Skip to content

Latest commit

 

History

History
69 lines (51 loc) · 2.79 KB

agents.md

File metadata and controls

69 lines (51 loc) · 2.79 KB

Agents and permissions

Lanzer works with any agent that speaks the Agent Client Protocol. Pick one in .env, which every command reads from the folder you run it in. --env .env.codex uses another file instead, and a variable set in your shell always wins:

::: code-group

LANZER_ACP_COMMAND=npx
LANZER_ACP_ARGS=["claude-agent-acp"]
LANZER_ACP_MODEL=sonnet
# sign in first: npx -y @openai/codex login
LANZER_ACP_COMMAND=npx
LANZER_ACP_ARGS=["-y","@agentclientprotocol/codex-acp"]
LANZER_ACP_PROVIDER=codex
LANZER_ACP_MODEL=gpt-5.5
LANZER_ACP_COMMAND=npx
LANZER_ACP_ARGS=["-y","@google/gemini-cli","--acp"]

:::

Variables

Variable Example Meaning
LANZER_ACP_COMMAND, LANZER_ACP_ARGS npx, ["claude-agent-acp"] The agent to start.
LANZER_ACP_PROVIDER codex Only matters for Codex.
LANZER_ACP_MODEL sonnet, gpt-5.5 The agent's own alias, not an API id.
LANZER_ACP_EFFORT medium Reasoning effort.
LANZER_ACP_MAX_ATTEMPTS 2 Prompts per session.
LANZER_ACP_ALLOW read,edit,search,think,other What the agent may do.
LANZER_ACP_ISOLATED 1 Run without your own agent setup.

::: warning Model names Use the agent's alias (sonnet, opus, haiku), not an API id like claude-sonnet-4-6. An unknown alias is skipped and the agent's default is used. :::

Permissions

A run is unattended, so by default the agent may only read, edit, search, think and other (where skills load). No shell, no deleting or moving files, no network.

# also allow shell commands
lox-lanzer generate ./examples/hello.lanzer --allow read,edit,search,think,other,execute
# no policy at all
lox-lanzer generate ./examples/hello.lanzer --allow-all

The list is read literally: naming kinds limits the run to exactly those. Refusals print as [perm] denied execute: … and the next fix prompt names them.

What confines files:

  • Writes stay in the campaign's workspace; reads in the workspace plus the grammar, reference files and skill. Claude uses its own file tools, so it keeps to its session roots instead: the same folders, but all writable.
  • After the run, a declared file left unwritten, or a changed support file, fails it. Other new files are reported.
  • Claude runs in acceptEdits mode, never bypassPermissions.

::: danger Shell access An agent that can run shell commands steps around every file check. At scale, run in a container. :::

Codex

Codex through codex-acp gets Lanzer's tools and real sessions, like Claude. Two limits: its token counts cover only the last model request, and whether it asks before running a shell command (so that leaving out execute is enforced) is not verified yet.