Remove the greenfield flow: a build always extends the existing codebase

packages/framework/prompts/README.md

@@ -6,7 +6,7 @@ is written in TypeScript any more, so prompting can change without touching the
 | file | what it is |
 |---|---|
 | `system_prompt.md` | The built-in system prompt (#326). Rom's doc. |
-| `build_prompt.md` / `extend_prompt.md` / `scaffold_prompt.md` | The prompts a build session opens with (#1347): greenfield, existing codebase (#185), and the scaffold retry (#182). `${{ tf.prompt }}` is the user's intent. |
+| `extend_prompt.md` | The prompt a build session opens with (#1347): deliver the intent within the existing codebase (#185). `${{ tf.prompt }}` is the user's intent. |
 | `protocols/await.md` | How to emit an awaited choice so the turn-boundary gate can detect it (#337/#339). |
 | `protocols/signal.md` | How to emit `setSessionName()` / `setReadyForMerge()` (#326). |
 | `presets/*.md` | One file per preset button: research (#331), readability (#360), maintainability (#361), security_audit (#461), ux (#472). |

packages/framework/prompts/SPEC.md

@@ -13,7 +13,7 @@ Every prompt The Framework sends an agent lives here as markdown. Nothing agent-
 - **The data-branch protocol** - where the framework's own data lives and how to read and write it without putting it on a code branch.
 - **The protocols** - how an agent signals: awaited choices, session name and ready for merge, plus the sections added only when it has a browser, when it runs hands-off, and when nothing can answer its questions.
 - **The presets** - one file per launcher button and per routine prompt.
-- **The build-opening prompts** - the three framings a build agent can open with: the greenfield build, the existing-codebase variant, and the scaffold retry for a build that produced nothing.
+- **The build-opening prompt** - the one framing a build agent opens with: deliver the work within the existing codebase.
 - **The on-before-mergeable prompt** - the optional extra turn a finished agent gets, queueing quality follow-ups and folding what it learned into the knowledge base.
 - **Prompts are reviewed before they land** - a prompt change goes through review like any other change.
 
@@ -47,7 +47,7 @@ See `## User story`.
 - **The triage scope rule** — the one-paragraph rule appended to both triage presets, that a triage only queues work and never implements it.
 - **The protocols** — how an agent signals to The Framework, and what this particular agent can do.
 - **The presets** — one file per preset: the launcher's buttons and the daemon's routine prompts.
-- **The build-opening prompts** — the greenfield build, the existing-codebase variant chosen when the workspace already holds source, and the scaffold retry sent when a build's opening turn left the workspace empty.
+- **The build-opening prompt** — the existing-codebase framing every build agent opens with: a project is a repo that already exists.
 - **The on-before-mergeable prompt** — the extra turn an agent gets when it signals ready for merge, if the user turned that on.
 
