kaizen-agent 1.33.0 → 1.33.1

package/README.md

@@ -2,11 +2,41 @@
 <img src="https://raw.githubusercontent.com/hfadhlullah/kaizen/main/assets/banner.png" alt="kaizen" width="760">
 </div>
 
-<p align="center"><b>See the plan before your AI starts. Get a second opinion when it is done.</b></p>
+<p align="center"><b>One tool to build and maintain your product with AI agents.</b><br>Requests in, a plan you approve, the work, an independent review, done.</p>
 
-**The problem:** an AI agent starts writing before you agree on what it should do, then grades its own work.
+<p align="center"><img src="assets/hero.gif" alt="Animated walkthrough of the real kaizen board with sample data. 1: a request gathered from a support sheet waits in the Icebox and is promoted. 2: it is an idea on the board, and Run opens Claude Code. 3: the agent works in the card's own terminal and writes a plan. 4: the plan panel shows what done and wrong mean, and it is approved. 5: it builds, then a reviewer that never saw the work reports a high finding, a double charge. 6: only that finding is ticked and fixed. 7: the reviewer re-checks it. 8: the run lands in Done, with the low finding left in the backlog." width="760"></p>
 
-**Kaizen:** it plans first and waits for your approval, then an agent that never saw the work reviews it.
+## The problem
+
+You ask an AI coding agent to retry a payment webhook when the bank times out. It starts editing at once, then says *"Done, tests pass."* The tests do pass. They were written by the same agent, from the same assumption: that a timeout means the charge failed. Sometimes the bank said yes and the reply got lost. Now a retry charges the customer twice, and nothing in the agent's report would make you look.
+
+None of this is a lack of intelligence. It is a missing workflow:
+
+| What goes wrong | What kaizen does |
+| :--- | :--- |
+| The agent decides what "done" means on its own, and you find out from the diff | It writes a plan first, including what *wrong* would look like, and waits for your approve or change |
+| The author grades its own work, so its blind spot is in the code *and* in the check | A second agent that never saw the work reviews it, from the request, the plan and the diff only |
+| Findings get fixed in an endless loop, or silently dropped | Serious findings are fixed and re-checked, at most twice; the rest come to you, and what is left goes to a backlog |
+| Requests live in sheets, docs and social posts; the plan in one chat, the review in another, the commit somewhere else | One board, from an icebox that gathers those requests to an idea, a plan, the build, the review, the fix loop and the commit, every step on file |
+
+Many assistants now ask for approval before they act, and some run a checker on their own draft. The difference here is *who* checks: a reviewer that shares the author's context shares its blind spots. Kaizen's reviewer starts cold, on purpose.
+
+Every stage is a file in `.kaizen/runs/<id>/`, so a run can be resumed later, by another session or another tool (Claude Code, Codex, Antigravity), and read back when someone asks why the code is the way it is.
+
+## Who it is for
+
+**Good fit if you:**
+
+- Build and maintain a product, and want one place for it: requests gathered from where they live, then plan, build, review and commit, without switching tools.
+- Already use an AI coding agent, and have caught it confidently wrong more than once.
+- Change things where a mistake is expensive: money, data, migrations, anything a customer sees.
+- Want to see and steer the plan before code is written, not reverse-engineer it from a diff.
+
+**Not for you if:**
+
+- Your changes are typos and one-liners. A run costs several times one prompt; kaizen takes a short path for small work, but it is not free.
+- You want the agent to ship without you. Kaizen is built to stop and ask.
+- You do not use an AI coding agent. Kaizen runs inside one; it is not one.
 
 ## Install
 

package/cli/sources.test.ts

@@ -101,6 +101,9 @@ test("a sheet row whose status is Done or Finish is skipped, not marked seen, an
   expect(split("sheet", "no\tdetail\tnotes\n1\tDark mode\tDone\n2\tExport\tfinish\n").map((i) => i.text)).toEqual(["Dark mode", "Export"]);
   // The whole cell names the column, not a word inside it.
   expect(split("sheet", "detail\tstatus notes\nDark mode\tDone\n")).toEqual([{ text: "Dark mode", notes: "Done" }]);
+  // A named column of page names loses to the column holding the requests.
+  expect(split("sheet", "NO\tPAGE\tDETAIL\tSTATUS\tNOTES\n1\tDash\tList\tDone\tAdd a chart to the list\n2\tPlan\tTimeline\t\tDropdown closes before an item can be picked\n3\tPlan\t\t\tRename the statuses to proposal and tender\n").map((i) => i.text))
+    .toEqual(["Dropdown closes before an item can be picked", "Rename the statuses to proposal and tender"]);
 
   addSource(state, SHEET);
   expect(await gatherSource(state, SHEET, text(sheet17(statuses)), pub)).toMatchObject({ ok: true, added: 5, note: "5 new, 4 done skipped" });

package/cli/sources.ts

@@ -385,14 +385,19 @@ export function sift(kind: Kind, body: string): { items: Item[]; done: number }
   const head = filled.slice(0, 5).find((row) => !/^\d{1,3}[.)]?$/.test(row.find(Boolean) ?? "") && (named(row, HEADER) >= 0 || named(row, STATUS) >= 0));
   const data = head ? filled.slice(filled.indexOf(head) + 1) : filled;
   const status = head ? named(head, STATUS) : -1;
-  const known = head ? named(head, HEADER) : -1;
-  // No request column named: one column for the whole sheet, the one holding the most
-  // text that reads like a request. Never a cell per row, which takes `Open` or a name.
-  let col = known;
-  if (col < 0) for (let i = 0, most = 0; i < Math.max(0, ...data.map((row) => row.length)); i++) {
-    const n = i === status ? 0 : data.reduce((sum, row) => sum + (wordy(row[i] ?? "") ? row[i]!.length : 0), 0);
-    if (n > most) { most = n; col = i; }
-  }
+  // One column for the whole sheet, the one holding the most text that reads like a
+  // request. Never a cell per row, which takes `Open` or a name.
+  // Finished rows do not vote: their replies outweigh the requests still open.
+  const open = data.filter((row) => status < 0 || !DONE.test(row[status] ?? ""));
+  const text = (i: number) => i === status ? 0 : open.reduce((sum, row) => sum + (wordy(row[i] ?? "") ? row[i]!.length : 0), 0);
+  let wordiest = -1;
+  for (let i = 0, most = 0; i < Math.max(0, ...data.map((row) => row.length)); i++) if (text(i) > most) { most = text(i); wordiest = i; }
+  // A named column loses to one holding far more text: `DETAIL` naming the page while
+  // `NOTES` holds the request.
+  let known = head ? named(head, HEADER) : -1;
+  // ponytail: a named column with no wordy text at all keeps it (one-cell title rows, pinned by test).
+  if (known >= 0 && wordiest >= 0 && text(known) > 0 && text(wordiest) > 3 * text(known)) known = -1;
+  const col = known >= 0 ? known : wordiest;
   let done = 0;
   const items = data.flatMap((row) => {
     const text = row[col] ?? "";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "kaizen-agent",
-  "version": "1.33.0",
+  "version": "1.33.1",
   "description": "Staged workflow for AI agents — code, writing, research, operations: plan, human approval, build, independent review, bounded fix loop.",
   "type": "module",
   "bin": {