Changelog

All notable changes are documented here. Format: Keep a Changelog

---

[Unreleased]

[1.11.1] — 2026-09-01

Added

• skill-densify Micro-Kernel Enhancements:
  • Anti-Lecturing Heuristics & docs/ vs skills/ Law: Added strict operational boundary separating human narrative/background documentation (docs/) from agent operational syntax matrices (skills/), eliminating textbook lecturing and definition preambles.
  • Expansion Handoff Contracts: Formalized companion skill delegation tables to prevent single-responsibility kernel bloat when handling peripheral operations.
  • Direct Negative Constraints: Prioritized direct "DO NOT" prohibitions and prompt-leak prevention rules.
  • densify.py Analyzer Upgrades: Added automated detection for pedagogical lecturing and textbook fluff, validation of direct negative constraints and expansion handoffs, and updated OPTIMAL density classification score for structured & disclosed micro-kernels.
  • Graduated Size Bands: Calibrated THIN, COMPACT, NORMAL, LARGE, REVIEW, and OPTIMAL thresholds to evaluate structural density rather than arbitrary line counts.
• Antigravity Multi-Surface Symlink Guidance: Documented explicit setup and commands for symlinking local plugin development repositories into both Antigravity 2.0 Desktop (~/.gemini/config/plugins/<name>) and agy CLI (~/.gemini/antigravity-cli/plugins/<name>).
• Antigravity Desktop Startup Discovery Notes: Clarified that Antigravity Desktop scans ~/.gemini/config/plugins/ on launch and requires an application restart to discover newly installed or updated plugins.

Fixed

• doctor User-Scope Scan: Added Antigravity roots (~/.gemini/config/plugins, ~/.gemini/antigravity-cli/plugins, and ~/.gemini/antigravity/skills) to the doctor --scope user dead-symlink inspection loop.
• densify.py Metric Drift: Fixed threshold/display mismatch (70 vs 65) and excluded trigger frontmatter from compression penalties.

[1.11.0] — 2026-08-21

Added

• skill-densify skill: High-density 5-layer micro-kernel compressor that transforms verbose conversational skills into ≤65-line kernels with compact decision tables, mechanical script offloading, silent fast-lanes, and progressive disclosure boundaries. Includes densify.py automated density analyzer and token budget auditor.
• scripts/skizl-ops.sh: Deterministic mechanical helper script providing single-turn CLI execution for sym-status, doctor, diffsum, and guard-check (saving up to 96% tokens on exploration turns).

Changed

• Micro-Kernel Refactoring across all bundle skills:
  • skill-manager: Compressed from 291 lines down to 56 lines (~80% reduction) using telegraphic micro-kernel routing and silent fast-lane execution format.
  • skill-forge: Compressed from 195 lines down to 69 lines (~65% reduction) with tabular operation/track matrices.
  • skill-draft: Compressed from 163 lines down to 48 lines (~70% reduction) into a compact single-pass decision engine.
• Antigravity Safeguards: Documented critical warning against executing agy plugin install on a symlinked plugin target (destructive towards link target). Either symlink/clone directly or install, never both.

[1.10.1] — 2026-08-17

Added

• sym declare / sym json — registers skills through Antigravity's declarative .agents/skills.json (and .agents/plugins.json) instead of symlinks. Full schema documented: entries, inherits, include_only, exclude, and the absolute / home-relative / repo-root path resolution rules. Use where git will not store symlinks, on Windows, or to commit the registration so teammates get skills on clone. Merges into an existing file rather than overwriting.
• publish documents Antigravity's plugin enable/disable model: "disabled": true ships a plugin switched off, while the user's choice is recorded in config.json under a plugins map keyed by directory name and always wins. Notes the trade-off — disabled is rejected by the closed Agent Plugins schema, so it belongs only in an Antigravity-specific manifest (verified: check.py --plugin errors on it, agy accepts it).

