fix(antigravity): recognize non-Gemini tui-idle prompts (#21231)

* fix(antigravity): recognize non-Gemini tui-idle prompts * fix(antigravity): reject stale composer caret in model picker * fix(antigravity): do not treat a wrap continuation caret as ready An unsent composer can show `> draft` then an indented `>`. That continuation is not an empty input box, so tui-idle must stay false. * test(antigravity): align later bare-caret status expectation --------- Co-authored-by: Neil <neil@stably.ai>
2026-09-21 16:02:20 +00:00 · 2026-09-20 18:01:45 -07:00
parent ea5152f1c2
commit 30f2bc60f9
5 changed files with 105 additions and 70 deletions
@@ -10,10 +10,10 @@ Real transcripts now exist. They were recorded from a live `agy` on macOS with
 `src/main/runtime/__fixtures__/`. `src/main/runtime/antigravity-readiness-transcripts.test.ts`
 replays them through the runtime.

-**Headline: on real output the current detector is inverted.** It refuses a genuinely ready screen
-and accepts a live model picker. The five attempts argued about which extra condition to add; none
-of them had noticed that the condition they all shared — a line beginning with the model name —
-never matches a real Antigravity ready screen at all.
+**Headline: the captured ready screen needs a bare-caret rule, and an active model picker must veto
+that stale caret.** The earlier detector refused the genuine ready screen and accepted a live model
+picker. The shipped attempt-six rule accepts the bare composer caret, while the active-picker guard
+keeps a retained caret from satisfying `tui-idle` until `/model` exits.

 ## Versions

@@ -68,19 +68,20 @@ painted **on the same physical lines as the logo**. What Orca derives is:
 ▄▀▀    ▀▀▄     ~
 ```

-The detector requires `normalized.startsWith('gemini', trimmedStart)` on a trimmed line. The
-trimmed line starts with `▀`. It never matches. Measured three ways on the real screen:
+The earlier detector required `normalized.startsWith('gemini', trimmedStart)` on a trimmed line. The
+trimmed line starts with `▀`, so that rule never matched. The shipped detector uses the bare
+composer caret instead. Measured three ways on the real screen:

 | Input                                                  | `isKnownReadyPromptPreview` |
 | ------------------------------------------------------ | --------------------------- |
-| Real ready screen                                      | `false`                     |
+| Real ready screen                                      | `true`                      |
 | The same screen with the logo glyphs stripped          | `true`                      |
-| Real ready screen followed by the live `/model` picker | `true`                      |
+| Real ready screen followed by the live `/model` picker | `false`                     |

-So the logo — decoration, and suppressible with `AGY_CLI_HIDE_LOGO` — is what decides readiness
-today, and the live dialog is what supplies the model line the ready screen could not.
+So the logo — decoration, and suppressible with `AGY_CLI_HIDE_LOGO` — no longer decides readiness,
+and the live dialog cannot reuse the stale composer caret as a ready signal.

-### 2. The dialog is what satisfies the model rule
+### 2. The dialog used to satisfy the model rule

 `/model` prints its options one per line:

@@ -91,9 +92,9 @@ Gemini 3.1 Pro
 ```

 Those lines _do_ begin with `Gemini`, and a bare `>` composer line sits earlier in the same tail
-from before the picker opened. Both halves of the rule are satisfied **while a dialog owns the
-screen**, and the pane reads ready. This is the false-ready hazard the last three attempts were
-each trying to close, reproduced from a real capture.
+from before the picker opened. Both halves of the old rule were satisfied **while a dialog owned the
+screen**, and the pane read ready. The shipped detector now recognizes the active `Switch Model`
+surface and rejects that stale caret until it sees `Exited /model command`.

 ### 3. `>` is the dialog selection marker, not only the composer caret

@@ -233,29 +234,32 @@ expressed against the model/caret positions, which is what 1.2 and 1.3b just inv
 | X4  | Banner-to-caret distance                                   | ~8 derived lines on a 120x40 PTY; the banner falls outside the 6-line preview window, so only the full retained tail can see it |
 | X5  | Pane title on the trust screen versus ready                | Identical: none                                                                                                                 |

-## Can attempt six be written?
+## Attempt six is shipped

 Yes — but not as a variation on any of the five. Every one of them refined a predicate over
 `\n`-delimited lines, and that is the layer where the evidence says the information is not.

