PAACM (Portable Adaptive Agentic Coding Method) is a method and a kit for agentic coding in Claude Code and Cursor. You give the agent a request, plus any supporting data: a reference repository, documents, screenshots, logs. You get back a committed change whose acceptance criteria are each marked passed, failed or untested with executed evidence. Or you get a precise BLOCKED: report that says what the agent needs from you.
The method is derived from a study of current agentic-coding research, vendor guidance and tooling, and specified in ALGORITHM.md.
request + supporting data
→ 0 orient → 1 intake and triage (route FAST | STANDARD | DEEP, autonomy L0–L3)
→ 2 specify → 3 explore → 4 plan → 5 tests first → 6 implement
→ 7 verify (the gate) → 8 independent review → 9 report
- One agent writes. A read-only explorer and a read-only reviewer work in fresh contexts. No swarms, and no two agents in one working tree.
- Process scales with risk. A FAST fix skips the specification, plan and review. DEEP work gets all of them, and high-stakes work also gets held-out tests.
- Done means executed evidence. The stop gate keeps the agent from finishing while the checks in
.agent/checks.conffail. After three failures in a row it must stop and ask. A missing check never counts as a pass. - Approved tests are frozen. After you approve the tests, hooks refuse edits to them and hold edits to executable configuration for your approval, and an audit of the diff catches changes made any other way.
- Seven escalation triggers make the agent stop instead of guessing: materially different approaches, missing access, behaviour-changing ambiguity, a spec–test conflict, a failing retry cap, an edit outside the plan, and anything consequential.
- State lives on disk. Each task has a folder with its specification, plan, progress log and report, and each completed plan task is a commit, so work survives the end of a session.
You need git and Node.js 20 or later.
Claude Code plugin:
/plugin marketplace add Hyper-AI-Lab/hyper-agents
/plugin install paacm@hyper-agents
Cursor plugin (local; a Teams or Enterprise admin can add this repository as a team marketplace instead):
npx github:Hyper-AI-Lab/hyper-agents init --cursor-userProject install, committed with your code and used by both IDEs (--layer power adds the subagents, hooks and permission profiles):
npx github:Hyper-AI-Lab/hyper-agents init --layer power --dry-runThen tell PAACM how to check your project. detect proposes commands from your package.json, Makefile, pyproject.toml, Cargo or Go files; review them, then keep the ones you trust in .agent/checks.conf:
npx github:Hyper-AI-Lab/hyper-agents detectAnd make a request: /paacm <request> in a project install (either IDE), /paacm:paacm <request> with the Claude Code plugin, or the paacm skill in Cursor's Agent with the Cursor plugin.
/paacm Idle sessions never expire. Expire a session after 30 minutes without requests and return 401 after that; add a regression test.
The user guide covers setup, writing requests, supporting data, autonomy levels, reading the report, worked examples and troubleshooting.
| Path | Content |
|---|---|
| docs/guide/ | The user guide |
| docs/method/ | The specification, with decision records |
| docs/research/ | The study behind it |
| docs/project/requirements.md | Requirements and their evidence |
src/content/ |
Everything the kit installs: instructions, skills, subagents, hooks, permission profiles |
src/runtime/ |
The runtime the agent and the hooks use: task state, the check runner, the audit, the gate |
src/cli/ |
The paacm command-line tool: init, upgrade, uninstall, doctor, detect, check, eval |
plugins/paacm/ |
The plugin for both IDEs, generated from src/content/ and src/runtime/ |
PAACM 3.4.1 is a complete implementation of the specification for Claude Code and Cursor, including project extensions for adapting it to a repository's own conventions (guide). Markdown instructions are guidance; the enforcement comes from the hooks, the IDE's permissions and its sandbox. Use the power layer or a plugin when enforcement matters. The method's evidence comes from published studies and vendor guidance, not from your code: evaluate it on tasks from your own history before relying on a claim that it helps.
Upgrading from PAACM 2.0.0: see upgrading-from-2.md.
See CONTRIBUTING.md. Apache-2.0; see LICENSE.