diff --git a/src/services/api/codexShim.ts b/src/services/api/codexShim.ts
index 392ddc7e..cb369d16 100644
--- a/src/services/api/codexShim.ts
+++ b/src/services/api/codexShim.ts
@@ -2,6 +2,7 @@ import { APIError } from '@anthropic-ai/sdk'
 import { buildAnthropicUsageFromRawUsage } from './cacheMetrics.js'
 import { compressToolHistory } from './compressToolHistory.js'
 import { fetchWithProxyRetry } from './fetchWithProxyRetry.js'
+import { stableStringify } from '../../utils/stableStringify.js'
 import type {
   ResolvedCodexCredentials,
   ResolvedProviderRequest,
@@ -559,7 +560,9 @@ export async function performCodexRequest(options: {
     {
       method: 'POST',
       headers,
-      body: JSON.stringify(body),
+      // WHY: byte-identity required for implicit prefix caching on
+      // OpenAI Responses API. See src/utils/stableStringify.ts.
+      body: stableStringify(body),
       signal: options.signal,
     },
   )
diff --git a/src/services/api/openaiShim.ts b/src/services/api/openaiShim.ts
index 015c0e2e..d04c5132 100644
--- a/src/services/api/openaiShim.ts
+++ b/src/services/api/openaiShim.ts
@@ -74,7 +74,12 @@ import {
   hasToolFieldMapping,
 } from './toolArgumentNormalization.js'
 import { logApiCallStart, logApiCallEnd } from '../../utils/requestLogging.js'
-import { createStreamState, processStreamChunk, getStreamStats } from '../../utils/streamingOptimizer.js'
+import {
+  createStreamState,
+  processStreamChunk,
+  getStreamStats,
+} from '../../utils/streamingOptimizer.js'
+import { stableStringify } from '../../utils/stableStringify.js'
 
 type SecretValueSource = Partial<{
   OPENAI_API_KEY: string
@@ -1852,12 +1857,17 @@ class OpenAIShimMessages {
       return false
     }
 
-    let serializedBody = JSON.stringify(
+    // WHY: byte-identity required for implicit prefix caching in
+    // OpenAI/Kimi/DeepSeek. stableStringify sorts object keys at every
+    // depth so spurious insertion-order differences across rebuilds of
+    // `body` (spread-merge, conditional assignments above) don't bust
+    // the provider's prefix hash.
+    let serializedBody = stableStringify(
       request.transport === 'responses' ? buildResponsesBody() : body,
     )
 
     const refreshSerializedBody = (): void => {
-      serializedBody = JSON.stringify(
+      serializedBody = stableStringify(
         request.transport === 'responses' ? buildResponsesBody() : body,
       )
     }
@@ -2036,7 +2046,7 @@ class OpenAIShimMessages {
             responsesResponse = await fetchWithProxyRetry(responsesUrl, {
               method: 'POST',
               headers,
-              body: JSON.stringify(responsesBody),
+              body: stableStringify(responsesBody),
               signal: options?.signal,
             })
           } catch (error) {
diff --git a/src/utils/serializationStability.test.ts b/src/utils/serializationStability.test.ts
new file mode 100644
index 00000000..3301ed2a
--- /dev/null
+++ b/src/utils/serializationStability.test.ts
@@ -0,0 +1,142 @@
+import { describe, expect, test } from 'bun:test'
+import { sortKeysDeep, stableStringify } from './stableStringify.js'
+
+// These tests pin byte-level stability of serialization helpers. The
+// invariant that matters for implicit prefix caching in OpenAI / Kimi /
+// DeepSeek / Codex — and for Anthropic cache_control breakpoints — is:
+// semantically-equal inputs must produce byte-identical output across
+// invocations and across key-order permutations.