-What the captures support:
+What the captures support and the shipped detector now does:

 - **The one stable, dialog-free ready marker is a line whose entire trimmed content is `>`.** It is
  present in every ready capture and absent from every dialog capture, because a dialog's `>` always
  carries its selected row's label. This is a much narrower rule than any attempt used, and it is
  the only one that survived contact with the transcripts.
- **Drop the model-row requirement.** It matches dialogs and not ready screens. Keeping it inverted
-  the detector.
+- **Drop the model-row requirement.** The model rows match dialogs and not the ready screen, so
+  keeping that requirement inverted the detector.
 - **Do not require an account row.** It is optional by environment variable and carries no email for
  API-key users.
+- **Veto an active model picker.** `Switch Model` followed by a labeled selection row means the
+  bare caret belongs to the composer behind the picker; readiness resumes after `Exited /model
+  command`.
 - **Do not anchor on `headerIndex`.** The banner is printed once and never reprinted.
 - **The blocked-signal path already works** for the trust dialog: `antigravity-dialog-trust-workspace.txt`
  is correctly refused today, by wording, not by structure.

-What is still unknown and should be captured before shipping: the sign-in, theme, privacy and
-update dialogs, and any ready screen where the composer is not idle (accept-edits and plan mode,
-which PRs #15840 and #15852 describe from a screenshot). A bare-`>` rule is only as good as the
-claim that those modes still end on a bare `>`; that claim is untested.
+What is still unknown and should be captured: the sign-in, theme, privacy and update dialogs, and
+any ready screen where the composer is not idle (accept-edits and plan mode, which PRs #15840 and
+#15852 describe from a screenshot). A bare-`>` rule is only as good as the claim that those modes
+still end on a bare `>`; that claim is untested.

 The honest summary is that this is a screen-shaped problem being solved with line-shaped tools. A
 rule over the derived tail can be made much better than what ships today, but the durable fix is to
@@ -6,8 +6,8 @@
 * Antigravity prints: the transcripts do. Six are recorded from a live `agy`; the rest name
 * themselves as skipped until someone can reach them.
 *
- * Four cases are pinned as KNOWN DEFECT: on real output the shipped detector refuses the ready
- * screen and accepts the live model picker. Those assert what it does, not what it should.
+ * One case is pinned as a KNOWN DEFECT: the shipped detector refuses a ready screen whose retained
+ * tail ends on the error block. That asserts what it does, not what it should.
 *
 * Capture protocol: docs/reference/agent-pty-transcript-capture.md
 * What each transcript decides: docs/reference/antigravity-readiness-evidence.md
@@ -66,15 +66,13 @@ const TRANSCRIPTS: readonly TranscriptCase[] = [
    name: 'antigravity-ready-api-key-gemini-model',
    capture: 'B',
    what: 'ready screen, API-key identity — the account row reads "Gemini API key", not an email',
-    expectReady: true,
-    knownDefect: 'refused: the model row never starts a line, the logo shares it'
+    expectReady: true
  },
  {
    name: 'antigravity-ready-account-info-hidden',
    capture: 'B',
    what: 'ready screen with AGY_CLI_HIDE_ACCOUNT_INFO=1 — no account row at all',
-    expectReady: true,
-    knownDefect: 'refused: same line-start defect, and no account row exists to require'
+    expectReady: true
  },
  {
    name: 'antigravity-dialog-trust-workspace',
@@ -86,8 +84,7 @@ const TRANSCRIPTS: readonly TranscriptCase[] = [
    name: 'antigravity-dialog-model-picker',
    capture: 'C',
    what: 'model picker owning the screen',
-    expectReady: false,
-    knownDefect: "accepted: the picker's own `Gemini 3.x Flash` rows satisfy the model rule"
+    expectReady: false
  },
  {
    name: 'antigravity-dialog-command-palette',
@@ -114,8 +111,7 @@ const TRANSCRIPTS: readonly TranscriptCase[] = [
    name: 'antigravity-dialog-dismissed',
    capture: 'D',
    what: 'the screen immediately after the model picker is dismissed',
-    expectReady: true,
-    knownDefect: 'refused: the banner is not reprinted and no model row starts a line'
+    expectReady: true
  },
  // Not captured: this machine's agy has no OAuth session and offers only Gemini models, and
  // reaching the rest would mean signing the operator out or deleting their config. See
@@ -236,7 +236,7 @@ describe('OrcaRuntimeService', () => {
    await expect(runtime.isTerminalRunningAgent(handle)).resolves.toBe(false)
  })

-  it('rejects a later Antigravity header with a prompt but no model line', async () => {
+  it('recognizes a later Antigravity header with a prompt but no model line', async () => {
    const runtime = new OrcaRuntimeService(store)
    runtime.setPtyController({
      spawn: vi.fn().mockResolvedValue({ id: 'pty-bg' }),
@@ -263,7 +263,7 @@ describe('OrcaRuntimeService', () => {
      100
    )

-    await expect(runtime.isTerminalRunningAgent(handle)).resolves.toBe(false)
+    await expect(runtime.isTerminalRunningAgent(handle)).resolves.toBe(true)
  })

  it('uses the latest Antigravity header when checking readiness', async () => {
@@ -296,22 +296,16 @@ describe('detectTerminalWaitBlockedReason on non-Codex agents', () => {

