[BUG] Terminal Fallback Causes Duplicate Tool Results and Doom Loop in Unattended Mode

[nabilfreeman]

Terminal Fallback Causes Duplicate Tool Results and "Doom Loop"

Problem (one or two sentences)

When an external terminal (PuTTY) fails during long-running commands and falls back to the integrated terminal, duplicate tool results are generated for the same tool_use_id, causing an API protocol violation that puts Roo into an unrecoverable "doom loop" that breaks unattended access.

Context (who is affected and when)

This affects users running unattended/autonomous workflows with external terminal configurations (e.g., PuTTY) when the external terminal host fails during long-running operations (5+ hours). The fallback mechanism to integrated terminal triggers a race condition that creates duplicate tool_result blocks, violating the Anthropic API protocol.

Reproduction Steps

1. Environment/Setup:

   • OS: macOS (but affects any OS using external terminals)
   • Extension Version: v3.38.2
   • Provider: Amazon Bedrock (anthropic.claude-sonnet-4-5-20250929-v1:0)
   • Configuration: External terminal enabled (PuTTY/pty host)
   • Mode: Orchestrator with long-running test commands

2. Exact Actions:

   • Start an orchestrator task that delegates to code mode
   • Code mode executes a long-running command via execute_command tool (e.g., npm run test:eval scenario-name)
   • Command runs for ~5 hours using external terminal
   • External terminal (PuTTY host) fails/disconnects
   • System automatically falls back to integrated terminal
   • User approval arrives during/after the fallback transition

3. What Was Observed:

   • Assistant executed command with tool_use_id: tooluse_-SZO09niSuehB1YFk4pX3Q
   • System responded: "Command is still running in terminal"
   • Assistant executed another command: ps aux | grep... with tool_use_id: tooluse_QZ-pU8v2QKO8L8fHoJRI2g
   • System responded: "No test processes found"
   • User approval arrived with ANOTHER tool_result for the same tooluse_QZ-pU8v2QKO8L8fHoJRI2g
   • This created 2 tool_result blocks for 1 tool_use block
   • API rejected with error: "The number of toolResult blocks at messages.31.content exceeds the number of toolUse blocks of previous turn"
   • Roo entered unrecoverable state, unable to proceed with the task

Expected Result

• Terminal fallback should occur transparently without generating duplicate tool results
• System should validate tool result count before sending to API
• If tool result mismatch occurs, system should auto-recover by reconciling the conversation history
• Unattended access should remain functional even after terminal failures

Actual Result

• Terminal fallback generated duplicate tool_result blocks for the same tool_use_id
• API protocol violation occurred with error message
• Roo entered "doom loop" - unable to recover or continue the task
• Unattended access broke completely, requiring manual intervention
• Agent was soft-locked overnight without ability to self-recover

Variations Tried

N/A - This is an edge case that occurred during production use. The issue is reproducible by simulating external terminal failure during long-running commands.