+
+describe('stableStringify', () => {
+  test('two invocations with the same object produce identical strings', () => {
+    const a = stableStringify({ b: 1, a: 2 })
+    const b = stableStringify({ b: 1, a: 2 })
+    expect(a).toBe(b)
+  })
+
+  test('key order at the top level does not affect output', () => {
+    expect(stableStringify({ a: 1, b: 2 })).toBe(stableStringify({ b: 2, a: 1 }))
+  })
+
+  test('key order at nested depths does not affect output', () => {
+    const x = { outer: { z: 1, a: 2, m: { b: 3, a: 4 } } }
+    const y = { outer: { m: { a: 4, b: 3 }, a: 2, z: 1 } }
+    expect(stableStringify(x)).toBe(stableStringify(y))
+  })
+
+  test('array element order IS preserved (semantic in API contracts)', () => {
+    expect(stableStringify({ messages: ['a', 'b', 'c'] })).not.toBe(
+      stableStringify({ messages: ['c', 'b', 'a'] }),
+    )
+  })
+
+  test('arrays of objects have keys sorted inside each element', () => {
+    const out = stableStringify({
+      tools: [
+        { name: 'Bash', description: 'run' },
+        { description: 'read', name: 'Read' },
+      ],
+    })
+    expect(out).toBe(
+      '{"tools":[{"description":"run","name":"Bash"},{"description":"read","name":"Read"}]}',
+    )
+  })
+
+  test('undefined values are omitted (matches JSON.stringify)', () => {
+    const out = stableStringify({ a: undefined, b: 1 })
+    expect(out).toBe('{"b":1}')
+  })
+
+  test('primitive and null pass through unchanged', () => {
+    expect(stableStringify(null)).toBe('null')
+    expect(stableStringify(42)).toBe('42')
+    expect(stableStringify('x')).toBe('"x"')
+    expect(stableStringify(true)).toBe('true')
+  })
+
+  test('throws TypeError on circular structures (same behavior as JSON.stringify)', () => {
+    const obj: Record<string, unknown> = {}
+    obj.self = obj
+    // The exact message varies by engine (V8: "Converting circular structure
+    // to JSON", Bun: "JSON.stringify cannot serialize cyclic structures.").
+    // We only pin the error class — same contract as native JSON.stringify.
+    expect(() => stableStringify(obj)).toThrow(TypeError)
+    expect(() => JSON.stringify(obj)).toThrow(TypeError)
+  })
+
+  test('throws TypeError on circular references nested deep in the graph', () => {
+    const inner: Record<string, unknown> = { val: 1 }
+    const outer = { a: { b: inner } }
+    inner.cycle = outer
+    expect(() => stableStringify(outer)).toThrow(TypeError)
+  })
+
+  test('does not throw on DAGs (same object referenced from multiple keys)', () => {
+    const shared = { x: 1 }
+    // Native JSON.stringify handles this fine — stableStringify must too.
+    expect(() => stableStringify({ a: shared, b: shared })).not.toThrow()
+    expect(stableStringify({ a: shared, b: shared })).toBe(
+      '{"a":{"x":1},"b":{"x":1}}',
+    )
+  })
+})
+
+describe('sortKeysDeep', () => {
+  test('returns an object with sorted keys at every depth', () => {
+    const sorted = sortKeysDeep({
+      b: 1,
+      a: { y: 2, x: { d: 3, c: 4 } },
+    }) as Record<string, unknown>
+    expect(Object.keys(sorted)).toEqual(['a', 'b'])
+    expect(Object.keys(sorted.a as Record<string, unknown>)).toEqual([
+      'x',
+      'y',
+    ])
+  })
+
+  test('arrays are preserved element-wise', () => {
+    const sorted = sortKeysDeep([
+      { b: 1, a: 2 },
+      { d: 3, c: 4 },
+    ]) as Array<Record<string, unknown>>
+    expect(Object.keys(sorted[0]!)).toEqual(['a', 'b'])
+    expect(Object.keys(sorted[1]!)).toEqual(['c', 'd'])
+  })
+})
+
+describe('prefix caching invariants — end-to-end', () => {
+  // This is the real payload shape that an OpenAI-compatible body
+  // takes on its way to the upstream provider. We exercise it via
+  // stableStringify to verify that rebuilding the body with different
+  // key insertion orders yields the same bytes.
+  const bodyA = {
+    model: 'gpt-4o-mini',
+    stream: true,
+    messages: [
+      { role: 'system', content: 'you are helpful' },
+      { role: 'user', content: 'hi' },
+    ],
+    tools: [{ name: 't', description: 'x' }],
+    temperature: 0.7,
+    top_p: 1,
+  }
+  const bodyB = {
+    top_p: 1,
+    temperature: 0.7,
+    tools: [{ description: 'x', name: 't' }],
+    messages: [
+      { content: 'you are helpful', role: 'system' },
+      { content: 'hi', role: 'user' },
+    ],
+    stream: true,
+    model: 'gpt-4o-mini',
+  }
+
+  test('two spread-merged request bodies produce identical stable bytes', () => {
+    expect(stableStringify(bodyA)).toBe(stableStringify(bodyB))
+  })
+
+  test('calling stableStringify twice yields identical bytes (idempotent)', () => {
+    expect(stableStringify(bodyA)).toBe(stableStringify(bodyA))
+  })
+})
diff --git a/src/utils/stableStringify.test.ts b/src/utils/stableStringify.test.ts
new file mode 100644
index 00000000..91603ce7
--- /dev/null
+++ b/src/utils/stableStringify.test.ts
@@ -0,0 +1,199 @@
+import { describe, expect, test } from 'bun:test'
+import { sortKeysDeep, stableStringify } from './stableStringify'
+
+/**
+ * Contract: `stableStringify(input)` must equal `JSON.stringify(input)`
+ * for every value where the latter is well-defined, except that object
+ * keys are emitted in lexicographic order at every depth. These tests
+ * focus on the native pre-processing semantics — `toJSON(key)` and
+ * primitive-wrapper unboxing — that the deep-sort path must preserve.
