Summary

Closes KR-2. Ships the final ST4 pieces:

• Recent chain events read against hivex_foundation.event_log (the substrate's one genuine tenant_id UUID-keyed table). Resolves Kora's workspace_id TEXT to tenant_id via JOIN hivex_foundation.tenant ON t.clerk_org_id = $1. Filters event_type LIKE 'kora.%'. Truncates jsonb_pretty(payload) to 300 chars + ellipsis to match TS-side KoraRecentChainEvent. Default LIMIT 50.
• Active Constitution revision read against kronicle.workspace_constitution_revisions. Schema gotcha verified: no superseded_at column; "active" = ORDER BY revision_number DESC LIMIT 1 (riding the idx_constitution_revisions_workspace_current index). rules_hash BYTEA hex-encoded to match the TS-side constitutionVersionHashFromRulesHash helper so K-6 (Constitution pre-screen middleware) sees the same canonical hex form. Returns None for fresh-workspace bootstrap state.
• KoraSessionContext shape + assembler mirroring TS-side packages/sea-mcp-server/src/kora/context-assembler/types.ts:130. Six load-bearing reads + two identity fields (workspace_id, assembled_at). assemble_session_context fans out the six reads via asyncio.gather.
• System prompt block extended with §6 "Recent kora. activity"* surfacing the last 5 events as readable bullets (size-bounded; full history queryable via read_recent_kora_events directly).
• All remaining ABC stubs replaced with real implementations: on_session_end, on_session_switch (cache flush on reset=True), on_pre_compress (returns ""), on_delegation (scratchpad mirror + emit), save_config (no-op per env-var config), handle_tool_call (inherits ABC default for the no-tools case). No NotImplementedError stubs remain in the provider.

Rule-3 STOP-gate findings during recon

• Spec SQL § ST4 § 2 used WHERE workspace_id = $1 against hivex_foundation.event_log. The actual schema (foundation/0003:73) keys off tenant_id UUID NOT NULL REFERENCES tenant(tenant_id) — no workspace_id column. PM's ST4 dispatch caught this correctly: this IS the substrate's one tenant_id exception. Implementation uses the TS reader's JOIN pattern verbatim (verified against packages/sea-mcp-server/src/kora/context-assembler/index.ts:287).
• No kora__append_event MCP tool exists. Same probe as the ST3 scratchpad-write recon: only kora__propose_convention, kora__read_escalation_queue, kora__propose_policy_change registered in Sea MCP src/tools/. Per spec § ST4 § 1: "If kora__append_event is missing: defer behind ChainEventEmitNotAvailableError, file BUILD_DEVIATIONS D-kr2-st4-no-chain-emit-mcp-tool". Done — third deferred-write surface, same pattern as ST2 capability-row mirror + ST3 scratchpad write.
• C2 capability mirror was 1 cap stale. When the ST2 parity test ran against current substrate main, it caught cap_unbless_convention (operator-only, added 2026-05-20) missing from the Python mirror. Updated the mirror + counts (24+25=49; Kora granted=22, denied=27). This is exactly the drift the C2 parity guard was built to catch — closing the C2 deviation still gated on K-7 substrate dispatch.

What landed

Key SQL (verbatim, mirrors TS readers)

-- read_recent_kora_events: event_log is tenant_id-UUID-keyed
SELECT el.event_id::text, el.event_type, el.occurred_at,
       jsonb_pretty(el.payload)::text AS payload_text
FROM hivex_foundation.event_log el
JOIN hivex_foundation.tenant t ON t.tenant_id = el.tenant_id
WHERE t.clerk_org_id = $1
  AND el.event_type LIKE 'kora.%'
ORDER BY el.occurred_at DESC
LIMIT $2

-- read_active_constitution_revision: no superseded_at column
SELECT revision_id::text, rules_hash
FROM kronicle.workspace_constitution_revisions
WHERE workspace_id = $1
ORDER BY revision_number DESC
LIMIT 1

_prefetch_all now gathers 7 reads in one asyncio.gather

Single round-trip latency for the full session-warm. Caches populated:

