Skip to content

CHANGELOG.md

Latest commit

 

History

History
597 lines (433 loc) · 61.8 KB

CHANGELOG.md

File metadata and controls

597 lines (433 loc) · 61.8 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

Contributors add user-facing entries under [Unreleased] in the same PR. Maintainers rename that section to a version and date when cutting a PyPI release. See CONTRIBUTING.md.

[Unreleased]

Added

  • CLI / packaging: Harden version display when editable and PyPI installs overlap — get_package_version_display() never prints None; detect_install_conflicts() via assess_install_health(); startup stderr hint; skillware doctor --install with recovery commands; install summary in skillware config show (#333).
  • Scripts: scripts/dev_install.sh and scripts/dev_install.ps1 — uninstall overlapping installs and reinstall editable dev deps.
  • Core / CLI: EVM operator config layer — bundled skillware/data/evm_defaults.yaml, writable user evm.yaml, skillware.core.evm_config merge helpers (load_merged_evm_config, resolve_chain, resolve_rpc_url, normalize_evm_address), and skillware evm commands (init, chains list, chain add, rpc enable, validate, open). Project/global YAML uses top-level evm: (or read-only web3: alias) — distinct from orchestration chains: (#379).
  • Docs: EVM operator config — shared defi RPC setup guide; cross-links from CLI, API keys, DeFi catalog, evm_tx_handler, and token_security_scanner (#379).
  • Core / CLI: Shared address book public_0x field, relaxed contact validation (email or wallet), top-level skillware addressbook commands (list, edit, set-wallet, remove, open), skillware config open, skillware evm token add, and skillware evm tokens list for operator token registry (#373, #374).
  • Docs: Address book operator config — shared contact identity guide (mail + defi, schema, CLI, cross-links); glossary public_0x / shared address book terms; transfer-by-name usage examples on evm_tx_handler (#373, #374).
  • Examples: gemini_gmail_minimal.py — README quick-start interactive Gemini + Gmail loop; pay_and_notify_demo.py — reproducible SkillContext pay + notify recipe with documented prerequisites.
  • Skill (defi/evm_tx_handler v0.3.0): Central address book recipient resolution with needs_input disambiguation; merges operator EVM chain/token config from skillware.core.evm_config (#373).
  • Docs: Structured hub architecture for catalog discovery — compact root README category table, docs/sitemap.md, and landing pages at docs/skills/<category>/README.md with catalog pages beside them (docs/skills/<category>/<skill_name>.md). Runtime skills/<category>/<skill_name> IDs and SkillLoader.load_skill() are unchanged (#370).
  • Skill (security/prompt_injection_firewall v0.2.0): OWASP LLM01 Layer-1 trust-boundary input defense upgrade — local evasion detection engine (leetspeak deobfuscation, multi-token ROT13 and token-reversal, typoglycemia scrambled keywords, mixed-script homoglyphs, markdown and HTML image exfiltration channels), academic/advisory mention-vs-use false-positive controls, operator policy telemetry (policy_action: allow | flag | block, removed_span_count, sanitized_length_delta), finding enrichments (decode_chain, decoded_preview), and DoS soft resource caps failing closed safely (#273, #361).
  • Skill (defi/token_security_scanner v0.1.0): Read-only GoPlus Token Security scan (scan, supported_chains) with normalized risk_tier / signals JSON for pre-trade agent checks; chain resolution via skillware.core.evm_config (#365, #368).

Changed

  • Docs: CONTRIBUTING.md — editable vs PyPI on the same Python; CLI docs for doctor --install (#333).
  • Docs: Catalog pages include a standardized intent header (problem solved, host agents, runtime / key requirements) for search matching without stuffing (#370).
  • Docs / CI: Doc-drift guards follow nested catalog paths; CONTRIBUTING, PR/issue templates, and the skill template point at category hubs and the sitemap (#370).
  • Skill (defi/evm_tx_handler): Setup envelope docs URL now points at docs/skills/defi/evm_tx_handler.md (#370).
  • Skill (office/pdf_form_filler): short_description no longer implies a Claude-only mapper (#370).
  • Card UI (security/prompt_injection_firewall): Expose policy_action, removed_span_count, and sanitized_length_delta in card.json and sync sample fixture (#273, #361).
  • Docs: Glossary cross-links on hub and catalog pages (#363).
  • README: Quick-start usage example is Gemini + office/gmail_handler (gemini_gmail_minimal.py) instead of the prompt-injection firewall snippet.

[0.5.6] - 2026-09-20

Added

  • Skill (linguistics/korean_slang v0.1.0): New linguistics/ category and offline Korean Gen-Z slang pack — interpret / suggest / lookup over a September 2026 curated KB, honorific audience gates, and constitution filters for slurs (#34).
  • GitHub: cat: linguistics in .github/labels.json and the New Skill Proposal category dropdown; GitHub UI labels sync on merge to main (#34).
  • Examples: korean_slang_demo.py — offline interpret, suggest, lookup, unmatched, and blocked-term paths (#34).
  • Scripts: korean_slang_stress_sim.py — offline interpret/suggest stress harness (#34).
  • Docs: Glossary and inclusive language — operator / contributor / host agent / end user; skill bundle vs PyPI package; anatomy roles (#252).
  • Core: Pluggable secret providers — SecretProvider, EnvSecretProvider, MappingSecretProvider, CallableSecretProvider, and SkillLoader.resolve_env_vars() inject manifest env_vars into BaseSkill(config=...) without requiring global os.environ mutation (#39).
  • Core: BaseSkill.credential() — config-first credential lookup with os.environ fallback for local .env workflows (#39).
  • Core: SkillContext(secret_provider=...) resolves credentials on each execute() and passes them via config (supports ephemeral tokens from custom providers) (#39).
  • CLI: skillware doctor ENVS column — reports missing required manifest env_vars (#39).
  • Examples: secret_provider_demo.py — host-injected credentials without os.environ (#39).
  • Core: SkillLoader.to_bedrock_tool() for AWS Bedrock Converse tool-use API — maps manifest parameters to toolSpec / inputSchema.json (#262).
  • Docs: Enterprise cloud usage guides — enterprise_cloud.md (hosting + adapter routing), bedrock.md, azure_openai.md, vertex.md (#262).
  • Examples: bedrock_tos_evaluator.py — Bedrock Converse loop with to_bedrock_tool() (#262).
  • Skill (creative/deck_builder v0.2.0): Enterprise presentation platform baseline — procedural Pillow placeholders, image fit policies, layouts (timeline, metrics, comparison), lint_deck quality gates, suggest_outline archetypes, governance ribbons/footers, and deck_builder_chain_demo.py (#336, #337).
  • Examples: deck_builder_chain_demo.py — suggest_outline → bg_remover → lint_deck → render via SkillContext (#337).
  • Skill (optimization/context_optimizer v0.1.0): Query-aware extractive context selection — local fastembed chunk scoring against agent_goal, traceable chunks_selected, fail-closed empty_result, and constitution-bound extractive-only output (#44).
  • Examples: context_optimizer_demo.py, context_optimizer_chain_demo.py (firewall → optimizer), optional context_optimizer_gemini_loop.py, context_optimizer_claude_loop.py (#44).
  • Skill (finance/uk_companies_house_handler v1.3.0): Phase v2c upgrade with deterministic matchers, filing helpers, and pagination controls — pure deterministic officer role matchers (officer_role) supporting 3 canonical categories (directors, secretaries, corporate) and 29 official Companies House statutory roles in terminology_map.yaml without NLU in Python; case-insensitive officer name substring matching (officer_name / officer_filter); multi-page officer scanning up to 10 pages (1,000 records max) in 100-item chunks for filtered queries; deterministic filing history sorting descending by date; filing helpers latest_only: true (single most recent filing) and latest_per_category: true (latest filing per category); direct host agent run_pipeline orchestration with <from_resolve> parameter substitution for multi-intent queries, removing redundant map_intent from manifest actions and agent directives; new decoupled composite action resolve_company_officer (resolve company -> halt on needs_input if ambiguous -> filter matching officer(s) by role and/or name in one turn); active vs. resigned transparency and disclaimers in envelopes (active_only, matched_count, terminology_note); strict context isolation preventing sticky filter contamination across turns (#310, #220).

