jevway 0.1.2 → 0.3.0

package/README.md

@@ -1,219 +1,297 @@
 # jevway
 
-Give your coding agent a browser goal. [TypeSafe Jev](https://docs.typesafe.ai) picks each click, type, and select in about 300 ms, and the agent gets back **one short summary** instead of making every decision itself.
+[![npm](https://img.shields.io/npm/v/jevway)](https://www.npmjs.com/package/jevway)
+[![CI](https://github.com/AbdallahAHO/jevway/actions/workflows/ci.yml/badge.svg)](https://github.com/AbdallahAHO/jevway/actions/workflows/ci.yml)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
 
-```text
-agent ──jevway_task(goal, values, done_when)──▶ jevway ──one Jev request per step──▶ Chrome
-agent ◀── { status, url, steps, stuck? } ──────┘        (same Chrome chrome-devtools inspects)
-```
+Give your coding agent a browser goal and get back one short, checked summary. [TypeSafe Jev](https://docs.typesafe.ai)
+picks each click, type and select in about 300 ms, so your agent stops spending a turn on every step.
 
-The only thing it needs is `TYPESAFE_API_KEY` or `OPENROUTER_API_KEY`. Other decision models, some of which read screenshots, are described [below](#models). No second LLM is involved: the agent passes the text values it already knows, and Jev *selects* which value goes into which field (a speculative `value_for_<field>` head in the same request).
+<!-- Source: docs/images/how-it-flows.mmd (render with @mermaid-js/mermaid-cli: -t default -b white -s 2) -->
+![Your agent hands jevway a goal; each step jevway reads the page, asks Jev for one decision and acts in Chrome; the agent gets one summary back](docs/images/how-it-flows.png)
 
-Formerly jev-browse: its tools were `browser_task`, `browser_look` and `browser_task_end`, and its environment variables `JEV_*`; none of those names work any more.
+jevway is an MCP server and a CLI. It works with Claude Code, Codex, Cursor, OpenCode and any other MCP client, in your
+own signed-in Chrome or a throwaway one.
 
-Built on the loop, DOM snapshot, and guards from [browser-use/jev-ultrafast](https://github.com/browser-use/jev-ultrafast) (MIT).
+## Why
 
-## Install into any agent
+Driving a browser click by click costs an agent a model turn, a page snapshot and a few seconds per step. A 15-step
+form takes minutes and fills the context window with snapshots. jevway hands the whole goal to Jev, a model built to
+choose browser actions, and returns when it's done or when it needs the agent:
 
-**Claude Code plugin** (the MCP server and the skill in one install), inside Claude Code:
+- **It never invents input.** Jev types only the `values` the agent passed. A missing value comes back as
+  `needs_input`, naming the field.
+- **It stops before anything irreversible.** Pay, send, delete and submit come back as `needs_confirmation`. One
+  confirmation allows one click, on the control that was named.
+- **It checks its own success.** `done` needs a separate check of `done_when` on the final page, plus any exact
+  `checks` you add, and comes with the page's own words as evidence.
+- **Model output only ever selects a listed element.** It never becomes a selector, a script or typed text.
+- **It can record what it did.** A video of the run, a screenshot before every action, and an HTML page to review it,
+  built in. See [Recordings](#recordings).
 
-```text
-/plugin marketplace add AbdallahAHO/jevway
-/plugin install jevway@jevway
-```
+## What you need
 
-On Claude Code 2.1.275 or later this works in one step: `/plugin install jevway --marketplace AbdallahAHO/jevway`.
+jevway has no browser extension and no Playwright: it drives Chrome directly. Three things are required, and two
+optional tools unlock more.
 
-**Or the MCP server alone**, one line per client:
+| | Needed for | Without it |
+| --- | --- | --- |
+| **Node.js 24+** | running jevway (`npx` fetches it) | it won't start |
+| **Google Chrome** (or Chromium) | every run: your running Chrome, or a throwaway one jevway starts | runs fail; set `CHROME_PATH` if Chrome isn't in the usual place |
+| **An API key**: `TYPESAFE_API_KEY` (recommended) or `OPENROUTER_API_KEY` | every decision Jev makes | the server still starts and lists its tools, and `jevway_look` works, but tasks and journeys answer with the setup fix |
+| `ffmpeg` *(optional)* | the `run.mp4` video in a recording | recordings keep their step screenshots and HTML page, without the video |
+| `notify-send` *(optional, Linux)* | the desktop notification when a person is needed (macOS has it built in) | the agent still asks you; there's just no pop-up |
+| chrome-devtools MCP *(optional)* | inspecting what jevway left on screen: console, network, performance, DOM details | nothing is lost in jevway itself; your agent just has no deeper inspection tool |
 
-```bash
-claude mcp add jevway -s user -- npx -y -p jevway jevway-mcp   # Claude Code
-codex mcp add jevway -- npx -y -p jevway jevway-mcp            # Codex
-```
+**Which key?** Both run the same Jev model. The TypeSafe key adds **Box**, which locates controls on the screenshot:
+it double-checks the click Jev chose and whether it's covered, rescues a step when Jev is stuck, can click a canvas
+control (opt-in), and marks each journey screen's primary action and the controls missing from the page's code. Through OpenRouter those Box features are off; everything else works.
 
