@namzu/sdk 0.6.0 → 1.0.0

package/src/constants/tools/index.ts

@@ -1,3 +1,34 @@
-export const DANGEROUS_PATTERNS = [/rm\s+-rf\s+\//, /mkfs/, /dd\s+if=/, /:(){ :\|:& };:/]
+// Patterns the verification gate's `deny_dangerous_patterns` rule
+// matches against the JSON-serialised tool input. The list is
+// intentionally short and high-signal: the goal is to catch the
+// canonical "I will brick the host" mistakes (filesystem wipes,
+// disk reformat, fork bomb) plus the most common shell-side
+// privilege/escape patterns (root sudo, world-writable chmod, the
+// classic curl|bash / wget|bash exfil-then-exec pipe, raw eval).
+//
+// This is NOT a security boundary — Cursor learned the hard way
+// that bash denylists are bypassed via shell tricks like `e""cho`
+// (see Backslash Security 2025). Sandbox enforcement (FS isolation,
+// network egress proxy) is the real boundary; these patterns only
+// catch the most blatant attempts and turn them into an explicit
+// review prompt instead of a silent execute.
+export const DANGEROUS_PATTERNS = [
+	// Filesystem wipe / fork bomb / raw disk write.
+	/rm\s+-rf\s+\//,
+	/mkfs/,
+	/dd\s+if=/,
+	/:(){ :\|:& };:/,
+	// Privilege escalation + world-writable chmod on /.
+	/\bsudo\b/,
+	/\bsu\s+-/,
+	/chmod\s+(?:-R\s+)?777\s+\//,
+	// Pipe-to-shell from network — exfil-then-exec staging.
+	/\bcurl\b[^|]*\|\s*(?:sh|bash|zsh)\b/,
+	/\bwget\b[^|]*\|\s*(?:sh|bash|zsh)\b/,
+	// Remote shell / outbound SSH.
+	/\bssh\s+\S+@/,
+	// Raw eval of dynamic strings.
+	/\beval\s+["'`$]/,
+]
 
-export const FILESYSTEM_TOOLS = new Set(['glob', 'read_file', 'write_file', 'bash'])
+export const FILESYSTEM_TOOLS = new Set(['glob', 'read', 'write', 'bash'])

package/src/manager/run/persistence.ts

@@ -1,8 +1,9 @@
 import { EMPTY_TOKEN_USAGE } from '../../constants/limits.js'
+import { AUTO_CONTINUATION_USER_MESSAGE } from '../../runtime/query/continuation.js'
 import { RunDiskStore } from '../../store/run/disk.js'
 import { type CostInfo, type TokenUsage, accumulateTokenUsage } from '../../types/common/index.js'
 import type { RunId, SessionId, TenantId } from '../../types/ids/index.js'
-import type { AssistantMessage, Message } from '../../types/message/index.js'
+import type { Message } from '../../types/message/index.js'
 import type { EmergencySaveData } from '../../types/run/emergency.js'
 import type { Run, RunPersistenceConfig, StopReason } from '../../types/run/index.js'
 import type { ProjectId, ThreadId } from '../../types/session/ids.js'
@@ -169,12 +170,39 @@ export class RunPersistence {
 	}
 
 	private resolveResult(): void {
-		const lastAssistant = [...this.run.messages]
-			.reverse()
-			.find((m): m is AssistantMessage => m.role === 'assistant' && m.content !== null)
+		// Walk the tail of the message log to assemble the final
+		// assistant output. The iteration loop's auto-continuation
+		// path (see `runtime/query/iteration/index.ts`) inserts a
+		// synthetic user prompt — exactly equal to
+		// `AUTO_CONTINUATION_USER_MESSAGE` — between two assistant
+		// messages whenever a turn ended with
+		// `stop_reason: max_tokens` mid-text. Treat that synthetic
+		// user as transparent: keep collecting assistant content past
+		// it so the run's persisted `result` carries the full
+		// multi-turn output, not just the trailing continuation
+		// chunk. Stops at the first non-assistant, non-marker
+		// message (e.g. the real user prompt that started the run,
+		// or a tool message between turns).
+		const chunks: string[] = []
+		for (let i = this.run.messages.length - 1; i >= 0; i--) {
+			const msg = this.run.messages[i]
+			if (!msg) continue
+			if (msg.role === 'assistant') {
+				if (msg.content !== null) chunks.push(msg.content)
+				continue
+			}
+			if (msg.role === 'user' && msg.content === AUTO_CONTINUATION_USER_MESSAGE) {
+				// Synthetic continuation prompt — skip and keep
+				// collecting the partial that preceded it.
+				continue
+			}
+			break
+		}
 
