canvas agent guard

canvas agent guard¶

Generate agent-safety config that blocks destructive canvas operations

Synopsis¶

guard generates the permission rules and hooks that stop an AI agent from running destructive canvas operations, derived from the live command tree so the list is always complete.

By default it hard-blocks the irreversible actions (delete, remove, conclude, crosslist, deactivate, unpublish, cancel, close, merge, split, reset, ...) and makes everything else that isn't a known read (create, update, grade, and any verb the guard doesn't recognize) require approval. Classification is fail-safe: only an explicit allowlist of read verbs (get, list, show, ...) stays allowed, so a new or non-obvious mutating command is gated rather than slipping through. Pass --all-writes to block writes too.

IMPORTANT: the "canvas api" escape hatch can issue any HTTP verb. The guard blocks "canvas api DELETE/PUT/POST/PATCH" patterns on the Bash surface but cannot enumerate arbitrary path arguments. For a hard guarantee, run the agent MCP-only (no Bash tool) or inside a read-only sandbox.

Output is printed for review by default; pass --write to install it. See the Agent Safety guide: https://jjuanrivvera.github.io/canvas-cli/user-guide/agent-safety/

canvas agent guard [flags]

Examples¶

  canvas agent guard --host claude-code
  canvas agent guard --host codex
  canvas agent guard --host opencode --all-writes
  canvas agent guard --host claude-code --write

Options¶

      --all-writes    Also block create/update/grade (default: those require approval)
  -h, --help          help for guard
      --host string   Target agent host: claude-code, codex, opencode
      --write         Write the config/hook files instead of printing them

Options inherited from parent commands¶

      --as-user int       Masquerade as another user (admin feature, requires permission)
      --columns strings   Select specific columns to display (comma-separated)
      --config string     config file (default is $HOME/.canvas-cli/config.yaml)
      --dry-run           Print curl commands instead of executing requests
      --filter string     Filter results by text (case-insensitive substring match)
      --instance string   Canvas instance URL (overrides config)
      --limit int         Limit number of results for list operations (0 = unlimited)
      --no-cache          Disable caching of API responses
  -o, --output string     Output format: table, json, yaml, csv (default "table")
      --quiet             Suppress informational messages, only output data and errors
      --show-token        Show actual token in dry-run output (default: redacted)
      --sort string       Sort results by field (prefix with - for descending, e.g., -name)
  -v, --verbose           Enable verbose output

canvas agent guard