@akshayram1/omnibrowser-agent 0.2.29 → 0.3.0

package/.env

@@ -0,0 +1 @@
+HF_TOKEN=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

package/README.github.md

@@ -0,0 +1,330 @@
+# omnibrowser-agent
+
+[![npm](https://img.shields.io/npm/v/@akshayram1/omnibrowser-agent)](https://www.npmjs.com/package/@akshayram1/omnibrowser-agent)
+[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
+
+Local-first browser AI operator. Plans and executes DOM actions entirely in the browser — no API keys, no cloud costs, no data leaving your machine.
+
+[Live Demo](https://omnibrowser-agent.vercel.app/examples/chatbot/) · [Embedding Guide](docs/EMBEDDING.md) · [Architecture](docs/arch.md) · [Deployment](docs/DEPLOYMENT.md) · [Roadmap](docs/ROADMAP.md)
+
+---
+
+## Architecture
+
+```mermaid
+flowchart TB
+    subgraph DELIVERY["Delivery Layer"]
+        EXT["🧩 Chrome Extension\npopup + background worker"]
+        LIB["📦 npm Library\ncreateBrowserAgent()"]
+    end
+
+    subgraph ORCHESTRATION["Orchestration"]
+        BG["background/index.ts\nSession & tick loop"]
+        BA["BrowserAgent class\nrunLoop() / resume() / stop()"]
+    end
+
+    subgraph CORE["Core  (src/core/)"]
+        PL["planner.ts\nheuristicPlan() / webllm bridge\nplanNextAction()"]
+        OB["observer.ts\ncollectSnapshot()\nDOM candidates + visibility filter"]
+        EX["executor.ts\nexecuteAction()\nclick / type / navigate\nscroll / focus / wait"]
+    end
+
+    subgraph SHARED["Shared  (src/shared/)"]
+        CT["contracts.ts\nAgentAction · PageSnapshot\nAgentSession · PlannerResult"]
+        SF["safety.ts\nassessRisk()\nsafe / review / blocked"]
+        PA["parse-action.ts\nparseAction()\nparsePlannerResult()"]
+    end
+
+    subgraph OUTCOMES["Action Outcomes"]
+        direction LR
+        OK["✅ safe → execute"]
+        RV["⚠️ review → needs approval"]
+        BL["🚫 blocked → stop"]
+    end
+
+    subgraph PLANNERS["Planner Modes"]
+        direction LR
+        HP["Heuristic\nzero deps · offline\nregex patterns"]
+        WL["WebLLM\non-device · WebGPU\nwindow.__browserAgentWebLLM"]
+    end
+
+    EXT --> BG
+    LIB --> BA
+    BG -. "chrome.tabs.sendMessage" .-> CORE
+    BA --> CORE
+
+    PL --> OB
+    PL --> SHARED
+    OB --> SHARED
+    EX --> SHARED
+
+    SF --> OUTCOMES
+    PL --> PLANNERS
+```
+
+---
+
+## How it works — one tick
+
+```
+goal + history + memory
+        │
+        ▼
+observer.collectSnapshot()  ──→  PageSnapshot (url, title, candidates[])
+        │
+        ▼
+planner.planNextAction()    ──→  PlannerResult { action, evaluation?, memory?, nextGoal? }
+        │
+        ▼
+safety.assessRisk(action)   ──→  safe | review | blocked
+        │
+   ┌────┴─────────────────────┐
+blocked             review (human-approved mode)
+   │                          │
+  stop              pause → user approves → resume
+                              │
+                         safe / approved
+                              │
+                              ▼
+             executor.executeAction(action)  ──→  result string
+                              │
+                              ▼
+                   session.history.push(result)
+                   → next tick
+```
+
+The planner uses a **reflection loop** before each action: it evaluates what happened last step, maintains working memory across steps, and states its next goal — giving the agent much better multi-step reasoning.
+
+---
+
+## Install
+
+```bash
+npm install @akshayram1/omnibrowser-agent
+```
+
+---
+
+## Quick start
+
+```ts
+import { createBrowserAgent } from "@akshayram1/omnibrowser-agent";
