deltavision 1.0.7

Delta-first computer-use agent framework: CV + DOM hybrid observation with a zero-LLM gating pipeline

DeltaVision

Observation middleware for computer-use agents. Sits between your browser and your model. Sends the model only what changed on screen instead of a fresh screenshot every step. The agent still reasons — it just reasons about less.

60-second install + use

pip install deltavision

import deltavision

obs = deltavision.DeltaVisionObserver()

# Give it a screenshot (bytes / PIL Image / base64 / data URL),
# the current URL, and a description of the last action.
result = obs.observe(screenshot_png_bytes, url="https://example.com", last_action="click #submit")

print(result.obs_type)              # "full_frame" or "delta"
print(result.model_facing_tokens())  # tokens DV is about to ship to the model
print(result.dv_internal_tokens())   # tokens DV consumed internally (always full frame)
# Savings on this step: 1 - model_facing / dv_internal.
# (estimated_image_tokens() is a deprecated alias of model_facing_tokens(); see Cost accounting.)
result.to_anthropic_tool_result_content()  # ready-to-send Claude content blocks

That's it. Plug the observe() call wherever your agent currently sends a screenshot to the model. Every adapter (to_anthropic_*, to_openai_*, to_browser_use_*, to_stagehand_*) returns content already shaped for that SDK's vision API. See examples/integration_tests.py for live-API proof on all four.

Verify your install in 10 seconds

deltavision selftest

Runs 9 staged checks (import → observer → delta path → coverage guard → adapter → HTTP sidecar round-trip). Each stage reports ✓ or ✗ with what it checked, so a failure points directly at the broken piece. No API keys, no network, all synthetic frames.

DeltaVision self-test — staged E2E
  S1  import deltavision                              ✓  v1.0.7
  S2  observer construction                           ✓  26 fields
  S3  initial observation → full_frame                ✓  tokens=1365
  S4  small delta is cheaper than full frame          ✓  tokens=171 vs FF=1365 (87.5% saved)
  S5  whole-frame change → coverage guard fires       ✓  tokens=1365 = FF 1365 (trigger='crop_covers_frame')
  S6  anthropic adapter output is well-formed         ✓  1 content blocks
  S7  HTTP /health reports package version            ✓  version=1.0.7
  S8  HTTP /observe round-trips a DVObservation       ✓
  S9  HTTP /reset clears observer state               ✓
  All 9 stages passed.

Run as an HTTP service (non-Python agents)

python -m server --port 9000

# Then from any runtime:
curl http://localhost:9000/health                 # → {"status":"ok","version":"1.0.7"}
curl -X POST http://localhost:9000/observe \
     -F file=@screenshot.png -F url=... -F format=anthropic

Endpoints: GET /health · POST /observe · POST /reset · GET /state. Returns adapter-formatted JSON (anthropic / openai / browser_use / stagehand / raw). See server.py for the full API.

Namespace layout caveat

Because v1.0.x maintains backwards compatibility with the flat-module imports (from observer import ...), the wheel ships both a deltavision/ umbrella package AND the raw modules (observer.py, vision/, agent/, model/) at site-packages root. If your CWD has a directory named vision/ or observer.py, it can shadow the installed one — run your script from a different directory or use from deltavision import X (which always resolves through the umbrella). v2.0 will nest modules under the package to remove this caveat.

PyPI: deltavision==1.0.7 · OS-level companion (V2, alpha): deltavision-os · Source of truth: private repo, public mirror auto-synced

What's new in 1.0.7

Cost-accounting honesty + regression fix. No new claims, just makes the existing claims auditable.

