Roadmap¶
This is the active 2.0 roadmap. The 1.x research history remains in Git history and release notes; its tool counts and protocol claims are not the current contract.
Current published prerelease: 2.0.0-rc.2¶
The latest published 2.0 prerelease is v2.0.0-rc.2. Its tag identifies the immutable tested artifact; compare later main commits with that tag before assuming a feature is in the download.
- Packaged HTTP profile and logging configuration verified from the built jar.
- Explicit project roots and root-relative aliases; traversal and symlink-escape tests.
- HTTP bearer enforcement by default, no built-in key, explicit per-tool scopes, and a scope-coverage test for the 24 exposed tools.
- Shared output policy with bounded, redacted diagnostics and generic tool errors.
- Credential, audit, and stored-plan execution tools removed from the MCP catalog pending redesign.
- SDK input validation and fail-closed schema setup.
- Cached callbacks and credential digests.
- Explicit error for markerless and hybrid project auto-detection.
- Nonroot Docker image with verified Gradle/sbt downloads and reusable Maven build cache.
- Bounded, offline Maven POM structural validation with model-visible fixed configuration diagnostics and packaged HTTP/stdio privacy probes.
Release candidate gate: rc.1 and later¶
The checks below apply to each candidate, including rc.2, on its final tag commit.
- Finish full Java 21/23/25 verification, packaged stdio and HTTP protocol smoke, and strict MkDocs build.
- Red-team prompt injection, path races, scope bypass, malformed and oversized JSON-RPC, and synthetic privacy canaries through both transports.
- Reconcile every public tool description and configuration page with the 24-tool runtime catalog; generate a machine-checked catalog in CI.
- Measure callback overhead, process memory, bounded output behavior, and build latency on synthetic Maven, Gradle, and sbt fixtures. Set budgets only from measured baselines.
- Run two independent exact-head PR reviews and address every inline finding. Tag the prerelease only when all final-head gates are green; record the tag and release outcome separately.
The release evidence matrix is a living gate, not a list of assumed passes. Attach the exact commit, environment, command, and summarized result to the release PR; a result on one machine is not a universal performance guarantee.
| Gate | Evidence to attach | State at roadmap update | Owner |
|---|---|---|---|
| JDK 21/23/25, packaged stdio/HTTP, strict docs | CI checks, packaged protocol smoke, mkdocs build --strict | Recheck on final release head | Release engineer |
| MCP protocol | Pinned official runner, advertised-capability scenarios and DNS-rebinding case on final jar | Recheck on final release head | Protocol engineer |
| Privacy and adversarial cases | Synthetic canary matrix on both transports; counts/status only | Implemented; recheck on final release head | Security reviewer |
| Public contract | Generated runtime catalog and docs drift check | Implemented in CI; recheck on final release head | Docs engineer |
| Performance | scripts/benchmark-release-gate.py aggregates for Maven, Gradle, sbt, callback and bounded-output stress, with warm-cache/offline provenance | Five-run offline candidate-tree baseline recorded in workflow; pinned-runner budget still pending | Performance engineer |
| Dependencies | Green keyless PR Dependency Review, then 0 open Dependabot alerts on the final default-branch commit using an authorized maintainer credential | Keyless gate implemented; recheck after final merge | Release engineer |
| PR review | Two fresh-checkout role-tagged reviews on final SHA; all inline threads answered | Pending final SHA | Quality and adversarial reviewers |
Before any release tag, the release engineer checks this matrix against the exact final default-branch commit and records the outcome. Do not infer a pass from an earlier PR head. A PR dependency review is not the final-branch alert audit; record both on the release head, and fail closed if alert access is unavailable. The performance script deliberately has no invented latency or memory budget. Compare repeated runs on a pinned runner and cache state before proposing one.
Stable 2.0 gate¶
- Publish a migration guide for required roots, HTTP keys and scopes, catalog removals, result shape, and supported protocol revision.
- Verify official MCP 2025-11-25 conformance with real stdio and Streamable HTTP clients. The MCP 2026-07-28 specification is published, but the current Java SDK 2.0.1 targets 2025-11-25; do not advertise 2026 support until the SDK and tests cover it.
- Use the optional full requirement audit to track this gap: the current packaged HTTP baseline passes 6 of 30 scored server scenarios, fails 23, and has one warning; three other server scenarios are unscored. This audit is nonblocking and is not a conformance claim. The six selected official scenarios plus adversarial probes remain the release gate.
- Test the published Docker image with a clean project mount, no host secrets, nonroot UID, and a read-only Maven cache.
- Verify the dormant annotations are absent from the public catalog; decide separately whether their retained services and methods should be deleted.
- Ensure privacy scanning on changed files and generated docs, without printing matched values to CI output.
- Replace the retained private
SupplyChainServiceOSV parser with the bounded lookup client before exposing its supply-chain tools.
Next development slices¶
Order the next slices by observed user friction, with a measured baseline or test before choosing an implementation:
The first post-RC increment already added MCP structuredContent and a bounded outputSchema for analyze_build_output while preserving its JSON text result. The same bounded contract now covers execute_build_command, preserving its existing JSON text and one-build execution behavior. Both transports, schema conformance, privacy, and adapter overhead were checked; see the result contract. The assertion-versus-compilation classification bug was also fixed and tested for Maven, Gradle, and sbt.
