fix(hooks): read hook input from stdin JSON instead of environment variables

[sergiobuilds]

pre-tool-use.sh and post-tool-use.sh read TOOL_NAME, FILE_PATH, and TOOL_OUTPUT from the environment. Claude Code delivers hook input as a JSON object on stdin and sets none of those, so both conditions test an empty string. The guards are installed, registered in hooks.json, and unreachable.

The PLOON-overwrite guard is the one that matters — it is the thing standing between an agent and a clobbered interview log, which is the one file in this project that cannot be regenerated.

Measured

Same input, two contracts:

$ echo '{"tool_name":"Write","tool_input":{"file_path":"personas/testuser/persona.md"}}' \
    | bash hooks/pre-tool-use.sh
(no output)

$ TOOL_NAME=Write FILE_PATH=personas/testuser/persona.md bash hooks/pre-tool-use.sh
WARNING: Writing to persona file personas/testuser/persona.md

session-start.sh is unaffected — it reads files, not the environment, and works today.

Second defect, found on the way

The path test was $FILE_PATH == personas/*, anchored at the start. Hooks receive an absolute path, so that pattern would still never match even after the input was fixed. Widened to *personas/*.

After

input	expected	actual
stdin JSON, Write to a persona file	warns	warns
stdin JSON, absolute persona path	warns	warns
stdin JSON, Read	silent	silent
stdin JSON, Write to README.md	silent	silent
post-tool-use, slop in tool_response	warns	warns
post-tool-use, string tool_response	warns	warns
malformed stdin	exit 0, silent	exit 0, silent

Parsing goes through node, which the project already requires, so no jq dependency is added. Malformed input falls through to empty rather than aborting under set -euo pipefail, so a hook can never take a session down with it.