-		if (lastAssistant?.content) {
-			this.run.result = lastAssistant.content
+		if (chunks.length > 0) {
+			// chunks were collected newest-first; reverse so the
+			// assembled string is chronological.
+			this.run.result = chunks.reverse().join('')
 		}
 	}
 

package/src/persona/assembler.ts

@@ -23,14 +23,9 @@ export function assembleSystemPrompt(persona: AgentPersona, skills?: Skill[]): s
 		}
 	}
 
-	if (skills && skills.length > 0) {
-		const loadedSkills = skills.filter((s) => s.body)
-		if (loadedSkills.length > 0) {
-			const skillSections = loadedSkills.map(
-				(s) => `### ${s.metadata.name}\n**Skill directory:** \`${s.dirPath}\`\n\n${s.body}`,
-			)
-			sections.push(`## Skills\n\n${skillSections.join('\n\n')}`)
-		}
+	const skillsSection = renderSkillsSection(skills)
+	if (skillsSection) {
+		sections.push(skillsSection)
 	}
 
 	if (persona.reflexes?.outputDiscipline) {
@@ -48,6 +43,34 @@ export function assembleSystemPrompt(persona: AgentPersona, skills?: Skill[]): s
 	return sections.join('\n\n')
 }
 
+export function renderSkillsSection(skills?: Skill[]): string | null {
+	if (!skills || skills.length === 0) return null
+
+	const available = skills
+		.map((s) => {
+			const details = [`description: ${s.metadata.description.trim()}`]
+			if (s.metadata.compatibility) details.push(`compatibility: ${s.metadata.compatibility}`)
+			if (s.metadata.license) details.push(`license: ${s.metadata.license}`)
+			if (s.metadata.allowedTools) details.push(`allowed-tools: ${s.metadata.allowedTools}`)
+			return `- ${s.metadata.name} (${details.join('; ')})\n  directory: ${s.dirPath}`
+		})
+		.join('\n')
+
+	const loadedSkills = skills.filter((s) => s.body)
+	const sections = [
+		`## Available Skills\nThese Agent Skills are available through progressive disclosure. Use a skill only when the task matches its description. If a skill is not already loaded below, activate/read its SKILL.md from the listed directory when the runtime provides filesystem or skill-loading access.\n\n${available}`,
+	]
+
+	if (loadedSkills.length > 0) {
+		const skillSections = loadedSkills.map(
+			(s) => `### ${s.metadata.name}\n**Skill directory:** \`${s.dirPath}\`\n\n${s.body}`,
+		)
+		sections.push(`## Loaded Skills\n\n${skillSections.join('\n\n')}`)
+	}
+
+	return sections.join('\n\n')
+}
+
 export function renderOutputDiscipline(discipline: OutputDiscipline): string {
 	const lines: string[] = []
 

package/src/provider/collect.test.ts

@@ -57,7 +57,7 @@ describe('collect()', () => {
 					id: 'm',
 					delta: {
 						toolCalls: [
-							{ index: 0, id: 'toolu_a', function: { name: 'Read' } },
+							{ index: 0, id: 'toolu_a', function: { name: 'read' } },
 							{ index: 1, id: 'toolu_b', function: { name: 'WebSearch' } },
 						],
 					},
@@ -84,7 +84,7 @@ describe('collect()', () => {
 			{
 				id: 'toolu_a',
 				type: 'function',
-				function: { name: 'Read', arguments: '{"file_path":"/a"}' },
+				function: { name: 'read', arguments: '{"file_path":"/a"}' },
 			},
 			{
 				id: 'toolu_b',