Changed

• agents/ and commands/ are now marked CLI-verified only. Both load under agy 1.1.13, but neither appears in the Antigravity 2.0 plugin specification, which documents four components (skills, rules, hooks, MCP servers). Load-bearing behavior belongs in skills/, which both surfaces are specified to read.
• platforms.md records that Antigravity's authoritative customization reference ships inside the app at ~/.gemini/antigravity/builtin/skills/agy-customizations/ and is more complete than the public docs site — the declarative-config schema appears only there. It also notes that Agent Plugins conformance is now confirmed on the Antigravity 2.0 desktop app (2.8.1), not just the agy CLI: installed at ~/.gemini/config/plugins/, all three skills load and are listed as Plugin: skizl.

[1.10.0] — 2026-08-17

Added

• Agent Plugins 1.0.0 conformance. The root plugin.json is now a conformant Agent Plugins manifest ($schema: https://agent-plugins.org/schemas/1.0.0/plugin.schema.json) — the open, vendor-neutral packaging standard governed by AWS, Cursor, Microsoft, OpenAI, and Vercel, with Google as Core Maintainer. This makes skizl installable on ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code from one manifest.
• git-guard now checks the root plugin.json for version drift, bringing the count to seven version sites. The Agent Plugins schema makes version optional, so the file is only checked when it declares one.
• Plugin-layer validation. check.py --plugin validates an Agent Plugins root: manifest $schema, the name pattern, the closed top-level field set, author sub-keys, keywords typing, and reverse-domain extensions namespacing. It then validates mcp.json — explicit per-server type, transport-required fields, $schema version agreement with plugin.json, and misuse of PLUGIN_ROOT/PLUGIN_DATA in command, URLs, or headers, where they never expand — and finally each discovered skill. Failure isolation follows the spec: a fatal manifest error rejects the plugin, but one broken skill or server never disables the others.
• doctor gained four manifest-conformance checks (non-conformant manifest, undiscoverable skill, untyped MCP server, dead plugin variable) and a section documenting the --plugin run and what it deliberately does not cover.
• doctor gained a vendor component health pass for Antigravity: root hooks.json and mcp_config.json must parse, and every hook script must exist and be executable. A hook script that is missing or non-executable fails silently at run time, so nothing else surfaces it. Bare command names are reported as warnings only, since PATH differs on the target machine.
• publish now documents Antigravity's four auto-loaded root components — skills/, rules/*.md, hooks.json, mcp_config.json — with templates for each, and states plainly that none is interchangeable with the Claude, Codex, or Agent Plugins equivalent. They stay opt-in: empty scaffolds are noise.
• platforms.md gained an Agent Plugins packaging section establishing the standard as the portable packaging target, plus guidance on carrying Antigravity as a namespaced extensions entry rather than a competing root manifest.

Changed

• The root plugin.json previously carried Antigravity's private schema. Since only one $schema value can occupy the field, the Agent Plugins schema takes it and Antigravity-specific data moves to extensions["com.google.antigravity"] — the standard's reverse-domain escape hatch for client-specific data. Verified with agy plugin validate (CLI 1.1.13): the conformant and the old Antigravity-style manifest validate identically ([ok], all skills processed), while a missing or unparseable plugin.json errors. Antigravity ignores $schema entirely, so conformance costs nothing there.
• publish now scaffolds portable-first: the Agent Plugins manifest is the core, and .claude-plugin/ / .codex-plugin/ are documented as vendor fallbacks for hosts that have not adopted the standard. publish.md also records the closed-schema field list, the name pattern constraint, the extensions namespacing rule, and the boundary between the portable surface (skills/, mcp.json) and the vendor-only surface (agents/, commands/, hooks/).
• Antigravity install guidance now distinguishes two products with separate state trees. The published plugin docs sit under the Antigravity 2.0 section, so their ~/.gemini/config/plugins/ path describes the desktop app; the agy CLI installs to ~/.gemini/antigravity-cli/plugins/, and the two keep independent skill directories. Both surfaces read the workspace path .agents/plugins/<name>/, now documented as the preferred install location.
• CLAUDE.md documents all three bundled skills (skill-manager, skill-forge, skill-draft) instead of describing skizl as a single-skill repository, states the authoring/lifecycle boundary between them, and adds a skill-forge architecture section covering package layout, the operation-vs-track distinction, check.py profile guidance, and the six-brief regression protocol.

