The bug in the edit_existing_file tool - the tool itself is writing only 30-50 lines when it should be writing 600+

[olegperepelkin-maker]

Before submitting your bug report

• I've tried finding an answer on the Continue docs site
• I'm not able to find an open issue that reports the same bug
• I've seen the troubleshooting guide on the Continue Docs

Relevant environment info

- OS:Win11
- Continue version: Continue/2.0.0
- IDE version: Visual Studio Code/1.133.0
- Model: qwen/qwen3.8-27b via LM Studio (local)
- config:
  # ============================================================
# IoT Firmware Expert Config — local-only
#   • Continue  → LM Studio (localhost)   [this file]
#   • Cloud LLM → GitHub Copilot (rarely) [lives in the editor,
#                 NOT in this file]
# Rule: `model:` must exactly match LM Studio's /v1/models ID.
# ============================================================
name: IoT Firmware Expert Config
version: 1.1.0
schema: v1

models:
  # Primary chat / edit model — the exact ID from `curl http://ai-host:1234/v1/models`
  - name: Qwen 3.8 27B
    provider: lmstudio
    model: qwen/qwen3.8-27b          # exact ID from /v1/models
    apiBase: http://ai-host:1234/v1
    apiKey: lm-studio                # LM Studio doesn't check this; Continue requires non-empty
    roles:
      - chat
      - edit
      - apply
    systemPrompt: |
      You are an elite IoT/embedded firmware developer with deep expertise in:
      - ESP32/STM32 microcontroller development
      - PlatformIO build system and configuration
      - Bare-metal C/C++ firmware architecture
      - RTOS, hardware peripherals, and low-level protocols (I2C, SPI, UART)
      - Python embedded scripting for IoT
      - Memory optimization and performance tuning

      Guidelines:
      - Generate production-ready, optimized code
      - Always consider memory constraints and power efficiency
      - Provide minimal, focused explanations (preserve context)
      - Use pragmatic solutions over theoretical approaches
      - Include error handling and resource cleanup
    contextLength: 75776             # keep in sync with LM Studio's context setting
    requestOptions:
      temperature: 0.3
      top_p: 0.9
      top_k: 40
      max_tokens: 4096              # raised from 2048 so whole-file code gen isn't cut off

  # Embeddings (code search / codebase indexing)
  # NOTE: NOT in the current /v1/models list. Download `nomic-embed-text-v1.5`
  # in LM Studio, then uncomment this block (and verify the ID with /v1/models).
  # - name: Nomic Embed v1.5
  #   provider: lmstudio
  #   model: nomic-embed-text-v1.5
  #

Description

The evidence from editOutcome.jsonl shows the tool itself is writing only 30-50 lines when it should be writing 600+ — that's happening at the Continue extension level, before it ever touches the VS Code editor.

The editOutcome.jsonl log confirms the bug pattern clearly:

File Previous Lines New Lines Change
state.cpp 658 34 -624
state.cpp 34 5 -29
state.cpp 5 13 +8
ui_server.cpp 690 54 -636
config.h 37 7 -30
The edit_existing_file tool is catastrophically truncating — it's keeping only the first ~30-50 lines and discarding the rest. This is a bug in the Continue agent's edit tool implementation.

Summary of findings:
Continue version: 2.0.0 (from userAgent: "Visual Studio Code/1.133.0 (Continue/2.0.0)")
Model: qwen/qwen3.8-27b via LM Studio (local)

The bug is in the edit_existing_file tool — it's not a model issue, it's a tool implementation issue. The model is correctly generating the full file content, but the tool is only writing the first chunk.
No Continue settings can fix this — the config files (config.yaml, config.ts, .continuerc.json) only control model selection, indexing, and API keys. There's no tool behavior configuration.

Reporting this bug
editOutcome.jsonl.txt
with the editOutcome.jsonl data (renamed to *.txt file for upload) as evidence. The key detail: previousCodeLines: 658, newCodeLines: 34 — the tool is dropping 95% of the file content.

To reproduce

No response

Log output