hyper-agentic-engineering

Hyper Agentic Engineering

Portable memory, lean context, and reliable handoffs for coding agents.

https://github.com/user-attachments/assets/aca2e712-1bbe-4c4d-97c8-1f3a47f98baa

34-second overview · Play with sound. Setup prompt below.

Public alpha · engineering first · Python 3.11+ · no runtime dependencies.

HAE adds continuity to an existing Claude Code, Codex or Copilot setup. It preserves current instructions and hooks, keeps readable Markdown memory under your control, and records what another agent needs to resume. Obsidian is an optional viewer. An optional general-work mode uses project folders and deliverables without coding requirements.

Engineering setup includes generalized working practices and task routing: assumptions, simple focused changes, debugging, verification, review and team handoffs. Optional installed skills are selected by task rather than loaded as a whole library. See native Windows setup for PowerShell commands; macOS/Linux use the commands below.

Give this to your agent

Open Claude Code, Codex or Copilot in the project you want to work on, then paste:

Set up Hyper Agentic Engineering for this project. Read and follow
https://github.com/niomartinez/hyper-agentic-engineering/blob/v0.1.0-alpha.3/SETUP.md
Use the pinned release, inspect my existing setup, and preserve my instructions,
hooks and memories. Use engineering mode unless I ask for general work.
Show a concise plan, carry out the setup I have authorized, and verify save,
recall and recovery. Ask only for consequential choices you cannot establish.
Keep providers and automatic capture off unless I choose them.

Your agent does the installation. You do not need to run the commands below or read the technical guides. It needs terminal and file access, Python 3.11+, and access to the project. If a prerequisite or native trust action needs your input, it should tell you exactly what is missing. Browser-only chats cannot install local files.

Engineering is the default. Add “Use general-work mode” for research, operations or other project work. Obsidian is optional. Cowork packages are experimental; real sandbox install/resume is not yet verified.

The agent checks existing memory tools before choosing one authority. It never migrates, disables or deletes your existing memory automatically. Existing styles stay in place; lean responses and small-change practices are available with --style lean.

Manual setup, upgrades and removal: technical guide. Native Windows: PowerShell guide. Packaged runtime and plugins: alpha release.

Optional full-team setup from development source

The published v0.1.0-alpha.3 release supports the core setup above. The full-team profile is available in development source on main; it has no tagged release and has not been validated as stable. Resolve main to an exact commit, use a reviewed checkout of that commit, and record any local changes. Do not give the alpha.3 runtime a full-team command.

Open your agent in the actual project you want to work on, then paste:

Set up Hyper Agentic Engineering full-team for my current project. Read
https://github.com/niomartinez/hyper-agentic-engineering/blob/main/SETUP.md
and docs/team-setup.md. Use development source: resolve main to an exact commit,
clone that commit into an unused tool folder outside my project and memory,
and verify its origin and commit before running it. Inspect my existing setup.
Use plan --profile
full-team for my selected hosts and verified project scope. Preserve my
instructions, hooks and memory; resolve any competing memory authority first.
Install the reviewed full skill snapshot and portable team practices, then
verify save, recall, recovery and skill discovery. Finish Obsidian setup for
this host, reporting any missing app, registration or optional CLI action.
Offer the optional graphify and ctx7 steps in SETUP.md and ask before
installing anything. Keep hooks, Jev, backups and model changes off. Do not download an app or
publish anything without authorization.

Full-team includes a pinned manifest matching 123 upstream skill entries across eleven repositories, with 119 eligible skill bundles materialized by default, a cold skill library, curated host entries, portable adapters, clean Obsidian configuration for a new vault, and Caveman/Ponytail full defaults with session switches. Graphify is model-invocable under a graphify-first rule (read or build the repository's local graph before broad file reads), Context7 supplies current library documentation through a pinned ctx7, and nine addyosmani agent-skills are task-routed. The full community Karpathy method remains explicit-request only. Optional agent steps in SETUP.md install pinned graphify and ctx7 and, only on request, a hardened MCP-only ICM store; HAE itself installs none of them and writes no MCP configuration. Four Anthropic document skills remain catalog references because their license restricts retention outside the service; native document tools handle those tasks. The author's private vault, custom release/account workflows and plugins are excluded. Full-team setup and boundaries explains what is installed and how each host is verified.

What happens during work

1. Resolve the current project once. Retrieve only relevant notes.
2. The main conversation interprets a remember/override request or milestone.
3. Local code writes a small approved packet and returns a receipt. No second language model is needed.
4. At a cutoff or handoff, save verified results and the next unfinished step.
5. Another session uses the same folder; another person receives a sanitized project handoff.

Ordinary replies do not need curated notes. Account changes do not establish authorship or change the selected vault. Historical captures never become permission or current truth.

Optional capabilities

Codex hooks require native trust review. Copilot parent stop events may contain only metadata, so direct checkpoints remain necessary. There is no background generative janitor, transcript importer, dashboard or model override.

Commands

plan / apply      Preview and apply reviewed configuration changes
context           Resolve a mapped project or report unresolved scope
memory save       Persist approved prose, with idempotent receipts
memory recall     Bounded local search and optional Jev ranking
memory health     Report missing notes, pending captures and aging review dates
memory status     Show local provider usage and capture state
memory backend    Explicitly enable Jev or return to local-only mode
handoff           Write an approved brief inside the mapped project
backup            Create an explicit text-only local archive
catalog           Inspect optional upstream sources
obsidian status   Check the vault and native Obsidian app
obsidian open     Dispatch native vault manager (current checkout)
skills route      Select a relevant installed skill, with a local workflow fallback
rollback          Reverse the latest setup transaction, preserving memory
uninstall         Remove owned integrations, preserving memory and later edits

--home and --state-dir allow disposable profiles. They are global options and go before the command. Setup uses the selected profile's standard agent folders; it does not alter CODEX_HOME, CLAUDE_CONFIG_DIR or COPILOT_HOME. Custom agent home layouts need explicit validation before use.

Build and verify

python3 -m unittest discover -s tests -v
python3 tools/build_plugins.py
python3 tools/benchmark.py

Generated self-contained skills and per-host plugins are under dist/. Source skill folders are authoring inputs; use generated packages when distributing a skill alone. Do not copy only SKILL.md and assume the runtime exists.

Read setup and recovery, memory and privacy, compatibility and evidence, upstream licenses, and release status.

The 71-test suite passed on native Windows, macOS and Linux with Python 3.11 and 3.12. Generated command execution, install/upgrade/removal, Unicode memory and scope isolation are covered. All six jobs also passed the 1,000-note recall benchmark with zero provider calls and no irrelevant metadata matches. See measured evidence for latency results and limits. Live authenticated host acceptance, Cowork support and independent adoption remain unverified; this is an early evaluation release. See security boundaries.