@namzu/sdk 39.0.0 → 41.0.0

package/src/public-runtime.ts

@@ -274,9 +274,17 @@ export {
 // parses, which is the wire's property. Rendered once, converted at the driver.
 export { findDraft07Only, toSchemaDialect } from './registry/tool/dialect.js'
 export type { JsonSchemaDialect } from './registry/tool/dialect.js'
+// Better than converting per wire: emit the INTERSECTION of the dialects, so
+// there is nothing left for a driver to convert and an unmeasured wire is safe
+// by construction. Exported so a driver or a CI gate can assert it.
+export {
+	findPortableSchemaViolations,
+	toPortableToolSchema,
+} from './registry/tool/portable.js'
+export type { PortableSchemaViolation } from './registry/tool/portable.js'
 // The renderer itself, so a driver or a contract test can ask what a tool will
 // actually put on the wire without reaching into the registry.
-export { renderToolSchema } from './registry/tool/schema.js'
+export { renderToolSchema, toolWireSchema } from './registry/tool/schema.js'
 export type { StrictSchemaViolation } from './provider/strict-schema.js'
 export type { ModelIdGrammar, ModelVersion } from './provider/model-version.js'
 export { drainQuery, query } from './runtime/query/index.js'
@@ -717,12 +725,29 @@ export {
 	hasDrift,
 	HttpSseTransport,
 	HybridExecutionContext,
+	buildEnvelope,
+	classifyModernHttpFailure,
+	createMcpEraCache,
+	decodeResult,
+	defaultMcpEraCache,
+	encodeMcpHeaderValue,
+	isHeaderMismatchError,
+	isMissingRequiredClientCapabilityError,
+	isRecognizedModernError,
+	isResourceNotFoundError,
+	isUnsupportedProtocolVersionError,
 	LocalExecutionContext,
+	mcpEraCacheKey,
+	MCPHttpStatusError,
+	MCPInputRequiredError,
+	MCPInvalidResultTypeError,
 	MCPClient,
 	MCPMethodNotFound,
 	MCPConnectorBridge,
+	MCPProtocolError,
 	MCPServer,
 	MCPToolDiscovery,
+	resolveMcpEra,
 	mcpPromptToToolDefinition,
 	renderPromptMessages,
 	mcpJsonSchemaToZod,
@@ -741,10 +766,20 @@ export {
 	toolDefinitionToMCPTool,
 	toolResultToMCPToolResult,
 	toolsHash,
+	validateMcpHeaderAnnotations,
 	WebhookConnector,
 	zodToMCPJsonSchema,
 } from './connector/index.js'
 export type {
+	McpEnvelope,
+	McpEnvelopeInput,
+	McpEraProbe,
+	McpEraProbeAnswer,
+	McpEraResolution,
+	McpEraResolutionInput,
+	McpHeaderAnnotationVerdict,
+	McpParamHeaderBinding,
+	MCPDecodedResult,
 	MCPToolDiscoveryOptions,
 	MCPToolDrift,
 	MCPToolPolicy,
@@ -1221,6 +1256,7 @@ export {
 export {
 	CODING_AGENT_DELEGATION_DOCTRINE,
 	CODING_AGENT_DOCTRINE_CONTRIBUTION_ID,
+	CODING_AGENT_ORCHESTRATE_DOCTRINE,
 	CODING_AGENT_WORKING_DOCTRINE,
 	PLAN_MODE_DOCTRINE,
 	PromptContributionCollisionError,

package/src/public-tools.ts

@@ -43,6 +43,11 @@ export { GrepTool } from './tools/builtins/grep.js'
 // capability: an id with nothing that reads it is the unbacked suggestion
 // that was removed from bash's schema.
 export { JobTool } from './tools/builtins/job.js'
+// Blocks until a job `JobTool` started ends, instead of reading it in a
+// loop. Ships in the default builtin set alongside `job` and `bash`, for
+// the same reason `job` does: a job with nothing that can block on it is
+// the same unbacked suggestion `job` itself exists to fix.
+export { WaitForJobTool } from './tools/builtins/wait-for-job.js'
 // Loads a skill's instructions, and adopts the tool scope it declares.
 // NOT in the default builtin set: a run with no skills has nothing for it
 // to do, and offering a tool that can only refuse is worse than not
@@ -128,5 +133,18 @@ export {
 } from './connector/tools/index.js'
 
