Context — child of #14164
#14164 established (V-B-A) that the tool-tier projection mechanism is already built (ToolService x-neo-tool-tier + x-neo-harness-tool-projection → BaseServer --tool-projection-mode), and neural-link is already fully tagged. The non-NL servers are untagged — this slice tags knowledge-base, the smallest/cleanest, to prove the replication pattern.
Scope (this leaf — delivered by the PR)
- knowledge-base openapi: every operation tagged
x-neo-tool-tier + the root x-neo-harness-tool-projection policy (defaultVisibleTiers: [read], operatorOnlyTiers: [admin]), mirroring neural-link.
- Tier split (13 tools → 7 read / 4 extended / 2 admin):
- read (always-on core):
ask_knowledge_base, query_documents, get_document_by_id, list_documents, get_class_hierarchy, get_mcp_tool_handbook, healthcheck
- extended (on-demand via handbook):
get_deployment_state_snapshot, get_ingestion_progress, inspect_deployment, list_agent_faqs
- admin (operator-only, mutating):
manage_knowledge_base, ingest_source_files
- Compliance tests lock the tagging (policy + full classification + dangerous-not-read), mirroring the neural-link pattern in
OpenApiValidatorCompliance.spec.mjs.
→ A harness spawned with --tool-projection-mode harness-embedded loads 7 of 13 kb tools (a 46% cut); the 6 withheld stay reachable on demand via get_mcp_tool_handbook — no capability removed.
Out of scope (sibling #14164 work)
- gh (24) + mc (36) tier tagging (next slices, same scheme).
- The operator flipping
--tool-projection-mode on in the maintainer launch config (harness-local — the highest-leverage lever, NL already tagged).
Acceptance Criteria
Test Evidence
OpenApiValidatorCompliance.spec.mjs — 3 new kb tier tests + the existing kb per-server loop all green (34 passed). The "core line" (read vs extended) is review-adjustable — flagged for peer sanity-check on #14164.
Authored by Vega (Claude Opus 4.8, Claude Code). Origin session 1bb8a27b-ae0d-4668-a9a2-acbbe2387512.
Context — child of #14164
#14164 established (V-B-A) that the tool-tier projection mechanism is already built (
ToolServicex-neo-tool-tier+x-neo-harness-tool-projection→BaseServer --tool-projection-mode), and neural-link is already fully tagged. The non-NL servers are untagged — this slice tags knowledge-base, the smallest/cleanest, to prove the replication pattern.Scope (this leaf — delivered by the PR)
x-neo-tool-tier+ the rootx-neo-harness-tool-projectionpolicy (defaultVisibleTiers: [read],operatorOnlyTiers: [admin]), mirroring neural-link.ask_knowledge_base,query_documents,get_document_by_id,list_documents,get_class_hierarchy,get_mcp_tool_handbook,healthcheckget_deployment_state_snapshot,get_ingestion_progress,inspect_deployment,list_agent_faqsmanage_knowledge_base,ingest_source_filesOpenApiValidatorCompliance.spec.mjs.→ A harness spawned with
--tool-projection-mode harness-embeddedloads 7 of 13 kb tools (a 46% cut); the 6 withheld stay reachable on demand viaget_mcp_tool_handbook— no capability removed.Out of scope (sibling #14164 work)
--tool-projection-modeon in the maintainer launch config (harness-local — the highest-leverage lever, NL already tagged).Acceptance Criteria
x-neo-tool-tier; thex-neo-harness-tool-projectionpolicy declaresdefaultVisibleTiers: [read].manage_knowledge_base,ingest_source_files) are NOTread.Test Evidence
OpenApiValidatorCompliance.spec.mjs— 3 new kb tier tests + the existing kb per-server loop all green (34 passed). The "core line" (read vs extended) is review-adjustable — flagged for peer sanity-check on #14164.Authored by Vega (Claude Opus 4.8, Claude Code). Origin session 1bb8a27b-ae0d-4668-a9a2-acbbe2387512.