diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..ec19cd8
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,106 @@
+# AGENTS.md
+
+## What this is
+
+`auto-reverse` is a **conversational CLI** where an LLM agent and a human collaborate to reverse-engineer a website's API. The agent drives a headed browser; an embedded mitmproxy captures all traffic; endpoints are documented into OpenAPI + markdown in real time.
+
+The critical design point: **the agent and user work together in a loop**. The agent doesn't run autonomously — it pursues intent, reports findings, and asks the user what to do next. The user steers, redirects, or takes over the browser at any time.
+
+## Agent-user interaction model
+
+The REPL loop in `repl.py` is the interface. Non-`/` lines go to the Claude agent (`agent.py`). The agent has three tool groups:
+
+- **`browser_*`** (navigate, click, type, snapshot) — act on the page, return compact snapshots (url + title + up to 40 interactive elements, not raw HTML)
+- **`flows_search`** — query the FlowStore for captured endpoints
+- **`doc_document`** — enrich an endpoint with summary/description/tag, triggers spec rewrite
+
+Each agent turn runs up to 25 tool iterations (`MAX_ITERATIONS` in `agent.py`). The agent should: pursue intent → inspect traffic → enrich endpoints → summarize → ask user what next.
+
+The system prompt is in `cli.py:27-34`. It's intentionally brief — the tools shape behavior more than the prompt.
+
+## Commands
+
+```bash
+uv sync                                  # install deps
+uv run playwright install chromium       # one-time browser install
+uv run pytest -q                         # run all tests
+uv run pytest tests/test_store.py -q     # run one test file
+uv run pytest -k test_name -q            # run one test
+uv run auto-reverse https://example.com  # run the tool
+```
+
+Lint/typecheck (from nix shell or install separately):
+```bash
+ruff check src/
+pyright src/
+```
+
+## Architecture (read before editing)
+
+Single Python process, four concurrent roles:
+
+| Thread | Role | Key files |
+|--------|------|-----------|
+| Main | REPL + Claude tool-use loop | `repl.py`, `agent.py` |
+| Main (startup) | CLI arg parsing, wiring | `cli.py`, `config.py` |
+| mitmproxy thread | Embedded DumpMaster, capture addon | `proxy.py` |
+| Doc Worker thread(s) | Schema inference + LLM enrichment on new signatures | `doc/engine.py`, `doc/schema.py` |
+| Node subprocess | Playwright browser | `browser.py` |
+
+Data flow: user intent → agent acts via browser → proxy captures → FlowStore deduplicates by signature → DocWorker writes openapi.yaml + API.md → agent reads flows and reports back.
+
+**Signature dedup**: `(method, host, path_template, status_class)`. Path templating collapses numeric IDs, UUIDs, hex tokens to `{id}`. Same signature = no new LLM call, just merges samples. LLM cost scales with distinct endpoints, not request volume.
+
+**Scope filter** (`store.py`): same-host XHR/fetch only. Static assets and analytics hosts (google-analytics, segment, stripe-js, sentry, doubleclick) dropped by default. Everything still goes to raw archive.
+
+## Testing quirks
+
+- `tests/fixture_site.py` — stdlib `ThreadingHTTPServer` with JSON endpoints. Used by `conftest.py` as `fixture_site` fixture. No external deps.
+- Browser tests (`test_browser.py`, `test_e2e_smoke.py`) **skip** if Playwright/Chromium isn't installed. This is expected in CI without browsers.
+- Agent tests (`test_agent.py`) use a **mocked Anthropic client** with scripted responses. Never call the real API in tests.
+- `test_proxy.py` tests `flow_from_mitm` with `SimpleNamespace` doubles — no real proxy needed.
+- `test_e2e_smoke.py` spins up fixture site + real proxy + real browser. Skips if either is unavailable.
+- `asyncio_mode = "auto"` in pytest config — write `async def test_*` directly, no decorator needed.
+
+## Runtime constraint
+
+**Standard CPython 3.14, not free-threaded.** mitmproxy's `aioquic` and `mitmproxy-rs` deps only ship `abi3` wheels that the free-threaded build rejects. The workload is I/O-bound anyway, so this doesn't matter. Don't target `3.14t`.
+
+## Nix environment
+
+`flake.nix` provides a dev shell with uv, ruff, pyright, google-chrome. The shell hook sets:
+- `PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH` — points to system Chrome
+- `PLAYWRIGHT_SKIP_VALIDATE_HOST_REQUIREMENTS=true`
+- `UV_PYTHON_PREFERENCE=managed`
+
+Use `direnv allow` or `nix develop` to enter the shell.
+
+## Outputs
+
+All outputs go to `./auto-reverse-out/<host>-<timestamp>/`:
+- `openapi.yaml` — incremental OpenAPI 3.1 spec
+- `API.md` — human-readable endpoint docs
+- `archive.log` — raw captured traffic (one line per flow)
+- `client/` — optional typed httpx client (with `--gen-client`)
+
+## REPL commands (local, not sent to LLM)
+
+```
+<text>      state intent in natural language
+/flows [q]  list/search discovered endpoints
+/spec       show spec path + endpoint count
+/help       help
+/quit       exit
+```
+
+The design spec describes `/take`, `/stop`, `/save` — these are planned but not yet implemented in `repl.py`.
+
+## Key files to read first
+
+- `cli.py` — entrypoint, wiring, system prompt, all startup logic
+- `agent.py` — Claude tool-use loop (small, read the whole thing)
+- `store.py` — FlowStore, ScopeFilter, path templating
+- `tools/__init__.py` — tool registry assembly
+- `doc/engine.py` — DocEngine, the event-driven doc writer
+- `docs/superpowers/specs/2026-05-31-auto-reverse-design.md` — full design spec
+- `docs/superpowers/plans/2026-05-31-auto-reverse.md` — implementation plan with task breakdown