diff --git a/src/shared/agent-title-identity-characterization.test.ts b/src/shared/agent-title-identity-characterization.test.ts
new file mode 100644
index 00000000000..f018f41db8b
--- /dev/null
+++ b/src/shared/agent-title-identity-characterization.test.ts
@@ -0,0 +1,112 @@
+import { describe, expect, it } from 'vitest'
+import { getAgentLabel } from './agent-title-identity'
+
+/**
+ * Characterization of `getAgentLabel` before the identity refactor.
+ *
+ * `getAgentLabel` is an ordered first-match-wins scan of substring predicates over a display
+ * title, so chain position — not evidence strength — decides identity. These tests pin the
+ * current answers, including the wrong ones, so the resolver change shows up as a reviewable
+ * diff of assertions rather than as silent behavior drift.
+ *
+ * Cases marked DEFECT are minimized from real recorded pane titles
+ * (`terminal-history/<id>/checkpoint.json` -> `lastTitle`). At the time of writing the live
+ * corpus held 2,846 checkpoints / 1,084 populated titles / 747 distinct, of which 5 real titles
+ * resolved to the wrong agent. Task text is minimized here; the corpus stays local.
+ */
+
+/** Orca's own owner suffix: the agent that owns the pane is named after the final `- `. */
+const ownerSuffix = (task: string, agent: string): string => `${task}… - ${agent}`
+
+describe('getAgentLabel — characterization (pre-refactor)', () => {
+  describe('the owner suffix loses to a foreign name in task text', () => {
+    // DEFECT. In each case the pane owner is Grok, named by Orca's own `- grok` suffix
+    // grammar, while the competing agent appears only inside free-form task text. Codex is
+    // checked before Grok, so the weaker evidence wins.
+    it.each([
+      ['Switch Claude and Codex off the load balancer', 'Codex'],
+      ['Codex structured chat revalidation', 'Codex'],
+      ['Swap Codex off the load balancer', 'Codex']
+    ])('reads %j as %s instead of Grok', (task, current) => {
+      expect(getAgentLabel(ownerSuffix(task, 'grok'))).toBe(current)
+    })
+
+    it('reads a spinner-prefixed Grok pane as Codex', () => {
+      // DEFECT. Real shape: a status spinner and phase precede the task text.
+      expect(getAgentLabel(`⠸ - Thinking - ${ownerSuffix('Codex native-chat work', 'grok')}`)).toBe(
+        'Codex'
+      )
+    })
+
+    it('reads a Grok pane as Gemini CLI when its task text names two other agents', () => {
+      // DEFECT, and the inverse of the reported Antigravity bug: the pane is Grok, the task
+      // text mentions Antigravity and Gemini, and Gemini CLI is checked earliest of the three.
+      expect(
+        getAgentLabel(ownerSuffix('Electron QA: Antigravity tab vs Gemini label', 'grok'))
+      ).toBe('Gemini CLI')
+    })
+
+    it('resolves the owner suffix correctly only when no earlier agent is named', () => {
+      // Why this passes today: nothing earlier in the chain matches, so position never comes up.
+      expect(getAgentLabel(ownerSuffix('Fix the sidebar row', 'grok'))).toBe('Grok')
+    })
+  })
+
+  describe('a hyphenated worktree name is correctly not identity', () => {
+    // Not a defect — pinned so the resolver does not start claiming these.
+    it.each(['review-14600-codex', 'sta4779-review-codex', 'codex-split-core'])(
+      'declines %j',
+      (title) => {
+        expect(getAgentLabel(title)).toBeNull()
+      }
+    )
+  })
+
+  describe('chain position, not evidence strength, decides between two names', () => {
+    // The pairwise property: whichever agent is checked first wins, regardless of which one
+    // the title's grammar actually identifies. Both orderings of the same pair agree, which is
+    // the tell — the title carries no signal that distinguishes them.
+    it.each([
+      ['codex', 'grok', 'Codex'],
+      ['grok', 'codex', 'Codex'],
+      ['gemini', 'antigravity', 'Gemini CLI'],
+      ['antigravity', 'gemini', 'Gemini CLI'],
+      ['copilot', 'devin', 'GitHub Copilot'],
+      ['devin', 'copilot', 'GitHub Copilot']
+    ])('%s + %s both resolve to %s', (first, second, winner) => {
+      expect(getAgentLabel(`${first} and ${second}`)).toBe(winner)
+    })
+  })
+
+  describe('a vendor glyph is claimed by whichever check sits earliest', () => {
+    it('claims a Claude glyph even when the task text names another agent', () => {
+      // Correct today (the pane really is Claude) but for the wrong reason: the glyph is not
+      // consulted as vendor evidence, the `✳ ` prefix branch simply sits first.
+      expect(getAgentLabel('✳ Fix Codex false attention notifications on Windows')).toBe(
+        'Claude Code'
+      )
+    })
+
+    it('gives a bare agent name no way to outrank an earlier glyph check', () => {
+      // DEFECT. `agy` is the entire undecorated remainder — the strongest possible name
+      // evidence — and still loses to Claude's prefix glyph.
+      expect(getAgentLabel('✳ agy')).toBe('Claude Code')
+    })
+  })
+
+  describe('Antigravity model names are read as Gemini CLI', () => {
+    it('reads an Antigravity model title as Gemini CLI', () => {
+      // DEFECT. Antigravity models are named `Gemini <n.n> <Name>`, so an agy pane's own title
+      // carries a whole `gemini` token, and Gemini CLI is checked first.
+      expect(getAgentLabel('Gemini 3.7 Flash · high')).toBe('Gemini CLI')
+    })
+
+    it('still reads Gemini CLI even when the Antigravity name is also present', () => {
+      // DEFECT. `AGY_AGENT_NAME_RE` is checked in the Antigravity branch at position 12, so an
+      // explicit `agy` identity segment cannot outrank the model name at position 4. This is the
+      // case PR #15535 patches by teaching the Gemini detector to decline; the resolver instead
+      // treats the agy segment as the anchored identity and `Gemini <n.n> <Name>` as metadata.
+      expect(getAgentLabel('agy · Gemini 3.7 Flash')).toBe('Gemini CLI')
+    })
+  })
+})