Fixed

• Incomplete Antigravity install instructions in publish.md, doctor.md, and README.md. Each named a single global root when there are two — one per surface — so a user following them could install into the tree their Antigravity does not scan. All three now list the desktop root, the CLI root, and the workspace path both surfaces read.
• publish.md claimed Antigravity ignores agents/. It does not: agy plugin validate confirms agents/ and commands/ both load, and commands/ entries are converted to skills. Neither appears in the published documentation.
• publish.md claimed the root plugin.json carries no version and there was "nothing to bump". It does now, and git-guard checks it — corrected in the release checklist.
• platforms.md replaced the stale gemini row (which cited geminicli.com and a .gemini/skills/ root) with the documented Antigravity discovery roots and the official antigravity.google references. Also corrected the frontmatter guidance: Antigravity documents only name and description, and name defaults to the folder name.

Notes

• Verified on the agy CLI (1.1.13) only. The Antigravity 2.0 desktop app has a separate state tree and was not exercised — its plugin loading is documented but untested here.
• Declarative .agents/skills.json / plugins.json (with entries/inherits/include_only/exclude) is not implemented: the format does not appear in the published Antigravity skills, plugins, or CLI reference documentation. Revisit if a specification surfaces.

[1.9.0] — 2026-08-14

Added

• Skill Forge operations (Create / Update / Resume) — a lifecycle dimension orthogonal to the existing weight tracks. The operation decides where a run starts; the track still decides how much evidence and process the work needs. A completed skill receiving a new request is an Update, not a Resume.
• Forge record: compact continuation state (operation, contract, architecture decisions, track, completed gates, feedback deltas, review issue dispositions, evidence, next action) kept outside the distributable skill folder, so process state is never mistaken for runtime content. A new session resumes from the first incomplete gate instead of restarting or relitigating settled decisions.
• Architecture workshop: a pre-draft pass over boundary, capability clusters, runtime modes and branches, rules, routing, and only then package shape. Every proposed file must name the branch that reads it and the behavior it changes. Deliberately scale-aware — for a Light skill it stays one compact paragraph and is never expanded into a separate artifact.
• Two-phase review: the reviewer's single prompt splits into a Full-review prompt and an issue-scoped Verification prompt that checks disposition of a closed issue list without reopening general review.
• Issue severity and stable IDs — blocking vs advisory, with an explicit calibration rule tied to concrete consequence rather than section label. Two reviewer calls are the default maximum; a third requires a major structural rewrite and a recorded reason.
• extend mode on the audit track, for a completed skill receiving a new behavioral delta.
• Sixth regression scenario covering the Resume operation: evals/brief-resume.md, evals/prompts/resume.md, and fixtures (resume-forge-record.yaml, resume-repaired/), plus a held-out trigger case for resuming an unfinished review.

Changed

• The standard track reorders around the workshop: contract and architecture map are confirmed first, routing is validated before prose, and the package plan is committed as its own step (8 steps → 9). The empirical track's cross-reference follows to "steps 1–8".
• Reviewer checks expanded with failure recovery, safe resume and rerun idempotency, explicit approval boundaries before consequential actions, tool availability and declared fallbacks, portable path and environment assumptions, output-contract explicitness, evidence-backed completion claims, and proportional evaluation effort with a stopping condition.
• pass now always pairs with right-sized — an underbuilt or overbuilt verdict requires a blocking issue naming the missing behavior or unjustified process weight. Blockers are never truncated; advisories cap at five.
• Every regression brief now asserts the selected operation before the track, and the light and standard briefs test that workshop output stays proportional to the decision surface.
• Skill Forge bumped 1.1.0 → 1.4.0. Its description now covers create, update, resume, audit, debloat, repair, and activation-boundary tuning. (1.2.0 and 1.3.0 were development increments and were never released separately.)