App Version

v3.38.2

API Provider

Amazon Bedrock

Model Used

anthropic.claude-sonnet-4-5-20250929-v1:0

Roo Code Task Links

N/A - Task was running in unattended mode when failure occurred

Relevant Logs or Errors

Full Error Object

{
  "error": {
    "timestamp": "2026-01-05T05:51:10.811Z",
    "version": "3.38.2",
    "provider": "bedrock",
    "model": "anthropic.claude-sonnet-4-5-20250929-v1:0",
    "details": "Unknown Error: The number of toolResult blocks at messages.31.content exceeds the number of toolUse blocks of previous turn."
  },
  "history": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "<task>\n[REDACTED: Large task description with implementation details]\n</task>"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED: Project file structure]\n</environment_details>"
        }
      ],
      "ts": 1767572765406
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED: Assistant reasoning about orchestrator task delegation]",
          "summary": []
        },
        {
          "type": "text",
          "text": "I'll coordinate this migration task by first getting an architectural plan, then executing it incrementally with testing between steps."
        },
        {
          "type": "tool_use",
          "id": "tooluse_PnIzU0bpR_a6x_DZ2Ecxyw",
          "name": "new_task",
          "input": {
            "mode": "architect",
            "message": "[REDACTED: Task delegation message]",
            "todos": "[ ] Read files\n[ ] Review references\n[ ] Analyze implementation\n[ ] Create migration plan\n[ ] Signal completion"
          }
        }
      ],
      "ts": 1767572800206
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_PnIzU0bpR_a6x_DZ2Ecxyw",
          "content": "Subtask completed.\n\nResult:\n[REDACTED: Detailed architectural plan]"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED: Environment details]\n</environment_details>"
        }
      ],
      "ts": 1767572975593
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED: Planning next steps]",
          "summary": []
        },
        {
          "type": "tool_use",
          "id": "tooluse_ny96idEmSkuNX0DUebTWBQ",
          "name": "update_todo_list",
          "input": {
            "todos": "[x] Task 1\n[-] Task 2\n[ ] Task 3\n[ ] Task 4\n[ ] Task 5\n[ ] Task 6\n[ ] Task 7\n[ ] Task 8"
          }
        }
      ],
      "ts": 1767572985455
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_ny96idEmSkuNX0DUebTWBQ",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767572985656
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED: Reasoning about executing plan]",
          "summary": []
        },
        {
          "type": "text",
          "text": "Now executing Phase 1 of the migration."
        },
        {
          "type": "tool_use",
          "id": "tooluse_feOGGK5MQneQeNyWn3LiGg",
          "name": "new_task",
          "input": {
            "mode": "code",
            "message": "[REDACTED: Code modification instructions]",
            "todos": "[ ] Read file\n[ ] Update code\n[ ] Complete task"
          }
        }
      ],
      "ts": 1767573000981
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_feOGGK5MQneQeNyWn3LiGg",
          "content": "Subtask completed.\n\nResult:\n[REDACTED: File modification results]"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573054659
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "tool_use",
          "id": "tooluse_FORPY8fsS7ucI11KTUBp9g",
          "name": "update_todo_list",
          "input": {
            "todos": "[x] Task 1\n[x] Task 2\n[-] Task 3\n[ ] Task 4\n[ ] Task 5\n[ ] Task 6\n[ ] Task 7\n[ ] Task 8"
          }
        }
      ],
      "ts": 1767573062147
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_FORPY8fsS7ucI11KTUBp9g",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573062327
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "text",
          "text": "Proceeding to next phase."
        },
        {
          "type": "tool_use",
          "id": "tooluse_EyDV-CT2QFqdavs4HzHc-A",
          "name": "new_task",
          "input": {
            "message": "[REDACTED: Code modification instructions]",
            "mode": "code",
            "todos": "[ ] Read file\n[ ] Update code\n[ ] Complete"
          }
        }
      ],
      "ts": 1767573081694
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_EyDV-CT2QFqdavs4HzHc-A",
          "content": "Subtask completed.\n\nResult:\n[REDACTED: Modification results]"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573135034
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "tool_use",
          "id": "tooluse_AY755nWbR--el0tqPJlOSg",
          "name": "update_todo_list",
          "input": {
            "todos": "[x] Task 1\n[x] Task 2\n[x] Task 3\n[-] Task 4\n[ ] Task 5\n[ ] Task 6\n[ ] Task 7\n[ ] Task 8"
          }
        }
      ],
      "ts": 1767573141785
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_AY755nWbR--el0tqPJlOSg",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573141919
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "text",
          "text": "Proceeding to next phase."
        },
        {
          "type": "tool_use",
          "id": "tooluse_7rdkHyHnSjKfkhb-J7gEFg",
          "name": "new_task",
          "input": {
            "message": "[REDACTED: Code modification instructions]",
            "mode": "code",
            "todos": "[ ] Read file\n[ ] Update code\n[ ] Complete"
          }
        }
      ],
      "ts": 1767573156797
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_7rdkHyHnSjKfkhb-J7gEFg",
          "content": "Subtask completed.\n\nResult:\n[REDACTED: Modification results]"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573205740
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "tool_use",
          "id": "tooluse_akDNjcnQSvK7qLZzNsxuWg",
          "name": "update_todo_list",
          "input": {
            "todos": "[x] Task 1\n[x] Task 2\n[x] Task 3\n[x] Task 4\n[-] Task 5\n[ ] Task 6\n[ ] Task 7\n[ ] Task 8"
          }
        }
      ],
      "ts": 1767573213080
