Agent profiles
knoxchat.agentProfile picks how the agent loop behaves after it is allowed to run: how stubborn it is with repeated tool calls, and what it runs to check an edit. It does not replace Ask / Edits / Auto or Plan.
Setting: default | systems | rust | auto (default: default). Change it in Knox: Open Knox Settings or settings.json.
GUI chat stays unlimited tool rounds. Stop with Cancel. Background jobs and /autonomous are capped (40 / rust 60 / systems 80, absolute 200). Details: Background agents.
When the resolved profile is systems or rust, the agent meter shows that chip.
What each profile does
| Setting | Doom-loop after | Post-edit check |
|---|---|---|
| default | 3 identical / failing calls | LSP diagnostics (knoxchat.enablePostEditVerification) |
| rust | 4 | cargo check --workspace --all-targets |
| systems | 5 | make (kernel / QEMU) |
| auto | (resolved) | systems on kernel/QEMU trees, rust on a root Cargo.toml, otherwise default |
A doom-loop chip offers Change strategy (one-line instruction to pick a different approach). Stop any run with Cancel.
systems / rust ignore knoxchat.enablePostEditVerification (the compile oracle wins). A non-empty knoxchat.verifyCommand also skips LSP and selects command mode.
What auto looks at
auto lists the workspace root (not nested folders) and classifies, first match wins:
- kernel —
Kconfig(orKconfig.*) plusarch/,Makefile, orKbuild - qemu —
meson.buildplustarget/,accel/, orhw/ - cargo — root
Cargo.toml(a nestedrust/Cargo.tomldoes not count, so rust-for-linux stays systems) - Otherwise default (app tree)
A kernel/QEMU tree that also has a root Cargo.toml stays systems. Explicit systems / rust / default are never changed. With Jev on, auto can be confirmed; that overlay still does not override an explicit setting.
Bare Makefile / CMakeLists.txt is not enough for systems (app repos often have those).
Workspace card
On classified trees Knox injects a short Codebase Card on submit (not a dump of .config or lockfiles):
| Kind | Detected from |
|---|---|
| kernel / qemu | Kconfig + arch, or meson + target/accel/hw |
| cargo | root Cargo.toml |
| editor | this KnoxCoder / VS Code fork (src/ + extensions/ + product.json / gulp) |
| node | package.json |
| python | pyproject.toml, pytest.ini, requirements.txt, setup.py / setup.cfg |
| go | go.mod |
| cmake | CMakeLists.txt |
Order if several match: kernel > qemu > cargo > editor > node > python > go > cmake. The editor fork wins over a bare Node package.json.
You do not paste this. /instructions is still the place to see token cost of rules and AGENTS.md.
After a red build
If the verify command (or make / cargo test / npm test on the product path) is still red, Knox holds instead of treating the turn as done and keeps trying to fix. A hold chip can show on the last assistant message.
- Same compiler-error signature:
knoxchat.verifyMaxIterations(default 8) skips further auto-builds. - Extra no-tool “oracle still red” rounds cap at 8; 3 in a row with the same diagnostics stop that loop.
Set knoxchat.verifyMode to off if you do not want a post-edit oracle. diagnostics is LSP; command is knoxchat.verifyCommand (examples: make -j8, cargo check --workspace --all-targets). A non-empty verify command selects command.
Settings recap
| Setting | Effect |
|---|---|
knoxchat.agentProfile | default / systems / rust / auto |
knoxchat.enablePostEditVerification | LSP check + attempted fixes after mutating tools (ignored on systems/rust or when verifyCommand is set) |
knoxchat.verifyCommand | Compile oracle; empty = LSP on default |
knoxchat.verifyMode | diagnostics / command / off |
knoxchat.verifyMaxIterations | Circuit breaker for identical compiler errors (1–20, default 8) |