@pmelab/gtd 15.1.0 → 15.2.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pmelab/gtd",
-  "version": "15.1.0",
+  "version": "15.2.1",
   "private": false,
   "description": "Git-aware CLI that emits the next prompt for an autonomous coding agent based on the current repository state",
   "bin": {
@@ -38,6 +38,7 @@
     "build": "tsdown --filter web && npx @tailwindcss/cli -i src/web/styles.css -o dist/web/main.css --minify && node scripts/inline-web-client.mjs && tsdown --filter gtd",
     "postbuild": "jiti scripts/generate-schema.ts && node scripts/assert-no-test-doubles.mjs",
     "watch": "tsdown --watch",
+    "bench:replay": "node scripts/bench-replay.mjs",
     "test": "turbo run format:check typecheck lint lint:sh lint:boundaries test:unit test:e2e:inmem test:e2e:live test:web analyze",
     "test:unit": "vitest run --project unit",
     "test:watch": "vitest",

package/src/workflows/prose.ts

@@ -126,13 +126,36 @@ export const questionBar = `- The goal is shared understanding, not a quota or a
   plus a one-line rationale in prose, no checkboxes. That heading
   always comes last, after every other \`##\` section
 - Above the bar, write it under \`## Open Questions\` — always the first
-  \`##\` section — as \`### <question>\` plus a checkbox list: two
-  concrete options and a free-text slot, all unticked:
+  \`##\` section — as \`### <question>\` plus a body plus a checkbox list.
+  The count of concrete options is your call, taken from how many
+  genuinely distinct answers the fork has — never padded to a fixed
+  number. A floor still applies: at least two real options plus the
+  free-text slot, or the question is not a fork
+- The heading is followed by one or two lines of body naming the fork
+  and what actually differs, THEN the option list. The body frames the
+  fork; each option's own content carries that option's impacts — they
+  never repeat each other. A body that lists the options back is
+  padding, and an option line that restates the fork is padding too
+- Each option is one short line plus its own impacts nested under it as
+  a bullet list: one to four bullets, free-form — what it costs, what
+  it buys, how it differs from the others. Write the dimensions this
+  fork actually has and skip the ones it does not; no \`Costs:\`/\`Buys:\`/
+  \`Differs:\` labels, no fixed bullet count. The free-text slot is last,
+  mandatory, identified by position rather than by label:
 
       ### <the question>
 
-      - [ ] <first option — a concrete answer, a few words of rationale>
-      - [ ] <second option>
+      <one or two lines: what the fork is, what actually differs>
+
+      - [ ] <first option — a short line>
+        - <impact>
+        - <impact>
+      - [ ] <second option — a short line>
+        - <impact>
+      - [ ] <third option — a short line>
+        - <impact>
+        - <impact>
+        - <impact>
       - [ ] _your answer_
 
 - Never tick a box yourself — the human ticks exactly one per question`

package/src/workflows/text.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from "vitest"
-import { summaryPrompt, withSkills } from "./text.js"
+import { architectureAuthorPrompt, designTriagePrompt, summaryPrompt, withSkills } from "./text.js"
 import { renderText } from "./text.fixture.js"
 
 describe("withSkills", () => {
@@ -18,6 +18,45 @@ describe("withSkills", () => {
   })
 })
 
+describe("the shared open-question instruction", () => {
+  it("leaves the option count to the agent, floored at two real options", () => {
+    const prompt = renderText(() => designTriagePrompt("base"))
+    expect(prompt).toContain("count of concrete options is your call")
+    expect(prompt).toContain("at least two real options plus the")
+    expect(prompt).not.toContain("plus a checkbox list: two")
+  })
+
+  it("requires a body between the heading and the option list", () => {
+    const prompt = renderText(() => designTriagePrompt("base"))
+    expect(prompt).toContain("followed by one or two lines of body naming the fork")
+    expect(prompt).toContain("what actually differs, THEN the option list")
+  })
+
+  it("requires per-option impacts, free-form with no fixed count", () => {
+    const prompt = renderText(() => designTriagePrompt("base"))
+    expect(prompt).toContain("its own impacts nested under it")
+    expect(prompt).toContain("one to four bullets, free-form")
+    expect(prompt).toContain("labels, no fixed bullet count")
+  })
+
+  it("is the identical block, full text, in both the design gate and the architecture gate", () => {
+    // Anchored at questionBar's own first and last lines (not a bullet
+    // partway in) so the whole ~3.4k-char block is compared — a divergence
+    // anywhere inside, including its opening bullets, fails this test the
+    // moment either gate's copy drifts from the shared source.
+    const start = "The goal is shared understanding, not a quota or an empty section"
+    const end = "Never tick a box yourself — the human ticks exactly one per question"
+    const sharedBlock = (prompt: string): string =>
+      prompt.slice(prompt.indexOf(start), prompt.indexOf(end) + end.length)
+
+    const designBlock = sharedBlock(renderText(() => designTriagePrompt("base")))
+    const architectureBlock = sharedBlock(renderText(() => architectureAuthorPrompt()))
+
+    expect(designBlock.length).toBeGreaterThan(3000)
+    expect(designBlock).toBe(architectureBlock)
+  })
+})
+
 describe("summaryPrompt", () => {
   it("puts the total and every model's cost on a line of its own", () => {
     const prompt = summaryPrompt({