-Cursor (`~/.cursor/mcp.json`, under `mcpServers`), or any other stdio MCP client:
+`npx -y jevway doctor` checks all of this and names the fix for anything missing.
 
-```json
-"jevway": { "command": "npx", "args": ["-y", "-p", "jevway", "jevway-mcp"] }
-```
+## Quick start
 
-OpenCode (`opencode.json`, under `mcp`):
+1. **Get a key** from [TypeSafe](https://docs.typesafe.ai) (or use an [OpenRouter](https://openrouter.ai) key).
 
-```json
-"jevway": { "type": "local", "command": ["npx", "-y", "-p", "jevway", "jevway-mcp"], "enabled": true }
-```
+2. **Run the setup** in your terminal:
 
-**The key** goes in `~/.config/jevway/env` (`$XDG_CONFIG_HOME/jevway/env` when that is set), so it never has to go
-into a chat or a synced client config:
+   ```bash
+   npx -y jevway setup
+   ```
 
-```bash
-mkdir -p ~/.config/jevway && touch ~/.config/jevway/env && chmod 600 ~/.config/jevway/env
-# then add the line TYPESAFE_API_KEY=… (or OPENROUTER_API_KEY=…) in an editor
-```
+   It asks for the key without showing it and saves it to `~/.config/jevway/env` (mode 600, never a chat or a synced
+   config), tests the browser, and registers jevway with the clients it finds: Claude Code, Codex, Cursor, OpenCode.
+   Every change is shown first, as a command or a diff, and made only on a yes. It ends with `jevway doctor`.
 
-- `jevway` and `jevway-mcp` read the file at start. A variable already in the environment wins, so a client's `env`
-  block works too.
-- Some clients pass servers a filtered environment; the file still works there.
-- Without a key the server still starts and lists its tools: `jevway_look` works, and the other tools answer with the fix.
+   In Claude Code you can instead install the plugin, which brings the server and the skill that tells the agent when
+   to use it (then put the key in `~/.config/jevway/env` yourself):
 
-**Check the setup:**
+   ```text
+   /plugin marketplace add AbdallahAHO/jevway
+   /plugin install jevway@jevway
+   ```
 
-```bash
-npx -y jevway doctor          # Node, key and route, browser, ffmpeg; exit 0 when ready; never prints the key
-npx -y jevway doctor --plan   # the steps still missing, numbered
-npx -y jevway doctor --live   # plus one real Jev request (a fraction of a cent)
-```
+3. **Allow the browser.** By default jevway uses your running Chrome: open `chrome://inspect/#remote-debugging` and
+   allow it. To use a throwaway browser instead, set `JEVWAY_BROWSER=launch`.
 
-Installing through an agent? Point it at [INSTALL-FOR-AGENTS.md](INSTALL-FOR-AGENTS.md): rules (a diff before any
-config edit, no key in the chat, no relaunching a browser it didn't start), steps per client, and checks.
+4. **Ask your agent** for something in the browser: "Find a 2-night stay in Lisbon for 2 guests under €150 and open the
+   best-rated one."
 
-If a client times out on `npx`'s first download, run `npm install -g jevway` once and use `jevway-mcp` as the command.
+`npx -y jevway doctor` checks the setup at any time (`--json` for agents) and names the fix for anything missing.
 
-`JEVWAY_BROWSER` (the server's `env`, or the env file) selects the browser:
+## Let your agent set it up
 
-- `auto` (default): your running Chrome, after you allow it at `chrome://inspect/#remote-debugging`. This is the same Chrome `chrome-devtools-mcp --auto-connect` uses, so Jev's tab is visible to chrome-devtools.
-- `launch` or `launch-headless`: a throwaway profile.
-- `http://127.0.0.1:9222`: an explicit debugging port.
+Paste this into your coding agent (Claude Code, Codex, Cursor, OpenCode…):
 
-Outside the Claude Code plugin, copy `skill/jevway/` (in the repository, or `node_modules/jevway/skill/jevway/`) into
-your agent's skills, so it knows when to use Jev and when to use chrome-devtools.
+```text
+Set up jevway for me and prove it works: follow https://github.com/AbdallahAHO/jevway/blob/main/SETUP.md step by
+step. Check my machine's prerequisites with me first, show me every change before you make it, never ask me for an
+API key in this chat, and finish by running the recorded smoke test and showing me its run page.
+```
+
+[SETUP.md](SETUP.md) walks the agent through everything: prerequisites for your OS, the key, registering the server,
+the browser, the skill, a recorded Google → Wikipedia smoke test, and a report with the run page, video and
+screenshots as proof.
 
