Decisions and Outlook
Current marks behavior in this release line. Direction records intended
work and its readiness criteria. Research marks unresolved design questions.
Roadmap labels carry no release dates. Entries link to the subsystem chapters
that describe the current implementation and limits.
Settled decisions
Section titled “Settled decisions”Keep one analysis contract below every protocol
Section titled “Keep one analysis contract below every protocol”Status: Current. MCP, LSP, CLI, and library hosts call shared analysis services. Those services define the canonical result and its proof semantics; protocol adapters handle transport, lifecycle, and error mapping.
Trade-off: Protocol-specific shortcuts are constrained so the same operation has the same meaning in an editor and an agent.
Derive exact identity from structure and resolution
Section titled “Derive exact identity from structure and resolution”Status: Current. Parsed nodes and language resolvers provide exact identity for declarations, calls, formals, receivers, imports, and members. Display strings, source snippets, regular expressions, and delimiter parsing remain presentation or discovery data.
Trade-off: Bifrost reports ambiguity or unknown when structured evidence leaves more than one possible target.
Persist durable facts; materialize expensive semantics on demand
Section titled “Persist durable facts; materialize expensive semantics on demand”Status: Current. The relational store holds queryable declarations and structural relations. Bifrost builds semantic artifacts, bounded ICFG views, and solver state for a request, then retains them only under complete validity keys.
Trade-off: The first deep query may have to construct its semantic view. Ordinary navigation avoids the cost of maintaining every possible dataflow derivation.
Key caches by semantic inputs
Section titled “Key caches by semantic inputs”Status: Current. Reuse keys cover content, workspace generation, language and grammar epochs, adapter and IR semantics, configuration, dependencies, and active models as applicable. Age has no bearing on validity.
Trade-off: Keys and invalidation logic carry more detail. They permit safe reuse across processes and compatible worktrees.
Restrict complete caches to complete values
Section titled “Restrict complete caches to complete values”Status: Current. Complete caches accept only successful, complete computations. Cancelled, budget-exhausted, stale, failed, and semantically partial results remain outside the cache.
Trade-off: A later request may repeat partial work. Publishing that work as complete would give subsequent requests a false completeness claim.
Put language rules in language crates
Section titled “Put language rules in language crates”Status: Current. Language crates implement imports, visibility, overloads, hierarchy, call shapes, and ecosystem conventions. Shared engines consume their typed facts and outcomes through common contracts.
Trade-off: Each language needs a substantive adapter. Clients can still handle evidence consistently through the common result contract.
Use a bounded distributive interprocedural kernel
Section titled “Use a bounded distributive interprocedural kernel”Status: Current. IFDS-family fact propagation and IDE-style edge functions share an explicit ICFG plus contracts for budgets, termination, path quality, summaries, and witnesses. Analyses outside the finite distributive domain use a separately named abstraction.
Trade-off: The solver covers a narrower domain than a general abstract interpreter. That domain has testable rules for behavior, reuse, and incompleteness.
Include evidence quality and coverage in results
Section titled “Include evidence quality and coverage in results”Status: Current. Client-visible results carry proof, completeness, provenance, work, scope, and termination. An empty partial response does not support a clean result.
Trade-off: Results require more structure than a list of locations. Clients use those fields to describe the result accurately.
Active directions
Section titled “Active directions”Compose immutable usage facts
Section titled “Compose immutable usage facts”Status: Direction. The proposed usage engine stores immutable file-local resolution material under its content identity. A query composes those facts for the selected workspace, avoiding a mutable global graph and chains of parent deltas. See the usage-analysis design for the composition contract and current implementation, and the storage direction for the finer-grained cache question. Shipping requires parity with current resolvers, bounded memory, accurate cancellation and cap reporting, and useful reuse across linked worktrees.
Close known semantic gaps
Section titled “Close known semantic gaps”Status: Direction. Dataflow conformance tests expose recurring sources of incomplete results: unresolved dispatch, field identity, destructuring and unpacking, language hooks, and missing dependency behavior. Each gap needs structured semantics and positive and near-miss tests. Existing cases remain incomplete until that support exists. See Dataflow Maturity and Limits for the current experimental boundary.
Persist reusable class-set and procedure knowledge
Section titled “Persist reusable class-set and procedure knowledge”Status: Direction. Complete content-keyed summaries could reduce repeated receiver, type, and procedure analysis. Shipping them requires stable identity, dependency closure, complete witness semantics, and safe invalidation. Type-flow results remain experimental and are often partial or inconclusive on broad corpora. Summary Identity and Reuse specifies the model-side identity gate, while Summary Lifetimes separates authored models from solver reuse.
Complete incremental policy evaluation
Section titled “Complete incremental policy evaluation”Status: Direction. The store already represents policy evaluation units. The general coordinator still needs to select reusable units, recompute affected units, and prove parity with a fresh full evaluation. The schema alone does not provide incremental execution.
Automate dependency discovery
Section titled “Automate dependency discovery”Status: Direction. Automatic dependency discovery and semantic-pack activation could improve coverage. Every discovered artifact, pack, producer, compatibility decision, and rejection must remain observable. Activation must be based on identified dependencies and compatibility evidence. The current diagnosable pipeline is described under Current Boundaries in the semantic-model chapter.
Add stable public analysis capabilities
Section titled “Add stable public analysis capabilities”Status: Direction. Typestate and other deeper analyses can enter the serializable extension API after their identities, limits, capability negotiation, and completion semantics stabilize. An internal or policy-only implementation is insufficient for a public extension capability. The current experimental boundary is summarized under Dataflow Maturity and Limits.
Research questions
Section titled “Research questions”Remote service deployment
Section titled “Remote service deployment”Status: Research. A remote MCP or shared analysis service changes the trust model. The local-first service boundary describes the current baseline. A hosted design still needs explicit policies for source custody, authentication, tenancy, cache isolation, scheduling, cancellation, version skew, audit logging, and deletion.
Richer heap and context precision
Section titled “Richer heap and context precision”Status: Research. Heap abstraction, aliases, context selection, and dynamic dispatch can improve precision, enlarge the state space, and invalidate summary identities. Candidate designs need independent oracles and measurements on real corpora. Finding more paths in one fixture is insufficient evidence.
Concurrency guarantees
Section titled “Concurrency guarantees”Status: Research. Structured summaries could support typed concurrency effects and focused race checks. A “race free” result would need a declared guarantee boundary for thread creation, synchronization, aliasing, native calls, and unmodeled code. Positive findings and complete absence claims require different evidence.
Non-goals
Section titled “Non-goals”Bifrost does not currently aim to:
- replace language compilers or claim compiler-complete semantics for every valid program;
- persist every expanded interprocedural graph and derivation for every query;
- hide unsupported semantics behind text or name heuristics;
- interpret may-reachability as runtime path feasibility;
- claim equal precision for every language because the adapters share an API; or
- turn a deterministic fixture or narrow demonstration into a global accuracy or scalability claim.
Open product questions
Section titled “Open product questions”- Should a hosted service remain an explicitly separate product architecture, or should the local runtime expose tenancy primitives in anticipation of it?
- Which analysis capabilities should receive long-term wire compatibility first: semantic relations, value flow, typestate, or retained policy reports?
- How much roadmap detail belongs in public docs before a direction has an executable acceptance test?
- Should public architecture snapshots be versioned per release, or should this guide always describe only the current release line?
Discuss or contribute
Section titled “Discuss or contribute”To propose a direction or describe a missing use case, open a feature request. An observed problem is enough; an implementation proposal is optional. For an informal discussion, join the Bifrost Discord.