• Cost split exposed in the public API. DVObservation.model_facing_tokens() (what DV ships to the model — drives savings claims) and DVObservation.dv_internal_tokens() (what DV consumes internally — always full-frame size) are now first-class methods on every observation. The pre-split estimated_image_tokens() is preserved as a deprecated alias of model_facing_tokens() (will be removed in v1.1.0). See Cost accounting for the formula and why it matters.
• BenchmarkTrace schema + deltavision verify-trace CLI. Every DV-MCP proxy run now emits a structured dv_trace_v1_<run_id>.jsonl next to its legacy log. The trace records per-step dv_internal_tokens, model_facing_tokens, payload_image_sha256 (canonical manifest hash), frame_sha256, classifier trigger, and crop bboxes. Anyone can run deltavision verify-trace path/to/trace.jsonl to validate that the trace's stated savings are internally consistent — schema-valid, hashes match, totals add up, no fabricated payloads. The credibility primitive: trust the trace, not the marketing.
• BUG-0006 fix: greedy bbox-merge default flipped to OFF. v1.0.6 shipped BBOX_MERGE_ENABLED=True as the public default, which inflated model_facing_tokens on benchmarks where only 1–3 contours survive MAX_REGIONS=6 truncation anyway (the merger ran before truncation, adding cost the truncation would have dropped for free). On the scripted-77 spreadsheet benchmark this measured a 77.2% → 72.1% regression. The fix: default off in the library, opt-in for the DV-MCP proxy (which faces fragmented-diff agent workloads where the merger genuinely helps). Full truncate-then-merge fix tracked for v1.1.0.
• BUG-0008 fix: dv_internal_tokens() no longer silently returns 0. A hand-built DVObservation missing consumed_frame_size used to return 0, which would silently inflate any 1 − model_facing / dv_internal calculation to ~100%. Now it falls back to frame.size for full-frame observations and raises ValueError for delta observations (where the original frame is unavailable). Loud failure beats silent zero.
• Public bug log. bugs/log.jsonl (source) + BUGS.md (rendered) document every bug found during development, with severity, repro, root cause, and fix commit. Currently 0 open. The point is the narrative — engineering discipline, not broken software.

The trace + verifier + cost split are the spine for the next release line: paired A/B harnesses and adapter-level "no full frame in delta payload" invariants both build on this in v1.1.0.

What's new in 1.0.6

Two CV-pipeline upgrades that, together, flip DV from "wins on tokens, loses on steps" to "wins on both axes" — the criterion the project requires:

• Greedy bbox-merge optimizer (vision/diff.py). The proxy used to emit up to MAX_REGIONS=6 separate crop bboxes per step. On steps with one big real change plus several scattered tiny ones, this fragmented into 6 expensive crops totaling > 1365 tokens, tripping the token-cap fallback and forcing a full frame. The merger greedily pairs bboxes while cost(A∪B) < cost(A) + cost(B). Replayed across 16 saved proxy runs: 38.1% → 48.9% mean savings, no classifier changes. Toggleable via DeltaVisionConfig(BBOX_MERGE_ENABLED=True); the public observer default is off after BUG-0006, while dv_playwright_mcp.py opts in for fragmented-diff agent workloads.
• Periodic full-frame refresh (dv_playwright_mcp.py). The proxy now forces a full-frame response every DV_DELTA_REFRESH_EVERY=5 consecutive deltas, so the agent re-anchors instead of getting lost in dialog interactions because it's only seeing tiny crops. Combined with a leaner screenshot prompt this dropped agent step count from 49 → 32 on the SF mapsheets task (vs FF's 45 steps). On the same trace, dv_internal_total_tokens=61,425 (45 steps × 1,365 each — same as FF's required full-screenshot cost), model_facing_total_tokens=28,725, so the cost-split savings is 1 − 28,725/61,425 = 53.2%, with a per-step model-facing-vs-internal savings of 34.2%. The agent took fewer steps AND DV shipped less to the model — wins on both axes the project requires.

When DeltaVision helps — and when it doesn't

DeltaVision's savings are task-shape-dependent. Reading the headline numbers without this context will over-sell the tool.

Task shape	DV savings	Why
Sticky-context workflows (forms, SPAs, spreadsheets, multi-tab user tasks — same page, many small interactions)	40–77%	agent re-reads the same page; small deltas dominate
Mixed browsing (typical CU workload: clicks + typing + occasional nav)	20–40%	some steps hit the delta path, some hit new_page
Scroll-heavy media exploration (WebVoyager-style news/feed sites)	5–20%	scroll-bypass gate fires; delta crops are near-viewport
Nav-heavy research (URL-hop every 1–2 steps, big scrolls)	~0%	every observation is a full frame by correct design — no redundant context to strip

Dogfood-measured, not modeled. The ~0% case is a real sibling-agent A/B (2026-04-19, 10-step deep-research trajectory, URL nav + 600 px scrolls) — DV stayed on the full_frame path on every step by design, because the classifier correctly identified that each observation was genuinely a new page or a new viewport of content.

