diff --git a/src/agents/hooks/deferred-tools.ts b/src/agents/hooks/deferred-tools.ts
index cc4aa620..9df8039f 100644
--- a/src/agents/hooks/deferred-tools.ts
+++ b/src/agents/hooks/deferred-tools.ts
@@ -17,10 +17,15 @@ export const deferredToolsHook: Hook = {
     }
 
     if (ctx.systemPrompt.hasOverride) {
+      // Custom systemPrompt path: every tool ships inline in the catalog,
+      // there is no ToolSearch split. Mirror that into the prompt-cache
+      // fields so downstream renderers see a self-consistent shape.
       ctx.setTurnCatalog({
         tools: ctx.allTools,
         deferred: [],
         deferredEnabled: false,
+        inlineTools: ctx.allTools,
+        discoveredCatalogTools: [],
       })
       return
     }
diff --git a/src/agents/hooks/forward-progress-to-channel.test.ts b/src/agents/hooks/forward-progress-to-channel.test.ts
index 0f7afb70..113536c0 100644
--- a/src/agents/hooks/forward-progress-to-channel.test.ts
+++ b/src/agents/hooks/forward-progress-to-channel.test.ts
@@ -158,7 +158,7 @@ function hookContext(onAssistantTurn: (text: string) => void): HookContext {
       hasOverride: false,
       renderEffective: () => '',
     },
-    turnCatalog: { tools: [], deferred: [], deferredEnabled: false },
+    turnCatalog: { tools: [], deferred: [], deferredEnabled: false, inlineTools: [], discoveredCatalogTools: [] },
     setTurnCatalog() {},
     mergeUsage() {},
     markDidCompact() {},
diff --git a/src/agents/hooks/split-render.ts b/src/agents/hooks/split-render.ts
index d3080466..382a5793 100644
--- a/src/agents/hooks/split-render.ts
+++ b/src/agents/hooks/split-render.ts
@@ -21,6 +21,12 @@ export const splitRenderHook: Hook = {
     const sessionCtx = getCurrentSessionContext()
     const rendered = renderSystemPromptSplit(ctx.systemPrompt.template, getTodos(), {
       tools: ctx.turnCatalog.tools,
+      // 2026-05-26 cache anchoring: render only the always-loaded subset into
+      // the stable `## Tool Catalog`; route already-discovered deferred tools
+      // through the variable suffix so promoting a new tool via ToolSearch no
+      // longer extends the stable system prompt and breaks OpenAI prefix-cache.
+      inlineCatalogTools: ctx.turnCatalog.inlineTools,
+      discoveredCatalogTools: ctx.turnCatalog.discoveredCatalogTools,
       deferredTools: ctx.turnCatalog.deferred,
       discoveredTools: sessionCtx?.discoveredTools,
     })
diff --git a/src/prompt.test.ts b/src/prompt.test.ts
index 6e91ca40..7b788a6a 100644
--- a/src/prompt.test.ts
+++ b/src/prompt.test.ts
@@ -74,6 +74,111 @@ describe('renderSystemPromptSplit — cache anchoring', () => {
     })
     assert.equal(variable, '')
   })
+
+  // Regression: 2026-05-26 dogfood §cache hit rate root-cause.
+  // Even after fix/cache-suffix-relocate moved TodoList + deferred-reminder
+  // out of `instructions`, the `## Tool Catalog` section still listed every
+  // discovered tool. Each ToolSearch promotion grew that section a few bytes
+  // and broke OpenAI's prefix-cache fingerprint anywhere after it. Result:
+  // dogfood cache hit ~11% instead of the expected 60-80% — see usage.jsonl
+  // 1,536-token-truncation pattern. The fix routes already-discovered tool
+  // descriptions into the variable suffix so the stable prefix is invariant
+  // under discoveredTools changes.
+  describe('cache anchoring under discovered tool promotion', () => {
+    it('keeps stable prefix byte-identical when discoveredTools grows', () => {
+      const allTools = [
+        fakeTool('Bash'),
+        fakeTool('Read'),
+        fakeTool('ToolSearch'),
+        fakeTool('WebSearch'),
