v0.12.0

Added

• Added a dependency-free, root-bounded static-PNG and animated-GIF exporter for trusted repository-owned README logo recipes, with non-exporting contract checks, strict XML and artifact inspection, commit-time no-clobber protection, exact per-frame GIF timing validation, sanitized failures, deterministic fake-tool coverage, and transactional replacement.
• Added explicit target-contract and install-host data to the generated catalog so contract-backed cross-host installs are discoverable without changing runtime-specific artifacts.
• Added focused Claude and Cursor lifecycle evals with competing interviewer skills and repository-owned persistence checks.

Changed

• Updated animated-readme-logo to 0.4.0 so product repositories retain only brand-specific source, motion, timing, palette, and path decisions while the skill owns reusable export mechanics.
• Updated drawio-diagrams to 0.4.0 with default native animation for new directed runtime, process, and data flows, explicit opt-out, a technical-geominimalist default plus four bounded adaptive design profiles, bounded reference-style adaptation, architecture content gating, compact ER/UML/C4/BPMN/SysML/ML notation recipes, icon-first logo/service coverage, Lobe Icons and Simple Icons provider routing, a single user-responsibility rights notice, improved local shape search, conditional discovery, bounded three-pass self-review, and a fair head-to-head benchmark protocol.
• Updated claude-spec-interviewer and cursor-spec-interviewer to 0.2.2 so planning, question, transition, and plan-exit instructions follow the execution host while target-specific evidence and execution output remain unchanged.
• Updated incubator skillopt-setup to 0.1.1, keeping the Codex OpenAI-compatible gateway local to that workflow until reuse and isolation justify extraction.
• Updated incubator handoff to 0.1.1 with agent-neutral trigger wording.

Fixed

• Superseded ADR-0026 with ADR-0028 so both proven reuse and fail-closed backend isolation are required before a gateway can be extracted from its owning skill.
• Removed natural-language directive classification from validation; OpenAI prompts now use a small fail-closed foreign-control-name check.
• Made complete skill frontmatter validation YAML-aware, including nested metadata, comments, quoted keys, and flow mappings.
• Parsed agents/openai.yaml as YAML before validating nested types and checking the resolved default prompt.
• Separated target runtime from installation host throughout catalog labels and commands, synchronized the documented host-ready install sets with the portable public catalog, and required every generated install command to have a host-specific accessible copy label.
• Hardened drawio-diagrams export verification so exit-zero Desktop runs cannot publish missing, stale, symlinked, structurally invalid, or concurrently changed artifacts; both outputs validate before commit, destination installation is no-clobber, and interrupted or replacement commits preserve recovery backups instead of deleting raced data. Portable-image validation now checks every HTML/CSS/srcset source, rejects missing and fragment-only image cells plus obfuscated active schemes, ignores navigation links, and stays in parity with SkillOpt. Also removed unsupported Desktop Mermaid-import and --layout assumptions, corrected eval regexes, and added guarded implicit-activation coverage.
• Hardened animated-readme-logo export with descriptor-bound output-parent creation, cumulative frame-byte limits, two complete deterministic render passes with verified-byte reuse, digest-bound multi-output commits, and retained replacement recovery that cannot delete same-inode concurrent updates.
• Fixed SkillOpt Codex CLI target, judge, and reflector prompt transport to use explicit UTF-8 stdin, redact prompt echoes from captured streams, and reap large no-read timeouts, so large skill/resource snapshots no longer exceed argument limits or leak through process arguments and logs.

Deprecated

Removed

Security

• Kept fail-closed filesystem, process, tool, network, and environment isolation mandatory for every extracted or shared backend gateway.

v0.11.0

Added

Changed

• Updated animated-readme-logo to 0.3.0 with an approval-gated minimal exporter preflight and a configured Chromium/agent-browser preview fallback before any browser download.

Fixed

• Separated browser-preview failures from raster-export readiness and provider approval from local-tool installation approval.

Deprecated

Removed

Security

v0.10.1

Added

• Added ADR-0026 guidance that separates the executing client from the target agent ecosystem, keeps independently valuable skills unconstrained by catalog size, and leaves backend gateways with their owning workflow until reuse and isolation are proven.
• Added real disposable Codex, Cursor, and Claude Code destination-install smoke coverage plus live Claude- and Cursor-target-on-Codex activation evidence.

Changed

• Updated claude-spec-interviewer to 0.2.1 so cross-host execution adapts planning and question controls without changing Claude Code evidence or output artifacts.
• Updated cursor-spec-interviewer to 0.2.1 with execution-host control translation and a matching OpenAI default prompt, and updated architecture-compass to 0.2.1 with valid concise OpenAI metadata.
• Made catalog install commands target-correct by category and replaced manual Claude skill copies with supported -a claude-code installation commands.

Fixed

• Validated every existing agents/openai.yaml file and rejected OpenAI prompts that require foreign-host-only planning or question controls.