Changed

  • Docs: Terminology pass on current anatomy (Contract / Effect / Directive / Assurance / Presentation) — Skill bundle standard, calling agent → host agent, trust-model “Most operators”, README domain wording, MiCA example “Available tools”; tests/CI use placeholder names instead of dummy (#252).
  • Docs (context_optimizer): Catalog execute snippets use a tight token budget so copy-paste runs demonstrate selection; card UI fixture aligned to sample policy; optimize_document_context named chain in .skillware.yaml.example (#44).
  • Docs: Revamp skill trust model — trust-forward operator guide aligned with secret providers, SkillContext, doctor/paths tooling, and untrusted-input chains; soften README and usage index credential callouts.
  • Docs: Add host-context guidance for choosing full Directives, brief registry lines, and host-managed progressive loading (#348).
  • Core: SkillContext.tools("bedrock") exposes Bedrock Converse tool specs for multi-skill hosts (#262).
  • Skills: Bundled skills with env_vars now use self.credential() for config-first secret resolution (#39).

Removed

  • Skill (finance/uk_companies_house_handler v1.3.0): Excised redundant map_intent action, internal helper methods (_map_intent, _normalize_keyword, _lookup_terminology), and intent_to_action from terminology_map.yaml in favor of direct host agent run_pipeline step composition (#220).

Fixed

  • Skill (finance/uk_companies_house_handler v1.3.0): Officer name filtering (officer_name) matches all word parts across candidate name, deterministically resolving natural order ("Firstname Lastname") vs UK registry inverted order ("SURNAME, Firstname") (#220).
  • Skill (finance/uk_companies_house_handler v1.3.0): run_pipeline halts <from_resolve> substitution upon encountering subsequent company resolution steps, preventing cross-company parameter corruption in multi-company pipelines; context company_name inheritance guarded against mismatched company numbers (#220).

[0.5.5] - 2026-09-14

Added

  • Tests: Five-provider Usage Examples guard in tests/test_skill_docs.py — every catalog page must expose Gemini, Claude, OpenAI, DeepSeek, and Ollama snippets with load_skill and skill.execute (#104).
  • CLI: skillware theme [pastel|ocean|mono] subcommand — set or interactively choose the global presentation theme; --help topic index now includes Context, Chains, and Theme alongside existing groups (#335).
  • Skill (creative/deck_builder v0.1.0): Deterministic Microsoft PowerPoint (.pptx) presentation assembly from structured JSON deck specifications — 10 slide layout types (title, section, bullets, two-column, image, image with caption, quote, table, chart, blank), 3 bundled 16:9 widescreen master templates (pitch, corporate, minimal), theme token customization, pre-flight validation with soft-limit truncation warnings, directory traversal defenses, and inspection actions (#276).
  • Skill (data_engineering/semantic_web_proxy v0.1.0): Semantic web proxy that reduces a live page or raw HTML to token-efficient Markdown, plain text, or JSON via trafilatura — boilerplate, script, and navigation stripping, opt-in comment threads, document metadata, estimated token savings with optional context-window share, an SSRF guard that re-validates every redirect hop, and a page_likely_requires_javascript warning instead of a silently empty payload for client-rendered pages (#42).
  • Examples: semantic_web_proxy_demo.py — offline fixture-backed demo of boilerplate stripping, comment inclusion, the render warning, and the SSRF guard (#42).
  • Examples: claude_uk_companies_house_handler.py — interactive UK Companies House v1.2.1 Claude loop with multi-turn tool chaining and disambiguation hints (#346).
  • Scripts: uk_companies_house_host_simulation.py — maintainer stress harness for host/Gemini/Claude NLP loops against live Companies House API (not run in CI).

Fixed

  • Core: Sanitize registry IDs in SkillLoader.to_claude_tool() so Claude API tool names match the ^[a-zA-Z0-9_-]{1,128}$ pattern (#104).
  • Examples: Claude demo scripts match SkillLoader.to_claude_tool() names instead of raw manifest IDs (#104).
  • Docs: Align Claude tool-name guidance in agent_loops.md and cli.md with sanitized adapter names (#104).
  • CI: Format tests/test_extras_sync.py with Black after the install-extras guard landed (#340 follow-up).

Changed

  • Skill (finance/uk_companies_house_handler v1.2.1): Stabilization upgrade for multi-turn agent loops — single-step stack pop execution for run_pipeline (steps.pop(0)) preserving turn-level steering and intermediate visibility; in-flight <from_resolve> parameter substitution across remaining steps; decoupled composite actions (resolve_and_get_officers, resolve_and_get_filings) executing directly without wrapping _run_pipeline and skipping profile fetch when company_number is already known; breaking rename of map_intent output suggested_pipeline to steps with parameter merging via action_params; removed hallucinated snippet_type from _resolve_company candidate parsing; strict status: "partial" lifecycle reserved exclusively for in-flight run_pipeline calls with remaining steps (standard 10-item previews return ready); sanitized session context to lean keys (company_number, company_name, last_action, selected_transaction_id) dropping persisted filter hints; excised deprecated next_actions envelope field; directive expanded for empty registry rows and mandatory user-facing replies; agent_hint on empty officers[] / filings[] (#341, #346).
  • Docs (skill catalog): Backfill runnable five-provider agent loops on all catalog pages — including gmail_handler, semantic_web_proxy, and stub Ollama sections — aligned with docs/usage/skill_usage_template.md (#104).
  • Docs (install_extras.md): Backfill deck_builder, gmail_handler, and [all] package rows; add CI guard comparing the install guide to pyproject.toml optional-dependencies.
  • Docs (skill catalog): Skill history sweep — merge SHAs for deck_builder and semantic_web_proxy, #345 rows for Gemini 3.5 bumps (mica_module, tos_evaluator, mental_coach, synthetic_generator, token_limiter).
  • Docs / examples: Default Gemini model IDs migrated from 2.5 Flash / Flash-Lite to gemini-3.5-flash and gemini-3.5-flash-lite across catalog pages, runnable examples, skill defaults, and docs/usage/gemini.md (#265).
  • Docs / examples: Default Claude Haiku snippets migrated to claude-haiku-4-5-20251001 across catalog pages, docs/usage/claude.md, and examples/claude_token_limiter.py; model_pricing.json updated (#346).
  • Docs (uk_companies_house_handler): Rate-limit links, agent-loop host guidance, multi-turn Claude snippet; UK pipeline/composite patterns moved from generic agent_loops.md to the catalog page; generic multi-turn guidance only in agent_loops.md (#346).
  • Examples: Gemini/Claude UK Companies House loops use bundle["class"]() and print needs_input disambiguation hints (#346).
  • Skills (compliance/tos_evaluator, compliance/mica_module, wellness/mental_coach, data_engineering/synthetic_generator v0.1.1): Default Gemini evaluator / generator model IDs updated to 3.5 Flash / Flash-Lite (#265).
  • Skill (monitoring/token_limiter): model_pricing.json updated with Gemini 3.5 Standard-tier list prices (Sep 2026) (#265).
  • CLI: Brief skillware --help topic list aligned with Context, Chains, and Theme command groups (matching docs/usage/cli.md) (#335).
  • Security: Bump support windows — >= 0.5.5 patched, 0.4.6–0.5.4 silent (no security fixes), < 0.4.6 unsupported with CLI advisory (SECURITY.md, version_policy.py).