 ### Prompts are reviewed before they land

packages/framework/src/agent.test.ts

@@ -267,7 +267,7 @@ test('a build turn that stops to ask fires a live gate and resumes on the pick (
     '```'
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
-      if (/Build this app end to end/.test(prompt)) return awaitBlock // the build stops to ask
+      if (/Work within the existing codebase/.test(prompt)) return awaitBlock // the build stops to ask
       if (/You paused to ask/.test(prompt)) return 'Built it with Postgres. Done.' // the resume
       return 'done'
     },
@@ -302,7 +302,7 @@ test('a run with no preset and no serve config reviews nothing (#1372)', async (
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
       prompts.push(prompt)
-      if (/Build this app end to end/.test(prompt)) return 'Built it. Done.'
+      if (/Work within the existing codebase/.test(prompt)) return 'Built it. Done.'
       return 'done'
     },
     sessionId: 'blackbox1372',
@@ -325,7 +325,7 @@ test('a build turn that stops to showMultiSelect fires a checklist gate and resu
     '```'
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
-      if (/Build this app end to end/.test(prompt)) return awaitBlock
+      if (/Work within the existing codebase/.test(prompt)) return awaitBlock
       if (/You paused to ask/.test(prompt)) return 'Added the picks to TODO. Done.'
       return 'done'
     },
@@ -361,7 +361,7 @@ test('a build turn that stops for plan approval resumes on Approve (#358)', asyn
     '```'
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
-      if (/Build this app end to end/.test(prompt)) return awaitBlock
+      if (/Work within the existing codebase/.test(prompt)) return awaitBlock
       if (/You paused to ask/.test(prompt)) return 'Built the plan out. Done.'
       return 'done'
     },
@@ -402,7 +402,7 @@ test('a declined plan stops the session cleanly instead of building on it (#358)
   let resumed = false
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
-      if (/Build this app end to end/.test(prompt)) return awaitBlock
+      if (/Work within the existing codebase/.test(prompt)) return awaitBlock
       if (/You paused to ask/.test(prompt)) resumed = true
       return 'done'
     },
@@ -438,7 +438,7 @@ test('an unmarked option is an ordinary answer, whatever it is labelled (#358)',
   let resumedWith = ''
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
-      if (/Build this app end to end/.test(prompt)) return awaitBlock
+      if (/Work within the existing codebase/.test(prompt)) return awaitBlock
       if (/You paused to ask/.test(prompt)) resumedWith = prompt
       return 'Understood — waiting for your instructions.'
     },
@@ -508,7 +508,7 @@ test('with nobody to ask, a session takes the recommended option and carries on
   let resumed = false
   const driver = new FakeDriver({
     respond: (prompt: string): string => {
-      if (/Build this app end to end/.test(prompt)) return 'built it\n```await-choices\n{ "options": [{ "label": "A" }] }\n```'
+      if (/Work within the existing codebase/.test(prompt)) return 'built it\n```await-choices\n{ "options": [{ "label": "A" }] }\n```'
       if (/You paused to ask/.test(prompt)) resumed = true
       return 'done'
     },
@@ -965,47 +965,6 @@ test('a build extends an existing project instead of rebuilding it (#185)', asyn
   }
 })
 
-test('a build uses greenfield framing for an empty workspace (#185)', async () => {
-  const cwd = await mkdtemp(join(tmpdir(), 'fw-greenfield-'))
-  try {
-    const { driver, prompts } = realNamedDriver([{ text: 'scaffolded it' }, { text: 'scaffolded it properly' }])
-    await runAgent({ prompt: 'a blog', driver, cwd, todoLoop: false })
-    assert.match(prompts[0]!, /Build this app end to end/i)
-    assert.doesNotMatch(prompts[0]!, /existing codebase/i)
-  } finally {
-    await rm(cwd, { recursive: true, force: true })
-  }
-})
-
-test('a build re-prompts to scaffold from scratch when the workspace stays empty (#182)', async () => {
-  const cwd = await mkdtemp(join(tmpdir(), 'fw-scaffold-'))
-  try {
-    const { driver, prompts } = realNamedDriver([
-      { text: 'thinking about the stack' },
-      { text: 'scaffolded the whole app' },
-    ])
-    const { text } = await runAgent({ prompt: 'a blog', driver, cwd, todoLoop: false })
-    // Two turns: the build, then the hard from-scratch retry once nothing landed on disk.
-    assert.equal(prompts.length, 2)
-    assert.match(prompts[1]!, /from scratch|empty/i)
-    assert.equal(text, 'scaffolded the whole app')
-  } finally {
-    await rm(cwd, { recursive: true, force: true })
-  }
-})
-
-test('a build does not re-prompt when it produced files (#182)', async () => {
-  const cwd = await mkdtemp(join(tmpdir(), 'fw-noscaffold-'))
-  await writeFile(join(cwd, 'package.json'), '{}')
-  try {
-    const { driver, prompts } = realNamedDriver([{ text: 'built it' }])
-    await runAgent({ prompt: 'a blog', driver, cwd, todoLoop: false })
-    assert.equal(prompts.length, 1)
-  } finally {
-    await rm(cwd, { recursive: true, force: true })
-  }
-})
-
 test('a prompt session runs its text without build framing, and works no backlog (#353)', async () => {
   const cwd = await mkdtemp(join(tmpdir(), 'fw-promptkind-'))
   await writeFile(join(cwd, 'TODO_AGENTS.md'), '- [ ] leftover task\n')