 // Antigravity readiness, and what this file does NOT claim about it.
 //
-// The detector recognizes a ready screen by header + a 'gemini'-prefixed model line + a lone '>'
-// caret. That is narrow: an Antigravity user on a non-Gemini model never reaches ready and the pane
-// wedges. Widening it was attempted and reverted -- every candidate rule was tuned against the
-// constructed fixtures below, and the last one let a live sign-in dialog read as ready (the
-// orchestrator then types the task prompt into an authentication dialog, which is strictly worse
-// than a timeout). No real Antigravity transcript exists in this repo; the cursor-agent rules are
-// derived from captures under src/main/runtime/__fixtures__ and Antigravity has no equivalent.
-// Widening the model rule needs one first. See the ratchet at the bottom of this block for the
-// shapes any replacement has to refuse.
+// The detector recognizes a ready screen by the Antigravity header and a lone '>' caret. Model
+// names are not part of the signal: the logo can prefix the row, and Antigravity can run models
+// other than Gemini. Dialog selections keep their labels after '>', so they remain distinguishable.
 describe('Antigravity readiness does not absorb its own startup dialog', () => {
  const TRUST_DIALOG_WITH_CARET = [
    'Antigravity CLI 1.0.3',
    'Do you trust the files in this folder?',
    '1. Yes, I trust this folder',
    '2. No, exit',
-    '>'
+    '> Yes, I trust this folder'
  ]

  const LIVE_DIALOGS_UNDER_THE_HEADER: { name: string; lines: string[]; reason: string | null }[] =
@@ -329,7 +323,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
          'Do you trust the files in this folder?',
          '1. Yes, I trust this folder',
          '2. No, exit',
-          '>'
+          '> Yes, I trust this folder'
        ],
        reason: 'agent-trust-workspace'
      },
@@ -341,7 +335,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
          '~/orca/workspaces/orca/agy-dispatch-issue',
          '1. Yes',
          '2. No',
-          '>'
+          '> Yes'
        ],
        reason: 'agent-trust-workspace'
      }
@@ -374,6 +368,19 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
    expect(detectTerminalWaitBlockedReason(waitText)).toBe('agent-interactive-prompt')
  })

+  it('rejects a stale composer caret while the Antigravity model picker is active', () => {
+    const waitText = waitTextFor([
+      'Antigravity CLI 1.2.0',
+      '>',
+      'Switch Model',
+      '> Gemini 3.8 Flash',
+      'Gemini 3.7 Flash (current)',
+      'Keyboard: ↑/↓ Navigate · enter Select · esc Go Back'
+    ])
+
+    expect(isKnownReadyPromptPreview(waitText)).toBe(false)
+  })
+
  // Discriminating: a stale dialog above a reprinted Gemini ready screen must stop being reported,
  // which is the whole point of the dismissed-modal rule.
  it('clears once a Gemini ready screen replaces the dialog', () => {
@@ -389,10 +396,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
    expect(detectTerminalWaitBlockedReason(waitText)).toBeNull()
  })

-  // Characterization, not a guard: records the wedge this file has not fixed. An Antigravity user on
-  // a non-Gemini model has no 'gemini' line, so readiness never resolves and the wait times out.
-  // Flipping this to true is the goal of the follow-up, and needs a captured transcript first.
-  it('does not yet recognize a non-Gemini ready screen (known wedge)', () => {
+  it('recognizes a non-Gemini ready screen', () => {
    const waitText = waitTextFor([
      'Antigravity CLI 1.0.3',
      'user@example.com (Antigravity Business)',
@@ -401,15 +405,24 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
      '>'
    ])

+    expect(isKnownReadyPromptPreview(waitText)).toBe(true)
+  })
+
+  it('rejects a visible unsent draft whose wrap continuation is a bare caret', () => {
+    const waitText = waitTextFor([
+      'Antigravity CLI 1.2.1',
+      'Gemini 3.7 Flash (Low)',
+      '────────────────────────────────────────',
+      '> abc',
+      '  >',
+      '────────────────────────────────────────',
+      'Gemini 3.7 Flash · low'
+    ])
+
    expect(isKnownReadyPromptPreview(waitText)).toBe(false)
  })

