Summary

KR-3 ST2 ships the iso_link_* typed-edge tool family — 3 model-facing MCP tools over the relationlink substrate (ADR-0033/ADR-0034 typed edges).

Combined typed-graph surface is now 7 tools (4 iso_node_* + 3 iso_link_*) returned by get_tool_schemas.

Rule-3 STOP-gate findings during recon

Verified against packages/db/migrations/0058_relationlink.sql on substrate main 41ddc208:

• Table is relationlink (bare, no schema prefix) — lives in public. Same workspace_id TEXT pattern as ST2/ST3 reads (not tenant_id).
• No RLS on the table — workspace_id filter in WHERE is the application-layer isolation. Direct asyncpg reads work without GUC.
• link_type column is un-CHECK'd TEXT — V1 vocabulary (21 entries: 11 sea/idea + 10 platform-wide) is informational and lives in sb1-substrate-shapes/src/relationlink.ts. The JSON Schema enum on iso_link_create enforces it client-side; the application-layer per-pair gate config is the normative source of truth.
• Three concurrent blockers gate the write path (new BUILD_DEVIATIONS entry D-kr3-st2-no-relationlink-write-mcp-tool):
  1. created_by_actor_kind CHECK lacks 'kora'. The CHECK covers 8 entries (operator, oracle, critic, claude_pm, hermes, platform_seal, platform_rollback, platform_session_expiry). Migration 0058 predates 'kora''s introduction (added to actor_registry in 0076). A Kora-side INSERT would fail.
  2. No Sea MCP write tool exposes the path. Same probe as ST3 scratchpad-write recon — only kora__propose_convention, kora__read_escalation_queue, kora__propose_policy_change registered. No kora__create_relationlink equivalent.
  3. chain_event_id UUID NOT NULL requires substrate-side SECDEF that emits + binds the chain event in one transaction (same pattern as kronicle.compact_scratchpad from Plan 02). Direct INSERT either fails (no chain_event_id) or, if filled in client-side, breaks the chain witness invariant.

Per ST2 spec § "Read vs write paths": "Writes → through Sea MCP (authorization + audit)". Deferred per the spec's BUILD_DEVIATIONS-or-defer pattern. PM coordinates the substrate-side bucket that addresses all three blockers in one shot.

What landed

Traverse SQL (recursive CTE per direction)

For outgoing:

WITH RECURSIVE walk AS (
  SELECT to_entity_id::text AS entity_id, to_entity_kind AS entity_kind,
         link_type AS via_link_type, 1 AS depth
  FROM relationlink
  WHERE workspace_id = $1 AND validity_state = 'active'
    AND from_entity_id = $2 AND link_type = ANY($3::text[])
  UNION ALL
  SELECT rl.to_entity_id::text, rl.to_entity_kind, rl.link_type, walk.depth + 1
  FROM walk
  JOIN relationlink rl ON rl.workspace_id = $1 AND rl.validity_state = 'active'
    AND rl.link_type = ANY($3::text[])
    AND rl.from_entity_id::text = walk.entity_id
  WHERE walk.depth < $4
)
SELECT DISTINCT entity_id, entity_kind, via_link_type, depth FROM walk

Mirror for incoming walks to_entity_id → from_entity_id. direction='both' issues both queries and the Python side dedupes by (entity_id, entity_kind, via_link_type) keeping MIN depth.

Capability checks

Reuses the iso_node assert_kora_can_perform stub (D-kr3-st1-capability-check-deferred). Per-tool capability mapping:

• iso_link_create → cap_sea_link_authoring
• iso_link_traverse → cap_read_unfiltered_relationlink
• iso_link_list_for_node → cap_read_unfiltered_relationlink

Test plan

22 new tests, all passing:

Schemas (6):

• All 3 schemas well-formed (OpenAI function-call shape)
• V1_LINK_TYPES = 11 sea/idea + 10 platform-wide = 21, no overlap
• iso_link_create schema's link_type enum matches V1_LINK_TYPES tuple
• iso_link_traverse schema caps max_depth at 3
• iso_link_list_for_node schema caps limit at 100
• Combined typed-graph surface is 7 tools

Read SQL (2):

• read_relationlink_for_node binds (workspace_id, entity_id, limit); OR on from/to; filters validity_state='active'
• List-for-node caps limit at MAX_LIST_LIMIT=100

Traverse (5):

• Outgoing runs one query (anchors at from_entity_id)
• direction='both' runs two queries (outgoing + incoming variants)
• max_depth capped at 3
• Rejects empty link_types (would sprawl)
• Dedupes by MIN depth when same node reached via multiple paths
• Rejects unknown direction values

Deferred write (1):

• create_relationlink raises RelationLinkWriteNotAvailableError; message names all 3 blockers verbatim (operators can grep the message + see exactly what substrate needs to ship)

Tool handlers (8):

• iso_link_create returns deferred envelope with the right deviation_id
• iso_link_create rejects invalid link_type
• iso_link_create rejects invalid node_kind
• iso_link_traverse returns results from mocked CTE
• iso_link_traverse validates link_types against V1 vocabulary
• iso_link_list_for_node returns rows
• iso_link_list_for_node enforces server-side limit cap
• provider.handle_tool_call routes iso_link_* by prefix

Capability check (1):

• Shared assert_kora_can_perform stub fires + logs D-kr3-st1-capability-check-deferred + the cap name on the iso_link traverse path (proves both families share the gate)

Gates

• ty check — 7,337 diagnostics, zero-delta vs KR-3 ST1 baseline.
• pytest tests/plugins/memory/ — 294/294 passing (22 new ST2 + 22 ST1 + 250 pre-ST2).
• Full suite via xdist (-n auto): 24,617 passed / 142 failed / 129 skipped vs KR-2 ST4 merge baseline (24,570/143/129): +47 passed, −1 fail. Same tests/tools/* + tests/skills/* + tests/tui_gateway/* xdist isolation noise; none touch plugins/memory/isokron/.

Rule-6 / BUILD_DEVIATIONS / Open asks

New deviation D-kr3-st2-no-relationlink-write-mcp-tool — 3 substrate blockers spelled out in BUILD_DEVIATIONS.md. PM needs to dispatch the substrate-side bucket that handles all three in one shot (CHECK extension + MCP tool + chain-event SECDEF).

Existing deviations carry forward (5 total open now):

Per PM's pipeline-order note: if no substrate buckets are ready to swap to after ST3, KR-6 is next on this lane — it's the cheapest unlock (~50 LOC; closes D-kr3-st1 without needing CC#1 substrate work).

Standing by for KR-3 ST3 (tool registration polish + Hermes memory deprecation + system prompt updates).

🤖 Generated with Claude Code