agentic-stack is a portable layer of agent memory, skills, and protocols stored in a .agent/ folder that can be used across multiple coding-agent tools. It is for people who switch between agent harnesses while keeping shared knowledge and monitoring data. The catalogue entries provide skills and rules for this shared agent layer.
Borrowing it
Nothing to install: this file belongs to codejunkie99/agentic-stack. 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/codejunkie99/agentic-stack/master/.agent/skills/evolver/SKILL.mdgit clone --depth 1 https://github.com/codejunkie99/agentic-stackWrote 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/codejunkie99/agentic-stack/evolver)<a href="https://agentmods.dev/skills/codejunkie99/agentic-stack/evolver"><img src="https://agentmods.dev/badge/skills/codejunkie99/agentic-stack/evolver/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/codejunkie99/agentic-stack/evolver"><img src="https://agentmods.dev/badge/skills/codejunkie99/agentic-stack/evolver.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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 | $0.00002 | $0.00822 |
| Opus 5.5 | $0.00001 | $0.00329 |
| Sonnet 5.5 | $0.00000 | $0.00164 |
| Haiku 4.5 | $0.00000 | $0.00082 |
Grade A, and why
evolver 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 — 63 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Evolver — change the harness only when a paired screen says it is better
Recursive self-improvement without self-deception: a candidate edits the
policy around the model (prompt, knowledge, runtime seat, config), carries
a diagnosis, and is installed only after beating its parent on the same
tasks with a 1.96 paired-t margin. Tool: tools/evolve.py.
Loop
evolve.py diagnose— failures from episodic memory grouped into (skill, pathology) with symptom, trace events, edit location, expected change. Pick the top one.evolve.py scaffold <dx_id> --where prompt --marker <TOKEN>— a candidate directory with the file(s) to edit. Editevolve/candidates/<cid>/files/.... Put the marker where the new behaviour fires (a sentence in the prompt, a log line in a seat) so the activation gate can prove the change ran.evolve.py screen <cid> --n 4— validity gate, activation gate, paired gain. Readrounds/<id>.json. No pass, no install; write a better candidate or a better eval card.evolve.py round <cid> ... --keep 2— survivors get full-training utility and enter the gene bank cell if they beat the incumbent (ties keep the incumbent). Stagnation and max-rounds stop the loop.evolve.py install <cid>— applies files to the live.agent/with a backup underevolve/backups/.--forceexists for humans and is logged.
Evals are the bottleneck
Add task cards to evolve/evals/ whenever a diagnosis has no task that
reproduces it: {id, domain, prompt, check} with contains, regex, or a
command check. Two cards minimum; more cards, less noise.
Seats
harness/seats.py splits the loop into memory / planning / capability /
action. A runtime candidate ships harness/seats.json pointing a seat at
a new module; the conductor needs no change.
Example
Pathology "rejected-tests-failing-before-push" in git-proxy → candidate
edits skills/git-proxy/SKILL.md to run the test command before git push
and prints [evolved:pretest]. Screen on 4 push-flavoured cards: gain 0.75,
t=3.0 → pass → round → install.
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 · 63 lines · 2 tokens per session scan A d53ba1f3a549
evolver is a skill published in the GitHub repository codejunkie99/agentic-stack (2,302 stars, last pushed 9d ago), licensed Apache-2.0. It adds 2 tokens to every session and 822 once invoked, about $0.0000 per session on Opus 5.5. 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 2026-10-13.
Other skills, from other repositories
debug-inference
Debug inference clients that use an attached provider and its native endpoint, including hosted APIs and host-local Ollama, vLLM, SGLang, TRT-LLM, LM Studio, or NIM. Use for provider attachment, endpoint policy, credential substitution, topology, and migration from the removed managed inference endpoint. Trigger…
torchforge-rl-training
Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.
llamaguard
Meta's 7-8B specialized moderation model for LLM input/output filtering. 6 safety categories - violence/hate, sexual content, weapons, substances, self-harm, criminal planning. 94-95% accuracy. Deploy with vLLM, HuggingFace, Sagemaker. Integrates with NeMo Guardrails.
notebooklm
Install, authenticate, troubleshoot, and operate Gemini Notebook through the notebooklm-py CLI or typed async Python API. Use for notebook and source management, grounded chat and research, and artifact generation or download when the user mentions Gemini Notebook, notebooklm-py, the notebooklm CLI, or its Python API.…
oh-my-posh
Install, configure, or troubleshoot Oh My Posh/ohmyposh: shell init, themes, segments, Nerd Font icons, and prompt setup on PowerShell, zsh, bash, or fish.
spark-engineer
Use when writing Spark jobs, debugging performance issues, or configuring cluster settings for Apache Spark applications, distributed data processing pipelines, or big data workloads. Invoke to write DataFrame transformations, optimize Spark SQL queries, implement RDD pipelines, tune shuffle operations, configure…