diff --git a/src/subagent.ts b/src/subagent.ts
index cb1462ee7..bf1c48ce3 100644
--- a/src/subagent.ts
+++ b/src/subagent.ts
@@ -1702,8 +1702,13 @@ ${runtimeSnapshotContext('user', plan.entries)}`
       const completionPersona = `${persona}
 
 Completion protocol: call \`${resultToolName}\` exactly once with the final result matching its parameter schema. This is the only completion channel for this run. Do not finish with a plain-text answer.`
-      const perOpMaxTokens = operation === 'migration' || operation === 'compaction' || operation === 'document-archive' ? 8_192
-        : operation === 'metadata-maintenance' ? 4_096
+      // Maintenance delegates emit small structured results, but reasoning-style
+      // models can burn a large completion budget on analysis before emitting
+      // them. Undersized static caps abort the whole operation with
+      // stopReason=max-tokens (e.g. dense CJK capacity archives, see #70), so
+      // keep generous ceilings instead of tight ones.
+      const perOpMaxTokens = operation === 'migration' || operation === 'compaction' || operation === 'document-archive' ? 32_768
+        : operation === 'metadata-maintenance' ? 16_384
         : undefined
       const fixed = this.taskAgentModelResolver?.()
       const baseAgentOptions = perOpMaxTokens === undefined ? undefined : { maxTokens: perOpMaxTokens }
diff --git a/tests/subagent.spec.ts b/tests/subagent.spec.ts
index a0af11702..9038fce8e 100644
--- a/tests/subagent.spec.ts
+++ b/tests/subagent.spec.ts
@@ -815,7 +815,7 @@ describe('Mnemon memory subagent coordinator', () => {
     })
     expect(host.start).toHaveBeenCalledWith('spawn', expect.objectContaining({
       toolFilter: { allow: [expect.stringMatching(/^mnemon_subagent_result_/)] },
-      agentOptions: { maxTokens: 4_096 },
+      agentOptions: { maxTokens: 16_384 },
       persona: expect.stringContaining('fastest bounded metadata-sampling path'),
     }))
     expect(memoryService.metadataSample).toHaveBeenCalledWith('product', expect.any(AbortSignal))
@@ -1001,7 +1001,7 @@ describe('Mnemon memory subagent coordinator', () => {
     })
     expect(host.start).toHaveBeenCalledWith('spawn', expect.objectContaining({
       toolFilter: { allow: [expect.stringMatching(/^mnemon_subagent_result_/)] },
-      agentOptions: { maxTokens: 8_192 },
+      agentOptions: { maxTokens: 32_768 },
     }))
     const migrationCall = (host.start.mock.calls[0] as unknown as [string, { prompt: Array<{ text: string }>; persona: string; toolFilter: { allow: string[] } }])[1]
     const migrationPrompt = migrationCall.prompt[0]!.text
@@ -1244,7 +1244,7 @@ describe('Mnemon memory subagent coordinator', () => {
     })
     expect(host.start).toHaveBeenCalledWith('spawn', expect.objectContaining({
       toolFilter: { allow: [expect.stringMatching(/^mnemon_subagent_result_/)] },
-      agentOptions: { maxTokens: 8_192 },
+      agentOptions: { maxTokens: 32_768 },
       persona: expect.stringContaining('local USER.md compactor'),
     }))
     const compactionCall = (host.start.mock.calls[0] as unknown as [string, { prompt: Array<{ text: string }>; persona: string }])[1]
@@ -1384,7 +1384,7 @@ describe('Mnemon memory subagent coordinator', () => {
       maintenance: { kind: 'local-compaction' },
     })
     expect(host.start).toHaveBeenCalledWith('spawn', expect.objectContaining({
-      agentOptions: { provider: 'pinned-provider', model: 'pinned-model', maxTokens: 8_192 },
+      agentOptions: { provider: 'pinned-provider', model: 'pinned-model', maxTokens: 32_768 },
     }))
   })
 })