diff --git a/.claude/skills/amy-expert/SKILL.md b/.claude/skills/amy-expert/SKILL.md
index bf423a0250..1677296d17 100644
--- a/.claude/skills/amy-expert/SKILL.md
+++ b/.claude/skills/amy-expert/SKILL.md
@@ -104,9 +104,15 @@ Wire-up checklist:
 4. Extend `printUsage()` in `Main.kt`.
 5. Add the row to `cli/README.md`'s command table.
 6. Update `cli/ROADMAP.md` — move the row from 🆕 / 📦 to ✅.
+7. If the verb changes observable wire behaviour (a new event kind,
+   a new relay-routing rule, a new JSON discriminator), add a case
+   in the appropriate harness under `cli/tests/` — `cli/tests/marmot/`
+   for MLS flows, `cli/tests/dm/` for NIP-17, or a new sibling suite
+   if it's neither.
 
 If you change output shape: note it in the commit message, bump the
-example in `README.md`, update any interop fixtures.
+example in `README.md`, update any interop fixtures under
+`cli/tests/`.
 
 ## Where things live
 
@@ -116,6 +122,11 @@ cli/
 ├── DEVELOPMENT.md         # touch-the-code: architecture, conventions, testing
 ├── ROADMAP.md             # parity matrix + ordered milestones
 ├── plans/                 # dated design docs (use for new subsystems)
+├── tests/                 # end-to-end shell harnesses against a local relay
+│   ├── lib.sh             # shared logging + result tracking
+│   ├── headless/          # shared amy wrappers + assertions
+│   ├── marmot/            # MLS group-messaging interop (vs whitenoise-rs)
+│   └── dm/                # NIP-17 DM interop (two amy clients)
 └── src/main/kotlin/…/cli/
     ├── Main.kt            # argv dispatch
     ├── Args.kt            # flag parser
diff --git a/cli/DEVELOPMENT.md b/cli/DEVELOPMENT.md
index 2d5b990c2a..6b855761aa 100644
--- a/cli/DEVELOPMENT.md
+++ b/cli/DEVELOPMENT.md
@@ -197,8 +197,8 @@ Amy-specific layer still needs its own coverage:
 | Error / exit-code contract (bad args → 2, await timeout → 124, runtime → 1) | Table-driven tests invoking `main(argv)` with captured stdout/stderr. |
 | JSON output shape (each command's keys and types) | Snapshot tests: run a command against a throwaway data-dir, assert the JSON matches a golden file. |
 | File layout on disk (`identity.json`, `relays.json`, `groups/*.mls`, `keypackages.bundle`) | Structural assertions after a command sequence. |
-| Round-trip between two data-dirs on a local relay | End-to-end shell scripts under `cli/src/test/resources/scripts/`. Spin up `nostr-rs-relay`, run Alice + Bob, assert await verbs resolve. |
-| Interop with other clients | External harness consumes Amy as a binary; out of scope here but the JSON contract is what keeps it stable. |
+| Round-trip between two data-dirs on a local relay | End-to-end shell harnesses under `cli/tests/`. Each harness spins up a local `nostr-rs-relay`, bootstraps two or more fresh identities in their own `--data-dir`s, and drives a scenario via `amy` (+ `wn` for Marmot interop against whitenoise-rs). Today there are two suites: `cli/tests/marmot/` (13 MLS scenarios vs whitenoise-rs) and `cli/tests/dm/` (NIP-17 DM round-trips between two `amy` clients). |
+| Interop with other clients | Covered by `cli/tests/marmot/marmot-interop-headless.sh` (drives Amy against whitenoise-rs `wn`/`wnd`). Add new scenarios there or start a new sibling under `cli/tests/`. |
 
 **What not to test here:** event signing, filter assembly, MLS
 correctness, NIP-44 encryption. Those belong in `quartz`/`commons`.
@@ -207,6 +207,11 @@ If an Amy bug can only be caught here, it's a contract violation
 
 **Interop-test script template:**
 
+The canonical examples live under `cli/tests/` — read
+[`cli/tests/README.md`](./tests/README.md) for the layout, then
+crib from `cli/tests/dm/tests-dm.sh` or `cli/tests/marmot/tests-create.sh`.
+At the byte-banging level, a minimal round-trip looks like:
+
 ```bash
 set -euo pipefail
 TMP=$(mktemp -d)
diff --git a/cli/README.md b/cli/README.md
index a1077f37e3..abc2141470 100644
--- a/cli/README.md
+++ b/cli/README.md
@@ -98,7 +98,10 @@ GID=$(amy --data-dir ./alice marmot group create --name "Test" | jq -r .group_id
 ```
 
 For an interop-test script template, see
-[DEVELOPMENT.md § Testing](./DEVELOPMENT.md#testing).
+[DEVELOPMENT.md § Testing](./DEVELOPMENT.md#testing). The runnable
+harnesses live under [`cli/tests/`](./tests/README.md) —
+`cli/tests/marmot/` for MLS group messaging vs whitenoise-rs,
+`cli/tests/dm/` for NIP-17 DMs between two `amy` clients.
 
 ---