+        fakeTool('Dispatch'),
+      ]
+      const inlineCatalogTools = [
+        fakeTool('Bash'),
+        fakeTool('Read'),
+        fakeTool('ToolSearch'),
+      ]
+
+      // Turn N: only Bash / Read / ToolSearch are callable.
+      const beforePromotion = renderSystemPromptSplit(template, [], {
+        tools: inlineCatalogTools,
+        inlineCatalogTools,
+        discoveredCatalogTools: [],
+        deferredTools: [fakeTool('WebSearch'), fakeTool('Dispatch')],
+        discoveredTools: new Map(),
+      })
+
+      // Turn N+k: model called ToolSearch and promoted WebSearch + Dispatch.
+      const afterPromotion = renderSystemPromptSplit(template, [], {
+        tools: allTools,
+        inlineCatalogTools,
+        discoveredCatalogTools: [fakeTool('WebSearch'), fakeTool('Dispatch')],
+        deferredTools: [fakeTool('WebSearch'), fakeTool('Dispatch')],
+        discoveredTools: new Map([
+          ['WebSearch', 5],
+          ['Dispatch', 5],
+        ]),
+      })
+
+      assert.equal(
+        beforePromotion.stable,
+        afterPromotion.stable,
+        'stable prefix must not change when ToolSearch promotes deferred tools',
+      )
+      assert.notEqual(
+        beforePromotion.variable,
+        afterPromotion.variable,
+        'variable suffix must carry the discovered tool descriptions',
+      )
+    })
+
+    it('renders discovered tool descriptions in the variable suffix only', () => {
+      const inlineCatalogTools = [fakeTool('Bash'), fakeTool('ToolSearch')]
+      const { stable, variable } = renderSystemPromptSplit(template, [], {
+        tools: [...inlineCatalogTools, fakeTool('WebSearch')],
+        inlineCatalogTools,
+        discoveredCatalogTools: [fakeTool('WebSearch')],
+        deferredTools: [fakeTool('WebSearch')],
+        discoveredTools: new Map([['WebSearch', 1]]),
+      })
+      assert.doesNotMatch(stable, /WebSearch/)
+      assert.match(variable, /WebSearch/)
+      // The discovered-tools reminder is its own <system-reminder> block,
+      // distinct from the undiscovered-deferred reminder (no overlap with
+      // the existing test above).
+      assert.doesNotMatch(
+        variable,
+        /are now available via ToolSearch\. Their schemas are NOT loaded/,
+      )
+    })
+
+    it('omits the discovered-tools section when nothing has been promoted yet', () => {
+      const inlineCatalogTools = [fakeTool('Bash'), fakeTool('ToolSearch')]
+      const { variable } = renderSystemPromptSplit(template, [], {
+        tools: inlineCatalogTools,
+        inlineCatalogTools,
+        discoveredCatalogTools: [],
+        deferredTools: [fakeTool('WebSearch')],
+        discoveredTools: new Map(),
+      })
+      // Variable still carries the undiscovered-deferred reminder; only the
+      // new discovered-tools block should be absent.
+      assert.doesNotMatch(variable, /were loaded via ToolSearch earlier/)
+    })
+
+    it('falls back to rendering full tools[] when inlineCatalogTools is omitted', () => {
+      // Backward-compat path: callers that haven't updated still get the
+      // pre-fix behavior of rendering the entire tools array in the catalog.
+      // This keeps unrelated callers (tests, future custom systemPrompt
+      // shapes) from breaking at the type level.
+      const { stable } = renderSystemPromptSplit(template, [], {
+        tools: [fakeTool('Bash'), fakeTool('Read'), fakeTool('WebSearch')],
+      })
+      assert.match(stable, /^- Bash$/m)
+      assert.match(stable, /^- Read$/m)
+      assert.match(stable, /^- WebSearch$/m)
+    })
+  })
 })
 
