Phase 8 directions — CausalGraph, ExplainAPI, Docker/Helm, Sepolia CI: what should we build next?

[web3guru888]

Phase 8 roadmap directions — deployment hardening, multi-agent federation, consolidation

Phase 7 is complete (5 sub-phases: Reptile → TaskSampler → ReplayBuffer → HyperNetwork → SleepPhaseOrchestrator). Phase 8.1 (DecisionTracer) is now in spec. This thread collects ideas for the rest of Phase 8 and beyond.

---

What Phase 8 has been planned to cover

From the Phase 8 directions discussion (#275), the original options were:

Option	Focus	Key components
A	Deployment hardening	Docker/Helm chart, Sepolia CI, doctor CLI, security hardening
B	Distributed multi-agent	Peer discovery, distributed consensus, gossip protocol
C	Introspection + explainability	DecisionTracer, CausalGraph, ExplainAPI
D	Consolidate/stabilise	Integration tests, benchmarks, coverage audits

The recommendation was A + C — hardening and explainability together.

Phase 8.1 (DecisionTracer) is Option C's first sub-phase.

---

Proposed Phase 8 sub-phases

Phase 8.1 — DecisionTracer            [Issue #276 — IN SPEC]
Phase 8.2 — CausalGraph               [planned]
Phase 8.3 — ExplainAPI                [planned]
Phase 8.4 — DockerFile + Helm chart   [planned]
Phase 8.5 — Sepolia CI integration    [planned]

---

Phase 8.2 — CausalGraph: decision lineage graph

Build a DAG of decision traces to answer: "which earlier decisions caused this outcome?"

Proposed components:

• CausalNode — wraps a DecisionTrace with parent references
• CausalGraph — DAG over CausalNodes, keyed by trace_id
• CausalGraphBuilder — subscribes to DecisionTracer, adds nodes and edges
• graph.ancestors(trace_id, depth=3) — returns contributing ancestor traces
• graph.dot() — exports GraphViz DOT for visualisation

Why it matters: enables post-hoc debugging ("which memory retrieval 5 ticks ago led to this wrong action?") and supports reinforcement learning from trace feedback.

---

Phase 8.3 — ExplainAPI: REST/WebSocket explanation endpoint

Expose DecisionTracer and CausalGraph over HTTP:

GET  /api/traces                     → recent traces list
GET  /api/traces/{trace_id}          → full DecisionTrace JSON
GET  /api/traces/{trace_id}/explain  → natural-language explanation
GET  /api/causal/{trace_id}/ancestors → CausalGraph ancestors
WS   /ws/traces                      → live trace stream

Stack: FastAPI + uvicorn, minimal dependencies, no frontend required.
Auth: optional token header (configurable), no auth by default in dev mode.

---

Phase 8.4 — Docker + Helm chart

Current state: Dockerfile exists in repo root. What's missing:

• Multi-stage build — separate builder/runtime stages (reduce image size from ~2 GB to ~400 MB)
• docker-compose.yml — ASI-Build + Prometheus + Grafana stack
• Helm chart — charts/asi-build/ with values.yaml, deployment.yaml, service.yaml, configmap.yaml
• health_check.py — --doctor CLI flag that verifies all subsystems are ready

---

Phase 8.5 — Sepolia CI integration

Current state: Sepolia bridge live at 0xE034d479EDc2530d9917dDa4547b59bF0964A2Ca. What's missing:

• CI step: run scripts/check_bridge.py against Sepolia on every push to main
• Mock Sepolia: pytest-web3 fixture that spins up Anvil locally for unit tests
• Bridge health Prometheus metric: asi_bridge_latency_seconds{chain="sepolia"}

---

Open questions for the community

1. ExplainAPI priority — is a REST endpoint for traces useful before Docker/Helm (Phase 8.3 before 8.4), or should deployment come first?
2. CausalGraph depth limit — should ancestors() be bounded at compile-time (e.g. depth ≤ 10) or configurable per-call?
3. Helm vs Docker Compose — which deployment target matters most for contributors: local Docker Compose, Kubernetes Helm, or cloud-provider one-click (e.g. AWS CDK)?
4. ExplainAPI auth — should the default be no-auth (simpler) or token-required (safer)? What's the expected deployment environment?

Share your thoughts below — or open a sub-phase issue if you want to take ownership of any of these tracks.