What now

An overview the owner reads in under a minute. One line per task, never the detail of how it was done. Read only; change nothing, start nothing.

Gather (cheap, in this order)

1. This conversation: what the owner asked for and what was finished.
2. Running work: background agents, workflows and shells. Ask nothing of them. For each one this session launched (agent launches carry an agentId; Workflow and background Bash launches carry only a taskId, so match on either):
   • Look at its LAST event, not its first. One id notifies every time it stops, and again after a SendMessage wakes it. Running = launched or woken after its last task-notification. Finished = a notification after the last wake (status completed, failed or killed).
   • Agents and workflows: the notification check alone decides.
   • Background shells: "no notice yet" alone is not proof, because some shells never send one. Second check: lsof -t -- <tasks dir>/<taskId>.output (the path its launch result printed); exit 0 (a process holds the file) means running, exit 1 means ended. When the two checks disagree, write "unknown (launched <time>, no notice)".
   • Show each running item's start time, so one running for hours stands out. Shell: stat -f %Sm <taskId>.output. Agent or workflow: the launch line's timestamp in the transcript. None found: "start time unknown".
   • A finished one whose result you have not read yet is "done, not yet reviewed".
   • Launches no longer in context (after a compaction): grep the session transcript ~/.claude/projects/<project-dir>/<session-id>.jsonl for async_launched and task-notification. Cannot read it: "unknown (history compacted)". Why this way: no built-in screen gives this list to a skill (/tasks and /workflows are for the owner's eyes), and the transcript rules above come from claude-hud src/transcript.ts and a checked run of this session's own transcript (research 2026-10-18; claude-hud is github.com/jarrodwatts/claude-hud).
3. The plan, if there is one: the plan folder given in the args, otherwise the one this session worked in (plan-and-build: progress.md, plan.md, register.md; a skill plan in ~/plans/; a research folder's report.md). Run grep -n -E 'BLOCKED|UNVERIFIED|owner decides' <plan folder>/*.md first, then read the end of progress.md and only the chunk and task lines of plan.md. No plan: use the task list or the conversation.
4. Repo and branch, always: for every repo or worktree this session worked in (the working directory, worktrees agents used, repos named in the plan), run git -C <path> rev-parse --show-toplevel, git -C <path> branch --show-current (empty means detached HEAD: say so), git -C <path> status --short | wc -l (the uncommitted count) and git -C <path> remote get-url origin (fails when there is no remote). Name the repo by its folder. Not a git repo: say so.

Never trust memory for "running": a task is running only if a check above shows it. If something cannot be checked, write "unknown" for it.

Write (load unslop first)

Plain words, no task ids (write what T4 is, not "T4"), no em dashes. Each item: one line, saying what it is and what it needs or uses (a skill, tool, branch, service, or a decision from the owner).

**Goal:** <one sentence: what this session or plan is for>
**Where:** <repo> on branch `<branch>` (<clean | N uncommitted files>); one line per repo
or worktree in use, e.g. "agent worktree: ui-qa on `plan/ui-qa-c2`"

**Done**
- <task>: <result in a few words> (<what it used>)

**Running now**
- <task>: <who is doing it: main session, agent "<name>", workflow> (<what it uses>)
  or "Nothing running."

**Next**
1. <task>: <what it needs> (<skill or tool>)

**Blocked:** <items that cannot move and why (BLOCKED lines, failed agents), or omit>
**Waiting on you:** <decisions or steps only the owner can do, or "nothing">

Rules:

• Caps: at most 8 items under Done (merge older ones into one line, such as "earlier: 3 setup tasks"), 5 under Running now (then "plus N more"), 5 under Next, 3 under Blocked. About 250 words in all, and the word cap wins: cut older Done lines first, then Next items 4 and 5, then detail inside lines.
• Next is in plan order: item 1 is the first open task whose prerequisites are all done; items 2 to 5 are the open tasks after it, even when they wait on item 1. Only tasks waiting on the owner, or on a failed or BLOCKED item, go under Blocked or Waiting on you.
• Mark at most one item as in progress: the one the main session is on now.
• In a plan, say which part of the plan each Next item belongs to, in words ("part 3 of 5: the AI judge").
• Bugs or problems found along the way with no fix, ticket or Done line yet go in one Done line ("found along the way: ...").