Release/v6.8.5

[code-crusher]

Release v6.8.5

Fixed

• Placeholder-tagged tool arguments are repaired instead of dropped. When a model emits malformed native tool-call JSON with a placeholder tag in value position (e.g. "limit": <longcat_arg_value>180 or "limit": <longcat_arg_value>), AssistantMessageParser now strips the tag and keeps the trailing value, or drops the dangling key entirely when no value follows so the tool's default applies. The repair runs only after strict JSON.parse fails, so valid JSON is never altered. useMcpToolTool applies the same cleanup when validating MCP tool arguments, and streaming partial previews are tag-cleaned as well.

Includes version bump to 6.8.5 and changelog entry.

[matterai-app]

✅ Reviewed the changes: Solid, well-tested placeholder-tag repair with good streaming coverage. One logging issue: the repair warning fires on every failed candidate (including every streaming delta) instead of only on successful repair. Reviewed src/core/assistant-message/tests/AssistantMessageParser.spec.ts: no issues found. Reviewed src/core/tools/useMcpToolTool.ts: no issues found. Reviewed src/package.json: version bump only, no issues.

[matterai-app]

🧪 PR Review is completed: Well-structured JSON-repair layer with strict-parse-first semantics and strong test coverage; the prior per-delta console.warn spam is fixed. One concern to verify: the new repair-note push in presentAssistantMessage may create a second tool_result block with a duplicate tool_use_id. Reviewed src/core/assistant-message/AssistantMessageParser.ts: no other issues found. Reviewed src/utils/jsonRepair.ts: no issues found. Reviewed src/utils/tests/jsonRepair.spec.ts: no issues found. Reviewed src/core/assistant-message/tests/AssistantMessageParser.spec.ts: no issues found. Reviewed src/core/tools/useMcpToolTool.ts: no issues found. Reviewed src/core/tools/kilocode.ts: no issues found. Reviewed src/core/tools/readFileTool.ts: no issues found. Reviewed src/services/search-files/types.ts: no issues found. Reviewed src/shared/tools.ts: no issues found.

Skipped files

• CHANGELOG.md: Skipped file pattern

⬇️ Low Priority Suggestions (1)

src/core/assistant-message/presentAssistantMessage.ts (1 suggestion)

Location: src/core/assistant-message/presentAssistantMessage.ts (Lines 962-975)

🟠 API Contract / Tool-Result Pairing

Issue: This block runs only when toolResultPushed is already true — i.e. the tool handler has already pushed a tool_result for block.toolUseId. Unless pushToolResult_withToolUseId_kilocode (defined ~line 320) merges content into the existing tool_result for that id, this pushes a second tool_result block with the same tool_use_id into the conversation history. Anthropic (and most OpenAI-compatible providers) enforce 1:1 tool_use/tool_result pairing — a duplicate tool_use_id in tool_result blocks can cause the next API request to be rejected with a 400, killing the task loop. Since this fires on every repaired tool call (the exact scenario this PR enables), please verify the helper's behavior.

Fix: Append the repair note to the already-pushed tool_result for block.toolUseId instead of pushing a separate tool_result block.

Impact: Guarantees exactly one tool_result per tool_use, keeping the assistant/user message pairing valid across all providers.

-  				if (block.repaired && !block.partial && toolResultPushed && !cline.didRejectTool) {
-  					try {
-  						const executedArguments =
-  							block.name === "use_mcp_tool"
-  								? String(block.params.arguments ?? "{}")
-  								: JSON.stringify(block.params)
-  						pushToolResult_withToolUseId_kilocode({
-  							type: "text",
-  							text: formatArgumentRepairNote(executedArguments),
-  						})
-  					} catch (error) {
-  						console.error("[presentAssistantMessage] Failed to append argument repair note:", error)
-  					}
-  				}
+  				if (block.repaired && !block.partial && toolResultPushed && !cline.didRejectTool) {
+  					try {
+  						const executedArguments =
+  							block.name === "use_mcp_tool"
+  								? String(block.params.arguments ?? "{}")
+  								: JSON.stringify(block.params)
+  						const note = formatArgumentRepairNote(executedArguments)
+  						const buffer = options.resultBuffer ?? cline.userMessageContent
+  						const existing = [...buffer]
+  							.reverse()
+  							.find(
+  								(m) =>
+  									(m as { type?: string }).type === "tool_result" &&
+  									(m as { tool_use_id?: string }).tool_use_id === block.toolUseId,
+  							) as { content: unknown } | undefined
+  						if (existing && Array.isArray(existing.content)) {
+  							;(existing.content as Array<{ type: "text"; text: string }>).push({ type: "text", text: note })
+  						}
+  					} catch (error) {
+  						console.error("[presentAssistantMessage] Failed to append argument repair note:", error)
+  					}
+  				}