tool-interception

A Claude Code plugin that detects when Claude edits or deletes files outside your request. Clear drift pauses on a question dialog; milder drift is flagged above the prompt with pull-back, fine, and allow-folder keys.

A Claude Code plugin that pairs Chrome browser profiles with cswap accounts and blocks Chrome tool actions when the active account does not match the paired profile.

Denies the third consecutive poll or wait command when no edit has landed between them.

Claude Code plugin that intercepts Edit/Write/NotebookEdit calls and asks before Claude edits a file another open chat changed in the last 30 minutes, offering Proceed, Move to a worktree, or Cancel via /guard.

Probe mod for the reins project: a Claude Code plugin that registers tool.call hooks to study how the mod engine treats slow, failing or dialog-opening hooks and where mods sit against settings hooks.

Files you mark as yours: Claude cannot edit them until you hand them back.

A Claude Code plugin that hooks Edit, Write and NotebookEdit tool calls, counting successful edits per file and turn; at the third edit it warns the user and at the fifth it sends the model a note to stop editing, re-read the code path and state the root cause. Counts reset each turn and are tracked separately for the main loop and subagents.

A Claude Code TypeScript mod that intercepts tool calls (Bash, Read) to block, escalate or rewrite them before execution, demonstrated within a broader agent safety pipeline POC.

A Claude Code plugin that refuses or rewrites tool calls according to policy and, when observation is enabled, records hook events as database rows per lineage.

A Claude Code mod that renders a Tokyo Night dashboard pane showing the callers, impact and recorded decisions of the code the agent reads and edits, with an animated pixel mascot and a status band above the prompt.

Secrets found by betterleaks reach the model as ‹secret:…› tokens; Write/Edit restore them, other tools refuse them

A Claude Code Mod that adds debugger-style breakpoints to tool calls, with Continue/Step/Reject controls, a live dashboard, timeline, inspector and Error Lens failure diagnosis — all local, no network or model calls.