The gap between "finished" and "recorded"

An AI agent working through MCP closes a task the same way a human would: it marks it done, or — more commonly, since a human should still approve an agent's work — it submits the task for review and ends its session. At that exact moment, three things are still true that won't be true an hour later:

Nothing forces any of these to get resolved. So most of the time, they don't — not because anyone decided to skip them, but because nothing in the loop asked.

What WKFGo does about it

Three MCP tools close out work on a task: mark_done, submit_for_review, and end_work. As of this release, every one of their responses can carry two structured nudges back to the connected agent:

docReminder — always present when a task is involved. It instructs the agent to ask, in plain language, whether to write up what was just done as a wiki page via create_wiki_page: the approach taken, the decisions made, anything the next person — human or agent — would otherwise have to rediscover from scratch.

gitReminder — present when the project has a linked Git repository. It instructs the agent to ask whether to commit and push now (with TASK-<id> in the message, so the webhook links it automatically), and, separately, whether an already-pushed commit should be linked after the fact with attach_commit.

Both are phrased as instructions to ask, not to act. The agent's system instructions are explicit: don't write the wiki page or touch git unasked — wait for a yes.

Why "ask" and not "always do"

An earlier version of this same idea just told the agent to remind the user to commit. That's one-directional — a notice, not a decision point. Two things changed:

A silent write is worse than no write. A wiki page the agent decided on its own to publish, in your name, without your review, is a liability, not documentation. The value of create_wiki_page output depends entirely on someone having actually wanted it written.

Committing on someone's behalf assumes facts the server doesn't have. WKFGo doesn't know if you're mid-experiment, if the change is only half done, or if you'd rather squash five commits into one later. It knows a task closed and Git exists on the project — that's a cue to ask, not a green light to act.

What this looks like in practice

Connect Claude Code or Cursor to WKFGo over MCP (see the full setup guide). Work a task normally: start_work → make changes → submit_for_review. The tool call's response now includes something like:

{
  "ok": true,
  "docReminder": { "askUser": true, "message": "Ask the user whether to write this up as a wiki page…" },
  "gitReminder": { "askUser": true, "action": "ask_user_to_commit", "commitRef": "TASK-442", "message": "Ask the user whether to commit and push now…" }
}

The connected agent, following WKFGo's server instructions, turns that into three real questions in the chat — in your language, not English by default:

  1. "Want me to write up what I just did as a wiki page?"
  2. "Ready to commit and push? I'll use TASK-442 in the message."
  3. "You already pushed — want me to link that commit to the task?"

Answer yes to any of them and the agent calls the matching tool (create_wiki_page, a local git commit/push, or attach_commit). Answer no, and nothing happens — the reminder was the whole feature.

What it doesn't do

It doesn't write documentation for you, judge whether your commit message is good, or nag on every single tool call — only at the three points where work on a task actually concludes. And it doesn't require Git to be linked; the documentation question fires on its own.

FAQ

Does this work with any MCP client? Yes — any client that reads and follows tool descriptions and server instructions, including Claude Code and Cursor.

What if I say no to documenting? Nothing gets written. There's no follow-up nag on the same task.

Does the agent ever write the wiki page without asking? No — create_wiki_page is a separate tool call the agent only makes after you say yes.

Is this only for AI agents, or does it help human developers too? Human-driven sessions through an MCP client get the same nudge — the questions aren't agent-specific, they're task-lifecycle-specific.

Bottom line

The fix for "nobody documents anything" was never a stronger policy — it was a better-timed question. WKFGo's MCP server now asks it automatically, right when the answer is still cheap to give.