Changelog

All notable changes to this project are recorded here.

Unreleased

• Added an Agent Plugins 1.0.0 portable core with root plugin.json, root mcp.json, canonical Agent Skills discovery, and the existing read-only stdio server.
• Added offline canonical-schema validation with source digests, path and symlink containment tests, a real portable MCP startup check, archive assertions, and public install and security guidance.
• Added the mobile-first-website-design Agent Skill at metadata.version: "0.1.0" with a byte-exact 228-file native evidence packet, complete 12,029-line source mapping, all eight native cases, deterministic punctuation normalization for public prompt scalars, three worked examples, skill-local CI, and repository evaluation coverage.
• Updated Python plugin, skills.sh, Aider, Hermes Agent, package, repository, and MCP discovery routes for mobile-first-website-design.
• Added the prompt-enhancer Agent Skill at metadata.version: "0.1.0" with all sixteen native behavior cases, five packaged validation scripts, worked examples with real script output, complete nonblank-line coverage of the eight-file native source packet under evidence/ports/, package-local CI, and repository evaluation coverage.
• Updated Python plugin, skills.sh, Aider, Hermes Agent, package, repository, and MCP discovery routes for prompt-enhancer.
• Raised timebox to 0.1.1, restored its native hard-acceptance, admission, incomplete-work, verification, and evaluation-owner semantics, replaced count-only source mapping with executable public-target assertions, and aligned the published evaluation-manifest schema with the live parser and all 15 manifests.
• Closed final review follow-ups by splitting oversized validation code and regressions, documenting target-scaffolding, and labeling mixed visual-design source evidence explicitly.
• Refreshed the visual-design live-font fixture and raised that skill to 0.2.1 after catalog drift invalidated its recorded ranks.
• Closed repository-wide logic gaps in evaluation completeness, fixture claim scope, source evidence verification, public-file lineage coverage, special-entry rejection, catalog identity, empty trigger rejection, and dependency audit state.
• Added committed source-evidence manifests for all published skills. Historical native packets stay byte-bound where available; legacy packages without recoverable source bytes now declare a complete repository_baseline instead of unverifiable historical hashes.
• Added timebox at metadata.version: "0.1.0" with byte-exact native evidence, complete source-line coverage, all six source eval cases, deadline-bound validation, package-local CI, and collection integration.
• Added meaning-preserving-rewrite, simplify-skill, and goal-prompt at metadata.version: "0.1.0" with byte-exact native evidence, complete source mappings, portable dependency reconciliation, package-local CI, and repository evaluation coverage.
• Updated Python plugin, skills.sh, Aider, Hermes Agent, package, repository, and MCP discovery routes for the three dependency-campaign units.
• Added a regression that requires the Python plugin to register every canonical skill with its current frontmatter description.
• Added the dedupe Agent Skill at metadata.version: "0.1.0" with six adapters, eight native behavior cases, sixteen trigger cases, complete nonblank-line coverage, and a byte-exact 36-file native source packet under evidence/ports/.
• Updated Python plugin, skills.sh, Aider, package, repository, and MCP discovery routes for dedupe.
• Renamed the prior date-only skill to always-current-datetime at metadata.version: "0.1.0", added the acquired local time to every visible reply prefix, preserved the fail-closed clock helper, and expanded native lineage to all twelve date-time pressure cases.
• Updated every published skill body and active collection discovery surface to use always-current-datetime; raised logic-audit to metadata.version: "0.1.1" for its renamed dependency.
• Rebuilt the byte-exact native evidence packet, source lineage, examples, offline evals, package registration, adapters, and MCP catalogs for always-current-datetime.
• Added the logic-audit Agent Skill at metadata.version: "0.1.0" with capability-neutral web evidence, all ten native pressure cases, complete nonblank-line coverage, and a byte-exact native source packet under evidence/ports/.
• Updated Python plugin, skills.sh, Aider, Hermes Agent, package, repository, and MCP discovery routes for logic-audit.
• Fixed the date-only predecessor at metadata.version: "0.1.1" so the clock runs from the concrete skill working directory without relying on a shell variable, and added a regression case for loaders that supply no directory variable.
• Added the outcome-bounded-work Agent Skill at metadata.version: "0.1.0" with portable conversation and artifact-audit modes, all ten native pressure cases, complete nonblank-line coverage, and a byte-exact native source packet under evidence/ports/.
• Updated Python plugin, skills.sh, Aider, and Hermes Agent discovery routes for outcome-bounded-work.
• Added the date-only predecessor at metadata.version: "0.1.0" with a portable timezone script, the full registry eval artifact set, source-lineage hashes, a complete semantic mapping, and a byte-exact native source packet under evidence/ports/.
• Updated package, Python plugin, MCP, skills.sh, Aider, Continue, and Hermes Agent discovery routes for the date-only predecessor.
• Excluded the local Claude settings file and frozen source packets from Prettier so local CI does not inspect private machine settings or rewrite source evidence.

