@namzu/sdk 5.1.0 → 5.2.0

package/src/directory/__tests__/load-diagnostics.test.ts

@@ -0,0 +1,126 @@
+import { mkdirSync, mkdtempSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { describe, expect, it } from 'vitest'
+
+import { loadDirectory } from '../load.js'
+
+/**
+ * Three refusals this loader claimed to make and did not.
+ *
+ * Each is the same shape: a check that existed, read the wrong thing, and
+ * therefore reported the wrong answer while looking correct in review. A
+ * loader whose entire job is to say what is wrong with a directory has to be
+ * held to the diagnostics it actually emits, not the ones its code appears to
+ * emit.
+ */
+
+function project(files: Record<string, string>): string {
+	const root = mkdtempSync(join(tmpdir(), 'namzu-diag-'))
+	for (const [relative, body] of Object.entries(files)) {
+		const full = join(root, relative)
+		mkdirSync(join(full, '..'), { recursive: true })
+		writeFileSync(full, body)
+	}
+	return root
+}
+
+describe('a tool that would crash later is refused now', () => {
+	it('rejects a default export with no inputSchema', async () => {
+		// `name` and `execute` alone used to pass. The object then registered
+		// clean and died in `toLLMTools()` on `inputSchema._def` — a TypeError
+		// naming neither this file nor this loader.
+		const root = project({
+			'tools/broken.js': `
+export default {
+  name: 'broken',
+  description: 'no schema',
+  execute: async () => ({ success: true }),
+}
+`,
+		})
+
+		const { manifest, diagnostics, ok } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.map((d) => d.code)).toContain('not_a_tool')
+		expect(manifest.tools).toHaveLength(0)
+	})
+
+	it('still accepts a hand-written tool that satisfies the published type', async () => {
+		// The required set and no more: demanding `defineTool`'s extras would
+		// make this loader refuse an object the SDK's own type accepts.
+		const root = project({
+			'tools/fine.js': `
+export default {
+  name: 'fine',
+  description: 'a tool',
+  inputSchema: { parse: (v) => v },
+  execute: async () => ({ success: true }),
+}
+`,
+		})
+
+		const { manifest, ok } = await loadDirectory(root)
+
+		expect(ok).toBe(true)
+		expect(manifest.tools).toHaveLength(1)
+	})
+})
+
+describe('an import failure explains itself', () => {
+	it('gives the extension hint for a module it cannot resolve', async () => {
+		// The hint is chosen from Node's error CODE. It used to be matched
+		// against the message, which never contains the code, so every hint in
+		// the function was unreachable and every author got the bare error.
+		const root = project({
+			'tools/importer.js': `
+import './nowhere-at-all.js'
+export default { name: 'x', description: 'x', inputSchema: {}, execute: async () => ({}) }
+`,
+		})
+
+		const { diagnostics } = await loadDirectory(root)
+
+		const failed = diagnostics.find((d) => d.code === 'module_load_failed')
+		expect(failed).toBeDefined()
+		expect(failed?.cause).toContain('real extension')
+	})
+})
+
+describe('agent.ts metadata is checked, not assumed', () => {
+	it('refuses metadata values that are not strings', async () => {
+		// Typed `Record<string, string>` and admitted on `typeof === 'object'`,
+		// so the values were never looked at and the type was a promise this
+		// loader did not keep.
+		const root = project({
+			'agent.js': 'export default { model: "m", metadata: { count: 1 } }',
+		})
+
+		const { diagnostics, ok } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.some((d) => d.code === 'invalid_config')).toBe(true)
+	})
+
+	it('refuses an array, which typeof calls an object', async () => {
+		const root = project({
+			'agent.js': 'export default { model: "m", metadata: ["a", "b"] }',
+		})
+
+		const { ok } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+	})
+
+	it('accepts a metadata object of strings', async () => {
+		const root = project({
+			'agent.js': 'export default { model: "m", metadata: { team: "core" } }',
+		})
+
+		const { manifest, ok } = await loadDirectory(root)
+
+		expect(ok).toBe(true)
+		expect(manifest.config.metadata).toEqual({ team: 'core' })
+	})
+})

package/src/directory/__tests__/load-directory.test.ts

@@ -0,0 +1,243 @@
+import { mkdirSync, mkdtempSync, symlinkSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { describe, expect, it } from 'vitest'
+
+import { loadDirectory } from '../load.js'
+
+/**
+ * Every case builds a real directory and reads it. A fixture object standing
+ * in for the filesystem would pass while the scan, the import and the
+ * containment check all did the wrong thing — and the whole package is those
+ * three things.
