Changelog

All notable changes to this project are documented in this file.

The format follows Keep a Changelog, and this project uses Semantic Versioning.

[Unreleased]

[1.1.0] - 2026-08-27

Added

• Strict mypy and branch-aware coverage quality gates for source and maintainer scripts, including direct CLI-wrapper and changed-line regression tests.
• CPython 3.14 hosted coverage and a stable event-aware CI Quality gate across eight supported endpoint and intermediate lanes.
• A resumable draft-first GitHub Release publisher that accepts exactly two ZIPs, two checksums, and two SPDX 2.3 SBOMs and verifies published assets with bounded retries.

Changed

• The documented portable runtime boundary now covers CPython 3.11–3.14.
• Tagged publication depends on the stable summary gate, uses Python 3.14, and treats an existing published Release as read-only.

Fixed

• Legacy Review migration no longer deletes a pre-existing user-owned temporary validation artifact when a migration is refused.

[1.0.4] - 2026-08-10

Changed

• Release provenance, Codex package SBOM, and Agent Plugins package SBOM attestations now use the verified actions/attest 4.2.2 commit pin.

[1.0.3] - 2026-08-09

Added

• Extracted Knowledge Selection smoke tests for both the native Codex archive and the portable Agent Plugins archive.
• Per-run facts, Profile, Knowledge Selection, and compact-context paths for AI-agent, mobile, and solution-decision workflows.
• A canonical repository roadmap and a time-bound host compatibility evidence report for release validation.
• Two-model, three-trial-per-case 1.0.3 behavior evidence with strict source, runtime, execution-log, and fixture provenance.

Changed

• Tag publication now runs only after the complete Linux, macOS, and Windows quality matrix succeeds for the same workflow and commit.
• Remediation routing cases now require an accepted Architecture Decision and preserve the separate authority for solution selection.
• The full Solution Advisor benchmark treatment now supplies the declared decision styles, guides, and case studies required by its decision corpus.

Fixed

• ZIP smoke validation now checks original entry names before Windows path normalization and rejects Windows drive-relative or drive-absolute entries.
• Specialized audits now bootstrap missing project governance only when persistence is allowed and preserve read-only requests without mutation.

[1.0.2] - 2026-08-08

Added

• Versioned installation and invocation instructions for Codex and ChatGPT desktop, Cursor, VS Code and GitHub Copilot, and Kiro in both English and Simplified Chinese READMEs.

Changed

• The compatibility boundary now distinguishes full Agent Plugins installation from Kiro's complete Agent Skills and shared-resources projection, and names the host-specific controls that remain outside Hengmu's portable core.

[1.0.1] - 2026-08-08

Added

• A host-compatibility matrix and accepted cross-host equivalence decision that distinguish portable workflow outcomes from client-specific lifecycle and permission controls.
• An extracted Agent Plugins archive smoke test that runs the primary audit path through deterministic Knowledge Selection in CI and release builds.

Changed

• The Agent Plugins discovery manifest is now checked in at repository root, host-neutral, and identity-aligned with the native Codex manifest.

Fixed

• Portable archives now retain the native manifest as inert provenance data, preventing Knowledge Selection from failing after extraction while preserving existing selector hashes and artifact contracts.

[1.0.0] - 2026-08-06

Added

• Constrained Greenfield design through the existing architecture-solution-advisor: Brief 1.1 records required, preferred, and prohibited constraints, and Decision 1.4 records their assessments and the complete target architecture.
• Greenfield target planning in Plan 1.3, bound directly to the accepted Brief and Decision with no fabricated Findings, mapped to runtime/deployment units, data ownership, interfaces, trust boundaries, critical flows, operations, and constraints.
• Routing and evaluation coverage for open/constrained design, specification, constraint conflicts, Chinese equivalents, target-architecture completeness, and Greenfield plan source validity.
• Compatibility and migration guidance for the one coordinated 1.0 artifact release, plus an accepted constrained-target architecture ADR.

