semantic-browser 1.3.0

Deterministic semantic runtime over live Chromium for LLM planners

Semantic Browser

Semantic Browser turns live Chromium pages into compact semantic "rooms" for LLM planners.

Release: v1.3.0 (Alpha)
Latest release tag format: see docs/versioning.md

Make browser automation feel less like parsing soup and more like an old BBC Micro text adventure.

• Live page -> structured room state
• DOM distilled into meaningful objects, not soup
• Built for agentic browser automation
• Token-efficient, deterministic, inspectable

@ BBC News (bbc.co.uk)
> Home page. Main content: "Top stories". Navigation: News, Sport, Weather.
1 open "News" [act-8f2a2d1c-0]
2 open "Sport" [act-c3e119fa-0]
3 fill Search BBC [act-0b9411de-0] *value
+ 28 more [more]

The planner replies with one action ID and the runtime executes deterministically. This means less confusion, less hallucination and ultimately significantly less cost.

What's New in v1.3.0

v1.3 is a substantial extraction overhaul focused on making Semantic Browser work reliably on modern framework-heavy, live-updating websites — the kind of pages (betting, trading, SPA dashboards) that previously caused timeouts, stale actions, or missing elements.

• Framework-agnostic element discovery — AngularJS (ng-click, ng-model), Vue (v-on:click, @click), Alpine.js (x-on:click), and arbitrary custom elements (<abc-tab>, <sbk-input>) are now discovered via a universal hyphenated-tag pass with framework attribute inference.
• Fuzzy structural settle — Pages with live-updating content (odds feeds, stock tickers, chat) no longer cause settle timeouts. A configurable tolerance (default 5%) replaces exact count matching, with auto-escalation to 10% after repeated resets.
• Stable fingerprints — Action IDs no longer include pixel position (rect.y), using DOM id and CSS selector instead. Eliminates stale action IDs from minor layout shifts.
• Smarter locator resolution — Volatile framework state classes (ng-pristine, ng-untouched, etc.) are stripped from CSS selectors before resolution. Input elements resolve via sanitized CSS first, avoiding custom element wrapper traps like <sbk-input>.
• Better result classification — Actions that produce positive side-effects alongside newly-appearing blockers (e.g. a betslip opening with role="dialog") are now correctly classified as "success" instead of false "blocked".
• Robust modal detection — Three-tier detection covering ARIA (with visibility checks), class-based heuristics, custom-element modals, and viewport-coverage heuristics. Dialog blockers now require >30% screen coverage to trigger.
• Increased budgets — Curated actions raised from 15→25, room budget 1K→2K chars, max elements 2K→4K. Complex pages no longer exhaust the action surface.
• SPA navigation awareness — URL changes during settle reset structural counters. SPA navigations always classify as "success".

Validated headful against Paddy Power (cookie dismiss → football navigation → add bet → fill stake → "Login & Place Bets" button found → match page → back) with 98.8% fingerprint stability on live betting pages.

See CHANGELOG.md for full release notes.

Why this is different (and why it now works)

Other browser tools give the LLM the same data in a different wrapper. Semantic Browser gives it a fundamentally different view.

• Plain-text room descriptions - prose, not JSON.
• Curated actions first - top 25 useful actions, then more if needed.
• Progressive disclosure - more gives full action list without flooding every step.
• Tiny action replies - action IDs, nav, back, done.
• Narrative history - readable previous steps, not noisy machine dump.
• Guardrails for reality - anti-repeat fallback, nav hardening, transient extract retry.
• Honest failure mode - if a site throws anti-bot gates, we say so and show evidence.

Cross-method comparator (shared 25-task pack)

Method	Success rate	Failures	Median speed (ms)	Planner input median (billable)	Planner output median (billable)	Payload token-est median (estimated)	Total effective context median (estimated)	Median browser/runtime calls	Indicative planner cost/request (USD)
Standard browser tooling	24% (6/25)	19	11,819.8	10,118	74	6,918	17,224	6.0	0.041005
OpenClaw browser tooling	72% (18/25)	7	10,514.2	6,833	66	5,219	12,078	6.0	0.022053
Semantic Browser	100% (25/25)	0	9,353.3	540	14	310	879	5.0	0.004036

To put costs into context, at 5 complex browser tasks/day over a year (1,825 tasks), the estimated planner cost is about $74.83/year for the standard browser approach vs $7.37/year for Semantic Browser, a difference of about $67.47/year.

This is a dramatic jump in a reference harness run, not a universal guarantee.