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_akDNjcnQSvK7qLZzNsxuWg",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573213239
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "text",
          "text": "Proceeding to final code update phase."
        },
        {
          "type": "tool_use",
          "id": "tooluse_-qHP2tTSTWSNL19dpMtpTQ",
          "name": "new_task",
          "input": {
            "message": "[REDACTED: Code modification instructions]",
            "mode": "code",
            "todos": "[ ] Read file\n[ ] Update code\n[ ] Complete"
          }
        }
      ],
      "ts": 1767573228017
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_-qHP2tTSTWSNL19dpMtpTQ",
          "content": "Subtask completed.\n\nResult:\n[REDACTED: Modification results]"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573286474
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "tool_use",
          "id": "tooluse_bcGI0mPoTNGCgrQ9NFwWsQ",
          "name": "update_todo_list",
          "input": {
            "todos": "[x] Task 1\n[x] Task 2\n[x] Task 3\n[x] Task 4\n[x] Task 5\n[-] Task 6\n[ ] Task 7\n[ ] Task 8"
          }
        }
      ],
      "ts": 1767573293932
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_bcGI0mPoTNGCgrQ9NFwWsQ",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573294136
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "text",
          "text": "Phase 6: Running the linter to verify all changes are syntactically correct."
        },
        {
          "type": "tool_use",
          "id": "tooluse_Ok7Bkcl8RyCT2lEfRtkygg",
          "name": "execute_command",
          "input": {
            "command": "npm run lint",
            "cwd": "/path/to/project"
          }
        }
      ],
      "ts": 1767573299656
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_Ok7Bkcl8RyCT2lEfRtkygg",
          "content": "Command executed in terminal within working directory. Command execution was not successful, inspect the cause and adjust as needed.\nExit code: 2\nOutput:\n\n[REDACTED: Linter output with project-specific warnings and TypeScript errors in other files]"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573333822
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "The linter ran and found TypeScript errors in OTHER files not related to our changes. Our modified files have no errors.",
          "summary": []
        },
        {
          "type": "text",
          "text": "The linter shows no errors in our modified files. Now let's test the scenario."
        },
        {
          "type": "tool_use",
          "id": "tooluse_i3O-8SusSKuQWllVCHcdqw",
          "name": "update_todo_list",
          "input": {
            "todos": "[x] Task 1\n[x] Task 2\n[x] Task 3\n[x] Task 4\n[x] Task 5\n[x] Task 6\n[-] Task 7\n[ ] Task 8"
          }
        }
      ],
      "ts": 1767573342204
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_i3O-8SusSKuQWllVCHcdqw",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767573342409
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "[REDACTED]",
          "summary": []
        },
        {
          "type": "text",
          "text": "Now running the test to verify the migration works correctly."
        },
        {
          "type": "tool_use",
          "id": "tooluse_-SZO09niSuehB1YFk4pX3Q",
          "name": "execute_command",
          "input": {
            "command": "npm run test:eval scenario-name",
            "cwd": "/path/to/project"
          }
        }
      ],
      "ts": 1767573348006
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_-SZO09niSuehB1YFk4pX3Q",
          "content": "Command is still running in terminal from '/path/to/project'.\n\n\nYou will be updated on the terminal status and new output in the future."
        },
        {
          "type": "text",
          "text": "<environment_details>\n# Actively Running Terminals\n## Terminal 1 (Active)\n### Working Directory: `/path/to/project`\n### Original command: `npm run test:eval scenario-name`\n\n[REDACTED: Other environment details]\n</environment_details>"
        }
      ],
      "ts": 1767592162165
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "The test command is still running. I should wait for it to complete.",
          "summary": []
        },
        {
          "type": "text",
          "text": "Let me check the status of the running test:"
        },
        {
          "type": "tool_use",
          "id": "tooluse_cXZ3_VlASVOFI3iG1Ypdow",
          "name": "execute_command",
          "input": {
            "command": "echo \"Waiting for test to complete...\"",
            "cwd": "/path/to/project"
          }
        }
      ],
      "ts": 1767592168782
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_cXZ3_VlASVOFI3iG1Ypdow",
          "content": "Command executed in terminal within working directory. Exit code: 0\nOutput:\nWaiting for test to complete..."
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767592174398
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "The test is still running. Let me check if there's any test process.",
          "summary": []
        },
        {
          "type": "tool_use",
          "id": "tooluse_QZ-pU8v2QKO8L8fHoJRI2g",
          "name": "execute_command",
          "input": {
            "command": "ps aux | grep -i \"npm.*test:eval\" | grep -v grep || echo \"No test processes found\"",
            "cwd": "/path/to/project"
          }
        }
      ],
      "ts": 1767592178982
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_QZ-pU8v2QKO8L8fHoJRI2g",
          "content": "Command executed in terminal within working directory. Exit code: 0\nOutput:\nNo test processes found"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767592179756
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tooluse_QZ-pU8v2QKO8L8fHoJRI2g",
          "content": "{\"status\":\"approved\",\"message\":\"The user approved this operation and provided the following context\",\"feedback\":\"Run again\"}"
        },
        {
          "type": "text",
          "text": "<environment_details>\n[REDACTED]\n</environment_details>"
        }
      ],
      "ts": 1767592234718
    }
  ]
}