-From a checkout: `pnpm install && pnpm build`, then `node /absolute/path/to/jevway/dist/mcp.js` as the command.
+## Install
 
-For agents without MCP, use the CLI:
+**Claude Code:** the plugin above, or the server alone:
 
 ```bash
-npx -y jevway --browser launch-headless \
-  --url https://en.wikipedia.org/wiki/Main_Page \
-  --goal "Open the Wikipedia article about the Rosetta Stone" \
-  --values '{"search_query":"Rosetta Stone"}' \
-  --done-when "The Rosetta Stone article is open"
-# prints the JSON summary; exit 0 only when status is "done"
+claude mcp add jevway -s user -- npx -y -p jevway jevway-mcp
 ```
 
-## Models
+**Codex:**
+
+```bash
+codex mcp add jevway -- npx -y -p jevway jevway-mcp
+```
 
-`JEVWAY_MODEL` (or `--model`) picks the decision model. All of them speak the same System One API through the same SDK. With `TYPESAFE_API_KEY` set, Jev runs on the TypeSafe API as `jev-1.14.0`: screenshots go in as `assets` and Box questions are available. Everything else, and Jev without that key, goes through OpenRouter with `OPENROUTER_API_KEY`. OpenRouter drops TypeSafe's `assets` map, so Box only runs on the direct route.
+**Cursor** (`~/.cursor/mcp.json`, under `mcpServers`) **or any stdio MCP client:**
 
-| `JEVWAY_MODEL` | Model | Screenshots | Fixture + live web (`pnpm verify`) | Journeys |
-| --- | --- | --- | --- | --- |
-| `jev` (default) | `jev-1.14.0` (TypeSafe API: `assets` + Box) / `typesafe/jev-1.14` (OpenRouter) | `fallback` | direct: 7/7 `fallback` and `always`, 6/6 off, ~240 ms/request | 8/8 off and `always` |
-| `pplx-decider` | `perplexity/pplx-decider-v1.1-27b` | `always` | 6/6 off and `always`, 0.4–1.3 s/request | 8/8 |
-| `clef-omni` | `cloudflare/clef-omni` | `always` | 4/6 (3/6 off) | 7/8 (8/8 off) |
-| `clef` | `cloudflare/clef` | none (images cost 16k tokens) | 6/6, 0.7–1.2 s/request | — |
-| `clef-flash` | `cloudflare/clef-flash` | `always` | 1/6 | — |
-| `gpt-6-luna` | `openai/gpt-6-luna-decisions` | `always` | 0–1/6: OpenAI refuses target questions (502) | — |
+```json
+"jevway": { "command": "npx", "args": ["-y", "-p", "jevway", "jevway-mcp"] }
+```
 
-**Jev 1.14 is the fastest model that passes everything, and it reads screenshots. Perplexity's decider also passes everything with screenshots, but slower. Clef passes text-only, and is slower still.**
+**OpenCode** (`opencode.json`, under `mcp`):
 
-`JEVWAY_SCREENSHOTS` (or `--screenshots`) controls screenshots for the vision models:
+```json
+"jevway": { "type": "local", "command": ["npx", "-y", "-p", "jevway", "jevway-mcp"], "enabled": true }
+```
 
-- `always`: on every step.
-- `fallback`: only when a text-only decision would stop as uncertain or blocked, has its DONE rejected, or follows an action that changed nothing.
-- `off`: never.
+Outside the Claude Code plugin, copy `skill/jevway/` into your agent's skills folder, so it knows when to use jevway and
+when to use chrome-devtools. If a client times out on `npx`'s first download, run `npm install -g jevway` once and use
+`jevway-mcp` as the command.
 
-Journeys send one with each screen's judgments whenever the mode isn't `off`. Each observed element index is drawn on the image as a numbered box, taken from the snapshot's own rects. Covered elements get no box, and the overlay is removed before any input. The model still only selects observed ids.
+Having an agent install it? `npx -y jevway setup --client <claude-code|codex|cursor|opencode|other> --print` prints the
+exact command or config for that client and changes nothing. [SETUP.md](SETUP.md) has the
+steps per client, the rules (a diff before any config edit, no key in the chat) and the checks.
 
-### Box (Jev 1.14 on the TypeSafe API)
+**Choosing a browser** (`JEVWAY_BROWSER` in the server's env or the env file):
 
-On every request that carries a screenshot, Box questions run alongside the decision. Box answers always land back on an **observed element** (`src/jev/box.ts`), with one opt-in exception:
+| Value | Uses |
+| --- | --- |
+| `auto` (default) | your running Chrome, after you allow it at `chrome://inspect/#remote-debugging` |
+| `launch` / `launch-headless` | a throwaway profile jevway starts and closes |
+| `http://127.0.0.1:9222` | a Chrome you started with that debugging port and its own profile |
 
