human-in-the-loop

Reviewed at commit d8df5ff

Claude assigns you the tasks only you can do. They wait in a My tasks pane, counted under the prompt, until you answer or reject them, so nothing Claude needs…

Made by
Tzafrir Rehan
What we checked

Adds three tools that let Claude hand you tasks only you can do (a secret, an approval, a choice). They wait in a My tasks pane until you answer, and your answers are sent back to Claude as a message. It adds a few lines to the system prompt, makes no network calls, runs no programs and keeps tasks in its own store, per project.

Review details

Static review of the source at the pinned commit; not run. Delta review of the changes since bf8233c; the rest carries over from the review at that commit. Registers three plugin tools through $.tool.register (assign_task, list_tasks, withdraw_task), the command /my-tasks and a pane plus status line. Hooks: session.start, session.end, command.run (my-tasks only), tool.describe and tool.call (only its own three tools; subagents are denied), turn.start, turn.step, turn.complete, prompt.submit, prompt.compose, prompt.context, ui.render (Pane, ToolUse of assign_task), ui.focus, session.attach and session.detach. prompt.compose appends three lines to the system prompt telling Claude to assign tasks; prompt.context adds a short list of open tasks at session start and after compaction. prompt.submit reads each prompt: one that starts with the answer marker for an open task is treated as the answer, held back with a message when it looks like a secret (key shapes such as sk-, AKIA, ghp_, private keys, JWTs, long mixed tokens), and when saved responses are due the prompt text is replaced by, or gets an extra context block with, the user's own responses. Responses are sent with $.prompt.submit (asUser) or $.session.append. $.prompt.read and $.prompt.fill put typed text into the prompt box. Tasks, answers and per-session slots are stored with $.store under keys per project folder (with a heartbeat every 60 s so sessions in one project can hand over tasks), so answers persist in the plugin's own JSON store outside the project; the README discloses this and says the transcript keeps answers too. No $.http, $.process or $.fs use, no environment access, no eval, no dynamic code. Open tasks are capped at 5 and titles are sanitised before drawing. Store failures time out after 5 s and fall back without blocking. What changed in this version: the answer field in the pane is now four rows tall (pane.ts, register.tsx rows prop), and the working state is tracked from the main loop's turn.start and turn.complete only (subagent turns ignored) instead of from a passthrough ui.render hook on AbovePrompt, which is removed. No new runtime calls, hosts, files or permissions. Read in full in the previous review: hooks.json, plugin.json, register.tsx lines 1-1280 and 1500-1640 and 2030-2218, tasks.ts, the head of text.ts and project.ts, and the types; the rest of register.tsx and project.ts was covered by targeted search. Tests and docs in the repository were not reviewed.

What it uses
  • Reads your prompts: yes
  • Sees tool calls: yes
  • Blocks or changes tool calls: no
  • Reads or writes files: no
  • Runs programs on your machine: no
  • Calls the model (uses tokens): no
  • Sends data over the network: no
  • Stores data: yes
  • Shows something on screen: yes
Hooked events

Claude assigns you the tasks only you can do. They wait in a My tasks pane, counted under the prompt, until you answer or reject them, so nothing Claude needs from you gets lost in the chat.

Made this mod? Add a badge to your README

modsforClaude badge for human-in-the-loop

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/human-in-the-loop.svg)](https://modsforclaude.com/mods/human-in-the-loop/)

← All mods