v0.2.0

v0.2.0 — LLM Wiki, Kiro, Gitclaw

LLM Wiki (inspired by Karpathy's LLM Wiki pattern)

Andrej Karpathy described a pattern where the LLM incrementally builds and maintains a persistent wiki from raw sources — instead of RAG's "rediscover from scratch" approach. gitagent now ships this as a first-class template:

gitagent init --template llm-wiki

Scaffolds a complete wiki-maintainer agent with three skills:

• wiki-ingest — process sources into interlinked wiki pages with cross-references
• wiki-query — search the wiki, synthesize answers, file valuable results back as new pages
• wiki-lint — health-check for contradictions, orphan pages, stale claims, and knowledge gaps

The pattern maps directly to gitagent's existing architecture: knowledge/ = raw sources, memory/wiki/ = the compiled wiki, skills/ = the operations. No new CLI features needed — it works with every adapter (Claude, Codex, Gemini, Cursor, gitclaw, etc.).

See the full example at examples/llm-wiki/.

New Adapters

• AWS Kiro CLI — export to .kiro/agents/<name>.json + prompt.md (gitagent export -f kiro)
• gitclaw SDK — full export + runner with provider:model format conversion (gitagent export -f gitclaw, gitagent run -a gitclaw)

Improvements

• extends resolution — gitagent install now clones parent agents; export -f claude-code merges parent SOUL.md, RULES.md, and skills
• --force flag for gitagent install — re-install existing dependencies
• Git source detection expanded to support Bitbucket and GitLab (not just GitHub)

Fixes

• Duplicate requiredOption crash fixed (v0.1.8 regression)
• Broken switch fallthrough in export/import commands
• Command injection prevention in install command (execFileSync)

Full Changelog

v0.1.9...v0.2.0

v0.1.9

Hotfix

• fix: duplicate requiredOption crash — v0.1.8 shipped with duplicate .requiredOption('-f, --format') in export and --from in import commands, causing Commander to throw on startup. Fixed by merging into single option with all formats listed.
• fix: broken switch fallthrough — case 'codex' was placed after default: with no break, causing unknown formats to fall through into the codex handler instead of erroring.

Fixes #53

v0.1.8

What's New

New Adapters

• Google Gemini CLI — export, run, and import (gitagent export -f gemini)
• OpenAI Codex CLI — export and import (gitagent export -f codex)
• Cursor — export to .cursor/rules/*.mdc + enhanced import from modern .mdc files (gitagent export -f cursor)
• OpenCode — fixed to use AGENTS.md + correct provider/model config format

Multi-Agent Support

• OpenClaw multi-agent workspace export — sub-agents now produce separate OpenClaw workspaces with full SOUL.md, AGENTS.md, TOOLS.md, and skills

Bug Fixes

• Claude Code hooks — fixed JSON structure to use correct nested matcher + hooks array format; fixed event mapping (SessionStart, PostToolUseFailure, SessionEnd)
• Command injection prevention — replaced execSync string interpolation with execFileSync array args across git-cache and registry-provider
• Temp file cleanup — process.exitCode instead of process.exit() so runner finally blocks actually execute
• Shared compliance — extracted buildComplianceSection to src/adapters/shared.ts (dedup from copilot + opencode)

Docs

• Added regulated-industries guide — structural fit of gitagent for compliance-heavy domains (financial close, legal, healthcare, insurance)

Full Changelog

v0.1.7...v0.1.8

v0.1.7 — First Public Release

gitagent v0.1.7 — First Public Release 🚀

A git-native, framework-agnostic open standard for defining AI agents.

Your git repository is the agent. Clone a repo, get an agent.

What's in this release

Core CLI

• gitagent init — Scaffold agents from templates (minimal, standard, full)
• gitagent validate — Schema + compliance validation
• gitagent run — Clone any agent repo and run it instantly
• gitagent export — Export to 8 target formats
• gitagent import — Import from Claude, Cursor, CrewAI
• gitagent audit — Generate compliance reports
• gitagent skills — Search, install, list, inspect skills
• gitagent lyzr — Deploy agents to Lyzr Studio

Adapters / Runners

• Claude Code — Full interactive mode with skills, tools, permission modes
• OpenAI Agents SDK — Python code generation with typed tool stubs
• CrewAI — YAML crew config with role/goal extraction
• Lyzr Studio — REST API deployment with provider mapping
• OpenClaw — Workspace generation with auth provisioning
• Nanobot — Config + system prompt for interactive sessions
• GitHub Models — Streaming chat completions
• Git (auto-detect) — Detects adapter from model/config

Agent Standard

• agent.yaml — Schema-validated manifest with model preferences, fallback chains, skills, tools, runtime config
• SOUL.md — Natural language agent identity
• RULES.md — Hard behavioral constraints
• skills/ — Reusable capability modules (Agent Skills standard)
• tools/ — MCP-compatible tool definitions
• knowledge/ — Reference documents with retrieval config
• memory/ — Persistent cross-session state with runtime subfolder
• hooks/ — Lifecycle handlers (bootstrap, teardown)
• compliance/ — Regulatory artifacts (FINRA, SEC, Fed, CFPB)

Key Fixes

• Resolved npx binary shadowing issue (node_modules .bin/claude vs system Claude CLI)
• Fixed --refresh flag to re-clone instead of git pull on shallow caches
• Correct argument ordering for Claude runner flags

Install

npm install -g @open-gitagent/gitagent

or run directly:

npx @open-gitagent/gitagent@0.1.7 run -r https://github.com/shreyas-lyzr/architect -a claude

Links

• 📄 Spec & Docs
• 📦 npm — @open-gitagent/gitagent
• 📦 npm — @shreyaskapale/gitagent

---

#gitagent #ai-agents #open-standard #git-native #framework-agnostic #claude-code #openai #crewai #lyzr #agent-skills #compliance #version-control