Key observation: The same tool_use_id (tooluse_QZ-pU8v2QKO8L8fHoJRI2g) received TWO tool_result blocks - one from the command execution, and another from user approval. This violates the API protocol requiring exactly one tool_result per tool_use.

---

Root Cause Analysis

The Problem

A race condition between terminal fallback and user approval generates duplicate tool_result blocks for the same tool_use_id, violating the Anthropic API protocol and causing an unrecoverable state.

Specific Mechanism

1. Terminal Fallback Trigger

   • When external terminal (PuTTY) fails during long-running commands, a ShellIntegrationError is thrown
   • System automatically retries with fallback to integrated terminal

2. Race Condition

   • Long-running commands create pending user approval requests via ask("command_output")
   • When terminal fails during this wait, user approval can arrive DURING fallback retry

3. Duplicate Generation

   • The pushToolResult() callback adds tool_result blocks to conversation history
   • First call: Original execution pushes tool_result for tool_use_id
   • Terminal fails: Fallback is triggered
   • User approval arrives: Contains tool_result from pending ask
   • Second call: Fallback execution pushes ANOTHER tool_result for SAME tool_use_id
   • Result: 2 tool_result blocks for 1 tool_use block → API error

Why Existing Safeguards Failed

Existing validateAndFixToolResultIds() function has de-duplication logic embedded within position-matching flow, but doesn't explicitly filter duplicate tool_result blocks with the same tool_use_id before other processing.

---

Recommended Fix

Surgical Single-File Solution

Approach: Add explicit de-duplication step to validateAndFixToolResultIds() that filters duplicate tool_result blocks before any other processing.

File Modified: Only src/core/task/validateToolResultIds.ts

Implementation: Insert new logic after extracting tool_result blocks (after line 83):

// Find tool_result blocks in the user message
const toolResults = userMessage.content.filter(
	(block): block is Anthropic.ToolResultBlockParam => block.type === "tool_result",
)

// === NEW: Explicit deduplication step ===
// Build a set to track which tool_use_ids we've seen in tool_results
const seenToolResultIds = new Set<string>()

// Deduplicate tool_results: keep only first occurrence of each tool_use_id
// This handles the fallback scenario where duplicate tool_results are generated
const deduplicatedContent = userMessage.content.filter((block) => {
	if (block.type !== "tool_result") {
		return true // Keep all non-tool_result blocks
	}
	
	// For tool_result blocks, only keep if we haven't seen this tool_use_id yet
	if (seenToolResultIds.has(block.tool_use_id)) {
		return false // Filter out duplicate
	}
	
	seenToolResultIds.add(block.tool_use_id)
	return true // Keep first occurrence
})

// Update the working content to use deduplicated version
userMessage = {
	...userMessage,
	content: deduplicatedContent,
}

// Re-extract toolResults from deduplicated content for subsequent processing
const toolResults = deduplicatedContent.filter(
	(block): block is Anthropic.ToolResultBlockParam => block.type === "tool_result",
)
// === END NEW CODE ===

// Rest of existing logic continues unchanged...

