fix(agent-prompt): observe the composer before and after Enter instead of a blind submit clock

src/main/runtime/agent-prompt-composer-pending.test.ts

@@ -0,0 +1,126 @@
+import { describe, expect, it } from 'vitest'
+import {
+  buildAgentPromptFragment,
+  detectAgentPromptComposerVerdict
+} from './agent-prompt-composer-pending'
+
+const PROMPT = [
+  'You are working on task task_123 for dispatch ctx_456.',
+  '',
+  'Send worker_done when finished.'
+].join('\n')
+
+describe('detectAgentPromptComposerVerdict', () => {
+  it('is unknown when no rendered screen exists', () => {
+    expect(detectAgentPromptComposerVerdict(null, PROMPT)).toBe('unknown')
+    expect(detectAgentPromptComposerVerdict({ lines: [] }, PROMPT)).toBe('unknown')
+  })
+
+  it('is pending when the emulator extracted a draft holding the payload', () => {
+    expect(
+      detectAgentPromptComposerVerdict(
+        { lines: ['›'], draft: '[Pasted Content 5033 chars]' },
+        PROMPT
+      )
+    ).toBe('pending')
+    expect(
+      detectAgentPromptComposerVerdict(
+        { lines: ['›'], draft: 'You are working on task task_123 for dispatch ctx_456.' },
+        PROMPT
+      )
+    ).toBe('pending')
+  })
+
+  it('is clear when the extracted draft is something else entirely', () => {
+    expect(
+      detectAgentPromptComposerVerdict({ lines: ['›'], draft: 'notes the operator typed' }, PROMPT)
+    ).toBe('clear')
+  })
+
+  it('is pending for a Codex pasted-content placeholder on the prompt line', () => {
+    const lines = [
+      ' >_ OpenAI Codex (v0.152.0)',
+      ' model:       gpt-5.6-sol medium',
+      '› [Pasted Content 5033 chars]',
+      '  tab to queue message',
+      ' gpt-5.6-sol · 100% context left'
+    ]
+    expect(detectAgentPromptComposerVerdict({ lines }, PROMPT)).toBe('pending')
+  })
+
+  it.each([
+    '❯ [Pasted text #1 +91 lines]',
+    '❯ [Pasted: 91 lines]',
+    '> [[ You are working.. [91 lines] .. ]]'
+  ])('is pending for the placeholder %s', (composerLine) => {
+    expect(detectAgentPromptComposerVerdict({ lines: ['history', composerLine] }, PROMPT)).toBe(
+      'pending'
+    )
+  })
+
+  it('is pending when the prompt text itself sits after the prompt glyph', () => {
+    const lines = ['› You are working on task task_123 for dispatch ctx_456.', 'Send worker_done']
+    expect(detectAgentPromptComposerVerdict({ lines }, PROMPT)).toBe('pending')
+  })
+
+  it('is clear when the last prompt line is empty even if history echoes the prompt', () => {
+    const lines = [
+      '> [Pasted text #1 +2 lines]',
+      '✻ Thinking…',
+      '────────',
+      '❯',
+      '────────',
+      '? for shortcuts'
+    ]
+    expect(detectAgentPromptComposerVerdict({ lines }, PROMPT)).toBe('clear')
+  })
+
+  it('is clear when the prompt line holds unrelated text', () => {
+    expect(
+      detectAgentPromptComposerVerdict({ lines: ['› something the operator typed'] }, PROMPT)
+    ).toBe('clear')
+  })
+
+  it('is unknown when no prompt glyph line exists on screen', () => {
+    expect(detectAgentPromptComposerVerdict({ lines: ['plain output', 'more'] }, PROMPT)).toBe(
+      'unknown'
+    )
+  })
+
+  it('ignores a prompt glyph line that scrolled far above the bottom of the screen', () => {
+    const lines = [
+      '› [Pasted Content 5033 chars]',
+      ...Array.from({ length: 20 }, (_, i) => `l${i}`)
+    ]
+    expect(detectAgentPromptComposerVerdict({ lines }, PROMPT)).toBe('unknown')
+  })
+
+  it('does not mistake the Codex banner for a composer line', () => {
+    expect(
+      detectAgentPromptComposerVerdict(
+        { lines: [' >_ OpenAI Codex (v0.152.0)', 'model: gpt'] },
+        PROMPT
+      )
+    ).toBe('unknown')
+  })
+
+  it('does not treat a shell prompt ending in > as a composer', () => {
+    expect(detectAgentPromptComposerVerdict({ lines: ['PS C:\\repo> codex'] }, PROMPT)).toBe(
+      'unknown'
+    )
+  })
+})
+
+describe('buildAgentPromptFragment', () => {
+  it('uses the first non-empty line, collapsed and capped', () => {
+    expect(buildAgentPromptFragment('\n\n  You   are working\ton task\nrest')).toBe(
+      'You are working on task'
+    )
+    expect(buildAgentPromptFragment('x'.repeat(100))?.length).toBe(48)
+  })
+
+  it('refuses fragments too short to be distinctive', () => {
+    expect(buildAgentPromptFragment('ok')).toBeNull()
+    expect(buildAgentPromptFragment('   \n\n')).toBeNull()
+  })
+})

