Semantic Contracts Installer

Install a small, stable set of Semantic Contracts so coding agents start each session with the project's shared vocabulary and rules — not just terms the model already knows, but the compositions and rendered templates it cannot guess.

Why contracts first

A semantic anchor (BLUF, MECE, arc42) already works in any chat — the model knows it; installing a list of anchors adds little. A semantic contract is different: it pins which anchors apply in this project and carries a project-local template (e.g. the "Specification" contract = Cockburn use cases + Gherkin + EARS + activity diagrams, with the full template text). That template is information the model cannot reconstruct on its own. Installing contracts is therefore the primary mode; installing a bare anchor block is the --anchors-only fallback.

This is the Anchor / Contract / Skill taxonomy (ADR-007) in practice: the installer's job is to put the always-on contracts into the agent's instruction file.

Workflow (contracts mode — default)

1. Decide scope and portability. Prefer project-local onboarding for "works across as many agents as possible"; use user-level memory only for agents that support it natively. Read references/agent-support.md before writing files.

2. Present the contract catalog. Read data/contracts.json (a bundled snapshot of the published catalog) and show the user the available contracts by title + description. Keep the installed set small — usually 2 to 6 contracts that match the project's actual practice.

3. Render the selected contracts into a block. For each chosen contract, emit its template (or templateDe when the project works in German) under a clear heading, followed by its resolved anchors[] as a one-line supporting-vocabulary list. Start from assets/templates/contracts-block.md. Do not add the markers — the installer adds them.

4. Run the installer instead of editing by hand.

   • scripts/install.sh --source /path/to/contracts-block.md --target-dir <repo> --scope project for repository onboarding.
   • scripts/install.sh --source /path/to/contracts-block.md --scope home for personal onboarding.
   • It injects into the first suitable existing markdown file, else creates a minimal AGENTS.md, and writes an idempotent marker block (<!-- semantic-anchors:start/end -->) so updates replace the previous installation — including any block left by the former semantic-anchor-onboarding skill.

5. Add native behavior only where needed.

   • The marker block lives in AGENTS.md / CLAUDE.md, which Claude Code loads natively — so do not install a SessionStart hook for CLAUDE.md; it would double-load the block.
   • Only use --claude-hook when the block lives in a file Claude Code does not auto-load. The hook is registered with matcher: "startup" (not resume), so it does not re-inject on every resume.
   • For Gemini CLI / Copilot / Cursor, rely on the shared file the installer chose; mirror into the native file only when that tool's workflow needs it.

6. Validate on a fresh session. Start a new session and ask the agent to list the active contracts and how they will shape the next answer. If it misses them, shorten the file and make the active list more explicit.

Sub-mode: --anchors-only

For projects that want only a bare anchor list (the former onboarding behavior): select 3-7 anchors, render them from assets/templates/anchor-block.md, force a single primary choice inside the conflict groups in references/anchor-selection.md, and run the same installer. This is a fallback — prefer contracts.

Keeping the snapshot fresh

data/contracts.json is a bundled snapshot so the skill installs offline and reproducibly. Refresh it from the canonical source with scripts/refresh-contracts.sh (fetches the published contracts.json). Re-run it before installing if the project needs the very latest contract templates.

Writing rules

• Install contracts, not essays — each rendered template is already terse; do not pad it.
• Prefer AGENTS.md as the cross-agent baseline; prefer project onboarding over user onboarding when the user says "all agents."
• State that explicit user instructions override the installed contracts.
• Do not make exploratory or adversarial techniques always-on unless the user asks.

Assets, references, scripts

• data/contracts.json — bundled contract catalog snapshot.
• assets/templates/ — contracts-block.md, anchor-block.md, and per-tool starter files (AGENTS.md, CLAUDE.md, GEMINI.md, copilot-instructions.md).
• references/agent-support.md — agent loading rules and portability.
• references/anchor-selection.md — category prompts and conflict groups (for --anchors-only).
• scripts/install.sh — idempotent block injection with multi-agent file detection.
• scripts/refresh-contracts.sh — refresh the bundled snapshot.
• scripts/claude-session-start.sh — SessionStart hook for non-native targets only.

Requirements

• Runtime: python3 (used by install.sh for settings/JSON manipulation and path resolution); curl for refresh-contracts.sh.
• Platform: Unix/macOS (WSL on Windows).

Maintenance

• skill/ is the canonical source; scripts/sync-claude-plugin.sh regenerates the plugin copies. Do not edit both by hand.
• The former semantic-anchor-onboarding skill is a deprecated alias that forwards here; it will be removed after one release.