Tiers and safe mode
Minimalist, balanced and full, the features each one turns on, per-feature overrides and safe mode.
The balanced/full package_tools feature offers the local packages tool only after tool_search discovery, even when general tool discovery is off. Its mutations always require approval; full permission mode does not bypass package review.
Balanced/full also includes pre-limit compaction checkpoints, native graph orientation/impact, focused handoff, instruction includes and static package capability previews. Caveman style and PDF/GitHub fetch modes require explicit selection. Full defaults MCP discovery on unless explicitly disabled. The event-stream model-spending watcher (I28) is excluded by lead decision.
One tier setting (minimalist, balanced, full) selects a bundle of switches that already exist. It adds no new meaning to the setup presets, the chat view, density or prompt modes: a tier only chooses which of their existing values to set and which feature modules are on. Every feature stays individually switchable. With no tier set, nothing changes: every feature keeps the default it had before tiers.
Names, reconciled. minimal/balanced/full already mean three different things, and they keep those meanings:
| Where | Values | What a tier does with it |
|---|---|---|
Setup preset (/setup, setup) | default, minimal, full, custom | Nothing. A tier never writes the preset or its memory and approval choices |
Chat view (view, layout) | minimal, balanced, full | /tier sets it to the tier's view |
Prompt mode (/prompt-mode, prompt modes) | minimal, focus, guide, full, custom | /tier sets the project's mode: minimalist minimal, balanced focus, full full; guide is never set by a tier |
| Density (terminal profile) | compact, comfortable, spacious | /tier sets compact for minimalist, comfortable otherwise; minimalist also hides the right bar; balanced and full turn showEffort on |
| Footer (the row under the prompt) | follows the view | minimal has none; balanced and full place their segments (layout) |
| Feature modules | the manifest below | follow the tier live |
The bundle is written once by /tier <name> (or the Tier row of /setup), like a preset: a later change to one switch stays. Feature modules are different: they follow tier every time they are read, so a tier change needs no restart.
Commands
/tierlists the bundles, what each tier includes ([x - -]is minimalist, balanced, full) and what is on now and why./tier minimalist|balanced|fullapplies a tier and says which features it added or removed./tier <feature> on|off|defaultoverrides one feature (features.<id>in settings).defaultremoves the override./tier offclears the tier./experimentalshows the experimental flags with the tier applied, plus the active overrides./setuphas a Tier row: left/right cycles none, minimalist, balanced, full and shows what the next one adds or removes.- Local only: the daemon refuses
tierandfeaturesfrom remote terminals and peers.
Starting tier
The installers take --tier minimalist|balanced|full (releases); the choice travels in the install receipt and is
read once, by the first launch. It is a starting configuration, not a lock: it is used only while setup was never completed and no tier
is saved, so it never overwrites a tier or settings that already exist, mendcode upgrade and a reinstall do not apply it again, and /tier
changes the tier whenever you like. minimalist applies the bundle and configures itself (setup); balanced and full
open the setup page with that tier selected. With no receipt (a source run, an older install) nothing changes.
Slash commands
Minimalist lists twenty-one built-in slash commands, three fewer than pi's twenty-four (settings model tree thinking scoped-models export import share bug copy name session changelog hotkeys fork clone trust login logout new compact resume reload quit). The set is MINIMALIST_SLASH in
packages/protocol/src/tier-data.ts: new sessions model connect compact settings fork export copy rename usage tree undo redo plan permissions tier trust help shortcuts quit. tree is there because it is how a message is found, jumped to and forked at in every tier. tier and settings are in it so everything else is one step away; connect is how a provider is added.