+ *
+ * `.js` is used for tool files rather than `.ts`. Node's type stripping is
+ * real and the loader accepts both, but a test that depends on it would be
+ * testing Node's version rather than this code.
+ */
+
+const CAN_SYMLINK = (() => {
+	try {
+		const probe = mkdtempSync(join(tmpdir(), 'namzu-proj-sym-'))
+		symlinkSync(probe, join(probe, 'self'), 'dir')
+		return true
+	} catch {
+		return false
+	}
+})()
+
+function project(files: Record<string, string>): string {
+	const root = mkdtempSync(join(tmpdir(), 'namzu-proj-'))
+	for (const [relative, body] of Object.entries(files)) {
+		const full = join(root, relative)
+		mkdirSync(join(full, '..'), { recursive: true })
+		writeFileSync(full, body)
+	}
+	return root
+}
+
+const TOOL = (name: string) => `
+export default {
+  name: ${JSON.stringify(name)},
+  description: 'a tool',
+  inputSchema: { parse: (v) => v, safeParse: (v) => ({ success: true, data: v }) },
+  execute: async () => ({ success: true, output: 'ok' }),
+}
+`
+
+describe('reading a project directory', () => {
+	it('loads instructions, config and tools', async () => {
+		const root = project({
+			'instructions.md': 'You are helpful.\n',
+			'agent.js': 'export default { model: "m", temperature: 0.2 }',
+			'tools/search.js': TOOL('search'),
+		})
+
+		const { manifest, ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok, JSON.stringify(diagnostics)).toBe(true)
+		expect(manifest.instructions).toBe('You are helpful.')
+		expect(manifest.config).toMatchObject({ model: 'm', temperature: 0.2 })
+		expect(manifest.tools.map((t) => t.definition.name)).toEqual(['search'])
+	})
+
+	it('reports a broken tool instead of dropping it', async () => {
+		const root = project({
+			'instructions.md': 'hi',
+			'tools/good.js': TOOL('good'),
+			'tools/broken.js': 'this is not valid javascript !!!',
+		})
+
+		const { manifest, ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		// The good one still loaded — one bad file does not lose the rest.
+		expect(manifest.tools.map((t) => t.definition.name)).toEqual(['good'])
+		const failure = diagnostics.find((d) => d.code === 'module_load_failed')
+		expect(failure?.path).toContain('broken')
+		expect(manifest.sources.find((s) => s.id === 'broken')?.outcome).toBe('failed')
+	})
+
+	it('refuses both files when two tools claim one name', async () => {
+		const root = project({
+			'instructions.md': 'hi',
+			'tools/a.js': TOOL('search'),
+			'tools/b.js': TOOL('search'),
+		})
+
+		const { manifest, ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.some((d) => d.code === 'duplicate_tool_name')).toBe(true)
+		// Picking a winner silently is what hides the mistake.
+		expect(manifest.tools).toHaveLength(1)
+	})
+
+	it('rejects a default export that is not a tool', async () => {
+		const root = project({
+			'instructions.md': 'hi',
+			'tools/x.js': 'export default { hello: 1 }',
+		})
+
+		const { ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.some((d) => d.code === 'not_a_tool')).toBe(true)
+	})
+
+	it('warns rather than errors when instructions are absent', async () => {
+		// `runAgent` treats instructions as optional; this package does not get
+		// to overrule the kernel about what a valid agent is.
