3tears-scrape 0.23.0

AI-driven, self-healing web scraping: pluggable render backends + LLM-proposed extraction candidates validated and judged against real page content, never hand-rolled per-site parsers

3tears-scrape

AI-driven, self-healing web scraping. Onboarding a new site is a config addition (a ScrapeTarget row), never a hand-written parser: an LLM proposes extraction candidates, each is structurally validated against the real page, an LLM judge picks the winner by comparing candidate values against real page content, and the winning strategy is persisted as a recipe reused on every later fetch until it stops working.

pip install 3tears-scrape

What you get

• ScrapeDriver -- a pluggable, backend-agnostic render contract (RenderedPage, NavStep, NetworkCall). The drivers/ package is the current set -- a list here goes stale the first time one is added. Today: NodriverSidecarDriver (a real headful Chromium against an Xvfb display, in an isolated HTTP sidecar -- see sidecar/; headful rather than headless because real sites treat the two differently, and because a headful display is what a human can be handed when a page needs one), CamoufoxDriver (in-process stealth Firefox, no sidecar needed), DocumentDriver (PDF/DOCX/XLSX/CSV/TXT/MD/LaTeX via 3tears-agent-tools' parse_document), ApiDriver (stateless JSON API GET), NetworkCaptureDriver (authenticated in-session XHR capture for pages whose real data never reaches a statelessly-fetchable API), MultiDocumentDriver (a listing of links that each point at one whole document rather than a row, fetched via an injected inner driver and concatenated into one combined page, skipping documents already seen on a prior poll), ListingDetailDriver (a listing table whose rows each link to a per-record detail page, resolved row-scoped and merged back into one flat synthetic table -- detail value preferred, listing value as fallback, neither fabricated), and NodriverDownloadDriver (a document behind a real bot challenge a plain HTTP client can't pass, fetched through the sidecar's forced-download endpoint).
• Extraction-strategy shapes (eval_loop.StrategyType) -- css selector candidates for real (or synthesized) HTML tables; regex candidates for text-block/prose pages with no table structure at all; per_document for independently-worded documents that share no template a single cached pattern could ever generalize across, which get a fresh extraction each instead of a reused recipe, plus a judge; and multi_row_vision for a table whose own structure defeats text-based table extraction and needs the whole table read visually at once. css and regex run the propose -> structurally-validate -> judge -> persist cycle, and their recipe is genuinely reused: while it keeps validating, a poll costs no LLM call at all. The judge only ever sees extracted values and real page content, never which mechanism produced them. per_document and multi_row_vision deliberately do NOT: neither has a cached pattern to reuse, so each pays LLM calls every poll, and more than one apiece. per_document spends, per document, an extraction plus an unconditional grounding judge -- so two calls at minimum. A born-digital document can cost more, because that path chunks the extraction by field count; an OCR'd one is a single vision call regardless of schema size. Read the chunk size in extraction.py rather than assuming, since it decides the multiplier. multi_row_vision spends two calls over the whole table: an extraction, then the same kind of judge. Budget from the code rather than from a number here: a schema-scaled cost is easy to under-read as one call per item. Both still persist a marker recipe, for operational visibility rather than for reuse, so a recipe row existing is not the same as a recipe being reused. That is the cost of a target whose structure nothing learned once can serve, and the reason a strategy is chosen deliberately. Chosen explicitly per target, never auto-detected -- two superficially identical page shapes have needed opposite strategies.
• run_eval_loop / run_eval_loop_multi_row -- the self-healing cycle itself, for the css and regex strategies: reuse a target's existing ScrapeRecipe with zero LLM calls while it keeps validating (the per_document and multi_row_vision branches have no reusable pattern and pay several calls per poll, as costed above); regenerate candidates once consecutive_validation_failures crosses a threshold, or immediately when a failure is classified as the page having genuinely changed (see challenge.py), since waiting two more polls for evidence already in hand is pure latency.
• ScrapeTarget / ScrapeRecipe / ScrapeExtraction / ScrapeTargetHealth -- the domain-agnostic persisted entities (three-tier L1/L2/L3-backed collections). ScrapeTargetHealth is per-target FETCH health, deliberately separate from the recipe: it exists for targets that never had a recipe at all, such as one walled off before it ever extracted successfully. The core never learns what a field means -- only that whatever a candidate extracts for a caller-supplied field name parses as its declared type.
• Pluggable target config -- TargetSource (StaticTargetSource / YamlTargetSource / CollectionTargetSource) and bootstrap_targets() for seeding a database-backed target collection from a git-tracked YAML file, never overwriting a row already present.
• ScrapeTool -- an ad-hoc, single-call MCP-exposed entry point: give it a URL and a field schema, get back real extracted data through the same eval loop, no target pre-configuration required.
• enrichment.py -- a secondary, separate LLM pass capturing free-form context a structured schema has no field for, kept distinct from validated structured data.
• parse_form() / build_form_post() -- deterministic, browser-free HTML form parsing and postback serialization: reduce a form to its POST target, its default "successful controls", and its submit controls, then serialize a replayable body. For the many server-rendered portals (the classic case being ASP.NET WebForms' __VIEWSTATE/__EVENTVALIDATION) that only return rows after a form POST, this replaces driving a browser with a plain HTTP round-trip. Pure and side-effect-free -- no network, no LLM, no hardcoded framework or field meaning.
• find_target_page() -- a bounded-turn research agent: give it a plain-language query ("Ohio WARN Act notices"), it searches and fetches candidate pages (WebSearch/WebFetch, capped turns), deterministically verifies the winner has real extractable structure (a table, a document link, a JSON API response -- never a browser fetch, so it never needs the sidecar running), and returns a PageFinderResult (url, a driver_backend guess, verified). Independently callable -- it never persists a ScrapeTarget or forces extraction to follow; a caller with a known URL never needs this at all.
• discover_candidates() / discover_row_candidates() -- "capture every variable on this page": the inverse of generate_candidates/generate_row_candidates, no caller-supplied field_schema required. An LLM proposes field names/types/selectors it finds on the page, each validated by the exact same validate_candidate/validate_row_candidate every schema-driven candidate already goes through -- a discovered field that doesn't structurally validate is dropped, never included. Returns a DiscoverySchemaResult whose field_schema/strategy are already shaped exactly as ScrapeTarget.field_schema/ScrapeRecipe.extraction_strategy expect. Never persists anything -- a caller decides what to do with the result.
• capture_request_shape() -- the "how" sibling to find_target_page()'s "what": drive a real browser session at a target and report back every XHR/fetch call it actually made, so an authenticated in-session API's real request shape is captured rather than guessed from remembered conventions. Deliberately does not pick "the one real call" out of the results -- a target's real data call is reliably neither the largest nor the first, so it returns everything and the caller decides.