[0.5.4] - 2026-09-03

Added

  • Core: SkillContext — one-line registry host context with discovery filters, progressive prepare() / execute(), and provider tool adapters (#330).
  • Core: Named skill chains — chains: in YAML, run_chain(), step when: conditional skip, skillware chain list|show|validate|run|dry-run, and skillware context show (#330).
  • Docs: Skill chaining and registry context — closes #297; implements #330.
  • Examples: skill_context_gemini_loop.py (SkillContext + Gemini), sanitize_input_chain_demo.py (run_chain local); ollama_skills_test.py refactored to use SkillContext.

Changed

  • Docs: SkillContext modes, edge cases, and Ollama multi-skill guidance in skill_chaining.md, ollama.md, introduction.md, cli.md; one-line chain pointers on middleware skill catalog pages.
  • CLI: User-configurable pastel, ocean, and mono presentation themes; interactive menu selection persists globally, project config can override it, and unknown values fall back to pastel (#248).
  • CLI: The mail submenu and direct mail commands now follow the active presentation theme (#248).
  • Skill (security/deceptive_ui_guard v0.2.0): Upgraded deceptive UI scanner with DOM zone classification (checkout, modal, cmp, navigation, general) and severity multipliers, KB allowlists (screen-reader accessibility, CMP consent banners, SEO metadata), expanded taxonomy (prechecked opt-ins, drip pricing, fake scarcity timers, nag loops), mobile profile heuristics, session fingerprint tracking, optional Playwright computed-style render diffing lane, and 23 golden HTML test corpus fixtures (#314, #327).
  • Security: Bump support windows — >= 0.5.4 patched, 0.4.6–0.5.3 silent (no security fixes), < 0.4.6 unsupported with CLI advisory (SECURITY.md, version_policy.py).

Fixed

  • Examples: Derive agent-loop tool dispatch names from the loaded manifest in claude_tos_evaluator.py, ollama_tos_evaluator.py, and ollama_novelty_extractor.py (#178).

[0.5.3] - 2026-09-01

Added

  • Skill (monitoring/kpi_gate v0.1.0): Deterministic business-KPI gate evaluating a metrics snapshot against an operator-maintained policy charter and optional versioned benchmark data — error/warning findings plus honest insufficient_data refusals with reason codes, fail-closed contract errors from a closed registry, stdlib-only validation, and the catalog/docs surface (#317).
  • Tests (tests/test_examples_smoke.py): Added automated CI smoke test suite for local-execute offline demo scripts under examples/, catching import regressions and SkillLoader dispatch errors without requiring live API keys (#237).
  • Docs (docs/TESTING.md): Documented the example smoke testing layer and skip policy for live model provider loops in CI (#237).

Changed

  • Docs (README.md, docs/introduction.md): Refresh opening narrative, tighten Quick Start, shorten Gemini hero example (monitoring/token_limiter), rectangular How it works Mermaid, simplify introduction diagrams, remove legacy Body/Mind/Conscience aliases, and align Presentation (card.json) with registry bundle standard (#326).
  • Security: Bump support windows — >= 0.5.3 patched, 0.4.6–0.5.2 silent (no security fixes), < 0.4.6 unsupported with CLI advisory (SECURITY.md, version_policy.py).
  • Docs: Document issuer.org design-ownership policy; align ARPA-driven registry skills (prompt_injection_firewall, bg_remover, novelty_extractor) and catalog Issuer lines (#295).

Fixed

  • Docs (README.md, docs/introduction.md): How it works Mermaid — rectangular AnyHost["Any Host"] node aligned with Registry and Loader blocks (#326).
  • Tests: Isolate pytest tests/ from the operator's global config.yaml via autouse SKILLWARE_CONFIG_DIR in tests/conftest.py; add configured-mode discovery and loader coverage alongside legacy-order tests (#302).

[0.5.2] - 2026-08-27

Added

  • Skill (security/deceptive_ui_guard v0.1.0): Deterministic deceptive UI surface scanner for web HTML — dual DOM vs visible extraction, structural/lexical heuristics, corroboration gates, trust score, and agent guidance for pre-click and pre-context guardrails (#78).

  • Documentation: Skill catalog pages include a Version header and Skill history table (commit links, dates, versions, and linked GitHub contributor handles); docs/skills/README.md adds a Version column to category tables.

  • Tests: test_catalog_pages_have_version_and_history_blocks() and test_skill_library_index_has_version_column() in tests/test_registry_docs.py guard catalog version/history coverage.

  • Skill (finance/uk_companies_house_handler v1.2.0): Phase v2b upgrade — run_pipeline for sequential multi-step execution with automatic halting on needs_input (disambiguation) or error, step-slicing resumption (steps[prior_completed:]), placeholder auto-substitution (<from_resolve>), and cumulative data payload merging (#220).

  • Skill (finance/uk_companies_house_handler v1.2.0): Composite actions resolve_and_get_officers and resolve_and_get_filings to resolve companies and fetch target records in single-turn operations (#220).

  • Skill (finance/uk_companies_house_handler v1.2.0): Default limit 10 previews with partial response status and agent_hint metadata for officers and filing history (#220).

  • Skill (finance/uk_companies_house_handler v1.2.0): Session context carry-forward across sequential turns (company_number, company_name, role_hint, officer_filter, next_actions) (#220).

  • Documentation: docs/skills/finance/uk_companies_house_handler.md — updated architecture reference for 9 action handlers, partial response previews, and pipeline orchestration (#220).

