Skip to main content

Setup & configuration

Settings by Mode: what actually applies where

Mode is not a property of a setting: there is no per-setting Chat or Code switch. Whether a setting applies in Chat Mode depends on what the agent actually does in that mode, and for several settings the answer is the opposite of what the Settings panel's grouping suggests. This page lists the ones that surprise people.

How to read this page

There are three ways a setting ends up mode-specific:

Basis What it means How sure we are
Mode-gated The app explicitly checks the mode before using the setting. Certain.
Code-only in practice Nothing checks the mode, but the setting only matters in situations that arise in Code Mode (a panel that only exists there, a tool only available there, a session type only Code Mode creates). Accurate today, but not enforced.
Undetermined Not yet checked. Unknown.

The Settings panel shows a small badge on a page in the left rail only when every relevant setting has been checked, and a group header counts the pages under it that apply only in the other mode: Code + Chat needs the checked settings on that page to reach both modes, and a single-mode badge such as Code-only in practice needs every setting in the section checked. Today that is Agent → Verification (QEL) and Agent → Routing (Code + Chat) and Appearance & Editor → Editor (Code-only in practice). Automations → Background runs (Fleet) has no badge: its fleet.* and worktree settings are code-only in practice, but it also shows agent.max_background_sessions, which has not been checked and also caps sub-agents. Individual rows carry their own badge when the setting is checked: "Code only", "Differs by mode", "Code + Chat edits" (the Verification checks: always in Code Mode, and in Chat Mode only on turns that edited a file), or "Both modes" for a setting with one shared value (temperature, max tokens, reasoning effort), where changing it changes Chat and Code together. No badge means "not checked yet", never "both modes".

QEL in Chat Mode runs only when the turn wrote a file

  • Code Mode: any creation, modification or bug-fix task is verified, with the usual repair budget.
  • Chat Mode: the same, but only when the turn successfully wrote or edited a file. A write that was refused does not count. In Chat Mode a write needs an attached folder.

So a Chat Mode request like "fix this snippet I pasted" with no folder attached is not put through build verification, because there is nothing on disk to check. The same rule applies to the final check when a turn ends. Each decision appears in the debug log as a qel_gate event with the mode, the task type, the write count and the reason.

So these settings (qel.contract_review, qel.semantic_judge, qel.execution_proofs, qel.modification_proofs, qel.code_review, qel.tsgo, qel.grader_model) apply in Code Mode, and in Chat Mode only on turns that wrote files.

`qel.answer_judge` applies on the opposite turns to every other QEL setting

QEL grades a turn one of two ways: as a build or fix (every other qel.* setting) or as an answer (a check that the reply does not cite files it never read, plus the optional qel.answer_judge second model). The two never overlap, so every turn gets at most one of them.

The answer check runs in either mode, only on turns that used tools and produced an answer. It is advisory: it adds at most one note and never blocks or repairs. A Chat Mode fix request that wrote nothing is graded here rather than by the build check. A fix for code pasted into a chat with no folder linked is answered in one reply without tools, and so is a request for new code there ("Build a React component for ..."), so neither check runs on them.

So turning qel.answer_judge off does not quieten the rest of QEL, and turning the rest off does not quieten it.

Settings that are code-only in practice

Nothing checks the mode for these. They only take effect in Code Mode because of what else has to be true:

  • qel.code_review needs changed files, and Chat Mode normally has no file-writing tools. The exception: Chat Mode with a folder attached can write, and QEL runs on a chat turn that did (see above), so the review can too.
  • fleet.* and the worktree settings (agent.worktree_max_age_days, agent.worktree_disk_warning_gb, agent.worktree_disk_max_gb) apply because Fleet sessions are always code sessions.
  • editor.* configures the code editor panel, which only exists in Code Mode.
  • memory.auto_extract is one setting with two effects: the off-switch stops memory extraction in both modes, while decision-log extraction only ever happens in Code Mode.

Other settings that surprise people

# Setting What surprises
1 llm.smart_auto vs llm.routing_rules Adjacent in the Routing section, opposite mode behaviour. Smart Auto does nothing in Chat Mode; routing rules apply in both modes and can use mode as a rule condition.
2 llm.smart_auto Not project-overridable, while its neighbours llm.default_model / llm.code_model are. A .bodega/config.json that sets it is silently ignored.
3 agent.dual_model Has no effect in Chat Mode, with no marker in the UI.
4 skills.auto_activate Code-only. Skills you trigger with a slash command are not: they work in both modes.
5 Project rules (.bodega-rules / AGENTS.md / CLAUDE.md) Weaker in Code Mode. Both modes include the rules, but only Chat Mode adds an extra reminder telling the model to follow them.
6 persona.system_prompt Used in both modes, but wrapped in different surrounding instructions per mode, so the same custom prompt can behave differently.
7 browser.widened_enabled Checked when tools are set up for a turn and again before each navigation, so turning it off mid-session blocks the next widened navigation straight away.

Every audited setting, by mode

This table is generated from the same registry the Settings panel reads its badges from, so the two cannot disagree. A setting that is not listed has not been checked: it is not a claim that it works the same in both modes.

