Read-side access to Claude Code's on-disk background-job state.

Claude Code ships a supervisor daemon (claude daemon run) that orchestrates background agent tasks launched via the claude agents TUI. Per-task state lives under ~/.claude/jobs/<short_id>/:

• state.json -- current snapshot: state, intent (original prompt), session id, link to the session JSONL, timestamps, auto-generated name, etc.
• timeline.jsonl -- append-only event log: at each state transition the daemon writes a line carrying timestamp, new state, a one-line detail, and (often) the full text body.

The session content itself is a normal ~/.claude/projects/<slug>/<session_id>.jsonl -- the same format ClaudeWrapper.History parses. Each summary's session_path points at it for direct cross-linking.

This module is read-only on purpose. The dispatch protocol (how the TUI launches new tasks) is undocumented and version-sensitive; mirroring it would defeat the drift defenses the wrapper relies on.

Liberal parsing

Like ClaudeWrapper.History, parsing is forgiving. list/1 skips entries that are not directories (e.g. the daemon's pins.json), that carry no state.json (a spare worker dir), or whose state.json fails to parse -- malformed jobs are dropped rather than failing the whole listing. The timeline parser skips blank and malformed lines.

Example

{:ok, root} = ClaudeWrapper.Jobs.home()
{:ok, summaries} = ClaudeWrapper.Jobs.list(root)

for s <- summaries do
  IO.puts("#{s.short_id}  [#{s.state}]  #{s.intent}")
end

{:ok, job} = ClaudeWrapper.Jobs.get(root, "90c961c7")

for event <- job.timeline do
  IO.puts("#{event.at}  #{event.state}")
end