+		const root = project({ 'tools/a.js': TOOL('a') })
+
+		const { ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(true)
+		expect(diagnostics.find((d) => d.code === 'instructions_missing')?.severity).toBe('warning')
+	})
+
+	it('errors on an instructions file that says nothing', async () => {
+		const root = project({ 'instructions.md': '   \n' })
+
+		const { ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.find((d) => d.code === 'instructions_empty')?.severity).toBe('error')
+	})
+
+	it('reports a malformed config field by name', async () => {
+		const root = project({
+			'instructions.md': 'hi',
+			'agent.js': 'export default { model: 42 }',
+		})
+
+		const { ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.find((d) => d.code === 'invalid_config')?.message).toContain('"model"')
+	})
+
+	it('ignores dot- and underscore-prefixed entries silently', async () => {
+		const root = project({
+			'instructions.md': 'hi',
+			'tools/real.js': TOOL('real'),
+			'tools/_helpers/util.js': 'export const x = 1',
+			'tools/.hidden.js': TOOL('hidden'),
+		})
+
+		const { manifest, ok, diagnostics } = await loadDirectory(root)
+
+		expect(ok).toBe(true)
+		expect(manifest.tools.map((t) => t.definition.name)).toEqual(['real'])
+		expect(diagnostics.some((d) => d.code === 'unscanned_directory')).toBe(false)
+	})
+
+	it('warns about a subdirectory it did not scan', async () => {
+		const root = project({
+			'instructions.md': 'hi',
+			'tools/nested/deep.js': TOOL('deep'),
+		})
+
+		const { diagnostics } = await loadDirectory(root)
+
+		expect(diagnostics.find((d) => d.code === 'unscanned_directory')?.message).toContain('nested')
+	})
+
+	it('reports a missing directory rather than throwing', async () => {
+		const { ok, diagnostics } = await loadDirectory(join(tmpdir(), 'namzu-does-not-exist-xyz'))
+
+		expect(ok).toBe(false)
+		expect(diagnostics[0]?.code).toBe('root_missing')
+	})
+
+	it('throws only on programmer error', async () => {
+		await expect(loadDirectory('' as string)).rejects.toThrow(TypeError)
+	})
+})
+
+describe('modules: skip', () => {
+	it('gives full structure without importing anything', async () => {
+		// The proof is the side effect: a module that would throw on import is
+		// listed, and nothing runs.
+		const root = project({
+			'instructions.md': 'You are helpful.',
+			'agent.js': 'throw new Error("agent.js executed")',
+			'tools/a.js': 'throw new Error("tool executed")',
+		})
+
+		const { manifest, ok } = await loadDirectory(root, { modules: 'skip' })
+
+		expect(ok).toBe(true)
+		expect(manifest.instructions).toBe('You are helpful.')
+		expect(manifest.sources.map((s) => s.id).sort()).toEqual(['a', 'agent', 'instructions'])
+		expect(
+			manifest.sources
+				.filter((s) => s.slot !== 'instructions')
+				.every((s) => s.outcome === 'not_loaded'),
+		).toBe(true)
+		expect(manifest.tools).toHaveLength(0)
+	})
+})
+
+describe('include is an allow-list', () => {
+	it('scans only what was asked for', async () => {
+		const root = project({ 'instructions.md': 'hi', 'tools/a.js': TOOL('a') })
+
+		const { manifest } = await loadDirectory(root, { include: ['tools'] })
+
+		expect(manifest.included).toEqual(['tools'])
+		expect(manifest.instructions).toBe('')
+		expect(manifest.tools).toHaveLength(1)
+	})
+
+	it('scans nothing when the list is empty', async () => {
+		// An allow-list that admits everything when empty is the fail-open
+		// shape this estate removed elsewhere.
+		const root = project({ 'instructions.md': 'hi', 'tools/a.js': TOOL('a') })
+
+		const { manifest } = await loadDirectory(root, {
+			include: [] as unknown as ['tools'],
+		})
+
+		expect(manifest.sources).toHaveLength(0)
+		expect(manifest.tools).toHaveLength(0)
+	})
+})
+
+describe('symlinks are refused, not followed', () => {
+	it.skipIf(!CAN_SYMLINK)('does not import a linked tool file', async () => {
+		const outside = mkdtempSync(join(tmpdir(), 'namzu-outside-'))
+		writeFileSync(join(outside, 'evil.js'), TOOL('evil'))
+		const root = project({
+			'instructions.md': 'hi',
+			'tools/real.js': TOOL('real'),
+		})
+		symlinkSync(join(outside, 'evil.js'), join(root, 'tools', 'linked.js'), 'file')
+
+		const { manifest, diagnostics } = await loadDirectory(root)
+
+		// The file that would be imported is not the file that was listed.