Quickstart

from threetears.scrape.drivers.nodriver_sidecar import NodriverSidecarDriver
from threetears.scrape.eval_loop import run_eval_loop_multi_row
from threetears.scrape.collections import ScrapeRecipeCollection, ScrapeExtractionCollection

driver = NodriverSidecarDriver("http://localhost:8088")
page = await driver.render("https://example.gov/warn-notices")

extraction = await run_eval_loop_multi_row(
    "example_warn_notices", page.html, page.final_url,
    schema={"employer": str, "county": str, "affected_count": int},
    recipe_collection=ScrapeRecipeCollection(registry, config),
    extraction_collection=ScrapeExtractionCollection(registry, config),
    api_key=openrouter_api_key,
)
print(extraction.validation_status, extraction.structured_fields["records"])

The nodriver sidecar

NodriverSidecarDriver talks HTTP to a genuinely separate process/container -- this is the AGPL isolation boundary nodriver (AGPL-3.0) requires, since this package itself is MIT-licensed. Build and run it via docker buildx bake nodriver-sidecar from the repo root; the container definition, its pinned AGPL-3.0 licence, and its own contract tests all live in sidecar/. If you don't want the sidecar dependency at all, CamoufoxDriver is a fully in-process, MPL-2.0-safe alternative.

Architecture

Lifted from faidh's src/faidh/scrape/ into this package (docs/scrape-task-01-lift-core-package.md, 2026-07-15) as a directory move -- zero logic changes. faidh's WARN Act plugin is used throughout as the running example of a consumer; it remains a faidh-side module (faidh/src/faidh/intake/plugins/warn_act.py), not part of this package, and is the only place WARN-domain meaning exists.

Why this exists

WARN Act notices are published by ~24 US state labor departments in wildly inconsistent formats -- real HTML tables, PDF/DOCX/XLSX filings, prose text blocks, JSON APIs, and at least one state gated behind an authenticated search form with no stable URLs. The naive answer is "write a scraper per state." That's 24+ bespoke, brittle, un-composable parsers that all rot independently.

The actual answer: two orthogonal, config-selected axes -- how to fetch a page and how to extract from it -- cover every state as a combination of the two, with zero jurisdiction-specific code in the core.