- The slash menu (browse and search),
/helpand the command palette list that set only. Balanced and full list every built-in command whose feature is on, as before, and with no tier nothing is hidden. Your own commands, skills, extension commands and MCP prompts are never hidden by a tier. - A hidden command typed by hand runs nothing and answers in one line, the tier that has it and how to get there:
/review · balanced · /tier balancedThe line is not kept in the prompt history. Commands of a feature (/panel,/mcps,/fanout,/voice,/pet...) are absent from browsing while their feature is off; typed exactly they show one row naming the switch (/pr off · /tier balanced), and Enter on it runs nothing (commands)./panel("presets need the balanced or full view") and/sidebarstill answer for themselves./tutorialis not in the set; add it back (below) to keep the tour. - Customizing stays free. The per-command
hiddenpreference of the slash menu decides regardless of tier:hidden: falselists a command minimalist would not (and it runs),hidden: trueremoves one from browsing as it always did (still searchable and runnable). In/settings slashCommands(the Slash menu → Edit commands row, also under/customize)hflips the selected command and the row saysshown,hiddenorhidden by tier; the preference is written only where it differs from the tier, so a command flipped back follows the tier again. A feature you switch on yourself (/tier <feature> on,features.<id>: true) also brings its command into minimalist. A command whose feature is off stays out of the menu even withhidden: false.
Feature manifest
The README contains every manifest entry, generated by
bun scripts/readme-tiers.ts (stdout only). Replace its marked block with that output after
manifest changes. bun scripts/readme-tiers.ts --check and bun test scripts/readme-tiers.test.ts
reject drift in labels, tier inclusion, modes, switches or guide links.
The unverified OpenCode serve and Cursor SDK adapters and their tier switches have been withdrawn; see optional engines. Saved switches are ignored while other feature overrides are preserved.
packages/protocol/src/tier-data.ts declares every feature once: its tier, where it costs (tools and prompt on every request, chrome, or calls of its own), the existing switch that controls it, and the tools it owns. Adding a feature means adding its entry; the budget test below checks which tools each tier offers.
| Feature | Tier | Mode | Own switch |
|---|---|---|---|
extensions, packages, hooks | minimalist | kernel | only --safe-mode, MENDCODE_EXTENSIONS=0, MENDCODE_HOOKS=0, disabledPackages |
usage_resume (auto-resume at a usage limit, usage-limit) | minimalist | tier | /tier usage_resume off; armed by default in every tier |
codegraph (graph_*, code_search) | balanced | tier | MENDCODE_CODEGRAPH=0 |
subagents (task) | balanced | tier | tools.task |
script_workflows (workflow_script tool, /workflows run; script workflows) | balanced | tier | tools.workflow_script; the tool is deferred (found through tool_search), so it costs no context until used |
threads (threads tool; sessions) | balanced | tier | experimental.thread_tools (off without a tier) |
pr_watch (pr_watch tool, /pr; pr-watch) | balanced | permit | experimental.pr_watch (network through gh; never on by itself) |
pr_review (/pr review, /pr watch-repo; pr-review) | balanced | permit | experimental.pr_review (network through gh, model calls; off with no tier; posting also needs the repo's enabled: true) |
output_filters (command output filters, "tokens saved"; advanced-config) | balanced | tier | tool_output.filters (off without a tier) |
forecast (footer ~N turns to compact; layout) | balanced | tier | /footer +forecast (off without a tier) |
lean_tools (only the coding core in every request, the rest through tool_search; tool discovery) | minimalist | tier | /tier lean_tools off (off without a tier) |
aux_tools (the situational tools of the core in every request: background command output, sub-agent status, web fetch, questions, plan mode; tool discovery) | balanced, opt-in | tier | /tier aux_tools on (on without a tier; off in every tier, where each one is listed when its situation arises) |
harness_tools (MendCode's tools inside Claude Code, Codex and HTTP-MCP ACP agents: code graph, computer use, memory, settings, image generation; without thread_mcp, one line on how sub-agents on MendCode's other models are turned on) | balanced | tier | /tier harness_tools on (off without a tier and in minimalist) |
plan_tools (todo and the goal tools in every request) | full, opt-in | tier | /tier plan_tools on (on without a tier; deferred in every tier) |
find_tools (glob and grep in every request; they keep minimalist's first request above the provider's cacheable minimum, tool discovery) | minimalist | tier | /tier find_tools off (on without a tier; off, shell searches with rg and tool_search finds them) |