  • Tests: Expanded unit test suite in skills/finance/uk_companies_house_handler/test_skill.py covering pipeline resumption, composite actions, partial previews, map_intent guards, and punctuation-safe queries (#220).

Changed

  • Skill (security/deceptive_ui_guard v0.1.0): Production deception lexicon with taxonomy metadata and eight categories; standalone agent instructions; demo fixtures recreating documented dark patterns; install_extras security category row (#78).
  • Docs: Updated contributor instructions guidance in CONTRIBUTING.md, ai_native_workflow.md, and templates/python_skill/ to emphasize concise, append-only skill context over persona starters (#258, #284).
  • Docs: Aligned ai_native_workflow.md and templates/python_skill/README.md checklists with catalog Version and Skill history requirements.
  • Meta: GitHub issue templates refreshed for v0.5.1 CLI (doctor, config show, paths/mail submenus) and office/gmail_handler filing paths; label taxonomy adds all registry cat: <category> labels (shared pastel color) so category filters never collide with repo-wide labels such as security (#294).
  • Skill (finance/uk_companies_house_handler v1.2.0): get_officers default active_only is now true (v1.x default was false) — resigned officers are excluded unless active_only: false (#220).
  • Skill (finance/uk_companies_house_handler v1.2.0): Removed conversational prefix stripping from composite actions; agents must pass clean query / company_query plus optional role_hint. map_intent returns needs_input when a company name is required but missing (#220).
  • Skill (finance/uk_companies_house_handler v1.2.0): map_intent wires document_types, entity_types, and status_mappings from terminology_map.yaml into terminology_map output; richer agent_hint on rate-limit, timeout, and connection errors (#220).
  • Documentation: skills/finance/uk_companies_house_handler/instructions.md — skill-context instructions, disambiguation playbook, and parameter-hygiene guidance aligned with #258/#284 (#220).
  • Examples: examples/uk_companies_house_handler_demo.py and shared helpers updated for v2b composites, pipelines, partial previews, and disambiguation resume (#220).

Fixed

  • Meta: Shorten cat: security label description to fit GitHub's 100-character limit; enforce max length in label sync and tests.

[0.5.1] - 2026-08-19

Added

  • Skill (office/gmail_handler v0.2): Reply signature parity, preview metadata (signature_applied, signature_source, signature_profile); attachment list on read_message, download_attachment, and outbound send/reply attachments via flexible paths (local, file://, HTTP(S), optional cloud URIs with fsspec); multi-profile signatures via mail.signatures and skillware mail signature profiles|add-profile|set-profile (#293).
  • Config: Top-level mail: section in .skillware.yaml and global config.yaml — address book path, plain-text signature (inline or file), scan state, and send ledger paths for office/gmail_handler (#292).
  • CLI: skillware mail submenu — address book init/show/validate/set-path/add; signature init/show/set/validate/clear; default HTML signature with Skillware logo (40px) and links; interactive menu option 7 / mail (#292).
  • CLI: skillware config show displays merged resolved mail.* settings including signature source (#292).
  • Skill: office/gmail_handler reads signature from merged config when skill-local default_signature_plain is unset; ships data/config.yaml.example (#292).
  • Documentation: docs/skills/office/gmail_handler.md — attachments, multi-profile signatures, untrusted attachment disclaimer; docs/usage/cli.md and docs/usage/api_keys.md updated (#293).
  • Packaging: PyPI metadata — keywords, expanded classifiers (development status, intended audience, Python 3.10-3.12, AI and library topics), and Documentation, Issues and Changelog entries in [project.urls]. Metadata only; wheel contents are unchanged (#299).

Changed

  • Packaging: PyPI Homepage → https://skillware.site; canonical ARPAHLS casing on GitHub URLs in [project.urls] (release chore).

[0.5.0] - 2026-08-18

Removed

  • Meta: Removed maintainer-only GitHub release draft (.github/RELEASE_v0.4.9.md); use CHANGELOG.md + GitHub Releases UI instead (#290).

Added

  • Skill: office/gmail_handler — deterministic Gmail IMAP/SMTP handler for agent mail workflows: recipient resolution via editable address book, preview/confirm send and reply, inbox search/read, sent-folder and send-ledger search, scan cursor, and context carry-forward (#208, #291).
  • Examples: examples/gmail_handler_demo.py — mocked resolve, preview, search, and read flow (no live Gmail credentials); examples/gemini_gmail_handler.py — interactive Gemini tool loop (demo mode via GMAIL_HANDLER_EXAMPLE_DEMO=1).
  • Documentation: docs/skills/office/gmail_handler.md — integration guide, env setup, action reference, and address book schema.
  • CI: tests/test_registry_identity.py — CI guard asserting every registry-layout skill's manifest.name matches its path-derived registry ID and that all manifest names are globally unique (#280, #288).

Changed

  • Documentation: README, introduction, vision, skill trust model, and CLI usage docs — clarify legacy vs config path resolution, tier order, and that bundled registry skills remain available without a local skills/ tree (#289).

[0.4.9] - 2026-08-12

Added

  • Config: Persistent YAML configuration (global config.yaml and project .skillware.yaml); paths section drives skill root discovery when present; bundled registry always included (#246).
  • CLI: skillware config show prints merged configuration (read-only); skillware paths tips updated for config files (#246).
  • CLI: Interactive paths submenu (menu 4) — view bundled registry, persist project/external paths to .skillware.yaml, shadowing and flat-layout diagnose; tiered help topics (menu 6); universal navigation (0 exit, b back); doctor spinner while diagnosing (#247).
  • Loader: SkillLoader.load_skill(..., execute_module=False) inspect-only load (manifest, instructions, card, requirement pre-flight) without executing skill.py; clearer ImportError when skill.py import fails after pre-flight (#235).
  • CLI: skillware doctor checks manifest deps and skill.py import readiness per skill (DEPS / LOAD table); optional skill ID, --category, and --skills-root (#235).

Changed

  • CLI: skillware examples table — wider EXTRA column; GITHUB shows script filename as a clickable link (full URL on ctrl+click) (#247).
  • CLI: Interactive menu option 6 description reflects grouped help topics (#247).
  • Loader: Skill-not-found errors list searched roots with tier labels (project, external, bundled) (#247).
  • Loader: SkillLoader.load_skill() validates manifest requirements version specifiers (for example web3>=6.0.0) against installed package versions before loading skill.py; unpinned entries still require importability only (#14).
  • GitHub: New Skill Proposal template category dropdown synced with the registry (creative, security); added creative label (#279, #277).

[0.4.8] - 2026-08-03

Added

  • Skill: security/prompt_injection_firewall — offline-only deterministic pre-flight scanner (no LLM path) with local kb/ detectors for hidden text, Unicode/confusable evasion, nested encodings, instruction overrides, corroboration-based sensitivity, and sanitization output (#46).
  • dev_tools/issue_resolver: Caller-fetched ISSUE_RESOLVER.md repository profiles — ordered profile discovery URLs, load_repository_profile parsing with provenance-labelled context, profile standard doc, and Skillware dogfood profile under .github/ISSUE_RESOLVER.md (#145, #271).
  • Documentation: docs/contributing/issue_resolver_profile.md — normative contract and examples for repository-maintained agent context files.