1. Data flow: source → Postgres (faidh's WARN Act consumer, as an example)

flowchart TD
    A[ScrapeTarget config<br/>YAML seed row] --> B["ScrapeDriver.render(url, ...)"]
    B --> C[RenderedPage<br/>html, status, final_url, timing_ms, network_calls]
    C --> D{run_eval_loop /<br/>run_eval_loop_multi_row}
    D -->|reuse existing recipe| E[ScrapeRecipe]
    D -->|regenerate via LLM candidates + judge| E
    E --> F[ScrapeExtraction persisted<br/>structured_fields = records]
    F --> G["WarnActPlugin._produce()<br/>generic records → Tier-2 fields"]
    G --> H[ArbitrarySignalEntity.create<br/>signal_type=warn_act_notice]
    H --> I["PluginBase.collect()<br/>throttle-gate → _produce() → save_entity() → yield"]
    I --> J[(Postgres:<br/>faidh_arbitrary_signals)]
    I --> K["publish_arbitrary_signals()"]
    K --> L[NATS subject: arbitrary_signals]
    L --> M[Downstream: scoreboard, retrospect, ...]

    F -.-> N[(Postgres:<br/>scrape_targets / scrape_recipes /<br/>scrape_extractions / scrape_target_health -- provenance)]

Two Postgres-writing paths run in parallel, not sequentially:

• Provenance (scrape_targets, scrape_recipes, scrape_extractions, scrape_target_health) -- what was fetched, what strategy won, when, and whether the fetch itself is healthy. Domain-agnostic, lives in this package.
• Signal (faidh_arbitrary_signals) -- what faidh's forecasting actually reads. Faidh-specific, mapped by WarnActPlugin._produce(), stays in faidh.

2. The abstraction: ScrapeTarget is the entire adapter surface

Zero jurisdiction-specific Python exists anywhere in this package. In faidh's WARN Act consumer, every state is a ScrapeTarget row (faidh/src/faidh/intake/plugins/seeds/warn_act_targets.yaml), bootstrapped into the DB.

Field	Controls
url	what to fetch
driver_backend	which of the 8 ScrapeDriver implementations renders it
wait_for	CSS selector to settle on (async-loading pages)
nav_steps	ordered click/fill/wait_for/wait_ms/scroll_into_view/scroll_page/evaluate actions (search-form-gated, lazy-render, or in-page-JS-state targets)
timeout_seconds	per-target render budget
field_schema	field name → Python type, e.g. {"employer": str, "county": str}
extraction_strategy_type	"css", "regex", "per_document", or "multi_row_vision"
api_results_path / api_fragment_field	JSON traversal (api driver; also multi_document's JSON discovery mode)
link_selector	CSS selector matching the document links on a listing page (multi_document's HTML discovery mode)
multi_row	one record vs. many per page

The core (driver.py, extraction.py, eval_loop.py, collections.py) never branches on target identity -- only on ScrapeTarget fields and a caller-supplied field_schema. It doesn't know what employer means; it only enforces that whatever a candidate extracts for that field name parses as the declared type. Stated in extraction.py's own docstring: "this module never hardcodes what a field means."

3. Fetch × Extract = many combinations, not one scraper per site

flowchart LR
    subgraph Fetch["Axis 1 -- driver_backend (drivers/)"]
        D1[nodriver_sidecar<br/>headful Chromium on Xvfb, sidecar]
        D2[camoufox<br/>in-process stealth Firefox]
        D3[document<br/>PDF/DOCX/XLSX/CSV → parse_document]
        D4[api<br/>stateless JSON GET]
        D5[network_capture<br/>authenticated XHR capture]
        D6[multi_document<br/>link list → N whole documents<br/>combined into one page]
        D7[listing_detail<br/>listing rows + per-row detail pages<br/>merged row-scoped]
        D8[nodriver_download<br/>bot-challenged document<br/>via sidecar forced download]
    end
    subgraph Converge["Everything converges here"]
        R[RenderedPage.html<br/>real or synthetic -- identical downstream]
    end
    subgraph Extract["Axis 2 -- extraction_strategy_type"]
        S1[css<br/>real/synthesized table → LLM proposes selectors]
        S2[regex<br/>prose text block → LLM proposes named-group patterns]
        S3[per_document<br/>no shared template → extraction<br/>+ grounding judge per document, no cached recipe]
        S4[multi_row_vision<br/>table structure defeats text extraction<br/>→ vision read of the whole table]
    end
    D1 & D2 & D3 & D4 & D5 & D6 & D7 & D8 --> R
    R --> S1
    R --> S2
    R --> S3
    R --> S4
    S1 & S2 & S3 & S4 --> J[Identical propose → validate →<br/>judge → persist cycle, eval_loop.py]

