profile | choice: low | normal | max | how much of the computer the jobs take • low: 1 thread per job, nice 19, ionice idle, half the usable cores at once • normal: 2-4 threads, nice 10, one job per usable core • max: the share of cores, nice 5, 1.5 jobs per core |
When one is detectedorchestrator.detect_action | choice: off | notice | tell | notice: a notice and the jobs list; tell: its session gets a prompt too |
Orchestratororchestrator.enabled | bool | the heavy jobs of the sessions (builds, tests, containers / media / GPU) wait in a queue and take turns: cluster-run, heavy_run, jobs_add. Off: cluster-run runs its command at once |
Memory guardorchestrator.memory_guard | bool | every second, under the memory kept free, the biggest of what claude-cluster runs is killed - a command of a session (its shells, background tasks, agents), a heavy job, at worst an agent grown abnormal; never the core - before the system kills anything itself; its session is told what and why (run it again with less memory), with a notice and a notification |
Kill on out-of-memoryorchestrator.kill_on_oom | bool | under the out-of-memory threshold the newest job is killed (never by default: the jobs are only slowed) |
Sub-sessions may add jobs (MCP)orchestrator.sub_add | bool | heavy_run, jobs_add for the sub-sessions: off by default, nothing asks the permission of their mode on this path (their shell cluster-run is not concerned) |
Sub-sessions may kill their jobsorchestrator.sub_kill_own | bool | job_kill on their own jobs |
Sub-sessions see the other jobsorchestrator.sub_see_others | bool | jobs_list / job_status of the jobs of the other sessions |
Notify a killed joborchestrator.notify_killed | bool | a notification (desktop, app) when a job is killed |
Notify a finished joborchestrator.notify_done | bool | a notification (desktop, app) when a job finishes |
Heavy jobs at onceorchestrator.max_jobs | int | running at once (0: from the hardware, the profile, the live load and memory) |
Jobs per sessionorchestrator.max_per_session | int | running at once for one session (0: no limit) |
Cores left to meorchestrator.reserved_cpu | int | kept free for you and the interface |
Memory left to me (MB)orchestrator.reserved_ram_mb | int | kept free for you and the interface |
Threads per joborchestrator.threads | int | the parallelism granted to a job (-j, --parallel, CARGO_BUILD_JOBS...); 0: from the profile (low: 1) |
Aging (seconds)orchestrator.aging_seconds | int | a waiting job gains one point every so many seconds, a priority step is 10: nothing starves |
Session shown firstorchestrator.focus_boost | int | points of the session you look at: its jobs go first |
Detect heavy commands (CPU s)orchestrator.detect_cpu_seconds | int | a command outside the queue that burned this many CPU seconds is listed and told (0: off) |
Hold: memory available (MB)orchestrator.hold_mem_mb | int | below it nothing new starts |
Reduce: memory available (MB)orchestrator.reduce_mem_mb | int | below it the running jobs are slowed |
Out of memory (MB)orchestrator.oom_mem_mb | int | below it, with Kill on out-of-memory, the newest job is killed |
Hold: swap used (MB)orchestrator.hold_swap_mb | int | above it, with the memory short too, nothing new starts |
Resume after (seconds)orchestrator.resume_after | int | calm for so long before the stopped jobs resume (hysteresis) |
Memory kept free (MB)orchestrator.guard_free_mb | int | the memory guard acts below it (0: 5 % of the RAM, at least 500 MB) |
Smallest command killed (MB)orchestrator.guard_min_mb | int | a command using less is never the one the memory guard kills |
Finished jobs keptorchestrator.history | int | the list of the finished jobs |
Heavy commandsorchestrator.patterns | list | more commands that go through the queue, comma separated: "text" (class custom) or "text = class" (build, tests, container, gpu, custom); a command containing it is heavy |
Session prioritiesorchestrator.session_prio | list | "s3=5": the jobs of a session come first (higher) or last (negative) |
Session quotasorchestrator.session_quota | list | "s3=1": jobs of a session running at once |
| Réglage | Type | Rôle |
