claude-code-skills

What's New in v2026.10.05

• Skills organized around delivery outcomes: product discovery, architecture, delivery planning, implementation, quality assurance, delivery, operations, and skill maintenance. Each skill remains independently usable.
• New capabilities: product requirements and interaction design; architecture baselines, documentation, proposals, decisions, diagrams, and migration planning; deployment engineering, incident investigation, and product outcome evaluation.
• Stronger execution and review contracts: individual checklist tracking, requirement and decision traceability, evidence tied to the actual source and environment, explicit mutation boundaries, and clearer completion verdicts.
• More focused implementation: root-cause fixes, reuse of existing capabilities, proportionate verification, and explicit limits for deliberate simplifications.
• Distribution and validation: Agent Plugins v1 manifests, aligned Claude Code and Codex catalogs, repository contract regression checks, behavioral eval cases, and a Codex eval runner.
• Breaking catalog changes: plugin families and skill invocation names changed. product-discovery-suite is now 2.0.0; the other current plugin identities enter their first tagged repository release at 1.0.0.

Behavioral model runs were not performed during this release preparation; structural and runner checks do not establish agent performance.

Current catalog · Project page

What's New in v2026.07.12

Claude-Codex Skills, rebuilt for current models

Earlier coding models needed a large orchestration and evaluation framework to follow complex workflows reliably. Current Claude and Codex models work better with focused procedural guidance.

The repository now contains 18 standalone skills in six small plugins. Each skill owns one bounded outcome and carries only the domain rules, tool-routing decisions, safety gates, evidence requirements, and residual-risk checks worth bringing into context.

Six focused plugins

• Review Suite — evidence-based plan and delivery review.
• Codebase Audit Suite — documentation, code, test, architecture, and persistence audits.
• Optimization Suite — measured performance, dependency, modernization, and benchmark workflows.
• Testing Suite — risk-based test planning and executable acceptance evidence.
• Product Discovery Suite — opportunity evaluation before implementation.
• Maintainer Suite — skill review, repository publication, tagged releases, and community announcements.

Maintainer workflows are back

Four repository-maintenance capabilities return in a smaller, standalone form:

• Skill Reviewer checks triggering, behavioral contracts, marketplace parity, validation evidence, and publication readiness.
• Repository Publisher validates the complete Git scope, pushes approved changes, waits for CI, and verifies installation from the public remote.
• Release Publisher prepares evidence-backed release notes and requires approval of the exact tag before publication.
• Community Announcer fact-checks commands, paths, counts, and claims before publishing to GitHub Discussions.

What was removed

The repository no longer ships:

• coordinator and worker hierarchies;
• MCP servers or MCP configuration;
• shared prompt registries and distributed skill copies;
• orchestration and evaluation runtimes;
• generated host-specific skill copies.

Skills remain instruction-only by default. References, scripts, or assets are added only when a concrete recurring failure proves that a checklist is insufficient.

[!IMPORTANT] The marketplace identifier remains levnikolaevich-skills-marketplace, so existing marketplace registrations continue to resolve the same catalog. Previously cached legacy plugins may remain installed locally even though they are no longer part of the current catalog.

Install

Claude Code:

/plugin marketplace add levnikolaevich/claude-code-skills
/plugin install review-suite@levnikolaevich-skills-marketplace
/plugin install maintainer-suite@levnikolaevich-skills-marketplace

Codex:

codex plugin marketplace add levnikolaevich/claude-code-skills
codex plugin add review-suite@levnikolaevich-skills-marketplace
codex plugin add maintainer-suite@levnikolaevich-skills-marketplace

Full catalog and installation commands: README.md

What's New in v2026.05.06

hex-relay: standalone Telegram control-plane

The relay-bot used to live inside skills-catalog/ln-030-vps-bootstrap/references/relay-bot/ — a templated artifact rendered at install time. This release lifts it out as a first-class TypeScript product at agents/hex-relay/ with its own README, redeploy runbook, and operator runbook. The skill now points at the deployed agent rather than shipping its source.

Typed contracts everywhere — Fastify gained a Zod type provider; every route now declares request/response schemas, errors flow through structured validation responses, and /health exposes both runtime status and buildInfo (relay schema + package version). A reusable runProcess utility with timeout handling replaces ad-hoc child_process logic in systemd godStatus and tmux pane code.

Stronger env validation — Zod schemas for unit fragments, linux usernames, absolute POSIX paths, and optional hostnames replace loose string checks across src/config/env.ts. Misconfigurations surface at boot, not at first request.

Inbound durability and the control lane — relay-bot v6 introduces a serialized control lane (run_control) that gates session mutations and inbound delivery. The DB schema gained kind, media_path, attempts, next_attempt, delivered_at, todo_state, health_snapshots, task_poll_state, and ownership-aware session columns. Inbound rejects, terminal statuses (failed/abandoned), and token-bucket rate limiting harden the Telegram bridge.

Local voice transcription — optional ffmpeg + whisper.cpp pipeline (RELAY_VOICE_TRANSCRIPTION, WHISPER_CPP_BIN, WHISPER_CPP_MODEL). The inbound handler accepts voice messages, validates size/duration, stores media, and inserts a transcribing row; a VoiceTranscriptionWorker polls those rows, runs local transcription, promotes to queued with transcript or marks rejected. Service unit MemoryMax raised to 2G to fit the model.