+
+const agent = createBrowserAgent({
+  goal: "Search for contact Jane Doe and open her profile",
+  mode: "human-approved",        // or "autonomous"
+  planner: { kind: "heuristic" } // or "webllm"
+}, {
+  onStep:             (result, session) => console.log(result.message),
+  onApprovalRequired: (action, session) => console.log("Review:", action),
+  onDone:             (result, session) => console.log("Done:", result.message),
+  onError:            (err,    session) => console.error(err),
+  onMaxStepsReached:  (session)         => console.log("Max steps hit"),
+});
+
+await agent.start();
+
+// After onApprovalRequired fires:
+await agent.resume();
+
+// Cancel at any time:
+agent.stop();
+```
+
+---
+
+## Planner modes
+
+| Mode | Description | When to use |
+|---|---|---|
+| `heuristic` | Zero-dependency regex planner. Works fully offline. | Simple, predictable goals — navigate, fill, click |
+| `webllm` | On-device LLM via WebGPU. Fully private, no API calls. | Open-ended, multi-step, language-heavy goals |
+
+### WebLLM with a custom system prompt
+
+```ts
+const agent = createBrowserAgent({
+  goal: "Fill the checkout form",
+  planner: {
+    kind: "webllm",
+    systemPrompt: "You are a careful checkout assistant. Never submit before all required fields are filled."
+  }
+});
+```
+
+See [docs/EMBEDDING.md](docs/EMBEDDING.md) for the full WebLLM bridge wiring guide.
+
+### Recommended WebLLM models
+
+- `Llama-3.2-1B-Instruct-q4f16_1-MLC` — fast, ~600 MB
+- `Llama-3.2-3B-Instruct-q4f16_1-MLC` — better quality, ~1.5 GB
+- `Phi-3.5-mini-instruct-q4f16_1-MLC` — strong quality, ~2 GB
+- `Mistral-7B-Instruct-v0.3-q4f16_1-MLC` — balanced quality, ~4.1 GB
+- `Qwen2.5-7B-Instruct-q4f16_1-MLC` — strongest quality, ~4.3 GB
+- `Llama-3.1-8B-Instruct-q4f16_1-MLC` — strong reasoning, ~4.8 GB
+
+Model availability can vary by WebLLM release/build; if one fails to load, use a smaller fallback like `Llama-3.2-1B-Instruct-q4f16_1-MLC`.
+
+### Bring Your Own Model (Colab + MLC)
+
+Use the notebook workflow to fine-tune and quantize a planner model for OmniBrowser:
+
+- [Notebook assets](notebook/README.md)
+- [Google Colab notebook](notebook/custom_quantized_llm_colab.ipynb)
+- [Starter dataset JSONL](notebook/data/omnibrowser_planner_train.jsonl)
+
+The workflow:
+
+1. Generate/expand a dataset in OmniBrowser prompt format (`notebook/scripts/generate_dataset.mjs`).
+2. Validate selector + action correctness (`notebook/scripts/validate_dataset.mjs`).
+3. QLoRA fine-tune `Qwen/Qwen2.5-1.5B-Instruct` in Colab.
+4. Merge and quantize with `mlc_llm convert_weight --quantization q4f16_1`.
+5. Upload to Hugging Face and load in WebLLM via custom `appConfig`.
+
+```ts
+import * as webllm from "@mlc-ai/web-llm";
+import { createWebLLMBridge } from "@akshayram1/omnibrowser-agent";
+
+const appConfig: webllm.AppConfig = {
+  model_list: [
+    {
+      model: "https://huggingface.co/your-account/omnibrowser-planner-q4f16_1-MLC",
+      model_id: "omnibrowser-planner-q4f16_1",
+      model_lib:
+        webllm.modelLibURLPrefix +
+        webllm.modelVersion +
+        "/Qwen2.5-1.5B-Instruct-q4f16_1-ctx4k_cs1k-webgpu.wasm",
+    },
+  ],
+};
+
+const engine = await webllm.CreateMLCEngine("omnibrowser-planner-q4f16_1", { appConfig });
+window.__browserAgentWebLLM = createWebLLMBridge(engine);
+```
+
+---
+
+## Agent modes
+
+| Mode | Behaviour |
+|---|---|
+| `autonomous` | All `safe` and `review` actions execute without pause |
+| `human-approved` | `review`-rated actions pause and emit `onApprovalRequired` — call `resume()` to continue |