In faidh's WARN Act consumer, every one of the ~24 onboarded states is one config row picking a combination from this matrix. A new driver/strategy is only written when a target needs a fetch mechanism or extraction shape that doesn't exist yet -- and each one that was written has a named real target behind it: the regex strategy (a prose listing with no table), ApiDriver (a JSON search endpoint), native CSV support, NetworkCaptureDriver (an authenticated Aura POST), MultiDocumentDriver plus per_document (one PDF letter per notice, no two worded alike), NodriverDownloadDriver (those PDFs behind a Cloudflare challenge), multi_row_vision (a born-digital PDF table that text extraction mis-split), and ListingDetailDriver (a listing whose records are genuinely split across two pages).

4. "The signal must carry the value; the model must never infer it" -- where the rule actually lives

Checkpoint	Mechanism	Proof (faidh's WARN Act consumer)
Structural validation, all-or-nothing per record	validate_candidate / validate_row_candidate (+ regex variants), extraction.py -- no coercion, no defaults, a record missing even one requested field is dropped entirely	Oklahoma: 217 real records exist, only 128 survived -- the other 89 genuinely lack a workforce-region value on the page
Schema omits what a page can't provide	field_schema per target -- never forces a value that isn't there	New York schema doesn't request affected_count/effective_date because its page has neither
Judge does semantic verification, not just structural plausibility	_judge_candidates / _judge_row_candidates, eval_loop.py -- sees real page HTML + candidate values, told structural validity is already checked, job is semantic correctness	Georgia: an earlier schema mistakenly asked for county; judge rejected the resulting hallucinated-mapping candidate rather than accepting it
Confirmed vs. best-guess survives to the signal	ScrapeExtraction.validation_status ∈ validated / needs_review / failed / blocked	WarnActPlugin._produce() logs a distinct warning for needs_review yields rather than treating them as confirmed
"We never received the page" is not "the page was wrong"	challenge.classify_failed_page, challenge.py -- asked only once extraction has already failed, and only about a page that differs from the last one that worked; a blocked verdict persists validation_status="blocked" with no records and leaves the recipe untouched	A bot wall used to increment the same failure counter a redesign does, so three polls later the recipe was discarded and an LLM round spent learning to extract data from a challenge page

One value appears in ScrapeTool's JSON payload under the same key and is deliberately not a ValidationStatus: backoff, reported when the fetch circuit suppressed the poll. It is never stored, because a suppressed poll persists nothing -- the other four all describe a page we did or did not receive, where backoff describes a fetch we declined to attempt. A consumer that lumps it in with blocked will record a bot wall for a target that was merely being backed off, which is the exact conflation the circuit was built to stop.

Open tension, not fully resolved: a needs_review record still gets yielded as a real signal -- "surface for review, never silently drop" beats "silently withhold," by deliberate design. But nothing downstream currently filters on validation_status. The signal carries the distinction; no consumer currently reads it. A consumer treating every signal as equally trustworthy today would be wrong to. blocked widens that gap slightly: it is the one value that means no data was produced and none should have been, so a consumer that treats it as a failed extraction will over-count failures for a target that is merely walled.

5. Where new capabilities plug in

flowchart TD
    Core["This package<br/>driver.py / extraction.py / eval_loop.py / collections.py"]
    T02["find_target_page() -- shipped<br/>page_finder.py<br/>outputs a ScrapeTarget-shaped guess<br/>(URL + driver_backend, verified flag)<br/>docs/scrape-task-02-page-finder-agent.md"]
    T03["discover_candidates() / discover_row_candidates() -- shipped<br/>extraction.py<br/>schema OUT instead of schema IN<br/>docs/scrape-task-03-schema-discovery-mode.md"]
    T04["capture_request_shape() -- shipped<br/>request_shape_finder.py<br/>real captured request shapes<br/>for in-session XHR targets"]
    Core -.plain data in/out, no hidden state.-> T02
    Core -.plain data in/out, no hidden state.-> T03
    Core -.plain data in/out, no hidden state.-> T04
    T02 --> BT["bootstrap_targets()<br/>target_source.py"]
    T03 --> BT
    BT --> Core

All three front-end stages have shipped, and each landed without touching driver.py, eval_loop.py, or collections.py -- which is the point. They produce and consume the same plain data shapes (RenderedPage, FieldSchema, PageFinderResult, DiscoverySchemaResult) the core already uses, and none of them persists anything: a caller decides whether a discovered page or schema becomes a real ScrapeTarget.