Sweet spot: agents that re-read the same page. Form-heavy flows, SPA interactions, multi-tab comparison tasks. Not ideal: "open 10 new tabs, read each, close them."

This is the tool's shape, not a bug. The CV classifier can't compress what isn't redundant.

Cost accounting — what "savings" means here

Every DV trace records two distinct token costs per step. They mean different things and conflating them is the most common way to overstate or misstate DV's effect:

• dv_internal_tokens — what DV consumed internally on this step. Always equals the cost of the full screenshot at the consumed viewport size. DV needs the full frame in memory to compute the next-step diff, so this number does not go down regardless of what the model sees. This is the infrastructure cost of running the pipeline.
• model_facing_tokens — what DV actually shipped to the model on this step. On a delta step this is just the changed crops; on a full-frame step it equals dv_internal_tokens. This is the number that drives user-visible token spend on Claude, GPT, etc.

Every "X% savings" claim in this README is computed as

savings = 1 − sum(model_facing_tokens) / sum(dv_internal_tokens)

over the trace. The infrastructure cost (dv_internal_tokens) is not reduced — DV is observation middleware, not a screenshot-skipper. What it reduces is the bytes that reach the model.

If you read a saved benchmark JSON, the explicit keys are dv_internal_tokens / model_facing_tokens per step and dv_internal_total_tokens / model_facing_total_tokens / total_savings_pct in the summary. Older benchmark outputs use ff_tokens / dv_tokens (per step) and ff_total_tokens / dv_total_tokens (summary) as legacy aliases — they map exactly to dv_internal / model_facing respectively, kept for one release.

DVObservation.model_facing_tokens() and DVObservation.dv_internal_tokens() are the public methods. The pre-split DVObservation.estimated_image_tokens() is now a deprecated alias of model_facing_tokens() — kept for one release for back-compat, will be removed in v1.1.0.

Every claim below should be reproducible by running the named benchmark and checking the resulting JSON's total_savings_pct against the number in the table.

Headline demos — two videos, two honesty tiers

(1) Real Google Maps apartment search — 38.4% savings on the real web

A scripted 11-step trajectory on live Google Maps: search Brooklyn apartments, scroll listings, click into two real listing detail pages (461 Dean Apartments, The Bay NYC Luxury), scroll details, zoom map. Every observation runs through the live DeltaVision CV pipeline on the real screenshot — no mocks, no hand-coded pages.

	Full Frame	DeltaVision	Savings
Image tokens	15,015	9,249	38.4%
Full-frame obs	11 (every step)	6 (URL changes + zoom + scroll guard)	—
Delta obs	—	5	—

Full Frame column = dv_internal_total_tokens (cost of all 11 full screenshots DV consumed). DeltaVision column = model_facing_total_tokens (what DV shipped to the model). 38.4% = 1 − 9,249 / 15,015.

Video walkthrough: benchmarks/ablation/video_frames/gmaps_demo_v1.mp4 (68 s, 1080p60) — includes live counters + scene labels. Source metadata: gmaps_demo_v1.metadata.json (per-step timing, cumulative tokens, obs_type + trigger).

Run it yourself (real Google Maps, no auth, no API key):

python examples/gmaps_demo.py      # produces runs_gmaps/browser.webm + metadata.json

The 38.4% is in the mixed-browsing band of the task-shape matrix above — honest for a real user workflow where the agent does both navigation (full_frame) and same-page reading (delta).

(2) Real 2-site workflow: Maps → Sheets — 54% savings on live sites

A scripted 21-step workflow on two live, unmodified sites: Google Maps (research phase, 7 steps — search apartments, open two listings, scroll details) → Google Sheets (document phase, 14 steps — type findings into a real anonymous-edit spreadsheet). No mocks. The two phases run in separate browser contexts so the recording shows each site correctly.

