Skip to content

autodev-local-setup.md

Latest commit

 

History

History
130 lines (99 loc) · 6.2 KB

autodev-local-setup.md

File metadata and controls

130 lines (99 loc) · 6.2 KB
Raw

Setting up autodev for local development

This sets up a personal autodev instance that acts as you, on your own repo or a test repo. It's for developing and testing autodev. For the shared service account, see "Run it as the group service" in notes/autodev-runbook.md. Back to the README.

What you're starting

pixi run mia up reads mia.yaml (copy mia.yaml.example) and starts, in tmux:

tmux session window runs where
temporal temporal Temporal dev server (gRPC localhost:7233, UI :8233); shared by every service host
<service> broker csub-broker, applying your ~/.config/csub/broker.toml policy host
<service> worker autodev worker, wrapped in agentic-sandbox sandbox

For a personal instance mia.yaml lists one service, your autodev.yaml. The worker polls GitHub, drives the workflows, and asks the broker to submit Claude Code jobs to LSF.

Prerequisites

Run these from a Janelia cluster submit host.

  1. pixi on your PATH, and tmux.

  2. agentic-sandbox checked out next to this repo (../agentic-sandbox), or set sandbox_dir in mia.yaml to its scripts directory. It needs rootless podman or bwrap set up, per its README.

  3. csub with Claude support. The pixi env pins csub to krokicki/csub branch deploy-all, so pixi install provides it, including csub-broker.

  4. Broker policy in ~/.config/csub/broker.toml:

    • allow_claude = true
    • allowed_hosts includes api.anthropic.com, claude.ai, platform.claude.com, plus everything in agent.allow_hosts from autodev.yaml
    • allowed_roots contains your paths.work_root

    Check with csub probe.

  5. Claude login for your account: claude auth login. Check: csub submit --claude --walltime 5 --wait -- claude -p "Reply OK" --model sonnet prints OK.

  6. A GitHub repo you can write to, to hold test issues. The default in the example config is this repo.

Steps

1. Install

pixi install

2. Configure

cp autodev.yaml.example autodev.yaml    # git-ignored

Edit autodev.yaml:

  • deployment: autodev-<you>. This prefixes your labels, task queue, workflow IDs, branches and tmux session, which keeps your instance apart from the service and from colleagues.
  • github.repo: the repo to watch.
  • paths.work_root: agent workspaces, which must be inside your broker's allowed_roots.
  • paths.state_dir: the worker's logs and state. Keep it apart from work_root. work_root is bound read-write into the sandbox, so it must not be an ancestor of state_dir, and <state_dir>-host (the broker log) and temporal.state_dir from mia.yaml (the Temporal binary and database) must stay outside every writable bind.

3. GitHub token

mkdir -p ~/.config/autodev
gh auth token > ~/.config/autodev/token && chmod 600 ~/.config/autodev/token

Or save a fine-grained PAT there with Contents, Pull requests and Issues read/write, and Checks and Commit statuses read. Then set github.token_file: ~/.config/autodev/token in autodev.yaml. The sandboxed worker reads the token from this file because the sandbox masks .env. Host-side commands such as autodev check also accept .env or $GITHUB_TOKEN.

4. Check, create labels, start

cp mia.yaml.example mia.yaml            # keep only the `autodev` service, pointing at your autodev.yaml; keep temporal.managed: true
pixi run mia exec autodev check         # the `temporal` row is BAD until `mia up` starts the dev server; fix any other BAD first
pixi run mia exec autodev init-labels   # creates the autodev-<you>:* labels (idempotent)
pixi run mia up                         # tmux sessions temporal and autodev; applies the poll schedule
pixi run mia exec autodev check         # now every row should say ok

5. Try it

Open a small issue in the repo (for example "add a --version flag"), add the label autodev-<you>:go, and watch:

pixi run mia attach autodev     # windows: broker, worker (Temporal: mia attach temporal)

The Temporal UI is at http://localhost:8233. From your laptop use ssh -L 8233:localhost:8233 <submit-host>. Workflows are named per issue.

autodev opens PR autodev-<you>/issue-N and labels the issue autodev-<you>:review. Because you are both the bot and the reviewer, GitHub won't let you approve your own PR. Comment /approve <first 7+ chars of the head sha> instead. Only /approve comments posted after the bot's latest round summary count. Use /revise <what> or a Request changes review to run another round.

Day to day

pixi run mia ps                     # every service, or: mia ps autodev
pixi run mia logs autodev [worker|broker] [-n N] [-f]
pixi run mia logs temporal
pixi run mia attach autodev
pixi run mia restart autodev        # after changing code or config; leaves Temporal and other services alone
pixi run mia restart                # everything, Temporal included
pixi run mia down                   # everything, Temporal last
pixi run mia exec autodev check     # any app command, with this service's config
pixi run test                       # unit tests; no GitHub, LSF or Temporal server needed

Logs: worker logs are in <state_dir>/logs, the broker log in <state_dir>-host/logs, and Temporal's in temporal.state_dir/logs from mia.yaml (mia logs temporal).

Gotchas

  • Workflow code is replay-sensitive. Changing autodev/workflows.py or autodev/decide.py can break in-flight workflows. test_replay_recorded_history catches it. If the change is intentional, guard it with workflow.patched() and re-record the history (see the runbook's known gaps).
  • Agent tests need tools in the job. For example pixi needs path_prepend and extra_mounts, and package hosts need to be in agent.allow_hosts. Nothing installs them for you.
  • The Temporal dev server is unauthenticated. Anyone with a login on the host can signal or terminate workflows, so don't run this on a shared host.
  • agent.api_key_env is refused by mia up for now. Use a Claude login.
  • No cancel command. temporal workflow terminate -w <id> stops a workflow but leaves its job running. Also run csub kill <job>.