Full designs live in docs/: scrape-lift-design.md (the overall shape and its decision log), scrape-task-01-lift-core-package.md (the lift itself and the AGPL isolation boundary), scrape-task-02-page-finder-agent.md, scrape-task-03-schema-discovery-mode.md, and scrape-task-04-multi-document-driver.md (the multi-document driver plus the sidecar's browser-forced-download mode).

6. Module map

packages/scrape/src/threetears/scrape/
├── driver.py                 ScrapeDriver ABC, RenderedPage, NavStep, NetworkCall
├── drivers/
│   ├── nodriver_sidecar.py   HTTP → headful Chromium sidecar on Xvfb (AGPL boundary)
│   ├── camoufox.py           in-process stealth Firefox (MPL-2.0-safe)
│   ├── document.py           parse_document() + document_text_to_html()
│   ├── api.py                stateless JSON GET, _resolve_path()
│   ├── network_capture.py    wraps another driver, _find_largest_record_list()
│   ├── multi_document.py     link list → N documents → one combined page, seen_urls dedup
│   ├── listing_detail.py     listing rows + per-row detail pages, merged into one table
│   └── nodriver_download.py  sidecar POST /v1/download → parse_document_bytes_to_html()
├── extraction.py             generate_candidates/generate_row_candidates (+ regex),
│                              validate_candidate/validate_row_candidate (+ regex),
│                              html_to_text(), strip_boilerplate()
├── eval_loop.py               run_eval_loop()/run_eval_loop_multi_row(),
│                              _judge_candidates()/_judge_row_candidates(),
│                              _persist_extraction(), _save_recipe()
├── challenge.py               PageVerdict, classify_failed_page() -- asks what a page that
│                              failed extraction actually is; no vendor markers
├── health.py                  ScrapeTargetHealth + collection, content_fingerprint(),
│                              record_validated_fetch(), record_classification(),
│                              record_circuit_state(), record_robots_block(),
│                              list_walled() -- the only non-primary-key query here:
│                              which targets need a human, walls and robots blocks both
├── circuit.py                 TargetCircuit, BackoffPolicy -- gates the FETCH of a target
│                              that keeps coming back walled, so its fetch rate and its
│                              classification rate both decay. Transitions come from
│                              models.circuit_breaker.CircuitBreaker via restore()
├── session_state.py           seal_session_state()/open_session_state()/
│                              usable_session_state() -- a human's cleared browser state,
│                              sealed under an operator key, handed back to later
│                              unattended renders so one solve is not paid for twice
├── robots.py                  RobotsGate, RobotsPolicy -- waits a site's Crawl-delay and
│                              escalates a Disallow to a human. Both on by default
├── reprobe.py                 ScheduledJobsReprobeScheduler -- books the next probe as a
│                              scheduled-jobs relative_delay job. Needs the [reprobe] extra
├── collections.py             ScrapeTarget/ScrapeRecipe/ScrapeExtraction (BaseEntity)
│                              + BaseCollection subclasses (L1/L2/L3 via
│                              threetears.core.backends.protocol.DurableStore)
├── migrations.py               v001-v012, PACKAGE_NAME="3tears_scrape"
├── target_source.py            TargetSource ABC, YamlTargetSource, CollectionTargetSource,
│                              StaticTargetSource, bootstrap_targets()
├── page_finder.py              find_target_page() -- bounded-turn search/fetch research agent
├── request_shape_finder.py     capture_request_shape() -- real captured XHR/fetch shapes
├── forms.py                    parse_form()/build_form_post() -- browser-free postback replay
├── tool.py                     ScrapeTool -- ad-hoc single-call MCP entry point
├── enrichment.py                secondary free-form LLM notes pass (separate from structured_fields)
└── llm_retry.py                 bounded_retry_structured_call() -- shared by extraction.py + eval_loop.py

Example consumer (faidh repo, not part of this package):
faidh/src/faidh/intake/plugins/warn_act.py               WarnActPlugin -- the ONLY place WARN-domain meaning exists
faidh/src/faidh/intake/plugins/seeds/warn_act_targets.yaml   the ScrapeTarget config rows
faidh/src/faidh/intake/runner.py                          publish_observations(), publish_arbitrary_signals(), poll_scrape_targets()
faidh/src/faidh/intake/plugins/__init__.py                PluginBase.collect() -- persist-then-yield template method
faidh/src/faidh/intake/signals/arbitrary.py               ArbitrarySignalEntity/Collection -- the actual Postgres sink