package/src/public-runtime.ts

@@ -113,7 +113,11 @@ export {
 	ManagedRegistry,
 	PluginRegistry,
 	Registry,
+	ToolCatalog,
 	ToolRegistry,
+	createToolCatalogFromRegistry,
+	loadingFromAvailability,
+	toolDefinitionToCatalogEntry,
 } from './registry/index.js'
 
 export {
@@ -129,8 +133,11 @@ export {
 	EmergencySaveManager,
 	PlanManager,
 	RunPersistence,
+	ThreadManager,
 } from './manager/index.js'
 
+export { InMemoryThreadStore } from './store/thread/memory.js'
+
 export { LocalTaskGateway } from './gateway/local.js'
 
 // ─── providers, sandbox, vault ───────────────────────────────────────────
@@ -220,7 +227,12 @@ export {
 	FileLockManager,
 } from './bus/index.js'
 
-export { evaluateRule, VerificationGate } from './verification/index.js'
+export {
+	defaultSandboxedGateConfig,
+	defaultSandboxedShellGateConfig,
+	evaluateRule,
+	VerificationGate,
+} from './verification/index.js'
 
 // ─── probe (typed observation over AgentBus + RunEvent stream) ───────────
 
@@ -254,6 +266,7 @@ export {
 	DefaultPathBuilder,
 	GitWorktreeDriver,
 	parseWorktreeList,
+	SharedRunWorkspace,
 	WorkspaceBackendRegistry,
 } from './session/workspace/index.js'
 

package/src/public-tools.ts

@@ -24,6 +24,7 @@ export { GlobTool } from './tools/builtins/glob.js'
 export { GrepTool } from './tools/builtins/grep.js'
 export { LsTool } from './tools/builtins/ls.js'
 export { SearchToolsTool } from './tools/builtins/search-tools.js'
+export { VerifyOutputsTool } from './tools/builtins/verify-outputs.js'
 export {
 	createStructuredOutputTool,
 	STRUCTURED_OUTPUT_TOOL_NAME,
@@ -44,6 +45,7 @@ export {
 export { buildAdvisoryTools } from './tools/advisory/index.js'
 export { buildMemoryTools } from './tools/memory/index.js'
 export { buildCoordinatorTools } from './tools/coordinator/index.js'
+export { buildAgentTool, type AgentToolOptions } from './tools/coordinator/agent.js'
 
 // ─── RAG tool builder ────────────────────────────────────────────────────
 

package/src/public-types.ts

@@ -17,6 +17,7 @@ export type * from './types/ids/index.js'
 export type * from './types/message/index.js'
 export type * from './types/common/index.js'
 export type * from './types/tool/index.js'
+export type * from './types/toolset/index.js'
 export type * from './types/permission/index.js'
 export type * from './types/run/index.js'
 export type * from './types/provider/index.js'
@@ -44,6 +45,7 @@ export type * from './types/verification/index.js'
 export type * from './types/bus/index.js'
 export type * from './types/probe/index.js'
 export type * from './types/doctor/index.js'
+export type * from './types/workspace/index.js'
 
 // Session-hierarchy type surface (ses_010 moved entities here).
 export type * from './types/session/index.js'
@@ -118,6 +120,11 @@ export type { AdvisoryToolsOptions } from './tools/advisory/index.js'
 
 export type { CoordinatorToolsOptions, TaskLaunchedCallback } from './tools/coordinator/index.js'
 