+
+---
+
+## Supported actions
+
+| Action | Description | Risk |
+|---|---|---|
+| `navigate` | Navigate to a URL (http/https only) | safe |
+| `click` | Click an element by CSS selector | safe / review |
+| `type` | Type text into an input or textarea | safe / review |
+| `scroll` | Scroll a container or the page | safe |
+| `focus` | Focus an element | safe |
+| `wait` | Pause for N milliseconds | safe |
+| `extract` | Extract text from an element | review |
+| `done` | Signal task completion | safe |
+
+---
+
+## AbortSignal support
+
+```ts
+const controller = new AbortController();
+const agent = createBrowserAgent({ goal: "...", signal: controller.signal });
+agent.start();
+
+controller.abort(); // cancel from outside
+```
+
+---
+
+## Chrome Extension
+
+1. Build:
+
+```bash
+npm run build
+```
+
+2. Open `chrome://extensions`, enable **Developer Mode**, click **Load unpacked**, select `dist/`.
+
+3. Open any tab, enter a goal in the popup, pick a mode, and click **Start**.
+
+See [docs/DEPLOYMENT.md](docs/DEPLOYMENT.md) for publishing and CI pipeline details.
+
+---
+
+## Project structure
+
+```
+src/
+├── background/      Extension service worker — session management
+├── content/         Extension content script — runs in page context
+├── core/            Shared engine (planner, observer, executor)
+│   ├── planner.ts
+│   ├── observer.ts
+│   └── executor.ts
+├── lib/             npm library entry — BrowserAgent class
+│   └── index.ts
+├── popup/           Extension popup UI
+└── shared/          Types, safety, and parse utilities
+    ├── contracts.ts
+    ├── safety.ts
+    └── parse-action.ts
+```
+
+---
+
+## Changelog
+
+### v0.2.6
+
+- Reflection-before-action pattern (`evaluation → memory → next_goal → action`) — agent reasons about each step before acting
+- Working memory carried across ticks for better multi-step goals
+- `parsePlannerResult()` exported from the library
+- `systemPrompt` option in `PlannerConfig` — pass your own prompt without rewriting the bridge
+- Thought bubble (💭) messages in the live demo chat showing the agent's next intent
+
+### v0.2.4 — v0.2.5
+
+- CI pipeline: auto version bump on push to main
+- Removed page-agent dependency — reflection pattern implemented natively
+- Chatbot demo redesign: right-aligned user messages, typing indicator, tab navigation (CRM + Task Manager)
+- `parsePlannerResult()` and `PlannerResult` type exported from library
+
+### v0.2.2
+
+- SDK/extension separation: core logic in `src/core/` shared between extension and npm library
+- 22 unit tests across planner and safety modules
+- Action verification in executor (disabled-check, value-verify, empty-check)
+- `CandidateElement.label` from associated `<label>` elements
+- Retry loop with `lastError` fed back to planner on failure
+
+### v0.2.0
+
+- New actions: `scroll` and `focus`
+- Smarter safety: risk assessment checks element label/text
+- Improved heuristic planner with regex pattern matching
+- Better page observation: filters invisible elements, up to 60 candidates
+- Library API: `resume()`, `isRunning`, `hasPendingAction`, `onMaxStepsReached`, `AbortSignal`
+
+### v0.1.0
+
+- Extension runtime loop, shared action contracts, heuristic + WebLLM planner, human-approved mode
+
+---
+
+## Docs
+
+- [Embedding Guide](docs/EMBEDDING.md) — integrate into any web app
+- [Architecture](docs/arch.md) — layer-by-layer breakdown
+- [Deployment](docs/DEPLOYMENT.md) — npm publish, Vercel, Chrome extension, CI
+- [Roadmap](docs/ROADMAP.md) — planned features
+
+---
+
+## License
+
+MIT © Akshay Chame

