- A key's name is judged on its own: anything that looks like a key is refused whatever key it's for, so Rename can't be used to test guesses - Previews can't add a key, rename an account or forget a GitHub login; a test checks every action the code refuses is blocked - A Pi switch over a Claude login it can't identify is refused, and the live entry is compared before it's written - Hermes key add: the child dies with the request; the name goes as --label - Switching from the composer reloads the agent's models - changed_logins dedups through a sorted set - Hermes model names: plain files up to 64 MB, naming waits at most 5 s; Hermes .env read is bounded - The API key dialog's Cancel waits while a key is checked - stores.rs module docs match what each store does Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
History
|
|
|---|---|---|
| .. | ||
| examples | ||
| src | ||
| tests | ||
| Cargo.toml | ||
| README.md | ||
Explicit agent installation
Settings → Providers → Install runs on the selected device, as its user. Merely
listing or detecting agents never invokes these CLI installers. The table below
was checked with curl -fsSL against the linked vendor documentation on
2026-09-21. Commands download the vendor's current release, except for the existing
pinned Antigravity archive.
| Agent / documentation | Unix command (Linux and macOS) | Native Windows command |
|---|---|---|
| Claude Code | curl -fsSL https://claude.ai/install.sh | bash |
irm https://claude.ai/install.ps1 | iex |
| Codex | curl -fsSL https://chatgpt.com/codex/install.sh | sh |
irm https://chatgpt.com/codex/install.ps1 | iex |
| Cursor | curl https://cursor.com/install -fsS | bash |
irm 'https://cursor.com/install?win32=true' | iex |
| OpenCode | curl -fsSL https://opencode.ai/install | bash |
npm.cmd install -g @opencode/cli |
| Pi | curl -fsSL https://pi.dev/install.sh | sh |
npm.cmd install -g --ignore-scripts @earendil-works/pi-coding-agent |
| Grok | curl -fsSL https://x.ai/cli/install.sh | bash |
npm.cmd install -g @xai-official/grok |
| Hermes | curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash |
irm https://hermes-agent.nousresearch.com/install.ps1 | iex |
| Devin | macOS, when brew resolves: brew install --cask devin-cli; otherwise curl -fsSL https://cli.devin.ai/install.sh | bash |
irm https://static.devin.ai/cli/setup.ps1 | iex |
| Muse Code | curl -fsSL https://dev.meta.ai/install.sh | sh |
irm https://dev.meta.ai/install.ps1 | iex |
| fx | curl -fsSL https://fx.sh/setup.sh | bash |
Manual setup (no native Windows installer) |
| Antigravity | Existing verified 1.1.1 archive in acp::antigravity_archive; no shell command |
Same archive path on supported architectures |
Unix shell commands run through sh -c, with PATH augmented from the user's login
shell and npm toolchain. Native Windows irm commands run through
powershell -NoProfile -NonInteractive -ExecutionPolicy Bypass -Command.
Windows npm commands use the resolved npm launcher (normally npm.cmd).
If shell prerequisites are absent, these documented npm alternatives are used
when npm resolves: npm install -g @openai/codex,
npm install -g @opencode/cli,
npm install -g --ignore-scripts @earendil-works/pi-coding-agent, and
npm install -g @xai-official/grok. Codex also falls back to npm on Windows when
PowerShell is unavailable. Grok's npm route is documented in its
enterprise guide; its
published package metadata
includes Windows binaries. No action is offered when none of the documented
methods' prerequisites resolves, or for Mock. The row then keeps a manual hint;
shell hints on Windows can be used in WSL.
Install processes receive CI=1, NONINTERACTIVE=1, TERM=dumb, no stdin/TTY,
and no inherited ZERON_* or nested Claude environment markers. Output is bounded
and credentials are redacted before errors reach Settings. Cancel, timeout
(15 minutes), or dropping the request kills the owned process tree. There is no
automatic elevation: an installer that needs sudo or an interactive prompt may
fail. Cancelling does not undo files an installer already wrote.
After installation, executable version caches for the agent are invalidated and the normal installed probe runs again. A missing binary produces PATH guidance; the running app's login-shell snapshot is not refreshed. Authentication remains a separate step. Hermes may additionally require its documented ACP Python extra for agent sessions; the Install action checks CLI detection only.
Muse Code
muse::MuseHarness drives muse serve over the Muse Session Protocol (MSP):
newline-delimited JSON-RPC 2.0 on stdio, the wire Muse's own TypeScript SDK
(@muse-code/sdk) speaks. Every state-changing method is a command carrying a
client-minted UUIDv7 commandId; a fresh turn's id is its turn/start command
id. A runtime is one host with one session (session/start, or
session/resume by id); turns end on turn/completed, including the wake turns
Muse starts when background subagents finish. Later prompts are
turn/start with ifBusy: "steer", stops are turn/interrupt, approvals are
answered from the host's own choices, and questions reach the user through
request_input. Models (model/list) and titles run on a read-only host
(--disable-shell --disable-write --no-session-log). The host's environment
drops an inherited META_API_KEY and sets MUSE_NO_AUTO_UPDATE=1.
tests/muse.rs replays sessions recorded from a real Muse Code 1.4.x host
(tests/fixtures/muse/*.ndjson, through fake-muse.py).
Verification
The installer-fixture feature exposes an explicit command injection seam for
integration tests. The engine's command environment override is compiled only
for unit tests, never production. Tests exercise successful detection, failure,
missing binaries, cancellation, duplicate requests, process descendants, bounded
redaction, platform prerequisites, and version-cache invalidation.
The opt-in engine test rpc::tests::installer_rpc_npm runs the documented Codex
npm alternative in a disposable prefix and isolated home, through InstallHarness.
It asserts that the returned Settings descriptor changes from uninstalled/disabled
to installed/enabled. It never runs a vendor shell installer. Set TMPDIR to a
scratch directory before running it with --ignored --nocapture.
The iOS app adds no Install action. HarnessDescriptor.can_install already has
#[serde(default)] (crates/engine/src/registry.rs); iOS's WireHarness: Decodable
in apps/ios/Zeron/Sync/WorkspaceStore.swift ignores unrecognized keys, including
canInstall. HarnessCatalog continues consuming the existing mapped fields.
fx
fx uses its native fx acp server. Install on macOS/Linux with
curl -fsSL https://fx.sh/setup.sh | bash, then run fx login or fx setup
in a terminal to configure the model provider. Zeron detects PATH and
~/.local/bin/fx; FX_EXECUTABLE overrides the executable.
Models and reasoning levels come from ACP. The provider selector is excluded
from model discovery. Explicit model selections are sent to fx and errors
propagate instead of silently running a different model. Skills use the existing
host discovery for .agents/skills, .claude/skills, and .fx/skills.
Zeron writes fx's ask policy before every prompt, including resumed sessions:
some fx releases report ask while retaining a saved automatic policy. Zeron then
decides permissions: Supervised presents requests to the user; Full access
approves them. Streaming, tool calls, cancellation and session loading use the
shared ACP driver. Login and updates remain CLI-managed; native Windows has no
one-click installer.