Setting Applies in Basis Why
llm.smart_auto Code only Mode-gated Smart Auto returns null outside Code Mode.
llm.routing_rules Both Mode-gated Routing rules evaluate in both modes and take mode as a rule condition: per iteration in Code Mode, once per message in Chat Mode.
chat.artifacts Both (differs by mode) Split Artifacts run in both modes (2026-11-09). Code Mode adds a files-first rule and opens artifacts in an Artifact tab in the editor area.
agent.dual_model Code only Mode-gated Dual-model mode only activates in Code Mode; Chat Mode always uses one model.
skills.auto_activation Code only Mode-gated Auto-activation by description runs only in Code Mode. Skills you trigger with a slash command run in both modes.
memory.auto_extract Both (differs by mode) Split Correction capture runs in both modes; decision extraction runs only in Code Mode, on turns that wrote files.
qel.contract_review Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.semantic_judge Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.execution_proofs Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.modification_proofs Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.tsgo Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.grader_model Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.code_review Code; Chat only on turns that wrote or edited a file Mode-gated QEL runs in Code Mode, and in Chat Mode only on turns that actually wrote or edited a file.
qel.answer_judge Both Mode-gated The answer-grounding check runs on the turns QEL skips, in either mode.
notifications.turn_end Both Shared (no mode check) Fires when the current chat finishes a turn while the window is unfocused, in either mode.
notifications.approval_needed Both Shared (no mode check) Fires when a tool approval appears while the window is unfocused, in either mode.
power.keep_awake_during_runs Both Shared (no mode check) Holds the machine awake (plugged in only) while any run is active, in either mode.
fleet.* Code only in practice Code-only in practice No mode check reads fleet.*. They are code-only because fleet sessions are created as code sessions. Nothing enforces it.
agent.worktree_max_age_days Code only in practice Code-only in practice No mode check reads the worktree settings. Worktrees exist only for sessions created as code sessions. Nothing enforces it.
agent.worktree_disk_warning_gb Code only in practice Code-only in practice No mode check reads the worktree settings. Worktrees exist only for sessions created as code sessions. Nothing enforces it.
agent.worktree_disk_max_gb Code only in practice Code-only in practice No mode check reads the worktree settings. Worktrees exist only for sessions created as code sessions. Nothing enforces it.
editor.* Code only in practice Code-only in practice No mode check reads editor.*. They configure the code editor panel, which only mounts in Code Mode.
llm.qel_auto_escalate Code only Mode-gated Escalating to the smart model after a failed check happens only in Code Mode.
llm.temperature Both Shared (no mode check) One value per model, read with no mode check: changing it affects Chat and Code.
llm.max_tokens Both Shared (no mode check) One value per model, read with no mode check: changing it affects Chat and Code.
llm.reasoning_effort Both Shared (no mode check) Global, with per-model overrides, read with no mode check: changing it affects Chat and Code.
persona.system_prompt Both (differs by mode) Split Used in both modes, but the scaffolding built around it differs by mode, so the same prompt can behave differently.

Using the mode lens in Settings

Under the search box in Settings there is a switch: All · Chat · Code. It changes what you see, never what is set.

  • All (the default) shows everything as usual.
  • Chat or Code shows what applies to that mode at full strength. A section whose every checked setting only applies in the other mode folds into one muted row at the end of its group, such as "Code-only (1)" for Editor under Chat. Click it to open it. Rows that only apply in the other mode read in muted text. Nothing is hidden, and every control still works.
  • Settings nobody has checked yet stay at full strength, with no badge. A line under the switch says so: no Chat, Code or Both tag means not checked yet.
  • Search still searches everything. Results that only apply in the other mode sort last, carry a badge, and a footer counts them.
  • Your choice is remembered on this computer. It does not follow the app's current mode by itself; the first time, Settings offers Match current mode.
  • A link that opens a folded section opens its fold. The guided tour shows All while it runs and puts your choice back afterwards.

Not classified yet

Loops, Observatory, hooks, GitHub, MCP, ACP agents, plugins and custom agents have not been classified yet, rather than guessed. Neither have agent.max_background_sessions (which is why Fleet has no section badge), agent.auto_verify, general.permission_mode, skills.enabled, llm.code_model, boost.enabled and agent.subagents_enabled.

Controls the app shows or hides depending on the current mode have not been reviewed either.

Mode is not the same axis as global-vs-project

Two different questions get confused with each other:

  • "Does this setting apply in Chat Mode?", this page.
  • "Can a project's committed .bodega/config.json change it?", a security boundary, covered on the Settings, Themes & Spending page under "What a project's committed config can and cannot change", and on the Workspace Trust page.

They are independent, and conflating them is dangerous in one specific direction: a setting being moved next to project-overridable settings in the UI is not a reason to make it project-overridable. The list of project-overridable settings is reviewed on its own and does not follow the panel layout.

One behaviour worth repeating here because it interacts with routing: llm.routing_rules does not shadow your global rules, project rules are added above them. Because rules are first-match, prepending is strictly more powerful than shadowing.

In the CLI

The bodega CLI reads the same settings through the same backend, so everything on this page applies there identically, the QEL rule, the routing split and the Chat Mode tool set work the same in both. What the CLI does not yet have is this annotation in bodega config output; it prints values without saying which mode they reach. That is a known gap, not a difference in behaviour.

This page mirrors the in-app docs hub for app version 1.0.0-beta.54.1. Found something unclear or out of date? Tell us on Discord. New here? Download the free beta and follow along.