package/README.md

@@ -5,96 +5,55 @@
 
 Local-first browser AI operator. Plans and executes DOM actions entirely in the browser — no API keys, no cloud costs, no data leaving your machine.
 
-[Live Demo](https://omnibrowser-agent.vercel.app/examples/chatbot/) · [Embedding Guide](docs/EMBEDDING.md) · [Architecture](docs/arch.md) · [Deployment](docs/DEPLOYMENT.md) · [Roadmap](docs/ROADMAP.md)
+[Live Demo](https://omnibrowser-agent.vercel.app/examples/chatbot/) · [GitHub](https://github.com/akshayram1/omnibrowser-agent) · [Embedding Guide](https://github.com/akshayram1/omnibrowser-agent/blob/main/docs/EMBEDDING.md) · [Roadmap](https://github.com/akshayram1/omnibrowser-agent/blob/main/docs/ROADMAP.md)
 
 ---
 
 ## Architecture
 
-```mermaid
-flowchart TB
-    subgraph DELIVERY["Delivery Layer"]
-        EXT["🧩 Chrome Extension\npopup + background worker"]
-        LIB["📦 npm Library\ncreateBrowserAgent()"]
-    end
-
-    subgraph ORCHESTRATION["Orchestration"]
-        BG["background/index.ts\nSession & tick loop"]
-        BA["BrowserAgent class\nrunLoop() / resume() / stop()"]
-    end
-
-    subgraph CORE["Core  (src/core/)"]
-        PL["planner.ts\nheuristicPlan() / webllm bridge\nplanNextAction()"]
-        OB["observer.ts\ncollectSnapshot()\nDOM candidates + visibility filter"]
-        EX["executor.ts\nexecuteAction()\nclick / type / navigate\nscroll / focus / wait"]
-    end
-
-    subgraph SHARED["Shared  (src/shared/)"]
-        CT["contracts.ts\nAgentAction · PageSnapshot\nAgentSession · PlannerResult"]
-        SF["safety.ts\nassessRisk()\nsafe / review / blocked"]
-        PA["parse-action.ts\nparseAction()\nparsePlannerResult()"]
-    end
-
-    subgraph OUTCOMES["Action Outcomes"]
-        direction LR
-        OK["✅ safe → execute"]
-        RV["⚠️ review → needs approval"]
-        BL["🚫 blocked → stop"]
-    end
-
-    subgraph PLANNERS["Planner Modes"]
-        direction LR
-        HP["Heuristic\nzero deps · offline\nregex patterns"]
-        WL["WebLLM\non-device · WebGPU\nwindow.__browserAgentWebLLM"]
-    end
-
-    EXT --> BG
-    LIB --> BA
-    BG -. "chrome.tabs.sendMessage" .-> CORE
-    BA --> CORE
-
-    PL --> OB
-    PL --> SHARED
-    OB --> SHARED
-    EX --> SHARED
-
-    SF --> OUTCOMES
-    PL --> PLANNERS
+```
+  Chrome Extension          npm Library
+  (popup + bg worker)       createBrowserAgent()
+         |                        |
+         +----------+-------------+
+                    |
+             Orchestration
+          (session & tick loop)
+                    |
+         +----------+----------+
+         |          |          |
+     observer   planner    executor
+    (DOM snap) (heuristic  (click/type/
+               /webllm)    navigate...)
+         |          |          |
+         +----------+----------+
+                    |
+                 safety
+           (safe/review/blocked)
 ```
 
----
-
-## How it works — one tick
+### One tick
 
 ```
 goal + history + memory
-        │
-        ▼
-observer.collectSnapshot()  ──→  PageSnapshot (url, title, candidates[])
-        │
-        ▼
-planner.planNextAction()    ──→  PlannerResult { action, evaluation?, memory?, nextGoal? }
-        │
-        ▼
-safety.assessRisk(action)   ──→  safe | review | blocked
-        │
-   ┌────┴─────────────────────┐
-blocked             review (human-approved mode)
-   │                          │
-  stop              pause → user approves → resume
-                              │
-                         safe / approved
-                              │
-                              ▼
-             executor.executeAction(action)  ──→  result string
-                              │
-                              ▼
-                   session.history.push(result)
-                   → next tick
+        |
+        v
+observer.collectSnapshot()  -->  PageSnapshot (url, title, candidates[])
+        |
+        v
+planner.planNextAction()    -->  PlannerResult { action, evaluation?, memory?, nextGoal? }
+        |
+        v
+safety.assessRisk(action)   -->  safe | review | blocked
+        |
+   blocked --> stop
+   review  --> pause (human-approved) --> user calls resume()
+   safe    --> executor.executeAction()
+        |
+        v
+   session.history.push(result) --> next tick
 ```
 
-The planner uses a **reflection loop** before each action: it evaluates what happened last step, maintains working memory across steps, and states its next goal — giving the agent much better multi-step reasoning.
-
 ---
 
 ## Install
@@ -135,10 +94,10 @@ agent.stop();
 
 ## Planner modes
 
-| Mode | Description | When to use |
-|---|---|---|
+| Mode        | Description                                         | When to use                                   |
+|-------------|-----------------------------------------------------|-----------------------------------------------|
 | `heuristic` | Zero-dependency regex planner. Works fully offline. | Simple, predictable goals — navigate, fill, click |
-| `webllm` | On-device LLM via WebGPU. Fully private, no API calls. | Open-ended, multi-step, language-heavy goals |
+| `webllm`    | On-device LLM via WebGPU. Fully private, no API calls. | Open-ended, multi-step, language-heavy goals |
 
 ### WebLLM with a custom system prompt
 
@@ -152,42 +111,44 @@ const agent = createBrowserAgent({
 });
 ```
 