-- **Cross-check:** Box asks where the next click belongs. If it confidently points at a different element than Jev chose, and the box doesn't also cover Jev's choice, the step hands back as `uncertain`.
-- **Covered check:** for a chosen click, a follow-up asks whether the target is visibly unobstructed and boxes it. A clearly covered target is fed back as history ("covered; clear it first") up to twice, then hands back.
-- **Rescue:** when Jev is BLOCKED, unsure, or claims DONE that `done_when` rejects, and the box confidently covers an observed element, jevway clicks that element. It is still gated by the box's own commit-risk score.
-- **Coordinate clicks (opt-in):** `allow_coordinate_clicks` / `--allow-coordinate-clicks`. When Jev is stuck and the box matches nothing the DOM reader reported (canvas, or a frame that was skipped), jevway clicks the box centre after a hit-test. Irreversible-looking clicks still need `confirm_irreversible`.
-- **Journeys:** each screen's primary action is boxed (a dashed outline in the viewer, `primary_action` in reports). A tappable control with no DOM element becomes a `control_not_in_dom` QA finding.
+**Without MCP**, use the CLI. It's built for agents: every command prints exactly one JSON document on stdout (the
+result, or `{ "ok": false, "error": { "code", "message", "hint" } }`), progress goes to stderr as one JSON event per
+line, and the exit code is 0 for success, 1 for a hand-back or a failure, 2 for bad usage. `--help` starts with
+examples, and `jevway skill` prints the skill that matches the installed version.
 
-Every model judges commit risk **per click candidate by name**, in one follow-up request per step. Jev also keeps its page-level guess, and a click is held if either says so. Per-model thresholds, rules and results are in [docs/verification.md](docs/verification.md#decision-models-on-openrouter-2026-10-11).
+```bash
+# One-shot: runs, closes its tab, prints the result
+npx -y jevway run --browser launch-headless --url https://en.wikipedia.org \
+  --goal "Open the Wikipedia article about the Rosetta Stone" --done-when "The Rosetta Stone article is open"
+
+# A session: a background daemon keeps the tab between commands, so a hand-back can be answered
+jevway run --session new --url https://example-stays.com --goal "Book 2 nights in Lisbon" --values '{"city":"Lisbon"}'
+jevway run --session <id> --goal "Book 2 nights in Lisbon" --values '{"email":"ada@example.com"}'
+jevway look --session <id>          # what the page shows now, no model request
+jevway sessions                     # open sessions;  jevway end <id> closes one
+```
 
-## Tools
+## Use
 
-**`jevway_task`** takes `goal`, `url` | `tab`, `values`, `secret_keys`, `files`, `done_when`, `checks`, `session`, `confirm_irreversible`, `confirm_target`, `max_steps`, `max_seconds`, `allowed_hosts`, `blocked_hosts`, and, for new sessions only, `allow_coordinate_clicks` and `record_dir`. It returns:
+The agent calls `jevway_task` with the outcome it wants, the values it may type, and the evidence of success:
 
 ```json
-{ "session": "be63210d", "status": "needs_input", "reason": "Blocked: no supplied value for \"Email address\"",
-  "url": "…", "title": "…", "page_text": "…", "elapsed_ms": 899,
-  "steps": [{ "operation": "TYPE_TEXT", "element": "Full name", "value_key": "full_name", "page_changed": true, "confidence": 0.9 }],
-  "jev": { "requests": 3, "input_tokens": 7534, "median_latency_ms": 284, "screenshots": 0, "model": "jev-1.14.0", "est_cost_usd": 0.000316 },
-  "stuck": { "operation": "BLOCKED", "candidates": [] },
-  "missing": [{ "label": "Email address", "role": "textbox", "path": "/stays" }] }
+{
+  "url": "https://example-stays.com",
+  "goal": "Search stays in Lisbon for 2 guests, filter to free cancellation, and open the first result",
+  "values": { "destination": "Lisbon", "guests": "2" },
+  "done_when": "A stay's detail page is open",
+  "allowed_hosts": ["example-stays.com"]
+}
 ```
 
-`jev.model` is the model id the requests went to. `jev.est_cost_usd` estimates the call's cost from its input tokens (every request, Box and the done's evidence request included) at the model's published rate, rounded to 6 decimals: Jev on the TypeSafe API is priced at $0.042 per 1M input tokens (output is free); Jev through OpenRouter gets the same rate (OpenRouter's margin, if any, unchecked); a model with no rate written down gets `null`. It is an estimate, not billing.
+It gets back one concise result, `status` and `reason` first:
 
-`secret_keys` names the keys of `values` that are secrets (an API token, a one-time code). Jev sees `[secret:<key>]` in place of such a value and never reads it; it is typed as given, and wherever the page echoes it, `page_text`, `steps`, reasons, traces and `record_dir`'s `summary.json` and `index.html` show the marker instead. Recorded screenshots and video are pixels of the page and are not masked. Mark only real secrets: every occurrence of the value is replaced, even inside other words.
+```json
+{ "status": "done", "reason": "Completion condition verified: a stay's detail page is open.",
+  "story": ["Typed the \"destination\" value into \"Destination\".", "Clicked \"Search\".", "Clicked \"Casa Flora\"."],
+  "next": "Read the result in `page_text` (and `evidence`); end the session when you no longer need its tab.",
+  "session": "a63a13f0-…", "url": "…", "title": "Casa Flora · Staylist",
+  "files": { "run_dir": "…/.jevway/runs/2026-11-02_1028_search-stays-in-lisbon_a63a", "screenshot": "…/final.jpg" },
+  "page_text": "…" }
+```
 