Changed

  • creative/bg_remover: Hardened to v0.2.0 — rembg session reuse, Base64 and file validation, 25 MB limit, path traversal rejection, parent directory creation for output_path, demo script, and expanded bundle tests (#257, #268).
  • dev_tools/issue_resolver: Profile discovery prefers .github/ISSUE_RESOLVER.md over a repo-root fallback; expanded Skillware dogfood profile aligned with contributor workflow and stage gates.
  • Version policy: Raise security support floor to >= 0.4.7; legacy band 0.3.5–0.4.6 (upgrade recommended, silent CLI); unsupported advisory for installs below 0.3.5.

[0.4.7] - 2026-07-25

Added

  • Citation: Root CITATION.cff and README Citing section for formal software citation; first Zenodo-archived GitHub release (#269, #270).

[0.4.6] - 2026-07-23

Added

  • Documentation: OpenAI-compatible host guide and runnable Groq example covering shared to_openai_tool() usage, provider base URLs, credentials, and local servers (#261).
  • Framework: skillware/core/ui_schema.py helpers to resolve dot paths and validate output-card field keys (#199).
  • CI: Parametrized guard tests/test_card_ui_schema.py — every registry skill with output-card ui_schema must ship a fixture under tests/fixtures/card_ui_schema/ whose samples cover all field keys (#199).

Fixed

  • compliance/mica_module, compliance/pii_masker: Align card.json output fields with actual execute() return shape (#199).

Changed

  • finance/wallet_screening: Paginate Etherscan txlist (up to 10k normal txs) and surface metadata.warnings when history is truncated or unavailable; bump skill version 1.0.0 → 1.0.1 (#214, parent #115).
  • Skills (finance/uk_companies_house_handler): Completed Phase v2a upgrade (#220). Added context parameter to carry forward session state (company_number, company_name, officer_filter, etc.). Added partial envelope status and optional pipeline state to support paused multi-step pipelines (noted as v2b prep in instructions). Updated _get_officers to fallback to company_name from context and added tests for fallback logic. Updated examples/gemini_uk_companies_house_handler.py into a fully interactive chat loop.
  • Documentation: CONTRIBUTING and contributor workflow checklists — update card.json and output fixtures together when changing execute() output (#199).

[0.4.5] - 2026-07-16

Added

  • Added creative/bg_remover, an offline background removal skill powered by rembg (#196).
  • Packaging: Category, per-skill, and [all] optional extras generated from skill manifests via scripts/sync_extras.py; hand-maintained dev, SDK extras, and meta [agents] (#236).
  • Documentation: Install extras — canonical guide for pip install skillware[...] targets (#236).
  • Framework: skillware/core/extras.py and richer ImportError hints from SkillLoader when manifest requirements are missing (#236).
  • CI: PyPI wheel packaging smoke test — builds a wheel, installs it in a fresh venv (base deps only), and verifies every bundled registry skill is present and loadable via scripts/wheel_smoke_test.py (#182).
  • Documentation: Cross-linked wheel-smoke CI job in CONTRIBUTING.md and docs/contributing/ai_native_workflow.md (#182).
  • CLI: skillware paths shows skill root resolution order (external → project → bundled), tier labels, shadowing summary, and operator tips; interactive menu option 4 wired (#81).
  • Framework: skillware/core/discovery.py — shared skill root discovery for SkillLoader and the CLI (tier labels, shadowing, registry ID listing; foundation for config-driven paths in #246).
  • Framework: BaseSkill.validate_params() — optional helper to validate tool arguments against manifest parameters JSON Schema; raises SkillwareParamValidationError on mismatch. Not called automatically by the loader or execute() (#125). Reference: examples/claude_wallet_check.py, examples/gemini_tos_evaluator.py.
  • Framework: Registry manifests standardize on outputs: (legacy singular output: removed from finance/wallet_screening) (#125).

Changed

  • Framework: Map manifest requirement pillow to import name PIL in SkillLoader dependency checks (#196).
  • Packaging (breaking): Removed legacy extras cli, embeddings; [all] is skill runtime deps only (not SDK packages). Use [agents] for Gemini + Claude + OpenAI SDK deps (#236).
  • CLI: skillware list prints a pointer to the install extras guide (#236).
  • Documentation: README, CONTRIBUTING, TESTING, skill catalog pages, and examples index updated for the new extra taxonomy; every skill doc recommends its per-skill extra (#236).
  • Framework: SkillLoader delegates root resolution to discovery.get_skill_roots() so load, list, test, and paths use the same order (#81).
  • CLI: Help and menu list paths as available (no longer “coming soon”) (#81).
  • Documentation: README — trim redundant Contributing cross-links, add skillware paths to install verification, and link the skill trust model in the docs table; cross-link skillware paths in docs/introduction.md and docs/usage/README.md.
  • Documentation: Add operator expectations (bundled vs external, out-of-band changes, where to report issues) and an instruction-only content note to docs/security/skill-trust-model.md; add trust-model cross-links from README configuration and CONTRIBUTING Skill Package Standard (#243).

[0.4.3] - 2026-07-10

Added

  • Framework: Added _sanitize_gemini_tool_name() to skillware/core/loader.py to explicitly map provider tool naming constraints.
  • Tests: Added test_sanitize_gemini_tool_name() to tests/test_loader.py to verify safe translation and test_skill_docs_gemini_anti_patterns() to tests/test_registry_docs.py to enforce documentation hygiene by preventing manual tool wrapping and mutation anti-patterns in skill catalog pages.
  • Documentation: Added canonical dispatch guidelines to docs/usage/gemini.md and docs/usage/skill_usage_template.md.
  • Documentation: Category selection guidance for contributors; issue-first policy for new top-level folders in CONTRIBUTING.md (#204).

Changed

  • GitHub: Overhauled issue templates (CLI, Skill Upgrade, Examples, Packaging), issue chooser config.yml, PR template, and label taxonomy with pastel colors; labels sync automatically from .github/labels.json via CI on merge to main (#227).
  • Framework: SkillLoader.to_gemini_tool() now returns a google.genai.types.Tool object instead of a raw dictionary, ensuring compatibility with the google-genai SDK when passing tools to GenerateContentConfig.
  • Tests: Updated tests/test_loader.py to assert against the properties of the google.genai.types.Tool object for to_gemini_tool().
  • Documentation: Updated Gemini integration snippets across all skill catalog pages and introduction.md to reflect the to_gemini_tool() API change and correct tool name sanitization.
  • Documentation: Aligned agent_loops.md and cli.md with Gemini sanitized tool-name dispatch after #229 (#230 follow-up).
  • Examples: Removed manual types.Tool wrapping, and consistently utilize SkillLoader._sanitize_gemini_tool_name() for derived tool names in gemini examples.

Fixed

  • Framework: SkillLoader.to_gemini_tool() returned a plain dictionary instead of google.genai.types.Tool, causing runtime failures when passed to GenerateContentConfig(tools=...) with the google-genai SDK. Fixes #223.

