From 54baf579b30b072f5fca34f5cbe6a39c5db6440d Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 21:35:10 +0100
Subject: [PATCH 1/9] build: declare prompt-toolkit and markdown-it-py, which
 the session imports

---
 pyproject.toml | 2 ++
 uv.lock        | 4 ++++
 2 files changed, 6 insertions(+)

diff --git a/pyproject.toml b/pyproject.toml
index a9ebdb1..c92a882 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -47,6 +47,8 @@ dependencies = [
     "click>=8.3.3",
     "claude-agent-sdk>=0.2.174",
     "rich>=13.0.0",
+    "markdown-it-py>=2.2.0",
+    "prompt-toolkit>=3.0.40",
     "questionary>=2.0.0",
     "watchdog>=3.0.0",
     "setproctitle>=1.3.0",
diff --git a/uv.lock b/uv.lock
index e39d339..5b58e6e 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1732,8 +1732,10 @@ dependencies = [
     { name = "click" },
     { name = "cryptography" },
     { name = "httpx2" },
+    { name = "markdown-it-py" },
     { name = "multidict" },
     { name = "opentelemetry-api" },
+    { name = "prompt-toolkit" },
     { name = "pygments" },
     { name = "pyjwt" },
     { name = "python-multipart" },
@@ -1780,11 +1782,13 @@ requires-dist = [
     { name = "cryptography", specifier = ">=50.0.0" },
     { name = "github-copilot-sdk", marker = "extra == 'copilot'", specifier = ">=1.0.26,<2" },
     { name = "httpx2", specifier = ">=2.12.0" },
+    { name = "markdown-it-py", specifier = ">=2.2.0" },
     { name = "markdownify", marker = "extra == 'collector'", specifier = ">=0.14.1" },
     { name = "multidict", specifier = ">=6.9.1" },
     { name = "opentelemetry-api", specifier = ">=1.45.1" },
     { name = "playwright", marker = "extra == 'manual'", specifier = ">=1.40.0" },
     { name = "playwright-stealth", marker = "extra == 'manual'", specifier = ">=1.0.0" },
+    { name = "prompt-toolkit", specifier = ">=3.0.40" },
     { name = "pygments", specifier = ">=2.20.0" },
     { name = "pyjwt", specifier = ">=2.15.0" },
     { name = "python-multipart", specifier = ">=0.0.31" },

From 1981e8735ed14ce7b78f72e65fc97b0b3e0a10f3 Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 21:35:10 +0100
Subject: [PATCH 2/9] fix(cli): accept /quit and /commands in the session

---
 src/reverse_api/repl/app.py | 2 ++
 tests/test_repl.py          | 9 +++++++++
 2 files changed, 11 insertions(+)

diff --git a/src/reverse_api/repl/app.py b/src/reverse_api/repl/app.py
index fcb3bbb..b0426df 100644
--- a/src/reverse_api/repl/app.py
+++ b/src/reverse_api/repl/app.py
@@ -75,6 +75,7 @@
     "/help": "keys and commands",
     "/exit": "quit rae",
 }
+ALIASES = {"/quit": "/exit", "/commands": "/help"}
 HISTORY_RUNS = 20
 # Commands that open a picker are echoed once a choice shows something, not when the picker opens.
 PICKER_COMMANDS = {"/model", "/effort", "/mode", "/history", "/messages", "/resume"}
@@ -280,6 +281,7 @@ def cycle_effort(self) -> None:
     def run_command(self, text: str) -> None:
         cmd, _, arg = text.partition(" ")
         cmd, arg = cmd.lower(), arg.strip()
+        cmd = ALIASES.get(cmd, cmd)
         st, view = self.st, self.view
         if cmd not in PICKER_COMMANDS | {"/exit", "/clear"}:
             view.command(cmd, arg)
diff --git a/tests/test_repl.py b/tests/test_repl.py
index c007685..ff3f87a 100644
--- a/tests/test_repl.py
+++ b/tests/test_repl.py
@@ -417,6 +417,15 @@ def test_commands_are_echoed(self, app_session):
         session.run_command("/help")
         assert "› /help" in plain(session.transcript)
 
+    def test_quit_and_commands_are_aliases(self, app_session, monkeypatch):
+        session, _history = app_session
+        quits = []
+        monkeypatch.setattr(session, "quit", lambda: quits.append(True))
+        session.run_command("/quit")
+        assert quits == [True]
+        session.run_command("/commands")
+        assert "› /help" in plain(session.transcript) and "unknown command" not in plain(session.transcript)
+
     def test_an_exact_command_wins_over_longer_ones(self, app_session):
         session, _history = app_session
         session.buf.text = "/mode"

From d9d464fb6cb49b02f5cd3b43d2b5bc4bfbf05566 Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 21:35:10 +0100
Subject: [PATCH 3/9] docs: describe the full-screen session, the current
 models and effort

---
 README.md                                     |  4 +-
 website/content/docs/cli/commands.mdx         | 62 +++++++++++++------
 website/content/docs/configuration/models.mdx | 38 +++++++++---
 website/content/docs/index.mdx                | 10 +--
 website/content/docs/modes/agent.mdx          | 20 +++++-
 website/content/docs/modes/collector.mdx      |  2 +-
 website/content/docs/modes/manual.mdx         |  6 +-
 website/content/docs/quick-start.mdx          |  5 +-
 8 files changed, 105 insertions(+), 42 deletions(-)

diff --git a/README.md b/README.md
index 8f9c257..bd9bad2 100644
--- a/README.md
+++ b/README.md
@@ -125,14 +125,14 @@ Settings live in `~/.reverse-api/config.json` and can be edited via `/settings`
 
 ## CLI
 
-`reverse-api-engineer` opens a full-screen session; `agent`, `engineer` and `collector` open it too when run in a terminal (without `--json` or `--no-interactive`). Type `/` for the commands:
+`reverse-api-engineer` opens a full-screen session; `agent`, `engineer` and `collector` open it too when run in a terminal (`agent` and `engineer` stay plain with `--json` or `--no-interactive`). Type `/` for the commands:
 - `/model`, `/effort`, `/mode`: switch the model, how hard it thinks, and the mode (also Shift+Tab, ctrl+e).
 - `/settings`: SDK, agent provider, output and sync settings.
 - `/history`: past runs, with their details, messages, and re-engineer or resume.
 - `/messages [run_id]`: what the agent did in a run.
 - `/resume [run_id]`: continue a past run's conversation where it stopped, on any SDK. Works for runs that saved their SDK session: agent and engineer runs from this version on. Collector runs cannot be resumed.
 - `/output` (ctrl+o): everything from the session, unfolded. `/open`: the latest client.
-- `/new`, `/clear`, `/help`, `/exit`.
+- `/new`, `/clear`, `/help` (or `/commands`), `/exit` (or `/quit`).
 
