Skip to content
This repository was archived by the owner on Sep 8, 2026. It is now read-only.

Latest commit

 

History

History
57 lines (50 loc) · 5.87 KB

setup-reference.md

File metadata and controls

57 lines (50 loc) · 5.87 KB

ContextGuard setup reference

This file is generated from context-guard-kit/setup_wizard.py. Do not edit it manually. Regenerate it with python3 scripts/generate_setup_reference.py --write and verify it with --check.

Agent capability matrix

verified means ContextGuard has a bounded project write path. report-only means setup reports guidance and does not claim a verified write target. Only Claude Code currently has a verified user-scope write path; every user-scope write requires explicit agent selection and --yes.

Key Agent Project scope User scope Rules Skill MCP Hooks Notes
claude Claude Code verified verified write yes — — yes Installs project-local hooks, denies, and statusline in .claude/settings.json.
codex OpenAI Codex CLI verified report-only yes yes yes — Reads AGENTS.md and supports an optional project skill and project-scoped stdio MCP configuration.
gemini Gemini CLI verified report-only yes — yes — Reads GEMINI.md and supports project .gemini/settings.json MCP setup.
cursor Cursor verified report-only yes — yes — Reads project rules and supports project .cursor/mcp.json setup.
windsurf Windsurf verified report-only yes — — — Reads project rules; add an advisory ContextGuard block with --with-init.
cline Cline verified report-only yes — — — Reads project rules; add an advisory ContextGuard block with --with-init.
copilot GitHub Copilot Coding Agent verified report-only yes — yes — Reads repository instructions and supports project .vscode/mcp.json setup.
opencode OpenCode verified report-only — yes yes — Supports a project skill and project opencode.json MCP setup; no hooks are auto-written.
forgecode ForgeCode verified report-only — — yes — Supports project-local .mcp.json setup; other helpers remain shell-driven.
generic Other / unknown agent shell only report-only — — — — No automated setup surface; run ContextGuard helpers from the shell as needed.

Setup flags

--plan and --dry-run are read-only. Writes require --yes; user scope additionally requires explicit --agent or --only selection.

Flag Default Choices Description
--root — — project root to configure (default: nearest git root, else current directory)
--scope project project, user, global setup scope: project-local by default; user/global targets only known user-level paths and requires explicit --agent for writes
--profile recommended minimal, recommended, max which guardrails to enable: minimal (deny rules + Read guard), recommended (adds Bash trim/escrow and model defaults), max (adds the deprecated statusline and the failed-attempt nudge); individual --no-* flags still remove items
--yes false — apply the recommended/selected setup without prompts
--plan false — show the setup plan without writing files
--verify false — run a read-only setup health check; never writes or prompts
--json false — print machine-readable result
--rules-only false — run an isolated rule-file operation without reading or changing settings/hooks
--narration-mode — quiet, default with --rules-only, add quiet Claude narration guidance or restore default behavior
--no-backup false — do not create .bak-* before modifying existing settings
--no-denies false — skip recommended permissions.deny rules
--no-statusline false — skip the token statusline (deprecated: it duplicates Claude Code /usage; only --profile max installs it)
--no-bash-hook false — skip Bash trim/sanitize hook
--bash-reference-v1 false — opt in to 7-day scoped receipt references in the Bash hook; handles are provider-visible
--no-bash-reference-v1 false — disable/remove the optional Bash receipt-reference hook flag (default)
--no-read-guard false — skip large Read guard hook
--no-model-defaults false — skip model/effort defaults
--no-diet-scan false — skip the read-only diet scan summary after applying setup
--allow-path-helper-fallback false — allow trusted PATH helper resolution only after bundled/repo helpers are missing and identity validation passes
--agent — — adapter key(s) to configure; comma-separated or repeatable. Alias for --only.
--only — — restrict cross-agent setup/plan to adapter key(s); comma-separated or repeatable (e.g. --only codex,gemini). Default: claude plus any detected agents.
--with-init false — also write advisory ContextGuard rule files for repo-rule agents (AGENTS.md, GEMINI.md, .cursorrules, etc.) when applying; safe and idempotent.
--with-skill false — also generate optional project-local skill files where supported, currently Codex and OpenCode .agents/skills/context-guard/SKILL.md.
--with-mcp false — also configure a verified project-scoped stdio MCP server for Codex, Gemini, Cursor, Copilot/VS Code, OpenCode, or ForgeCode.
--brief-mode — lite, standard, ultra, off plan/apply advisory brief-mode snippets in project rule files; choose lite, standard, ultra, or off to remove.
--list-adapters false — print the cross-agent adapter registry and exit
--failed-attempt-nudge — — enable Bash terminal-event hooks that inject a one-line strategy-switch hint when the same command fails twice in a row (off by default; also enabled by --profile max)
--no-failed-attempt-nudge — — skip the failed-attempt nudge hook even under --profile max