|---|---|---|
Profileorchestrator.profile | choice: low | normal | max | how much of the computer the jobs take • low: 1 thread per job, nice 19, ionice idle, half the usable cores at once • normal: 2-4 threads, nice 10, one job per usable core • max: the share of cores, nice 5, 1.5 jobs per core |
When one is detectedorchestrator.detect_action | choice: off | notice | tell | notice: a notice and the jobs list; tell: its session gets a prompt too |
Orchestratororchestrator.enabled | bool | the heavy jobs of the sessions (builds, tests, containers / media / GPU) wait in a queue and take turns: cluster-run, heavy_run, jobs_add. Off: cluster-run runs its command at once |
Memory guardorchestrator.memory_guard | bool | every second, under the memory kept free, the biggest of what claude-cluster runs is killed - a command of a session (its shells, background tasks, agents), a heavy job, at worst an agent grown abnormal; never the core - before the system kills anything itself; its session is told what and why (run it again with less memory), with a notice and a notification |
Kill on out-of-memoryorchestrator.kill_on_oom | bool | under the out-of-memory threshold the newest job is killed (never by default: the jobs are only slowed) |
Sub-sessions may add jobs (MCP)orchestrator.sub_add | bool | heavy_run, jobs_add for the sub-sessions: off by default, nothing asks the permission of their mode on this path (their shell cluster-run is not concerned) |
Sub-sessions may kill their jobsorchestrator.sub_kill_own | bool | job_kill on their own jobs |
Sub-sessions see the other jobsorchestrator.sub_see_others | bool | jobs_list / job_status of the jobs of the other sessions |
Notify a killed joborchestrator.notify_killed | bool | a notification (desktop, app) when a job is killed |
Notify a finished joborchestrator.notify_done | bool | a notification (desktop, app) when a job finishes |
Heavy jobs at onceorchestrator.max_jobs | int | running at once (0: from the hardware, the profile, the live load and memory) |
Jobs per sessionorchestrator.max_per_session | int | running at once for one session (0: no limit) |
Cores left to meorchestrator.reserved_cpu | int | kept free for you and the interface |
Memory left to me (MB)orchestrator.reserved_ram_mb | int | kept free for you and the interface |
Threads per joborchestrator.threads | int | the parallelism granted to a job (-j, --parallel, CARGO_BUILD_JOBS...); 0: from the profile (low: 1) |
Aging (seconds)orchestrator.aging_seconds | int | a waiting job gains one point every so many seconds, a priority step is 10: nothing starves |
Session shown firstorchestrator.focus_boost | int | points of the session you look at: its jobs go first |
Detect heavy commands (CPU s)orchestrator.detect_cpu_seconds | int | a command outside the queue that burned this many CPU seconds is listed and told (0: off) |
Hold: memory available (MB)orchestrator.hold_mem_mb | int | below it nothing new starts |
Reduce: memory available (MB)orchestrator.reduce_mem_mb | int | below it the running jobs are slowed |
Out of memory (MB)orchestrator.oom_mem_mb | int | below it, with Kill on out-of-memory, the newest job is killed |
Hold: swap used (MB)orchestrator.hold_swap_mb | int | above it, with the memory short too, nothing new starts |
Resume after (seconds)orchestrator.resume_after | int | calm for so long before the stopped jobs resume (hysteresis) |
Memory kept free (MB)orchestrator.guard_free_mb | int | the memory guard acts below it (0: 5 % of the RAM, at least 500 MB) |
Smallest command killed (MB)orchestrator.guard_min_mb | int | a command using less is never the one the memory guard kills |
Finished jobs keptorchestrator.history | int | the list of the finished jobs |
Heavy commandsorchestrator.patterns | list | more commands that go through the queue, comma separated: "text" (class custom) or "text = class" (build, tests, container, gpu, custom); a command containing it is heavy |
Session prioritiesorchestrator.session_prio | list | "s3=5": the jobs of a session come first (higher) or last (negative) |
Session quotasorchestrator.session_quota | list | "s3=1": jobs of a session running at once |