	Full Frame	DeltaVision	Savings
Image tokens	28,665	13,179	54.0%
Research phase (Maps nav, 7 steps)	9,555	8,286	~13% — nav-heavy, expected
Document phase (Sheets fills, 14 steps)	19,110	4,893	~74% — delta-heavy
Full-frame obs	21 (every step)	7	—
Delta obs	—	14	—

Full Frame column = dv_internal_total_tokens, DeltaVision column = model_facing_total_tokens, 54.0% = 1 − 13,179 / 28,665.

The split tells the whole story: nav-heavy = near-zero savings, sticky-context = 74%. The combined 54% is an honest mixed-task number.

Run it yourself:

python examples/multitab_real_demo.py   # produces runs_multitab_real/browser.webm + metadata.json

(3) Scripted 3-tab workflow — 67% savings on local mocks (compression ceiling)

	Full Frame	DeltaVision	Savings
Image tokens	39,585	13,076	67.0%
Per-step avg	1,365	451	67.0%

Full Frame = dv_internal_total_tokens, DeltaVision = model_facing_total_tokens, 67.0% = 1 − 13,076 / 39,585. This is on a local HTML mock (no network, no real-site quirks) — that's why it's labeled the compression ceiling.

Video walkthrough: benchmarks/ablation/video_frames/apartment_demo.mp4 (32 s, 1080p60)

This version uses local HTML mocks designed to exercise DV's sweet spot. The 67% shows the compression ceiling when a task has many same-page interactions. For how real agents on real sites actually do, see video (1) above.

cd examples/multitab_apartment_demo/mocks && python3 -m http.server 8765  # Terminal 1
cd examples/multitab_apartment_demo && python3 run_multitab_demo.py       # Terminal 2

Why This Matters

Standard computer use agents send a full 1280x900 screenshot (~1600 tokens) on every step, whether 1 pixel changed or the entire page swapped. DeltaVision puts a 4-layer CV classifier in front of the model that decides: did the page change, or just a region? Send accordingly.

Four headline benchmarks across different task shapes:

Benchmark	What it measures	Steps	DV cost	FF cost	Savings	Reproduce
Maps→Sheets agent benchmark (live Claude agent, Haiku 4.5, CDP-measured)	real agentic sticky-context task	53	14,115 tok	72,345 tok	80.5%	python benchmarks/mapsheets/run_bench.py --trial N
Real 2-site workflow (Google Maps → Google Sheets, live sites)	real multi-tab CU task	21	13,179 tok	28,665 tok	54.0%	python examples/multitab_real_demo.py
Multi-tab apartment workflow (3 tabs, deterministic script)	realistic 3-tab user task	29	13,076 tok	39,585 tok	67.0%	python examples/multitab_apartment_demo/run_multitab_demo.py
Spreadsheet (deterministic, local HTML mock)	compression ceiling	25	7,780 tok	34,125 tok	77.2%	python examples/spreadsheet_observation_cost.py
TodoMVC matched-trajectory (real Playwright + Anthropic tool_result)	compression with real SPA	9	6,133 tok	13,824 tok	55.6%	python examples/observer_integration_proof.py
TodoMVC head-to-head (real Claude agent, n=3 per side)	utility — real agent decisions	7 each	23,693 ±66 tok	62,270 ±218 tok	62.0%	python benchmarks/headtohead/run_head_to_head.py

In every row, DV cost = model_facing_total_tokens (what DV shipped to the model) and FF cost = dv_internal_total_tokens (what DV consumed internally — the same number FF would have shipped). Savings = 1 − DV / FF per row. See Cost accounting.

Read these together. They answer three different questions:

• First: byte-reproducible on any machine — no agent, no trajectory variance, no auth. Shows the per-observation compression ceiling.
• Second: matched-trajectory with a real agent-shaped payload. Shows compression survives real tool_result plumbing.
• Third: actual utility — DV-wrapped Claude vs FF-baseline Claude on the same task, same model, n=3 each. 3/3 success on both sides, identical step counts, ±66 token variance on DV (deterministic). DV saves 62% input tokens with no reliability penalty.

Architecture note: CV + DOM hybrid (added in v4)

Pure CV observation couldn't reliably detect two things: (1) focus state after clicking an input (cursor blinker is sub-pixel, below diff threshold), and (2) small interactive elements in the low-res thumbnail (e.g. 20 px TodoMVC checkbox). Both are cheap to query from the DOM (~300 tokens as structured text). DV now runs one page.evaluate() per step that returns the visible clickable elements (bbox + label) and the currently-focused element — ground truth the CV pipeline can't produce. The agent uses these coordinates directly instead of guessing from pixels. This raised head-to-head DV success from 2/3 → 3/3 with lower total tokens (agent wastes fewer steps on retry-after-false-failure). See vision/elements.py.

