diff --git a/.claude/AGENTS.md b/.claude/AGENTS.md
index 6fb4ae6..9acbb23 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 |
+  | Starting or continuing a long-running 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..d3a52a3
--- /dev/null
+++ b/docs/ai/long-running-workstreams.md
@@ -0,0 +1,33 @@
+> **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 before starting or continuing any multi-hour or multi-session autonomous workstream.
+
+## Contract
+
+- Maintain a `scratch.md` as the single control surface for any long-running autonomous workstream.
+- The scratch file has three permanent sections:
+  1. **Current Truth For Compaction** — the minimal state a fresh agent needs after context compaction: what is true now, what already happened, where we are.
+  2. **Next Steps** — an always-current ordered queue. The only required intervention should be "what's next? write it down, then do it."
+  3. **Archaeology** — older per-run logs and completed steps, linked not restated.
+- Treat the top section as the source of truth; Archaeology exists so Current Truth can stay short.
+- Route durable learnings through `/learn`; scratch.md is not a learning backlog.
+- Place the scratch file in the project's coordination directory (e.g. `docs/specs/<milestone>/scratch.md`), not a generic tmp location.
+
+## Compaction Cleanup
+
+After each context compaction, before continuing work:
+
+1. Update **Current Truth For Compaction** to reflect what is now true.
+2. Move completed Next Steps items into a short dated Archaeology entry.
+3. Delete duplicated detail from Current Truth that now lives in Archaeology.
+4. Compact the top section until it is shorter than before compaction.
+
+A scratch file that skips compaction cleanup grows into a transcript. Future agents will spend as much effort reading it as they would starting over.
+
+## Notes
+
+- Distinguish scratch from durable evidence: triage files, run logs, and test traces live in separate files; scratch.md holds only state and queue.
+- If the workstream needs a long-range plan (many tasks, dependencies), keep a separate plan file and point to it from Next Steps rather than embedding it inline.
+- Promote any durable principle discovered during the workstream through `/learn` so it survives into the learning store; don't leave it buried in Archaeology.