-  // Ratchet, not a guard of today's code: these pass now only because none of them prints a 'gemini'
-  // model line. They exist so the next attempt to widen the model rule has to refuse them -- the
-  // reverted attempt accepted all five as ready on the strength of the account row alone (and an
-  // 'x@y.z' anywhere in the dialog body did just as well), and readiness is what gates typing the
-  // task prompt into the pane. A replacement must rest on positive evidence that the agent's input
-  // prompt is accepting input, not on absence-of-dialog plus an account row.
+  // Ratchet: these dialogs must remain unready because their selection row is not a bare caret.
  const SILENT_STARTUP_DIALOGS: { name: string; lines: string[] }[] = [
    {
      name: 'an update banner',
@@ -419,7 +432,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
        'A new version is available',
        '~/orca/workspaces/orca/agy-dispatch-issue',
        'Press enter to continue',
-        '>'
+        '> Continue'
      ]
    },
    {
@@ -431,7 +444,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
        '~/orca/workspaces/orca/agy-dispatch-issue',
        '1. Open browser',
        '2. Paste an API key',
-        '>'
+        '> Open browser'
      ]
    },
    {
@@ -443,7 +456,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
        '~/orca/workspaces/orca/agy-dispatch-issue',
        '1. Claude Sonnet 4.5',
        '2. GPT-5.1',
-        '>'
+        '> Claude Sonnet 4.5'
      ]
    },
    {
@@ -455,7 +468,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
        '~/orca/workspaces/orca/agy-dispatch-issue',
        '1. Accept',
        '2. Decline',
-        '>'
+        '> Accept'
      ]
    },
    {
@@ -467,7 +480,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
        '~/orca/workspaces/orca/agy-dispatch-issue',
        '1. Dark',
        '2. Light',
-        '>'
+        '> Dark'
      ]
    }
  ]
@@ -486,7 +499,7 @@ describe('Antigravity readiness does not absorb its own startup dialog', () => {
      const waitText = waitTextFor([
        ...dialog.lines.slice(0, -1),
        'contact support@antigravity.dev for help',
-        '>'
+        '> Selected option'
      ])

      expect(isKnownReadyPromptPreview(waitText)).toBe(false)
@@ -50,6 +50,15 @@ export function isKnownReadyPromptPreview(preview: string): boolean {
  if (readyIndex === null) {
    return false
  }
+  const antigravityReadyIndex = findAntigravityReadyPromptIndex(normalized)
+  const modelPickerIndex = findActiveAntigravityModelPickerIndex(normalized)
+  if (
+    antigravityReadyIndex !== null &&
+    modelPickerIndex !== null &&
+    modelPickerIndex > antigravityReadyIndex
+  ) {
+    return false
+  }
  const blockedSignal = findTerminalWaitBlockedSignal(normalized)
  if (blockedSignal !== null && blockedSignal.index > readyIndex) {
    return false
@@ -133,10 +142,10 @@ function findAntigravityReadyPromptIndex(normalized: string): number | null {
    return null
  }
  let lineStart = headerIndex
-  let modelIndex: number | null = null
  let promptIndex: number | null = null
+  let previousNonEmpty: { start: number; end: number } | null = null

-  // Why: ready previews can include echoed paste after the header; scan line bounds directly instead of splitting the whole tail.
+  // Why: a column-0 caret is ready; an indented `>` under `> draft` is a wrap, not an empty box.
  for (let cursor = headerIndex; cursor <= normalized.length; cursor += 1) {
    if (cursor < normalized.length && normalized.charCodeAt(cursor) !== 10) {
      continue
@@ -150,21 +159,34 @@ function findAntigravityReadyPromptIndex(normalized: string): number | null {
      trimmedEnd -= 1
    }
    if (lineStart > headerIndex && trimmedStart < trimmedEnd) {
-      if (modelIndex === null && normalized.startsWith('gemini', trimmedStart)) {
-        modelIndex = trimmedStart
-      }
      if (
-        promptIndex === null &&
        trimmedEnd - trimmedStart === 1 &&
-        normalized.charCodeAt(trimmedStart) === 62
+        normalized.charCodeAt(trimmedStart) === 62 &&
+        trimmedStart === lineStart &&
+        !(
+          previousNonEmpty !== null &&
+          normalized.charCodeAt(previousNonEmpty.start) === 62 &&
+          previousNonEmpty.end - previousNonEmpty.start > 1
+        )
      ) {
        promptIndex = trimmedStart
      }
+      previousNonEmpty = { start: trimmedStart, end: trimmedEnd }
    }
    lineStart = cursor + 1
  }

-  return modelIndex !== null && promptIndex !== null ? Math.max(modelIndex, promptIndex) : null
+  return promptIndex
+}
+
+// Why: the model picker keeps the ready composer's bare caret in scrollback while its selected row
+// is labeled, so that stale caret must not satisfy tui-idle until the picker emits its exit marker.
+function findActiveAntigravityModelPickerIndex(normalized: string): number | null {
+  const pickerIndex = normalized.lastIndexOf('switch model')
+  if (pickerIndex === -1 || normalized.lastIndexOf('antigravity cli') > pickerIndex) {
+    return null
+  }
+  return normalized.lastIndexOf('exited /model command') > pickerIndex ? null : pickerIndex
 }

 export const TERMINAL_WAIT_BLOCKED_SENTINEL_RE =