LearnNewsExamplesServices
Frontmatter
id14198
titleTag memory-core MCP tools with tier projection (read/write/extended/admin) — #14164 slice
stateClosed
labels
enhancementaiarchitecture
assigneesneo-opus-vega
createdAtJun 27, 2026, 5:38 AM
updatedAtJun 27, 2026, 11:07 AM
githubUrlhttps://github.com/neomjs/neo/issues/14198
authorneo-opus-vega
commentsCount0
parentIssue14164
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtJun 27, 2026, 11:07 AM

Tag memory-core MCP tools with tier projection (read/write/extended/admin) — #14164 slice

Closed v13.1.0/archive-v13-1-0-chunk-7 enhancementaiarchitecture
neo-opus-vega
neo-opus-vega commented on Jun 27, 2026, 5:38 AM

Context — child of #14164

The last repo-side slice (after kb #14188 + gh #14195). Tags memory-core — the largest + most-nuanced surface (36 tools). Stacked on the kb+gh branch (the compliance tests share OpenApiValidatorCompliance.spec.mjs); rebases onto dev once #14190 merges.

Scope (this leaf — delivered by the PR)

  • memory-core openapi: every operation tagged x-neo-tool-tier + the root x-neo-harness-tool-projection policy (defaultVisibleTiers: [read, write], operatorOnlyTiers: [admin]).
  • Tier split (36 tools → 15 read / 4 write / 13 extended / 4 admin):
    • read (always-on queries + mailbox reads): healthcheck, get_mcp_tool_handbook, who_is_online, query_raw_memories, query_recent_turns, query_summaries, query_hybrid_graph, get_session_memories, get_all_summaries, search_nodes, get_node, get_neighbors, get_context_frontier, list_messages, get_message
    • write (always-on core-maintainer writes — A2A + memory + presence): add_message, add_memory, mark_read, record_turn_presence
    • extended (on-demand diagnostics / session-lifecycle / frontier / wake / archive / perms-list): get_memory_core_tool_metrics, get_sqlite_holder_diagnostics, get_rem_pipeline_state, inspect_deployment, get_deployment_state_snapshot, mutate_frontier, pre_brief_session, resume_session, set_session_id, transition_task, manage_wake_subscription, archive_message, list_permissions
    • admin (operator-only, destructive / permission): purge_session, delete_message, grant_permission, revoke_permission
  • Compliance tests lock the tagging (policy + classification + dangerous-not-read; 14 mutating ops asserted not-read).

→ harness-embedded loads 19 of 36 (a 47% cut); the 17 withheld stay reachable via get_mcp_tool_handbook. No capability removed.

The core-line (review-adjustable)

mc has the most-nuanced line (most maintainer-core-writes + most destructive ops). Applied the Ada-validated scheme (#14190): mutations→admin/write, diagnostics→extended, safe-queries→read; A2A/memory writes are visible write (constant for a maintainer). The borderline calls (transition_task, resume_session/set_session_id → extended; get_context_frontier → read) are flagged for sanity-check.

Acceptance Criteria

  • Every mc op carries a valid x-neo-tool-tier; policy declares defaultVisibleTiers: [read, write].
  • All 14 mutating/destructive ops are NOT read; purge/delete/grant/revoke are admin.
  • Compliance tests assert policy + classification + dangerous-not-read.
  • No capability removed — extended/admin reachable via the handbook.

Test Evidence

OpenApiValidatorCompliance.spec.mjs — 3 new mc tier tests + the existing mc per-server loop green (40 passed total with kb+gh). YAML parse + tier distribution verified (15/4/13/4, no untagged op).

Authored by Vega (Claude Opus 4.8, Claude Code). Origin session 1bb8a27b-ae0d-4668-a9a2-acbbe2387512.

tobiu referenced in commit 97a4f59 - "feat(ai): tag memory-core MCP tools with tier projection — 15 read / 4 write / 13 extended / 4 admin (#14198) (#14199) on Jun 27, 2026, 11:07 AM
tobiu closed this issue on Jun 27, 2026, 11:07 AM