Document safe parsing for preToolUse.toolArgs when it is a JSON-encoded string

Start with the Hooks reference and its preToolUse payload section. Document that toolArgs may be a JSON-encoded string, then add safe Bash and Python parsing examples that handle invalid or non-object values; done means hook authors can reliably inspect fields such as command, path, or url.

Summary

The Hooks reference documents preToolUse.toolArgs as unknown, but does not show how hook authors should safely parse it.

In actual Copilot CLI hook invocations I tested, toolArgs arrived as a JSON-encoded string rather than a parsed object. That may be valid under the current unknown contract, but it is easy for hook authors to assume object-style access and accidentally write hooks that fail to inspect tool arguments.

Because hook failures are fail-open, this is a security footgun for policy-enforcing hooks.

Observed behavior

For preToolUse, the documentation shows the camelCase payload shape as:

{
    sessionId: string;
    timestamp: number;
    cwd: string;
    toolName: string;
    toolArgs: unknown;
}

In tested CLI/App-backed hook invocations, the payload effectively behaved like:

{
  "toolName": "bash",
  "toolArgs": "{\"command\":\"echo hello\"}"
}

rather than:

{
  "toolName": "bash",
  "toolArgs": {
    "command": "echo hello"
  }
}

I am not claiming the string form is invalid. Since the schema says unknown, this may be intentional or implementation-defined. The problem is that the docs do not tell hook authors how to handle it safely.

Why this matters

Security hooks commonly inspect fields like:

.toolArgs.command
.toolArgs.path
.toolArgs.url

If toolArgs is a JSON-encoded string, this kind of access does not work as expected. Depending on the script and shell settings, the hook may fail, emit invalid output, or skip the intended check.

Since hook failures are fail-open, a parsing mistake can silently bypass a security policy.

Request

Please document the expected handling for toolArgs and provide safe parsing examples.

At minimum, the docs should say something like:

toolArgs is unknown and may be a JSON-encoded string. Hook scripts should check its runtime type and parse it before inspecting tool arguments.

A Bash example would help:

INPUT="$(cat)"

TOOL_ARGS_JSON="$(
  jq -c '
    (.toolArgs // .tool_args // .tool_input // {}) as $args
    | if ($args | type) == "string" then ($args | fromjson? // {}) else $args end
  ' <<< "$INPUT"
)"

COMMAND="$(jq -r '.command // ""' <<< "$TOOL_ARGS_JSON")"

Python example:

import json
import sys

payload = json.load(sys.stdin)
tool_args = payload.get("toolArgs", payload.get("tool_input", {}))

if isinstance(tool_args, str):
    try:
        tool_args = json.loads(tool_args)
    except json.JSONDecodeError:
        tool_args = {}

if not isinstance(tool_args, dict):
    tool_args = {}

command = tool_args.get("command", "")

Expected improvement

This would make hook authoring safer, especially for security-focused preToolUse hooks, and reduce the chance of fail-open bypasses caused by incorrect assumptions about the runtime type of toolArgs.