-See [docs/EMBEDDING.md](docs/EMBEDDING.md) for the full WebLLM bridge wiring guide.
-
 ### Recommended WebLLM models
 
-- `Llama-3.2-1B-Instruct-q4f16_1-MLC` — fast, ~600 MB
-- `Llama-3.2-3B-Instruct-q4f16_1-MLC` — better quality, ~1.5 GB
-- `Phi-3.5-mini-instruct-q4f16_1-MLC` — strong quality, ~2 GB
-- `Mistral-7B-Instruct-v0.3-q4f16_1-MLC` — balanced quality, ~4.1 GB
-- `Qwen2.5-7B-Instruct-q4f16_1-MLC` — strongest quality, ~4.3 GB
-- `Llama-3.1-8B-Instruct-q4f16_1-MLC` — strong reasoning, ~4.8 GB
-
-Model availability can vary by WebLLM release/build; if one fails to load, use a smaller fallback like `Llama-3.2-1B-Instruct-q4f16_1-MLC`.
+| Model ID | Size | Notes |
+|----------|------|-------|
+| `Llama-3.2-1B-Instruct-q4f16_1-MLC` | ~600 MB | fastest |
+| `Llama-3.2-3B-Instruct-q4f16_1-MLC` | ~1.5 GB | fast |
+| `Phi-3.5-mini-instruct-q4f16_1-MLC` | ~2 GB | quality |
+| `Mistral-7B-Instruct-v0.3-q4f16_1-MLC` | ~4.1 GB | balanced |
+| `Qwen2.5-7B-Instruct-q4f16_1-MLC` | ~4.3 GB | strong |
+| `Llama-3.1-8B-Instruct-q4f16_1-MLC` | ~4.8 GB | strong |
+| `Qwen3-8B-q4f16_1-MLC` | ~5 GB | latest Qwen |
+| `gemma-2-9b-it-q4f16_1-MLC` | ~5.5 GB | Google Gemma |
+| `DeepSeek-R1-Distill-Llama-8B-q4f16_1-MLC` | ~5 GB | reasoning |
+| `Llama-3.1-70B-Instruct-q3f16_1-MLC` | ~35 GB | most capable (needs 24+ GB VRAM) |
 
 ---
 
 ## Agent modes
 
-| Mode | Behaviour |
-|---|---|
-| `autonomous` | All `safe` and `review` actions execute without pause |
+| Mode             | Behaviour                                                                 |
+|------------------|---------------------------------------------------------------------------|
+| `autonomous`     | All `safe` and `review` actions execute without pause                     |
 | `human-approved` | `review`-rated actions pause and emit `onApprovalRequired` — call `resume()` to continue |
 
 ---
 
 ## Supported actions
 
-| Action | Description | Risk |
-|---|---|---|
-| `navigate` | Navigate to a URL (http/https only) | safe |
-| `click` | Click an element by CSS selector | safe / review |
-| `type` | Type text into an input or textarea | safe / review |
-| `scroll` | Scroll a container or the page | safe |
-| `focus` | Focus an element | safe |
-| `wait` | Pause for N milliseconds | safe |
-| `extract` | Extract text from an element | review |
-| `done` | Signal task completion | safe |
+| Action     | Description                        | Risk           |
+|------------|------------------------------------|----------------|
+| `navigate` | Navigate to a URL (http/https only) | safe           |
+| `click`    | Click an element by CSS selector   | safe / review  |
+| `type`     | Type text into an input            | safe / review  |
+| `scroll`   | Scroll a container or the page     | safe           |
+| `focus`    | Focus an element                   | safe           |
+| `wait`     | Pause for N milliseconds           | safe           |
+| `extract`  | Extract text from an element       | review         |
+| `done`     | Signal task completion             | safe           |
 