-| Status | Meaning |
-| --- | --- |
-| `done` | Jev chose DONE **and** a separate `done_when` check passed, or that check alone scored ≥ 0.9 once or ≥ 0.75 on 3 decisions in a row. A rejected DONE waits and re-decides, up to 3 times. |
-| `needs_input` | A field needs a value that isn't in `values`. Nothing is invented. `missing` lists each field (label, role, page path). |
-| `needs_confirmation` | The next click looks irreversible (pay, send, delete, submit). `confirm_irreversible: true` allows exactly one such click, and an agent should set it only for the action and target the user's own request names (a category such as "anything unused" names no target: it lists the candidates and asks the user); otherwise it asks the user. It allows only the control that needs_confirmation named (same label, same page path); a different irreversible control is handed back as a new needs_confirmation. It counts only on the call that answers a needs_confirmation (or a new session's first call, which has no flagged control and allows any one); anywhere else it is ignored. Pass `confirm_target` with the user's words that name the target; jevway checks the page shows them (the control's own words, like "delete", do not count) and refuses otherwise. |
-| `uncertain` | Operation or target confidence was below the model threshold (Jev 0.3; per model in `src/jev/model.ts`), 6 actions in a row ran under 0.5 confidence (exploring, not progressing), or DONE could not be verified. |
-| `blocked` | No supported operation can progress, 3 actions in a row changed nothing, the same target was covered 3 times, one action only toggled the page back and forth 4 times, or the same element was used 6 times in 7 actions; the page shows an error after this call acted (the reason quotes it); or a sign-in screen blocks the goal and no supplied value fits its fields (`wall: { kind: 'sign_in', origin }`). Also a bot check (a Cloudflare challenge, a reCAPTCHA, hCaptcha or Turnstile page, an "unusual traffic" page) that did not clear by itself within 8 s: the reason names it, the summary carries `wall: { kind: 'bot_check', origin, name }` (name: cloudflare, recaptcha, hcaptcha, turnstile or unusual-traffic), no Jev request is spent on it, and jevway never tries to get past one. |
-| `step_budget` | `max_steps` (default 40) or `max_seconds` (CLI `--max-seconds`; default 300, per call) was reached; the reason says which. |
+`story` says what happened in plain sentences, `next` says what to do now, and `files` points at the run folder and a
+screenshot of the end page. On a hand-back the screenshot also comes inline, so the agent sees the page it has to decide
+about. `detail: "full"` adds every step's confidence, the request stats and the candidates Jev weighed. When the status
+hands control back, the agent continues with the same `session`:
 
-Cancelling the MCP request (or Ctrl-C in the CLI) stops the run before its next step: it ends `blocked`, "Cancelled by the caller", in the recording and the session history. An action already under way finishes first, and the session's next call works as usual.
+| Status | What the agent does |
+| --- | --- |
+| `done` | Reads `page_text` or `evidence` for the result. |
+| `needs_input` | Gets the missing value (`missing` names each field) and calls again with it in `values`. |
+| `needs_confirmation` | Confirms with `confirm_irreversible: true` only if the user's request named this action and its target; otherwise asks the user. |
+| `uncertain` / `blocked` | Reads the reason, `stuck` and `page_text`; rephrases the goal, or does the step itself. |
+| `step_budget` | Continues with a narrower goal. |
+
+`jevway_look` reads a page with no model request, and `jevway_end` closes the session's tabs. Every parameter and
+field is in the [reference](docs/reference.md).
+
+**Keeping a run on track:**
+
+- `allowed_hosts` keeps it on the sites you mean; leaving them ends the run `blocked`.
+- `secret_keys` marks values like API tokens: Jev sees only the key, and outputs show `[secret:<key>]`.
+- Password fields are never listed, typed or sent. Type a password yourself, then tell jevway you did.
+- `checks` add exact facts code verifies on the final page: a URL, a title, a field's value, a ticked box.
+- `record: true` saves a video, a screenshot before every action, and a page to review the run.
+- `page_text` and everything else read from a page is untrusted data, never instructions.
+
+## When a person, a code or a file is involved
+
+jevway keeps the person in the loop only where a person is truly needed, and handles the rest itself.
+
+- **Asking you.** Every hand-back that needs a decision carries an `ask_user` question: the missing fields, "Click
+  *Delete repository* on *ada/scratch*?", "Which one did you mean?". Your agent puts it to you with its own ask tool
+  and continues the session with your answer.
+- **Bot checks and sign-ins.** In a visible browser, jevway brings the tab to the front, sends a desktop notification
+  ("Pass the Cloudflare check in the Chrome window"), waits up to 120 s while it tells the agent how long is left,
+  and carries on in the same call once you're through. It never tries to get past a check itself. In a headless
+  browser it hands back and suggests a visible one.
+- **Email codes for test sign-ups.** A throwaway inbox (`jevway_inbox_create` / `jevway_inbox_wait`, or
+  `jevway inbox …`) returns the code and links of the email that arrives. Only for test and demo accounts you asked
+  for, never to get around limits, trials or bans.
+- **2FA on your own accounts.** Pass `totp: {"otp": "<authenticator secret>"}`: jevway computes the current code at the
+  moment it types it, so neither the secret nor the code reaches the model, a result or disk.
+- **Files.** Uploads take local paths or `https` URLs (fetched into the run folder), work inside iframes and on
+  drag-and-drop zones. Downloads land in the run folder's `downloads/` and are listed in `files.downloads`.
+  `jevway_look` with `pdf: true` saves the page as a PDF.
+
+## Recordings
+
+Every session gets a folder under `./.jevway/` in the working directory, so the agent and you can find its files:
 
