Agentic Context Engine is an open-source engine that gives AI agents a persistent learning loop, helping them remember successful strategies and learn from failures across sessions. It is used to improve production agents, and also powers Kayba’s hosted service. Catalogue add-ons support workflows for operating and configuring the engine.
Borrowing it
Nothing to install: this file belongs to kayba-ai/agentic-context-engine. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/kayba-ai/agentic-context-engine/main/.claude/skills/kayba-pipeline/stage-6-hitl/SKILL.mdgit clone --depth 1 https://github.com/kayba-ai/agentic-context-engineWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/kayba-ai/agentic-context-engine/stage-6-hitl)<a href="https://agentmods.dev/skills/kayba-ai/agentic-context-engine/stage-6-hitl"><img src="https://agentmods.dev/badge/skills/kayba-ai/agentic-context-engine/stage-6-hitl/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/kayba-ai/agentic-context-engine/stage-6-hitl"><img src="https://agentmods.dev/badge/skills/kayba-ai/agentic-context-engine/stage-6-hitl.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 1 finding, up to medium
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- medium Excessive Agency · line 246 Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00087 | $0.02435 |
| Opus 5 | $0.00044 | $0.01218 |
| Sonnet 5 | $0.00017 | $0.00487 |
| Haiku 4.5 | $0.00009 | $0.00244 |
Grade A, and why
kayba-stage-6-hitl scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 259 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Stage 6: Human-In-The-Loop Gate
Present the action plan with enough context for an informed decision, collect the user's approval, and record the outcome.
The goal is not rubber-stamping. The user must receive enough information to genuinely evaluate, modify, or reject the plan -- even if they have not seen Stages 1-5.
Inputs
eval/action_plan.md-- the prioritized action plan from Stage 5eval/baseline_metrics.md-- the evaluation rubric with baseline valueseval/baseline_metrics.json-- raw metric data (for exact numerator/denominator counts)eval/stage1_insights_summary.md-- original insights (for trace evidence references)
Read all four files before starting.
Process
1. Build the executive summary
Compute and present the following counts from the action plan:
- Total insights analyzed (raw count before deduplication)
- Distinct actionable items after deduplication
- Breakdown: prompt fixes, code fixes, discarded
- Discard rate with one-line reason per discard (e.g., "5ac7f4ce: efficiency optimization, conflicts with turn discipline constraint")
Format:
EXECUTIVE SUMMARY
-----------------
Insights analyzed: 19 (raw) -> 12 distinct after dedup
Actionable: 9 (8 prompt fixes, 1 code fix)
Discarded: 3 (reasons listed below)
Discards:
- 5ac7f4ce (Upfront Info Collection): conflicts with higher-priority turn discipline
- fe2d51cb (Proactive Reservation Lookup): already default behavior, no failure evidence
- 1fa1b826 (Cancellation Denial Enumeration): subsumed into cancellation checklist
2. Present the top 3 highest-impact changes
For each of the top 3 fixes by priority, present:
Before/after behavior -- use concrete examples from actual traces referenced in the insights. Quote the specific agent behavior that was wrong (before) and describe what the agent should do instead (after). Reference the trace task ID.
Target metric delta -- which metric(s) this fix targets, the current baseline value, and the expected direction. Do not fabricate precise target numbers. Use the format: "M1: 41.4% -> higher (target: 90%+)" only when the action plan provides a target; otherwise use "M1: 41.4% -> up".
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 259 lines · 87 tokens per session scan A 394383c6c5ad
kayba-stage-6-hitl is a skill published in the GitHub repository kayba-ai/agentic-context-engine (2,565 stars, last pushed 11d ago), licensed Apache-2.0. It adds 87 tokens to every session and 2,435 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
titen-memory
Use the Titen MCP server to recall bounded evidence-grounded context, record verified durable signals, submit feedback, and coordinate checkpoints, leases, or handoffs. Use when work may benefit from prior project memory or when a verified outcome should be preserved for another agent; do not use it to capture raw…
open-source
Documentation reference for writing Python code using the browser-use open-source library. Use this skill whenever the user needs help with Agent, Browser, or Tools configuration, is writing code that imports from browseruse, asks about @sandbox deployment, supported LLM models, Actor API, custom tools, lifecycle…
writing
A writing guide for turning verified facts and calculations into finished text for a specific audience. It follows the requested language, structure, and length.
mnemon
Persistent memory CLI for LLM agents. Store facts, recall past knowledge, link related memories, manage lifecycle.
goai
GoAI is a Go SDK for AI applications. One unified API across 25+ LLM providers. Inspired by the Vercel AI SDK, adapted to Go idioms (generics, interfaces, channels).
mnemo-cortex
Installs and wires Mnemo Cortex (local-first persistent memory) into OpenClaw and other MCP-capable agents. Use for cross-session recall, decision history, or multi-agent shared memory.