diff --git a/.claude/AGENTS.md b/.claude/AGENTS.md
index 6fb4ae6..99b15cc 100644
--- a/.claude/AGENTS.md
+++ b/.claude/AGENTS.md
@@ -125,6 +125,7 @@ Do not run full validation suites after every small edit by habit. Run the narro
   | Editing AGENTS.md, .agents/skills/, docs/ai/ | writing-docs.md |
   | Maintaining or troubleshooting ask-intern | ask-intern.md |
   | Operating or changing the learning system | learning-system.md |
+  | Managing a multi-hour or multi-session autonomous workstream | long-running-workstreams.md |
 - Re-read the hard rules before implementation.
 
 ## Session Tools
diff --git a/docs/ai/long-running-workstreams.md b/docs/ai/long-running-workstreams.md
new file mode 100644
index 0000000..64576cc
--- /dev/null
+++ b/docs/ai/long-running-workstreams.md
@@ -0,0 +1,46 @@
+> **IMPORTANT: Before reading, check if you already read this file earlier in this session. If yes, skip the read and announce "Context already loaded: long-running-workstreams.md (re-using from earlier)". If no, read it and announce "Context loaded: long-running-workstreams.md".**
+
+# Long-Running Workstreams
+
+Read this when managing a multi-hour or multi-session autonomous workstream where the agent must resume after compaction or hand off context across runs.
+
+## Contract
+
+- Maintain a `scratch.md` control surface alongside any evidence or triage files. Evidence explains what happened; scratch state tells the next agent what to do without rediscovering the whole thread.
+- Keep the "Current Truth" section current and sharp. Archaeology grows below it and stays compressed.
+- Apply compaction-time maintenance every time context is compacted or a session resumes: update "Current Truth", move completed next-steps into a short archaeology entry, delete detail now covered by links or evidence files, and compact the top section before continuing.
+- Do not let the scratch file grow into a second transcript. Garbage-collect it on every re-read after compaction.
+
+## Scratch File Structure
+
+Use this shape for `scratch.md`:
+
+```
+## Current Truth (update at every compaction)
+<3-5 line summary of the actual state right now>
+
+## Next Steps
+- [ ] <next immediate action>
+- [ ] <next queued action>
+
+## Archaeology
+### <YYYY-MM-DD> - <brief session note>
+<single short paragraph; delete detail covered by evidence files or issue trackers>
+```
+
+Keep "Current Truth" and "Next Steps" together at the top; archaeology grows at the bottom. If "Next Steps" is empty, the workstream is done.
+
+## Canonical Files
+
+- `docs/specs/<feature>/scratch.md` — primary control surface for a feature workstream.
+- Separate evidence files (`triage.md`, `shakedown-log.md`, run artifacts) hold full detail; scratch summarizes.
+
+## Verification
+
+After any compaction or resume, check that "Current Truth" reflects the actual current state, not an earlier snapshot. A stale "Current Truth" is the sign that the compaction maintenance step was skipped.
+
+## Notes
+
+- Do not write the same content to both scratch.md and evidence files. Scratch summarizes; evidence files contain full detail.
+- This pattern is most valuable for: shakedown loops, experiment queues, autonomous workstreams, and any session expected to compact at least once.
+- If the workstream has a living triage file, keep them separate: triage captures what was found; scratch captures what to do next.