 Scriptable subcommands (pipe to `jq`):
 
diff --git a/website/content/docs/cli/commands.mdx b/website/content/docs/cli/commands.mdx
index 377037b..c9eed51 100644
--- a/website/content/docs/cli/commands.mdx
+++ b/website/content/docs/cli/commands.mdx
@@ -1,32 +1,52 @@
 ---
 title: Slash commands
-description: Commands available inside the interactive CLI.
+description: Commands and keys of the interactive session.
 ---
 
-While in the interactive prompt, the following slash commands are available:
+Running `reverse-api-engineer` opens a full-screen session: the conversation
+scrolls above an input bar fixed to the bottom. Type `/` to open the command
+menu; `Tab` completes and `Enter` runs the highlighted command.
 
-| Command                | What it does                                              |
-|------------------------|-----------------------------------------------------------|
-| `/settings`            | Configure model, agent provider, SDK, output directory.   |
-| `/history`             | View past runs with token usage and cost.                 |
-| `/messages <run_id>`   | View the full message log for a previous run.             |
-| `/help`                | List all commands. Alias: `/commands`.                    |
-| `/exit`                | Quit the CLI.                                             |
+| Command                | What it does                                                        |
+|------------------------|---------------------------------------------------------------------|
+| `/new`                 | End the open run and start a new capture.                           |
+| `/model`               | Choose the model. In collector mode, the collector's own model.     |
+| `/effort`              | How hard the agent thinks, for Claude models (`low` to `max`).      |
+| `/mode`                | Switch between agent, manual, engineer and collector.               |
+| `/settings`            | SDK, agent provider, output directory, language and sync settings.  |
+| `/history`             | Past runs with their cost; open one to resume, re-engineer or read its messages. |
+| `/messages [run_id]`   | What the agent did in a run, step by step.                          |
+| `/resume [run_id]`     | Continue a past run's conversation where it stopped, on any SDK.    |
+| `/output`              | Everything from the session, unfolded.                              |
+| `/open`                | Open the latest generated client.                                   |
+| `/clear`               | Clear the conversation and its spend.                               |
+| `/help`                | Keys and commands. Alias: `/commands`.                              |
+| `/exit`                | Quit. Alias: `/quit`.                                               |
 
-## Mode cycling
+`/resume` works for runs that saved their SDK session, which agent and engineer
+runs do from this release on. Older runs and collector runs cannot be resumed.
 
-Press **`Shift+Tab`** to cycle between modes:
+## Keys
 
-```text
-agent → manual → engineer → collector → agent ...
-```
+| Key                  | What it does                                               |
+|----------------------|------------------------------------------------------------|
+| `Enter`              | Send. While the agent works, the message is queued.        |
+| `Esc`                | Stop the run, or close a menu.                             |
+| `Shift+Tab`          | Cycle the mode: agent → manual → engineer → collector.     |
+| `Ctrl+E`             | Cycle the effort level.                                    |
+| `Ctrl+O`             | Show the full output of folded steps.                      |
+| `PageUp` / `PageDown`| Scroll the conversation. `End` jumps back to the latest.   |
+| `Ctrl+D`             | Quit.                                                      |
 
-The current mode is shown in the prompt. Each mode has its own behavior. See
-[Usage modes](/docs/modes/manual).
+Click a file or URL in the conversation to open it.
+
+When the agent asks a question, its options show under the input: pick one
+with the arrow keys, press `Space` to pick several when the question allows
+it, or type your own answer.
 
 ## Direct subcommands
 
-Outside the interactive CLI, you can use direct subcommands:
+Outside the session, you can use direct subcommands:
 
 | Command       | What it does                                               |
 |---------------|------------------------------------------------------------|
@@ -37,10 +57,16 @@ Outside the interactive CLI, you can use direct subcommands:
 | `show`        | Inspect one run; supports `--json`.                        |
 | `run`         | Execute a generated client with its language toolchain.    |
 
+In a terminal, `agent`, `engineer` and `collector` open the session in their
+mode and start the run, keeping options such as `--headless` and
+`--output-dir` for the prompts you type next. `agent` and `engineer` print
+plain output instead with `--json`, `--json-stream` or `--no-interactive`;
+`collector` does when it runs outside a terminal.
+
 Scripted flags are command-specific: `agent`, `engineer`, and `run` support
 `--json` and `--json-stream`; `list` and `show` support `--json`; and `run`
 also supports `--no-interactive` for script selection and dependency prompts. See
 [Scripted usage](/docs/cli/scripted-usage).
 