+export type {
+	RegisterSharedRunPlanInput,
+	SharedRunWorkspaceConfig,
+} from './session/workspace/index.js'
+
 export type {
 	ConnectorManagerConfig,
 	EnvironmentConnectorManagerConfig,

package/src/registry/index.ts

@@ -4,6 +4,13 @@ export type { ManagedRegistryConfig } from './ManagedRegistry.js'
 
 export { ToolRegistry } from './tool/execute.js'
 export type { ToolExecutionResult } from './tool/execute.js'
+export {
+	ToolCatalog,
+	createToolCatalogFromRegistry,
+	loadingFromAvailability,
+	toolDefinitionToCatalogEntry,
+} from './toolset/catalog.js'
+export type { ToolCatalogFromRegistryOptions, ToolCatalogSearchOptions } from './toolset/catalog.js'
 
 export { ConnectorRegistry } from './connector/definitions.js'
 export { ScopedConnectorRegistry } from './connector/scoped.js'

package/src/registry/tool/execute.test.ts

@@ -20,8 +20,10 @@
  *     reports true iff at least one tool is suspended.
  *   - `getAvailability(name)` returns 'active' as a default even for
  *     unknown names (this is non-obvious but is the current behavior).
- *   - `searchDeferred(q)` is a case-insensitive filter against name OR
- *     description of every DEFERRED tool.
+ *   - `searchDeferred(q)` filters DEFERRED tools: a useful whole query
+ *     matches name OR description (case-insensitive); a batched multi-term
+ *     query matches meaningful tokens against the NAME only. Generic/short
+ *     tokens (`clawtool`, `tool`, …) are ignored so they can't over-activate.
  *   - `assignTiers(mapping)` mutates `tool.tier` on existing tools;
  *     throws via `getOrThrow` on unknown name; throws if the tier id
  *     is not in `tierConfig.tiers`.
@@ -84,6 +86,15 @@ describe('ToolRegistry — register + availability', () => {
 		expect(r.getAvailability('b')).toBe('deferred')
 	})
 
+	it('register overloads: array w/o state defaults active, (id, tool) form, bad id throws', () => {
+		const r = new ToolRegistry()
+		r.register([makeTool('arr')])
+		expect(r.getAvailability('arr')).toBe('active')
+		r.register('byid', makeTool('byid'))
+		expect(r.get('byid')).toBeDefined()
+		expect(() => r.register('oops', 'not-a-tool' as never)).toThrow(/requires a ToolDefinition/)
+	})
+
 	it('getAvailability returns active for unknown names (current default)', () => {
 		const r = new ToolRegistry()
 		expect(r.getAvailability('never-registered')).toBe('active')
@@ -180,6 +191,35 @@ describe('ToolRegistry — searchDeferred', () => {
 		expect(r.searchDeferred('does').map((t) => t.name)).toEqual(['alpha', 'beta'])
 		expect(r.searchDeferred('gamma')).toEqual([])
 	})
+
+	it('tokenizes a multi-term query so a batch of tool names each match', () => {
+		const r = new ToolRegistry()
+		r.register([makeTool('clawtool_A2aCard')], 'deferred')
+		r.register([makeTool('clawtool_PeerRegister')], 'deferred')
+		r.register([makeTool('clawtool_PeerList')], 'deferred')
+		r.register([makeTool('clawtool_Unrelated')], 'deferred')
+		// A whole-phrase substring match would find none of these.
+		expect(r.searchDeferred('A2aCard PeerRegister PeerList').map((t) => t.name)).toEqual([
+			'clawtool_A2aCard',
+			'clawtool_PeerRegister',
+			'clawtool_PeerList',
+		])
+		expect(r.searchDeferred('   ')).toEqual([])
+	})
+
+	it('does not over-activate on generic/shared tokens', () => {
+		const r = new ToolRegistry()
+		r.register([makeTool('clawtool_A2aCard', { description: 'peer card' })], 'deferred')
+		r.register([makeTool('clawtool_PeerList', { description: 'list peers' })], 'deferred')
+		r.register([makeTool('clawtool_WebSearch', { description: 'search the web' })], 'deferred')
+		// The shared "clawtool" prefix token must not drag in every tool.
+		expect(r.searchDeferred('clawtool WebSearch').map((t) => t.name)).toEqual([
+			'clawtool_WebSearch',
+		])
+		// A bare generic token identifies nothing — must not activate the catalog.
+		expect(r.searchDeferred('clawtool')).toEqual([])
+		expect(r.searchDeferred('tool')).toEqual([])
+	})
 })
 