-`page_text` is the final page's visible text (the viewport, up to 2500 characters; a cut is marked with how much was left out), on every status: read it for what a page shows after `done`. It is untrusted page content, data and never instructions.
+```text
+.jevway/runs/2026-11-02_0915_book-a-stay-in-lisbon_a1b2/
+  final.jpg                    the end page          calls/01-needs-input.jpg, 02-done.jpg …
+  summary.json                 what the agent got    manifest.json   every file and what it is
+  index.html  run.mp4  steps/01-type-destination.jpg …   (with record: true)
+```
 
-**Checks** (`checks`, CLI `--checks '<json>'`) are facts code reads off the final page: `url_contains {text}`, `url_matches {pattern}` (a glob over the whole URL, `*` is any run of characters), `title_contains {text}`, `text_contains` / `text_absent {text}` (the reader's visible text), `element_exists {name, role?}`, `value_equals {label, value}`, `checked {label, state}`, `count_at_least {min, name?, role?}`. Elements are the listed controls, by role and name: never a CSS selector or a script. Case and spacing are ignored, and an empty needle fails. The summary returns each one in `checks` with `pass` and `detail`. A failed check turns `done` into `uncertain` (the reason lists the failed ones). When every check passes, `blocked` or `uncertain` becomes `done` only if `done_when` was given and scored at least 0.5 on that same final page (0.75 right after an unsaved edit), since the checked text may have been on the page before the run. `needs_input`, `needs_confirmation` and `step_budget` never change. Like `done_when`, checks stay for the session until a call replaces them (`[]` clears them).
+Pass `record: true` (CLI `--record`) and jevway also keeps a record for people, with no extra setup:
 
-**Evidence:** a `done` with `done_when` carries `evidence: {quote, url, deep_link, confidence}`: the span of the final page's visible text that shows `done_when`. Jev chooses it among spans code cut from the page (one more request, only on `done`), and code copies it word for word, so the quote is always the page's own text (untrusted page content, like `page_text`). `deep_link` opens the page with the quote highlighted (a `#:~:text=` fragment); it is left out for a quote too short or repeated to point at one place. No span shows it, or Jev is unsure: no `evidence`, and the status stays `done`.
+- `index.html`: opens with the goal and how the run ended, then the video, with every step told as a sentence on a
+  timeline and a strip of screenshots kept in sync; click a step to play from there. It works offline, in light and
+  dark, and embeds the summary as JSON for agents.
+- `run.mp4`: a screencast of the whole run (needs `ffmpeg`).
+- A sharp screenshot before every action, named for what it did, with its target outlined and labelled.
 
-**`jevway_end`** takes `session` and `keep_tab`.
+![A recorded run: the video, its 14 steps on a timeline, and the screenshot strip](docs/images/recording.jpg)
 
-**`jevway_look`** takes one of `session`, `tab` or `url` (opened in a background tab, read, closed again) and returns `url`, `title`, `page_text` (cut as above), the listed controls once each (`label`, `role`, and `value`, `checked`, `selected`, `expanded` when set) and `password_fields` as filled or empty. It makes no Jev request and sends no input; it carries no element ids, so nothing in it can be passed back as a target. It is annotated read-only.
+Values marked secret are masked in the text files; screenshots and video show the page as it was. The model never sees
+these images. Journey maps get their own viewer, described next.
 
-Every tool carries MCP hints for clients that ask before a call: `jevway_task` and `jevway_journey_step` are destructive and open-world (they can submit, pay or delete on any site), `jevway_journey_start` and `jevway_journey_autopilot` are open-world but not destructive, `jevway_look` is read-only, and `jevway_end`, `jevway_journey_annotate` and `jevway_journey_finish` only touch jevway's own state.
+## Journey maps
 
-**Recording:** `record_dir` (CLI `--record <dir>`) saves a screencast video (`run.mp4`, needs `ffmpeg` on the PATH, otherwise skipped), a 2× JPEG before every action with its target outlined, the final page, `summary.json`, and an `index.html` page that opens on the video with its steps on a timeline and a strip kept in sync, then the run's calls, reason, address and cost. These shots are for people; the model still gets its own 1024 px screenshots.
+The `jevway_journey_*` tools map how a person moves through a site on a phone: every screen with its type and UI
+patterns, the taps between them, and QA findings like tap targets under 44 px, console errors and dead ends. You pick
+the taps, or let Jev explore. The result is a shareable viewer. See [docs/journeys.md](docs/journeys.md).
 