Deprecated

Removed

Security

v0.10.0

Added

• Added once-per-task stable update checks for the selected CodeGraph/ast-grep analysis stack, with one itemized per-tool approval checkpoint, offline and opt-out handling, versioned actions, verification, and rollback disclosure.
• Added capability guidance for current and legacy CodeGraph releases, a focused optional-tool escalation guide, and expanded eval scenarios for update consent, runtime boundaries, legacy behavior, native LSP fallback, advanced codemods, security-policy routing, and bounded rewrites.
• Added deterministic codegraph-ast-grep contract validation, command-safety regression fixtures, and machine-regraded captured behavioral evals to the repository validation pipeline.

Changed

• Updated codegraph-ast-grep to 0.2.0 with installed-help-first command selection, consolidated semantic exploration when exposed, watcher-aware graph freshness, capability-gated ast-grep outline, evidence reconciliation, and reviewed rewrite staging.
• Reworked setup for exact package or checksum-verified release provenance, reproducible project-local launches, effective CODEGRAPH_DIR handling, and explicit Codex, Cursor, Claude Code, and generic MCP configuration boundaries.
• Expanded Claude Code install documentation to include the portable codegraph-ast-grep workflow.

Fixed

• Removed unconditional assumptions about codegraph_trace, CLI explore, codegraph init -i, universal auto-sync, npm-only CodeGraph installation, and unpinned persistent ast-grep MCP execution.
• Separated analysis-tool binary/package updates from runtime configuration refresh, prompt hooks, telemetry choices, graph operations, and unrelated application dependency updates.
• Made project-local ast-grep verification fail closed instead of allowing npm to fetch a missing command, and required affirmative approval before project-opening CodeGraph diagnostics that may migrate generated state.
• Preserved Team-pinned CodeGraph update scope and made native Windows checks, updates, and verification restore prior process environment state after telemetry/config guards.

Deprecated

Removed

Security

• Prevented automatic or blanket tool updates, pipe-to-execution installer defaults, unreviewed optional-tool installs, and rewrite snapshot/update-all shortcuts.
• Required stable-source provenance, exact update scope, checksum/build-policy disclosure, redaction, telemetry-disabled update checks unless separately approved, and explicit approval for every side-effect class.

v0.9.1

Added

• Added a concise public SkillOpt run summary for the accepted drawio-diagrams routing and readability optimization.

Changed

• Updated drawio-diagrams to require explicit orthogonal waypoints around text and callout obstacles and to report validator names, connector rails, label backgrounds, and label spacing for dense diagrams.

Fixed

Deprecated

Removed

Security

v0.9.0

Added

• Promoted animated-readme-logo 0.2.0 with live Recraft V4.1 discovery and cost approval, a deterministic local SVG fallback, explicit task/provider/export statuses, and README-safe delivery guidance.
• Added dependency-free strict SVG and GIF/APNG/WebP inspection with hidden-metadata rejection, a focused animated-logo fixture harness, and reusable bounded visual assertions shared with draw.io evaluation.

Changed

• Kept animated-logo routing portable across Codex, Cursor, Claude, and other Agent Skills hosts; provider capability and output contracts drive the route instead of agent-specific forks.
• Refactored draw.io visual evaluation behind a reusable six-assertion library without changing its public grammar.

Fixed

• Made README logo audits resolve assets inside an explicit repository root, parse bounded live Markdown/HTML image references, verify SVG and raster content, enforce meaningful reduced-motion/static fallbacks, and report readiness without fabricating unavailable artifacts.

Deprecated

Removed

Security

• Rejected absolute, UNC, root-escaping traversal, encoded scheme/entity, duplicate-attribute, and symlink-escaped README asset paths before reading files.
• Rejected symlinked or oversized visual-artifact trees, active/foreign/metadata-bearing SVG content, hidden raster metadata, and non-regular animated-image inputs before release proof.

v0.8.0

Added

• Added capability-detected Codex, Cursor, Claude, and unknown-host collaboration adapters to the portable architecture-compass workflow.
• Added Architecture Compass lifecycle evaluation cases for conditional planning, ADR conflicts, stack deviations, fallbacks, no-write decision phases, bounded execution, audits, and PR reviews.

Changed

• Updated architecture-compass to use a read-only planning phase for unresolved durable choices or broad, multi-boundary, behavior-changing, or phased refactors while keeping narrow behavior-preserving ADR-backed work direct, audits read-only, and diff reviews on the host review surface.
• Added explicit planning, read-only-enforcement, architecture-decision, and execution fields; a pending-write-permission gate; exact route-matching continuations; and repository-state rechecks before approved changes are applied.
• Documented portable Architecture Compass installation and usage for Codex, Cursor, and Claude.