packages/framework/src/agent.ts

@@ -4,7 +4,7 @@ import { createAgentControls, emitSessionStart, endStopDetail } from './agent-te
 import { createTurnSignalEmitter } from './turn-gate.js'
 import { runAwaitRounds } from './await-gate.js'
 import { runTodoLoop, type TodoLoopResult } from './todo-loop.js'
-import { buildPrompt, extendPrompt, isWorkspaceEmpty, scaffoldPrompt } from './steps.js'
+import { extendPrompt } from './steps.js'
 import { type ChoicePick, type ChoiceRequest, type FrameworkEvent } from './events.js'
 import type { AgentMessages } from './agent-messages.js'
 import { isHandsOff, type AgentLocation } from './agent-location.js'
@@ -195,7 +195,7 @@ export async function runAgent(opts: RunAgentOptions): Promise<RunAgentResult> {
   try {
     const rounds = await runAwaitRounds({
       session,
-      prompt: openingPrompt(opts, kind, resuming, session.cwd),
+      prompt: openingPrompt(opts, kind, resuming),
       emitTurnSignals,
       requestChoice: opts.requestChoice,
       emit,
@@ -215,16 +215,6 @@ export async function runAgent(opts: RunAgentOptions): Promise<RunAgentResult> {
     if (rounds.stopped) answerController.abort(new Error('[framework] stopped by your answer'))
     let text = rounds.text
 
-    // #182: a build must actually produce an app. If nothing landed on disk the agent stalled
-    // (e.g. sanity-checking the stack), so re-prompt once with a hard "create it from scratch"
-    // directive. Only for a real driver — the fake one writes nothing, so its workspace always
-    // reads empty — and only when the agent is not mid-question, which the gates just drained.
-    if (kind === 'build' && !resuming && opts.driver.id !== 'fake' && !rounds.stopped && isWorkspaceEmpty(opts.cwd)) {
-      const scaffolded = await session.prompt(scaffoldPrompt(opts.prompt), { signal: agentSignal })
-      emitTurnSignals(scaffolded.text)
-      text = scaffolded.text
-    }
-
     // The session controls (a Stop, an answer that said stop #358) abort between turns, and the
     // opening rounds do not observe the abort themselves, so look before treating this as a
     // success — otherwise an aborted session settles as done.
@@ -286,17 +276,15 @@ export async function runAgent(opts: RunAgentOptions): Promise<RunAgentResult> {
  * A resumed session (#720/#1467) gets the text verbatim: the `--resume`d transcript already
  * carries the framing, so composing it again would stack a second preamble onto a conversation
  * that lived through the first. So does a transparent or prompt-less session, which is what
- * "raw `claude -p`" means. A build is framed for the workspace it lands in: an existing codebase
- * is *extended*, not rebuilt from scratch (#185).
+ * "raw `claude -p`" means. A build extends the existing codebase (#185) — a project is a repo
+ * that already exists, so the greenfield from-scratch framing is gone (#1683 review).
  */
-function openingPrompt(opts: RunAgentOptions, kind: AgentKind, resuming: boolean, cwd: string): string {
+function openingPrompt(opts: RunAgentOptions, kind: AgentKind, resuming: boolean): string {
   if (resuming || opts.transparent) return opts.prompt
   if (kind === 'prompt') {
     return opts.vanilla ? opts.prompt : renderSystemPrompt({ prompt: opts.prompt }).user
   }
-  // Gated on a real driver, so the fake one (which writes nothing, so its workspace always reads
-  // empty) always takes the greenfield path and stays deterministic.
-  return opts.driver.id !== 'fake' && !isWorkspaceEmpty(cwd) ? extendPrompt(opts.prompt) : buildPrompt(opts.prompt)
+  return extendPrompt(opts.prompt)
 }
 