 export { createFileReadTracker } from './tools/file-read-tracker.js'
+export type { LedgerReplayReport } from './runtime/query/file-evidence-replay.js'
+/**
+ * Rebuild a tracker from a conversation's own history, for a host that keeps
+ * one per conversation across turns and has just restored one from a store.
+ * Reads no file's content — only the paths the history names, canonicalized
+ * the way the built-in tools canonicalize them so that the entries land where
+ * those tools will look for them. See the SDK's tool-execution documentation
+ * for what a replayed observation does and does not establish.
+ */
+export {
+	type ObservationSeedContext,
+	seedObservationLedger,
+} from './runtime/query/file-evidence-seed.js'
 
 export { buildResidentToolEvidenceTools } from './tools/resident-tool-evidence.js'

package/src/registry/tool/execute.ts

@@ -20,7 +20,7 @@ import type {
 import { toErrorMessage } from '../../utils/error.js'
 import { cloneJsonValue as clonePreparedInput } from '../../utils/json-snapshot.js'
 import { ManagedRegistry } from '../ManagedRegistry.js'
-import { renderToolSchema } from './schema.js'
+import { renderToolSchema, toolWireSchema } from './schema.js'
 import { ToolResultHalted, screenToolResult } from './screen.js'
 
 export type { ToolExecutionResult }
@@ -461,9 +461,7 @@ Executable tool names, descriptions, and JSON input schemas are attached through
 				function: {
 					name: tool.name,
 					description: describeWithOutput(description, tool.outputSchema),
-					parameters:
-						(tool.modelInputSchema ? structuredClone(tool.modelInputSchema) : undefined) ??
-						renderToolSchema(tool.inputSchema),
+					parameters: toolWireSchema(tool),
 				},
 			}
 		})