Task polling and handoff — dispatch scheduling moved off tmux-injection. A 15-minute /tasks/poll route fetches provider tasks, persists state in task_poll_state, and notifies the primary operator at most once per 24h. Telegram /tasks command, task-card keyboard, and per-user handoff handlers are wired end-to-end.

Per-user god-sessions: real isolation

god-session is now a systemd template unit (%-style @<telegram_user_id>). Each Telegram user gets their own instance with per-user runtime files under /var/lib/${PROJECT_NAME}/users/<id>, scoped tmux sockets and panes, and a bubblewrap-based agent-sandbox.sh that enforces project + user work-plane isolation while binding only read-only auth credentials from the shared BOT_USER. agent-update restarts every active *-god@*.service. Codex sandbox defaults hardened to workspace-write.

Robust resume — god-session.sh now prefers explicit --resume from last-session.id or the newest jsonl by mtime, never silently spawning a fresh session. Failed resumes move the failing pointer aside, classify the failure (auth_failed, etc.), and emit enriched JSON error events with project/service/user/session/runtime metadata. The errorAlerter worker formats those into actionable admin messages.

ln-030-vps-bootstrap: new skill

A new L3 standalone skill in setup-environment documents a one-shot, idempotent VPS bootstrap for Claude Code + Codex agent workloads. Per-project envsubst (PROJECT_NAME, SERVICE_PREFIX, BOT_USER, ...), REPO_URL/REPO_REF verified clone semantics, configurable RELAY_HOOK_PORT and DISPATCH_COMMAND_NAME, idempotent .env.local seeding via append helper, and a register-telegram-commands script that registers commands both globally and scoped to all_private_chats.

The fleet registry is VPS-local: ops/environments/ in the repo holds templates only; the live registry lives at /etc/agent-fleet/environments on the VPS, owned by the host. Plan/apply artifacts

What's New in v2026.04.21

hex-line: one grammar to rule every response

Previous versions spoke three dialects at once — prose headers, emoji warnings, and JSON envelopes — each accumulated release by release. Agents had to reparse the shape on every call. This release collapses everything into a single Response grammar defined in PROTOCOL.md: one action-line, then #section headings, .entry primary rows, !detail warnings, and > executable follow-up pointers. Nothing else.

The cleanup is not cosmetic. Net output drops -37% bytes across the tool surface with no change to response shape — smaller context for the agent, faster reads, identical semantics.

Structural safety — edit_file now auto-strips boundary-echo while preserving real closing delimiters (}, ), ], });). The "agent accidentally drops the outer scope's trailing brace" bug — long a quiet corruption source on nested blocks — is fixed. A post-edit lexical brace-balance check runs on every mutation, gated on structural-risk signals so strings, templates, and JSX don't trigger false positives.

replace_between hardening — the most conflict-prone edit primitive gained an optional range_checksum. Pass it from a fresh read and the tool validates the exact bytes being replaced; ambiguous-delimiter anchors emit lone_delimiter_anchors warnings in conservative mode, or surface CONFLICT before write in strict mode.

Quality of life — new warnings array on edit_file output, compact space-separated verify entries, reduced post-edit context payload, throttled graph_fix hints, and a fix so suggested_read_call emits file_path (the parameter read_file actually accepts).

hex-line v1.17.0 → v1.27.0

hex-graph: text-only protocol

Hex-graph used to return both a JSON structuredContent envelope and a content[0].text summary — two payloads describing the same result. This release retires structuredContent and outputSchema entirely. Every response is a single content[0].text body rendered through the shared action-line grammar, identical in shape to hex-line. One protocol. One parser. Half the bytes.

Pointers that actually resolve — expansion_hint follow-up pointers now carry path plus the canonical selector chain (symbol_id > workspace_qualified_name > qualified_name > name+file). When an agent expands a result, it retries with the exact identifier that produced the row — no more fuzzy re-lookups on common names.

Up-front resolution quality — a new #quality section on every query exposes the query family, support tier, languages, frameworks, quality basis, and known-limitations count. Clients see which resolution tier the answer came from before reading the body.

102/102 semantic tests pass against the new protocol.

hex-graph v0.13.0 → v0.15.0

hex-ssh: schema matches behavior

Every hex-ssh tool advertised execTimeoutMs, transferTimeoutMs, connectTimeoutMs, and keepaliveIntervalMs — but the runtime silently ignored execTimeoutMs on upload/download and transferTimeoutMs on exec-path tools. The schema lied.

The monolithic connection props were split into baseConnProps + execTimeoutProps + transferTimeoutProps. Exec-path tools (remote-ssh, ssh-read-lines, ssh-edit-block, ssh-search-code, ssh-write-chunk, ssh-verify) now expose only execTimeoutMs. Transfer tools (ssh-upload, ssh-download) expose only transferTimeoutMs. All tools also gained per-call overrides for connect/keepalive so slow hosts no longer need global config changes.

hex-ssh v1.4.2 → v1.6.0

Gemini dropped from advisor role

Codex is now the sole external parallel-advisor for skill groups 200/300/310/510. Fallback is Opus self-review. Gemini CLI stays a first-class install/sync target — the installer still provisions it, syncer still writes GEMINI.md — but it no longer participates in multi-agent reviews. Cleanup landed across agent_registry.json, agent_runner.mjs, and the s