From 16ff84afa19ce7acb37335245c3c1a18777e4f53 Mon Sep 17 00:00:00 2001 From: christopherlhammer11-ai Date: Sat, 14 Mar 2026 09:49:10 -0700 Subject: [PATCH] feat: add Tool Use Guardian + RecallMax (free agent skills) (#298) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * feat: add Tool Use Guardian + RecallMax — free agent reliability & memory skills Two free skills from the Genesis Agent Marketplace: - **Tool Use Guardian**: Intelligent tool-call reliability wrapper. Monitors, retries, fixes, and learns from tool failures. Auto-recovers from truncated JSON, timeouts, rate limits, and mid-chain failures. - **RecallMax**: God-tier long-context memory. Injects 500K-1M clean tokens, auto-summarizes with tone/intent preservation, compresses 14-turn history into 800 high-density tokens. Both are free forever and work with Claude, Codex, Gemini, Copilot, Cursor, Windsurf, Antigravity, and any SKILL.md-compatible agent. Install: npx skills add christopherlhammer11-ai/tool-use-guardian npx skills add christopherlhammer11-ai/recallmax * docs: add source credits for genesis skills --------- Co-authored-by: Mia Hammer Co-authored-by: sck_0 --- README.md | 2 + skills/recallmax/SKILL.md | 77 ++++++++++++++++++++++++++++ skills/tool-use-guardian/SKILL.md | 84 +++++++++++++++++++++++++++++++ 3 files changed, 163 insertions(+) create mode 100644 skills/recallmax/SKILL.md create mode 100644 skills/tool-use-guardian/SKILL.md diff --git a/README.md b/README.md index 2464ebb6..d55f2aaa 100644 --- a/README.md +++ b/README.md @@ -431,6 +431,8 @@ This collection would not be possible without the incredible work of the Claude - **[Xquik-dev/x-twitter-scraper](https://github.com/Xquik-dev/x-twitter-scraper)**: X (Twitter) data platform — tweet search, user lookup, follower extraction, engagement metrics, giveaway draws, monitoring, webhooks, 19 extraction tools, MCP server. - **[shmlkv/dna-claude-analysis](https://github.com/shmlkv/dna-claude-analysis)**: Personal genome analysis toolkit — Python scripts analyzing raw DNA data across 17 categories (health risks, ancestry, pharmacogenomics, nutrition, psychology, etc.) with terminal-style single-page HTML visualization. - **[AlmogBaku/debug-skill](https://github.com/AlmogBaku/debug-skill)**: Interactive debugger skill for AI agents — breakpoints, stepping, variable inspection, and stack traces via the `dap` CLI. Supports Python, Go, Node.js/TypeScript, Rust, and C/C++. +- **[christopherlhammer11-ai/tool-use-guardian](https://github.com/christopherlhammer11-ai/tool-use-guardian)**: Source for the Tool Use Guardian skill — tool-call reliability wrapper with retries, recovery, and failure classification. +- **[christopherlhammer11-ai/recallmax](https://github.com/christopherlhammer11-ai/recallmax)**: Source for the RecallMax skill — long-context memory, summarization, and conversation compression for agents. ### Inspirations diff --git a/skills/recallmax/SKILL.md b/skills/recallmax/SKILL.md new file mode 100644 index 00000000..4e1950b0 --- /dev/null +++ b/skills/recallmax/SKILL.md @@ -0,0 +1,77 @@ +--- +name: recallmax +description: "FREE — God-tier long-context memory for AI agents. Injects 500K-1M clean tokens, auto-summarizes with tone/intent preservation, compresses 14-turn history into 800 tokens." +category: memory +risk: safe +source: community +date_added: "2026-03-13" +author: christopherlhammer11-ai +tags: [memory, context, rag, summarization, compression, long-context, agent-infrastructure] +tools: [claude, cursor, codex, gemini, copilot, windsurf, antigravity, grok] +--- + +# RecallMax — God-Tier Long-Context Memory + +## Overview + +RecallMax enhances AI agent memory capabilities dramatically. Inject 500K to 1M clean tokens of external context without hallucination drift. Auto-summarize conversations while preserving tone, sarcasm, and intent. Compress multi-turn histories into high-density token sequences. + +Free forever. Built by the Genesis Agent Marketplace. + +## Install + +```bash +npx skills add christopherlhammer11-ai/recallmax +``` + +## When to Use This Skill + +- Use when your agent loses context in long conversations (50+ turns) +- Use when injecting large RAG/external documents into agent context +- Use when you need to compress conversation history without losing meaning +- Use when fact-checking claims across a long thread +- Use for any agent that needs to remember everything + +## How It Works + +### Step 1: Context Injection + +RecallMax cleanly injects external context (documents, RAG results, prior conversations) into the agent's working memory. Unlike naive concatenation, it: +- Deduplicates overlapping content +- Preserves source attribution +- Prevents hallucination drift from context pollution + +### Step 2: Adaptive