package/src/registry/tool/portable.ts

@@ -0,0 +1,264 @@
+/**
+ * The schema shape every wire namzu speaks reads the same way.
+ *
+ * `dialect.ts` solves the problem one wire at a time: render draft-07, and let
+ * each driver say it again in the dialect it knows its endpoint parses. That
+ * works only where a driver HAS a measurement. Of the ten driver packages,
+ * three convert; the other seven forward the rendering verbatim, because
+ * nobody had measured their wires.
+ *
+ * One of those seven then got measured, the hard way. Its gateway validates a
+ * tool's `parameters` against the JSON Schema 2020-12 metaschema, and `read`'s
+ * `readRange` — the only `z.tuple(...)` in the first-party tool surface —
+ * renders draft-07 as `items: [a, b]`:
+ *
+ *     [400] Tool 4 function has invalid 'parameters' schema:
+ *     [{'minimum': 1, 'type': 'integer'}, {'minimum': 1, 'type': 'integer'}]
+ *     is not of type 'object', 'boolean'
+ *
+ * The lesson is not "convert in five more drivers". Wiring each driver to a
+ * dialect requires a measurement per wire, and the two hardest to measure are
+ * precisely the ones nobody can measure on a user's behalf: the driver whose
+ * schema field the vendor documents nowhere, and every compatible endpoint a
+ * user points the generic HTTP driver at.
+ *
+ * So the kernel emits the INTERSECTION instead: the subset of JSON Schema that
+ * is valid, and means the same thing, in draft-07 and in 2020-12 alike. A
+ * schema in that subset needs no conversion anywhere, which is why this runs at
+ * the rendering boundary rather than at ten driver boundaries.
+ *
+ * `toSchemaDialect` is not replaced by this and does not become dead: a driver
+ * can still be handed a `parameters` object namzu never rendered (a host
+ * passing `ChatCompletionParams` straight in), and converting at the boundary
+ * remains the right answer there.
+ */
+
+export interface PortableSchemaViolation {
+	/** Dotted path to the offending keyword, e.g. `properties.readRange.items`. */
+	readonly path: string
+	readonly keyword: string
+	/** What to write instead. */
+	readonly remedy: string
+}
+
+const NO_TUPLES =
+	'say the array uniformly — one `items` schema plus `minItems`/`maxItems`; a tuple is spelled differently in draft-07 and 2020-12 and is invalid in one of them either way'
+
+/**
+ * Keywords outside the intersection, and why each one is out.
+ *
+ * A deny-list, like `findStrictSchemaViolations` and for the same reason: an
+ * allow-list would have to enumerate every annotation a schema may carry and
+ * would refuse a schema for saying something harmless.
+ *
+ * Every entry names a wire in the driver matrix that reads it wrong, not a
+ * style preference.
+ */
+const OUTSIDE_THE_INTERSECTION: ReadonlyMap<string, string> = new Map([
+	// 2020-12 moved the positional schemas to `prefixItems` and kept `items`
+	// for the tail, so an array-valued `items` does not merely look old — it
+	// means something else, and a 2020-12 validator rejects the whole request.
+	// Measured on Zen's Console gateway; the quote is in this file's header.
+	['additionalItems', NO_TUPLES],
+	// The other spelling of the same tuple. A draft-07 validator does not know
+	// the keyword, so it ignores it — which turns a constrained array into an
+	// unconstrained one SILENTLY, the worse failure of the two. The strict
+	// tool-input subset refuses it outright as well (`strict-schema.ts`), and
+	// the function-declaration schema on one of these wires is an OpenAPI 3.0
+	// subset that has never documented it.
+	['prefixItems', NO_TUPLES],
+	// A rendered tool schema has to stand alone. The renderer already runs with
+	// `$refStrategy: 'none'`, and the MCP bridge inlines refs before handing a
+	// schema out, because a wire that does not resolve them sees an empty
+	// constraint or refuses the document.
+	['$ref', 'inline the referenced schema; a tool schema has to stand alone'],
+	['$defs', 'inline the definitions; a tool schema has to stand alone'],
+	['definitions', 'inline the definitions; a tool schema has to stand alone'],
+	['$id', 'drop it; a tool schema is not a retrievable document'],
+	['$anchor', 'drop it; there are no references to anchor'],
+	['$dynamicRef', 'inline the referenced schema; a tool schema has to stand alone'],
+	['$dynamicAnchor', 'drop it; there are no references to anchor'],
+	// Nothing on any wire reads it, and asserting a dialect is the one thing a
+	// schema that must work in both must not do. It also rides in the tools
+	// block, which renders at position 0 inside the prompt-cache prefix.
+	['$schema', 'drop it; no provider reads it and it asserts a dialect'],
+	// 2019-09 vocabulary. Same silent-widening hazard as `prefixItems` on a
+	// draft-07 validator.
+	['unevaluatedItems', 'state the constraint with `items` and `maxItems`'],
+	['unevaluatedProperties', 'state the constraint with `additionalProperties`'],
+])
+
+/**
+ * Every place a schema leaves the intersection, with its exact path.
+ *
+ * Exported so a test can sweep every shipped tool and a driver can assert
+ * rather than hope. The path is the point: the vendor's error names the tool
+ * index and the offending fragment but never where inside the schema it sits.
+ */
+export function findPortableSchemaViolations(
+	schema: unknown,
+	path = '',
+): PortableSchemaViolation[] {
+	if (Array.isArray(schema)) {
+		return schema.flatMap((item, index) => findPortableSchemaViolations(item, `${path}[${index}]`))
+	}
+	if (typeof schema !== 'object' || schema === null) return []
+
+	const found: PortableSchemaViolation[] = []
+	for (const [keyword, value] of Object.entries(schema as Record<string, unknown>)) {
+		const here = path ? `${path}.${keyword}` : keyword
+		const remedy = OUTSIDE_THE_INTERSECTION.get(keyword)
+		if (remedy !== undefined) {