Changed

• Public positioning now presents Hengmu as an evidence-bound architecture design, decision, and governance system with two first-class entry paths: current-state assessment and open/constrained target design. The README, website, plugin metadata, workflow metadata, banners, governance diagrams, and editorial figures use the same model.
• The stable hengmu router routes design, specify, constrain, and their Chinese equivalents to the existing Solution Advisor. The public surface remains one router plus eight focused workflow Skills.
• Required constraints are challenged before becoming hard requirements; preferred constraints may lose with a recorded trade-off; prohibited options are hard-eliminated. Constraint inputs never substitute for proof.
• Technology-evolution remains a narrow evidence lens. Version pins require current official or repository evidence and are never supplied from memory.

Compatibility

• Brief 1.0, Decision artifacts through 1.3, and Plan artifacts through 1.2 remain readable. Brief 1.1, Decision 1.4, and Plan 1.3 ship together as the current 1.0 open/constrained target-design contract.

Other work included in 1.0.0

Added

• A stable $hengmu entry point that shows the full capability menu, accepts natural-language goals, and routes to one of eight focused workflow Skills without requiring users to remember their public names.
• An idempotent prepare-project-audit command that creates a facts-derived .architecture/ control plane when missing, validates existing governance, and refuses to overwrite partial user-owned state.
• Localized 青野/Hengmu icons and banners, an original 青野 editorial character, three paired English/Chinese article illustrations including the target-design hero, and paired Mermaid/Excalidraw/SVG/PNG governance flows.
• A full Simplified Chinese README that mirrors the English quick start, trust model, workflows, development commands, and project policies.
• Consumption-time validate-knowledge-context checks for the Selection lock, result hash, and exact ordered selected-entry projection.
• CI and release ancestry checks for the Selector source and latest reviewed implementation commits.

Changed

• The public surface now distinguishes one discoverability-only entry Skill from eight directly invocable workflow contracts; audit, verification, decision, planning, and gate authority remain separate.
• Dependabot now groups all Python version updates into one PR and caps routine Python and GitHub Actions update queues at two and one open PR respectively.
• GitHub workflows now use the pinned Node 24 releases of Checkout 7.0.1, Setup Python 7.0.0, and Dependency Review 5.0.0.
• project-architecture-audit now treats missing governance as a bootstrap condition and persists audits by default; Advisory mode requires an explicitly read-only request.
• The project, repository, installable plugin ID, release archive prefix, SBOM, and public tooling identity are now Hengmu. Pre-rename self-review artifacts retain their recorded identity as immutable internal evidence.
• The development lock now requires pytest 9.1.1 or newer within the 9.x series, removing the vulnerable 8.4.2 test-runner dependency.
• Runtime dependency bounds now require JSON Schema 4.26.0 and PyYAML 6.0.3, matching the verified lock files already used by installation and CI.
• Git-verified archived Selection locks remain historically readable but can no longer create new trusted Review, Decision, Plan, coverage, or Gate chains; current deterministic replay is required.

Fixed

• Profile construction now resolves facts stored at any depth under .architecture/ from the repository root and records portable relative bindings instead of treating a per-run reviews/inputs/ directory as the project.

[0.4.2] - 2026-07-29

Added

• Repository-facts schema 1.1 roles, role-aware Profile/Knowledge routing, and regression coverage so tests, examples, documentation, generated code, vendor trees, and benchmark fixtures cannot infer product architecture.
• Knowledge-selection schema 1.4 kind/maturity bindings, per-kind budgets, a source-anchored complete Selector Runtime Input Manifest, distinct project and plugin commits, current-runtime replay, safe archived Git-blob verification, and rejection of unverifiable locks from trusted chains.
• Compact model-facing Knowledge context sidecars that omit the full exclusion ledger while binding its machine-facing selection lock by SHA-256.
• Decision-intent namespaces for data-authority and plugin-runtime topology, preventing ambiguous local-first wording from loading unrelated client state guidance.
• A schema 1.5 Base/Full/Compressed behavior-benchmark treatment manifest, compact Skill assets, corpus-level declared-context proxy, and condition/ manifest-bound execution provenance.
• Optional informational governance-run manifests and validation for high-risk Governed/Enforced trajectory records. They are deliberately excluded from the trusted Gate evidence chain.
• Advisory, Governed, and Enforced operating-mode guidance, plus a 0.4.2 migration guide.
• A failure-driven Knowledge curation policy: new Golden knowledge must trace to a concrete decision-quality gap and add a representative regression.