ui_tool (customize_ui, the user's screen settings, in every request) | balanced | tier | /tier ui_tool on (on without a tier; deferred in minimalist) |
orientation (cwd, branch, test command, loaded instruction files and a top-level tree in the first user message; tool discovery) | balanced | tier | /tier orientation off (off without a tier) |
file_viewer (a click on a path opens the file in the right slot, /open; file viewer) | balanced | tier | pathClick in /customize (editor turns the click off, panel turns it on in any tier); /tier file_viewer off |
git_hub (/git: Changes, Commit, Worktrees and PR tabs, a local commit of the files you tick, and git init in a folder that is not a repository; git) | balanced | tier | git.defaultTab, git.newSessionWorktree; /tier git_hub off (off without a tier; no model tool, no prompt text) |
git_actions (the Git page's push, publish to GitHub as a private repository and create pull request, and the PR tab's read; git) | balanced | permit | experimental.git_actions (network through your own git and gh; never on by itself; no model tool, no prompt text) |
mcp (servers, mcp__*, mcp_resource_*) | full | tier | mcp.json |
thread_mcp (thread bridge, cross-engine delegation: a Claude Code, Codex or ACP session's sub-agents on MendCode's other models, which its instructions name by engine) | full | permit | threadMcp.claude / threadMcp.codex, off by default |
jev (adaptive reasoning) | full | tier | experimental.reasoning_auto |
idle_compact (one compaction of a large idle session just before its prompt cache expires; compaction) | full | tier | /tier idle_compact on|off; idleCompactAtTokens (off without a tier) |
cache_countdown (the warm footer segment: the minutes the prompt cache of the last call has left, cold after; layout) | full | permit | /footer +warm places it; no tier places it, and outside full it draws nothing unless /tier cache_countdown on (no call of its own beyond one context.recent per turn, no tool, no prompt text) |
wait_cache (a native session on Anthropic whose calls were twice 5 to 60 minutes apart writes its cache at the 1-hour lifetime from then on; prompt cache) | full | tier | /tier wait_cache on|off; ignored while promptCacheTtl is 1h (off without a tier; no call of its own, no tool, no prompt text) |
compact_notice (one muted transcript line when the context reaches 80% of the compaction point: tokens left, turns left, /compact; compaction) | balanced | tier | /tier compact_notice on|off; compactNoticePercent (off without a tier; no call, no prompt text) |
backlog (your own list of pending things for a project, kept across sessions: /backlog, a panel card, mendcode backlog; panel) | full | tier | /tier backlog on|off (off without a tier; no model tool, no prompt text; the model reads it and never writes it) |
task_worktree (task takes isolation: "worktree": the child works in a git worktree of HEAD and its changes end on a branch of their own, never merged; agents) | full | tier | /tier task_worktree on|off (off without a tier; the argument is in the task schema only while it is on) |
task_fork (experimental; task takes fork: true: a new child starts from the parent's latest request on the same model, so its first call reads the parent's cached prefix; every call then reads the parent's context; agents) | full | opt-in | /tier task_fork on|off (no tier turns it on; the argument is in the task schema only while it is on) |
pets | full | tier | pet |
pet_generator (experimental; customize) | full | permit | explicit /pet new and confirmation; requires enabled image_generation, features.pet_generator: false blocks it |
voice, browser, code_mode, snip, worktree_tool | full | permit | voice.enabled, MENDCODE_BROWSER=1, experimental.* |
one_tree (sessions sharing one checkout: one line on a tool result that names a file another session is changing; one tree) | balanced | tier | /tier one_tree on|off (off without a tier; no tool, no prompt text, nothing while a session works alone) |
computer (computer), graph_context (code graph) | balanced | opt-in | /tier computer on, /tier graph_context on; no tier turns them on |
pi (your own pi as an engine; pi-engine) | full | permit | pi.json or /tier pi on |
storage_compact (storage) | minimalist | permit | explicit offline compaction; features.storage_compact: false refuses it |
official_packages (local catalog) | balanced | permit | explicit installation; gates all three official packages |
autoresearch / web_routing | full / balanced | permit | explicit installation; also require official_packages |
Modes: tier features are on when the tier includes them. permit features are opt-in or risky: the tier only allows them and their own switch must still be on, so no tier ever turns on computer control, the browser or dictation by itself. kernel features are not tier-gated.
