Full parsed configuration. More...
#include <entropic/types/config.h>
Public Attributes | |
| ModelsConfig | models |
| Tiers + router. | |
| RoutingConfig | routing |
| Routing rules. | |
| GenerationConfig | generation |
| Default generation params. | |
| PermissionsConfig | permissions |
| Tool permissions. | |
| MCPConfig | mcp |
| MCP server settings. | |
| CompactionConfig | compaction |
| Auto-compaction settings. | |
| DelegationConfig | delegation |
| Delegation isolation (gh#160) | |
| LSPConfig | lsp |
| LSP integration. | |
| PromptCacheConfig | prompt_cache |
| Prompt KV cache settings. | |
| StorageConfig | storage |
| Storage backend settings (v1.8.8) | |
| std::string | log_level = "INFO" |
| Log level string. | |
| std::optional< std::filesystem::path > | constitution |
| Constitution: nullopt = bundled default, disabled = explicit false. | |
| bool | constitution_disabled = false |
| true if constitution explicitly disabled | |
| std::optional< std::filesystem::path > | app_context |
| App context: nullopt = disabled by default. | |
| bool | app_context_disabled = false |
| true if app_context explicitly disabled | |
| std::optional< std::string > | app_context_content |
| Inline app_context text, supplied instead of a path (gh#141). | |
| std::filesystem::path | workspace_app_context |
| Per-workspace app_context file, relative to each workspace root (gh#179, v2.14.0). | |
| bool | concurrent_sessions = true |
| gh#158 (v2.13.0): may DIFFERENT session keys run together? | |
| bool | inject_model_context = true |
| Auto-inject model context into system prompt. | |
| int | vram_reserve_mb = 512 |
| Reserved VRAM headroom (MB, 0–65536) | |
| std::filesystem::path | config_dir |
| Config dir — base for bundled data discovery. | |
| std::filesystem::path | log_dir |
| Session log directory (session.log + session_model.log). | |
| bool | ggml_logging = false |
| Enable ggml/llama.cpp logging to llama_ggml.log in log_dir. | |
| std::filesystem::path | llama_log_path |
Override path for ggml/llama log when ggml_logging == true (gh#23 MVP item 12, v2.3.24). | |
| bool | console_logging = true |
| Emit engine spdlog output to the stderr console sink. | |
| ConstitutionalValidationConfig | constitutional_validation |
| Constitutional validation pipeline settings. | |
| InferenceConfig | inference |
| Inference-side knobs (currently speculative decoding only). | |
Detailed Description
Full parsed configuration.
Aggregates all config sections. C++ equivalent of Python's LibraryConfig (engine-only fields, no TUI).
- Version
- 1.8.1
Member Data Documentation
◆ app_context
| std::optional<std::filesystem::path> entropic::ParsedConfig::app_context |
◆ app_context_content
| std::optional<std::string> entropic::ParsedConfig::app_context_content |
Inline app_context text, supplied instead of a path (gh#141).
Set by the app_context: {content: ...} object form. Takes precedence over app_context when both are present. Exists because a consumer may hold this text in memory and be unable to write it to disk — the file is a provenance boundary for them, and on Android there is no stable writable path to point at.
◆ app_context_disabled
◆ compaction
| CompactionConfig entropic::ParsedConfig::compaction |
◆ concurrent_sessions
gh#158 (v2.13.0): may DIFFERENT session keys run together?
ON by default since the audit that decision #66 records was completed: every piece of per-handle mutable state a turn touches is now per-run or locked. The flag shipped false in the first half of gh#158 precisely because that audit was outstanding — a racy default is worse than honest serialization — so turning it on is the audit's conclusion, not a change of mind about the risk.
It stays as a KILL SWITCH, not as an opt-in. Set concurrent_sessions: false to restore v2.12.0 semantics exactly: a second run on ANY key returns ENTROPIC_ERROR_ALREADY_RUNNING and the handle serializes every turn. That is the escape hatch for a consumer who hits a concurrency defect in the field and needs a one-line configuration change rather than a version pin.
With it on (the default), only a second run on the SAME key is refused — that one is a genuine conflict, because the two runs share one conversation.
YAML: concurrent_sessions: false.
◆ config_dir
| std::filesystem::path entropic::ParsedConfig::config_dir |
◆ console_logging
Emit engine spdlog output to the stderr console sink.
Default true (operators reading stderr see everything). TUI consumers that paint to fd 2 must set this false so engine log lines don't corrupt the screen — with it off, logs route to the per-handle session.log file sink only. (gh#59 follow-up, v2.3.7)
◆ constitution
| std::optional<std::filesystem::path> entropic::ParsedConfig::constitution |
◆ constitution_disabled
◆ constitutional_validation
| ConstitutionalValidationConfig entropic::ParsedConfig::constitutional_validation |
◆ delegation
| DelegationConfig entropic::ParsedConfig::delegation |
◆ generation
| GenerationConfig entropic::ParsedConfig::generation |
◆ ggml_logging
◆ inference
| InferenceConfig entropic::ParsedConfig::inference |
◆ inject_model_context
◆ llama_log_path
| std::filesystem::path entropic::ParsedConfig::llama_log_path |
Override path for ggml/llama log when ggml_logging == true (gh#23 MVP item 12, v2.3.24).
When non-empty, the orchestrator installs llama_log_set against this exact path. When empty (default), keeps the pre-v2.3.24 hardcoded <log_dir>/llama_ggml.log — bit-identical for callers not opting in. Accepts absolute or relative paths; relative paths resolve from the process CWD (not log_dir).
- Version
- 2.3.24
◆ log_dir
| std::filesystem::path entropic::ParsedConfig::log_dir |
Session log directory (session.log + session_model.log).
When non-empty, the facade creates a SessionLogger on configure.
◆ log_level
| std::string entropic::ParsedConfig::log_level = "INFO" |
◆ lsp
◆ mcp
| MCPConfig entropic::ParsedConfig::mcp |
◆ models
| ModelsConfig entropic::ParsedConfig::models |
◆ permissions
| PermissionsConfig entropic::ParsedConfig::permissions |
◆ prompt_cache
| PromptCacheConfig entropic::ParsedConfig::prompt_cache |
◆ routing
| RoutingConfig entropic::ParsedConfig::routing |
◆ storage
| StorageConfig entropic::ParsedConfig::storage |
◆ vram_reserve_mb
| int entropic::ParsedConfig::vram_reserve_mb = 512 |
◆ workspace_app_context
| std::filesystem::path entropic::ParsedConfig::workspace_app_context |
Per-workspace app_context file, relative to each workspace root (gh#179, v2.14.0).
Empty = off (the default).
When set (e.g. .entropic/ENTROPIC.md), a workspace whose root holds that file uses its text as the app_context for every session bound to it, in place of the handle's. A workspace without the file keeps the handle's. Off by default because it puts repository-supplied text into the system prompt — the same "the repository decides" exposure gh#199 closed for .mcp.json. Must be relative and stay inside the root (no ..). Read once, when the workspace is created.
The documentation for this struct was generated from the following file:
- entropic/types/config.h
