Canonicalize targets around labels

apps/cli/src/commands/eval/commands/bundle.ts

@@ -91,7 +91,9 @@ function definitionsWithEvalTargetRefs(
 
   const result = [...definitions];
   for (const ref of targetRefs) {
-    if (ref.use_target && !result.some((definition) => definition.name === ref.name)) {
+    if (ref.definition && !result.some((definition) => definition.name === ref.name)) {
+      result.push(ref.definition);
+    } else if (ref.use_target && !result.some((definition) => definition.name === ref.name)) {
       result.push({ name: ref.name, use_target: ref.use_target } as TargetDefinition);
     }
   }

apps/cli/src/commands/eval/targets.ts

@@ -307,7 +307,9 @@ export async function selectMultipleTargets(
   const definitions = [...fileDefinitions];
   if (targetRefs) {
     for (const ref of targetRefs) {
-      if (ref.use_target && !fileDefinitions.some((d) => d.name === ref.name)) {
+      if (ref.definition && !fileDefinitions.some((d) => d.name === ref.name)) {
+        definitions.push(ref.definition);
+      } else if (ref.use_target && !fileDefinitions.some((d) => d.name === ref.name)) {
         definitions.push({ name: ref.name, use_target: ref.use_target } as TargetDefinition);
       }
     }

apps/cli/src/commands/eval/task-bundle.ts

@@ -43,6 +43,27 @@ const SKIPPED_DIR_NAMES = new Set([
   '.beads',
   '.DS_Store',
 ]);
+const AUTHORING_TOP_LEVEL_TARGET_FIELDS = new Set([
+  'label',
+  'provider',
+  'prompts',
+  'transform',
+  'delay',
+  'env',
+  'model',
+  'use_target',
+  'fallback_targets',
+  'grader_target',
+  'max_budget_usd',
+  'workers',
+  'provider_batching',
+  'subagent_mode_allowed',
+  'max_retries',
+  'retry_initial_delay_ms',
+  'retry_max_delay_ms',
+  'retry_backoff_factor',
+  'retry_status_codes',
+]);
 
 export interface TaskBundleTargetSelection {
   readonly evalFileAbsolutePath?: string;
@@ -591,6 +612,44 @@ function uniqueTargetDefinitions(
   return selected;
 }
 
+function serializeTargetDefinition(definition: TargetDefinition): Record<string, unknown> {
+  const target: Record<string, unknown> = { label: definition.name };
+  if (definition.id !== undefined) {
+    target.id = definition.id;
+  }
+  const config: Record<string, unknown> = {};
+
+  for (const [key, value] of Object.entries(definition)) {
+    if (value === undefined || key === 'name' || key === 'id' || key === 'config') {
+      continue;
+    }
+    if (AUTHORING_TOP_LEVEL_TARGET_FIELDS.has(key)) {
+      target[key] = value;
+    } else {
+      config[key] = value;
+    }
+  }
+
+  if (isRecord(definition.config)) {
+    for (const [key, value] of Object.entries(definition.config)) {
+      if (value !== undefined) {
+        config[key] = value;
+      }
+    }
+  }
+  if (Object.keys(config).length > 0) {
+    target.config = config;
+  }
+
+  return target;
+}
+
+function serializeTargetDefinitions(
+  definitions: readonly TargetDefinition[],
+): readonly Record<string, unknown>[] {
+  return definitions.map((definition) => serializeTargetDefinition(definition));
+}
+
 function uniqueTargetNames(selections: readonly TaskBundleTargetSelection[]): readonly string[] {
   const names: string[] = [];
   const seen = new Set<string>();
@@ -962,7 +1021,7 @@ export async function materializeTaskBundle(
     target: options.targetName,
     tests: [evalCase],
   });
-  await writeYamlFile(targetsPath, { targets: targetDefinitions });
+  await writeYamlFile(targetsPath, { targets: serializeTargetDefinitions(targetDefinitions) });
 
   return {
     testDir,
@@ -1033,7 +1092,7 @@ export async function materializeEvalBundle(
     tests: options.tests.map((test) => buildPortableEvalCase(test, rewrites)),
   });
   await writeYamlFile(targetsPath, {
-    targets: uniqueTargetDefinitions(options.targetSelections),
+    targets: serializeTargetDefinitions(uniqueTargetDefinitions(options.targetSelections)),
   });
 
