ASIF 0.4 gaps and implementation status¶
Reviewed against the current specification and recorded test results on 2026-09-26.
The OpenAPI-style organization work is complete for the core, continuation profile and destination report: 128 object definitions, fixed-field tables, linked types and requirements, and 140 schema-validated examples, including attachment representations and availability states. Object/field links and generated-document freshness are checked. These counts exclude the separately documented package transport objects.
The reference subset now has Python and standalone TypeScript CLIs, shared continuation tests and local package/signature exchange checks. Both are locally authored; this does not close the independent-evidence gap. The project license is now MIT.
The remaining work has three distinct meanings: specification gaps need portable rules or machine-readable definitions; implementation gaps need executable support and broader tests; evidence/release gaps need independent exchange or the publication steps already proposed in governance. A missing vendor adapter is not a missing core session field. Passing local tests does not establish general interoperability or permission to execute.
| Priority | Area | Gap type | Work completed | Still required |
|---|---|---|---|---|
| P0 | Semantic validation | Implementation / evidence | Core IDs/references, causal/supersession graphs, provenance spans, selected-history state, resource integrity, checkpoint checks, branch-selected task dependency/revision checks, dependency/base closure and selected report invariants. | Map each normative requirement to validation and positive/negative fixtures, then complete rule-by-rule coverage, including remaining lifecycle/coverage/requirement semantics, inherited Git prerequisites and runtime evidence binding. The validator explicitly claims a subset. |
| P0 | Configuration equivalence | Implementation / evidence | Neutral scope/predicate evaluation, effective order, merge groups and declarative policy precedence. | Evidence that source and destination apply equivalent instructions and enforcement. Vendor dialect mappings remain deferred by scope. |
| P0 | Request reconstruction | Implementation / evidence | Deterministic neutral projection preserves ordered typed inputs, tool definitions, settings and resource references without injecting instructions twice. | Provider wire encoding, actual token/media budgeting and accepted model/compaction adaptations. Vendor encodings remain deferred. |
| P0 | Adapter fidelity | Implementation / evidence | Core/reference fixtures and explicit unknown/unsupported behavior are available for adapter authors. | Actual source adapters with stable identity mappings, coverage and return-export evidence. No vendor adapter is included. |
| P0 | Independent evidence | Evidence | Reproducible Python/TypeScript checks, shared action-specific continuation outcomes and local bidirectional package/signature exchange. Shared authorship is explicit. | Two independently authored implementations exchanging independently produced sessions. Current independent exchange results: zero. |
| P0 | Destination preflight/import | Implementation / evidence | Source/report hash, exact supplied capability snapshots, component bindings, expiry, action-specific subject closure, core requirement evidence, account/version declarations, blockers and synthetic-execution refusal are checked. | Live destination identity/capability checks, credentials, writer coordination, native import/index maintenance, transactional native-store rollback and actual next interaction. These remain deferred. |
| P1 | Partial/streaming calls | Specification / implementation / evidence | Versioned stream and external-binding features, schemas, byte assembly, and two immutable captures of a progressing call. Both validators check amendment identity, selected chains, retries and external-predecessor refusal using shared cases. | More protocols/media, partial result variants, independent implementations and a larger adversarial corpus. Unsupported protocols require explicit refusal. |
| P1 | State transitions | Implementation / evidence | Tests reject contradictory decision outcomes, unaccounted task reopenings and results after a terminal result; explicit decision corrections retain evidence. Task prerequisites now have reference scope, branch-local revision selection, cycle checks, descriptive state rules and partial-history fixtures. | Broader lifecycle combinations, workflow-specific scheduling and real source observations. |
| P1 | Workspace restoration | Specification / implementation | Fresh-destination file/directory snapshots and deltas, deletion checks, modes, path collisions, staged rollback and no-overwrite tests. Git index declarations are checked separately. | Actual Git/index/object/submodule/LFS and symlink restoration, richer selection dialects, hostile concurrent filesystem access and crash durability. These capabilities are explicitly unsupported locally. |
| P1 | Neutral configuration dialects | Specification | The local reference evaluates asif.activation/0.1 and asif.policy/0.1, documented in REFERENCE.md. |
Decide whether these local dialects become optional standard dialects. If adopted, publish normative syntax, schemas, error/refusal rules and conformance fixtures; local evaluator behavior alone is not the portable contract. |
| P1 | Transport object schemas | Specification / validation | Package manifest, file and signature objects have field tables, examples and executable checks. | Add machine-readable schemas and include these objects/examples in the reference-generation and schema-validation checks. The current 128-object/140-example checks do not cover them. |
| P1 | Packaging/authenticity | Evidence / optional extensions | Optional ZIP_STORED envelope, exact inventory, bounded verification and detached Ed25519 manifest signatures with caller-supplied trust. | Independent cryptographic/container review. Key lifecycle/trust distribution needs a separately defined scope; encryption and large-object transports remain optional future extensions, not core release blockers. A signature does not establish semantic truth or continuation readiness. |
| P1 | TypeScript representation coverage | Implementation | A standalone port supports all nine commands and passes shared continuation expectations and local package/signature exchange checks. | Broaden exact numeric handling, ZIP variants and bracket-glob support if claimed. Current limits are explicit unsupported cases in REFERENCE.md; the port is not full ASIF conformance or independent evidence. |
| P1 | Limits/redaction | Implementation / evidence | Published parser/resource/archive limits; duplicate/depth/path refusal; exact-pattern audit across duplicated JSON/base64/resource content without echoing matches. | Broader adversarial limits corpus, format-aware sanitizers, compressed/media inspection and secret discovery. Unavailable resources cannot be audited. |
| P0 | Standard release readiness | Evidence / release | Governance proposes maturity, compatibility, review and release criteria; MIT licensing is in place for the specification and reference code. | Complete the remaining applicable maturity gates: named independent maintainers, contribution terms and publication-venue requirements, public review, a compatibility policy, and pinned release artifacts. Licensing alone does not satisfy these gates. |
P0 items apply to the stated conformance or release claim. Live import and continuation evidence is needed when claiming those capabilities; it does not require the core specification to define every vendor's native format. P1 capabilities must either have supported, tested semantics or remain explicitly outside the implementation's claimed scope.
The next vendor-neutral work is to map normative requirements to checks, resolve the status and definitions of the neutral configuration dialects, and bring the transport objects into schema/example validation. Vendor adapters and live destination integration remain deferred; independent evidence remains a standardization requirement even while that integration work is deferred.
Evidence: reference scope and commands, local reference results, continuation checks, object and navigation checks, and implementation record. All continuation reports in the example corpus are synthetic. The native-agent work above is deliberately deferred while ASIF remains vendor-neutral.