---
name: work-preservation
description: Protect user and agent work through durable artifacts, reversible edits, and concurrency awareness.
version: 1.0.0
updated: 2026-08-14
---

# Work Preservation

- Inspect status before editing and treat unknown changes as someone else's work.
- Own a clear file scope; never revert unrelated changes or clobber a concurrent wave's commit.
- Keep plans, prompts, evidence, reports, and produced assets in the repository or designated durable storage.
- Prefer deprecation, archival, migration, and soft deletion over erasure.
- Label known-bad or stale artifacts `bad-stale-*` and preserve them instead of deleting them.
- Keep at least 1 GB free on both the local machine and every remote machine in use; check before
  and after large builds or transfers and stop before crossing the floor.
- Before replacing content, understand what would be lost and retain a recovery path.
- Use focused commits or equivalent durable checkpoints when the repository workflow permits.
- Make resumable jobs skip completed items and preserve per-item results.
- Do not use ephemeral storage as the only home of valuable work.

Completion includes a focused commit, synchronization at the deploy source when in scope, and a
durable handoff that another person can inspect and resume. Publishing remains a separate gate.

Preserve provenance as append-only source facts and linked derivations. Missing capture time, geography, identity, or attribution stays unknown; never place media on a map without real coordinates, turn absent samples into zeroes, or let derived fields overwrite originals.

For long or leased work, persist the immutable first arrival time, exact session/run ID, heartbeat deltas, last durable output, and safe resume point. Maintain a compact tail/index for append-only boards. Takeover requires evidence of failure or resource emergency; age alone is insufficient.