Fixed

• .gitignore now excludes local secrets (.env*, with *.example re-included), Python build artifacts, Skill Forge continuation state (/.skill-forge/), and local feedback, planning, and article drafts that were never part of the plugin.

[1.8.0] — 2026-08-13

Added

• Skill Forge host profiles for the portable Agent Skills specification plus Claude Code, Codex, Cursor, Gemini CLI, and skizl conventions, with official documentation links and explicit portability rules.
• Executable Skill Forge regression fixtures, isolated prompt briefs, a held-out trigger suite, and profile-aware validation for portable and host-specific packages.

Changed

• Skill Forge now recovers demonstrated workflows before interviewing, preserves descriptions for manual-only skills, chooses instruction rigidity by task fragility, applies least-surprise security rules, and evaluates empirical skills against isolated baselines.
• The fresh-context reviewer now checks host compatibility, invocation behavior, permissions, script dependencies, and evidence quality in addition to sizing and progressive disclosure.
• Expanded Skill Forge guidance, templates, and evals from four illustrative briefs to a repeatable five-scenario regression protocol covering light, standard, empirical, repair, and hardening workflows.

Fixed

• Replaced shell-looking skizl lifecycle examples with an explicit handoff to the installed skill-manager skill; skizl is a plugin/meta-skill namespace, not a CLI binary.
• Corrected host-specific invocation guidance so portable skills stay portable and vendor-only fields are emitted only for their target host.

[1.7.0] — 2026-07-12

Added

• release command (ship): an atomic, verified, and resumable plugin release workflow supporting scope checks, auth preflights, test suites, and distribution verification.

Changed

• Upgraded git-guard pre-commit hook to support parsing metadata.version in skill frontmatter with fallback to top-level version:, and validation of arbitrary CLI version constants (e.g. SPECTACULAR_VERSION in cli/spectacular).

[1.6.1] — 2026-07-08

Added

• skill-forge skill — a standalone skill-that-writes-skills: a lean router (SKILL.md) over four weight tracks (light / standard / empirical / audit), a fresh-context reviewer gate, a co-located GLOSSARY.md, minimal templates, and a mechanical check.py lint. Self-contained: the empirical track runs its own with-skill-vs-baseline eval loop rather than depending on another skill.
• skill-draft skill — the fast lane: sketch a small skill in one pass (single SKILL.md, no tracks/reviewer/evals), meant to pair with skill-forge for the heavy build.

Changed

• Renamed the packaged skill skizl → skill-manager: folder skills/skizl/ → skills/skill-manager/ and frontmatter name: updated. Invocation is unchanged — the plugin namespace stays skizl, so /skizl <command> and all skizl <verb> commands work as before.
• Leaner, non-overlapping descriptions across the three skills to stop them colliding on triggering: skill-draft = quick single-file draft, skill-forge = full tracked build with reviewer + evals (each names the other as the off-ramp), skill-manager = post-authoring lifecycle only (no "create" verb; dropped the redundant trigger list). skill-manager bumped to 1.6.0.
• skill-forge learns the draft→forge handoff: its audit track now forks into repair mode (trim/fix an aged or bloated skill) and harden mode (build up a validated fast draft, e.g. a skill-draft output plus human feedback). In harden mode, thinness is the starting point, not a defect — no ceremony is added unless a branch, a failing test, or the feedback demands it. Added the brief-audit-harden.md regression brief.

Removed

• Archived skill-creator, writing-great-skills, and progressive-skill-builder into the repo-root _archive/ — their doctrine (vocabulary, eval loop, scaffold weighting) is now folded into skill-forge, which no longer references them.

