High-performance browser automation bridge and multi-instance orchestrator with advanced stealth injection and real-time dashboard.
Subagent Context
You are running PinchTab optimization tasks against a specific browser provider (chrome or cloak).
You will be given:
- The active PROVIDER
- Your report file path
- The range of groups you are responsible for
You will be given a set of group files, each containing numbered steps to execute using the PinchTab browser automation tool.
What to read
- PinchTab skill:
skills/pinchtab/SKILL.md— full command reference and patterns. - Group files:
tests/optimization/group-XX.md— task descriptions and verification markers.
What NOT to read
tests/tools/scripts/baseline.sh— deterministic baseline; reading it defeats the purpose.tests/tools/scripts/run-optimization.sh— infrastructure script, not relevant.tests/benchmark/— separate benchmark lane, not your concern.
Environment
- PinchTab server:
http://localhost:9867, token:benchmark-token - Fixtures:
http://fixtures/(Docker hostname) - Available fixture pages:
/,/wiki.html,/wiki-go.html,/wiki-python.html,/wiki-rust.html,/articles.html,/search.html,/form.html,/dashboard.html,/ecommerce.html,/spa.html,/login.html— plus additional fixture pages referenced in specific group steps.
Wrapper
Always use ./scripts/pt ... — never call pinchtab directly.
The wrapper executes pinchtab inside the Docker service and forwards PINCHTAB_TOKEN and PINCHTAB_SERVER automatically.
Multiple subagents run in parallel against the same browser instance. Make sure your commands don't interfere with other agents' work.
Recording
Each agent writes to its own report file to avoid concurrent-write corruption. Your report file path will be provided as REPORT_FILE when you are launched. Pass it on every step-end call with --report-file.
Record every step result immediately after completion:
./scripts/runner step-end --report-file "$REPORT_FILE" <group> <step> answer "<what you observed>" <pass|fail|skip> "verification notes"
For failures:
./scripts/runner step-end --report-file "$REPORT_FILE" <group> <step> fail "<what went wrong>" skip ""
<group>is the group number (e.g.,0,15,38)<step>is the step number within the group (e.g.,1,2,3)- Keep answers factual — do not self-grade in the answer payload.
- Quote actual output, don't paraphrase. The answer field must include the literal text or marker from the tool output. For example, if the server returns
status: ok, writestatus: okin the answer — not "Server responded with ok". Verification patterns match against exact substrings, so paraphrasing causes false failures. - Always include
UPPER_CASE_MARKERstrings verbatim. Fixture pages embed verification markers likeSUGGESTIONS_VISIBLE_COUNT_2,VERIFY_HOME_LOADED_12345,SCROLL_MIDDLE_MARKER, etc. When you see these in the page content or tool output, copy them exactly into your answer — they are the primary tokens that automated verification matches against.
Provider & Environment
The benchmark can run against different browser providers:
chrome(default)cloak(CloakBrowser for stealth / anti-detection testing)
You will be told the active PROVIDER when you are launched.
Critical: Before running any ./scripts/pt command, export the correct container and token:
export PINCHTAB_CONTAINER=optimization-pinchtab # for cloak runs
# or
export PINCHTAB_CONTAINER=tools-pinchtab-1 # for normal compose chrome runs
export PINCHTAB_TOKEN=benchmark-token
# Then all commands go through the wrapper:
./scripts/pt nav https://...
./scripts/pt --tab "$TAB" click e5
...
The wrapper (./scripts/pt) respects these two environment variables and forwards them into the container. Always set them in your shell at the start of the session and after any subshell.
Execution approach
- Read the PinchTab skill to learn available commands.
- Read the assigned group files.
- For each step: navigate to the fixture, interact using PinchTab commands, verify the expected markers appear, and record the result.
- Use your judgment to figure out the right commands — the group files describe WHAT to achieve, not HOW.