-Manual mode is available only inside the interactive REPL because it opens a
+Manual mode is available only inside the interactive session because it opens a
 visible browser for human-driven capture.
diff --git a/website/content/docs/configuration/models.mdx b/website/content/docs/configuration/models.mdx
index f4d2275..2a36b0c 100644
--- a/website/content/docs/configuration/models.mdx
+++ b/website/content/docs/configuration/models.mdx
@@ -7,32 +7,43 @@ description: Choose a model for Claude, OpenCode, Copilot, or Cursor.
 
 | Model         | When to pick it                                                          |
 |---------------|--------------------------------------------------------------------------|
-| **Sonnet 4.6** *(default)* | Balanced cost and quality. Right answer for most captures.    |
-| **Opus 4.6**  | Complex APIs (many endpoints, nested auth, polymorphic responses).       |
-| **Haiku 4.5** | Cheapest and fastest. Good for simple endpoints or quick iterations.     |
+| **Sonnet 5.5** | Balanced cost and quality, faster and cheaper than Sonnet 4.6.          |
+| **Opus 5.5**  | Complex APIs (many endpoints, nested auth, polymorphic responses).       |
+| **Opus 4.6**  | The previous Opus, for comparison or when a run needs it.                |
+| **Haiku 5.5** | Cheapest and fastest. Good for simple endpoints or quick iterations.     |
 
-## Change the model
+Sonnet 4.6 stays the default until you choose another model. It and Haiku 4.5
+are no longer offered in the pickers, but `--model` and `config.json` still
+accept them.
 
-In the CLI:
+## Change the model
 
-```text
-/settings then "Claude Code Model"
-```
+In the session, type `/model`. In collector mode, `/model` sets the
+collector's own model (`collector_model`).
 
 Or pass a model to a direct command:
 
 ```bash
-reverse-api-engineer agent --prompt "capture the jobs api" --model claude-sonnet-4-6
+reverse-api-engineer agent --prompt "capture the jobs api" --model claude-sonnet-5-5
 ```
 
 Or edit `~/.reverse-api/config.json` directly:
 
 ```json
 {
-  "claude_code_model": "claude-opus-4-6"
+  "claude_code_model": "claude-opus-5-5"
 }
 ```
 
+## Effort
+
+Claude models take an effort level, from `low` to `max`: how hard the agent
+thinks before acting. Higher effort is slower and costs more, and helps on hard
+captures. Set it with `/effort` or `Ctrl+E` in the session, or with
+`claude_effort` in the config. Leaving it unset keeps each model's default:
+`medium` for Opus 5.5 and Haiku 5.5, `high` for the others. Haiku 4.5 has no
+effort setting.
+
 ## Config file reference
 
 The full default config:
@@ -43,6 +54,8 @@ The full default config:
   "agent_browser_notes": "",
   "agent_browser_npx_package": "agent-browser@0",
   "claude_code_model": "claude-sonnet-4-6",
+  "claude_setting_sources": [],
+  "claude_effort": null,
   "collector_model": "claude-sonnet-4-6",
   "cursor_model": "composer-2.5",
   "cursor_web_search": true,
@@ -50,6 +63,11 @@ The full default config:
   "copilot_model": "gpt-5",
   "opencode_model": "big-pickle",
   "opencode_provider": "opencode",
+  "opencode_auto_start": true,
+  "opencode_base_url": "http://127.0.0.1:4096",
+  "opencode_npx_package": "opencode-ai@latest",
+  "ollama_auto_start": true,
+  "ollama_base_url": "http://127.0.0.1:11434",
   "output_dir": null,
   "output_language": "python",
   "real_time_sync": true,
diff --git a/website/content/docs/index.mdx b/website/content/docs/index.mdx
index 33914f0..abb3a4c 100644
--- a/website/content/docs/index.mdx
+++ b/website/content/docs/index.mdx
@@ -98,11 +98,13 @@ Use **Anthropic's API directly** (default) or route through **OpenCode**
 running locally. Copilot and Cursor SDK backends are also available when their
 optional dependencies and credentials are configured.
 
-### Interactive CLI with mode cycling
+### Full-screen interactive CLI
 
-Press **`Shift+Tab`** to cycle between modes. Slash commands (`/settings`,
-`/history`, `/messages`) keep configuration and run inspection one keystroke
-away.
+The agent's steps, messages and reasoning stream into one conversation, with
+the input bar always at the bottom. Press **`Shift+Tab`** to cycle between
+modes; slash commands (`/model`, `/effort`, `/history`, `/resume`,
+`/settings`) keep configuration, past runs and their conversations one
+keystroke away. See [Slash commands](/docs/cli/commands).
 
 ### Session history & cost tracking
 
diff --git a/website/content/docs/modes/agent.mdx b/website/content/docs/modes/agent.mdx
index 193ae64..bd53440 100644
--- a/website/content/docs/modes/agent.mdx
+++ b/website/content/docs/modes/agent.mdx
@@ -23,12 +23,28 @@ English; providers such as Playwright MCP or Chrome DevTools MCP route browser a
 
 ```bash
 reverse-api-engineer
-# the REPL starts in agent mode
-> click on the first job listing
+# the session starts in agent mode
+> all open jobs on jobs.ashbyhq.com/openai
 ```
 
+Or start from the subcommand: in a terminal, `agent` opens the same session
+and starts the run right away, with its options kept for the prompts you type
+next.
+
+```bash
+reverse-api-engineer agent --prompt "all open jobs on jobs.ashbyhq.com/openai" --headless
+```
+
+The agent's steps, messages and reasoning stream into the conversation. Type a
+follow-up when it finishes, answer its questions in the input bar, and press
+`Esc` to stop it. `/resume` picks a past run back up where it stopped.
+
 ## Running non-interactively
 