Call chain, one live poll cycle (faidh's WARN Act consumer): runner.publish_observations() → WarnActPlugin.collect() (inherited PluginBase) → WarnActPlugin._produce() → resolves self._drivers[target.driver_backend] → driver.render(...) → run_eval_loop_multi_row(...) (or run_eval_loop if multi_row=False) → _run_reuse_cycle (re-runs the stored strategy; on a miss, classifies the page and routes) or _regenerate → _persist_extraction() writes scrape_extractions → back in _produce(), each record becomes an ArbitrarySignalEntity → collect()'s save_entity() writes faidh_arbitrary_signals → publish_arbitrary_signals() re-drives collect() and publishes each yielded entity to arbitrary_signals() on NATS.

7. When a target needs a human: the whole loop

3tears returns outcomes and stores state. It does not own a queue, a conversation, or an operator -- your platform does. This is every seam it offers, in the order a platform uses them, because reconstructing it from four modules is how a step gets missed.

1. Scrape as normal. Most targets return a recipe and reuse it forever with no model call. A walled one returns validation_status="blocked" with the classifier's own words, and its recipe is left byte-identical -- a wall says nothing about whether your selectors still work.

2. Stop paying for it. Repeated blocks open the target's circuit, after which the fetch does not happen at all: subsequent calls return "backoff" with retry_after_seconds and never touch the browser. That is what stops a walled target costing a classifier call on every poll forever.

3. Ask what is stuck. ScrapeTargetHealthCollection.list_walled() returns the targets a human can actually help, with the evidence and the backoff. Targets whose host merely stopped answering are excluded -- the circuit opens on those too, and a person sent to one arrives with nothing to clear. This is the only non-primary-key query in the package, and it exists so you do not have to fetch fifty targets to find four.

It returns two kinds of queue item, and they need different decisions from the operator. Check which one you have before routing it:

	last_blocked_at set	robots_blocked_at set
What happened	a bot wall stood where the content should be	the site's robots.txt disallows this path
What the human does	clears the challenge in the browser; the solved session is reused unattended after that	decides whether we may fetch it AT ALL -- which is a policy call, not a puzzle
Reaching for the browser	is the answer	is not, on its own. A person working the page over VNC is exempt from Disallow because the exclusion protocol governs automated agents, but that exemption is for a session a human is genuinely in -- not for opening one and letting the robot drive through it
The other exit	RobotsPolicy.overrides records a per-origin agreement with a site, which is the honest way to say "we have permission for this one" without turning robots off everywhere

Both are cleared by record_human_cleared. Treating them as one queue sends an operator to solve a challenge that is not there, and -- worse -- quietly converts a policy decision into a puzzle someone clicks through.

4. Nothing is held while it waits. No browser, no session, no tab. The queue item carries url plus nav_steps, and the target is re-driven from those when an operator actually arrives, so waiting costs nothing.

5. A human arrives. Your platform opens a session on the sidecar (POST /v1/hitl/session), pulls the target in (POST .../tab -- its own isolated browser context, so one target never sees another's cookies), and points the operator at the router you mounted. Bounded slots, a hard TTL, and a token the sidecar minted. The sidecar authenticates nobody: who was entitled to that token is your platform's decision, evaluated where identity lives.

The operator's half is a router you mount, not a service we run. build_operator_router() returns a FastAPI APIRouter carrying the operator page, the vendored noVNC client and the WebSocket that relays RFB from the display. Mount it under any prefix you like, at any depth: every URL it emits is relative, so it never needs to know where it ended up.

Install the extra it needs -- pip install "3tears-scrape[hitl]", which adds FastAPI and 3tears-nats. It is an extra because a deployment that never needs a human never needs a web framework, and nothing else in this package imports one.

from threetears.scrape.operator import build_operator_router

app.include_router(
    build_operator_router(
        authorize=my_session_authorizer,   # is this token entitled to this session?
        display=my_display_endpoint,       # where is this session's RFB server?
        claim=my_claim_lookup,             # does THIS pod still hold it? (omit on a single pod)
    ),
    prefix="/scrape/hitl",                 # yours to choose; the router never learns it
)

Pass claim on any deployment running more than one pod. Without it a pod never learns it has lost a session, so two things fail to happen: the operator's stream does not end, and the pod does not release what it is holding -- an x11vnc on a live display and browser contexts carrying the cookies of whatever the human was part way through. Both survive until the sidecar's own session TTL reaps them, while the pod that actually owns the session serves somebody else against the same id. With claim supplied, the stream ends and the display is released at the moment ownership moves.

