Problem	Fix
The Teamwork Graph CLI will not install or connect.	Run `twg doctor` and fix the first failed check. To reinstall, run the installer for your device in Install and set up.
The Benchmark CLI will not install, or `benchmark update` fails, including on Windows.	Run the installer for your device in Install and set up, then confirm with `benchmark version --verbose`. Reinstalling over an existing install is safe.
`benchmark run browser` does not open a browser.	Run `benchmark run browser --no-open-browser` and open the URL it prints. This is also the right option over SSH or in a remote shell.
A prerequisite row will not turn green.	Select Try again on that row, then Run checks again. If the row offers no button, the fix has to happen outside the benchmark, and the row's own text says what is missing.
Continue to prompts stays disabled.	Each of Work, Knowledge, and Code needs at least one connected source. Newly connected sources can take a few minutes to sync, so select Check connection again before assuming a failure.
A prompt's checkbox is disabled.	Its sources are not connected. The row names the source it needs, and you connect it back in Connect data sources. An empty detail never disables a prompt.
Continue to review stays disabled.	A prompt you selected is missing a detail, such as a topic, project, or person. The footer names the detail, and the row's warning opens the prompt so you can fill it in.
Pairing status reports that setup is incomplete.	Finish setup in the benchmark app, or run `benchmark pair setup --agent codex`.
The run stopped partway.	Start a new run rather than resuming. Both arms must answer the same prompts from the same fresh state for the comparison to hold.
The benchmark app closed while a run was going.	The run belongs to the terminal, not the tab. If the terminal is still running, reopen the URL it printed. If the terminal is gone, start a new run.
The report will not open, or looks empty.	Run `benchmark doctor`, then record the run ID from the report header before contacting support.
The Teamwork Graph arm answers, but the Siloed context arm stops and reports that an MCP server is already defined in a narrower scope.	From the directory you started the run in, remove the narrower Claude Code entry with `claude mcp remove atlassian -s local`, then rerun. Use `-s project` when `claude mcp get atlassian` reports a project scope.
`claude mcp list` shows the expected Atlassian server, but the Siloed context arm still cannot reach it.	Scope depends on the working directory. From the directory you start the run in, run `claude mcp get atlassian` and confirm the scope is user config and the URL matches the endpoint the benchmark configures. A stale entry often points at an older `/sse` endpoint.
On Windows, Claude Code denies every `twg` command, often with "Permission to use PowerShell has been denied because Claude Code is running in don't ask mode."	In the same shell you start the run from, set `$env:BENCHMARK_CLAUDE_PERMISSION_MODE = "auto"` (PowerShell) or `export BENCHMARK_CLAUDE_PERMISSION_MODE=auto` (bash or zsh). Then start a new run. Results using `auto` are less comparable than the default mode.
You need to inspect a failed command.	Run `benchmark <command> --help`, then share the exact command, exit code, stdout, and stderr.

Command reference

Run benchmark <command> --help for the complete flag list on any command.

Command	Purpose
`benchmark run browser`	Open the benchmark app and do the whole benchmark in it: setup, prompts, progress, blind review, and the report. This is the recommended path.
`benchmark version --verbose`	Confirm the installed version and setup state.
`benchmark update`	Update an existing installation. If it fails, run the installer for your device in Install and set up.
`benchmark doctor`	Run diagnostics on the local installation.
`benchmark skills install`	Install or repair the Teamwork Graph AI Context Benchmark skill for your agents.
`benchmark connectors status`	Report connector authentication state.
`benchmark connectors login`	Sign in to a connector the benchmark uses.
`benchmark pair status`	Read-only setup readiness for a manual run.
`benchmark pair setup`	Prepare and validate access for a manual run.
`benchmark live`	Start the live view for a manual three-terminal run.
`benchmark run control`	Run the Siloed context arm on its own.
`benchmark run test`	Run the Teamwork Graph context arm on its own.
`benchmark report`	Rebuild a local report from a completed run.
`benchmark report zip`	Package a run's redacted evidence for sharing.
`benchmark share browser`	Review and package a redacted report from a recent `benchmark run browser` session.
`benchmark report unzip`	Open a packaged bundle again.

Options for `benchmark run browser`

The defaults suit a first run. Change one only when your environment needs it.

Option	When to use it
`--agent codex` or `--agent claude`	Choose which agent runs both arms. By default the benchmark app infers it from the terminal you started in.
`--no-open-browser`	Print the URL instead of opening a browser, for example over SSH.
`--port 8080`	Serve on a fixed port. The default is an unused port chosen at start.
`--inactivity-timeout 1800`	Seconds without browser activity before the server shuts down. The default is 1800, or 30 minutes.

Manual three-terminal run

Use the three-terminal commands only when the benchmark app is unavailable, or when support asks you to reproduce a problem. Complete setup first with benchmark pair setup. Keep all three terminals open until the run finishes.

1
2
3
4
benchmark live
benchmark run control --agent codex --suite cc --parallelism 1
benchmark run test --agent codex --suite cc --binary --parallelism 1

Both arm commands must use the same agent and exactly one selector: --suite, --scenario, or --prompt. --binary belongs to the test arm only.

Option or argument	Rule
`--reconfigure-connectors`	Force a fresh sign-in and revisit optional connector choices.
`--reconfigure-inputs`	Update the saved example values used in prompts. No agent or connector check is needed.
`--details`	Print full readiness, connector setup, and validation output.
`--orchestrator-host <host>`	Name the tool coordinating the run, such as `cursor`. Omit it for Codex and Claude Code.
`--no-lookback`	Ignore the default scenario time windows. Pass it to both arms or neither.