Open-source AI Tools

Cline Shipped Two Releases in One Day: One Fixes a Silent File-Read Cutoff, One Lets Unattended Runs Approve Themselves

November 22, 20261 views阅读中文原文
Cline Shipped Two Releases in One Day: One Fixes a Silent File-Read Cutoff, One Lets Unattended Runs Approve Themselves

The short version

On November 21, 2026, the open-source coding agent Cline shipped two versions on the same day: extension 4.1.27 (06:05 UTC) and CLI 3.0.75 (05:30 UTC). Two changes matter most:

  1. The extension fixed a hidden second truncation. The file-reading tool was already allowed to return up to 2,000 lines, roughly 48,000 characters. But before that result reached the model, a later stage cut it down to an 8,000-character budget — and it cut from the middle. The agent could see the top of a function and its return statement while missing a branch in between. File reads now use the full file-content allowance.
  2. The CLI now starts one-shot runs in yolo mode by default. Previously, running something like cline "fix this bug" with nobody at the keyboard still defaulted to act mode, which waits for approval. Now it assumes no one is there to approve.

In short: one change means the agent was reading less of your code than it appeared to. The other means the agent approves itself. Both touch the trust boundary, in opposite directions.

This article is compiled from Cline's official GitHub release notes and from reporting by Freedom.Tech, Code×Radar, AIDev.news and other outlets.

How the 8,000-character cutoff happened

This is the least flashy item in the release and the one most likely to affect day-to-day work.

Cline's read_files tool has an explicit limit: 2,000 lines or about 48,000 characters per read. Anything larger gets read in ranges, and the tool tells the agent which lines it received so it can ask for the next range. That part was fine.

The problem sat one step later. When Cline assembled the message for the model provider, it treated a file read like ordinary command output, applied the 8,000-character tool-result budget, and truncated by keeping the head and tail while dropping the middle.

The contributor who fixed it tested a 751-line Markdown file. The read produced 47,384 characters, but only 7,967 reached the provider. In other words, the agent nominally read the whole file and actually saw the beginning and the end.

The practical damage in a code review is specific: the agent may see a function's entry point and its final return, yet miss an if branch in the middle. Its conclusion looks reasonable, but it was reached on incomplete context. Worse, the tool's footer still reports which lines it "saw," so the gap does not announce itself.

The fix routes file reads through the existing file-content allowance so they arrive whole, up to the read tool's own cap. Command output, search results, MCP results and custom tool results keep their own 8,000-character limit, and an explicitly configured tool-result cap still applies.

Yolo by default: convenient, but know the edges

CLI 3.0.75 changes a default: one-shot runs — a prompt passed as an argument, or piped through stdin — now start in yolo mode instead of act mode. The release notes give the reason plainly: nobody is at the keyboard to approve tool calls.

Five exceptions still keep act mode:

  • Auto-approve is turned off (--auto-approve false, or turned off in saved settings)
  • A mode flag is passed explicitly
  • Plan mode is the saved mode
  • The prompt is a /team prompt
  • You need team tools — yolo mode has none, so pass --act

The release also fixes an older problem: --timeout and Ctrl+C previously could not interrupt the first turn of a one-shot run. Now they can. And one quieter security-relevant change: the CLI no longer reads .env or bunfig.toml from the directory it launches in, removing a path where files in a working directory could quietly change behavior.

If you call Cline CLI from scripts or CI, read these together. No approval by default means no human catches a bad step mid-run. Being interruptible at least gives you a timeout backstop.

What else changed

Model catalog refresh. 94 providers were merged in from models.dev, and Lilac and Tempr were dropped. The recommended list adds Muse Ember; the free list drops Reflection Beam, Laguna S 2.1 and Solar Mini 4. Default models changed for 67 other providers, mostly toward Kimi K3.1 Pro, Gemini 4 Neon and DeepSeek V4.1 Flash 1017. ClinePass dropped a batch of models as well.

A cluster of reasoning-effort fixes. Picking "None" is now saved and actually sent (previously GPT-6 on Azure kept its default effort and rejected tool calls). Turning reasoning off now reaches custom models that carry saved overrides. On the native OpenAI provider, the reasoning level is fitted to what each model supports, so max and Extra High are no longer rejected by GPT-5.x and o-series models. Reflection Beam no longer fails every turn with a 400 Missing required parameter error after compaction or a model switch.

Smaller items. LM Studio settings gained an API key field for servers with "Require Authentication" turned on. The MCP tab now shows a "read-only" badge on tools whose server declares them read-only. The account view links to privacy settings, and the extension package includes third-party license notices. On the CLI side, the CLINE_CAPTURE_* debug variables — which wrote full provider request bodies to disk — were removed, and undici was raised to 7.30.1 to fix two process-crash issues.

Desktop 0.0.50 and SDK 0.2.1 also shipped the same day, carrying largely the same changes.

Who should update now

A quick read on who this is a good fit for:

  • Anyone who has Cline review large files. If you have seen the agent declare a function fine while quietly missing its middle, reopen that read after updating and let it fetch the relevant range fresh.
  • Anyone running Cline CLI in scripts or CI. Yolo-by-default is a behavior change, not a bug. Either accept it with --timeout as a backstop, or pass --auto-approve false to get approval back.
  • Anyone on third-party or local models. The catalog refresh and the reasoning-effort fixes mostly benefit this group.

Limitations and things to watch

  • This does not fix "large files still do not fit." The 2,000-line / 48,000-character tool limit is unchanged; big files still need ranges. What changed is that a read which succeeded no longer gets cut again on the way out.
  • Do not accept the yolo default blindly on shared or production machines. No approval means one bad shell command executes. Use --auto-approve false, or at minimum isolate the working directory and credentials.
  • Do not forget --act in flows that need team tools. Yolo mode has no team tools, and a script will simply not get them.
  • Expect input costs to move. File reads are no longer cut to 8,000 characters, so the same read now sends more tokens. Long-file work costs more on input — not because prices rose, but because you were previously saving money on incomplete context.

How to use it: three steps

Start with these next steps; ten minutes covers all three.

  1. Update the extension to 4.1.27 and the CLI to 3.0.75 (npm i -g cline or your usual install path), then restart the editor.
  2. Pick a long file where a past review skipped a section, run a fresh read, and confirm the returned character count matches the file's size.
  3. Audit your scripts: decide explicitly between yolo and act. Add --timeout, and add --act wherever team tools are needed.

Sources

  • Cline official release v4.1.27: https://github.com/cline/cline/releases/tag/v4.1.27
  • Cline official release cli-v3.0.75: https://github.com/cline/cline/releases/tag/cli-v3.0.75
  • Freedom.Tech, "Cline 4.1.27": https://freedom.tech/posts/2026-11-21-cline-4-1-27/
  • Code×Radar, "Cline 4.1.27 fixes a second cutoff that hid the middle of file reads": https://codexradar.dev/changes/cline-4-1-27-file-read-truncation/
  • AIDev.news, "Cline v4.1.27 Adds API Key Field, Expands Model Catalog, and Improves File Handling": https://www.aidev.news/article/cline-v4-1-27-adds-api-key-field-expands-model-catalog-and-improves-f

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...