Changed

• The Solution Advisor Skill now keeps behavioral rules in its lean entry point and moves artifact-command detail into a focused reference.
• Gate policy schema 1.2 declares product_mode; its label cannot weaken or bypass deterministic Gate outcomes.
• Benchmark scoring reports tool-call telemetry when an invoked surface emits it, and distinguishes missing field-level usage telemetry from zero.

Fixed

• Historical Knowledge selections are no longer re-evaluated with a future Selector or Knowledge tree; exact deterministic replay occurs only when the recorded creation runtime still matches.
• Python dependency inspection now parses normalized PEP 508 package names and applies exact aliases, avoiding nextcloud-client, agentscope, and pgvector substring false positives.

[0.4.0] - 2026-07-29

Added

• Twenty-one curated golden knowledge entries with named options, concrete operating mechanisms, failure and exit semantics, claim-to-source mappings, and maintainer curation provenance.
• A Greenfield Design Brief contract, template, validator, decision bindings, and Architecture Decision schema 1.3 path that does not manufacture a source review.
• Solution-decision benchmark observations and scoring for recommendation accuracy, over-design, trade-off coverage, knowledge citation validity, rejected-option explanations, migration actionability, and stability.
• A read-only Codex benchmark adapter with machine Rule IDs, canonical decision vocabulary, structured-output validation, and one bounded evidence-only correction that never receives ground truth.
• Benchmark schema 1.4 provenance that binds source, environment, dependencies, configuration, plugin manifest and Skill version, fixtures, runner and adapter bytes, reconstructible command templates, exact trial commands, external runtime fingerprints, and hash-verified execution logs.
• A Git-bound archived benchmark verification mode for cross-platform reviewers; strict current-runtime replay remains the default.
• Severity- and stage-aware verification policy: critical V3, high V2, medium/low V1, and V4 for risk acceptance and release.

Changed

• Generated knowledge now defaults to draft and cannot become active without a non-generated curation record.
• Knowledge selection uses canonical domain IDs, includes plugin-platform, labels required/recommended/optional context, performs bounded one-hop relation expansion, and downweights generic reference-architecture matches.
• Knowledge validation rejects golden entries that are template-similar, omit named option trade-offs, or leave claims without authoritative sources.
• Backend API and web frontend domain guidance now cites relevant HTTP, OpenAPI, WCAG, and web-performance sources.
• Adversarial fixture paths and prose are outcome-neutral, with a regression test preventing expected decisions from leaking into model-visible inputs.
• Benchmark scores distinguish absent usage telemetry from actual zero token/cost consumption.
• Base-commit gates now require re-review only when a classified critical or security path changed after the reviewed commit; later governance-only records no longer create an impossible review/HEAD self-reference.
• An accepted keep-current decision with explicit migration slices, rollback, validation, and exact affected-path coverage now satisfies compatible migration governance without manufacturing a remediation plan for a non-risk Finding.

Security

• Greenfield decisions bind the exact Design Brief and knowledge-selection bytes; remediation decisions retain verified-review provenance.
• Persistent risk acceptance and release evidence can no longer rely on the same global V1 verification floor as low-risk findings.

[0.3.2] - 2026-07-29

Fixed

• Made the benchmark command-rendering assertion compare against the native path representation so the cross-platform safety test passes on Windows without weakening its argument-boundary guarantee.

[0.3.1] - 2026-07-29

Fixed

• Added the Windows-only colorama dependency to the exact development lock so hash-enforced CI installation works on every supported runner.
• Bound SBOM generation, attestation, and release upload to one exact artifact path instead of passing an unexpanded glob to actions/attest.