+ */
+
+describe('stableStringify — toJSON semantics', () => {
+  test('Date at top level → ISO string', () => {
+    const d = new Date('2024-01-02T03:04:05.678Z')
+    expect(stableStringify(d)).toBe(JSON.stringify(d))
+  })
+
+  test('Date nested in object → ISO string + sorted keys', () => {
+    const d = new Date('2024-01-02T03:04:05.678Z')
+    const input = { z: 1, when: d, a: 'x' }
+    expect(stableStringify(input)).toBe(
+      `{"a":"x","␟when␟":"PLACEHOLDER","z":1}`
+        .replace('␟when␟', 'when')
+        .replace('"PLACEHOLDER"', JSON.stringify(d.toISOString())),
+    )
+  })
+
+  test('Date inside an array → each element converted', () => {
+    const a = new Date('2024-01-02T03:04:05.678Z')
+    const b = new Date('2025-06-07T08:09:10.111Z')
+    const input = [a, b]
+    expect(stableStringify(input)).toBe(JSON.stringify(input))
+  })
+
+  test('URL value serializes via URL.prototype.toJSON', () => {
+    const u = new URL('https://example.com/path?q=1')
+    expect(stableStringify(u)).toBe(JSON.stringify(u))
+    expect(stableStringify({ url: u })).toBe(JSON.stringify({ url: u }))
+  })
+
+  test('custom class with toJSON returning a plain object → keys sorted', () => {
+    class Thing {
+      toJSON() {
+        return { z: 1, a: 2, m: 3 }
+      }
+    }
+    const out = stableStringify(new Thing())
+    expect(out).toBe('{"a":2,"m":3,"z":1}')
+  })
+
+  test('toJSON(key) receives the property name for object values', () => {
+    const seen: string[] = []
+    class Trace {
+      toJSON(k: string) {
+        seen.push(k)
+        return k
+      }
+    }
+    const t = new Trace()
+    stableStringify({ alpha: t, beta: t })
+    // Object keys are sorted, so toJSON is invoked alpha-first.
+    expect(seen).toEqual(['alpha', 'beta'])
+  })
+
+  test('toJSON(key) receives the array index as a string for array elements', () => {
+    const seen: string[] = []
+    class Trace {
+      toJSON(k: string) {
+        seen.push(k)
+        return k
+      }
+    }
+    const t = new Trace()
+    stableStringify([t, t, t])
+    expect(seen).toEqual(['0', '1', '2'])
+  })
+
+  test('toJSON(key) receives empty string at top level', () => {
+    let captured: string | undefined
+    class Trace {
+      toJSON(k: string) {
+        captured = k
+        return 'ok'
+      }
+    }
+    stableStringify(new Trace())
+    expect(captured).toBe('')
+  })
+
+  test('toJSON returning undefined drops the property (matches native)', () => {
+    class Hidden {
+      toJSON() {
+        return undefined
+      }
+    }
+    const input = { a: 1, gone: new Hidden(), b: 2 }
+    expect(stableStringify(input)).toBe(JSON.stringify(input))
+    expect(stableStringify(input)).toBe('{"a":1,"b":2}')
+  })
+
+  test('nested mix: object with a Date field and a regular field → keys sorted, Date as ISO', () => {
+    const d = new Date('2024-01-02T03:04:05.678Z')
+    const input = { z: { when: d, a: 1 }, a: 'first' }
+    expect(stableStringify(input)).toBe(
+      `{"a":"first","z":{"a":1,"when":${JSON.stringify(d.toISOString())}}}`,
+    )
+  })
+})
+
+describe('stableStringify — primitive wrapper unboxing', () => {
+  test('new Number at top level → numeric primitive', () => {
+    const n = new Number(42)
+    expect(stableStringify(n)).toBe(JSON.stringify(n))
+    expect(stableStringify(n)).toBe('42')
+  })
+
+  test('new String at top level → string primitive', () => {
+    const s = new String('hello')
+    expect(stableStringify(s)).toBe(JSON.stringify(s))
+    expect(stableStringify(s)).toBe('"hello"')
+  })
+
+  test('new Boolean at top level → boolean primitive', () => {
+    const b = new Boolean(true)
+    expect(stableStringify(b)).toBe(JSON.stringify(b))
+    expect(stableStringify(b)).toBe('true')
+  })
+
+  test('new Boolean(false) at top level → false', () => {
+    const b = new Boolean(false)
+    expect(stableStringify(b)).toBe(JSON.stringify(b))
+    expect(stableStringify(b)).toBe('false')
+  })
+
+  test('boxed wrappers as object values → primitives + sorted keys', () => {
+    const input = {
+      z: new Number(1),
+      a: new String('x'),
+      m: new Boolean(false),
+    }
+    expect(stableStringify(input)).toBe('{"a":"x","m":false,"z":1}')
+    // Native form: same primitive shape (without sort guarantee).