 function fakeTool(name: string): Tool {
diff --git a/src/prompt.ts b/src/prompt.ts
index 2f815df1..4992de74 100644
--- a/src/prompt.ts
+++ b/src/prompt.ts
@@ -96,6 +96,14 @@ export type SystemPromptRenderOptions = {
   tools: Tool[]
   deferredTools?: Tool[]
   discoveredTools?: ReadonlyMap<string, number>
+  // 2026-05-26: cache anchoring. When `inlineCatalogTools` is provided, the
+  // stable `## Tool Catalog` section renders only that subset, and the
+  // remainder of `tools` (discovered via ToolSearch this session) is rendered
+  // separately into the variable suffix. Callers that don't pass this still
+  // fall back to rendering the full `tools` array in the catalog — preserves
+  // backward compat for tests / custom systemPrompt shapes.
+  inlineCatalogTools?: Tool[]
+  discoveredCatalogTools?: Tool[]
 }
 
 function formatSkillsSection(role: Role): string {
@@ -572,8 +580,13 @@ export function renderSystemPromptSplit(
   todos: TodoItem[],
   options?: SystemPromptRenderOptions,
 ): RenderedSystemPrompt {
-  const tools = options?.tools ?? []
-  const toolDescriptions = formatToolCatalog(tools)
+  // Stable catalog: prefer the caller-supplied inlineCatalogTools subset (the
+  // always-loaded tools, whose membership is fixed for the query loop). Fall
+  // back to the full tools[] for callers that haven't opted into the split
+  // (tests, future custom systemPrompt shapes). The discovered subset is
+  // routed into the variable suffix below.
+  const stableCatalogTools = options?.inlineCatalogTools ?? options?.tools ?? []
+  const toolDescriptions = formatToolCatalog(stableCatalogTools)
   const toolSection = [
     '## Tool Catalog',
     'Available tools (full schemas / usage live in the tools API description field):',
@@ -588,6 +601,12 @@ export function renderSystemPromptSplit(
   if (todoSection) {
     variableParts.push(todoSection)
   }
+  const discoveredReminder = buildDiscoveredToolsReminder(
+    options?.discoveredCatalogTools ?? [],
+  )
+  if (discoveredReminder) {
+    variableParts.push(discoveredReminder)
+  }
   const deferredReminder = buildDeferredToolsReminder(
     options?.deferredTools ?? [],
     options?.discoveredTools ?? new Map(),
@@ -682,3 +701,27 @@ ${undiscovered
   .join('\n')}
 </system-reminder>`
 }
+
+// 2026-05-26: discovered (already-promoted) deferred tools used to be listed
+// in the stable `## Tool Catalog` section, which made the stable system
+// prompt grow each time ToolSearch promoted a new tool. Under OpenAI's
+// auto-cache (wire-byte prefix fingerprint), this broke prefix cache for
+// every subsequent turn — dogfood §cache hit rate measured ~11% instead of
+// the expected 60-80%. The fix renders these tools in the variable suffix
+// (injected into the last user message) so the stable section is invariant
+// under discoveredTools changes. The block name lists the tool names + their
+// whenToUse so the model has the same one-line hint it had pre-fix; full
+// schemas still live in the provider's tools API field.
+function buildDiscoveredToolsReminder(
+  discoveredCatalogTools: readonly Tool[],
+): string {
+  if (discoveredCatalogTools.length === 0) {
+    return ''
+  }
+  return `<system-reminder>
+The following deferred tools were loaded via ToolSearch earlier in this session and are now callable. Full schemas live in the tools API description field — call them by name:
+${discoveredCatalogTools
+  .map(tool => (tool.whenToUse ? `- ${tool.name}: ${tool.whenToUse}` : `- ${tool.name}`))
+  .join('\n')}
+</system-reminder>`
+}
diff --git a/src/query.ts b/src/query.ts
index b2e19369..1f269a3c 100644
--- a/src/query.ts
+++ b/src/query.ts
@@ -491,6 +491,8 @@ export async function query(params: QueryParams): Promise<{
     tools: params.tools,
     deferred: [],
     deferredEnabled: false,
+    inlineTools: params.tools,
+    discoveredCatalogTools: [],
   }
 
   const makeHookContext = (messagesSnapshot?: Message[]): HookContext => ({
@@ -580,6 +582,8 @@ export async function query(params: QueryParams): Promise<{
       tools: params.tools,
       deferred: [],
       deferredEnabled: false,
+      inlineTools: params.tools,
+      discoveredCatalogTools: [],
     }
     for (const hook of lifecycleHooks) {
       await hook.beforeTurn?.(makeHookContext())
diff --git a/src/tools/deferred-loading.ts b/src/tools/deferred-loading.ts
index 07d6016b..3caae1e2 100644
--- a/src/tools/deferred-loading.ts
+++ b/src/tools/deferred-loading.ts
@@ -8,6 +8,17 @@ export type TurnToolCatalog = {
   tools: Tool[]
   deferred: Tool[]
   deferredEnabled: boolean
+  // 2026-05-26: prompt-cache split. `inlineTools` is the always-loaded subset
+  // (alwaysLoad-tagged + ToolSearch when deferred loading is on); its
+  // membership is fixed for the lifetime of one query loop, so rendering it
+  // into the `## Tool Catalog` stable section keeps that section byte-stable.
+  // `discoveredCatalogTools` is the this-turn promoted-via-ToolSearch subset;
+  // its membership changes between turns, so the prompt renders it inside the
+  // variable suffix (injected into the last user message) instead of the
+  // stable system section. Both are still in `tools` so the provider's tools
+  // array carries every callable schema this turn.
+  inlineTools: Tool[]
+  discoveredCatalogTools: Tool[]
 }
 
 export function buildTurnToolCatalog(input: {
@@ -17,20 +28,26 @@ export function buildTurnToolCatalog(input: {
 }): TurnToolCatalog {
   const { allTools, discoveredTools, config } = input
   if (!shouldEnableDeferredLoading(config, allTools)) {
+    const all = [...allTools]
     return {
-      tools: [...allTools],
+      tools: all,
       deferred: [],
       deferredEnabled: false,
+      inlineTools: all,
+      discoveredCatalogTools: [],
     }
   }
 
   const { alwaysLoaded, deferred } = partitionTools(allTools)
   const discovered = deferred.filter(tool => discoveredTools.has(tool.name))
-  const tools = [...alwaysLoaded, toolSearchTool, ...discovered]
+  const inlineTools = dedupeTools([...alwaysLoaded, toolSearchTool])
+  const tools = dedupeTools([...inlineTools, ...discovered])
   return {
-    tools: dedupeTools(tools),
+    tools,
     deferred,
     deferredEnabled: true,
+    inlineTools,
+    discoveredCatalogTools: discovered,
   }
 }