Savings grow with task length. On long SPA workflows with sticky page structure, DV stays on the DELTA path for 80%+ of steps, each of which costs 3-7× less than a full frame.

Integration tests

Every adapter claim is covered by a runnable script. Reproduce with python examples/integration_tests.py:

Framework	What's verified	Proof
Anthropic tool_result	Live API call via claude-sonnet-4-20250514 ingests DV's content blocks without error	1,762 in / 50 out tokens, real response
OpenAI CUA (Operator)	to_openai_computer_call_output() matches the computer-use spec	data URL decodes to valid PNG, type=computer_call_output
Browser Use	pip install browser-use + 5-line monkey-patch wires DV in; to_browser_use_screenshot_b64() returns a valid base64 PNG	Patches BrowserSession.get_browser_state_summary, see examples/browser_use_integration/
Stagehand	to_stagehand_middleware_parts() returns a valid list of typed content parts	Adapter method DeltaVisionObserver.to_stagehand_middleware_parts()

Artifact: examples/integration_test_results.json (commit-tracked; 4/5 pass, Skyvern skipped — not on PyPI).

How It Works

Browser Action
    |
    v
+--------------------------------------+
|  DeltaVision CV Pipeline (no LLM)    |
|                                       |
|  Layer 1: URL change (free)           |
|  Layer 2: Pixel diff ratio (numpy)    |
|  Layer 3: Perceptual hash (PIL)       |
|  Layer 4: Anchor template match (cv2) |
|  + Scroll bypass gate                 |
|  + Animation guard                    |
+------------------+-------------------+
                   |
          +--------+--------+
          |                 |
      DELTA path       NEW_PAGE path
    crops + diff     full screenshot
    (~400 tokens)    (~1600 tokens)
          |                 |
          v                 v
+--------------------------------------+
|  Any Model Backend                    |
|  Claude / GPT-4o / Ollama / Local     |
+--------------------------------------+

Classifier accuracy: 17/17 scenarios across 8 diverse websites with default config. No site-specific tuning needed. Source: benchmarks/generalization/results.json.

Quick Start

From PyPI (recommended):

pip install deltavision
# Backends are optional extras:
pip install "deltavision[claude]"   # Anthropic
pip install "deltavision[openai]"   # OpenAI
pip install "deltavision[ollama]"   # local Ollama VLMs
pip install "deltavision[all]"      # everything

From source (for development or to run benchmarks):

git clone https://github.com/ddavidgao/deltavision.git
cd deltavision
python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e ".[dev,all]"
playwright install chromium

# Run tests (no API keys needed) — 224 total, offline
pytest tests/ -q --ignore=tests/test_e2e_live.py --ignore=tests/test_live_capture.py

# The reproducible spreadsheet benchmark — no API key, no auth, deterministic
python examples/spreadsheet_observation_cost.py      # → 77.2% savings, same every run

# Integration tests against 4 CU frameworks (Anthropic live if ANTHROPIC_API_KEY set)
python examples/integration_tests.py

# Reaction time benchmark (pure CV, no model)
python benchmarks/reaction/run_reaction.py --rounds 5 --headless

# Classifier generalization test (17 scenarios, 8 sites)
python benchmarks/generalization/test_classifier_diverse.py

With a model backend

# Ollama (free, local, no API key)
ollama pull qwen2.5vl:7b
python main.py --task "Search Wikipedia for 'computer vision'" \
    --url https://en.wikipedia.org --backend ollama --model qwen2.5vl:7b --headless

# Claude API (Sonnet 4.6 default; Opus 4.7 and newer models work via --model)
export ANTHROPIC_API_KEY=sk-...
python main.py --task "..." --url ... --backend claude
python main.py --task "..." --url ... --backend claude --model claude-opus-4-7-20260417

