asalazarvaldiviaocg/handoff

/handoff is a Claude Code skill for when a session has to end mid-task. It lets the work already running finish, commits locally, and writes one HANDOFF.md that a cold session can resume from without guessing. Then it stops. It runs in any agent that reads SKILL.md.
Without it, hitting the usage limit mid-task means tomorrow's session starts from zero: it re-reads the code, re-runs the tests and retries what already failed yesterday.
/plugin marketplace add asalazarvaldiviaocg/handoff
/plugin install handoff@handoff
## What happens when you run it
Type /handoff, or say "wrap up", "let's stop here" or "I'm out of usage".
| Phase | What it does | What it refuses to do |
| 0. Freeze | Tells you in one line what is still running and how long it needs. Lets that unit finish: a code review delivers its findings, a workflow its diffs, a test suite its numbers. | Start the next bug, the next review round, or "one quick fix". |
| 1. Evidence | Reads the real state: branch, HEAD, dirty files, unpushed commits, secrets in changes and commits, the latest test results. | Trust its memory of the conversation. |
| 2. Commit | Commits the files it edited this session, by name, in every repo it touched. Red tests become an honest WIP: commit. | Deploy. Push, unless you asked in that turn. Touch files it didn't edit. Stage a .env. |
| 3. Write | Writes HANDOFF.md with five sections and runs a linter on it. | Hand you a file with empty sections, {{template}} holes or placeholders. |
| 4. Stop | Tells you where the file is, what you need to decide, and the exact phrase to resume. | Keep working after that message. |
## The five sections
Every handoff has the same shape, written in your language:
| # | Section | It answers |
| 1 | Goal | What you asked for, quoted, and the rules you set (no push, no deploy, review-round cap). |
| 2 | Where everything ended | Per repo: branch, HEAD, clean or not, pushed or not. Tests with numbers. What was verified versus what an agent only claimed. |
| 3 | What changed | File by file, with the commit. What was deliberately left alone. |
| 4 | What we tried that failed | Each dead end with its real cause and what to do instead. |
| 5 | Exact next steps | Numbered, with full commands and absolute paths. Your open decisions first, as A/B. |
Section 4 is why this exists. It lists each dead end with its cause so the next session doesn't retry it. From the example handoff:
> False premise about the function timeout. Took "10 s per request" from an old forum answer, so I cut the provider call timeout to 8 s → legitimate slow charges got retried → that caused F2. The official platform docs say 60 s. Instead: read the official limits page before designing around a limit.
## Pick up where you left off
Open a session in the same folder and say:
continue from HANDOFF.md
The agent checks the file against reality first (does HEAD still match section 2? is the tree clean?), reads section 4 before touching code, asks your pending A/B questions, then starts at step 1. With the plugin installed, a new session tells you a handoff is waiting before you ask:
Handoff waiting: ledgerline: idempotent payment webhook (STOPPED AT USAGE LIMIT, 9 h ago). To pick it up, say: continue from HANDOFF.md