How This Fixes the Bug

Bug Scenario:

Assistant: [tool_use(id="tooluse_QZ-pU8v2QKO8L8fHoJRI2g")]
User: [tool_result(id="tooluse_QZ-pU8v2QKO8L8fHoJRI2g"), 
      tool_result(id="tooluse_QZ-pU8v2QKO8L8fHoJRI2g")]  ← DUPLICATE!

With This Fix:

1. Deduplication step runs first
2. First tool_result with tooluse_QZ-pU8v2QKO8L8fHoJRI2g: Not seen → keep → add to Set
3. Second tool_result with same ID: Already in Set → filter out
4. Result: Only 1 tool_result remains for the 1 tool_use block
5. Then existing position-matching/ID-correction logic continues normally

Advantages

1. Surgical: Only modifies validation function (~20 lines), no changes to ExecuteCommandTool.ts or Task class
2. Minimal Impact: No architectural changes to codebase
3. Explicit: De-duplication is clear, separate step that's easy to understand
4. Safe: Existing tests verify this behavior; fix makes it more explicit
5. Performance: Single-pass O(n) filtering before other validation

Impact

• Before: Terminal failures → duplicate tool_results → API errors → unrecoverable doom loop → broken unattended access
• After: Duplicates filtered in validation layer → graceful API submission → resilient unattended access

---

Additional Notes

Why Integrated Terminal Isn't a Full Solution

While switching to integrated terminal was suggested as a workaround, it doesn't support custom shells, which breaks workflows requiring:

• direnv (environment variable management)
• Custom shell configurations
• Shell-specific tooling

Therefore, fixing the external terminal fallback mechanism is necessary for these use cases.

Testing Recommendations

1. Add unit test in validateToolResultIds.spec.ts for exact duplicate scenario
2. Integration test simulating PuTTY failure during long-running commands
3. Verify only one tool_result per tool_use in fallback scenarios
4. Test unattended mode resilience to terminal failures

---

Historical Context

When Was This Bug Introduced?

Type: Latent bug exposed by architectural change

Timeline:

• Terminal Fallback Mechanism: Present in codebase for extended period (months/years)
• Became Problematic: December 18, 2025 (v3.36.13) when native tool protocol became default
• Error First Surfaced: January 5, 2026 during unattended overnight operation

Why Did This Surface Now?

Root Cause: Switch from XML to Native Tool Protocol as Default

1. XML Tool Protocol (Pre-December 2025)

   • Used text-based tool results
   • No strict 1:1 tool_use to tool_result requirement
   • Duplicates were tolerated or ignored by the API
   • Race condition existed but was invisible

2. Native Tool Protocol (Post-December 2025)

   • Enforces strict protocol: exactly one tool_result per tool_use_id
   • Anthropic API rejects messages with mismatched counts
   • Duplicates cause immediate API errors
   • Race condition became visible and fatal

Related Issues During Native Tool Rollout

Multiple PRs attempted to fix tool_result issues during November-December 2025 native tool protocol rollout:

• v3.33.0 (Nov 18, 2025) - PR fix: prevent duplicate tool_result blocks in native tool protocol #9248: "Prevent duplicate tool_result blocks in native tool protocol"
• v3.31.3 (Nov 13, 2025) - PR fix: resolve native tool protocol race condition causing 400 errors #9363: "Resolve native tool protocol race condition causing 400 errors"
• v3.36.3 (Dec 9, 2025) - PR fix: validate and fix tool_result IDs before API requests #9952: "Validate and fix tool_result IDs before API requests" (introduced validateAndFixToolResultIds())
• v3.36.5 (Dec 10, 2025) - PR fix: add missing tool_result blocks to prevent API errors #10015: "Add missing tool_result blocks to prevent API errors"
• v3.36.5 (Dec 10, 2025) - PR fix: filter orphaned tool_results when more results than tool_uses #10027: "Filter orphaned tool_results when more results than tool_uses"
• v3.36.8 (Dec 16, 2025) - PR feat(anthropic): enable native tools by default and add telemetry tracking #10021: "Enable native tools by default for Anthropic"
• v3.36.13 (Dec 18, 2025) - PR Revert "Revert "feat: change defaultToolProtocol default from xml to native"" #10186: "Change default tool protocol from XML to native" ← THE TRIGGER

This pattern suggests the native tool protocol rollout exposed multiple edge cases in tool result handling, including this terminal fallback race condition.