GitHub release 0.1.4

• Made examples/ a required directory in every skill: AGENTS.md gains an Examples section, the generation contract adds examples/ to the required layout, and scripts/validate_skill.py fails a skill whose examples/ directory is missing or never referenced from its SKILL.md.
• Added 39 worked examples across the four skills, each one a complete run with the user's words, the reply, real command output, exit codes, and the files the run created.
• Invoked every skill the way a user would before and after the change, and fixed what those runs exposed.
• skill-factory: the scaffolder emits examples/; the new scripts/check_placeholders.py gate fails a boilerplate scaffold that previously passed every check; scripts/doctor.py reports a degraded mode with fallback commands instead of blocking when the task runner is absent; seed evals describe the requested skill instead of the scaffolder. Tests grew from 72 to 94.
• reify: one brief schema shared by the asset, the record reference, and the validator; a validator invocation that works from a user working directory, with branches for exit 0, 1, and 2; a fixed decision ID format and defined revert moves; a reply budget matching the evals; a rule against inventing first-person biographical detail.
• starting-point: a description that can fail to match and routes vague requests to reify; the new references/constraint-classes.md; proof thresholds by task type; a limit on how much analysis reaches the user; the handoff to nonexistent workflows removed.
• visual-design-system-extractor: font ranking on fit and legibility first with rarity only breaking ties inside a band; a set-level pairing check across seven dimensions that can veto a winner; a rarity measure that no longer saturates across the rare tail; a binding meta.rarity_floor; and a render, screenshot, and visual judgement loop that gates completion at meta.viability.
• Raised visual-design-system-extractor to metadata.version: "0.2.0". Kept reify, skill-factory, and starting-point at metadata.version: "0.1.0".

GitHub release 0.1.3

• Added the skill-factory Agent Skill at metadata.version: "0.1.0" with the full registry eval artifact set and source-lineage hashes.
• Unified the repository rules and the skill-factory generation contract in AGENTS.md.
• Raised the skill description limit to the 1024-character agentskills.io specification limit; descriptions still start with Use when.
• Adopted the one-line markdown layout for skill markdown and applied it to starting-point.
• Kept starting-point at metadata.version: "0.1.0".
• Added the platform-neutral reify skill with frozen source lineage.
• Added trigger, behavior, failure, recovery, and speed evaluations for reify.
• Updated package, plugin, MCP, skills.sh, Aider, Continue, and Hermes Agent discovery routes.
• Added the visual-design-system-extractor Agent Skill at metadata.version: "0.1.0", retrofitted from the standalone srinitude/visual-design-system-extractor package with frozen source lineage.
• Required every typeface the extractor selects to be a Google Fonts family whose rarity is measured against the live fonts.google.com/metadata/fonts feed during the run, recorded with rank, family count, percentile, trending rank, date added, source, and retrieval date.
• Added scripts/rare_google_fonts.py with catalog, discover, and verify commands, plus a live comparison inside the extraction validator that fails on stale ranks and absent families.
• Replaced the prose extraction schema with the machine readable references/extraction-schema.yaml contract, which now drives section order, confidence labels, evidence buckets, and the font rules from one file.

GitHub release 0.1.2

• Allowed verified zero-cost routes through the full ledger path.
• Added durable pending records that block automatic repayment after an ambiguous failure.
• Released rejected-request reservations only when route metadata proves no provider attempt occurred.
• Blocked unknown-price requests before creating a ledger reservation.
• Kept every skill version unchanged.

GitHub release 0.1.1

• Added a fixed-route, capped, resumable OpenRouter sweep runner.
• Added manifest-bound user approval and first-party route and cost checks.
• Removed repository validation that tied skill versions to the package version.
• Kept starting-point at metadata.version: "0.1.0".

0.1.0

Initial public package.

• Added the starting-point Agent Skill with source-lineage evidence.
• Added positive, rejection, behavior, failure, recovery, and speed evaluations.
• Added strict validators and JSON Schemas.
• Added a bundled read-only MCP server.
• Added plugin or adapter routes for ten clients.
• Added deterministic package and copy checks.