   const manifest = bundleManifest({

apps/cli/src/commands/runs/rerun.ts

@@ -122,6 +122,9 @@ async function readTargetDefinitions(
 }
 
 function targetName(definition: Record<string, unknown>): string | undefined {
+  if (typeof definition.label === 'string' && definition.label.trim().length > 0) {
+    return definition.label.trim();
+  }
   return typeof definition.name === 'string' && definition.name.trim().length > 0
     ? definition.name.trim()
     : undefined;

apps/cli/test/commands/eval/bundle.test.ts

@@ -54,11 +54,11 @@ describe('agentv eval bundle', () => {
     await writeFile(
       path.join(sourceDir, '.agentv', 'targets.yaml'),
       `targets:
-  - name: inherited
+  - label: inherited
     provider: mock
     response: '{"answer":"Mock provider response from inherited target"}'
     fallback_targets: [backup]
-  - name: backup
+  - label: backup
     provider: mock
     response: '{"answer":"Backup mock response"}'
 `,
@@ -153,8 +153,8 @@ tests: ../data/cases.yaml
     expect(input[0]?.content[0]).toEqual({ type: 'file', value: 'files/data/input.txt' });
 
     const bundledTargets = await readFile(path.join(bundleDir, 'targets.yaml'), 'utf8');
-    expect(bundledTargets).toContain('name: inherited');
-    expect(bundledTargets).toContain('name: backup');
+    expect(bundledTargets).toContain('label: inherited');
+    expect(bundledTargets).toContain('label: backup');
 
     await rm(sourceDir, { recursive: true, force: true });
     const run = await runCli(bundleDir, [
@@ -169,6 +169,43 @@ tests: ../data/cases.yaml
     await expectFileExists(path.join(bundleDir, 'run', 'index.jsonl'));
   }, 60_000);
 
+  it('preserves inline eval target object definitions in the bundled target graph', async () => {
+    const sourceDir = path.join(tempDir, 'inline-source');
+    const bundleDir = path.join(tempDir, 'inline-bundle');
+    await mkdir(path.join(sourceDir, '.agentv'), { recursive: true });
+    await mkdir(path.join(sourceDir, 'evals'), { recursive: true });
+    await writeFile(path.join(sourceDir, '.agentv', 'targets.yaml'), 'targets: []\n', 'utf8');
+    await writeFile(
+      path.join(sourceDir, 'evals', 'inline.eval.yaml'),
+      `targets:
+  - label: candidate
+    provider: mock
+    response: '{"answer":"inline bundled response"}'
+tests:
+  - id: inline-case
+    input: hello
+    assertions:
+      - type: contains
+        value: inline
+`,
+      'utf8',
+    );
+
+    const bundle = await runCli(sourceDir, [
+      'eval',
+      'bundle',
+      'evals/inline.eval.yaml',
+      '--out',
+      bundleDir,
+    ]);
+
+    expect(bundle.exitCode).toBe(0);
+    const bundledTargets = await readFile(path.join(bundleDir, 'targets.yaml'), 'utf8');
+    expect(bundledTargets).toContain('label: candidate');
+    expect(bundledTargets).toContain('provider: mock');
+    expect(bundledTargets).toContain('inline bundled response');
+  }, 30_000);
+
   it('reports unbundleable workspace references with their eval location', async () => {
     const sourceDir = path.join(tempDir, 'missing-source');
     const bundleDir = path.join(tempDir, 'missing-bundle');
@@ -177,7 +214,7 @@ tests: ../data/cases.yaml
     await writeFile(
       path.join(sourceDir, '.agentv', 'targets.yaml'),
       `targets:
-  - name: default
+  - label: default
     provider: mock
 `,
       'utf8',

apps/cli/test/commands/grade/grade-prepared.test.ts

@@ -83,7 +83,7 @@ console.log(JSON.stringify({
     path.join(root, '.agentv', 'targets.yaml'),
     `
 targets:
-  - name: codex
+  - label: codex
     provider: cli
     command: bun ./scripts/target.ts
 `,