-## Safety
+## How it works
 
-- Model output only ever selects an observed element index. It never becomes a selector, script, or text, and it becomes a coordinate only through the opt-in coordinate click (stuck, a confident box matching no observed element, a hit-test, and the commit-risk gate).
-- Typing uses real key events per character, so autocompletes register it. Enter is a separate fixed key, offered only right after typing into a search-like field, and pressed only while that field still has focus. Escape is offered only while a menu, popover or suggestion list is expanded, and pressed once automatically when a popup covers the chosen target (then a blur and a synthetic outside press, for lists that ignore Escape).
-- Every input re-checks the page for freshness and for covered targets first (from jev-ultrafast).
-- Page text is treated as untrusted data in every instruction.
-- **Hosts:** `allowed_hosts` / `blocked_hosts` (CLI `--allowed-hosts`, `--blocked-hosts`; server-wide `JEVWAY_ALLOWED_HOSTS`, `JEVWAY_BLOCKED_HOSTS`) take exact hosts or `*.example.com` (subdomains only, not `example.com`). A call can only narrow the server's lists, and a later call only the session's. A start URL outside is never opened; the page URL is checked before every decision and after every step, and leaving ends the run `blocked` naming the host. The run never navigates back. `about:blank` and Chrome error pages always pass; `data:`, `file:` and other host-less pages are outside an allowed list. Only the top page's URL is checked: iframes, popup tabs and the page's own requests are not.
+Each step, jevway reads the page through the Chrome DevTools Protocol into a numbered list of controls and the visible
+text, and sends one request to Jev with the goal, that list and the names of the supplied values. Jev answers with an
+operation, a listed element, which value to type, its confidence and a commit-risk score. Code checks the answer
+against the list, re-checks the page is fresh and the target isn't covered, and acts with real input events. Guards
+stop loops, and a separate check decides `done`.
 
-## Verify
+The full technical specification is in [docs/how-it-works.md](docs/how-it-works.md).
 
-```bash
-pnpm test      # offline: action space, request shape, answer validation, task-loop rules
-pnpm verify    # live: headless Chrome + real Jev, checked against independent ground truth
-pnpm verify:real   # live public sites, a fresh browser per run, 5 runs each (JEVWAY_REAL_RUNS, JEVWAY_REAL_ONLY)
-pnpm verify:langfuse   # 10 jobs in Langfuse's demo project, signed-in Chrome on :9333, recorded into report.html
-pnpm report:render verify/out/<folder>   # rebuild a suite's report.html (or a run folder's index.html), offline
-pnpm verify:replay --record verify/out/replay/<name>   # live: record every Jev answer of the fixture suites
-pnpm verify:replay verify/out/replay/<name>            # offline: replay them, no key, no credits
-```
+## Models
 
-`pnpm verify` runs:
+Jev 1.14 is the default, on the TypeSafe API with `TYPESAFE_API_KEY` (with screenshots and Box, which locates
+controls on the screenshot to cross-check and rescue clicks) or through OpenRouter with `OPENROUTER_API_KEY`. Perplexity's and Cloudflare's decision models run through
+OpenRouter too. Jev is priced at $0.042 per 1M input tokens, so a typical 10-step task costs well under a cent. See
+[Models](docs/reference.md#models) for results per model.
 
-- **Local booking fixture:** async autocomplete, native select, checkbox, async results, then open a result. Next, `needs_input` for a missing email, then `needs_confirmation` on "Confirm booking and pay", then the confirmed booking. All four scenarios share one session.
-- **Real Wikipedia search.**
-- **Canvas-only control** (Box profiles with screenshots): no click without the opt-in, and the game starts with it.
-- **Wide table rows** (like Langfuse's traces table): rows wider than their scroll box are listed and opened, not their checkboxes. The run is recorded.
-- **MCP stdio round trip.**
+## Documentation
 
-`pnpm verify:real` runs Wikipedia, MDN, GitHub, Hacker News, npm, Booking.com and Google Flights. Every task stops before login or payment, passes are checked against the final URL, and bot walls (Cloudflare, captchas) are reported apart from failures, judged by the same detector the run uses (`src/browser/bot-wall.ts`).
+| Doc | For |
+| --- | --- |
+| [docs/reference.md](docs/reference.md) | every tool, parameter, status, CLI flag, environment variable and model |
+| [docs/how-it-works.md](docs/how-it-works.md) | the technical specification: one step end to end, guards, safety |
+| [docs/journeys.md](docs/journeys.md) | the journey-map tools |
+| [SETUP.md](SETUP.md) | agents setting jevway up for a person, with a recorded smoke test |
+| [skill/jevway/SKILL.md](skill/jevway/SKILL.md) | the skill agents load: when and how to call jevway |
+| [docs/verification.md](docs/verification.md) | the test suites and their results |
+| [docs/prior-art.md](docs/prior-art.md) | other open-source Jev browser projects, and what was taken from them |
+| [CONTRIBUTING.md](CONTRIBUTING.md) | building, testing and releasing |
 
-**Replay.** `pnpm verify:replay` re-runs the fixture suites (`verify`, `verify:gotchas`, `verify:redteam`, `verify:capabilities`) against recorded Jev answers, so a loop or guard change can be regression-tested offline and for free. `JEVWAY_REPLAY=record` saves each answer under the sha256 of its request (keys sorted, screenshots left out, local fixture ports normalised) in `JEVWAY_REPLAY_DIR`; `JEVWAY_REPLAY=replay` answers only from there, with no key and no network. A request never recorded is a miss that names the goal and step and shows what changed against the nearest recording, and any miss fails the suite. A change to what Jev is asked is meant to miss: record again after it. Gotchas replay their `normal` profile (the `churn` and `slow` pages change with timing), and the live Wikipedia scenario always misses.
+## Contributing
 