Fixed

• Replaced the release workflow's cross-runtime wildcard verification command with the explicit Codex-ready public skill list.
• Fixed SkillOpt split generation to preserve wrapped expected-behavior bullets instead of truncating semantic judge requirements.
• Corrected Cursor Plan evidence classification so behavioral no-write instructions are not reported as an enforced filesystem sandbox.

Deprecated

Removed

Security

v0.7.0

Added

• Added the incubating skillopt-setup workflow with local Microsoft SkillOpt installation and readiness checks, per-skill train/validation/test preparation, provider-backed, hybrid, and Codex CLI execution profiles, run artifact verification and summaries, and guarded best_skill.md adoption.
• Added a local .agents/ artifact audit and a loopback-only Codex-backed OpenAI Chat Completions gateway with an end-to-end compatibility probe for SkillOpt experiments.
• Added deterministic draw.io visual-eval support with six PNG/SVG artifact cases, fixture-backed assertions, generated full and text-only split variants, page-index export, DRAWIO_BIN overrides, and diagrams.net executable discovery.
• Added runtime-specific native Plan-mode lifecycle and fallback eval coverage for the Codex, Cursor, and Claude spec interviewers.
• Added drawio-diagrams architecture-readability guidance and eval coverage for connector-label gutters, fan-out lanes, hierarchy, whitespace, dark-mode labels, and logo fidelity.
• Added repository validation for SkillOpt helper and adapter contracts and draw.io visual assertions.

Changed

• Updated all three public spec interviewers to require native Plan mode when supported, route transitions through each host's native controls, use structured clarification when available, and record conversational fallback only when Plan mode is unavailable or explicitly declined.
• Made Plan-mode interviewing read-only and moved approved spec and required ADR persistence to a save-only continuation after mode exit that stops before feature implementation.
• Strengthened SkillOpt setup and readiness reporting with strict training-ready checks, target/mode/profile-specific fresh manifests, active-split data floors, explicit model and credential blockers, and visual renderer readiness.
• Expanded drawio-diagrams XML authoring, routing, theming, icon, and verification guidance to improve dense architecture readability and preserve real logo artwork.
• Updated public usage examples for each runtime's native Plan-mode lifecycle.

Fixed

• Fixed save-only spec-interviewer persistence so a required ADR also receives the minimal index entry mandated by the repository's existing convention, while all other repo-facing documentation remains deferred to implementation.
• Fixed drawio-diagrams connector-crossing validation to honor explicit waypoints and side ports and calculate nested cell bounds before warning about callout overlaps.
• Fixed draw.io visual assertion globs so Markdown escapes such as \*.png and \*.svg retain wildcard semantics in both JavaScript and Python matchers.
• Fixed SkillOpt readiness to score the configured active split, ignore None visual sentinels, reject missing configured split directories, and block stale or mismatched adapter manifests.
• Fixed generated SkillOpt case metadata to distinguish text-only execution from isolated-artifact-write visual execution instead of advertising legacy broad workspace writes.

Deprecated

Removed

Security

• Isolated every Codex-backed SkillOpt target, judge, and reflection launch with strict network-disabled permissions and minimal runtime reads; visual cases alone receive bounded temporary-workspace artifact writes, with protected control output and symlink rejection.
• Restricted the bundled Codex gateway to loopback-only, text-only operation that denies workspace reads and rejects all writes, host config/rules, inherited environments, and profiles; it also terminates full Codex process trees on success, timeout, or client disconnect. Remote deployment remains prohibited without OS/container host-read isolation.
• Hardened provider endpoint probes, optimized-skill adoption, and local-artifact promotion against leaked URL/auth secrets, raw transcripts and private paths, stale proof, and regressing test scores.
• Pinned release publication and annotated tags to the exact validated main commit, making the workflow fail closed if main advances between readiness and publication.

v0.6.2

Added

Changed

Fixed

• Fixed drawio-diagrams route-crossing validation so transparent text boxes and callouts are still treated as connector obstacles.
• Added regression coverage for connectors crossing transparent fillColor=none callouts.

Deprecated

Removed

v0.6.1

Added

• Added strict XML preflight and browser URL helper coverage to the public drawio-diagrams skill.
• Added drawio-diagrams rules for real-logo defaults, icon source cascades, connector routing, fixed-aspect logos, and simplified/detailed diagram views.
• Added business AI workflow lookup terms to the drawio-diagrams icon catalog for AI, automation, ERP/CRM, cloud, data, delivery, and security icon needs.

Changed

• Expanded drawio-diagrams setup, delivery, verification, and icon-catalog guidance with concrete commands and curated style starters.
• Updated drawio-diagrams to ask for missing-logo fetch approval during initial setup and to use simplified icons consistently when approval or sources are unavailable.

Fixed

• Hardened drawio-diagrams regression coverage for forbidden XML constructs, icon catalog smoke checks, and browser URL generation.
• Added diagram-rule regression coverage for floating semantic edges and distorted image/logo cells.

Deprecated

Removed