Documentation review: 10 October 2026. Source: BC-Bench 0.15.0, commit
c84b793a8a9b7422af9512dfe67baeb2bc666d95. This update did not execute installation, agents, containers or benchmark runs.
Use src/bcbench/agent/shared/config.yaml.
| Control | Scope |
|---|---|
instructions.enabled |
Copies the entire profile, including skills and agents. |
skills.enabled |
Copies skills only for an isolated comparison. |
agents.enabled, agents.name |
Copies and selects an agent. |
plugins |
Per-session plugins, each with its own enabled. |
mcp.servers |
Declares servers; AL MCP has its own flag. |
Upstream profiles are placeholders. The old guide’s al-developer-bench and al-conductor-bench examples are not presented as current upstream agents. Microsoft Learn MCP is commented out and disabled by default; enabling it is an experimental change. Record AL MCP, LSP and BC MCP separately.
Upstream has a disabled, revision-pinned BCQuality plugin entry. Copy it from the selected revision and preserve its identity and pin. Inspect grant_dir_access when the skill needs corpus access. A loaded plugin does not prove skill invocation: retain discovery, loading and actual output evidence. Any instruction that encourages use is part of the experimental arm.
Select a compatible ALDC distribution and a published commit. Claude Code uses claude-plugin; Copilot CLI uses copilot-cli-plugin. Verify the selected manifests. This is a proposed experiment setup, not a runtime-validated BC-Bench integration.
Treat the BCQuality upstream and Javier’s fork as separate arms. The fork adds al-knowledge for knowledge questions; the reviewed upstream documents al-code-review. Do not attribute the fork extension to upstream.
Sources: experiments, ALDC, BCQuality, fork. Next: comparisons.