+With `--json`, `--json-stream` or `--no-interactive`, `agent` never opens the
+session: it prints plain output and exits, which is what scripts and other
+agents want.
+
 ```bash
 reverse-api-engineer agent \
   --prompt "capture the public jobs api" \
diff --git a/website/content/docs/modes/collector.mdx b/website/content/docs/modes/collector.mdx
index 8a54878..94c3d2e 100644
--- a/website/content/docs/modes/collector.mdx
+++ b/website/content/docs/modes/collector.mdx
@@ -3,7 +3,7 @@ title: Collector mode
 description: Use an AI agent to collect structured web data as JSON and CSV.
 ---
 
-Collector mode is an interactive REPL mode for data collection tasks. Instead
+Collector mode is a mode of the interactive session for data collection tasks. Instead
 of generating an API client from a HAR, it runs a Claude-powered collector that
 uses web search and fetch tools, writes collected items, and exports structured
 artifacts.
diff --git a/website/content/docs/modes/manual.mdx b/website/content/docs/modes/manual.mdx
index d1c68c6..d1d6db8 100644
--- a/website/content/docs/modes/manual.mdx
+++ b/website/content/docs/modes/manual.mdx
@@ -25,14 +25,14 @@ playwright install chromium
 
 <Steps>
   <Step>
-    Start the interactive REPL:
+    Start the interactive session:
     ```bash
     reverse-api-engineer
     ```
 
     Press **`Shift+Tab`** once from the default agent prompt to switch into
-    manual mode. Manual mode is REPL-only because it requires a human in front
-    of a visible browser.
+    manual mode. Manual mode only runs from the interactive session because it
+    requires a human in front of a visible browser.
   </Step>
   <Step>
     Enter a task description (e.g. *"Fetch Apple job listings"*).
diff --git a/website/content/docs/quick-start.mdx b/website/content/docs/quick-start.mdx
index 01ee80c..cb6ba62 100644
--- a/website/content/docs/quick-start.mdx
+++ b/website/content/docs/quick-start.mdx
@@ -11,8 +11,9 @@ description: From a fresh install to a working API client in under five minutes.
     reverse-api-engineer
     ```
 
-    You land in a minimalist interactive prompt. Cycle between modes with
-    **`Shift+Tab`**:
+    You land in a full-screen session: the conversation above, an input bar
+    fixed at the bottom. Type `/` for the commands, and cycle between modes
+    with **`Shift+Tab`**:
 
     - **agent**: autonomous AI browser agent (default)
     - **manual**: browser capture + AI generation (you drive)

From 457aeca90b094288fad38b32616bc14ed5139b38 Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 22:41:40 +0100
Subject: [PATCH 4/9] fix(cli): keep engineer's --output-dir for later prompts,
 and find commands by alias in the menu

---
 src/reverse_api/cli.py      | 11 +++++++++++
 src/reverse_api/repl/app.py | 17 ++++-------------
 tests/test_repl.py          | 37 ++++++++++++++++++++++++++++++++++++-
 3 files changed, 51 insertions(+), 14 deletions(-)

diff --git a/src/reverse_api/cli.py b/src/reverse_api/cli.py
index 504aacb..00ad0ea 100644
--- a/src/reverse_api/cli.py
+++ b/src/reverse_api/cli.py
@@ -1874,6 +1874,7 @@ def engineer(
             lambda: run_engineer(
                 run_id, prompt=main_prompt, additional_instructions=additional, model=model, output_dir=output_dir, is_fresh=fresh
             ),
+            lambda p, m: engineer_from_prompt(p, m or model, output_dir),
         )
         return
     if not machine_output:
@@ -1922,6 +1923,16 @@ def engineer(
     sys.exit(0 if payload["status"] == "ok" else 1)
 
 
+def engineer_from_prompt(text: str, model: str | None = None, output_dir: str | None = None) -> dict[str, Any] | None:
+    """A prompt in engineer mode: a known run id re-engineers that run; anything else is an instruction for the latest run."""
+    if session_manager.get_run(text):
+        return run_engineer(text, model=model, output_dir=output_dir)
+    latest = session_manager.get_history(limit=1)
+    if not latest:
+        raise RuntimeError("no runs yet · start one in agent or manual mode")
+    return run_engineer(latest[0]["run_id"], model=model, output_dir=output_dir, additional_instructions=text)
+
+
 def run_engineer(
     run_id: str,
     har_path: Path | None = None,
diff --git a/src/reverse_api/repl/app.py b/src/reverse_api/repl/app.py
index b0426df..ed11399 100644
--- a/src/reverse_api/repl/app.py
+++ b/src/reverse_api/repl/app.py
@@ -215,7 +215,9 @@ def palette_active(self) -> bool:
 
     def palette_items(self) -> list[tuple[str, str]]:
         typed = self.buf.text.strip().lower()
-        return sorted(((c, d) for c, d in COMMANDS.items() if c.startswith(typed)), key=lambda item: item[0] != typed)
+        aliased = {target for alias, target in ALIASES.items() if alias.startswith(typed)}
+        items = [(c, d) for c, d in COMMANDS.items() if c.startswith(typed) or c in aliased]
+        return sorted(items, key=lambda item: item[0] != ALIASES.get(typed, typed))
 
     def choose(self) -> None:
         """Enter on an open picker: its highlighted item, or for an agent's question, whatever was typed meanwhile."""
@@ -498,7 +500,7 @@ def start(self, prompt: str, mode: str | None = None, echo: bool = True) -> None
         elif mode == "collector":
             run = lambda: cli.run_collector(prompt=prompt, model=model)  # noqa: E731
         elif mode == "engineer":
-            run = lambda: self.engineer(prompt, model)  # noqa: E731
+            run = lambda: cli.engineer_from_prompt(prompt, model)  # noqa: E731
         else:
             run = lambda: cli.run_agent_capture(prompt=prompt, model=model)  # noqa: E731
         self.launch(prompt, run, mode, echo)
@@ -534,17 +536,6 @@ def check() -> None:
 
             threading.Thread(target=check, name="rae-update-check", daemon=True).start()
 
-    def engineer(self, text: str, model: str | None) -> dict[str, Any] | None:
-        """A known run id re-engineers that run; anything else is an instruction for the latest run."""
-        sm = self.cli.session_manager