 describe('ToolRegistry — tier mutation + guidance', () => {
@@ -227,12 +267,23 @@ describe('ToolRegistry — toPromptSection + toLLMTools', () => {
 		r.register(makeTool('a'))
 		r.register([makeTool('b')], 'deferred')
 		const s = r.toPromptSection()
+		expect(s).toContain('<tool_runtime_contract>')
+		expect(s).toContain('runtime tools parameter')
 		expect(s).toContain('<available_tools>')
 		expect(s).toContain('- a: a tool')
 		expect(s).toContain('<deferred_tools>')
+		expect(s).toContain('Deferred tools are discoverable')
 		expect(s).toContain('- b')
 	})
 
+	it('toPromptSection references search_tools only when it is active', () => {
+		const r = new ToolRegistry()
+		r.register(makeTool('search_tools'))
+		r.register([makeTool('b')], 'deferred')
+		const s = r.toPromptSection()
+		expect(s).toContain('Use search_tools to load these before use')
+	})
+
 	it('toLLMTools: converts active + suspended tools', () => {
 		const r = new ToolRegistry()
 		r.register(makeTool('a'))
@@ -303,7 +354,41 @@ describe('ToolRegistry — execute', () => {
 		)
 		const result = await r.execute('strict', { required: 123 }, makeContext())
 		expect(result.success).toBe(false)
-		expect(result.error).toMatch(/Invalid input/)
+		expect(result.error).toMatch(/Validation failed for "strict"/)
+		expect(result.error).toContain('Expected string, received number')
+	})
+
+	it('empty-args validation lists required params with descriptions', async () => {
+		const r = new ToolRegistry()
+		r.register(
+			makeTool('needs', {
+				inputSchema: z.object({ q: z.string().describe('the query'), n: z.number() }),
+			}),
+		)
+		const result = await r.execute('needs', {}, makeContext())
+		expect(result.success).toBe(false)
+		expect(result.error).toMatch(/called with no arguments/)
+		expect(result.error).toContain('q: string — the query')
+		expect(result.error).toContain('n: number')
+	})
+
+	it('validation hint reports when there are no required params', async () => {
+		const r = new ToolRegistry()
+		r.register(makeTool('opt', { inputSchema: z.object({ k: z.string().optional() }) }))
+		const result = await r.execute('opt', { k: 123 }, makeContext())
+		expect(result.success).toBe(false)
+		expect(result.error).toContain('No required parameters known.')
+	})
+
+	it('validation hint tolerates a schema it cannot introspect', async () => {
+		const r = new ToolRegistry()
+		const bogusSchema = {
+			safeParse: () => ({ success: false, error: { issues: [{ path: [], message: 'nope' }] } }),
+		}
+		r.register(makeTool('weird', { inputSchema: bogusSchema as never }))
+		const result = await r.execute('weird', { a: 1 }, makeContext())
+		expect(result.success).toBe(false)
+		expect(result.error).toContain('Could not introspect required parameters.')
 	})
 
 	it('wraps thrown errors in the execute function', async () => {
@@ -320,6 +405,50 @@ describe('ToolRegistry — execute', () => {
 		expect(result.error).toMatch(/execution failed: boom/)
 	})
 