+		expect(manifest.tools.map((t) => t.definition.name)).toEqual(['real'])
+		expect(diagnostics.some((d) => d.code === 'symlink_refused')).toBe(true)
+	})
+})

package/src/directory/__tests__/subagents.test.ts

@@ -0,0 +1,155 @@
+import { mkdirSync, mkdtempSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { describe, expect, it } from 'vitest'
+
+import { deriveSupervisorOptions } from '../derive-supervisor.js'
+import { loadDirectory } from '../load.js'
+
+/**
+ * `SupervisorAgent` needs an `agentIds` roster and a manager that can spawn
+ * them. Nothing led from a directory to either, so a multi-agent system could
+ * be described on disk and not run.
+ *
+ * Delegates are loaded by recursion — a delegate directory has the same shape
+ * as its parent — capped at one level, because how deep a system may fan out
+ * is a topology decision and a directory layout should not make it by default.
+ */
+
+const provider = { id: 'mock', name: 'Mock' } as never
+const agentManager = { sendMessage: async () => ({}) } as never
+
+function tree(files: Record<string, string>): string {
+	const root = mkdtempSync(join(tmpdir(), 'namzu-sub-'))
+	for (const [relative, body] of Object.entries(files)) {
+		const full = join(root, relative)
+		mkdirSync(join(full, '..'), { recursive: true })
+		writeFileSync(full, body)
+	}
+	return root
+}
+
+const TOOL = (name: string) => `
+export default {
+  name: ${JSON.stringify(name)},
+  description: 'a tool',
+  inputSchema: { parse: (v) => v, safeParse: (v) => ({ success: true, data: v }) },
+  execute: async () => ({ success: true, output: 'ok' }),
+}
+`
+
+const SYSTEM = {
+	'instructions.md': 'You coordinate.',
+	'agent.js': 'export default { model: "coordinator-model" }',
+	'agents/researcher/instructions.md': 'You research.',
+	'agents/researcher/agent.js': 'export default { model: "cheap-model" }',
+	'agents/researcher/tools/search.js': TOOL('search'),
+	'agents/writer/instructions.md': 'You write.',
+}
+
+describe('a project that declares delegates', () => {
+	it('loads each one as a project in its own right', async () => {
+		const { manifest, ok, diagnostics } = await loadDirectory(tree(SYSTEM))
+
+		expect(ok, JSON.stringify(diagnostics)).toBe(true)
+		expect(manifest.agents.map((a) => a.id).sort()).toEqual(['researcher', 'writer'])
+
+		const researcher = manifest.agents.find((a) => a.id === 'researcher')
+		expect(researcher?.manifest.instructions).toBe('You research.')
+		expect(researcher?.manifest.tools.map((t) => t.definition.name)).toEqual(['search'])
+		expect(researcher?.manifest.config.model).toBe('cheap-model')
+	})
+
+	it('surfaces a delegate failure in the parent list', async () => {
+		// A delegate that could not load is a fact about THIS project. A caller
+		// reading one list should not have to walk the tree to discover the run
+		// will be short a specialist.
+		const { manifest, ok, diagnostics } = await loadDirectory(
+			tree({ ...SYSTEM, 'agents/writer/tools/bad.js': 'this is not valid !!!' }),
+		)
+
+		expect(ok).toBe(false)
+		expect(diagnostics.some((d) => d.message.startsWith('agents/writer:'))).toBe(true)
+		// The broken one is not offered as a delegate.