-        if sm.get_run(text):
-            return self.cli.run_engineer(text, model=model)
-        latest = sm.get_history(limit=1)
-        if not latest:
-            self.view.error("no runs yet · start one in agent or manual mode")
-            return None
-        return self.cli.run_engineer(latest[0]["run_id"], model=model, additional_instructions=text)
-
     def work(self, fn: Callable[[], Any]) -> None:
         """Run an engine in a worker thread; its UI calls stream into the conversation through the bridge."""
 
diff --git a/tests/test_repl.py b/tests/test_repl.py
index ff3f87a..d0aba50 100644
--- a/tests/test_repl.py
+++ b/tests/test_repl.py
@@ -5,7 +5,7 @@
 import threading
 from concurrent.futures import Future
 from pathlib import Path
-from unittest.mock import AsyncMock, MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock, call, patch
 
 import pytest
 from rich.console import Console
@@ -426,6 +426,13 @@ def test_quit_and_commands_are_aliases(self, app_session, monkeypatch):
         session.run_command("/commands")
         assert "› /help" in plain(session.transcript) and "unknown command" not in plain(session.transcript)
 
+    def test_the_menu_finds_commands_by_alias(self, app_session):
+        session, _history = app_session
+        session.buf.text = "/qu"
+        assert [c for c, _ in session.palette_items()] == ["/exit"]
+        session.buf.text = "/c"
+        assert {c for c, _ in session.palette_items()} == {"/clear", "/help"}
+
     def test_an_exact_command_wins_over_longer_ones(self, app_session):
         session, _history = app_session
         session.buf.text = "/mode"
@@ -879,6 +886,34 @@ def test_agent_in_a_terminal_opens_the_session_with_every_option(self, monkeypat
         }
         assert capture.call_args_list[1].kwargs == {"prompt": "later task", "model": "claude-opus-5-5", "output_dir": None, "headless": True}
 
+    def test_engineer_keeps_its_output_dir_for_later_prompts(self, monkeypatch):
+        from click.testing import CliRunner
+
+        from reverse_api import cli
+
+        monkeypatch.setattr(cli, "_in_terminal", lambda: True)
+        with patch.object(cli, "open_session") as session, patch.object(cli, "engineer_from_prompt") as later:
+            CliRunner().invoke(cli.main, ["engineer", "abc123", "-o", "/out"])
+            runner = session.call_args.args[4]
+            runner("add pagination", None)
+        later.assert_called_once_with("add pagination", None, "/out")
+
+    def test_an_engineer_prompt_is_a_run_id_or_an_instruction_for_the_latest_run(self, monkeypatch):
+        from reverse_api import cli
+
+        history = MagicMock()
+        history.get_run.side_effect = lambda text: {"run_id": text} if text == "abc123" else None
+        history.get_history.return_value = [{"run_id": "latest1"}]
+        monkeypatch.setattr(cli, "session_manager", history)
+        with patch.object(cli, "run_engineer") as engineer:
+            cli.engineer_from_prompt("abc123", "m", "/out")
+            cli.engineer_from_prompt("add pagination", "m", "/out")
+        assert engineer.call_args_list[0] == call("abc123", model="m", output_dir="/out")
+        assert engineer.call_args_list[1] == call("latest1", model="m", output_dir="/out", additional_instructions="add pagination")
+        history.get_history.return_value = []
+        with pytest.raises(RuntimeError, match="no runs yet"):
+            cli.engineer_from_prompt("add pagination")
+
     def test_without_a_terminal_a_prompt_is_required(self, monkeypatch):
         from click.testing import CliRunner
 

From 37c48a8bdcb9712ffedb8bc94713d0c58627a55b Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 22:41:40 +0100
Subject: [PATCH 5/9] docs: say which runs /resume can pick up

---
 website/content/docs/modes/agent.mdx | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/website/content/docs/modes/agent.mdx b/website/content/docs/modes/agent.mdx
index bd53440..f722e74 100644
--- a/website/content/docs/modes/agent.mdx
+++ b/website/content/docs/modes/agent.mdx
@@ -37,7 +37,8 @@ reverse-api-engineer agent --prompt "all open jobs on jobs.ashbyhq.com/openai" -
 
 The agent's steps, messages and reasoning stream into the conversation. Type a
 follow-up when it finishes, answer its questions in the input bar, and press
-`Esc` to stop it. `/resume` picks a past run back up where it stopped.
+`Esc` to stop it. `/resume` picks a past run back up where it stopped, for
+runs that saved their SDK session.
 
 ## Running non-interactively
 

From 1f1e40db5cb4b43d983b44b8b1d2f6c17b3bee59 Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 22:52:12 +0100
Subject: [PATCH 6/9] feat(cli): show the run id under the input bar, and the
 run's folders on the verified step

---
 src/reverse_api/base_engineer.py |  7 +++++++
 src/reverse_api/engineer.py      |  1 +
 src/reverse_api/repl/app.py      |  2 +-
 src/reverse_api/repl/bridge.py   |  8 +++++++-
 src/reverse_api/repl/view.py     |  8 ++++++++
 tests/test_engineer.py           | 11 ++++++++++
 tests/test_repl.py               | 35 ++++++++++++++++++++++++++++++++
 7 files changed, 70 insertions(+), 2 deletions(-)

diff --git a/src/reverse_api/base_engineer.py b/src/reverse_api/base_engineer.py
index 70a7228..0db5a42 100644
--- a/src/reverse_api/base_engineer.py
+++ b/src/reverse_api/base_engineer.py
@@ -97,6 +97,13 @@ def _report_usage(self) -> None:
         if isinstance(self.ui, LedgerUI):
             self.ui.usage(self.usage_metadata)
 
+    def _report_workspace(self) -> None:
+        """Tells the full-screen session where the run's files are, for its verified step."""
+        from .repl.bridge import LedgerUI
+
+        if isinstance(self.ui, LedgerUI):
+            self.ui.workspace(self.scripts_dir, self.local_scripts_dir)
+
     async def _opening_message(self, brief: str) -> str | None:
         """The first message of the run: its brief, or when resuming, the user's next message."""
         message = await self._prompt_follow_up() if self.resume_session else brief