[0.4.2] - 2026-07-08

Changed

  • CLI: Interactive splash — correct SKILLWARE ASCII logo, Rich 3-stop gradient on logo and tagline (#D4E4F1 → #79B6D8 → #EBD8DC), shared _package_version_str() for splash and --version, tagline Skillware v{version} — Skill Management Framework (#222).
  • Documentation: Compact architecture Mermaid diagrams — shorter README labels, model-agnostic adapter nodes in introduction, horizontal agent-loop layout with role table (#210 follow-up, #221).
  • Documentation: Docs sweep — README Mermaid Registry alignment; vision.md pinned to v0.4.x; skill catalog and usage guides prefer bundle["class"]() with explicit bundle["module"].ClassName() noted as still supported (#225).

[0.4.1] - 2026-07-08

Added

  • finance/uk_companies_house_handler: New skill for UK Companies House REST API — deterministic company search, profile, officers, PSC, filing history, and intent-to-operation mapping with UK corporate terminology translation; bundled endpoint index and terminology map; status-based response envelope (ready/needs_input/error) with disambiguation support (#172, #218).
  • Documentation: Add docs/security/skill-trust-model.md documenting the skill execution model, on-disk resolution order and shadowing, provenance tiers (Bundled / Project / External), and operator security guidance; wire links from SECURITY, usage, CONTRIBUTING, and CODE_OF_CONDUCT (#109).
  • Documentation: Add minimal Mermaid architecture flow diagrams to README, introduction, and agent loops; cross-links, Step 1 mini-pipeline, adapter fan-out, and direct-path footnotes (#210, #217).

Changed

  • Loader: SkillLoader.load_skill() auto-discovers the single BaseSkill subclass in each skill.py and exposes it as bundle["class"]; get_skill_class() helper added. Existing bundle["module"] usage is unchanged (#89).
  • Loader: SkillLoader.load_skill() validates that manifest.yaml name matches the path-derived registry ID for category/skill_name layouts; emits SkillwareIdentityWarning on mismatch (warn-only v1). Flat private skills under a skill root are unchanged. Bundles now include optional registry_id (#200).
  • Version policy: Raise security support floor to >= 0.3.5, legacy band 0.3.0–0.3.4 (silent CLI), unsupported advisory for installs below 0.3.0 (#192).
  • Documentation: Clarify skill ID vs manifest name vs provider tool names in docs/usage/cli.md, docs/usage/agent_loops.md, and docs/introduction.md; require full registry IDs in CONTRIBUTING manifest standard (#201).

Fixed

  • office/pdf_form_filler and defi/evm_tx_handler: Align manifest.yaml name with registry paths (office/pdf_form_filler, defi/evm_tx_handler); update examples and docs to use manifest-derived tool dispatch (#201).

[0.4.0] - 2026-06-30

Added

  • monitoring/token_limiter: New monitoring skill that evaluates cumulative token usage and returns CONTINUE, WARN, or FORCE_TERMINATE for autonomous agent loops; bundled indicative model pricing, ROI scaffold fields for v2, local and provider loop examples (#23, #207).
  • New monitoring skill category in CONTRIBUTING, ai_native_workflow.md, and docs/skills/README.md.
  • Tests: tests/test_registry_docs.py — CI doc-drift guards that verify skill catalog, examples index, and agent-loops reference matrix stay in sync with the registry (#183, #189).
  • Documentation: Cross-linked tests/test_registry_docs.py in ai_native_workflow.md and CONTRIBUTING.md so contributors know doc-drift guards run in CI; added explicit PR checklist reminder in CONTRIBUTING.md (#193, #197).

Changed

  • Documentation: docs/skills/monitoring/token_limiter.md — budget disclaimer callout, limitations clarity, and enterprise disclaimer (#23).
  • Documentation: README Stats section and live PyPI download badges (pepy / PyPI Stats dashboards, header DLs ↓ total) (#198).
  • Documentation: Aligned CLI and examples docs (docs/usage/cli.md, examples/README.md, docs/vision.md, README.md) with the skillware test / skillware examples behavior (#191, #194).

[0.3.9] - 2026-06-26

Fixed

  • CLI: skillware examples and list --examples fetch examples/README.md from GitHub when no local copy is found (pip installs and directories outside a checkout).

[0.3.8] - 2026-06-26

Changed

  • Tests: tests/test_skill_issuer.py now requires test_skill.py for every registry skill under skills/ (#160).
  • Documentation: Clarified that bundle tests must mock network calls and model downloads in CI (#170).
  • Documentation: Added a Status section to TESTING.md summarizing the current testing model (#179).
  • Documentation: Post-release alignment — category tables, Python 3.10+ badge, dev install ([dev,all] vs [dev]), README configuration via .env.example, DeFi env vars in .env.example, framework env vars in api_keys.md (#154).

Added

  • CLI: skillware test runs bundle tests via pytest — all roots, by skill ID, or by --category; supports -v and --no-header. Documented in cli.md, TESTING.md, and contributor guides (#83).
  • CLI: skillware list --examples shows per-skill example script counts; skillware examples [skill_id] lists indexed runnable scripts from examples/README.md with GitHub source links; interactive menu option examples (#126).

[0.3.7] - 2026-06-22

Added

  • wellness/mental_coach: Deterministic wellness coaching firewall with crisis triage, hard scope limits, embedded public KB, optional Gemini scope evaluator, and catalog documentation (#148).

Changed

  • Tests: Moved tests/test_mica_module.py to tests/skills/compliance/test_mica_module.py so maintainer skill tests follow the tests/skills/<category>/ layout; tests/ root is framework-only (#86).
  • wellness/mental_coach: Set real issuer contact email and add health disclaimer on the catalog page (PR #174 follow-up).

Fixed

  • finance/wallet_screening: Align examples and docs with finance/wallet_screening manifest tool name; fix gemini_wallet_check.py and claude_wallet_check.py to match tool name dynamically from manifest; correct card.json UI fields to match actual skill output schema; update instructions.md, provider snippets, Data Schema, and usage docs (#173).

[0.3.6] - 2026-06-15

Added

  • Tests: Backfilled test_skill.py for six registry skills (mica_module, pii_masker, synthetic_generator, wallet_screening, pdf_form_filler, prompt_rewriter); all registry skills now ship co-located bundle tests. Fixed prompt_rewriter package export so pytest can collect the bundle (#158).
  • CLI: skillware/__main__.py enables python -m skillware as a fallback when the skillware command is not on PATH (#135). Added cmd_help() for rich-formatted help, wired to skillware --help/-h and interactive menu option 4. Added --version/-V flag.

Fixed

  • novelty_extractor: Bundle tests mock fastembed embeddings so CI avoids HuggingFace downloads and rate limits (#159 follow-up).
  • finance/wallet_screening: WalletScreeningSkill.manifest loads manifest.yaml from the bundle directory (#165).