+			found.push({ path: here, keyword, remedy })
+			continue
+		}
+		// The draft-07 tuple. Reported once, at the tuple, rather than once per
+		// positional schema underneath it — one fixable finding, not a list.
+		if (keyword === 'items' && Array.isArray(value)) {
+			found.push({ path: here, keyword, remedy: NO_TUPLES })
+			continue
+		}
+		// A union of types is valid JSON Schema in both dialects, but the
+		// OpenAPI-3.0-shaped wires namzu speaks take a single type name and
+		// either refuse a list or ignore it.
+		if (keyword === 'type' && Array.isArray(value)) {
+			found.push({
+				path: here,
+				keyword,
+				remedy: 'use `anyOf` of single-typed schemas; several wires take one type name only',
+			})
+			continue
+		}
+		found.push(...findPortableSchemaViolations(value, here))
+	}
+	return found
+}
+
+/**
+ * Rewrite what can be rewritten, so a tuple cannot reach a wire.
+ *
+ * Only the tuple spellings and `$schema` are rewritten. The rest of the
+ * deny-list is reported and not repaired on purpose: inlining a `$ref` or
+ * choosing which member of a `type` list the author meant is a guess about
+ * intent, and a guess that silently changes what a tool accepts is worse than
+ * the 400 it avoids. Those are caught by the profile test at the source.
+ *
+ * Returns the input unchanged — the SAME reference — when nothing needs
+ * rewriting, which is the overwhelming common case. The tools block sits at
+ * position 0 of the prompt-cache prefix, so a fresh, equal object per request
+ * would invalidate the cache for the whole run.
+ */
+export function toPortableToolSchema(json: Record<string, unknown>): Record<string, unknown> {
+	return makePortable(json) as Record<string, unknown>
+}
+
+function makePortable(value: unknown): unknown {
+	if (Array.isArray(value)) {
+		let changed = false
+		const out = value.map((item) => {
+			const next = makePortable(item)
+			if (next !== item) changed = true
+			return next
+		})
+		return changed ? out : value
+	}
+	if (typeof value !== 'object' || value === null) return value
+
+	const node = value as Record<string, unknown>
+	const positional = Array.isArray(node.prefixItems)
+		? node.prefixItems
+		: Array.isArray(node.items)
+			? node.items
+			: undefined
+	if (positional) return collapseTuple(node, positional)
+
+	let changed = '$schema' in node
+	const out: Record<string, unknown> = {}
+	for (const [key, child] of Object.entries(node)) {
+		if (key === '$schema') continue
+		const next = makePortable(child)
+		if (next !== child) changed = true
+		out[key] = next
+	}
+	return changed ? out : node
+}
+
+/**
+ * Say a positional array uniformly.
+ *
+ * What is kept: the arity, and every constraint that applies to every member.
+ * What is lost: which member sits where. That loss is real and it is the price
+ * of a shape both dialects read identically — and it costs nothing in practice,
+ * because the tool's Zod schema still parses the call and still refuses a
+ * wrongly-ordered array with a message the model can act on. The wire schema is
+ * a hint to the model; the parser is the contract.
+ */
+function collapseTuple(
+	node: Record<string, unknown>,
+	positional: readonly unknown[],
+): Record<string, unknown> {
+	// The tail rule, in whichever dialect this node happens to be written in:
+	// draft-07 spells it `additionalItems`, 2020-12 spells it `items` and only
+	// calls it a tail when `prefixItems` holds the members.
+	const tail = Array.isArray(node.prefixItems) ? node.items : node.additionalItems
+	const arity = positional.length
+	const members = positional.map(makePortable)
+
+	// A tail that is itself a schema constrains every element past the members,
+	// so those elements join the uniform reading rather than closing it.
+	const tailSchema =
+		tail !== undefined && tail !== true && tail !== false ? makePortable(tail) : undefined
+	// `false` closes the array at the members' length; so does a `maxItems` the
+	// author already pinned there.
+	const closed =
+		tailSchema === undefined &&
+		(tail === false || (typeof node.maxItems === 'number' && node.maxItems <= arity))
+
+	// An open tail with no schema says "anything past the members", which no
+	// single `items` can express alongside the members themselves. Saying
+	// nothing about the elements is the only honest uniform reading; the bounds
+	// below still carry whatever the author pinned.
+	const candidates =
+		tailSchema !== undefined ? dedupe([...members, tailSchema]) : closed ? dedupe(members) : []
+
+	const element =
+		candidates.length === 0
+			? undefined
+			: candidates.length === 1
+				? candidates[0]
+				: { anyOf: candidates }
+
+	const out: Record<string, unknown> = {}
+	for (const [key, child] of Object.entries(node)) {
+		if (key === '$schema') continue
+		if (key === 'additionalItems') continue
+		if (key === 'prefixItems' || (key === 'items' && Array.isArray(child))) {
+			if (element !== undefined) out.items = element
+			continue
+		}
+		// The 2020-12 tail, already folded into `items` above.
+		if (key === 'items' && Array.isArray(node.prefixItems)) continue
+		if (key === 'maxItems' && closed && typeof child === 'number') {
+			out.maxItems = Math.min(child, arity)
+			continue
+		}
+		out[key] = makePortable(child)
+	}
+	// A tuple closed by `additionalItems: false` rather than by a bound loses
+	// its ceiling otherwise: a uniform `items` says nothing about length, so a
+	// schema written to forbid a third element would begin to allow any number.
+	if (closed && out.maxItems === undefined) out.maxItems = arity
+	return out
+}
+
+function dedupe(schemas: readonly unknown[]): unknown[] {
+	const seen = new Set<string>()
+	const out: unknown[] = []
+	for (const schema of schemas) {
+		const key = JSON.stringify(schema) ?? 'undefined'
+		if (seen.has(key)) continue
+		seen.add(key)
+		out.push(schema)
+	}
+	return out
+}

