Milestone 3.9.3 — Phase 3 Exit Gate and Handoff Document

Status: Planned
Goal: 3.9 — Phase 3 quality gate
Phase: 3 — Memory Engine and Operator Platform
Estimated effort: 1–2 days

Why This Milestone Exists

Phase 3 is the largest and most complex phase. A formal exit gate ensures nothing was shipped partially and that the system is ready for Phase 4 (multi-provider support, billing, and behavioural fingerprinting).

Phase 3 Exit Checklist

Functionality

A real conversation produces extractable memories within 30 seconds of session completion
Subsequent requests include injected memories verified via X-IBEX-Memories-Injected header
Context assembly fallback (service timeout) gracefully returns original messages without error
Dashboard shows real data for all main views (agents, memories, analytics, sessions)
Directive update propagates to proxy cache within 1 second
Session archive written to MinIO within 30 seconds of session completion
Memory export produces valid CSV/JSON for 1,000-memory dataset

Security

All Phase 1 and Phase 2 security integration tests (1.5.1) pass — zero regression
Cross-tenant memory isolation verified: Org A memories never in Org B context
GDPR deletion cascade removes all data (memories, sessions, archives) for a given org
Memory content is never logged, traced, or included in error messages
Directive content is never logged or included in error responses

Performance

Context assembly p95 < 50ms at 50 concurrent requests (3.9.2 load test passes)
Memory write p95 < 200ms (including embedding call) — measured in integration test
pgvector semantic search p95 < 30ms for 100,000 memories per agent
Proxy overhead p99 < 25ms (5ms over Phase 2 target; context assembly adds ~5ms via gRPC)
Embedding cache hit rate > 70% under sustained extraction load

Operations

make e2e-smoke-p3 exits 0 with real OpenAI API key
All 6 new services start, pass health checks, and have working /ready endpoints
Celery workers process extraction queue with < 10s lag under normal load
All services have .env.example files documenting every env var

Documentation

docs/roadmap/CURRENT_STATE.md updated with Phase 3 completion
docs/DEVELOPMENT_GUIDE.md updated with Phase 3 local setup (6 new services)
All ADRs (0032–0044) written and indexed in docs/adr/README.md
OpenAPI spec for management API exported to docs/api/openapi.json
GitHub release tag v0.3.0 created after this milestone merges

`make e2e-smoke-p3`

bash

#!/usr/bin/env bash
# infra/scripts/e2e_smoke_p3.sh
# Phase 3 end-to-end smoke: memory extraction + injection verified.
# Prerequisites: full local stack (make compose-dev-up), all services running.
 
set -euo pipefail
 
: "${OPENAI_API_KEY:?Required}"
 
PROXY="${IBEX_PROXY_ADDR:-http://localhost:8080}"
TOKEN="${IBEX_DEV_TOKEN:-ibex_dev_sk_LOCALDEVELOPMENTONLY}"
AGENT="${IBEX_DEV_AGENT_ID:-00000000-0000-0000-0000-000000000003}"
 
echo "=== Phase 3 E2E Smoke Test ==="
 
# ── Turn 1: Establish a fact ──────────────────────────────────────────────────
echo "Turn 1: Establishing fact..."
R1=$(curl -sf -X POST "$PROXY/v1/chat/completions" \
    -H "Authorization: Bearer $TOKEN" \
    -H "X-IBEX-Agent-ID: $AGENT" \
    -H "Content-Type: application/json" \
    -d '{"model":"gpt-4o-mini","messages":[
        {"role":"user","content":"My database host is db.smoketest.example.com. Please confirm you noted that."}
    ]}')
SESSION_ID=$(curl -sI -X POST "$PROXY/v1/chat/completions" \
    -H "Authorization: Bearer $TOKEN" \
    -H "X-IBEX-Agent-ID: $AGENT" \
    -H "Content-Type: application/json" \
    -d '{}' 2>/dev/null | grep -i "x-ibex-session-id" | awk '{print $2}' | tr -d '\r')
echo "Session ID: $SESSION_ID"
echo "PASS: Turn 1 complete"
 
# ── Wait for extraction (max 30s) ─────────────────────────────────────────────
echo "Waiting for memory extraction..."
EXTRACTED=0
for i in $(seq 1 30); do
    MEMORY_COUNT=$(curl -sf "$PROXY/../api/v1/agents/$AGENT/memories" \
        -H "Authorization: Bearer $TOKEN" | python3 -c "import json,sys; d=json.load(sys.stdin); print(len(d.get('items',[])))" 2>/dev/null || echo "0")
    if [ "$MEMORY_COUNT" -gt 0 ]; then
        EXTRACTED=1
        echo "PASS: $MEMORY_COUNT memories extracted after ${i}s"
        break
    fi
    sleep 1
done
[ "$EXTRACTED" = "1" ] || echo "WARN: No memories extracted within 30s — check worker logs"
 
# ── Turn 2: Verify memory injection ──────────────────────────────────────────
echo "Turn 2: Verifying memory injection..."
HEADERS=$(curl -sI -X POST "$PROXY/v1/chat/completions" \
    -H "Authorization: Bearer $TOKEN" \
    -H "X-IBEX-Agent-ID: $AGENT" \
    -H "X-IBEX-Session-ID: $SESSION_ID" \
    -H "Content-Type: application/json" \
    -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"What database host should I use?"}]}')
 
INJECTED=$(echo "$HEADERS" | grep -i "x-ibex-memories-injected" | awk '{print $2}' | tr -d '\r' || echo "0")
if [ "${INJECTED:-0}" -gt "0" ]; then
    echo "PASS: $INJECTED memories injected in Turn 2"
else
    echo "WARN: 0 memories injected — extraction may not have completed"
fi
 
echo ""
echo "=== Phase 3 E2E Smoke: COMPLETE ==="

Phase 4 Readiness Checklist

Before Phase 4 begins, verify:

ibex_core.behavioral_fingerprints schema drafted (ADR written, migration in progress)
Rate limiter ready for Lua-script upgrade (interface unchanged since Phase 1)
Embedding service abstraction allows second model (ADR-0033 model upgrade path confirmed)
Dashboard extensible for multi-provider analytics without structural change
All 44 ADRs (0001–0044) indexed and linked from the architecture doc