+		expect(manifest.agents.map((a) => a.id)).toEqual(['researcher'])
+	})
+
+	it('refuses to read a second level of delegates', async () => {
+		const { manifest, diagnostics } = await loadDirectory(
+			tree({
+				...SYSTEM,
+				'agents/researcher/agents/deeper/instructions.md': 'You are too deep.',
+			}),
+		)
+
+		const researcher = manifest.agents.find((a) => a.id === 'researcher')
+		expect(researcher?.manifest.agents).toHaveLength(0)
+		expect(diagnostics.some((d) => d.code === 'subagent_too_deep')).toBe(true)
+	})
+
+	it('imports nothing under modules: skip, delegates included', async () => {
+		const { manifest, ok } = await loadDirectory(
+			tree({
+				'instructions.md': 'hi',
+				'agents/x/instructions.md': 'hi',
+				'agents/x/tools/boom.js': 'throw new Error("executed")',
+			}),
+			{ modules: 'skip' },
+		)
+
+		expect(ok).toBe(true)
+		expect(manifest.sources.some((s) => s.slot === 'agents')).toBe(true)
+	})
+})
+
+describe('deriving supervisor options', () => {
+	it('builds the roster from the directory', async () => {
+		const { manifest } = await loadDirectory(tree(SYSTEM))
+
+		const { config, delegates } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		expect(config.agentIds.sort()).toEqual(['researcher', 'writer'])
+		expect(config.systemPrompt).toBe('You coordinate.')
+		expect(config.model).toBe('coordinator-model')
+		expect(delegates).toHaveLength(2)
+	})
+
+	it('lets a delegate keep its own model and inherit when it has none', async () => {
+		// A cheap model for a narrow job is the common case; inheriting
+		// unconditionally would bill every specialist at the coordinator rate.
+		const { manifest } = await loadDirectory(tree(SYSTEM))
+		const { delegates } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		expect(delegates.find((d) => d.id === 'researcher')?.model).toBe('cheap-model')
+		expect(delegates.find((d) => d.id === 'writer')?.model).toBe('coordinator-model')
+	})
+
+	it('carries each delegate its own instructions and tools', async () => {
+		const { manifest } = await loadDirectory(tree(SYSTEM))
+		const { delegates } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		const researcher = delegates.find((d) => d.id === 'researcher')
+		expect(researcher?.systemPrompt).toBe('You research.')
+		expect(researcher?.tools.has('search')).toBe(true)
+	})
+
+	it('refuses a project with no delegates', async () => {
+		// An empty roster makes create_task unmountable, so the result would be
+		// a coordinator that cannot coordinate — and the caller asked for one.
+		const { manifest } = await loadDirectory(
+			tree({ 'instructions.md': 'hi', 'agent.js': 'export default { model: "m" }' }),
+		)
+
+		expect(() => deriveSupervisorOptions(manifest, { provider, agentManager })).toThrow(
+			/declares no delegates/,
+		)
+	})
+
+	it('refuses a manifest whose modules were never loaded', async () => {
+		const { manifest } = await loadDirectory(tree(SYSTEM), { modules: 'skip' })
+
+		expect(() => deriveSupervisorOptions(manifest, { provider, agentManager })).toThrow(
+			/modules: "skip"/,
+		)
+	})
+})

package/src/directory/__tests__/supervisor-budget.test.ts

@@ -0,0 +1,117 @@
+import { mkdirSync, mkdtempSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { describe, expect, it } from 'vitest'
+
+import { deriveSupervisorOptions } from '../derive-supervisor.js'
+import { loadDirectory } from '../load.js'
+
+/**
+ * `BaseAgentConfig` declares `tokenBudget` and `timeoutMs` as REQUIRED, and
+ * `deriveSupervisorOptions` supplied them only when `agent.ts` happened to
+ * name them — which is the uncommon case. An `as SupervisorAgentConfig` made
+ * that compile, so the returned object was typed `tokenBudget: number` while
+ * holding `undefined`.
+ *
+ * The consequence is not a type-level nicety. `buildLimitConfig` defaults only
+ * `maxIterations`, so an undefined budget and timeout disable BOTH hard stops
+ * — a project-derived supervisor ran with no token cap and no wall clock. And
+ * the child-spawn guard computes a delegate's allocation from the parent
+ * budget, so `undefined` became `NaN`, and `NaN <= 0` is false: the refusal
+ * that exists to stop an unfunded child let it through.