diff --git a/src/reverse_api/engineer.py b/src/reverse_api/engineer.py
index cac2ab2..a630df9 100644
--- a/src/reverse_api/engineer.py
+++ b/src/reverse_api/engineer.py
@@ -127,6 +127,7 @@ async def report_client_verified(args: dict[str, Any]) -> dict[str, Any]:
                     ]
                 }
             already_reported = True
+            self._report_workspace()
             # Deliberately no manual message_store/UI logging here — the
             # generic ToolUseBlock/ToolResultBlock handling in
             # _process_streaming_response already logs every tool call,
diff --git a/src/reverse_api/repl/app.py b/src/reverse_api/repl/app.py
index ed11399..ef077f3 100644
--- a/src/reverse_api/repl/app.py
+++ b/src/reverse_api/repl/app.py
@@ -335,7 +335,7 @@ def run_command(self, text: str) -> None:
                 view.flush_live()
                 self.transcript.clear()
                 st.scroll, st.folded, st.cost, st.tokens, st.step = 0, False, 0.0, 0, 0
-                st.last_client, st.client_name, st.host = None, "", ""
+                st.last_client, st.client_name, st.host, st.run_id = None, "", "", None
                 self.bridge.cost_base = 0.0
         elif cmd == "/settings":
             self.classic(lambda: self.cli.handle_settings(), then=self.reload_settings, pause=False)
diff --git a/src/reverse_api/repl/bridge.py b/src/reverse_api/repl/bridge.py
index ab2c3a7..d680e69 100644
--- a/src/reverse_api/repl/bridge.py
+++ b/src/reverse_api/repl/bridge.py
@@ -104,6 +104,7 @@ def __init__(self, view: LedgerView, invalidate: Callable[[], None], ask_user: C
         self.pending_card: Path | None = None
         self.cost_base = 0.0
         self.verified = False  # the agent called report_client_verified in this run
+        self.workspace: tuple[Path, Path | None] | None = None  # the run's folder and its synced copy
 
     def ui(self) -> LedgerUI:
         return LedgerUI(self)
@@ -203,7 +204,7 @@ def header(self, run_id: str, prompt: str, model: str | None = None, sdk: str |
         st.run_id, st.step, st.tokens = run_id, 0, 0
         st.running, st.waiting, st.verb, st.kind = True, False, "starting", "other"
         self.b.cost_base = st.cost
-        self.b.verified = False
+        self.b.verified, self.b.workspace = False, None
         # The run's own model and SDK; the selection in st.model stays for the next run.
         st.run_model, st.run_sdk = model or st.model, sdk or st.sdk
         claude = st.run_sdk == "claude" and not predates_effort(st.run_model)
@@ -250,6 +251,8 @@ def tool_result(self, tool_name: str, is_error: bool = False, output: str | None
         if name in HIDDEN_TOOLS:
             return
         if name == "report_client_verified" and not is_error:
+            if not self.b.verified and self.b.workspace:
+                self.view.workspace(*self.b.workspace)
             self.b.verified = True  # the card promises a client that ran, so only a confirmed report counts
         text = str(output or "").strip()
         lines = text.splitlines()
@@ -305,6 +308,9 @@ def progress(self, message: str) -> None:
         self.view.notice(message)
         self._changed()
 
+    def workspace(self, folder: Path, copy: Path | None) -> None:
+        self.b.workspace = (folder.resolve(), copy.resolve() if copy else None)
+
     def success(self, script_path: str, local_path: str | None = None) -> None:
         _ = local_path
         path = Path(script_path)
diff --git a/src/reverse_api/repl/view.py b/src/reverse_api/repl/view.py
index cb69a96..434c18c 100644
--- a/src/reverse_api/repl/view.py
+++ b/src/reverse_api/repl/view.py
@@ -482,6 +482,12 @@ def block(rows: list[str], more: int, cut: bool) -> Group:
         # The full copy is only styled if ctrl+o asks for it: tool output can run to thousands of lines.
         self.out(block(lines[:OUTPUT_PREVIEW], more, cut=True), Deferred(lambda: block(lines, 0, cut=False)))
 
+    def workspace(self, folder: Path, copy: Path | None) -> None:
+        """Where a verified client's files are, absolute so a click opens them."""
+        rows = [("in", folder), ("copy", copy)] if copy else [("in", folder)]
+        for label, path in rows:
+            self.out(f"{' ' * INDENT}[{HAIRLINE}]└[/] [{DIM}]{label}[/] [{INK_SOFT}]{link(str(path), path)}[/]")
+
     def code(self, source: str, language: str) -> None:
         """Written code, previewed from its first definition; ctrl+o shows all of it."""
         total = source.count("\n") + 1
@@ -683,6 +689,8 @@ def status(self) -> FormattedText:
             hint = [(f"fg:{DIM}", "type a follow-up · /new starts over" + (" · ⌃o expand " if st.folded else " "))]
         else:
             hint = [(f"fg:{DIM}", "/ commands" + (" · ⌃o expand " if st.folded else " "))]
+        if st.run_id:
+            hint += [(f"fg:{DIM}", f"· run {st.run_id} ")]
         used = sum(len(t) for _, t in left) + sum(len(t) for _, t in hint)
         return FormattedText(left + [("", " " * max(1, width() - used))] + hint)
 
diff --git a/tests/test_engineer.py b/tests/test_engineer.py
index 604bcc8..de51e25 100644
--- a/tests/test_engineer.py
+++ b/tests/test_engineer.py
@@ -909,6 +909,17 @@ async def test_tool_emits_client_executed_event_when_called(self, tmp_path):
         assert len(client_executed) == 1
         assert client_executed[0]["script_path"] == str(eng.scripts_dir / "api_client.py")
 
+    @pytest.mark.asyncio
+    async def test_tool_tells_the_session_where_the_files_are_once(self, tmp_path):
+        eng = self._make_engineer(tmp_path, output_language="python")
+        eng._report_workspace = MagicMock()
+
+        tool = eng._build_verification_tool()
+        await tool.handler({"summary": "first confirmation"})
+        await tool.handler({"summary": "second confirmation"})
+
+        eng._report_workspace.assert_called_once_with()
+
     @pytest.mark.asyncio
     async def test_tool_second_call_in_same_session_does_not_re_emit(self, tmp_path):