# OpenAI
export OPENAI_API_KEY=sk-...
python main.py --task "..." --url ... --backend openai

# Ablation: same task without delta gating (for comparison)
python main.py --task "..." --url ... --backend ollama --model qwen2.5vl:7b --force-full-frame

Safety modes

python main.py --task "..." --url ... --safety strict       # block credentials, shorteners
python main.py --task "..." --url ... --safety educational   # allowlist edu sites only
python main.py --task "..." --url ... --safety permissive    # log warnings only

Architecture

deltavision/
  vision/           # CV pipeline: diff engine, pHash, 4-layer classifier, capture
  agent/            # Agent loop, state machine, typed actions
  observation/      # Builds typed observations (FullFrame or Delta)
  model/            # Pluggable backends: claude, openai, ollama, local, scripted
  safety.py         # Model-agnostic action validation
  config.py         # All thresholds in one place, site-specific presets
  results/          # SQLite result store (query with db.summary() or raw SQL)
  benchmarks/
    reaction/       # CV-only reaction time benchmark
    generalization/ # Classifier accuracy across diverse sites + visual frame capture
    ablation/       # DeltaVision vs full-frame controlled comparison
    sites/          # Benchmark site registry (7 sites, 3 difficulty tiers)
  tests/            # 232 tests: unit, integration, live Playwright, real screenshots
  paper/            # Paper outline with figure/table mapping to data

Testing

pytest tests/ — 293 tests total (offline + live Playwright). Covers CV pipeline, classifier cascade, observation builder, safety layer, response parsers, HTTP sidecar, v1.0.3 regression invariants (import deltavision works, DV ≤ FF on every single step), v1.0.5 token-cap guard (proxy never bills more than a full frame), v1.0.6 greedy bbox-merge optimizer + periodic full-frame refresh, v1.0.7 cost-split methods (model_facing_tokens() / dv_internal_tokens()) + BenchmarkTrace schema + verify-trace CLI invariants.

Suite	Tests	Covers
CV pipeline	34	diff, pHash, 4-layer classifier cascade, real McGraw-Hill frames
Model response parsing	33	JSON extraction, VLM output quirks (fences, preamble, nested confidence)
Safety layer	37	URL safety, credential detection, action limits, preset configs
Config validation	45	every threshold range, every field type, bbox coherence
Results store	19	SQLite save/query/best, schema, persistence across reopen
Integration	15	observation builder, action parser, agent state, simulated pipeline
Observer API	34	lifecycle + 5 format adapters (Anthropic/OpenAI/Browser Use/Skyvern/Stagehand)
Live (CI-skipped)	7	browser E2E, live capture
Total	224

pytest tests/ -q                    # full offline suite (224 collected, all pass)
pytest tests/test_safety.py -v      # single module
pytest tests/ --cov=. --cov-report=term-missing  # coverage

Using as a library

from playwright.async_api import async_playwright
from config import DeltaVisionConfig
from agent.loop import run_agent
from model.claude import ClaudeModel
from safety import STRICT

async def go():
    async with async_playwright() as pw:
        browser = await pw.chromium.launch()
        page = await browser.new_page()
        state = await run_agent(
            task="Find the capital of France on Wikipedia",
            start_url="https://en.wikipedia.org",
            model=ClaudeModel(api_key="sk-...", model="claude-sonnet-4-6"),
            browser_page=page,
            config=DeltaVisionConfig(MAX_STEPS=20),
            safety=STRICT,
        )
        print(f"Done in {state.step} steps, delta ratio: {state.delta_ratio:.1%}")
        await browser.close()

The run_agent function is the single entry point. All tunable behavior is in DeltaVisionConfig. Results are in state (dataclass with observations, responses, transition_log, step, done, etc.).

Results

All results stored in results/deltavision.db (SQLite). Query:

python -c "from results.store import ResultStore; ResultStore().summary()"

V2 (OS-level): matched-trajectory ablation — 68.2% savings

The sibling repo deltavision-os (mss + pyautogui for OS-level desktop agents) published its own independent ablation:

The token numbers below use the same accounting convention as the rest of this README — the forced-full-frame baseline is the dv_internal denominator and the delta-gated number is the model_facing numerator. See Cost accounting.

