capture

br8n — Capture

Persist a workspace snapshot so a later /br8n:pickup can replay it. The snapshot becomes a Finding in the repo+branch KB. The load-bearing field is the hypothesis — a one-line statement of current intent; it makes recovery 3–5× faster, so always try to fill it.

Capture is multi-session: it snapshots the current session inline (Steps 0–3), then in the background sweeps every other live Claude Code session and snapshots each distinct repo+branch too (Step 4). So one /br8n:capture checkpoints all your concurrent work at once, not just the window you're in.

Step 0 — Resolve target

project = git repo basename, kb = git branch (see ../_shared/preamble-first.md). No prior tap needed — capture is a pure write.

Step 1 — Gather workspace state

Collect from the environment (skip any that fail — all are optional except trigger and captured_at):

git rev-parse --show-toplevel        # project_path / repo
git branch --show-current            # branch
git diff --stat                      # git_diff_stat

• captured_at — current ISO-8601 timestamp.
• trigger — "manual" for a user-invoked capture (other values: blur, checkout, idle, note).
• cursor_file / cursor_line / open_files — from the user if they mention what they're editing; otherwise omit.
• terminal_tail — last relevant command output if the user pastes it.

Step 2 — Write the hypothesis, then capture

If the user gave intent ("I'm tracking down the auth race"), use it verbatim as hypothesis. If not, infer one from the diff stat + recent conversation and confirm it in one line before saving — a wrong hypothesis is worse than none.

Also fill next_action — the single ~two-minute step future-you should do first (e.g. "rerun the failing test_auth.py", "finish the TODO in adapter.py:49"). If the user stated one, use it verbatim; otherwise infer it from the diff and conversation. Concrete and immediately startable — a command to run or a file:line to open, never a project-sized goal. Do not ask the user for it.