diff --git a/packages/release-bot/README.md b/packages/release-bot/README.md
index 52159ed..e9a4c52 100644
--- a/packages/release-bot/README.md
+++ b/packages/release-bot/README.md
@@ -27,7 +27,11 @@ as the truncation it is. The evidence roles have five turns and the planner and
 reviewer three. Every DAG task has `maxRetries: 0`, so a failed role is not
 silently rerun as a whole new analysis (OMA's one in-run structured-output
 correction still applies). The complete planning DAG has a thirty-minute
-wall-clock deadline.
+wall-clock deadline. Every request also sets DeepSeek's JSON output mode
+(`response_format: json_object`), so the provider guarantees that the answer
+parses as JSON. OMA still validates the answer against the role's schema, so
+the in-run correction only has to handle a schema mismatch, not a bracket
+error in several thousand characters of nested output.
 
 Repository diffs are untrusted evidence. The analyst and compatibility auditor
 receive only three custom read-only tools: immutable release evidence, a
diff --git a/packages/release-bot/src/orchestrator.ts b/packages/release-bot/src/orchestrator.ts
index 08f0c9d..dff5785 100644
--- a/packages/release-bot/src/orchestrator.ts
+++ b/packages/release-bot/src/orchestrator.ts
@@ -74,7 +74,7 @@ export async function generateReleaseDecision(
   const model = options.model ?? DEFAULT_MODEL
   const shared: Pick<AgentConfig,
     'model' | 'provider' | 'adapter' | 'apiKey' | 'temperature' | 'thinking' | 'maxTokens' |
-    'parallelToolCalls' | 'maxToolOutputChars' | 'compressToolResults'> = {
+    'parallelToolCalls' | 'extraBody' | 'maxToolOutputChars' | 'compressToolResults'> = {
       model,
       provider: options.adapter ? undefined : 'deepseek',
       adapter: options.adapter,
@@ -90,6 +90,15 @@ export async function generateReleaseDecision(
       // replace it because it is only checked after a call returns.
       maxTokens: 64_000,
       parallelToolCalls: false,
+      // DeepSeek's JSON output mode makes the provider guarantee that the
+      // answer parses. Two roles failed in 2026-09 on bracket mismatches deep
+      // inside long nested output, and the in-run correction failed the same
+      // way; #599 moved the nested object last as a workaround. The mode is
+      // accepted together with thinking and tool calls, and the structured
+      // output instruction already carries the word "json" that DeepSeek
+      // requires in the prompt. Schema conformance is still validated by OMA;
+      // this only removes the syntax failure class.
+      extraBody: { response_format: { type: 'json_object' } },
       maxToolOutputChars: 75_000,
       compressToolResults: { minChars: 2_000 },
     }
diff --git a/packages/release-bot/tests/orchestrator.test.ts b/packages/release-bot/tests/orchestrator.test.ts
index deaf04a..993c114 100644
--- a/packages/release-bot/tests/orchestrator.test.ts
+++ b/packages/release-bot/tests/orchestrator.test.ts
@@ -93,6 +93,14 @@ describe('OMA release orchestration', () => {
       ['release-planner', { enabled: true, effort: 'max' }],
       ['release-reviewer', { enabled: true, effort: 'max' }],
     ]))
+    // Every role asks DeepSeek for JSON output mode so the provider, not the
+    // in-run correction, guarantees the answer parses.
+    expect(adapter.extraBodyByRole).toEqual(new Map([
+      ['change-analyst', { response_format: { type: 'json_object' } }],
+      ['compatibility-auditor', { response_format: { type: 'json_object' } }],
+      ['release-planner', { response_format: { type: 'json_object' } }],
+      ['release-reviewer', { response_format: { type: 'json_object' } }],
+    ]))
     expect(run.tokenUsage).toEqual({ input_tokens: 40, output_tokens: 20 })
   })
 
@@ -193,6 +201,7 @@ class ReleaseScriptAdapter implements LLMAdapter {
   readonly toolSets: string[][] = []
   readonly maxTokensByRole = new Map<string, number | undefined>()
   readonly thinkingByRole = new Map<string, ThinkingConfig | undefined>()
+  readonly extraBodyByRole = new Map<string, Record<string, unknown> | undefined>()
   plannerMessages = ''
   reviewerMessages = ''
   private sequence = 0
@@ -203,6 +212,7 @@ class ReleaseScriptAdapter implements LLMAdapter {
     this.toolSets.push((options.tools ?? []).map(tool => tool.name))
     this.maxTokensByRole.set(role, options.maxTokens)
     this.thinkingByRole.set(role, options.thinking)
+    this.extraBodyByRole.set(role, options.extraBody)
     const messageText = JSON.stringify(messages)
     if (role === 'release-planner') this.plannerMessages = messageText
     if (role === 'release-reviewer') this.reviewerMessages = messageText