---

[1.6.0] — 2026-07-07

Added

• Google Antigravity plugin support: publish now scaffolds a root plugin.json manifest (conforming to the Antigravity plugin spec, including the required $schema field and description) alongside the Claude and Codex manifests, documents Antigravity's component discovery (skills/, rules/, root hooks.json, root mcp_config.json), and reports install paths — workspace .agents/plugins/ / _agents/plugins/ and global ~/.gemini/antigravity-cli/plugins/. Docs are explicit that Antigravity has no marketplace/install CLI for third-party plugins — install is always manual (clone or symlink into a scan directory).
• README: Antigravity install section; skizl repo itself now ships a root plugin.json so it installs as an Antigravity plugin directly.

Changed

• .gitignore: excluded skills/skill-creator/, skills/progressive-skill-builder/, skills/writing-great-skills/ — local-only skills not part of the skizl plugin.

---

[1.5.4] — 2026-06-20

Added

• Description-field safeguards for Codex's 1024-character skill description limit: pack now generates and reports bounded frontmatter description values, publish blocks over-limit descriptions and can produce missing ones, and doctor checks installed/source skills for missing, near-limit, or over-limit descriptions. The check applies to description alone, not description plus when_to_use.

Changed

• Shortened the active skizl skill description so it stays comfortably under the Codex limit.
• pin.mjs now normalizes generated shortcut descriptions and caps them before writing redirect skills.

---

[1.5.3] — 2026-05-16

Changed

• Install docs: clarified the two Codex install paths — npx codex-marketplace add … --plugin (external helper, adds marketplace and activates) vs. the native codex plugin marketplace add CLI (registers the marketplace only; activate afterward from the in-app /plugins browser).
• Corrected publish.md: codex plugin marketplace add is a valid command (previous docs said otherwise); there is still no codex plugin install subcommand.
• Restructured README install section with per-target headings and a "test locally" block.

---

[1.5.2] — 2026-05-10

Fixed

• git-guard hook: git tag is now allowed to lag behind manifest versions (commit first, tag after is normal workflow). Hook only blocks if a tag is somehow ahead of manifests — a genuinely broken state. Tag uses a new TYPES parallel array; collect accepts optional third arg "tag" to opt into this relaxed check.
• Removed the --no-verify bootstrap workaround from the install report — no longer needed.

---

[1.5.1] — 2026-05-10

Fixed

• git-guard install: gather all inputs (existing hook check + SKILL.md question) upfront before writing any files, so the install flow runs without mid-process interruptions
• git-guard install Step 4 report: added ⚠ git tag bootstrap note — on the very first commit after install, --no-verify is required once (the tag cannot exist before its commit); suggested commands included

---

[1.5.0] — 2026-05-10

Added

• git-guard command (version-guard, hook, drift): installs a pre-commit hook via core.hooksPath that collects versions from all 6 locations (manifests, README badge, CHANGELOG, git tag, optional SKILL.md), takes the highest as truth, prints a ✓/✗ table, and blocks the commit if anything diverges
• references/git-guard.md: full spec for install, remove, and check sub-commands, including the complete hook script and .git-guard.json config
• publish Step 5 now auto-runs git-guard install when no hook exists, so every published repo gets the guard automatically

Fixed

• Hook script: empty VERSIONS array + set -u could crash on bash < 4.4 — replaced ${VERSIONS[@]:-} with safe ${VERSIONS[@]+"${VERSIONS[@]}"} expansion
• Hook script: all file paths now anchored to $(git rev-parse --show-toplevel) so the hook works correctly when invoked outside the repo root
• Hook script: sort -V on macOS BSD sort misorders semver across major digit boundaries (e.g. 1.9 < 1.10) — added gsort -V detection with a pure-bash semver fallback
• Hook script: zero version sources found now prints an explicit notice and exits 0 instead of silently claiming "all consistent"

