sebastienrousseau/agtmls/handoff

Handoff

Trigger. Work is unfinished and continuity is about to break: context is running out, the session is ending, someone else is taking over, you are parking a branch, or a human asked where are we with this?

Reference material (the template, worked good/bad handoffs, and what to

capture for long-running or multi-agent work): see reference.md in this

directory.

The rule

**A handoff is judged by one thing: can the reader take the next action

without asking you a question?**

Not by how much it covers. A handoff that narrates three hours of work and leaves the reader unsure which command to run has failed; four lines that name the next command have succeeded.

The default failure is writing a diary instead of a state report. "First I looked at the parser, then I tried adding a guard, then I realised…" is your experience. The reader needs where things are, not how they got there — except where the journey contains traps (see below).

What a handoff must contain

Six parts. Omit any and the reader has to reconstruct it.

1. Goal

One sentence, in terms of the observable outcome. Not the ticket title — what "done" looks like.

2. State of the tree

The part people forget, and the one that costs the most when missing:

• Branch name, and what it is based on.
• Uncommitted changes — and whether they are coherent or mid-edit.
• Anything half-applied: a partial migration, a rename stopped halfway, a

dependency bumped but not rebuilt.

• Anything stashed, and what is in it.

If the tree is in a state that will not build, say so first. A reader who discovers that by running the tests has already lost the thread.

3. What is done — with evidence

Not "the parser is done" but "the parser change is done: pytest tests/test_parser.py reports 14 passed; it reported 2 failed before." A claim without evidence has to be re-verified by the reader, which means it was not really handed over. See verification-before-completion.

4. The next action

A command or an edit, not a direction. The single highest-value line in the document.

5. Dead ends and traps

The most valuable section, and the one always cut for length. Anything that cost you more than a few minutes and would cost the reader the same:

Do not try --parallel on the test runner: it passes locally and hangs in

CI. Spent 40 minutes on it; the fixtures share a temp dir.

Without this, your successor repeats your mistakes at full price. It is the only part of the document that cannot be reconstructed from the code.

6. Open questions

Decisions that are not yours to make, with who owns them and what is blocked until they are answered. Distinguish blocking from nice to know.

Anti-patterns

• The diary. Chronological narration of your process.
• "Should be straightforward." Either it is, and you should have finished

it, or it is not, and you have mislabelled the hard part.

• Unstated assumptions. You chose an approach for a reason; if it is not

written down, the reader will re-litigate or silently reverse it.

• Omitting the failures. A handoff listing only what worked implies the

obvious approaches are untried.

• Stale evidence. "Tests pass" from before the last three edits. Re-run,

or say when it was last observed.

• Linking without stating. "See the thread" makes the reader do the

synthesis you were supposed to do.

When not to use

• The work is finished — that is a completion report, not a handoff.
• Nothing is in flight and there is no tree state to describe.
• The project has its own status or session-continuity format; use that.