         """The tool's own description tells the agent to call it exactly
diff --git a/tests/test_repl.py b/tests/test_repl.py
index d0aba50..c73b66e 100644
--- a/tests/test_repl.py
+++ b/tests/test_repl.py
@@ -120,6 +120,35 @@ def test_no_card_until_the_agent_verified_the_client(self, session, tmp_path):
         ui.usage({"input_tokens": 10, "output_tokens": 5, "estimated_cost_usd": 0.01})
         assert "ready" not in plain(t), "a client path that does not exist gets no card"
 
+    def test_verified_shows_where_the_files_are(self, session):
+        _st, t, _view, b = session
+        ui = LedgerUI(b)
+        ui.header("abc123", "prompt", model="claude-sonnet-5-5")
+        folder, copy = Path("/runs/abc123"), Path("/work/scripts/jobs")
+        ui.workspace(folder, copy)
+        for _ in range(2):
+            ui.tool_start("mcp__verification__report_client_verified", {"summary": "ok"})
+            ui.tool_result("mcp__verification__report_client_verified", output="Recorded")
+        text = plain(t)
+        assert text.count(f"└ in {folder.resolve()}") == 1 and text.count(f"└ copy {copy.resolve()}") == 1
+
+    def test_no_folders_without_a_confirmed_verification(self, session, tmp_path):
+        _st, t, _view, b = session
+        ui = LedgerUI(b)
+        ui.header("abc123", "prompt", model="claude-sonnet-5-5")
+        ui.workspace(tmp_path / "runs" / "abc123", None)
+        ui.tool_result("mcp__verification__report_client_verified", is_error=True, output="failed")
+        assert "└ in" not in plain(t)
+        ui.header("def456", "prompt", model="claude-sonnet-5-5")
+        ui.tool_result("mcp__verification__report_client_verified", output="Recorded")
+        assert "└ in" not in plain(t), "a new run forgets the last run's folder"
+
+    def test_the_status_row_shows_the_run_id(self, session):
+        _st, _t, view, b = session
+        assert "run " not in "".join(text for _, text in view.status())
+        LedgerUI(b).header("abc123", "prompt", model="claude-sonnet-5-5")
+        assert "run abc123" in "".join(text for _, text in view.status())
+
     def test_folded_output_offers_ctrl_o(self, session):
         st, _t, view, b = session
         assert "expand" not in "".join(text for _, text in view.status())
@@ -450,6 +479,12 @@ def test_pickers_echo_once_something_is_picked(self, app_session):
         pick(session.st, "abc123")
         assert "› /history abc123" in plain(session.transcript)
 
+    def test_clear_forgets_the_run_id(self, app_session):
+        session, _history = app_session
+        session.st.run_id = "abc123"
+        session.run_command("/clear")
+        assert session.st.run_id is None
+
     def test_clear_starts_over_with_no_usage(self, app_session):
         session, _history = app_session
         st = session.st

From a39093528f4fd4d503bf71f7e91773562ad9d08c Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 22:52:12 +0100
Subject: [PATCH 7/9] docs: the run id in the status row and the verified
 step's folders

---
 website/content/docs/cli/commands.mdx | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/website/content/docs/cli/commands.mdx b/website/content/docs/cli/commands.mdx
index c9eed51..393ce6b 100644
--- a/website/content/docs/cli/commands.mdx
+++ b/website/content/docs/cli/commands.mdx
@@ -40,6 +40,11 @@ runs do from this release on. Older runs and collector runs cannot be resumed.
 
 Click a file or URL in the conversation to open it.
 
+The status row under the input shows the run's id, which `/resume`,
+`/messages` and `reverse-api-engineer show` take. Once the agent has verified
+the client, its step lists the run's folder and the synced copy in
+`./scripts`, as full paths.
+
 When the agent asks a question, its options show under the input: pick one
 with the arrow keys, press `Space` to pick several when the question allows
 it, or type your own answer.

From cb636cc9cd21a5874794da82735ebe0ea591dc68 Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 23:01:43 +0100
Subject: [PATCH 8/9] fix(cli): engineer prompts skip runs with no capture, and
 the run id only shows when it fits

---
 src/reverse_api/cli.py                | 19 +++++++++++--------
 src/reverse_api/repl/view.py          |  5 +++--
 tests/test_repl.py                    | 24 +++++++++++++++++++-----
 website/content/docs/cli/commands.mdx |  4 ++--
 4 files changed, 35 insertions(+), 17 deletions(-)

diff --git a/src/reverse_api/cli.py b/src/reverse_api/cli.py
index 00ad0ea..e9319b6 100644
--- a/src/reverse_api/cli.py
+++ b/src/reverse_api/cli.py
@@ -1927,10 +1927,16 @@ def engineer_from_prompt(text: str, model: str | None = None, output_dir: str |
     """A prompt in engineer mode: a known run id re-engineers that run; anything else is an instruction for the latest run."""
     if session_manager.get_run(text):
         return run_engineer(text, model=model, output_dir=output_dir)
-    latest = session_manager.get_history(limit=1)
-    if not latest:
-        raise RuntimeError("no runs yet · start one in agent or manual mode")
-    return run_engineer(latest[0]["run_id"], model=model, output_dir=output_dir, additional_instructions=text)
+    # Collector runs capture no traffic, so the latest run with a HAR.
+    latest = next((r for r in session_manager.get_history(limit=50) if _har_path(r["run_id"], r).exists()), None)
+    if latest is None:
+        raise RuntimeError("no captured runs yet · start one in agent or manual mode")
+    return run_engineer(latest["run_id"], model=model, output_dir=output_dir, additional_instructions=text)
+
+
+def _har_path(run_id: str, run: dict[str, Any]) -> Path:
+    """Where a run in history keeps its captured traffic."""
+    return Path((run.get("paths") or {}).get("har_dir") or get_har_dir(run_id, None)) / "recording.har"
 
 
 def run_engineer(
@@ -1962,10 +1968,7 @@ def run_engineer(
         else:
             if not prompt:
                 prompt = run_data["prompt"]
-            # Detect where it was saved
