This roadmap describes where the framework is, where it is going, and what is explicitly out of scope. Items move from exploratory to in progress to completed through issues and PRs; the roadmap is updated when those transitions happen.
Legend: ✅ completed · 🚧 in progress · 🧭 exploratory · ❌ explicit non-goal
✅ 19 framework modules under @workspace/framework:
ingestion, cognition, organization_intelligence, evaluation, intelligence_operations, collaboration, security, agents, skills, providers, memory, orchestrator, workflows, candidate_graph, organization_graph, reports, registry, benchmarks, shared.
✅ OpenAPI-first contract (lib/api-spec/openapi.yaml) with generated React Query hooks and Zod schemas.
✅ Deterministic cognition benchmark suite — 10 scenarios, 5 guard categories (determinism, provenance, calibration, snapshot, mutation guards), CI-runnable.
✅ Security layer — identity, multi-tenant isolation, capability-based authorization, append-only audit, sensitive-evidence redaction, session integrity, 8 security benchmarks.
✅ Operator console with sections for every framework module; confidence and provenance rendered as first-class visual primitives.
✅ Reference in-memory backends for memory, graphs, sessions, audit. No external services required to run the skeleton.
🚧 Postgres + pgvector memory adapter. Reference implementation of the memory backend using the schemas already scaffolded in lib/db. Goal: zero behavioral drift vs. the in-memory adapter; same benchmark outputs.
🚧 Cognition snapshot diffing UI. A console view that shows why two snapshots differ when a benchmark update is required.
🚧 Adversarial benchmark suite. Companion to the cognition suite, focused on inputs designed to confuse a single-agent system: contradictory references, manufactured signals, missing context.
🚧 Provenance integrity check in CI. Static analysis that fails the build if a public framework function produces an artifact without provenance.
🚧 Reviewer calibration export. A deterministic export format for calibration weights so they can be carried across deployments.
🧭 Plugin packaging guide. Third-party agents, skills, and ingestion adapters as separate packages that register themselves at boot.
🧭 Hosted memory layer reference. A reference deployment of the memory layer behind a thin authenticated HTTP surface, for teams that want a managed backend without giving up framework determinism.
🧭 Provider router. Routing decisions (cheap vs. strong model, structured-output vs. not) as a declarative policy, not in-line branches.
🧭 CLI for benchmark replay. pnpm framework replay --scenario <id> --until <step> for step-through cognition debugging.
🚧 Edge inference primitives. Confidence-weighted edge derivation in both the candidate and organization graphs.
🧭 Trajectory primitives. Time-aware reasoning over the candidate graph — career inflection detection, signal aging, evidence freshness.
🧭 Cross-graph fit explanations. Surfacing fit assessments as a small subgraph of supporting edges, not as a scalar.
🧭 Beyond per-reviewer calibration. Calibration as a function of the (reviewer, role family, organization) tuple, with explicit drift detection when reviewers move between organizations.
🧭 Outcome-feedback loops. Closed-loop learning from longitudinal outcomes, with snapshotted before/after weights so changes are auditable.
🧭 Disagreement attribution. Stable explanations for why specific reviewers disagree on specific kinds of evidence.
🧭 Multi-organization reasoning scenarios. Benchmarks that exercise cross-tenant carve-outs, calibration namespaces, and global benchmark contributions.
🧭 Long-horizon scenarios. Scenarios that run over simulated quarters, asserting that long-running calibration and memory updates remain deterministic.
🧭 Stress benchmarks for the security engine. Property-based tests that fuzz access requests against the rule set.
❌ An ATS or pipeline tool. Stages, kanban, applicant tracking belong in products built on top of the framework, not in the framework.
❌ A recruiting CRM. Outreach sequences, email templates, calendar integrations — out of scope.
❌ A CV matcher. Document → vector → score reductions discard everything the framework is designed to preserve.
❌ A chatbot recruiter. The framework's surface is data, not conversation.
❌ Black-box scoring models. Every score must be reducible to provenance-rooted signals.
- Pre-1.0 the framework will make breaking changes as needed; deprecations are documented in CHANGELOG.md and called out in PRs.
- The OpenAPI contract is versioned with the framework. Generated client/zod files are regenerated from the contract — never hand-edited.
- Snapshot files are part of the public interface in the sense that they represent the framework's committed behavior. Changes are intentional and reviewed.
- Open a discussion or issue using the
cognition_proposaltemplate. - Strong proposals come with a benchmark scenario that captures the desired behavior.
- "It would be cool if…" issues are welcome but lower priority than ones with reproducible evidence.