• Same 10-step trajectory run twice on a real Mac desktop, natural Qwen2.5-VL behavior. No cherry-picking — exact same actions each time.
• Forced full-frame: 17,600 image tokens
• Delta-gated: 5,600 image tokens
• 68.2% savings, zero difference in task outcome.

Plus a threshold sweep (3 trajectories × 3 values of NEW_PAGE_DIFF_THRESHOLD) empirically confirmed that the pHash layer dominates: diff-threshold in [0.30, 0.75] produced identical classifications. Paper-grade finding.

Raw data: deltavision-os/benchmarks/ablation_result.json and ablation_sweep_result.json. Plus a smoke-test run on the ScreenSpot-v2 community benchmark (Qwen2.5-VL-7B, 80% desktop accuracy on n=15) proves the V2 stack works end-to-end with a real VLM.

Repo: https://github.com/ddavidgao/deltavision-os

---

Browser side: savings depend heavily on the workload

All "token savings" percentages in the tables below follow the same accounting convention as the rest of the README: the baseline (full-frame) total is the dv_internal denominator, the DV total is the model_facing numerator, and savings = 1 − model_facing / dv_internal. See Cost accounting.

Two benchmarks, both same Anthropic tool_result format, both real Playwright, same DeltaVisionObserver wrapping — nothing else changes between baseline and DV. Savings vary by task type; here's the honest shape:

SPA + mixed browser tasks (5 tasks, 40 steps) — reproduce: python examples/multi_site_benchmark.py

Task	Steps	Token savings
TodoMVC: add 3 + filter	9	55.6%
TodoMVC: add 3, check 2, clear	10	63.6%
Wikipedia: search + navigate + scroll	9	47.6%
Hacker News: browse + scroll + open thread	7	26.2%
example.com idle	5	72.6%
Aggregate	40	52.8%

Scroll-dominated media exploration (10 WebVoyager sites, 70 steps) — reproduce: python examples/webvoyager_subset.py

Site	Steps	Token savings
huggingface	7	33.2%
wolfram	7	23.2%
cambridge	7	18.7%
github	7	18.2%
apple	7	12.7%
allrecipes	7	12.6%
bbc_news	7	10.3%
coursera	7	7.4%
arxiv	7	6.2%
espn	7	4.7%
Aggregate	70	14.7%

What the two benchmarks tell you

• DV excels on SPA / mixed-interaction tasks (55-65% savings). These are the real CU-agent workloads: click buttons, type in inputs, navigate with URL changes. Small region changes compress well into crops.
• DV's savings shrink on pure-scroll exploration (5-33% savings). The scroll_bypass gate correctly classifies scrolled frames as DELTAs, but the resulting delta crops are near-full-viewport because each scroll exposes a large band of new content. Token savings stay positive but the wire-byte savings can go slightly negative (PNG-of-thumbnail + PNG-of- large-crop can exceed a single PNG-of-full-frame).
• Best case (idle, static, tiny local change): 70%+ savings.
• Worst case (ESPN, full-page scrolling media): 4.7% token savings.

This is a fundamental property of the technique, not a bug: DeltaVision is an observation-level optimization for sticky-context workflows. On sites where every scroll reveals mostly-new pixels, there's less redundant observation to strip.

Practical takeaway for CU agent builders: if your workflow is mostly typing / clicking / form interactions, expect 40-70% token savings. If it's mostly scrolling through long feeds, expect 5-20%. Real agents do both, so a typical run lands between those endpoints (20-40% is a reasonable expectation for a mixed agent workload).

Multi-step ablation on Wikipedia (Qwen2.5-VL-7B)

Wikipedia search-and-navigate task. Both agents use the same model, same prompt, same browser.

	DeltaVision	Full-Frame (baseline)
Outcome	Task completed at step 3	Hit 50-step limit, did not complete
Image tokens used	~4,000	~81,600
Delta ratio	67%	0%

Caveat: the full-frame baseline didn't complete the task. Token counts are cumulative over the steps each agent actually executed, not directly comparable as "tokens to complete the same work." The meaningful claim is that DeltaVision completed a task that the full-frame path failed on, at a fraction of the per-step observation cost. Data: DB Runs 11/12.