-            paths = run_data.get("paths", {})
-            har_dir = Path(paths.get("har_dir", get_har_dir(run_id, None)))
-            har_path = har_dir / "recording.har"
+            har_path = _har_path(run_id, run_data)
 
     sdk = _session_sdk(resume) or config_manager.get("sdk", "claude")
     enable_sync = config_manager.get("real_time_sync", True)
diff --git a/src/reverse_api/repl/view.py b/src/reverse_api/repl/view.py
index 434c18c..3d8c9e8 100644
--- a/src/reverse_api/repl/view.py
+++ b/src/reverse_api/repl/view.py
@@ -689,9 +689,10 @@ def status(self) -> FormattedText:
             hint = [(f"fg:{DIM}", "type a follow-up · /new starts over" + (" · ⌃o expand " if st.folded else " "))]
         else:
             hint = [(f"fg:{DIM}", "/ commands" + (" · ⌃o expand " if st.folded else " "))]
-        if st.run_id:
-            hint += [(f"fg:{DIM}", f"· run {st.run_id} ")]
         used = sum(len(t) for _, t in left) + sum(len(t) for _, t in hint)
+        run = f"· run {st.run_id} " if st.run_id else ""
+        if run and used + 1 + len(run) <= width():  # left out rather than cut off
+            hint, used = hint + [(f"fg:{DIM}", run)], used + len(run)
         return FormattedText(left + [("", " " * max(1, width() - used))] + hint)
 
     def menu(self, rows: list[tuple[str, str, bool, bool]], title: str | None = None, limit: int = MENU_ROWS) -> FormattedText:
diff --git a/tests/test_repl.py b/tests/test_repl.py
index c73b66e..d9d5e46 100644
--- a/tests/test_repl.py
+++ b/tests/test_repl.py
@@ -149,6 +149,14 @@ def test_the_status_row_shows_the_run_id(self, session):
         LedgerUI(b).header("abc123", "prompt", model="claude-sonnet-5-5")
         assert "run abc123" in "".join(text for _, text in view.status())
 
+    def test_a_run_id_that_does_not_fit_is_left_out(self, session, monkeypatch):
+        from reverse_api.repl import view as view_module
+
+        _st, _t, view, b = session
+        LedgerUI(b).header("abc123", "prompt", model="claude-sonnet-5-5")
+        monkeypatch.setattr(view_module, "width", lambda: 40)
+        assert "run abc123" not in "".join(text for _, text in view.status())
+
     def test_folded_output_offers_ctrl_o(self, session):
         st, _t, view, b = session
         assert "expand" not in "".join(text for _, text in view.status())
@@ -933,20 +941,26 @@ def test_engineer_keeps_its_output_dir_for_later_prompts(self, monkeypatch):
             runner("add pagination", None)
         later.assert_called_once_with("add pagination", None, "/out")
 
-    def test_an_engineer_prompt_is_a_run_id_or_an_instruction_for_the_latest_run(self, monkeypatch):
+    def test_an_engineer_prompt_is_a_run_id_or_an_instruction_for_the_latest_capture(self, monkeypatch, tmp_path):
         from reverse_api import cli
 
+        captured = tmp_path / "har" / "capture1"
+        captured.mkdir(parents=True)
+        (captured / "recording.har").write_text("{}", encoding="utf-8")
         history = MagicMock()
         history.get_run.side_effect = lambda text: {"run_id": text} if text == "abc123" else None
-        history.get_history.return_value = [{"run_id": "latest1"}]
+        history.get_history.return_value = [
+            {"run_id": "collect1", "paths": {"har_dir": str(tmp_path / "har" / "collect1")}},
+            {"run_id": "capture1", "paths": {"har_dir": str(captured)}},
+        ]
         monkeypatch.setattr(cli, "session_manager", history)
         with patch.object(cli, "run_engineer") as engineer:
             cli.engineer_from_prompt("abc123", "m", "/out")
             cli.engineer_from_prompt("add pagination", "m", "/out")
         assert engineer.call_args_list[0] == call("abc123", model="m", output_dir="/out")
-        assert engineer.call_args_list[1] == call("latest1", model="m", output_dir="/out", additional_instructions="add pagination")
-        history.get_history.return_value = []
-        with pytest.raises(RuntimeError, match="no runs yet"):
+        assert engineer.call_args_list[1] == call("capture1", model="m", output_dir="/out", additional_instructions="add pagination")
+        history.get_history.return_value = history.get_history.return_value[:1]
+        with pytest.raises(RuntimeError, match="no captured runs yet"):
             cli.engineer_from_prompt("add pagination")
 
     def test_without_a_terminal_a_prompt_is_required(self, monkeypatch):
diff --git a/website/content/docs/cli/commands.mdx b/website/content/docs/cli/commands.mdx
index 393ce6b..690e232 100644
--- a/website/content/docs/cli/commands.mdx
+++ b/website/content/docs/cli/commands.mdx
@@ -42,8 +42,8 @@ Click a file or URL in the conversation to open it.
 
 The status row under the input shows the run's id, which `/resume`,
 `/messages` and `reverse-api-engineer show` take. Once the agent has verified
-the client, its step lists the run's folder and the synced copy in
-`./scripts`, as full paths.
+the client, its step lists the run's folder, and with real-time sync on, the
+synced copy in `./scripts`, as full paths.
 
 When the agent asks a question, its options show under the input: pick one
 with the arrow keys, press `Space` to pick several when the question allows

From 67820a28b9dbfc7368fedeb184e732ba22dd6bd2 Mon Sep 17 00:00:00 2001
From: kalil0321 <kalil.bouzigues@gmail.com>
Date: Wed, 28 Oct 2026 23:10:18 +0100
Subject: [PATCH 9/9] docs: chrome-mcp is the default agent provider in the
 config reference

---
 website/content/docs/configuration/models.mdx | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/website/content/docs/configuration/models.mdx b/website/content/docs/configuration/models.mdx
index 2a36b0c..6fabc7a 100644
--- a/website/content/docs/configuration/models.mdx
+++ b/website/content/docs/configuration/models.mdx
@@ -50,7 +50,7 @@ The full default config:
 
 ```json
 {
-  "agent_provider": "auto",
+  "agent_provider": "chrome-mcp",
   "agent_browser_notes": "",
   "agent_browser_npx_package": "agent-browser@0",
   "claude_code_model": "claude-sonnet-4-6",