+	it('wraps a non-Error throw', async () => {
+		const r = new ToolRegistry()
+		r.register(
+			makeTool('throws-string', {
+				async execute() {
+					throw 'plain string failure'
+				},
+			}),
+		)
+		const result = await r.execute('throws-string', {}, makeContext())
+		expect(result.success).toBe(false)
+		expect(result.error).toMatch(/execution failed/)
+	})
+
+	it('passes through a tool result that is unsuccessful with an error', async () => {
+		const r = new ToolRegistry()
+		r.register(
+			makeTool('soft-fail', {
+				async execute() {
+					return { success: false, output: '', error: 'soft failure' }
+				},
+			}),
+		)
+		const result = await r.execute('soft-fail', {}, makeContext())
+		expect(result.success).toBe(false)
+		expect(result.error).toBe('soft failure')
+	})
+
+	it('blocks a non-read-only tool in plan mode (no isReadOnly hint)', async () => {
+		const r = new ToolRegistry()
+		const execute = vi.fn(async () => ({ success: true, output: 'ok' }))
+		r.register(makeTool('mutate', { execute }))
+		const result = await r.execute(
+			'mutate',
+			{},
+			makeContext({
+				permissionContext: { mode: 'plan', runId: 'run_1', workingDirectory: '/tmp' },
+			}),
+		)
+		expect(result.success).toBe(false)
+		expect(result.error).toMatch(/plan mode/)
+		expect(execute).not.toHaveBeenCalled()
+	})
+
 	it('returns the tool result on happy path', async () => {
 		const r = new ToolRegistry()
 		r.register(makeTool('good'))

package/src/registry/tool/execute.ts

@@ -16,6 +16,11 @@ import { ManagedRegistry } from '../ManagedRegistry.js'
 
 export type { ToolExecutionResult }
 
+// Tokens too generic to identify a tool by name — ignored when matching a
+// batched `search_tools` query so they can't activate the whole catalog
+// (every bridged tool name shares the `clawtool` prefix, for instance).
+const SEARCH_STOP_TOKENS = new Set(['clawtool', 'tool', 'tools', 'mcp', 'the', 'and', 'for', 'use'])
+
 export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 	private availability: Map<string, ToolAvailability> = new Map()
 	private tierConfig?: ToolTierConfig
@@ -114,10 +119,27 @@ export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 	}
 
 	searchDeferred(query: string): ToolDefinition[] {
-		const q = query.toLowerCase()
-		return this.getByAvailability(['deferred']).filter(
-			(t) => t.name.toLowerCase().includes(q) || t.description.toLowerCase().includes(q),
-		)
+		const q = query.toLowerCase().trim()
+		if (q.length === 0) return []
+		// Per-token matching exists only so a batched query naming several tools
+		// at once ("A2aCard PeerRegister PeerList") activates each. Restrict it
+		// to the tool NAME and drop short/generic tokens — matching tokens
+		// against descriptions (or letting a shared word like "clawtool"/"list"
+		// through) would activate the whole catalog and defeat deferral.
+		const tokens = q.split(/\s+/).filter((tok) => tok.length >= 3 && !SEARCH_STOP_TOKENS.has(tok))
+		// A bare generic token ("clawtool") identifies nothing specific — skip the
+		// broad whole-query match for it so it can't activate the whole catalog.
+		const wholeQueryUseful = q.length >= 3 && !SEARCH_STOP_TOKENS.has(q)
+		return this.getByAvailability(['deferred']).filter((t) => {
+			const name = t.name.toLowerCase()
+			// Whole-query match (single-term capability search) against name or
+			// description — the deliberate, narrow behaviour.
+			if (wholeQueryUseful && (name.includes(q) || t.description.toLowerCase().includes(q))) {
+				return true
+			}
+			// Batched multi-name query: any meaningful token, name only.
+			return tokens.some((tok) => name.includes(tok))
+		})
 	}
 
 	assignTiers(mapping: Record<string, string>): void {
@@ -149,6 +171,9 @@ export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 		const deferred = this.getByAvailability(['deferred'], toolNames)
 
 		const parts: string[] = []
+		const contractNote = `<tool_runtime_contract>
+Executable tool names, descriptions, and JSON input schemas are attached through the runtime tools parameter. Treat that runtime schema as authoritative; this prompt section is a discoverability summary only.
+</tool_runtime_contract>`
 