 /** The live-chat phase a build reaches after its backlog, sharing the rounds' own loop. */

packages/framework/src/cli.test.ts

@@ -468,7 +468,7 @@ test('runCli runs a whole session offline', async () => {
   assert.equal(code, 0)
   const text = out.join('\n')
   assert.match(text, /▶ "a blog"/) // what it was asked for, said once as the session opens
-  assert.match(text, /Build this app end to end/)
+  assert.match(text, /Work within the existing codebase/)
   // #1372: nothing reviews the build, so there is no checklist pass — the build is the whole session.
   assert.doesNotMatch(text, /checklist pass/)
   assert.match(text, /\u2713 done/)
@@ -483,7 +483,6 @@ test('runCli --transparent runs a bare prompt raw, skipping the build flow + wra
     const text = out.join('\n')
     // Transparent = raw Claude Code: the build path's markers must NOT appear (contrast the test above).
     assert.doesNotMatch(text, /scope: full/) // no scope phase
-    assert.doesNotMatch(text, /Build this app end to end/) // no buildPrompt wrapping
     assert.doesNotMatch(text, /Work within the existing codebase/) // no extendPrompt wrapping
     assert.doesNotMatch(text, /production-grade/) // no synthesize / production-grade pass
   } finally {

packages/framework/src/steps.SPEC.md

@@ -1,9 +1,8 @@
-The prompts a build agent opens with, and the one check that decides between them: whether the workspace already holds an app.
+The opening prompt of a build agent, filled with the user's intent.
 
-## Business logic — TL;DR
+## Business logic
 
-- **Three build-opening prompts, filled with the user's intent** - the greenfield build prompt, the existing-codebase prompt (chosen when the workspace already holds source at build time), and the scaffold retry (sent when a build's opening turn left the workspace empty). Each prompt's text is authored as markdown in the prompts directory like every other agent-facing prompt; what happens here is only the choice between them and filling in the intent.
-- **Workspace-emptiness check** - a workspace counts as empty when it holds no source file the agent could have produced: lockfiles, dotfiles, and dependency/output directories (node_modules, .git, dist, build caches) do not count. Best-effort and cheap — it stops at the first real file, treats an unreadable or missing directory as empty, and never throws.
+A build agent's opening prompt is the existing-codebase framing — a project is a repo that already exists, so there is exactly one framing. Its text is authored as markdown in the prompts directory like every other agent-facing prompt; what happens here is only filling in the intent. The old greenfield and scaffold-retry framings (build an app in an empty folder) were removed: no product path leads a build agent into an empty workspace any more.
 
 ## Before modifying/creating SPEC.md files
 

packages/framework/src/steps.test.SPEC.md

@@ -1,4 +1,4 @@
-What the tests cover: the workspace-emptiness check calls an empty directory, a directory holding only noise (lockfiles, dotfiles, node_modules), and a missing directory empty, and a directory with a real source file not empty; the greenfield build prompt names the intent, frames an end-to-end build, and warns the workspace may be empty; the existing-codebase prompt names the work and the existing codebase and nothing else — none of the retired behavior rules, no scope ask, and no claim the workspace might be empty; the scaffold retry prompt is the hard from-scratch directive and says an empty directory is expected rather than a reason to refuse.
+What the tests cover: the existing-codebase opening prompt names the work and the existing codebase and nothing else — none of the retired behavior rules, no scope ask, and no claim the workspace might be empty.
 
 ## Before modifying/creating SPEC.md files