Changed

  • CI: GitHub Actions runs pytest skills/ then pytest tests/ after lint (bundle + framework/maintainer tests; closes #90) (#159).
  • CI: CodeQL GitHub Action upgraded from v3 to v4.
  • Dependencies: Extended [all] with registry skill runtime deps (web3, fastembed, numpy); added [defi] and [embeddings] optional extras. Documented manifest ↔ pyproject.toml convention in CONTRIBUTING and TESTING.md.
  • Documentation: TESTING.md, CONTRIBUTING.md, ai_native_workflow.md, and README architecture tree document the bundle / framework / maintainer / example testing model. Pytest collects tests/ and skills/ only (examples/ ignored).
  • Dependencies: Moved rich>=13.0 from [cli] extra to core dependencies; CLI is now available immediately after pip install skillware. The [cli] extra is kept as an empty deprecated alias for backward compatibility (#135).

[0.3.5] - 2026-06-05

Added

  • defi/evm_tx_handler (#142): Structured EVM agent wallet skill on Ethereum and Base — resolve, Uni V2 quote/preview/execute (approve + swap), transfer, balances, wallet_info, YAML registries, optional CoinGecko USD preview, max_trade_usd fail-closed cap, balance pre-flight checks, and mocked Web3 tests. Examples: examples/gemini_evm_tx_handler.py, examples/claude_evm_tx_handler.py.

Changed

  • CI: GitHub Actions installs from pyproject.toml only (pip install -e ".[dev,all]"); runs black --check, flake8, then pytest tests/ (#151, #153). Co-located skills/**/test_skill.py remains a local pre-PR step.
  • Documentation: COMPARISON.md and README updated for Agent Skills (SKILL.md) and fairer MCP framing (#123); TESTING.md and CONTRIBUTING.md aligned with CI and Black gate (#151, #153); defi skill category added to CONTRIBUTING.

Fixed

  • dev_tools/issue_resolver: Replaced wide emoji regex in commit-message validation with explicit Unicode ranges (CodeQL py/overly-large-range, #146).

[0.3.3] - 2026-05-29

Added

  • dev_tools/issue_resolver: GitHub issue workflow with sequential stage checklists, conditional verify/commit gates, and commit-message validation (#143).
  • Examples: gemini_issue_resolver.py, claude_issue_resolver.py, and ollama_issue_resolver.py for dev_tools/issue_resolver (#118).
  • Version policy: skillware/version_policy.py with supported-version thresholds; CLI prints one dim stderr advisory only for installs below 0.2.6 (#132).
  • Tests: tests/test_version_policy.py for advisory thresholds, opt-out, and CLI hook (#132).
  • Documentation: Added docs/vision.md with project story, roadmap, and agent discoverability (#133).

Changed

  • finance/wallet_screening: Unified TRM/scam transaction risk index for analysis (#140).
  • SECURITY.md: Supported-version table aligned with >= 0.3.1 security support and unsupported < 0.2.6 band (#132).
  • CLI: Calls version advisory once at main() startup, not on menu re-loops (#132).
  • Dependencies: Added packaging for semver comparisons (#132).
  • Documentation: README Mission links to vision.md; wallet screening comparison table lives in COMPARISON.md; docs table and cross-links updated (#133).

[0.3.2] - 2026-05-27

Added

  • Changelog: Added root CHANGELOG.md following Keep a Changelog, with retrospective release history from v0.2.0 and a README nav link (#108).
  • finance/wallet_screening: FTM publicKey matching and an ETH sanctions index (#128).

Changed

  • CLI: Visual redesign for skillware list, including pastel table, short_description column, interactive splash, and menu (#129).
  • CLI: Interactive polish - splash footer links, menu re-loop, stub labels for #81 / #83, width-aware table, shared terminal context (#130, #131).
  • Contributing: Aligned Code of Conduct, CONTRIBUTING, agent workflow, and PR template for CHANGELOG maintenance and co-authoring rules (#124).
  • Documentation: README documentation table and docs/introduction.md link to CHANGELOG.md; contributor template documents optional short_description (#130).

[0.3.1] - 2026-05-25

Added

  • Novelty Extractor Skill: Introduced the data_engineering/novelty_extractor skill (#116, fixes #24).
  • Examples Index: Added examples/README.md to serve as the canonical index of runnable provider scripts (#107).

Changed

  • SDK Migration: Migrated framework and examples from google-generativeai to the new google-genai SDK (#97) and updated all usage documentation snippets (#92).
  • Documentation: Improved README navigation and overall skill catalog discoverability (#98).
  • Documentation: Cross-linked runnable examples directly on skill catalog pages (#121) and synced agent_loops.md with the central examples index (#122).

[0.2.9] - 2026-05-22

Added

  • CLI Tool: Introduced the skillware command-line interface, starting with the skillware list command for skill discovery (implemented by contributor @rizzoMartin) (#84).
  • CLI Features: The list command prints a rich table of locally installed skills and supports filtering via --category, --issuer, and --skills-root flags.
  • Optional Extras: Added optional dependency groups in pyproject.toml ([cli], [gemini], [claude], [openai], [office], [all], [dev]) so users only install the SDKs their specific skills require (#87).

Changed

  • Leaner Core Install: Removed heavy SDKs (anthropic, google-generativeai, pymupdf, openai) from the default installation, reducing core requirements to just requests, pyyaml, python-dotenv, and beautifulsoup4 (#87).
  • Dependency Management: Consolidated dependency management entirely into pyproject.toml.
  • Requirements File: Transformed requirements.txt into a dev-convenience pointer (running pip install -e ".[dev,all]") rather than a duplicate flat dependency list.

[0.2.8] - 2026-05-22

Added

  • Issue Resolver Skill: Introduced the dev_tools/issue_resolver skill for universal GitHub issue analysis and resolution (#56).
    • Validates and normalizes any public GitHub issue URL and returns a structured payload containing pre-computed GitHub API and raw content URLs.
    • Guides calling agents through a 5-stage workflow (fetch issue, read repo context, analyze files, produce a ranked plan, implement after approval).
    • Includes optional github_token and extra_instructions parameters.
    • Compatible with all five provider adapters (Gemini, Claude, OpenAI, DeepSeek, Ollama).
    • Requires no network calls itself and has no runtime dependencies beyond PyYAML.
  • Dev Tools Category: Introduced the new dev_tools skill category.

Changed

  • Skill Catalog Revamp: Overhauled all pages under docs/skills/ to include a breadcrumb trail, per-provider Usage Examples, environment variable tables linking to the API keys guide, data schema blocks, and a limitations section (#82).
  • Documentation Polish: Removed emojis from all catalog pages and the main skills README index (#52).

Fixed

  • Metadata: Corrected the author name in pyproject.toml from ARPA Hellenic Logic Systems to ARPA Hellenic Logical Systems.

[0.2.7] - 2026-05-18

Added

  • Packaging: Full skill bundles on PyPI. Wheels now include manifest.yaml, instructions.md, card.json, and skill data files, rather than only .py modules.
  • Packaging: Configured MANIFEST.in to graft the skills/ tree and updated [tool.setuptools.package-data] so new registry skills do not require per-skill pyproject.toml edits.
  • Registry: Added empty __init__.py files under skills/ category packages (and skill folders where needed) to ensure setuptools packages the complete tree. This requirement is now enforced in tests for new registry skills.
  • Documentation: Added a "Finding skills on disk" usage guide.
  • Documentation: Added contributor notes for PyPI packaging in CONTRIBUTING.md and the skill template README.