src/main/runtime/agent-prompt-composer-pending.ts

@@ -0,0 +1,76 @@
+/** What the rendered screen says about an injected prompt: `pending` = the composer still holds
+ *  it, `clear` = a recognized composer is empty (or holds something else), `unknown` = no rendered
+ *  screen, or no composer line this detector recognizes — never grounds for another Enter. */
+export type AgentPromptComposerVerdict = 'pending' | 'clear' | 'unknown'
+
+export type AgentPromptComposerScreen = {
+  lines: readonly string[]
+  /** Composer text the emulator extracted from the cursor rows, when it recognized one. */
+  draft?: string
+}
+
+// Why: multi-line pastes collapse to a placeholder whose shape differs per TUI — Codex
+// `[Pasted Content 5033 chars]`, Claude `[Pasted text #1 +91 lines]`, Grok `[Pasted: 91 lines]`,
+// Hermes `[[ … [91 lines] .. ]]` — so match the shared "pasted"/"N lines" bracket, not a vendor.
+const PASTE_PLACEHOLDER_RE = /\[+\s*pasted\b[^\]]*\]+|\[\d+\s+lines\]/i
+// Why: composer prompts start the line with their glyph; a shell prompt like `PS C:\repo>` does not,
+// and a bare `>` only counts when spaced — Codex's own banner opens with `>_ OpenAI Codex`.
+const COMPOSER_PROMPT_LINE_RE = /^\s*(?:[❯›»]|>(?=\s|$))\s?(.*)$/
+// Why: the composer lives at the bottom of the frame; a glyph line further up is transcript.
+const COMPOSER_SCAN_LINES = 12
+const PROMPT_FRAGMENT_MIN_CHARS = 6
+const PROMPT_FRAGMENT_MAX_CHARS = 48
+
+/** The prompt's first non-empty line, normalized, or null when too short to be distinctive. */
+export function buildAgentPromptFragment(prompt: string): string | null {
+  for (const line of prompt.split('\n')) {
+    const collapsed = line.replace(/\s+/g, ' ').trim()
+    if (!collapsed) {
+      continue
+    }
+    return collapsed.length >= PROMPT_FRAGMENT_MIN_CHARS
+      ? collapsed.slice(0, PROMPT_FRAGMENT_MAX_CHARS)
+      : null
+  }
+  return null
+}
+
+export function detectAgentPromptComposerVerdict(
+  screen: AgentPromptComposerScreen | null | undefined,
+  prompt: string
+): AgentPromptComposerVerdict {
+  if (!screen || screen.lines.length === 0) {
+    return 'unknown'
+  }
+  const fragment = buildAgentPromptFragment(prompt)
+  // Why: a draft the operator typed is not the payload; another Enter would submit it, not ours.
+  const draft = screen.draft?.trim()
+  if (draft) {
+    return holdsInjectedPayload(draft, fragment) ? 'pending' : 'clear'
+  }
+  const composerText = findLastComposerLineText(screen.lines)
+  if (composerText === null) {
+    return 'unknown'
+  }
+  return holdsInjectedPayload(composerText, fragment) ? 'pending' : 'clear'
+}
+
+function holdsInjectedPayload(composerText: string, fragment: string | null): boolean {
+  return (
+    PASTE_PLACEHOLDER_RE.test(composerText) ||
+    (fragment !== null && composerText.includes(fragment))
+  )
+}
+
+// Why the last glyph line only: Claude keeps `> [Pasted text …]` in its transcript after submit,
+// while the empty `❯` below it is the composer that actually answers "is anything parked?".
+function findLastComposerLineText(lines: readonly string[]): string | null {
+  const tail = lines.filter((line) => line.trim().length > 0).slice(-COMPOSER_SCAN_LINES)
+  for (let index = tail.length - 1; index >= 0; index -= 1) {
+    const match = tail[index]!.match(COMPOSER_PROMPT_LINE_RE)
+    if (match) {
+      return match[1]!.trim()
+    }
+  }
+  return null
+}