The token must use the URL-safe alphabet. It rides a WebSocket subprotocol name, and RFC 6455 subprotocol names are HTTP tokens, so /, =, a comma or whitespace breaks the upgrade in the browser before anything server-side runs -- and the operator sees only "Could not open the display". secrets.token_urlsafe, a padding-stripped base64URL value, or a hex digest are all fine; a standard-base64 value is not.

The deployment this is shaped for is a Kubernetes pod with two containers: yours, which holds identity, coordination and the operator's socket, and the AGPL nodriver sidecar beside it, which holds Xvfb, Chromium and x11vnc and nothing else. They share a network namespace, so your container reaches the display on 127.0.0.1 and nothing outside the pod can.

That last clause cuts both ways, so the relay takes a transport. A pod carrying no Service and no Ingress dials out and is never dialled into, which is what keeps a live display off the network -- and it also means a process terminating the operator's WebSocket anywhere else has no address to connect to and no way to be given one. So relay_stream, the function that WebSocket runs, does not open the connection itself. Its transport does: it is handed the endpoint your display callable resolved, decides what reaching that endpoint means, and yields the reader and writer the bytes move between. Called with none it opens a plain TCP connection to that endpoint, which is the co-located pod above and is exactly what the relay has always done.

build_operator_router takes one too and passes it straight through, so a deployment whose display sits in a pod with no inbound network path supplies a transport there and changes nothing else. Omit it and the router behaves exactly as it always has.

The pod's own end of that is serve_display(), in threetears.scrape.operator_pipe. A pod holding a session's claim serves its display onto a NATS byte pipe, bridging the x11vnc beside it to a stream a caller anywhere can attach to. Which of the two routes a deployment uses is not a choice made per request: with the socket and the display in one pod, relay_stream connects to loopback and serve_display() never runs; with the socket terminated where nothing can reach the pod, the pod runs serve_display() and the process holding the socket supplies the transport. The loopback hop happens either way, in one process or the other.

The remaining seams are optional and all there for the same reason -- one display lives in one pod, and a request can arrive at any of them. claim_session() takes a lease so two pods cannot both serve one session, and serve_session() answers that session's control messages (open a tab, complete a tab, close the session, read its state) on a subject keyed to the session id, so a caller addresses the session and never a pod. The same claim is what entitles serve_display(), so a handover ends the operator's stream instead of leaving two live views on one session. A completed tab's reply carries the human's solve sealed, because a raw cookie jar is a live credential and a bus is not the place for one.

6. They clear it, and you keep the work. POST .../tab/{id}/complete returns the context's cookies and storage. Seal it with seal_session_state and store it with record_session_state; the sidecar holds no key and never seals anything, which is the point -- the container driving a browser for arbitrary targets is the one you least want holding a decryption key.

7. Lift the suppression. TargetCircuit.record_human_cleared(target_id). This step is not optional and not implied by step 6. Storing the solve does not reopen the target: its circuit is still open, so the next poll is still suppressed and the operator's work sits unused until a timer they know nothing about elapses. record_reachable cannot do it either -- it reports a fetch that succeeded, and a success from a request the breaker never admitted deliberately leaves the circuit open. A human's word is different evidence.

8. It rejoins the healthy ones. The next scrape reads the stored solve, sends those cookies with the request that would otherwise have been challenged, and the target extracts normally. When the solve expires, it is ignored rather than sent -- the target simply needs a human again, which degrades to asking for help rather than to bad data.

8. Which exit a request leaves by

A fetch can leave by a chosen exit -- the default route, a TOR circuit, a Cloudflare WARP tunnel, any proxy URL -- on the backends that honour one. The seam is threetears.core.egress, in core rather than here because an exit is not a scraping concept: EgressDriver answers both "what transport should httpx use" and "what does a browser need on its command line", since a deployment that got only one of those right would send its API calls one way and its scrapes another while reporting a single configured exit.

SocksEgress and WarpEgress produce socks5:// proxy URLs, which httpx routes through socksio -- absent, it raises ImportError at the first request rather than at import, so it surfaces as a runtime fault in a scrape rather than as a missing dependency. Install the extra:

pip install '3tears[socks]'

from threetears.core.egress import EgressRegistry, SocksEgress, WarpEgress

registry = EgressRegistry({"tor": SocksEgress("tor"), "warp": WarpEgress()})
tool = ScrapeTool(..., egress="tor", egress_registry=registry)

Worth knowing before reaching for it.

