Skip to main content

Confidence Score

Before Nimbus generates a plan, it retrieves relevant code chunks and shows you a confidence score.

What it looks like​

  confidence  ████████░░  82%
  retrieval   14 relevant chunks found
  ambiguity   low — task is well-scoped

How it's calculated​

The score combines two signals:

Retrieval quality — how many relevant code chunks were found using hybrid BM25 + vector search. The formula is:

confidence = min(95, 50 + chunk_count × 3)

So 14 chunks → 50 + 14×3 = 92%. The maximum score is 95%.

Ambiguity — whether the task description is specific enough:

  • low — task has 9 or more words. Nimbus considers it well-scoped.
  • medium — task has fewer than 9 words. Consider adding more detail.

The bar visualization​

The ████████░░ bar represents the confidence percentage. Each filled block is roughly 10%. A score above 80% generally means Nimbus found enough context to produce a good plan.

Auto-approval threshold​

By default, tasks at or above 92% confidence skip the approval prompt entirely and proceed directly to implementation. You can change this in ~/.nimbus/config.toml:

[local]
auto_approve_confidence = 85  # lower threshold
# auto_approve_confidence = 100  # always prompt

Set to 100 to always be prompted, regardless of score.

What low confidence means​

A score below 70% usually means:

  • The task references code that isn't in the index yet (new files, branches not yet indexed)
  • The task is very high-level ("refactor the whole codebase")
  • The task is in an area with sparse code coverage

You can still approve low-confidence tasks. The plan may be less precise, but Nimbus will still attempt it.