Future extensions / roadmap

skillsmith is architected so that most new capability arrives by adding an Adapter (new ecosystem) or a rule pack / hook (new analysis) — not by rewriting the engines. This page is the roadmap and the rationale for the seams already in place.

Ecosystem adapters

Shipped: claude-skill, claude-agent-sdk, mcp.

Stubs (interfaces present, TODO-marked in packages/adapters/.../stubs.py):

Adapter	What it will map onto core models
openai	OpenAI Assistants / function-tool schemas + system instructions.
langgraph	LangGraph node prompts + tool bindings.
crewai	CrewAI Agent/Task (role, goal, backstory, tools).
cursor	.cursor/rules and Cursor command/prompt files.

Each becomes a real adapter by following Extending. Because the engines are ecosystem-agnostic, audit/eval/augment/heal all light up for free once discover and parse exist.

Analysis & augmentation

• Model-backed generator for augment. Today TemplateGenerator emits a typed skeleton; the pipeline (validate-in-sandbox → variance/token gate → PR) is real. The next step is a Generator backed by a model that fills in the deterministic logic for a deterministic-candidate block, behind the same contract (stdin→stdout, nonzero on error) so the sandbox validator is unchanged.
• Semantic frontmatter diffing. A YAML-aware diff that classifies frontmatter changes by meaning (e.g. a description reword is a triggering-relevant change) to complement the JSON contract diff.
• DeepEval scored metrics. The eval package's scored extra wires DeepEval for pytest-native scored metrics; deterministic scorers stay ground truth and LLM-as-judge is never the sole gate.
• promptfoo eval gate in CI. ci.yml carries a promptfoo job; the wrapper already refuses configs whose only assertions are LLM-graded.

Self-healing

• Live MCP introspection. mcp.live_snapshot currently falls back to the committed *.mcpc.json. The next step spawns the configured server inside the sandbox and issues tools/list over stdio for a true live contract.
• Native MCP versioning (SEP-1575 / SEP-1400). When servers can declare a contract version, skillsmith will compare versions directly instead of structurally diffing snapshots. The diff module is written so this becomes an additional, preferred code path rather than a rewrite.
• Richer behavioral-drift fingerprinting. Beyond canary output hashes — structured output schema validation and tolerance bands for numeric drift.

Sandbox

• Firecracker/microVM path via microsandbox/e2b for the strongest isolation tier, with gVisor and bubblewrap/nsjail as the current layered fallbacks. Backend selection and limits are already config-driven (sandbox.backend_preference).

Portability

• The public contracts already emit JSON Schema (skillsmith schema export), so non-Python consumers (dashboards, other-language adapters, CI gates) can integrate without importing the Pydantic models. Cross-language adapters are a natural extension.