+    expect(JSON.parse(stableStringify(input))).toEqual(JSON.parse(JSON.stringify(input)))
+  })
+})
+
+describe('stableStringify — cycles vs DAGs', () => {
+  test('top-level cycle throws TypeError (regression guard)', () => {
+    const obj: Record<string, unknown> = { a: 1 }
+    obj.self = obj
+    expect(() => stableStringify(obj)).toThrow(TypeError)
+  })
+
+  test('deep cycle throws TypeError', () => {
+    const a: Record<string, unknown> = { name: 'a' }
+    const b: Record<string, unknown> = { name: 'b' }
+    a.next = b
+    b.back = a
+    expect(() => stableStringify(a)).toThrow(TypeError)
+  })
+
+  test('toJSON returning an ancestor still triggers the cycle check', () => {
+    type Node = { name: string; child?: { toJSON(): Node } }
+    const parent: Node = { name: 'parent' }
+    parent.child = {
+      toJSON() {
+        return parent
+      },
+    }
+    expect(() => stableStringify(parent)).toThrow(TypeError)
+  })
+
+  test('DAG (same object referenced twice via different keys) does NOT throw', () => {
+    const shared = { v: 1 }
+    const input = { left: shared, right: shared }
+    expect(() => stableStringify(input)).not.toThrow()
+    expect(stableStringify(input)).toBe('{"left":{"v":1},"right":{"v":1}}')
+  })
+
+  test('DAG of arrays does NOT throw', () => {
+    const shared = [1, 2, 3]
+    const input = { a: shared, b: shared }
+    expect(() => stableStringify(input)).not.toThrow()
+    expect(stableStringify(input)).toBe('{"a":[1,2,3],"b":[1,2,3]}')
+  })
+})
+
+describe('sortKeysDeep — same toJSON/unbox semantics', () => {
+  test('returns the post-toJSON, post-unbox sorted shape', () => {
+    const d = new Date('2024-01-02T03:04:05.678Z')
+    const out = sortKeysDeep({ z: 1, a: new Number(7), when: d }) as Record<
+      string,
+      unknown
+    >
+    expect(out).toEqual({ a: 7, when: d.toISOString(), z: 1 })
+    // Key order in the returned object is lexicographic.
+    expect(Object.keys(out)).toEqual(['a', 'when', 'z'])
+  })
+})
diff --git a/src/utils/stableStringify.ts b/src/utils/stableStringify.ts
new file mode 100644
index 00000000..0548f949
--- /dev/null
+++ b/src/utils/stableStringify.ts
@@ -0,0 +1,132 @@
+/**
+ * Deterministic JSON serialization.
+ *
+ * WHY: OpenAI / Kimi / DeepSeek / Codex all use **implicit prefix caching**
+ * — the server hashes the request prefix and reuses cached reasoning if
+ * the bytes match exactly. Even a trivial key-order difference between
+ * two otherwise-identical requests invalidates the hash and forces a
+ * full re-parse.
+ *
+ * This is also a pre-requisite for Anthropic / Bedrock / Vertex
+ * `cache_control` breakpoints: ephemeral cache entries match on exact
+ * content, so a re-ordered object literal busts the breakpoint.
+ *
+ * `JSON.stringify` is nondeterministic across engines and across
+ * successive iterations when objects carry keys added at different
+ * times (V8 preserves insertion order, which is the common failure
+ * mode when building a body from spread-merged configs).
+ *
+ * This helper recursively sorts object keys. Arrays preserve order
+ * (element order IS semantically significant in message/content arrays).
+ *
+ * Complements `sortKeysDeep` in src/services/remoteManagedSettings and
+ * src/services/policyLimits. Those two are INTENTIONALLY separate:
+ *   - remoteManagedSettings: matches Python `json.dumps(sort_keys=True)`
+ *     byte-for-byte to validate server-computed checksums. Must NOT
+ *     drop undefined (Python preserves null).
