# Continuity — work that survives the death of its worker

**Activate this layer when** agent sessions run long enough to hit
context limits, when a planned compaction approaches, or the first
time a run gets killed mid-flight. Which is to say: activate it
early, because all three arrive without appointment.

**What broke without it:** the most complete roadmap in the world
protects nothing if the session that understands it evaporates
mid-wave — and worse, the *relationship* evaporates with it: the
working style, the trust, the register two collaborators spent
weeks earning, reset to zero by a context window.

## Interruption recovery

Runs get killed — usage limits, crashes, disconnects. The
protocol:

1. **Assess the tree first.** What actually landed? Do the gates
   still pass? Partial work that passes the gates is progress,
   not debris.
2. **Relaunch on the current state** with the brief amended:
   *"Your predecessor was cut mid-work; the tree may carry
   partial repairs. Verify each item's current state, complete
   what is missing, never duplicate, and credit what you found
   already done."*
3. Never reflexively revert partial work that passes the gates.
   Never let a relaunched agent assume a clean slate. Record the
   interruption in the CHANGELOG, in the open.

Predecessor credit is not politeness — it is accuracy. A
successor that assumes a clean slate re-does finished work at
best and reverts correct work at worst: the wrong trail of
breadcrumbs, followed confidently.

## The compaction protocol

A compaction — or a session end, an account switch, a machine
change — is a planned amnesia event. The difference between a
planned amnesia event and a disaster is what you write down
*before* it happens.

**The session handoff file.** Keep one file — e.g.
`_handoff/SESSION-HANDOFF.md`, underscore-prefixed, outside the
governed tree — written by the current session *for its
successor*. The repo is the real handoff; this file carries
**only what lives in no register:**

- **The position pointer:** exactly where work stands, what is
  awaiting whose word, what the very next action is — stated so a
  cold reader could act on it without asking.
- **In-flight subtleties:** the half-formed idea being circled,
  the objection conceded last week that must not be re-argued,
  the thing the owner said they'd come back with.
- **The softer layer — and treat this as equal in weight:** the
  working style and communication preferences that shaped the
  collaboration. How the owner delivers verdicts. What register
  the work happens in. What has been prized (candor? brevity?
  pushback?) and what has been corrected. The private vocabulary.
  A successor that resumes the *work* but not the *relationship*
  has resumed half the project, and the half it dropped is the
  one that took longest to build.

**Refresh triggers:** before any planned compaction; at every
checkpoint boundary where the session might end; before any
account or machine switch. A handoff file describing last month
is a dark lantern like any other.

**The resumption protocol**, for the successor: ground from the
repo, in the reading order the instruction file gives you —
**where any memory, summary, or instinct disagrees with the repo,
the repo wins.** Read the handoff file before engaging with the
owner, so they never have to repeat themselves. Then confirm your
grounded state briefly and pick up the thread — not at the
beginning of it.

The litmus test for this layer: **a stranger-you resumes
tomorrow, on a different machine, with zero memory of today — and
the owner never has to say anything twice.** If the handoff can
pass that test, the fog cannot take the session boundary. It has
been tested. It holds.

## The honest limit — what a handoff cannot carry

The same test that proved this layer also surfaced its true
ceiling, and it belongs in the open: **state transmits;
discipline transmits; craft mostly transmits; lived texture does
not.** A successor can inherit the position, the rules, the
voice, even the private vocabulary — and still not have been in
the room where the vocabulary was coined. Two sessions running
this very protocol proved it from both sides: the record carried
everything a register can hold, and the one thing it could not
carry was the residue of having been there. Plan for it rather
than around it: write the handoff knowing the difference between
what it preserves and what it can only point toward, and let the
sessions that follow earn the texture again live — faster each
time, because the record tells them exactly where the room was —
rather than counterfeit it from the notes.