Configuration¶

All configuration uses environment variables with the CODE_CONTEXT_ prefix.

Environment Variables¶

Variable	Default	Description
`CODE_CONTEXT_MODEL_ID`	`global.anthropic.claude-opus-4-6-v1`	Bedrock model ID
`CODE_CONTEXT_REGION`	`us-east-1`	AWS region
`CODE_CONTEXT_TEMPERATURE`	`1.0`	Model temperature (must be 1.0 for thinking)
`CODE_CONTEXT_LSP_SERVERS`	See below	LSP server registry (JSON)
`CODE_CONTEXT_AGENT_MAX_TURNS`	`1000`	Max agent turns
`CODE_CONTEXT_AGENT_MAX_DURATION`	`1200`	Timeout in seconds (default: 20 min)
`CODE_CONTEXT_CONTEXT7_ENABLED`	`true`	Enable context7 MCP server for library documentation lookup
`CODE_CONTEXT_OTEL_DISABLED`	`true`	Disable OpenTelemetry tracing (avoids context detachment errors)
`CODE_CONTEXT_FULL_MAX_DURATION`	`3600`	Max duration for `--full` mode in seconds (300-14400)
`CODE_CONTEXT_FULL_MAX_TURNS`	`3000`	Max agent turns for `--full` mode (100-10000)
`CODE_CONTEXT_LSP_TIMEOUT`	`30`	LSP operation timeout in seconds (5-300)
`CODE_CONTEXT_LSP_STARTUP_TIMEOUT`	`30`	Max seconds to wait for LSP server init (5-120)
`CODE_CONTEXT_LSP_MAX_FILES`	`5000`	Max files before LSP analysis is skipped (100-50000)

LSP Server Registry¶

The CODE_CONTEXT_LSP_SERVERS variable accepts a JSON object mapping language keys to ordered lists of server commands. Each list is a fallback chain -- if the first server fails, the next is tried automatically.

{
  "py": ["ty server", "pyright-langserver --stdio"],
  "ts": ["typescript-language-server --stdio"],
  "rust": ["rust-analyzer"],
  "go": ["gopls serve"],
  "java": ["jdtls"]
}

For example, Python analysis first attempts ty server. If that fails to start (e.g., not installed), it falls back to pyright-langserver --stdio. If all servers in the chain fail, the agent compensates with other signal sources (see Tenet 6: Fail loud, fill gaps).

Tip

To add a new language, extend the JSON with its key and an ordered list of server commands. The language key is matched against file extensions detected during analysis.

Configuration via pydantic-settings¶

Configuration is implemented using pydantic-settings, which provides:

Environment variable parsing with the CODE_CONTEXT_ prefix
Type validation and coercion
Default values
JSON parsing for complex types (like LSP_SERVERS)

See the config module in the API reference for the full Settings model.

context7 Integration¶

When CODE_CONTEXT_CONTEXT7_ENABLED is true (the default), the analysis agent gains access to the context7 MCP server for looking up library documentation during analysis. This requires npx to be available on the system.

The context7 tools are prefixed with context7_ in the agent's tool namespace:

context7_resolve-library-id -- resolve a library name to a context7 ID
context7_query-docs -- query documentation for a resolved library

To disable: export CODE_CONTEXT_CONTEXT7_ENABLED=false

Full Mode Configuration¶

When --full is passed to the analyze command, the agent overrides several settings for exhaustive analysis:

Setting	Standard Default	Full Mode Override
`agent_max_duration`	1200 (20 min)	`full_max_duration` (default: 3600 / 60 min)
`agent_max_turns`	1000	`full_max_turns` (default: 3000)
`lsp_max_files`	5000	50,000

These overrides are applied via Settings.model_copy() at runtime. The original settings are not modified.

See Full Mode for complete details on exhaustive analysis.