Precedence, highest first: safe mode, features.<id>, the feature's own flag (for jev, an explicit experimental.reasoning_auto), the tier, today's default. A tools entry of your own beats the tier's denial of a tool.
Decisions taken: MCP belongs to full (the stated tier definition lists it there; /tier mcp on brings it to another tier). jev is on in full by default but still needs its provider configuration to do anything. Sub-agents and the code graph are in balanced as the savings tools. Tools are denied through the existing tools permission rules (see permissions), so a denied tool is out of the model's tool list and refused if called.
A tool declares a tier too: Tool.tier for built-in and project tools, and tier on an extension's registerTool (see extensions). A tool without a tier is in every tier; with a tier set, one that names a higher tier is not offered.
Safe mode
Peers cannot change execution switches through experimental, tools, disabled extension/package lists or package/MCP
installation endpoints. Safe mode remains a host startup choice, checked before stored feature overrides. Local overrides
outside safe mode keep their documented meaning.
mendcode --safe-mode (or MENDCODE_SAFE_MODE=1) starts with no extensions (daemon or terminal), no installed packages, no shell hooks, no MCP and no project custom tools, whatever the settings say. It builds on --pure: the daemon must have been started in safe mode, so use a separate --dir or --isolated when another daemon is running. Skills and commands in your own folders still load (they are prompt data, not code). /doctor shows a Mode section, and daemon.info reports safeMode. Settings are untouched, so restarting without the flag restores everything.
Baseline budget
packages/daemon/test/tier.test.ts builds the first request of a fresh session for each tier through the real engine, a fake provider and the real prompt modes, and counts the system prompt plus every tool definition (characters over four, a growth guard rather than an exact cost). Stored budgets: minimalist 3,500, balanced 5,000, full 6,400 estimated tokens for the test's tool pool (built-in, code graph, computer, browser and two fake MCP tools, with sub-agents on); measured 3,353, 4,808 and 6,111 (grep's context and edit's edits cost nothing: see tool discovery). The test also requires minimalist < balanced < full and checks which tools each tier offers. The budgets came down from 3,700 / 5,250 / 9,100 when the core tools' descriptions were cut to what a model needs (a limit, an output shape, a failure), stated once: an entry of read's paths no longer repeats the descriptions of the top-level keys. A change that makes a tier bigger fails it: raise the number on purpose in review, or move the cost to a higher tier.
Project context is measured separately because it depends on the project, not on the tier: packages/daemon/test/tier-instructions.test.ts composes the system prompt of each tier's prompt mode for a fixture project (a 1.6 KB AGENTS.md and five skills) and stores what that project adds. Every mode now carries the user's instructions and focus/full carry the compact skill catalog (prompt modes), so the fixture adds 204 (minimalist), 346 (balanced) and 346 (full) estimated tokens; before, it added 0, 0 and 346, because minimal and focus dropped AGENTS.md and focus dropped the catalog. The per-tier baseline numbers above are for a project without instruction files or skills and did not change. An AGENTS.md is capped at 16,000 characters (about 4,000 tokens) per file and 64,000 in total, so the worst case per mode is bounded.