Configuration¶

G6Solver uses pydantic-settings for configuration, loading values from environment variables and .env files.

Environment variables¶

Variable	Purpose	Required
`OPENROUTER_API_KEY`	LLM routing via OpenRouter	Yes (unless using Ollama)
`ALLOW_PAID_API`	Global opt-in for components that can call paid third-party APIs, including `ctx_vision`	No (prefer per-call consent for local use)
`VISION_DB_PATH`	SQLite persistence location for `ctx_vision` MCP analyses	No (default: `~/.vision_mcp/vision.db`)
`VISION_MODEL`	Override default OpenRouter vision model for `ctx_vision`	No
`ANTHROPIC_API_KEY`	Direct Claude API access for `agent_claude`	Optional
`PANDAS_DB_PATH`	Pandas MCP SQLite location	No (default: `~/.pandas_mcp/pandas.db`)
`DJANGO_SETTINGS_MODULE`	Django config module	No (auto-detected)
`POSTGRES_PASSWORD`	Database credential for Docker deployment	Yes (Docker only)
`BIAS_EVOLUTION_ENABLED`	Enables experimental bias evolution	No (default: `false`)
`BIAS_EVOLUTION_CONFIRM_TOKEN`	Human confirmation secret for mutating bias operations	Required only when bias evolution is enabled
`BIAS_EVOLUTION_TIER`	Local tier override for bias evolution (`trialing` or `premium`)	No
`BIAS_DB_PATH`	SQLite fallback path for bias evolution history	No (default: `~/.bias_mcp/bias.db`)
`G6_JOB_LIVE_SEARCH`	Allows job-agent handlers and recipes to call live web search instead of deterministic local fallback	No (default: `false`)
`G6_JOB_LIVE_LLM`	Allows job-agent debate/enrichment paths to call live LLM backends instead of deterministic/mock fallback	No (default: `false`)
`G6_JOB_ENABLE_GROUNDING`	Allows selected job agents, including `job_political`, to load optional grounding/context enrichment	No (default: `false`)
`G6_JOB_ENABLE_EXPERTA`	Allows selected job agents, including `job_political`, to load optional Experta rule-engine enrichment	No (default: `false`)
`G6_JOB_ENABLE_BAYESIAN`	Allows selected job agents, including `job_political`, to load optional Bayesian enrichment	No (default: `false`)
`LEAN_PROVER_DB_PATH`	SQLite persistence path for Lean prover sessions, theorem records, and strategy history	No (default: `~/.g6/lean_prover.sqlite3`)
`LEAN_PROVER_PROJECT_PATH`	Lake/Lean project path used for Lean project management and Mathlib-backed workflows	No (default: current directory)
`LEAN_BINARY_PATH`	Explicit path to the Lean executable when it is not discoverable on `PATH`	No
`LEAN_PROVER_TEMP_DIR`	Directory for bounded temporary Lean files during verification	No (default: project `.g6/lean_tmp` or OS temp)

Security notice

Never commit .env files to version control. Use environment variables or a secret manager in production. The .gitignore already excludes .env.

Setting up `.env`¶

If you want to pay for your own inference, add an OpenRouter key to your account on the hosted service:

Sign in at g6solver.com.
Open your account's LLM key page.
Paste your sk-or-... key and save it.

The key is sealed at rest, is never displayed again, and can be cleared from the same page. OpenRouter is currently the only provider you can bring a key for.

g6 --gui is not available

This section used to tell you to launch a desktop app and use its Settings panel. There is no desktop application to install; the hosted account page above is the supported path.

Create ~/.g6/.env with your keys:

# ~/.g6/.env
OPENROUTER_API_KEY=your_openrouter_key_here

Paid API consent

Some components require explicit opt-in before they call paid external APIs. For ctx_vision, image-processing calls require both OPENROUTER_API_KEY and either a per-call allow_paid_api=True flag or ALLOW_PAID_API=1 in the MCP server environment.

Prefer per-call allow_paid_api=True during local experimentation so consent is visible at the call site. Use ALLOW_PAID_API=1 only for a runtime where paid API usage is expected and budgeted.

Job-agent live integrations

Job agents default to local/deterministic behavior. Live web search, live LLM-backed debate, grounding, Experta, and Bayesian enrichment are disabled unless explicitly enabled. Enable these only when network access, API credentials, latency, dependency loading, and cost are acceptable for that runtime.

Political workflows

job_political outputs are decision support only. Do not treat them as legal, electoral, campaign-finance, compliance, lobbying, or publication approval advice. Built-in defaults are illustrative, jurisdiction matters, and constituent/campaign workflows may involve sensitive personal or political data.

