Monitor Log Agent — Development skill for Claude Code

monitor-log-agent

[简体中文](README.zh.md)

An example of a **read-only** Claude Code agent: a Python host prepares context, then [Claude Agent SDK](https://code.claude.com/docs/en/agent-sdk/overview) runs the agent loop and the host writes an HTML transcript.

Shape

Piece	Role
Host process	Load .env, sync source, build the prompt, cap tools, render HTML
Claude Code CLI	Spawned by the SDK; this is the actual agent
Skills repo	Filesystem skills (.claude/skills/). Session cwd points here
Source tree	Mounted with add_dirs so the agent can Read / Grep / Glob
Custom MCP tool	Read-only live checks (ssh_read: argv list + host allowlist, not a shell string)

Do not give the model Bash, Edit, or Write. Deny Bash again with a `PreToolUse` hook. `permission_mode="dontAsk"` and `strict_mcp_config=True` so only the MCP servers you pass in are loaded.

Skills live in a sibling repo, not inside the runner. A new investigation type should be a new skill plus source/SSH mapping, not a new agent binary.

Isolation

SDK default `setting_sources` is user + project + local. That loads `~/.claude/settings.json`, whose `env` **overwrites** process environment (including `ANTHROPIC_BASE_URL` / `ANTHROPIC_API_KEY` / `ANTHROPIC_MODEL`).

This project sets `setting_sources=["project"]` so only the skills repo’s `.claude/settings.json` is loaded.

The host calls `load_dotenv` on this repo’s `.env` and copies `ANTHROPIC_*` into `ClaudeAgentOptions.env`. You do not need `uv run --env-file`. `load_dotenv` does not override variables already in the shell.

Limits

20 turns, `$5` `max_budget_usd` (SDK `total_cost_usd`; may not match a third-party bill).

Requirements

• Python 3.11+ and uv
• Claude Code CLI on PATH

Run

uv run monitor-log-agent

Stdout prints `cost=$... subtype=...`. Report: `out/latest.html`.