Save ADE-bench results.

Save a run to the experiments_saved directory. If no run_id is provided, saves the most recent run with results.

Timeout for setup operations (task setup scripts and agent installation) in seconds.

Path to the shared config directory.

Path to the shared databases directory.

Path to the shared databases directory in the main repository root. This always points to the main repo's databases, even when running

Path to the shared DuckDB databases directory in the main repository.

Path to the shared migrations directory.

Path to the shared projects directory.

Path to the shared Snowflake databases directory in the main repository.

Check if file should be excluded from conversion.

Check if file should be included in conversion.

Get list of skill locations as strings (for result tracking).

Stop the live display and re-enable console handlers.

Keep containers alive for debugging without stopping services.

Get the list of task IDs in the dataset.

Path to the task's setup directory.

Path to the task's setup.sh script.

Get the list of tasks in the dataset.

Return the tasks directory path.

Create temporary output directory.

When every row differs in one column, pre-scan detects it.

Test basic snapshot capture without exclusions.

Test that hidden files (.*) are excluded by find command.

Test snapshot capture with path exclusions.

Rows with non-systematic diffs show in a compact table.

Create a test container with sample files.

Test diff creation between snapshots.

dbt MCP server detected by 'dbt-mcp' in args.

dbt MCP server detected by server name 'dbt'.

Non-zero exit on env file write logs warning but continues.

When set, the value is passed via -c model_reasoning_effort=<value>.

Value goes through shlex.quote so injection attempts are inert.

When unmatched rows > fuzzy_row_limit, skip fuzzy matching.

Test finding repo root from main directory. This test verifies that _get_repo_root finds the ade-bench repository root, not the task's tempor

Test finding repo root from worktree. This verifies that when code runs from a worktree, _get_repo_root correctly finds the main repository r

Tests with status 'error' should also be detected.

When no --plugin-set, use defaults.

Error when explicitly requested plugin set is incompatible.

Error when no plugin sets are compatible with agent.

Explicit --plugin-set names are resolved.

Plugin sets incompatible with agent are filtered out.

Non-zero exit on mcp add logs warning but doesn't raise.

Pre-scan catches the column-wide diff; post-scan catches sporadic diffs.

When OPENAI_CODEX_REASONING_EFFORT is unset, no -c flag is emitted.

No calls when mcp_servers is empty.

Columns differing in <90% of rows are not systematic.

When agents is None, compatible with all agents.

When agents is set, only compatible with listed agents.