fix(sdk): handle cyclic run-state on session resume

sdk/src/__tests__/run-state-git-changes.test.ts

@@ -1,6 +1,7 @@
 import { EventEmitter } from 'events'
 
 import { describe, expect, it } from 'bun:test'
+import z from 'zod/v4'
 
 import {
   applyOverridesToSessionState,
@@ -194,6 +195,78 @@ describe('repository snapshot persistence', () => {
       'AGENTS.md': 'Updated instructions',
     })
   })
+
+  it('clones a resumed state whose run state carries a cycle', async () => {
+    // Resuming used to die on the second interaction with "JSON.stringify
+    // cannot serialize cyclic structures". The state carries the previous
+    // turn's tool blocks, and a recursive zod `lazy` schema in one of them is
+    // self-referential, so the clone has to survive a cycle.
+    const baseSessionState = getInitialSessionState(getStubProjectFileContext())
+    const circular: Record<string, unknown> = {}
+    circular.self = circular
+    ;(baseSessionState.mainAgentState as any).output = circular
+
+    const continuedSessionState = await applyOverridesToSessionState(
+      '/repo',
+      baseSessionState,
+      { maxAgentSteps: 7 },
+    )
+
+    // The overrides still apply...
+    expect(continuedSessionState.mainAgentState.stepsRemaining).toBe(7)
+    // ...and the result is an independent copy whose cycle came across intact.
+    expect(continuedSessionState.mainAgentState).not.toBe(
+      baseSessionState.mainAgentState,
+    )
+    const clonedOutput = (continuedSessionState.mainAgentState as any).output
+    expect(clonedOutput.self).toBe(clonedOutput)
+  })
+
+  it('keeps a live zod schema intact when the fallback clone runs', async () => {
+    // The fallback runs exactly when the state is cyclic, and what makes it
+    // cyclic is a zod schema: mcp.ts stores MCP tools as live zod schemas, and
+    // a zod schema is self-referential. lodash copies own *enumerable*
+    // properties only, while zod keeps its internals on a non-enumerable
+    // `_zod`, so a plain cloneDeep hands back something that still looks like a
+    // schema -- `safeParse` comes from the prototype -- but throws on the first
+    // zod call with "undefined is not an object (evaluating '_zod.parent')".
+    const schema = z.object({ path: z.string() })
+    const baseSessionState = getInitialSessionState(getStubProjectFileContext())
+    const circular: Record<string, unknown> = { schema }
+    circular.self = circular
+    ;(baseSessionState.mainAgentState as any).output = circular
+
+    const continuedSessionState = await applyOverridesToSessionState(
+      '/repo',
+      baseSessionState,
+      {},
+    )
+
+    const clonedSchema = (continuedSessionState.mainAgentState as any).output
+      .schema
+    expect(clonedSchema._zod).toBeDefined()
+    expect(z.toJSONSchema(clonedSchema, { io: 'input' })).toMatchObject({
+      type: 'object',
+    })
+  })
+
+  it('falls back to a deep copy for values JSON.stringify rejects (BigInt)', async () => {
+    // The fallback is broader than "cycles only": BigInt makes JSON.stringify
+    // throw too, and cloneDeep preserves it rather than dropping the key.
+    const baseSessionState = getInitialSessionState(getStubProjectFileContext())
+    ;(baseSessionState.mainAgentState as any).output = { token: 10n }
+
+    const continuedSessionState = await applyOverridesToSessionState(
+      '/repo',
+      baseSessionState,
+      {},
+    )
+
+    expect((continuedSessionState.mainAgentState as any).output.token).toBe(10n)
+    expect(continuedSessionState.mainAgentState).not.toBe(
+      baseSessionState.mainAgentState,
+    )
+  })
 })
 
 describe('isTestFilePath', () => {

sdk/src/run-state.ts

@@ -13,7 +13,7 @@ import {
 } from '@codebuff/common/project-file-tree'
 import { getInitialSessionState } from '@codebuff/common/types/session-state'
 import { getErrorObject } from '@codebuff/common/util/error'
-import { cloneDeep } from 'lodash'
+import { cloneDeep, cloneDeepWith } from 'lodash'
 import z from 'zod/v4'
 
 import { loadLocalAgents } from './agents/load-agents'
@@ -1097,6 +1097,33 @@ export function withMessageHistory({
   return newRunState
 }
 
