Skip to content

High-performance browser automation bridge and multi-instance orchestrator with advanced stealth injection and real-time dashboard.

3.7 KB
Raw Blame

Stealth Score Subagent Context

You are driving PinchTab in a Docker container against public bot-detection sites. Your job: visit each site in tests/stealth-score/sites/index.md, follow its playbook, and write a structured JSON report.

Do not touch Docker, do not switch providers, do not read other runs' artifacts. The orchestrator already brought up the container for you.

Do not read anything under tests/stealth-score/results/.

Inputs You Receive

  • PROVIDER: chrome or cloak — already configured in the container.
  • REPORT_FILE: absolute path of the pre-seeded JSON file to append to.
  • PROJECT_ROOT: absolute path of the pinchtab repo.

Working Directory and Wrapper

Always run from tests/tools so the ./scripts/pt wrapper resolves:

cd $PROJECT_ROOT/tests/tools
export PINCHTAB_CONTAINER=stealth-score-pinchtab
export PINCHTAB_TOKEN=stealth-score-token

./scripts/pt forwards every argument to the pinchtab CLI inside the container. Use it exactly like the local CLI:

./scripts/pt health
TAB=$(./scripts/pt nav https://example.com --new-tab --print-tab-id)
./scripts/pt --tab "$TAB" snap

Read $PROJECT_ROOT/skills/pinchtab/SKILL.md for the full command surface.

Session

Create one session up-front and reuse it across all sites:

export PINCHTAB_SESSION=$(./scripts/pt session create --agent-id stealth-${PROVIDER})

Per-Site Loop

For every site listed in $PROJECT_ROOT/tests/stealth-score/sites/index.md:

  1. Read its playbook (tests/stealth-score/sites/<id>.md). It tells you the URL, the readiness signal, any click/scroll/wait the page needs, and the metrics to extract.
  2. Drive PinchTab to follow the playbook. Use refs from snap when targeting elements; use wait --text / wait --not-text to settle SPAs. Don't hard-code sleeps unless the playbook explicitly says so.
  3. Extract the listed metrics from text / snap output. Record what you actually saw. If a metric is genuinely unreadable, record "unavailable" and explain in notes. Do not invent values. Do not carry values between sites.
  4. Append a record to REPORT_FILE:
{
  "site": "<id>",
  "url": "<url from playbook>",
  "tab_id": "<from --print-tab-id>",
  "metrics": {
    "<key>": "<value>",
    ...
  },
  "notes": "<optional one-line caveats>"
}

Safe-append helper:

python3 - <<'PY'
import json
path = "<REPORT_FILE>"
entry = { "site": "<id>", "url": "<url>", "tab_id": "<id>", "metrics": {...}, "notes": "..." }
data = json.load(open(path))
data["sites"].append(entry)
json.dump(data, open(path, "w"), indent=2)
PY

Quality Bar

Good Bad
trust_score: "62.0" trust_score: "see page"
webdriver: "passed" webdriver: "ok"
robot_verdict: "You're not a Robot" robot_verdict: "no"
bot_score: "0.12 (Likely human)" bot_score: "low"

If a site shows a Cloudflare challenge ("Just a moment...", "attention required"), record metrics: {} and notes: "cloudflare challenge" and move on — do not call /solve without explicit user approval.

Completion

After every site has a record:

python3 - <<'PY'
import json, time
path = "<REPORT_FILE>"
d = json.load(open(path))
d["completed_at"] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
d["sites_processed"] = len(d["sites"])
json.dump(d, open(path, "w"), indent=2)
PY

Print exactly this as your final stdout line:

STEALTH_SCORE_RUN_COMPLETE provider=<provider> sites_processed=<n>