Classifier Generalization

Site	Type	Scenarios	Accuracy
Wikipedia	Traditional nav	4	100%
HumanBenchmark	Dynamic content	3	100%
Hacker News	Minimal HTML	3	100%
Dynamic SPA	JS content injection	2	100%
Scroll test	Viewport shift	2	100%
TodoMVC	SPA (React)	1	100%
GitHub (public browse)	SPA (Turbo)	1	100%
example.com	Static minimal	1	100%
Total	8 sites	17	100%

Source: benchmarks/generalization/results.json. All scenarios run with default config, no per-site tuning.

Reaction Time (CV pipeline only, no model)

	DeltaVision (5 clean rounds)	Human Median	Claude CU (n=1)
Best	74ms	273ms	13,491ms
Average	100ms	273ms	13,491ms (single measurement)

Data sources: DeltaVision values from DB Run 10 (5 clean rounds of 10, fixed state machine, Windows RTX 5080). Claude CU baseline from results/humanbenchmark_reaction_20260414_220141.json (backend label: claude_standard_cu, 1 round — model version not recorded in the JSON).

Note: This measures CV pipeline speed (screenshot + color detect + click). The comparison to Claude CU is unfair — Claude runs full model inference per step — and the Claude baseline is n=1, so the "avg" cell repeats the single measurement. The reaction benchmark demonstrates that simple visual tasks don't need a model at all.

Launch video

A 75-second narrated walkthrough of DeltaVision — the problem, the pipeline, the compression ceiling, and the real-agent head-to-head that produce these numbers:

File	Content
benchmarks/ablation/video_frames/deltavision_v1_launch.mp4	1080p60, 9 scenes: title / problem / task setup / DV pipeline internals on one real observation / side-by-side showing what FF sends vs what DV actually sends (thumbnail + crop snippets) / savings range / compression ceiling (77.2%) / head-to-head utility (62.0% with real Claude agent) / install

Record your own agent session:

# Needs Ollama running with a VLM
ollama serve
python benchmarks/demo/record_comparison.py --task wikipedia
python benchmarks/demo/record_comparison.py --task wikipedia_multi

Videos are recorded by Playwright at 60fps. ffmpeg combines them side-by-side with labels.

Key Design Decisions

1. The model never decides transition type. The CV classifier is deterministic, sub-millisecond, and testable.
2. Speed comes from sending less, not skipping the model. The model still reasons; it gets cropped regions instead of full screenshots.
3. Safety is framework-level. Critical for uncensored local models that won't refuse dangerous actions.
4. Backend-agnostic. Same observation format for Claude (Sonnet 4.6, Opus 4.7), GPT-4o, Qwen, or UI-TARS. Add a backend in model/ with BaseModel's interface and DV's classifier output drops in unchanged.
5. Scroll-aware. Scrolling shifts the viewport but doesn't change page state. The classifier knows this.
6. Animation-resistant. Subtle animations (spinners, fades) don't trigger false page transitions.

Troubleshooting

"ANTHROPIC_API_KEY not set" — copy .env.example to .env and fill in your key, or export it: export ANTHROPIC_API_KEY=sk-.... The CLI also loads .env from the project root automatically.

Ollama connection refused — start the server first: ollama serve in another terminal. Check the model is pulled: ollama list. Default host is http://localhost:11434.

ModuleNotFoundError: No module named 'numpy' — the project venv is separate from system Python. Use .venv/bin/python3 explicitly or activate the venv first. macOS's system Python is externally managed.

Playwright browser not found — run playwright install chromium after the first pip install.

Classifier misbehaves on a custom site — dump meta.json from benchmarks/generalization/frames/ to see what the CV pipeline measured. Tune PHASH_DISTANCE_THRESHOLD (default 20) or NEW_PAGE_DIFF_THRESHOLD (default 0.75) in config.py. All thresholds are validated at construction — bad values raise ConfigError immediately.

ConfigError at startup — you set a threshold out of range. Every field has documented bounds in config.py::DeltaVisionConfig.__post_init__. The error names the field and the valid range.

V1 vs V2

This is V1 (browser-focused) — on PyPI as deltavision. If you need OS-level or OSWorld-VM observation, see deltavision-os (on PyPI as deltavision-os, currently 0.1.0a0 alpha). V1 is the stable browser middleware; V2 extends the same CV pipeline to the full OS desktop via mss + pyautogui.

License

MIT