AL knowledge consultation
This is the public skill entry point: the file the host agent reads to start the skill. It tells the agent how to prepare the question and follow BCQuality's routing instructions. The same agent can execute all steps inline; no separate program or agent is required.
This file is distinct from skills/entry.md, which owns routing. Article
selection, precedence, and output policy belong to Entry, READ, DO, and the
dispatched internal action skill.
Execute
- Resolve
PLUGIN_ROOTfrom the installed plugin'splugin.json; this file isPLUGIN_ROOT/skills/al-knowledge/SKILL.md. Never assume the user's app root is the plugin root. - Build Entry's
task-context. Copy the caller's question verbatim intogoal. Setinputs-available: [knowledge-query]and bindknowledge-queryto that exact question. Passtechnologies,bc-version,countries, andapplication-areaonly when provided or reliably determined. Pass the optionalBCQUALITY_ENABLED_LAYERSandBCQUALITY_DISABLED_SKILLSsettings using the same parsing rules asal-code-review. - Read and execute
PLUGIN_ROOT/skills/entry.md, including Preparation, resolving all paths againstPLUGIN_ROOT. Ifpwshor index generation is unavailable, use the documented path-based discovery fallback. - Execute only the dispatched action skills with their exact input subsets.
Read
PLUGIN_ROOT/skills/read.mdandPLUGIN_ROOT/skills/do.mdon demand. Do not callal-code-reviewfor this knowledge question, synthesize a dispatch, or turn an explanation into a finding. - Validate the action skill's
knowledge-responseunder DO before accepting it: check the schema, exact bound question, and current-run full reads of every cited path. The consumer can usePLUGIN_ROOT/tools/validate_knowledge_response.pywith its own question and read evidence. Preserve the raw result when validation fails; record that failure separately instead of repairing or accepting the answer. Inline checks by the answering agent are self-attestation, not independent verification. Independent checks require a trusted consumer read collector. Return a valid response unchanged, or Entry'sno-match/faileddispatch record unchanged.
Where the host has no tool to invoke a skill and skills are instruction files the
agent reads (for example VS Code Copilot), invoking this skill means executing
steps 1–5 inline in the caller's own context. The knowledge-response the
dispatched action skill produces there is the response to return: producing it
that way is the protocol, not a fabrication. Lacking an invocation tool is never
a reason to skip this skill or to answer without its response.