package/src/registry/tool/schema.ts

@@ -1,5 +1,6 @@
 import type { z } from 'zod'
 import { zodToJsonSchema } from 'zod-to-json-schema'
+import { toPortableToolSchema } from './portable.js'
 
 /**
  * The single place a tool's Zod schema becomes the JSON Schema that goes on
@@ -50,17 +51,46 @@ export function renderToolSchema(schema: z.ZodType): Record<string, unknown> {
 }
 
 /**
- * Strip what no provider reads.
+ * Make the rendering something every wire reads the same way.
  *
- * Only `$schema` today, and only at the root — `zodToJsonSchema` does not
- * stamp it on nested nodes. Kept as a named function rather than an inline
- * `delete` so the next "providers ignore this" field has an obvious home
- * and a place to be justified.
+ * Two jobs, both at this one boundary:
+ *
+ * 1. **Strip what no provider reads.** `$schema` rides in the tools block,
+ *    which renders at position 0 inside the prompt-cache prefix, once per
+ *    tool, per request, forever — and it asserts a dialect, which is the one
+ *    thing a schema that has to work on every wire must not do.
+ *
+ * 2. **Leave the intersection of draft-07 and 2020-12 intact.** `renderToolSchema`
+ *    emits draft-07, where a tuple is `items: [a, b]`; a 2020-12 wire reads
+ *    that as invalid and rejects the whole request, taking every other tool in
+ *    the call down with it. Five of the ten drivers forward the rendering
+ *    verbatim, so the only place that can be fixed once is here. See
+ *    `portable.ts` for the measurement that put it here.
  */
 export function normalizeToolSchema(json: Record<string, unknown>): Record<string, unknown> {
-	if (!('$schema' in json)) return json
-	const { $schema: _dropped, ...rest } = json
-	return rest
+	return toPortableToolSchema(json)
+}
+
+/**
+ * The schema a tool actually puts on the wire.
+ *
+ * A tool may carry an explicit `modelInputSchema` — hand-written JSON Schema,
+ * not rendered from Zod — and that schema reaches providers through the same
+ * `tools` block as a rendered one. It therefore has to clear the same bar, and
+ * the two call sites that build that block (`ToolRegistry.toLLMTools` and the
+ * toolset catalog) had each spelled the fallback out on their own. Two paths
+ * that agree today are two paths that can disagree tomorrow.
+ *
+ * The clone is deliberate: a caller may edit the parameters it is handed —
+ * `hostModelSchema` does — and the definition's own object must not move
+ * underneath the next call.
+ */
+export function toolWireSchema(tool: {
+	readonly modelInputSchema?: Record<string, unknown>
+	readonly inputSchema: z.ZodType
+}): Record<string, unknown> {
+	if (tool.modelInputSchema) return toPortableToolSchema(structuredClone(tool.modelInputSchema))
+	return renderToolSchema(tool.inputSchema)
 }
 
 /**

package/src/registry/toolset/catalog.ts

@@ -9,7 +9,7 @@ import type {
 	ToolsetDefinition,
 	ToolsetPolicy,
 } from '../../types/toolset/index.js'
-import { renderToolSchema } from '../tool/schema.js'
+import { toolWireSchema } from '../tool/schema.js'
 
 export interface ToolCatalogSearchOptions {
 	readonly loading?: readonly ToolLoadingMode[]
@@ -245,17 +245,16 @@ function toolDefinitionToLLMTool(definition: ToolDefinition | undefined): LLMToo
 		function: {
 			name: definition.name,
 			description: definition.description,
-			// Through `renderToolSchema`, not a second inline conversion. The
+			// Through `toolWireSchema`, not a second inline conversion. The
 			// options were already identical, so this was not a different
 			// rendering — it was the same rendering without the guarantees:
 			// no `$schema` stripping (that key rides in the tools block, which
-			// sits at position 0 inside the prompt-cache prefix), no memoization,
-			// and no freeze. Two paths that agree today are two paths that can
-			// disagree tomorrow, and a tool reaching the wire through the catalog
-			// rather than the registry is not a different tool.
-			parameters:
-				(definition.modelInputSchema ? structuredClone(definition.modelInputSchema) : undefined) ??
-				renderToolSchema(definition.inputSchema),
+			// sits at position 0 inside the prompt-cache prefix), no portability
+			// normalisation, no memoization, and no freeze. Two paths that agree
+			// today are two paths that can disagree tomorrow, and a tool reaching
+			// the wire through the catalog rather than the registry is not a
+			// different tool.
+			parameters: toolWireSchema(definition),
 		},
 	}
 }