Roadmap¶

Honest list of what works, what's coming, and what's research.

Done (v0.1)¶

[x] sploink.wrap() monkey-patches Anthropic, Groq, OpenAI, Together, Ollama SDKs
[x] CallRecord + per-workflow trace + JSONL persistence
[x] Concurrent-workflow isolation via ContextVar
[x] Heuristic step classifier from observed call shape (no LLM)
[x] Static rule-based router (sploink/router.py)
[x] sploink.step(...) explicit-label context manager
[x] HTML report (python -m sploink.report)
[x] Force-directed canvas (python -m sploink.canvas)
[x] Graph data structure (DAG with topological-sort + validation)
[x] Architecture viewer — bipartite workflow ↔ substrate viz (python -m sploink.architecture)
[x] Bench against HotpotQA (preliminary results)

[ ] Graph.from_trace() — infer the workflow graph from observed traces. Critical for customers who don't use LangGraph/DSPy.
[ ] SubstrateGraph data structure — make the substrate side first-class data, like the workflow side.
[ ] Graceful fallback on routed-call failure (currently propagates exceptions).
[ ] Start/end timestamps on CallRecord so we can prove (or disprove) parallel execution from traces.
[ ] Tests for each wrapper — confirm sploink.wrap() actually wraps each SDK across SDK versions.
[ ] Rate-limit-aware Groq calls in the bench (so cloud results aren't contaminated).
[ ] Apples-to-apples bench comparison — F1 on the intersection of completed examples.

[ ] Telemetry-driven router — learn routing decisions from observed quality/cost outcomes per step type.
[ ] Workflow shape detection that drives shape-aware routing (RAG-shaped workflows route differently than coding-shaped ones).
[ ] External DAG consumption — first-class support for LangGraph and DSPy programs as input.
[ ] PII / prompt redaction in traces — opt-in for enterprise compliance.
[ ] Configurable trace storage backend (currently JSONL only).
[ ] Encryption at rest for traces.

[ ] Workflow design search — given a benchmark, search over which steps to include. (DSPy is the prior art here.)
[ ] Joint workflow + substrate optimization — propose new topologies AND new substrate assignments jointly. Open research problem.
[ ] Confidential compute — run cheap steps in TEEs / secure enclaves so enterprise customers can trust the edge.
[ ] Distributed compute network — user-contributed compute as a substrate type, with reputation and incentives.

Building our own inference engine (sploink dispatches to existing providers).
Building our own agent framework (use LangGraph or DSPy).
Building a vector database, retriever, or knowledge graph (orthogonal to compute routing).