Status: v0.28.0 published and verified.
PR #42 merged after every CI check
passed. The release workflow
passed its full gate and published v0.28.0
from commit fb8cddc. All downloadable binaries and installers were then
checksum-verified, and the published macOS arm64 binary passed the Studio smoke.
- External decision providers: native Laya and OpenAI-compatible JSON chat, with optional knowledge reranking and navigation/planning/review/repair guidance. See configuration, boundaries and fallbacks. Hosting and training remain external; no Python model scripts or weights are bundled.
configure --jsonnow persists before claiming success and returns nonzero for failed discovery. Regression tests cover both behaviors.- Live release checks exercise actual chat completion and JSON prediction. The squads check now rejects failed, unfinished and timed-out deliveries; structural team checks alone no longer produce a false green result.
- Vulnerability scanner installation works when Go's binary directory is outside
PATH. Dependency freshness includes the npm lockfile. - Studio's router and affected transitive dependencies are patched. Functional tests use bounded workers and scheduling headroom without removing assertions.
- A static frontend team supplies framework-free Node behavior tests. Planner and splitter prompts include the frozen API contract and required acceptance commands, so earlier architecture suggestions cannot silently replace its routes.
- Studio shows board-based delivery progress, delayed-connection feedback, readable mobile team shortcuts and a persistent Motion/Still setting. Phase scrolling stays inside its strip. Live insights no longer warn that an ongoing run lacks its final event.
Local checks ran on macOS arm64 with Go 1.27.0 and Node 24.21.0. Release CI ran on Linux with the module-selected Go toolchain and Node 22. The workflow stamps the final version, source commit and build time into every artifact.
| Check | Result |
|---|---|
Complete local gate (GOFLAGS=-p=4 make -j2 check) |
Passed; 72.2% coverage |
| Merged-tree CI and release gate | Passed, including package/integration race tests, lint, coverage and frontend checks; release-run coverage 71.2% against the 63% floor; 365 Studio tests |
| Studio browser checks | Passed on desktop and phone: layout, task navigation, motion preference persistence and 3D rendering; no page errors |
| Published macOS arm64 Studio | Downloaded release binary: all ten main pages navigate, authenticated API works and unauthenticated config access is denied; no browser exceptions or server errors |
| CLI configuration regression tests | Passed |
| Live delivery qualification guard regression tests | Passed; failed, unfinished, missing and timed-out results cannot qualify |
| Strict documentation build | Passed; internal broken links now fail the build |
| Documentation inventory | All 39 public help entries covered; runtime schema exposes 126 configuration fields |
| Version consistency and repository-reference checks | Passed |
| Go vulnerability scan | No known vulnerabilities found |
| Full npm dependency audit, including development dependencies | Zero reported vulnerabilities after updates |
| Native Laya Go-client smoke | Passed against upstream laya-serve 0.3.20, multilingual, CPU; temporary server stopped afterward |
| Live discovery, configure, chat and OpenAI-compatible decision smoke | Passed against oMLX with Qwen3-Coder-30B-A3B-Instruct-MLX-4bit |
| Release binaries | Six published macOS/Linux/Windows arm64/amd64 binaries; real Studio assets embedded; all downloaded checksums verified |
| Distribution metadata | All three installer checksums, four Homebrew binary hashes and both offline macOS bundles match the release |
| Binary identity | Linux amd64 smoke passed in CI; downloaded macOS arm64 reports v0.28.0, commit fb8cddc, build time and empty SourceRoot |
| Generated application, independently verified | Passed: Go race tests, current acceptance guard, real browser loading/refresh, 24 concurrent increments, HTTP/network failure states and no page exceptions |
| Live two-team application delivery | Passed: 3/3 tasks done, zero failed or unexecuted; correct frontend-static/backend-go teams; 771 seconds including setup |
Cross-compilation is not execution on all six platforms. Numeric model-response smokes establish protocol compatibility, not predictive accuracy or a speedup. The decision feature remains opt-in until evaluated on representative projects. Go lint reports zero findings. Studio lint passes with 41 existing warnings and zero errors; the warnings remain cleanup work.
The live suite ran before the final comment-only Node-file guard. Its retained application then passed that newer guard independently. The generated JavaScript checks validate source structure; the separate real-browser checks provide the behavioral evidence listed above. Upstream integration changed Studio and release metadata, and the merged Studio was rebuilt and checked again in Chromium.
The live scenario asked for a Go counter endpoint and a plain HTML/JavaScript
client, with no npm or framework. It activated backend-go and frontend-react
teams. Implementation tasks completed, but integration verification did not:
the run reported tester/QA rejection and an unresolved defect after about
25 minutes. The old test printed PASS because it only checked structural
invariants; its underlying success=false result is the qualification evidence.
The test has now been corrected so that outcome cannot be accepted.
The team-library mismatch is now addressed by frontend-static, whose checks
use Node's built-in test runner without npm. Follow-up inspection also found
that planning omitted the frozen contract and invented /counter while the
manager had specified /api/counter. Both planner and splitter now receive
that contract and its acceptance commands. Regression tests cover selection,
contract propagation, token bounds and the Node test runner's shell restrictions.
Follow-up runs also exposed a separate acceptance allowlist that rejected Node tests, a Node 24 zero-test exit that looked successful, and Go-only prompts on JavaScript tasks. Regression tests now exercise real passing, failing and absent Node suites, guard against unsafe flags, and keep task-specific verification consistent across workers, reviewers and correctors. The static team uses standard-library browser fakes instead of uninstalled third-party dependencies.
Further live verification exposed worker-only finalization instructions in the underlying agent library. SLMCode now replaces that exact generated tail with the active role schema before sending it to the provider. A regression test executes the real ReAct dependency and inspects the outgoing finalization request. Shared recovery/budget prompts also retain role contracts; tester success is not synthesized from file writes. Review and correction preserve the team charter, and a new dispatch can read files even if its predecessor exhausted its repetition guard.
One earlier attempt lost its local oMLX endpoint mid-run; it was restored and that interrupted attempt was not counted as passing. The successful rerun used the strict delivery guard, which requires the underlying result to report success with no failed or unfinished tasks.
To repeat qualification against a deployment's actual provider/checkpoint:
SLMCODE_E2E_ARTIFACTS=/tmp/slm-release make e2e-releaseDo not infer success from an earlier legacy PASS label, skip integration, or treat an unavailable check as passing.
The feature index maps public command groups and harness features to their guides: CLI/TUI/Studio, tasks and sessions, teams and pipelines, agents and skills, blocks/stacks/hooks/MCP, permissions, context and retrieval, memory, evolution, metrics, graph, calibration, evaluation, autoresearch and providers. Migration notes cover the CLI JSON behavior and updated source-build prerequisites. Maintainers should follow the release checklist.