overseer: Todo

[stamatron]

Closes #1

gh-overseer #1 — done

Shipped the concrete items; each is a separate subtask: commit. Defaults unchanged, so nothing new activates unless you set the config key. All 31 tests pass.

Done

• Local log preserved — log_file config appends the (redacted) log to disk as well as stdout.
• Verbose flag — --verbose (or verbose: true) streams the agent's live output into the log so you can watch a job.
• Optional acks + shorter reports — ack_start / ack_finish (both default true) gate the start/finish comments; prompt now asks the agent to keep its report short.
• Review phase first — prompt now tells the agent to review and ask for clarification before any code, making no changes if unclear.
• Impersonation — impersonate: true commits under your identity (commit_name/commit_email, else your login + noreply) instead of the bot's.
• Optional webfetch — allow_web: true appends WebFetch to allowed tools. (You could already do this by hand via allowed_tools; this is the convenience flag.)
• Session per Issue/PR (the priority) — work_root gives each issue/PR a persistent checkout dir. Claude keys its session transcripts by cwd, so a stable path is what lets a claude runner --resume the same session on every follow-up. First run mints a UUID --session-id (stored in state.json under sessions); follow-ups --resume it; a stale transcript falls back to a fresh session automatically. No work_root → old temp-dir behaviour, no session reuse.

Needs your call (no code guessed)

• mini runs as claude default — not an overseer bug. Two causes: (1) with no run_agent: mini line the first runner (claude) is used — that's by design; (2) the mini runner has no -m <model> so mini-swe-agent falls back to its default model (claude), and --cost-limit 0 makes it exit immediately. Tell me the model you want mini to drive and I'll fix the runner (add -m …, sane --cost-limit).
• Sandbox / Docker jail for arbitrary execution — deliberately not built yet; it's a real design fork. Rough options:
  • Docker-per-job: run the agent inside a container (repo bind-mounted, --network gated by allow_web, dropped caps, non-root, CPU/mem limits). Strongest isolation; needs Docker on the host and reworks run_agent to docker run.
  • Lightweight: --permission-mode + a tight allowed_tools (already in place) plus a seccomp/nsjail wrapper. Less isolation, no daemon.
  • The persistent work_root dirs are already an easy container mount point.
    Which way do you want (Docker dependency acceptable?), and should the jail be default-on or opt-in?

Note

• With sessions on, the prompt still ships the full thread each time (harmless redundancy alongside resume). Can trim later if you want leaner follow-up prompts.

$2.97 · 40 turns · 2,728,818 in (2,647,525 cached) · 35,133 out