| name | research-init |
| description | Use when installing, attaching, initializing, repairing, or validating ARC in a research workspace. |
Research init
Bring one target workspace to a verified ARC setup state. Treat setup as
infrastructure work, not research intent. Stop with a precise handoff to
research-onboard, research-brainstorm, or research-scope.
Read references/setup-contract.md before a
protocol merge, external RESEARCH_ROOT, or skill-path repair.
Boundary
Own setup orchestration only:
- inspect the toolbox, target workspace, research root, protocol files, skill
links, interface, and Server;
- install or repair repository-backed skill symlinks;
- attach
AGENTS.md and CLAUDE.md through managed blocks without replacing
project-owned text;
- initialize an empty managed root through
lib.research_state;
- validate the current managed state;
- build the interface and start or reuse a healthy Dashboard Server by default;
- report the resolved roots, mutations, Server URL, health, and next skill.
Do not infer or commit Project, Direction, or Experiment intent. Do not create
a Package, launch a Run, modify source/data/environments, write management
events directly, adopt unsupported legacy roots, or commit Git changes.
Procedure
1. Inspect before mutation
Run the read-only classifier:
python3 skills/research-init/scripts/research_init.py \
--workspace <workspace> \
inspect \
--agent codex
Use --agent claude or --agent both when requested. Resolve script paths
through the installed research-init symlink in a target project.
Interpret the state exactly:
ABSENT: safe greenfield setup candidate.CURRENT: reconcile protocols, skills, interface, and Server.INVALID: stop. Legacy data, an unknown version, or an unversioned managed
root is unsupported; do not adopt or rewrite it by guessing.
2. Review protocol conflicts
If either protocol is UNMANAGED, show the target file and the proposed
managed block. Obtain user confirmation before passing --merge-protocols.
Never overwrite project-owned text. A symlink, directory, or malformed managed
block is a hard conflict.
3. Set up a greenfield or current workspace
python3 skills/research-init/scripts/research_init.py \
--workspace <workspace> \
setup \
--agent codex
The command installs all repository skills for the selected agent, attaches
the protocols, initializes .research only when absent, builds the interface,
and starts or reuses the Dashboard Server. Use --merge-protocols only after
step 2. Use --no-serve only when the user explicitly requests a headless or
CI setup.
4. Validate and report
python3 skills/research-init/scripts/research_init.py \
--workspace <workspace> \
check \
--agent codex
Report all of the following:
workspace, research_root, and managed VERSION;- skill installation roots and any repaired links;
- protocol actions and confirmation that project-owned prefixes remain;
- interface root and Dashboard Server action,
started or reused;
- Dashboard URL, host, port, health, and SSH forwarding command;
- the exact Dashboard stop command;
READY_NO_PROJECT, READY_WITH_PROJECT, or REPAIR_REQUIRED;- the single next action.
Use research-onboard when no active Project exists. With an active Project,
use research-brainstorm for a vague Direction or research-scope for clear
intent.
Bootstrap note
An agent must discover research-init before this skill can install its
siblings. Keep one minimal bootstrap command in the toolbox README. After the
skill is available, route all setup, attach, and setup-repair work
through this skill.
