Configuration

OwlSight loads configuration from three sources, with later sources overriding earlier ones:

.owlsight/config.json  →  Environment variables  →  CLI arguments

---

Config File

The config file lives at .owlsight/config.json in your repository root. Create it with owlsight init or write it manually.

{
  "llm": {
    "baseUrl": "https://api.openai.com/v1",
    "model": "gpt-4o",
    "maxTokens": 4096,
    "temperature": 0.2,
    "maxToolRoundtrips": 15
  },
  "review": {
    "minSeverity": "Info",
    "excludePatterns": [
      "**/bin/**",
      "**/obj/**",
      "**/*.generated.cs",
      "**/node_modules/**"
    ],
    "maxFilesPerBatch": 10
  }
}

---

Config Reference

LLM Settings

Field	Type	Default	Description
llm.baseUrl	string	https://api.openai.com/v1	OpenAI-compatible API base URL
llm.apiKey	string	—	API key (prefer env var or CLI arg)
llm.model	string	gpt-4o	Model name
llm.maxTokens	int	4096	Maximum output tokens per LLM call
llm.temperature	float	0.2	Sampling temperature (lower = more deterministic)
llm.maxToolRoundtrips	int	15	Maximum tool-calling iterations per batch

Review Settings

Field	Type	Default	Description
review.minSeverity	string	Info	Minimum severity to report: Critical, Warning, Info, Nitpick
review.excludePatterns	string[]	[]	Glob patterns for files to skip
review.maxFilesPerBatch	int	10	Maximum files sent to the LLM per batch

---

Environment Variables

Variable	Maps to
OWLSIGHT_API_KEY	llm.apiKey
OWLSIGHT_BASE_URL	llm.baseUrl
OWLSIGHT_MODEL	llm.model

Environment variables override the config file but are overridden by CLI arguments.

export OWLSIGHT_API_KEY=sk-...
export OWLSIGHT_MODEL=gpt-4o-mini
owlsight review --base main

---

CLI Arguments

CLI arguments have the highest priority. See the review command reference for all options.

owlsight review --base main \
  --api-key sk-... \
  --base-url http://localhost:11434/v1 \
  --model llama3 \
  --min-severity Warning \
  --max-files-per-batch 5

---

Exclude Patterns

Use glob patterns to skip files from review. Patterns are matched against the relative file path.

{
  "review": {
    "excludePatterns": [
      "**/bin/**",
      "**/obj/**",
      "**/node_modules/**",
      "**/*.generated.cs",
      "**/*.Designer.cs",
      "**/Migrations/**",
      "**/*.min.js",
      "**/*.lock"
    ]
  }
}

---

Batching

When many files change, OwlSight groups them into batches to manage the LLM's context window. Each batch is reviewed independently, and findings are aggregated at the end.

The maxFilesPerBatch setting controls batch size. Lower values mean smaller context per call (better for models with limited context windows), while higher values provide more cross-file context.

Model	Recommended maxFilesPerBatch
GPT-4o (128k context)	10–15
GPT-4o-mini	8–10
Llama 3 (8k)	3–5
Local models	3–5

---

Precedence Example

Given this config file:

{ "llm": { "model": "gpt-4o", "baseUrl": "https://api.openai.com/v1" } }

And this environment:

export OWLSIGHT_MODEL=gpt-4o-mini

And this command:

owlsight review --base main --model llama3

The effective model is llama3 (CLI wins over env var, which wins over config file).