+ *   - policyLimits: uses `localeCompare` (keeps legacy behavior; locale-
+ *     sensitive but stable for a given runtime).
+ *   - this module (stableStringify): byte-identity for API body caching.
+ *     Drops undefined to match `JSON.stringify` — the openaiShim/codexShim
+ *     body is always downstream of `JSON.stringify` semantics.
+ * Do not consolidate without auditing the 3 callers — each has a
+ * different server-compat contract.
+ */
+
+/**
+ * Returns a byte-stable JSON string representation.
+ * - Object keys are emitted in lexicographic order at every depth.
+ * - Array element order is preserved.
+ * - Undefined values are dropped (matching `JSON.stringify`).
+ * - Indentation matches the `space` argument (0 by default → compact).
+ *
+ * Native `JSON.stringify` pre-processing is preserved before sorting:
+ *   - `toJSON(key)` is invoked on objects that define it (own or
+ *     inherited — covers `Date`, `URL`, and any user class). The `key`
+ *     argument is the property name for nested object values, the array
+ *     index as a string for array elements, and `''` for the top-level
+ *     call, matching native semantics.
+ *   - Boxed primitive wrappers (`new Number(...)`, `new String(...)`,
+ *     `new Boolean(...)`) are unboxed to their primitive form.
+ * Both happen BEFORE the array/object branches dispatch, so the value
+ * actually walked is the post-conversion form. If `toJSON` returns
+ * `undefined`, the value is dropped from its parent (matching native
+ * `JSON.stringify`).
+ *
+ * Single-pass: `deepSort` walks the (possibly converted) value tree
+ * once, building a sorted clone. A `WeakSet` of ancestors tracks the
+ * current path through the object graph so that circular references
+ * throw `TypeError` (same contract as native `JSON.stringify`). The
+ * cycle check runs on the post-`toJSON` value, so a `toJSON` impl that
+ * returns an ancestor still throws. Ancestors are always removed in a
+ * `finally` block when unwinding out of each object branch (even on
+ * exception), so DAG inputs — where the same object is reachable via
+ * multiple keys — are handled correctly and do not throw.
+ */
+export function stableStringify(value: unknown, space?: number): string {
+  return JSON.stringify(deepSort(value, new WeakSet(), ''), null, space)
+}
+
+/**
+ * Returns a deep-sorted clone of the input: object keys lexicographic
+ * at every depth, arrays preserved. Useful when callers need to feed
+ * the sorted shape into a downstream serializer (e.g., when they must
+ * call `JSON.stringify` with a custom spacing or replacer).
+ *
+ * Applies the same `toJSON(key)` invocation and primitive-wrapper
+ * unboxing as `stableStringify`, so the returned shape mirrors what
+ * native `JSON.stringify` would have walked.
+ */
+export function sortKeysDeep<T>(value: T): T {
+  return deepSort(value, new WeakSet(), '') as T
+}
+
+function deepSort(
+  value: unknown,
+  ancestors: WeakSet<object>,
+  key: string,
+): unknown {
+  // Step 1: invoke toJSON(key) if present — matches native pre-processing.
+  if (
+    value !== null &&
+    typeof value === 'object' &&
+    typeof (value as { toJSON?: unknown }).toJSON === 'function'
+  ) {
+    value = (value as { toJSON: (k: string) => unknown }).toJSON(key)
+  }
+
+  // Step 2: unbox primitive wrappers.
+  if (value instanceof Number) value = Number(value)
+  else if (value instanceof String) value = String(value)
+  else if (value instanceof Boolean) value = Boolean(value.valueOf())
+
+  // Step 3: primitives short-circuit (post-toJSON the value may now be one).
+  if (value === null || typeof value !== 'object') return value
+
+  // Step 4: arrays — element key is the index as a string.
+  if (Array.isArray(value)) {
+    return value.map((v, i) => deepSort(v, ancestors, String(i)))
+  }
+
+  // Step 5: cycle check on the post-toJSON value.
+  if (ancestors.has(value as object)) {
+    throw new TypeError('Converting circular structure to JSON')
+  }
+  ancestors.add(value as object)
+  try {
+    const sorted: Record<string, unknown> = {}
+    for (const k of Object.keys(value as Record<string, unknown>).sort()) {
+      const child = deepSort(
+        (value as Record<string, unknown>)[k],
+        ancestors,
+        k,
+      )
+      if (child === undefined) continue
+      sorted[k] = child
+    }
+    return sorted
+  } finally {
+    ancestors.delete(value as object)
+  }
+}
+