forge-init

name	forge-init
description	Repo bootstrap runbook — writes the local agent contract and captures repo specifics. User-invoked only.
source	https://raw.githubusercontent.com/weselben/RooForge/main/skills/forge-init/SKILL.md
disableModelInvocation	true

Forge Init

One-shot bootstrap. Runs once per repo. Prepares the agent's local contract (AGENTS.md) so every session starts aligned.

Steps

1. Create or smart-append AGENTS.md

Read ./agents.template.md. Use it as the base.

If AGENTS.md does not exist at repo root:

• Write the template content verbatim as AGENTS.md at repo root.
• Load use-git-identity to check ownership. If repo is under the configured identity, proceed. If unclear, ask: "Create AGENTS.md with forge contract at repo root?"

If AGENTS.md exists:

• Smart-append: preserve existing content, prepend the template's mandatory header (forge load rule + mandates rule) if absent, merge jargon entries (skip duplicates), keep letter-style prose.
• Do not overwrite user-written sections.

Always:

• Append at bottom (once):

  ---
  Do not commit or stage this file. Local changes only.
  

• Never commit or stage AGENTS.md.

Completion criterion: AGENTS.md at repo root contains the mandatory header, the jargon letter, the no-commit footer.

2. Grilling session for repo-specifics

Load grilling skill. Interview the user about this repo.

After every grilling wave (4 questions answered), adapt AGENTS.md:

• Amend only non-obvious facts not inferable from reading the repo.
• Interpret user's language into STE100.
• Never add deployment-specific or coding-stack-specific content (frameworks, ORMs, runtimes, languages) — the agent reads the code for that.
• Ask the user: "Is this as intended?"
• Only add statements the you judge as unambiguous and valuable.

During grilling:

• Answer user questions back in STE100.
• Never emit code blocks. Never explain what code does. Never document other files.
• Goal-driven prompting: frame every addition as "the goal is X, so the agent must Y" — no paths, no procedures, only the contract.

Completion criterion: grilling frontier empty; AGENTS.md captures repo-specific contract.

Rules

• One-shot. Run once. Idempotent re-runs only update.
• Never commit AGENTS.md. It is a local contract, not a repo artefact.
• STE100 throughout. Letter style; jargon: one flowing sentence per term.
• Mandatory header lives at top of AGENTS.md. Forge load rule + mandates rule. Never removed.
• Grilling is mandatory. Do not skip. User must confirm the contract.
• Goal-driven. Every addition to AGENTS.md states the goal, not the procedure.
• No stack-specific content. Frameworks, ORMs, runtimes, languages are inferable from the code — do not write them into AGENTS.md.

Template

The template lives at ./agents.template.md. Read it verbatim. It contains:

1. Letter header — "This is a letter from me to you"
2. Mandatory forge load rule
3. Mandates rule
4. Jargon — one flowing sentence per confusing term only

Boundaries

Forge Init does not:

• Create docs folders (forge-flow handles work surface)
• Set git identity (use-git-identity does that)
• Initialise labels (wayfinder/scripts/setup-repo-gh-cli.sh does that)
• Chart maps or resolve tickets (forge does that)

It only ensures AGENTS.md exists with the contract, then grills for repo specifics.