+/**
+ * True for a zod schema, which is the only kind of value a clone has to be
+ * careful with here.
+ */
+const isZodSchema = (value: unknown): boolean => {
+  if (typeof value !== 'object' || value === null) return false
+  const candidate = value as { safeParse?: unknown; _zod?: unknown }
+  return (
+    typeof candidate.safeParse === 'function' && candidate._zod !== undefined
+  )
+}
+
+/**
+ * Deep-clones a value without destroying the zod schemas inside it.
+ *
+ * zod installs its internals on a non-enumerable `_zod` property, and lodash
+ * copies own *enumerable* properties only. A plain `cloneDeep` therefore
+ * returns something that still looks like a schema -- `safeParse` lives on the
+ * prototype, so it survives -- but has lost `_zod`, and the next zod call on it
+ * throws "undefined is not an object (evaluating 'schema._zod.parent')".
+ * Schemas are immutable, so the clone keeps them as they are.
+ */
+export const cloneDeepPreservingZodSchemas = <T>(value: T): T =>
+  cloneDeepWith(value, (candidate) =>
+    isZodSchema(candidate) ? candidate : undefined,
+  )
+
 /**
  * Applies overrides to an existing session state, allowing specific fields to be updated
  * even when continuing from a previous run.
@@ -1115,10 +1142,33 @@ export async function applyOverridesToSessionState(
     maxAgentSteps?: number
   },
 ): Promise<SessionState> {
-  // Deep clone to avoid mutating the original session state
-  const sessionState = JSON.parse(
-    JSON.stringify(baseSessionState),
-  ) as SessionState
+  // Deep clone to avoid mutating the original session state. The JSON
+  // round-trip is the fast path - it matches what a persisted snapshot looks
+  // like - but this state can hold values JSON.stringify rejects. A resumed
+  // session carries the tool blocks of the turn before it, and a recursive zod
+  // `lazy` schema is cyclic: stringifying it throws "cannot serialize cyclic
+  // structures" and the resume dies. lodash cloneDeep handles cycles, so fall
+  // back to it and keep the clone working instead of losing the session.
+  //
+  // The catch is deliberately not narrowed, but it is narrower than "any JSON
+  // failure": in practice the throw comes from a cycle or a BigInt. Values JSON
+  // merely drops or coerces - `undefined`, functions, symbols, Date, URL - do
+  // not throw, so they keep taking the JSON path exactly as they did before and
+  // never reach cloneDeep. The fallback therefore only runs where this function
+  // used to throw: it repairs the resume that used to fail rather than changing
+  // one that used to work. On that path cloneDeep keeps the whole graph, which
+  // is the safe side to err on, since the clone is this turn's working copy.
+  // It goes through cloneDeepPreservingZodSchemas rather than plain cloneDeep:
+  // the values that make this state cyclic are the live zod schemas of the
+  // previous turn's MCP tools, and lodash drops a schema's non-enumerable
+  // internals, leaving a schema-shaped object that throws on first use.
+  // Mirrors cloneSessionState, which falls back the same way in run.ts.
+  let sessionState: SessionState
+  try {
+    sessionState = JSON.parse(JSON.stringify(baseSessionState)) as SessionState
+  } catch {
+    sessionState = cloneDeepPreservingZodSchemas(baseSessionState)
+  }
 
   // Apply maxAgentSteps override
   if (overrides.maxAgentSteps !== undefined) {