Fixed

  • Skill Loader: Fixed skill resolution paths after pip install (#13). SkillLoader.load_skill() no longer restricts searches to site-packages/skills/. It now falls back through the following order:
    1. An existing path on disk (absolute or relative)
    2. Roots defined in the SKILLWARE_SKILL_PATH environment variable
    3. A local skills/ folder in the current working directory (searching up to six parent directories)
    4. Bundled registry skills shipped with the package (Note: If nothing matches, the error now explicitly lists the paths that were tried).

[0.2.6] - 2026-05-17

Added

  • Framework: Added OpenAI adapter (SkillLoader.to_openai_tool()) for Chat Completions tool calling (#68).
  • Framework: Added DeepSeek adapter (SkillLoader.to_deepseek_tool()) as a separate public API (#70).
  • Framework: Added shared function-name sanitization for OpenAI-compatible providers.
  • Documentation: Added OpenAI and DeepSeek usage guides and corresponding integration examples (examples/openai_tos_evaluator.py, examples/deepseek_tos_evaluator.py) (#69, #70).
  • Documentation: Added usage guides index, agent loops, and skill usage template (#71).
  • Documentation: Added Usage Examples on all seven skill catalog pages (Gemini, Claude, OpenAI, DeepSeek, Ollama) (#71).
  • Documentation: Added generic setup guide for API keys for skills (#67).
  • Documentation: Added README links to usage index and agent loops (#71).
  • Contributing: Added Agent Contribution Workflow, an agent-directed guide (#64, #65).
  • Registry: Added Issuer attribution on all skills (manifest.yaml, card.json, docs/skills/*.md, catalog) (#63).
  • Registry: Added Enterprise disclaimer on ARPA catalog skill pages (#59, #62).
  • Tests: Added tests/test_skill_issuer.py for registry issuer validation (#63).

Changed

  • Documentation: Updated Ollama guide for current local models (#71).
  • Contributing: Restructured CONTRIBUTING.md for contribution types and skill standards (#64).
  • Contributing: Aligned the Usage Examples requirement in CONTRIBUTING and agent workflow (#71).

[0.2.5] - 2026-04-28

Added

  • TOS Evaluator Skill: Introduced the compliance/tos_evaluator skill for local-first website policy evaluation prior to automated access.
    • Checks robots.txt permissions for target URLs and user-agents.
    • Discovers candidate legal pages (Terms, Legal, Acceptable Use, API links).
    • Extracts and evaluates policy clauses related to automated behaviors (scraping, crawling, indexing, monitoring, etc.).
    • Returns structured verdicts (SAFE, UNSAFE, CAUTION, INSUFFICIENT_EVIDENCE) alongside evidence payloads and next-step guidance.
    • Features an optional, provider-configurable low-cost LLM fallback for ambiguous clauses.
  • Skill Infrastructure: Added the complete package contents for the TOS Evaluator under skills/compliance/tos_evaluator/ (including manifest, logic, and instructions).
  • Testing: Added central tests (tests/skills/compliance/test_tos_evaluator.py) and local skill tests.
  • Documentation: Added dedicated skill documentation (docs/skills/compliance/tos_evaluator.md) and updated the central skill catalog.
  • Examples: Added integration scripts (examples/gemini_tos_evaluator.py, examples/claude_tos_evaluator.py, examples/ollama_tos_evaluator.py).
  • Dependencies: Added beautifulsoup4 (bs4 in the manifest) to the project for deterministic HTML parsing.

[0.2.4] - 2026-04-11

Added

  • MiCA Compliance Module: Added the compliance/mica_module skill, featuring in-memory caching for ultra-low latency RAG (~1.7ms) and a weighted surgical router to prevent context window asphyxiation.

Changed

  • Pure Cognitive Framework: Realigned all MiCA examples (Gemini, Claude, Ollama) to follow a prompt-based cognitive pattern that avoids opaque native tool-calling obstacles.
  • Documentation: Comprehensive documentation updates for the new Compliance category and a refined core README.

Fixed

  • Quality Engineering: Resolved all PEP 8 and Flake8 violations across the registry and verified execution with 100% unit test success.

[0.2.3] - 2026-04-09

Added

  • Zero-Latency PII Masker Skill: Introduces the compliance/pii_masker component to act as a "Privacy Firewall", intercepting and scrubbing sensitive metadata (Names, Emails, Physical Addresses, Crypto Wallets) locally before external LLM dispatch.
  • Ollama Edge Interoperability: Leverages the 270M parameter arpacorp/micro-f1-mask structure for optimized, offline processing.
  • Dynamic Modalities: Added three processing modes for the masker:
    • mask: Preserves contextual entity tags (e.g., [PERSON_1]).
    • redact: Completely overwrites tokens with localized constants (xxxx).
    • remove: Intelligently drops strings from the payload to decrease token size.
  • Testing: Integrated rigorous Pytest mock structures intercepting the edge boundary.

Changed

  • API Manifests: Rewrote API compliance manifest parameters to match the internal JSON Schema architecture.

[0.2.2] - 2026-04-03

Added

  • New Skill: Introduced the data_engineering/synthetic_generator skill for bulk-generating high-entropy synthetic training data to combat model collapse (Resolves #22).
  • Model Agnosticism: Added internal routing support for the synthetic generator to use Ollama, Gemini, and Anthropic.
  • Zero-Dependency Entropy Scoring: Added a new zlib compression ratio heuristic to natively validate lexical entropy and block boilerplate outputs without heavy NLP dependencies.
  • New Documentation: Launched the Data Engineering category in the central skill registry along with comprehensive integration guides and integration scripts (examples/build_dataset_demo.py).

Fixed

  • Bug Fixes: Addressed all flake8 PEP8 linting issues across the module.

[0.2.1] - 2026-03-21

Added

  • Prompt Token Rewriter Skill: A new middleware skill (optimization/prompt_rewriter) that heuristically compresses bloated prompts into fewer tokens, supporting low, medium, and high aggression levels.
  • Optimization Category: Established a new domain in the skill registry for architectural and operational efficiency tools.
  • Skill Reference Card: Comprehensive documentation for the Rewriter at docs/skills/optimization/prompt_rewriter.md.
  • Interactive Demo: Added examples/prompt_compression_demo.py for offline testing of compression logic.

Changed

  • Middleware Patterns: Updated the Gemini usage guide with "Skill Chaining" examples demonstrating the rewriter as an automated pre-processor.
  • Standardized Manifests: Aligned all skill metadata with the new parameters and constitution standard.

Fixed

  • CI/CD Alignment: Fixed linting and formatting issues to ensure 100% flake8 compliance in core registry files.

[0.2.0] - 2026-03-21

  • Consolidated and rolled forward into v0.2.1.