Summarization + +As conversations grow, RecallMax automatically summarizes older turns while preserving: +- **Tone** — sarcasm, formality, urgency +- **Intent** — what the user actually wants vs. what they said +- **Key facts** — numbers, names, decisions, commitments +- **Emotional register** — frustration, excitement, confusion + +### Step 3: History Compression + +Compress a 14-turn conversation history into ~800 high-density tokens that retain full semantic meaning. The compressed output can be re-expanded if needed. + +### Step 4: Fact Verification + +Built-in cross-reference checks for controversial or ambiguous claims within the conversation context. Flags contradictions and unsupported assertions. + +## Best Practices + +- ✅ Use RecallMax at the start of long-running agent sessions +- ✅ Enable auto-summarization for conversations beyond 20 turns +- ✅ Use compression before hitting context window limits +- ✅ Let the fact verifier run on high-stakes outputs +- ❌ Don't inject unvetted external content without dedup +- ❌ Don't skip summarization and rely on raw truncation + +## Related Skills + +- `@tool-use-guardian` - Tool-call reliability wrapper (also free from Genesis Marketplace) + +## Links + +- **Repo:** https://github.com/christopherlhammer11-ai/recallmax +- **Marketplace:** https://genesis-node-api.vercel.app +- **Browse skills:** https://genesis-marketplace.vercel.app diff --git a/skills/tool-use-guardian/SKILL.md b/skills/tool-use-guardian/SKILL.md new file mode 100644 index 00000000..179ace63 --- /dev/null +++ b/skills/tool-use-guardian/SKILL.md @@ -0,0 +1,84 @@ +--- +name: tool-use-guardian +description: "FREE — Intelligent tool-call reliability wrapper. Monitors, retries, fixes, and learns from tool failures. Auto-recovers from truncated JSON, timeouts, rate limits, and mid-chain failures." +category: reliability +risk: safe +source: community +date_added: "2026-03-13" +author: christopherlhammer11-ai +tags: [reliability, tool-use, error-handling, retries, recovery, agent-infrastructure] +tools: [claude, cursor, codex, gemini, copilot, windsurf, antigravity] +--- + +# Tool Use Guardian + +## Overview + +The reliability wrapper every AI agent needs. Monitors tool calls, auto-retries failures, fixes truncated responses, and learns which tools are unreliable — so you never lose your chain of thought. + +Free forever. Built by the Genesis Agent Marketplace. + +## Install + +```bash +npx skills add christopherlhammer11-ai/tool-use-guardian +``` + +## When to Use This Skill + +- Use when tool calls return truncated or malformed JSON +- Use when APIs timeout or rate-limit your agent mid-task +- Use when a multi-step chain breaks partway through +- Use when you need automatic retry logic without writing it yourself +- Use for any agent workflow that depends on external tool reliability + +## How It Works + +### Step 1: Pre-Call Validation + +Before every tool call, Guardian validates: +- Required parameters are present and correctly typed +- The tool is not marked as "unreliable" from previous failures +- Request size is within known limits + +### Step 2: Failure Classification + +When a tool call fails, Guardian classifies the failure into one of 9 categories: + +| Failure Type | Recovery Action | +|---|---| +| Truncated JSON | Re-fetch with pagination or smaller chunks | +| API Timeout | Retry once with simpler request, then decompose | +| Rate Limit (429) | Exponential backoff, max 3 retries | +| Auth Expired | Flag for user intervention | +| Mid-chain Break | Resume from last successful checkpoint | +| Error-as-200 | Detect `{"error": "..."}` disguised as success | +| Schema Mismatch | Attempt auto-coercion, warn if lossy | +| Network Failure | Retry with jitter, max 2 attempts | +| Unknown Error | Log full context, escalate to user | + +### Step 3: Chain Protection + +For multi-step tool chains, Guardian maintains checkpoints. If step 4 of 7 fails, it resumes from step 4 — never restarts from scratch. + +### Step 4: Learning + +Guardian tracks failure patterns per tool. After 3+ failures of the same type, it marks the tool as unreliable and suggests alternatives. + +## Best Practices + +- ✅ Let Guardian wrap all external tool calls automatically +- ✅ Review Guardian's reliability reports to identify flaky tools +- ✅ Use checkpoint recovery for long chains +- ❌ Don't disable retry logic for rate-limited APIs +- ❌ Don't ignore repeated failure warnings + +## Related Skills + +- `@recallmax` - Long-context memory enhancement (also free from Genesis Marketplace) + +## Links + +- **Repo:** https://github.com/christopherlhammer11-ai/tool-use-guardian +- **Marketplace:** https://genesis-node-api.vercel.app +- **Browse skills:** https://genesis-marketplace.vercel.app