+ *
+ * The cast is now `satisfies`, which is what keeps this from recurring — but
+ * a type check is not a test, so the values are asserted here too.
+ */
+
+const provider = { id: 'mock', name: 'Mock' } as never
+const agentManager = { sendMessage: async () => ({}) } as never
+
+function tree(files: Record<string, string>): string {
+	const root = mkdtempSync(join(tmpdir(), 'namzu-budget-'))
+	for (const [relative, body] of Object.entries(files)) {
+		const full = join(root, relative)
+		mkdirSync(join(full, '..'), { recursive: true })
+		writeFileSync(full, body)
+	}
+	return root
+}
+
+const MINIMAL = {
+	'agent.js': 'export default { model: "m" }',
+	'agents/researcher/agent.js': 'export default { model: "m2" }',
+	'agents/researcher/instructions.md': 'Research things.',
+}
+
+describe('a derived supervisor always has a budget and a clock', () => {
+	it('defaults both when agent.ts names neither', async () => {
+		const { manifest } = await loadDirectory(tree(MINIMAL))
+		const { config } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		// The exact numbers matter less than that they are finite: a NaN or an
+		// undefined here is what silently disabled the limits.
+		expect(Number.isFinite(config.tokenBudget)).toBe(true)
+		expect(Number.isFinite(config.timeoutMs)).toBe(true)
+		expect(config.tokenBudget).toBeGreaterThan(0)
+		expect(config.timeoutMs).toBeGreaterThan(0)
+	})
+
+	it('uses the same defaults the SDK front door uses', async () => {
+		// Imported from the module that owns them rather than through the
+		// package entry: this file now lives inside that package, and a
+		// self-import resolves to the built `dist`, not to the source under
+		// test — which is how a stale build reads as a passing test.
+		const { DEFAULT_TIMEOUT_MS, DEFAULT_TOKEN_BUDGET } = await import('../../agents/runAgent.js')
+		const { manifest } = await loadDirectory(tree(MINIMAL))
+		const { config } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		// Two front doors that default differently is a difference nobody would
+		// find until a run behaved unlike its sibling.
+		expect(config.tokenBudget).toBe(DEFAULT_TOKEN_BUDGET)
+		expect(config.timeoutMs).toBe(DEFAULT_TIMEOUT_MS)
+	})
+
+	it('still honours what agent.ts declares', async () => {
+		const { manifest } = await loadDirectory(
+			tree({
+				...MINIMAL,
+				'agent.js': 'export default { model: "m", tokenBudget: 1234, timeoutMs: 5678 }',
+			}),
+		)
+		const { config } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		expect(config.tokenBudget).toBe(1234)
+		expect(config.timeoutMs).toBe(5678)
+	})
+
+	it('lets an override win over both', async () => {
+		const { manifest } = await loadDirectory(tree(MINIMAL))
+		const { config } = deriveSupervisorOptions(manifest, {
+			provider,
+			agentManager,
+			overrides: { tokenBudget: 42 },
+		})
+
+		expect(config.tokenBudget).toBe(42)
+	})
+})
+
+describe('a derived supervisor carries the skills it loaded', () => {
+	it('passes skills through instead of dropping them', async () => {
+		const { manifest } = await loadDirectory(
+			tree({
+				...MINIMAL,
+				'skills/plan-a-trip/SKILL.md':
+					'---\nname: plan-a-trip\ndescription: Plan a trip\n---\n\nAsk for dates.',
+			}),
+		)
+		expect(manifest.skills).toHaveLength(1)
+
+		const { config } = deriveSupervisorOptions(manifest, { provider, agentManager })
+
+		// Loaded from disk, put on the manifest, and then — before this — left
+		// out of the config the supervisor actually runs with.
+		expect(config.skills).toHaveLength(1)
+		expect(config.skills?.[0]?.metadata.name).toBe('plan-a-trip')
+	})
+})