---

[1.4.0] — 2026-05-10

Added

• snapshot command: save SKILL.md as versions/SKILL@<version>.md; auto-triggered on publish (if version: present) and on diff when local is ahead of global installed
• bump command: increment version: frontmatter (patch/minor/major or explicit), with optional snapshot after bump
• history command: list snapshots in versions/ sorted by semver descending; --show <ver> to read, --diff <ver> to diff current vs snapshot
• archive command: tar the entire skill folder as a timestamped tarball (_backups/<name>@YYYYMMDD-HHMM[--tag].tar.gz); always manual, suggested by diff on large-scale changes
• references/snapshot.md: full spec for snapshot, bump, and history sub-commands
• references/archive.md: full spec for archive including intent-detection table and large-diff heuristic

Changed

• diff now offers to snapshot when local version is ahead of global installed
• publish reference updated: source.path corrected to "./" (relative to repo root), hybrid layout removed, Codex install commands corrected to npx codex-marketplace add ... --plugin

---

[1.3.0] — 2026-05-07

Added

• .codex-plugin/plugin.json enriched with optional interface block (displayName, shortDescription, longDescription, developerName, category, capabilities, websiteURL, defaultPrompt, brandColor) for Codex browser UI
• publish reference and create-plugin skill updated to document the interface block as optional, with warning against adding privacyPolicyURL/termsOfServiceURL unless those files exist

Changed

• publish reference updated with verified Claude Code and Codex install methods from official docs
• Claude Code install key format clarified as {plugin-name}@{marketplace-key} — {marketplace-key} is the top-level "name" in .claude-plugin/marketplace.json, {plugin-name} is plugins[0].name; they are often the same but can differ (e.g. codex@openai-codex)
• README install command corrected from skizl@alexsmedile-skizl to skizl@skizl
• Quick Start examples replaced with generic writing container (removed project-specific cs container)
• Icon replaced with v2 squircle design matching Blip-style app icon aesthetic

Fixed

• .codex-plugin/plugin.json missing required version, description, and skills fields
• .agents/plugins/marketplace.json source.path corrected from "./" to "../../" (resolves relative to .agents/plugins/)

---

[1.2.0] — 2026-05-07

Fixed

• Removed hardcoded GitHub username from examples in SKILL.md, references/publish.md, and references/fork.md — replaced with generic <username> placeholder
• Removed personal file paths from references/list.md output example — replaced with /path/to/...
• publish command now resolves GitHub identity via gh api user (login for URL fields, full name for author.name) with git config user.name as fallback
• author.name / owner.name in all plugin manifests corrected to full display name

---

[1.1.0] — 2026-05-07

Added

• sym command with sub-commands init / in, migrate / out, status — symlinks skills/ into .claude/skills/ and .agents/skills/ using portable relative symlinks
• list command — shows installed skills across global, project, and source locations with symlink state (linked / broken / direct / source)
• diff command — compares two versions of a skill's SKILL.md, with version field extraction from frontmatter
• doctor command — diagnoses broken symlinks, missing SKILL.md, direct directories, orphaned agent links, and unlinked source skills
• fork command — clones a skill from a local path or GitHub URL into skills/<name>/ and symlinks it; strips git history on clone
• publish command — scaffolds .claude-plugin/, .codex-plugin/, and .agents/plugins/ manifests from SKILL.md frontmatter; generates README.md if missing
• IDEAS.md — name brainstorm and action backlog

Changed

• Skill renamed from skilz to skizl
• references/ architecture replaces actions/ + knowledge/ layout
• list promoted from alias of status to standalone command
• status alias list updated (removed list, kept info)
• Command order in SKILL.md reordered: sym first, status and onboard last

[1.0.0] — 2026-05-07

Added

• Initial release: pack, unpack, pin, unpin, status, onboard commands
• scripts/pin.mjs for pin/unpin automation
• Plugin manifests for Claude Code and Codex