• _charter_cache (RoleCharter)
• _policy_cache (List[PolicyRegistryEntry])
• _capability_cache (KoraCapabilityRow)
• _own_scratchpad_cache (List[ScratchpadEntry])
• _cross_agent_scratchpad_cache (List[ScratchpadEntry])
• _events_cache (List[RecentChainEvent])
• _constitution_cache (Optional[(revision_id, rules_hash_hex)])

Test plan

23 new tests, all passing:

test_events.py (9):

• Typed shape from happy path
• Default LIMIT 50 + custom limit binding
• SQL JOINs tenant on clerk_org_id + filters LIKE 'kora.%'
• Payload truncation at exactly 300 chars + ellipsis
• Short payload pass-through unchanged
• Empty payload handled
• Datetime occurred_at ISO-encoded
• Deferred emit raises ChainEventEmitNotAvailableError with tag + deviation ID

test_constitution.py (8):

• Typed shape with hex-encoded rules_hash
• Returns None for fresh-workspace bootstrap
• RLS-GUC call ordering (txn.enter → execute(set_config) → fetchrow → txn.exit)
• SQL uses revision_number DESC (NOT superseded_at)
• Hex encoder accepts bytes, bytearray, memoryview, hex str pass-through
• Hex encoder rejects unknown types with clear error

test_session_context.py (3):

• All 6 reads fire; typed shape with all fields populated
• None Constitution revision pair on fresh workspace
• Default limits (DEFAULT_SCRATCHPAD_LIMIT=100, DEFAULT_RECENT_EVENT_LIMIT=50)

test_provider_end_to_end.py (2):

• Full lifecycle walk — initialize → on_turn_start (warms 7 caches) → system_prompt_block (uses warm cache; all 6 sections present + Rule-6 verbatim + recent event_types in §6) → session_context() returns typed shape with constitution hex → sync_turn (cap_* token → attempt write → catch defer → log) → on_memory_write (mirrors to scratchpad) → on_delegation (scratchpad + emit, both deferred + logged) → on_pre_compress (returns "") → on_session_switch (no-reset keeps caches) → on_session_end (emit kora.session.ended deferred) → on_session_switch(reset=True) (flushes 7 caches) → shutdown. Zero NotImplementedErrors surface. 5+ deferred-write/emit WARNINGs logged.
• session_context() returns None when caches are cold (caller must call on_turn_start first)

Plus existing 64 isokron tests + parametrize-cross-product still passing.

Gates

• ty check — 7,337 diagnostics, same as ST3 baseline (Δ 0). Zero new diagnostics introduced.
• pytest tests/plugins/memory/ — 248/248 passing (23 new ST4 + 64 pre-ST4 isokron + 161 other memory providers).
• Full suite via xdist (-n auto): 24,570 passed / 143 failed / 129 skipped. Δ vs ST3 merge baseline (24,551/143/129): +19 passed, ±0 failed. Same tests/tools/* + tests/tui_gateway/* xdist isolation noise as ST2/ST3; none touch plugins/memory/isokron/.

Rule-6 / BUILD_DEVIATIONS / Open asks

Three deferred-MCP-tool surfaces still open (all gated on substrate-team dispatches; same pattern, same closure shape):

All three follow the same shape: signature is forward-stable, body swaps from raise <DeferredError> to mcp_client.invoke('kora__*', ...). The follow-on KR-N buckets on this lane are size O(60 LOC each) once the substrate tools land. PM coordinates the K-9 substrate dispatch.

KR-2 milestone closes here. The IsoKronMemoryProvider has:

• Full read coverage of the 6 load-bearing substrate surfaces
• A typed KoraSessionContext shape ready for K-6 (Constitution pre-screen middleware) to consume
• Honest deferred-write surfaces for the 3 substrate-side MCP tools we're waiting on
• ~/.kora/config.yaml memory.provider: isokron is enable-able in production with the known caveats documented in README "Operator pitfalls"

Standing by for KR-3 (beads-pattern consumer wiring) dispatch.

🤖 Generated with Claude Code