The last anti-bot loop in this pack now has a robust recovery path:

• capture challenge evidence (screenshot),
• try direct same-origin query route,
• then use a public read-only fallback endpoint when the primary UI is hard-blocked.

25 tasks across navigation, search, multi-step interaction, resilience, and speed.

If challenge/captcha is detected, the harness captures screenshot evidence and includes it in the LLM call.

Reproducibility artifacts:

• Protocol: docs/benchmark_protocol.md
• Manifest: benchmarks/manifest.json

Why Semantic Browser

• Semantic room output instead of DOM/JSON soup.
• Curated action surface for token-efficient planning.
• Deterministic action execution loop (observe -> act -> observe delta).
• Built-in blocker signaling and confidence reporting.
• Python API, CLI, and local service interfaces.

Install

Install directly from PyPI (no git clone required):

pip install --upgrade semantic-browser

Pin to this release explicitly:

pip install "semantic-browser==1.3.0"

Managed mode (recommended first run):

pip install "semantic-browser[managed]"
semantic-browser install-browser

Service mode:

pip install "semantic-browser[server]"

Quickstart

semantic-browser portal --url https://example.com --headless

Inside portal:

• observe summary
• actions
• act <action_id>
• back / forward / reload
• trace run-trace.json
• quit

More examples: docs/getting_started.md

Profile-Aware Runtime Modes

Persistent profiles are first-class in this release. Use them for serious long-running agent tasks where session continuity matters (SSO, cookies, extension state, trust signals).

• profile_mode=ephemeral: disposable context, best for stateless tasks.
• profile_mode=persistent: real reusable Chromium profile directory.
• profile_mode=clone: copy profile into temporary sandbox before run.

CLI launch examples:

# Ephemeral (default)
semantic-browser launch --headless

# Persistent profile (recommended agent mode)
semantic-browser launch --headless --profile-mode persistent --profile-dir "/path/to/chrome-profile"

# Clone mode (safe experimentation)
semantic-browser launch --headless --profile-mode clone --profile-dir "/path/to/chrome-profile"

Storage state can still be used in ephemeral mode:

semantic-browser launch --headless --profile-mode ephemeral --storage-state-path state.json

Note: storage state bootstrap is not equivalent to a real profile.

Breaking API Change (v1.1+)

Launch config no longer accepts user_data_dir.

• removed: user_data_dir
• added: profile_mode, profile_dir, storage_state_path

If you previously passed user_data_dir for storage state, migrate to storage_state_path in ephemeral mode. If you intended a true browser profile, use profile_mode=persistent with profile_dir.

Python Usage

import asyncio
from semantic_browser import ManagedSession
from semantic_browser.models import ActionRequest

async def demo() -> None:
    session = await ManagedSession.launch(headful=False)
    runtime = session.runtime
    await runtime.navigate("https://example.com")
    obs = await runtime.observe(mode="summary")
    first_open = next((a for a in obs.available_actions if a.op == "open"), None)
    if first_open:
        result = await runtime.act(ActionRequest(action_id=first_open.id))
        print(result.status, result.observation.page.url)
    await session.close()

asyncio.run(demo())

CLI Reference

semantic-browser version
semantic-browser doctor
semantic-browser install-browser
semantic-browser launch --headless
semantic-browser attach --cdp ws://127.0.0.1:9222/devtools/browser/<id>
semantic-browser observe --session <id> --mode summary
semantic-browser act --session <id> --action <action_id>
semantic-browser inspect --session <id> --target <target_id>
semantic-browser navigate --session <id> --url https://example.com
semantic-browser export-trace --session <id> --out trace.json
semantic-browser serve --host 127.0.0.1 --port 8765 --api-token dev-token

Ownership and Attach Safety

Runtime sessions now carry explicit ownership semantics:

• owned_ephemeral: runtime may close browser/context/page.
• owned_persistent_profile: runtime closes browser process only; never deletes profile data.
• attached_context: runtime does not close external browser/context by default.
• attached_cdp: runtime does not close external Chrome by default.

If you explicitly want destructive close behavior in attached modes, use force_close_browser().

Service Security Defaults

• Localhost-focused CORS defaults.
• Optional token auth via SEMANTIC_BROWSER_API_TOKEN / X-API-Token.
• Idle session TTL cleanup.

Benchmarks

Benchmark numbers are reference harness runs, not universal guarantees.

• Protocol: docs/benchmark_protocol.md
• Manifest: benchmarks/manifest.json

Open Source Docs

• docs/getting_started.md
• docs/real_profiles.md
• docs/versioning.md
• docs/publishing.md
• CHANGELOG.md
• LICENSE
• CONTRIBUTING.md
• SECURITY.md