Not every backend honours one. ApiDriver and NodriverSidecarDriver accept an EgressDriver; NetworkCaptureDriver and MultiDocumentDriver inherit whichever driver they wrap. CamoufoxDriver, DocumentDriver, ListingDetailDriver, and MultiDocumentDriver's own listing fetch reach the target on the container's default route regardless of what this tool is configured with. Configuring an exit and selecting one of those backends therefore proxies the robots.txt read and sends the page fetch direct. ScrapeTool names the offending drivers in a warning at construction rather than letting that happen quietly, so check your logs; there is no way for this package to route a backend that has no proxy support.

NodriverDownloadDriver is a third case. The warning names it -- it declares no exit, so it lands in the unproxied list like the others -- but the message understates it slightly. It posts to the sidecar's /v1/download, which accepts no per-request exit and reports none back, so its fetch leaves by whatever the CONTAINER's EGRESS_PROXY is set to rather than by the container's own address. That is still neither the default route nor this tool's configured exit: it is a per-container setting, so a deployment running one sidecar behind TOR gets TOR for downloads no matter what any driver was given.

It is wired in two places, and getting one right hides the other. Drivers take their own egress= because a driver may be shared between tools; ScrapeTool takes one for its own requests, including the robots.txt read that happens in FRONT of every fetch. Proxy the driver alone and the page leaves by TOR while the robots read discloses the container's real address to the same origin moments earlier. Proxy the tool alone and it is the page fetch that goes out direct, which is worse. ScrapeTool warns on both shapes.

An unknown name raises rather than falling back to the default route. A deployment that asked for tor and silently got direct would look correct in every log line while being wrong about the only property it configured this for.

TOR and WARP are for opposite problems. TOR is non-attribution, and its exits are public, enumerable and heavily challenged by exactly the bot walls this package works around -- expect a HIGHER challenge rate through it. WARP changes the address a site sees and is far less challenged, but Cloudflare can associate the traffic with the account, so it is not anonymity.

ScrapeTargetHealth.last_egress records which exit an observation came from, so "this target is walled" can be told apart from "this target is walled FROM THIS EXIT" -- without it, one blocked exit poisons a target permanently and the circuit backs it off having learned the wrong lesson. That value is what the render was CONFIGURED to use, not an observation of where the traffic went; confirming the latter needs an outside observer, which is why EgressDriver.health() exists and reports the address an exit actually presents.

Poll it. Nothing here does. EgressRegistry.health() probes every registered exit concurrently and never raises, but this library has no daemon to run it from, so wiring it is the consuming platform's job and there is no default that happens without you:

report = await registry.health()          # {name: EgressHealth}
down = [name for name, h in report.items() if not h.reachable]

The gap it closes is one you will otherwise diagnose the slow way. A dead TOR or WARP daemon fails every render transport-side, so every target's circuit opens -- and because an unreachable fetch deliberately never stamps last_blocked_at, list_walled() stays EMPTY while the whole fleet decays. The symptom is "everything stopped and nothing needs a human", which reads like a scraper problem and is a daemon problem. Render failures do log at WARNING with error_type, so it is not silent; it is just attributed to the wrong layer until somebody asks the exits directly.

License

MIT for everything this package's own authors wrote. See LICENSE.

The declared expression is MIT AND MPL-2.0 AND LicenseRef-noVNC-DES, because the wheel carries files that are not ours and saying "MIT" would be claiming otherwise:

• noVNC v1.7.0, vendored unmodified at src/threetears/scrape/operator_assets/novnc/, so a platform mounting the operator router gets a working display rather than instructions for installing one. MPL-2.0, which is file-level copyleft: it governs those files and does not reach the code around them. The upstream LICENSE.txt, AUTHORS and the licence texts they reference ship alongside the code and again in the wheel's dist-info/licenses/. Provenance -- version, source archive, digest, and the fact that nothing in the tree has been changed -- is recorded in novnc-provenance.json and held by tests/test_operator_assets.py. core/crypto/des.js carries its own permissive grants from AT&T Laboratories Cambridge and Widget Workshop, which match no listed SPDX licence; LicenseRef-noVNC-DES names them.
• pako, MIT, under operator_assets/novnc/vendor/pako/, which noVNC's compressed-encoding decoders import.

The bundled sidecar (sidecar/) wraps nodriver (AGPL-3.0) as a genuinely separate process, under its own sidecar/LICENSE. It is not part of the wheel.