Lean prover production gates

Before using Lean-backed verification as a production-critical gate, run lean_prover_lean_status and require readiness.ga_ready=true with no readiness.actions. Outputs with verified=false or verification_status="generated" are unverified source only, not proofs. verification_status="proved" is not enough on its own for policy approval; require verified=true or a stored verified theorem record. Use lean_prover_replay_artifact and lean_prover_enforce_policy for medium, high, regulated, or safety-critical workflows.

In production, set LEAN_PROVER_DB_PATH to a managed writable SQLite path. Configure LEAN_PROVER_PROJECT_PATH for workflows that import Mathlib or project-local Lean modules. Pantograph and research grounding are optional; if missing, status output reports warnings and batch Lean proof checking can still work.

Lean proving is an advanced reliability gate. It should be exposed to non-technical launch users through guided recipes or templates, not as a first-run requirement to install and debug Lean/Lake/Mathlib directly.

Full .env for all features:

OPENROUTER_API_KEY=your_openrouter_key_here
ANTHROPIC_API_KEY=your_anthropic_key_here
PANDAS_DB_PATH=~/.pandas_mcp/pandas.db

Bias evolution configuration¶

Bias evolution is experimental and can mutate component bias configuration. It is off by default and should not be part of the first-run onboarding path for a new non-technical user.

For a controlled local pilot:

BIAS_EVOLUTION_ENABLED=true
BIAS_EVOLUTION_CONFIRM_TOKEN=<random-human-reviewed-secret>
BIAS_EVOLUTION_TIER=trialing
BIAS_DB_PATH=~/.bias_mcp/bias.db

For paid production access, use BIAS_EVOLUTION_TIER=premium or rely on the normal license/subscription tier supplied by the gateway. Do not set G6_DEV_MODE=true in production; it bypasses tier checks.

Mutating operations (evolve, apply, rollback) require callers to pass a matching confirm_token. Snapshot, observe, evaluate, and grounding checks can run without it.

Claude-Mem configuration¶

ctx_claude_mem uses an upstream claude-mem worker for persistent Claude session memory. G6 can use an already running worker or the pinned source snapshot under vendor/claude-mem for controlled local testing.

Variable	Purpose
`CLAUDE_MEM_BASE_URL`	Explicit worker URL. Overrides port discovery.
`CLAUDE_MEM_PORT`	Legacy/G6-managed worker port, defaulting to `37777`.
`CLAUDE_MEM_WORKER_PORT`	Current upstream worker port variable. Used when `CLAUDE_MEM_PORT` is unset.
`CLAUDE_MEM_DATA_DIR`	Upstream data directory. G6 reads `.worker.port` from here when present.
`CLAUDE_MEM_WORKER_PATH`	Optional path used by the MCP `worker_start` operation.

Bundling caveat

The vendored claude-mem source is present for source availability, regression testing, and packaging evaluation. The root package is AGPL-3.0, and upstream's ragtime/ subtree carries PolyForm Noncommercial 1.0.0. Do not treat the vendored directory as automatically cleared for paid end-user distribution.

For local validation, use python scripts\claude_mem_vendor.py check, worker start, contract-test, and worker stop. Avoid silently running the upstream global installer from G6; it can rewrite user IDE hooks.

CSF configuration¶

The Computational Safety Framework (CSF) has dedicated configuration fields that control resource limits and safety bounds. These are set via the Settings class and can be overridden with environment variables.

Setting	Default	Description
`allowed_llm_base_urls`	`["https://openrouter.ai/api/v1"]`	Permitted LLM API base URLs
`default_llm_timeout_sec`	`60`	Default timeout for LLM calls (seconds)
`default_llm_max_retries`	`3`	Maximum retry attempts for failed LLM calls
`max_llm_tokens_cap`	`128,000`	Maximum token limit per LLM request
`max_horizon_cap`	`10,000`	Maximum planning horizon for goal decomposition
`default_state_limit`	`50,000`	Default state space limit

Environment variable mapping

All settings follow the pattern SETTING_NAME in uppercase. For example, max_llm_tokens_cap maps to the environment variable MAX_LLM_TOKENS_CAP.

These settings are read from the environment (and ~/.g6/.env) when G6 starts. The defaults above apply unless you override them — for example, add MAX_LLM_TOKENS_CAP=64000 to your ~/.g6/.env, or set it in the GUI Settings tab where exposed. You don't need to set any of these to get started; they are caps and timeouts for advanced tuning.

Per-goal resource bounds¶

For per-run resource control, declare resource_bounds directly in your goal's JSON — the engine enforces them, and they travel with the goal across every surface:

{
  "goal": "Summarize the attached report",
  "resource_bounds": {
    "max_execution_seconds": 30,
    "max_tokens_per_hour": 200000
  }
}

See Safety-Guarded Goals for bounds alongside guardrails, checkpoints, and breakpoints, and Security Architecture for the CSF safety framework.

Next steps¶

Troubleshooting -- common issues and solutions
Security Architecture -- CSF safety framework details
Component Model -- how components use configuration