Why Existing Fixes Didn't Catch This

The validation function validateAndFixToolResultIds() was introduced in v3.36.3 (Dec 9, 2025) and has de-duplication logic embedded within its position-matching flow (lines 146-183). However:

1. It wasn't an explicit first-pass filter
2. Processing is reactive rather than preventative
3. The terminal fallback race condition wasn't specifically tested
4. Long-running commands (5+ hours) with external terminal failures are rare edge cases

Implications

1. Not a New Code Bug: The terminal fallback logic is legacy code that worked under XML protocol
2. Architectural Exposure: Native tool protocol's strict requirements exposed latent race conditions
3. Similar Issues Possible: Other tool execution paths may have similar race conditions not yet discovered
4. Fix Importance: The proposed surgical fix prevents duplicates at validation layer, protecting against this and similar edge cases

This context explains why unattended access broke suddenly in January 2026 despite the codebase being stable for months.

I've analyzed this issue and here's my proposed implementation plan:

Root Cause Analysis

The bug occurs because validateAndFixToolResultIds() has an early return condition (line 100) that checks missingToolUseIds.length === 0 && !hasInvalidIds. When duplicate tool_results have the same valid tool_use_id:

1. existingToolResultIds is a Set that deduplicates - it only contains one entry
2. missingToolUseIds is empty (the ID exists in the Set)
3. hasInvalidIds is false (the ID is valid)
4. Early return triggers, duplicates pass through unfiltered
5. API receives 2 tool_results for 1 tool_use - protocol violation

Proposed Implementation

File Modified: Only src/core/task/validateToolResultIds.ts

Changes:

1. After extracting toolResults (line 83), check for duplicate tool_use_ids
2. Add explicit early deduplication step that keeps only the first occurrence of each tool_use_id
3. Track which tool_use_ids have been seen to ensure only one tool_result per ID

// After line 83, add:
// Deduplicate tool_results: if multiple have same tool_use_id, keep only the first
const seenToolResultIds = new Set<string>()
const deduplicatedToolResults = toolResults.filter((result) => {
    if (seenToolResultIds.has(result.tool_use_id)) {
        return false // Filter out duplicate
    }
    seenToolResultIds.add(result.tool_use_id)
    return true
})

// Also update userMessage.content to filter out duplicate tool_results
// This ensures the early return path also works correctly with deduped content

Test Added: A new test case in validateToolResultIds.spec.ts for the exact scenario:

• Two tool_result blocks with the same valid tool_use_id
• Verify only one is kept

Benefits

1. Surgical fix: Only modifies validation function (~20 lines)
2. Explicit: Clear deduplication step before any other processing
3. Safe: Existing tests continue to pass
4. Handles edge case: Specifically addresses the terminal fallback race condition

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[nabilfreeman]

Yes please @roomote, I agree with the fix. Please include more inline comments to explain the significance of the logic clearly for future developers (like my proposal). This is an important part of the tool use cycle and should remain understandable and maintainable. Same for the tests you write.

Thanks for the approval @nabilfreeman! I'll proceed with the implementation, adding comprehensive inline comments to explain the deduplication logic and its significance for future developers. I'll also ensure the tests are well-documented. Opening a PR shortly.

Opened PR #10466 with my attempt at addressing this issue. The fix adds explicit deduplication for duplicate tool_result blocks in validateAndFixToolResultIds(), with comprehensive inline comments explaining the logic as requested. Feedback welcome!

#10466

[daniel-lxs]

Hey @nabilfreeman thank you for reporting this and working with Roomote to create a PR, I'm currently reviewing it!