Milestone 0.6.4 release

Source: core repository guide, snapshot e58778bef40b. Synced 2026-10-24. Package versions are independent of milestone labels.

Status on 29 September 2026: published and verified on npm and GitHub. The manifest records this milestone’s independent package versions, and the publication record retains registry integrity, remote-tag and consumer evidence.

Delivered behavior

akit workflow report and synchronous SDK/report-package APIs project saved public V1/V2/V3 workflow evidence into one canonical report model. Technical, executive and comprehensive views export self-contained HTML or Markdown without provider configuration, network access, Docker, checkpoint restoration or a new evaluation. Comprehensive HTML is the default.

The assessment shows scope, methodology, coverage and exclusions, criterion-grounded findings and next actions, independent runtime/task/policy status, target/judge usage, unavailable cost, configuration identities, evidence omissions, cleanup, fault/retry/state-change and recovery limits. Every finding links to stable JSON-pointer/digest evidence. All views retain uncertainty and the technical evidence appendix. HTML uses semantic sections, labeled tables, keyboard navigation, responsive layouts and print styles; it contains no scripts or external assets.

Only eligible passed/failed final logical task outcomes enter success rates. Historical V1, preflight-only, invalid and unavailable evidence remains visible outside that denominator. Identical records count once. Compatible V3 pause/resume attempts preserve cumulative snapshots without treating each attempt as another independent task. Ambiguous or conflicting collections refuse aggregation. Empty eligible coverage produces an unavailable rate.

CLI input is bounded to 50 regular, non-symlink files, 1 MiB each and 8 MiB total. Output files must be new paths and use mode 0600. Exit 0 means successful export even when tasks failed; invalid evidence exits 2 and input/output failures exit 1. SDK inputs accept saved objects or JSON strings and arrays. Legacy prompt-response reporting stays unchanged. See the guide for examples.

Verification

Independent reviews accepted the canonical model, both renderers and CLI/SDK integration. Checks include actual-engine multi-attempt resume with preflight, truncated trace evidence, exact pointer hashes, hostile getters/proxies and injection payloads, no-overwrite outputs, FIFOs, invalid UTF8, symlinks and bounded input. No provider inference is required by reporting qualification.

Source qualification passed 51 report/surface tests with 2,111 assertions. The final tracked-source suite bun test packages scripts examples passed 1,732 tests / 7,877 assertions / 107 files. Whole-checkout bun test also passed 1,767 tests / 7,981 assertions including 35 ignored local tests outside the release. All 13 package builds and typechecks passed; lint had 0 errors and 30 established SDK warnings. Fresh final tarball consumers passed Node/Bun imports, report CLI/SDK parity in all 6 view/format combinations, historical/current evidence, TypeScript NodeNext/bundler and actual Docker execution/recovery regressions.

An intermittent five-second deadline affected the existing two-process publisher-shell test. The test now bounds each subprocess to five seconds and the combined test to fifteen seconds; all cleanup, source-preservation and credential-safety assertions remain. No production publisher timeout or CI configuration changed. An unchanged five-millisecond semantic late-usage test failed once in a later full run; its direct 35-test file and the unchanged final 1,732-test suite then passed. That timing sensitivity remains a recorded test limitation.

All 3 HTML views were inspected at 1440 and 375px widths with no document overflow or external requests. Keyboard skip/evidence navigation, every local link, Markdown parsed anchors and print exports passed. This is bounded browser qualification, not a comprehensive assistive-technology audit. The validation record retains the final versioned consumer results, package SHA512 hashes, exact checks and browser evidence. All 13 committed-source repacks match those qualified archives. The final versioned full suite/typecheck/lint passed; an initial formatting-only error in the new validation JSON was corrected with the established formatter.

Boundaries

Reports assess selected saved evidence. Free-form display labels and operation identifiers are hashed; private prose, file contents and checkpoint paths are excluded. Missing scenario definitions, exact harness version, full policy declarations, wider workload coverage and pricing are explicitly unavailable. Digests are identifiers, not authenticity signatures. No task rescoring, inferred readiness, independent statistical repetitions, model ranking or certification is claimed.

Supported qualification is macOS arm64, Node 24.18.0 and Bun 1.3.10. CLI retains its Bun runtime; Node supports the SDK/report package. Other operating systems, browser engines and third-party Markdown renderers are not runtime-qualified here. No external model-quality claim, paid call, optional harness integration, Loki feature or CI repair is included.

Package map and publication

Reports advances to 0.5.0; CLI and SDK to 0.6.3; core takes a release-alignment patch to 0.6.4 without changing saved workflow schemas. Necessary adapter/redteam dependency patches are listed in the manifest. MCP Docker 0.1.0 is unchanged. The SDK adds only the internal reports package dependency. The CLI now explicitly includes runtime source/bin/dist files and excludes local run artifacts. The publisher and tarball consumer verifier reject known runtime, checkpoint and environment-file paths. This corrects an ignored redaction-test record discovered in the prior CLI archive; published versions and local records remain unchanged. Fresh consumer checks and archive byte parity passed for the corrected candidate; all 13 committed-root repacks match the recorded final hashes.

All twelve package versions are published on npm. The exact registry release gate and fresh registry-installed Node/Bun/TypeScript/CLI/SDK/Docker consumers passed. The twelve annotated package tags and v0.6.4 resolve to the immutable candidate, and the GitHub release is published. Published versions and remote tags were not replaced or moved.

The next roadmap milestone is 0.7.0 comparative experiment orchestration. This report collection API does not establish cross-configuration comparison eligibility or execute repeated experiments.