[0.3.0] - 2026-07-29

Added

• Deterministic repository-facts inspection, provisional Profile construction, task-scoped knowledge selection, coverage validation, artifact fingerprinting, and safe legacy Review migration commands.
• Review, Finding, Architecture Decision, and Remediation Plan schema 1.2 bindings for facts, selected knowledge, critical flows, evidence fingerprints, finding fingerprints, assumptions, and migration slices.
• Ten Markdown/frontmatter Knowledge Packs containing 205 validated entries: foundations, domains, decision guides, architecture styles, patterns, technology profiles, reference architectures, migration guides, anti-patterns, and case studies.
• Knowledge manifest and entry schemas, relationship validation, source policy, freshness windows, stale-entry rejection, and explicit selection reasons and exclusions.
• Dedicated routing, knowledge-selection, decision-quality, false-positive, and artifact-validity evaluation corpora.
• Target-architecture, knowledge-authoring, 0.3 migration, and implementation documentation plus an accepted decision for the workflow/knowledge/script separation.

Changed

• The public surface now contains exactly eight workflow Skills. Knowledge curation moved to maintainer/skills/ because it is a release-maintenance role rather than an end-user architecture workflow.
• New project initialization records deterministic repository facts and keeps detected, declared, and inferred Profile inputs separate.
• New audits and decisions use schema 1.2; schema 1.0 and 1.1 remain readable and trusted 1.1 artifacts retain their 0.2 compatibility path.
• Architecture knowledge is selected per repository, task, and Skill instead of loading every bundled catalog into context.
• Plugin and portable CLI version are now 0.3.0.

Security

• Legacy verified Reviews are migrated only as candidates. Migration cannot synthesize independent verification, critical-flow coverage, or current trust.
• Fact, Profile, selection, knowledge-entry, Finding, Review, Decision, Plan, and completion-evidence hashes are checked at their owning boundaries.

[0.2.0] - 2026-07-28

Added

• Architecture Solution Advisor and Architecture Knowledge Curator Skills.
• Eight machine-readable quality, style, pattern, technology, reference architecture, migration, domain, and decision-guide catalogs with 128 sourced entries.
• Architecture Decision, Risk Acceptance, Rule Pack, Knowledge, Evidence Provider, and Benchmark schemas.
• Verified-review provenance bindings, verification levels, Finding fingerprints, Git evidence resolution, repository path containment, and machine-complete Rule Pack coverage.
• Contract, Finding, Change, and Release gate stages with exact, ancestor, and diff-aware freshness.
• Nineteen bundled core/domain Rule Packs plus validated repository-local organization Rule Packs.
• Eleven executable Evidence Provider adapters with shell-free invocation, executable/config/output hashing, timeout, safe environment propagation, and JSON/SARIF/JUnit structural validation.
• Role separation, human V3–V5 assurance requirements, deterministic evidence for V4–V5, and detached SSH review signatures at V5.
• Knowledge-bound three-option architecture decisions, hashed plan completion evidence, required-review enforcement, base-commit change classification, and review diffing.
• Ten adversarial architecture fixtures, repeated-trial stability, duration and optional usage metrics, deterministic benchmark scoring, and a caller-supplied forward-test harness.
• Fingerprint-bound baselines and waivers plus separately authorized, expiring risk acceptance.
• Fully hashed runtime/development locks, cross-platform checksum verification, license allow/deny audit, license-complete SPDX SBOM generation, dependency audit, SARIF output, and GitHub provenance/SBOM release attestations.

Changed

• Trusted enforcement now requires artifact schema 1.1; schema 1.0 remains readable for migration.
• Remediation planning consumes an accepted architecture decision instead of selecting technology and target architecture itself.
• CI now covers Python 3.11 and 3.13 on Ubuntu, macOS, and Windows.
• Plugin and portable CLI version are now 0.2.0.

[0.1.0] - 2026-07-28

Added

• Seven focused architecture-governance Skills for project, AI-agent, mobile, portfolio, verification, remediation-planning, and quality-gate workflows.
• Versioned review, finding, profile, portfolio, remediation, baseline, and policy schemas.
• A portable CLI for initialization, validation, and deterministic policy evaluation.
• Repository validation, activation eval corpus, deterministic plugin packaging, CI, release automation, and open-source governance documents.
• A dogfooded project architecture profile with explicit constraints and critical flows.