apps/cli/test/commands/prepare/prepare.test.ts

@@ -60,7 +60,7 @@ await Bun.write(\`\${payload.workspace_path}/\${step}.txt\`, \`\${payload.test_i
     path.join(root, '.agentv', 'targets.yaml'),
     `
 targets:
-  - name: codex
+  - label: codex
     provider: cli
     command: bun ./scripts/target.ts
 `,

apps/cli/test/commands/runs/rerun.test.ts

@@ -30,7 +30,7 @@ interface CliResult {
 }
 
 const DEFAULT_TARGETS = `targets:
-  - name: captured
+  - label: captured
     provider: mock
 `;
 
@@ -121,7 +121,7 @@ async function createBundleFixture(
   await writeFile(
     overrideTargetsPath,
     `targets:
-  - name: local
+  - label: local
     provider: mock
 `,
     'utf8',
@@ -272,7 +272,7 @@ describe('agentv runs rerun', () => {
 
   it('fails clearly for missing env and accepts an explicit env file', async () => {
     const created = await fixture(`targets:
-  - name: captured
+  - label: captured
     provider: cli
     command: \${{ LOCAL_AGENT_COMMAND }}
 `);

apps/cli/test/eval.integration.test.ts

@@ -32,13 +32,13 @@ async function createFixture(): Promise<EvalFixture> {
   const targetsPath = path.join(agentvDir, 'targets.yaml');
   const targetsContent = `$schema: agentv-targets-v2.2
 targets:
-  - name: default
+  - label: default
     provider: mock
-  - name: file-target
+  - label: file-target
     provider: mock
-  - name: cli-target
+  - label: cli-target
     provider: mock
-  - name: codex-target
+  - label: codex-target
     provider: codex
     model: gpt-5-default
 `;
@@ -90,7 +90,7 @@ async function createNestedEnvFixture(): Promise<EvalFixture> {
   const targetsPath = path.join(agentvDir, 'targets.yaml');
   const targetsContent = `$schema: agentv-targets-v2.2
 targets:
-  - name: default
+  - label: default
     provider: mock
 `;
   await writeFile(targetsPath, targetsContent, 'utf8');

apps/web/src/content/docs/docs/targets/cli-provider.mdx

@@ -14,7 +14,7 @@ Because the contract is "we invoke a command and read a file," almost any useful
 ```yaml
 # .agentv/targets.yaml
 targets:
-  - name: my_agent
+  - label: my_agent
     provider: cli
     command: python agent.py --prompt {PROMPT} --out {OUTPUT_FILE}
     grader_target: azure-base   # required if your evals use LLM graders
@@ -70,7 +70,7 @@ echo "Hello, world!" > {OUTPUT_FILE}
 
 | Field | Type | Required | Default | Description |
 |---|---|---|---|---|
-| `name` | string | yes | — | Target identifier used in eval configs. |
+| `label` | string | yes | — | AgentV target name used by eval `target`, CLI `--target`, and comparisons. |
 | `provider` | literal `"cli"` | yes | — | Selects this provider. |
 | `command` | string | yes | — | Shell command template. |
 | `timeout_seconds` | number | no | — | Kill the process if it runs longer than this. |
@@ -89,7 +89,7 @@ For targets where spin-up cost dominates per-case work (e.g. loading a model, au
 
 ```yaml
 targets:
-  - name: batched_agent
+  - label: batched_agent
     provider: cli
     provider_batching: true
     command: python agent.py --batch-in {PROMPT_FILE} --batch-out {OUTPUT_FILE}
@@ -106,12 +106,12 @@ AgentV has no dedicated "oracle" feature because the `cli` provider already comp
 ```yaml
 # .agentv/targets.yaml
 targets:
-  - name: my_agent
+  - label: my_agent
     provider: cli
     command: python agent.py --prompt {PROMPT} --out {OUTPUT_FILE}
     grader_target: azure-base
 
-  - name: oracle
+  - label: oracle
     provider: cli
     command: cp fixtures/{EVAL_ID}.expected.txt {OUTPUT_FILE}
     grader_target: azure-base