A synthetic 25.4 MB Maven failure showed that the old 32 KiB head / 96 KiB tail capture omitted the sole compiler error in the middle, leaving a failed build with errorCount: 0. Maven analysis captures at most 13 complete compiler lines of 2 KiB each while draining. Matching 28 MB synthetic Gradle/sbt failures now exercise bounded middle-candidate capture for execution and analysis, with both streams and the shared privacy projection. The unchanged head/tail capture and separate outputTruncated and diagnosticsTruncated signals remain.
Synthetic summary-only failed-test cases now produce a fixed test diagnostic for Maven, Gradle, and sbt even when the detailed assertion was outside retained output. Packaged HTTP and stdio checks require structured/text parity and suppress private canaries. This preserves a useful failure category without exporting raw test messages.
The first feature slice exposes three new static native workflows through prompts/list and prompts/get on both transports, while retaining the legacy prompt_* tools. HTTP requires prompt:read; prompt messages never interpolate paths, commands, arguments, or logs. The empty native resources capability remains advertised because its selected official conformance scenario failed when it was removed; resources/list returns no entries. Both-transport protocol, argument, scope, and synthetic privacy tests now cover the slice.
- Repair from private diagnostics. If redacted structured diagnostics cannot explain a real synthetic failure, provide a short-lived local artifact reference and a separate, explicit caller-approved share operation. Keep raw content local by default; scope artifact access to the caller and project, bound size and lifetime, and recheck the path at read time. Acceptance: injected instructions, private-path/email/secret canaries, cross-caller reads, symlink races, expiry, and oversized artifacts cannot reach model-visible results without approval. Avoid building a general file browser or log-export API.
- Reduce repair round trips. Observe a small set of representative repair sessions first. Then offer one constrained
diagnose → suggest → verifyoperation or build-plan primitive only if it demonstrably reduces calls and time. It requires caller ownership, cancelable process trees, TTL, revalidated project paths at execution, and no stored credentials. Acceptance: deterministic synthetic repair fixture, cancellation/race tests, privacy-safe result, and a measured comparison with the current tool sequence. Keep the first slice to a single build tool and command family. - Improve hot-path cost with evidence. Use the release-gate fixture and recorded workloads to profile subprocess capture, Maven/Gradle cache behavior, and parallel dependency lookup before tuning. Acceptance: the same tests and security/privacy cases pass, with an environment-matched latency/RSS comparison; no global budget inferred from a developer laptop.
- Adopt official Tasks when stable in the Java SDK. The 2026-07-28 core revision is published, while Tasks now lives in a separate extension. Wait for suitable Java SDK support and official conformance before replacing the dormant custom async protocol. Acceptance: cancellation, ownership, expiry, interoperability, and migration tests with real clients.
- Carry authoritative execution status (implemented after RC1). Built-in Maven, Gradle, and sbt
execute_build_commandresults now derivesuccessfrom the completed subprocess exit code, including when retained markers disagree. Plugins without a typed process result expose unknown status. The typed Gradle/sbt path also preserves bounded stdout diagnostics on failure, which the legacy Java method previously discarded when stderr was empty. The legacy method remains available for plans and async callers. - Preserve async middle diagnostics. Gradle/sbt synchronous capture now retains bounded middle candidates on failed calls. The dormant async path has the same class of loss. Extend its private collector only with tool-specific recognition, cancellation and ownership tests, and an environment-matched latency/RSS comparison. Reuse the existing privacy projection.
- Audit remaining file-access races and portability. Configuration validation and CVE scanning anchor reads, and the project guard checks markers through a single held handle where Java
SecureDirectoryStreamis supported. The guard's compatibility fallback on other providers retains canonical-path checks but remains vulnerable to concurrent swaps; validation and CVE scanning fail closed there. A regular-file-to-pipe swap between final-attribute check and open can occupy the scan's bounded worker until restart; other reads and execution tools also need a targeted path-swap audit. Before stable 2.0, choose a portable nonblocking file boundary or require isolated execution for untrusted projects, then test directory/file swaps on both transports and supported platforms. - Finish OSV severity analysis without false-clean results. Bounded batch lookup now reports vulnerability presence with unknown severity because OSV batch results are sparse, and fails incomplete when any recognized declaration was not checked. A vetted CVSS v3.1 base-vector scorer now supports single-coordinate OSV results. Sparse batch results still need bounded advisory-detail retrieval with explicit cost limits before promising severity-filtered scan counts; temporal/environmental vectors and other CVSS versions remain unknown. Add resolved transitive dependency graph coverage only after defining per-build-tool extraction and egress consent. Benchmark against representative projects and maintain synthetic no-network fixtures.
- Resolve optional single-dependency security output (implemented after RC1).
check_dependency_version(includeSecurityInfo=true)sends coordinates and version to OSV.dev only withcurrentVersionand exposes a bounded aggregate CVE count, lookup status, and known/unknown highest severity. Individual findings remain local. Keep malformed/failed lookup and both-transport privacy tests on the final release head.
Architecture debt stays visible alongside features: the 11 annotations on unregistered service methods have been removed, while their ordinary methods and beans remain for a separately justified deletion decision. Consolidate process execution and cache policies where measured duplication exists; preserve the single authorization/output-policy boundary for new tools; and keep the generated public catalog as the docs drift check. Each decision needs a failing test, a specific user or maintenance cost, or a measured hot path rather than a speculative rewrite.