 ---
 
@@ -203,19 +164,40 @@ controller.abort(); // cancel from outside
 
 ---
 
-## Chrome Extension
+## WebLLM bridge wiring
 
-1. Build:
+```ts
+import * as webllm from "@mlc-ai/web-llm";
+import { createBrowserAgent, parsePlannerResult } from "@akshayram1/omnibrowser-agent";
+
+const engine = await webllm.CreateMLCEngine("Phi-3.5-mini-instruct-q4f16_1-MLC");
+
+window.__browserAgentWebLLM = {
+  async plan(input) {
+    const { goal, history, lastError, memory, systemPrompt } = input;
+    const resp = await engine.chat.completions.create({
+      messages: [
+        { role: "system", content: systemPrompt || "You are a browser automation agent. Output only JSON." },
+        { role: "user",   content: `Goal: "${goal}"\nHistory: ${history.slice(-4).join(" -> ")}${memory ? "\nMemory: " + memory : ""}${lastError ? "\nLast error: " + lastError : ""}` }
+      ],
+      temperature: 0,
+      max_tokens: 200
+    });
+    return parsePlannerResult(resp.choices[0].message.content);
+  }
+};
 
-```bash
-npm run build
+const agent = createBrowserAgent({ goal: "Fill the checkout form", planner: { kind: "webllm" } });
+await agent.start();
 ```
 
-2. Open `chrome://extensions`, enable **Developer Mode**, click **Load unpacked**, select `dist/`.
+---
 
-3. Open any tab, enter a goal in the popup, pick a mode, and click **Start**.
+## Chrome Extension
 
-See [docs/DEPLOYMENT.md](docs/DEPLOYMENT.md) for publishing and CI pipeline details.
+1. `npm run build`
+2. Open `chrome://extensions`, enable **Developer Mode**, click **Load unpacked**, select `dist/`.
+3. Open any tab, enter a goal in the popup, pick a mode, and click **Start**.
 
 ---
 
@@ -226,68 +208,13 @@ src/
 ├── background/      Extension service worker — session management
 ├── content/         Extension content script — runs in page context
 ├── core/            Shared engine (planner, observer, executor)
-│   ├── planner.ts
-│   ├── observer.ts
-│   └── executor.ts
-├── lib/             npm library entry — BrowserAgent class
-│   └── index.ts
+├── lib/             npm library entry — createBrowserAgent()
 ├── popup/           Extension popup UI
 └── shared/          Types, safety, and parse utilities
-    ├── contracts.ts
-    ├── safety.ts
-    └── parse-action.ts
 ```
 
 ---
 
-## Changelog
-
-### v0.2.6
-
-- Reflection-before-action pattern (`evaluation → memory → next_goal → action`) — agent reasons about each step before acting
-- Working memory carried across ticks for better multi-step goals
-- `parsePlannerResult()` exported from the library
-- `systemPrompt` option in `PlannerConfig` — pass your own prompt without rewriting the bridge
-- Thought bubble (💭) messages in the live demo chat showing the agent's next intent
-
-### v0.2.4 — v0.2.5
-
-- CI pipeline: auto version bump on push to main
-- Removed page-agent dependency — reflection pattern implemented natively
-- Chatbot demo redesign: right-aligned user messages, typing indicator, tab navigation (CRM + Task Manager)
-- `parsePlannerResult()` and `PlannerResult` type exported from library
-
-### v0.2.2
-
-- SDK/extension separation: core logic in `src/core/` shared between extension and npm library
-- 22 unit tests across planner and safety modules
-- Action verification in executor (disabled-check, value-verify, empty-check)
-- `CandidateElement.label` from associated `<label>` elements
-- Retry loop with `lastError` fed back to planner on failure
-
-### v0.2.0
-
-- New actions: `scroll` and `focus`
-- Smarter safety: risk assessment checks element label/text
-- Improved heuristic planner with regex pattern matching
-- Better page observation: filters invisible elements, up to 60 candidates
-- Library API: `resume()`, `isRunning`, `hasPendingAction`, `onMaxStepsReached`, `AbortSignal`
-
-### v0.1.0
-
-- Extension runtime loop, shared action contracts, heuristic + WebLLM planner, human-approved mode
-
----
-
-## Docs
-
-- [Embedding Guide](docs/EMBEDDING.md) — integrate into any web app
-- [Architecture](docs/arch.md) — layer-by-layer breakdown
-- [Deployment](docs/DEPLOYMENT.md) — npm publish, Vercel, Chrome extension, CI
-- [Roadmap](docs/ROADMAP.md) — planned features
-
----
-
 ## License
 
 MIT © Akshay Chame