 		if (active.length > 0) {
 			const entries = active.map((t) => `- ${t.name}: ${t.description}`).join('\n')
@@ -157,13 +182,15 @@ export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 
 		if (deferred.length > 0) {
 			const entries = deferred.map((t) => `- ${t.name}`).join('\n')
-			parts.push(
-				`<deferred_tools>\nUse search_tools to load these before use:\n${entries}\n</deferred_tools>`,
-			)
+			const deferredIntro =
+				this.has('search_tools') && this.getAvailability('search_tools') === 'active'
+					? 'Use search_tools to load these before use:'
+					: 'Deferred tools are discoverable but not executable until the runtime activates them:'
+			parts.push(`<deferred_tools>\n${deferredIntro}\n${entries}\n</deferred_tools>`)
 		}
 
 		if (parts.length === 0) return ''
-		return parts.join('\n\n')
+		return [contractNote, ...parts].join('\n\n')
 	}
 
 	toLLMTools(toolNames?: string[]): LLMToolSchema[] {
@@ -254,8 +281,33 @@ export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 					.map((i) => `${i.path.join('.')}: ${i.message}`)
 					.join('; ')
 
+				// Distinguish "model sent an empty/no-arg call" from
+				// "model sent partial args" — the first is most often a
+				// streaming hiccup or a definition-test ping (Anthropic
+				// occasionally pings tool surfaces with `{}` while the
+				// schema is still loading), the second is a genuine
+				// programming mistake by the model. The model self-
+				// corrects MUCH more reliably when the error tells it
+				// (a) which fields are required, (b) their types, and
+				// (c) a minimal example call. Without these hints the
+				// downstream UI just shows a red "Failed" row and the
+				// model rarely retries with the right args.
+				const isEmptyInput =
+					rawInput === null ||
+					rawInput === undefined ||
+					(typeof rawInput === 'object' &&
+						!Array.isArray(rawInput) &&
+						Object.keys(rawInput as Record<string, unknown>).length === 0)
+
+				const requiredHint = describeRequiredInput(tool.inputSchema)
+
+				const enrichedMessage = isEmptyInput
+					? `Tool "${toolName}" was called with no arguments. ${requiredHint} Retry the call with the required parameters populated.`
+					: `Validation failed for "${toolName}": ${errorMessage}. ${requiredHint}`
+
 				this.log.error(`Tool input validation failed: ${toolName}`, {
 					errors: errorMessage,
+					empty: isEmptyInput,
 				})
 
 				span.setAttributes({
@@ -268,7 +320,7 @@ export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 				return {
 					success: false,
 					output: '',
-					error: `Invalid input for tool "${toolName}": ${errorMessage}`,
+					error: enrichedMessage,
 				}
 			}
 
@@ -319,3 +371,36 @@ export class ToolRegistry extends ManagedRegistry<ToolDefinition> {
 		return candidates.filter((t) => states.includes(this.getAvailability(t.name)))
 	}
 }
+
+/**
+ * Build a one-sentence "Required: <field>: <type>, <field>: <type>"
+ * hint from a Zod schema, used to enrich tool-input validation
+ * errors so the model can self-correct without round-tripping the
+ * full JSON schema again. Walks the schema's JSON-Schema rendering
+ * (already a dependency for tool registration) so we don't have to
+ * branch over Zod's internal type tree.
+ *
+ * Returns a fallback string for opaque/non-object schemas — the
+ * caller still ships the raw Zod issues separately, so the hint
+ * here is bonus context, not the only signal.
+ */
+function describeRequiredInput(schema: { _def?: unknown }): string {
+	try {
+		const json = zodToJsonSchema(schema as never) as {
+			properties?: Record<string, { type?: string; description?: string }>
+			required?: string[]
+		}
+		const required = json.required ?? []
+		if (required.length === 0) return 'No required parameters known.'
+		const props = json.properties ?? {}
+		const lines = required.map((name) => {
+			const def = props[name] ?? {}
+			const type = def.type ?? 'value'
+			const desc = def.description ? ` — ${def.description}` : ''
+			return `${name}: ${type}${desc}`
+		})
+		return `Required: ${lines.join(', ')}.`
+	} catch {
+		return 'Could not introspect required parameters.'
+	}
+}