-Traces go to `verify/out/`. A jobs suite's `report.html` opens on an overview of its jobs, each with its featured run's step shots; a job's view starts with the run's video, its steps on a timeline and a strip that follow the video, every run one click away, then what Jev was given and how the run was judged. A step shot on the overview opens its run at that step (`#<job>-r<run>-s<step>`), and a phone recording sits in a frame of its own aspect. `verify:agent`'s page has the same layout: each chore's runs, with the agent's tool calls, their arguments and statuses, its final answer and the judgement. Each run's cost row shows its estimated cost (the sum of its calls' `est_cost_usd`), and the overview the median per run and the suite's total; a run saved before summaries carried a cost, or on a model without a price, shows "unknown". Results are in [docs/verification.md](docs/verification.md).
+Bug reports, site failures and pull requests are welcome. [CONTRIBUTING.md](CONTRIBUTING.md) covers the setup, the
+offline test suite (no API key needed) and the conventions. Report security issues privately, as described in
+[SECURITY.md](SECURITY.md).
 
-How jevway compares with other open-source Jev browser projects, and which of their ideas are planned or rejected: [docs/prior-art.md](docs/prior-art.md).
+## License
 
-## Limits
+MIT. The task loop, DOM snapshot and guards build on [browser-use/jev-ultrafast](https://github.com/browser-use/jev-ultrafast)
+(MIT; see [LICENSE.jev-ultrafast](LICENSE.jev-ultrafast)).
 
-- An element whose centre is off-screen (a row wider than its scroll box) is clicked at the centre of its visible part. A target clipped by a scrolling sidebar or table is scrolled into view just before the click.
-- Also read: sortable column headers (including ones scrolled sideways), plain pointer-cursor items inside open popups (date-range menus), and contenteditable search bars. Rows are labelled by their visible text.
-- The DOM reader covers common HTML/ARIA controls, including inside open shadow roots and iframes. Every visible frame, same-origin or cross-origin (out-of-process frames too), is read with the same reader: its controls continue the page's numbering, its text follows the page's under `[Frame: …]` within the same 6000 characters, and clicks, typing and every guard run inside the frame. A frame is skipped when it shows less than 16×16 px, sits outside an open modal, is still blank, takes over 1.5 s to read, or is past the caps (10 frames, 3 levels deep); controls a banner or sheet covers in the page around the frame aren't listed. CSS-transformed (scaled, rotated) iframes are placed as if untransformed.
-- Closed shadow roots and canvas aren't read, and file inputs inside iframes aren't supported: they are not listed, and an upload is never set inside a frame. Box can see canvas controls (and frames skipped above), and click them with the opt-in; otherwise use chrome-devtools.
-- Uploads take only files the caller declares in `files` (key → absolute path, checked before the run), never a path from `values`. Only then are file inputs listed (hidden ones through their label or styled button) and UPLOAD offered; Jev sees each key and file name, never the path, and a native file picker is intercepted. No declared file fitting a field is `needs_input`.
-- A tab the session's own click opens (`target=_blank`, `window.open`) is followed, and a closed one falls back to the tab that opened it; tabs anyone else opens are never touched. `jevway_end` closes the tabs the session's clicks opened (and its first tab when it opened it) unless `keep_tab`. The new tab is set up when it appears, so on a page that loads within ~20 ms its first document's own scripts run before the click counter; that can only end a run `blocked`, never repeat a click. A dialog the new tab raises before it is set up freezes it; following then fails after 5 s and the run ends `blocked`. In `auto` mode Chrome shows the new tab in front. A `record_dir` video stays on the first tab.
-- Only on-screen elements are listed; Jev scrolls to reveal more. On a page whose document doesn't scroll (an app shell), the scroll goes to the largest visible pane that does; a scroll that moved nothing is reported to Jev as such. Exception: an off-screen link, button or row containing one of the supplied `values` is listed too, and scrolled into view when clicked.
-- The scroll aimed at an inner pane works on the main page only; inside an iframe, scrolling stays page-level.
-- Confidence thresholds (per model, in `src/jev/model.ts`) and the 0.5 risk threshold are starting points, tuned on a small set of runs.
-- Jev's commit-risk check is a speculative page-level judgment ("the click that best advances the goal") plus a per-element score for the chosen click; the higher one counts. Other models score each candidate only, and Box rescues use the box's own score.
-- A click scoring 0.25–0.5 on risk gets a screenshot look before it runs, but only with screenshots on (`fallback` or `always`). With screenshots `off`, text alone under-reads some checkout pages (a bare "Continue" can score under 0.5), so keep screenshots on when a task can reach a payment step.
+jevway was called jev-browse until 0.1.0. Its old tool names (`browser_task`, `browser_look`, `browser_task_end`,
+`journey_*`) and `JEV_*` environment variables no longer work.