Qredence Documentation

Fleet RLM reads non-secret defaults from config/fleet.toml. Copy config/fleet.toml.example to start one. The file is optional, but when it exists it must start with a [config] table that sets schema_version = 5.

Fleet resolves each setting from the first source that sets it: explicit overrides (such as CLI flags), the process environment, .env in the working directory, config/fleet.toml, then the built-in default.
Set FLEET_CONFIG in the process environment to load a different TOML file. Fleet ignores FLEET_CONFIG in .env.
Fleet reads the file once at startup. Restart after every change.
Fleet rejects unknown tables, unknown keys, and values of the wrong type. Credentials are environment-only, so setting one in the TOML file is also an error.
An invalid configuration stops the CLI with fleet: error: invalid configuration: … before any server, sandbox, or model call starts.
A Session can override execution_mode and the Root and Sub models. Fleet persists those selections per Session.

The Shipped column lists the value in config/fleet.toml.example. When the file omits a key, the built-in default applies.

`[config]`

Key	Shipped	Description
`schema_version`	`5`	Required. Any other value fails with `Fleet TOML requires [config] schema_version = 5`.

`[runtime]`

Key	Shipped	Range	Environment variable	Description
`execution_mode`	`"hybrid"`	`direct`, `flex`, `rlm`, `hybrid`	`FLEET_EXECUTION_MODE`	Default execution mode for new Sessions. See Reasoning and recursion.
`turn_timeout_seconds`	`0`	`0`, or 10 to 3600	`FLEET_TURN_TIMEOUT_SECONDS`	Wall-clock deadline for one Turn, shared by preparation, routing, sandbox execution, and every child agent. `0` means no Fleet deadline. See Turn deadlines.
`max_llm_calls_per_turn`	`150`	1 to 10000	`FLEET_MAX_LLM_CALLS_PER_TURN`	Shared model-call budget for one Turn. Root, Sub, routing, Skill selection, and Jev decision calls all count.
`max_agent_depth`	`3`	0 to 8	`FLEET_MAX_AGENT_DEPTH`	Maximum nesting depth of child agents. `0` disables child agents.
`max_child_sandboxes`	`8`	0 to 64	`FLEET_MAX_CHILD_SANDBOXES`	Child Sandboxes a Turn can allocate in total.
`max_active_children`	`4`	1 to 16	`FLEET_MAX_ACTIVE_CHILDREN`	Child agents that can run at the same time.
`max_active_sessions`	`64`	1 to 1024	`FLEET_MAX_ACTIVE_SESSIONS`	Session runtimes one server keeps in memory.
`linked_children`	`true`		`FLEET_LINKED_CHILDREN`	Create child Sandboxes as Daytona sandboxes linked to the root Sandbox.

Turn deadlines

turn_timeout_seconds must be an integer: exactly 0, or 10 to 3600. Fleet rejects booleans, floats, and other values before a Session runs. The environment variable also accepts integer text. Fleet resolves each Turn’s deadline once, when the Turn starts:

timeout_seconds in the Turn request, when present. The terminal’s /timeout command sends this field.
turn_timeout_seconds from TOML, .env, or the environment, when you set it.
The default of 0, which means no Fleet deadline.

Without a Fleet deadline, a Turn runs until it finishes, fails, is cancelled, or uses up its model-call budget. 0 turns off only the Fleet deadline. Daytona’s own sandbox lifetime and idle auto-stop bounds still apply. Fleet records the deadline source (request, configuration, or default) and the effective seconds on the stored Turn, the outcome event, and the fleet_turn trace span.

`[llm]`

One shared model connection serves every role: Root, Sub, Reflection, and the DSPy correctness Judge. Any DSPy model route works.

Key	Shipped	Environment variable	Description
`model`	unset (commented out)	`FLEET_LLM_MODEL`	DSPy model route, such as `openai/gpt-4o-mini` or `anthropic/claude-sonnet-4-5`. No whitespace, up to 300 characters. See Default models.
`api_base`	unset	`FLEET_LLM_BASE_URL`	Optional OpenAI-compatible endpoint. Must use HTTPS, or HTTP on a loopback host. Can’t contain credentials, a query string, or a fragment.
`api_key_env`	unset		Optional name of the environment variable that holds the API key.
`adapter`	`"chat"`	`FLEET_LLM_ADAPTER`	DSPy adapter: `chat` or `json`.
`response_format`	`"auto"`	`FLEET_LLM_RESPONSE_FORMAT`	`auto` or `json_object`. `json_object` requests the provider’s JSON mode and requires `adapter = "json"`.
`max_tokens`	unset		Shared maximum output tokens, at least 1.
`temperature`	unset		Shared sampling temperature, from 0 to 2.
`num_retries`	unset		Shared provider retries, from 0 to 10.
`cache`	unset		Shared DSPy response cache setting.

Set the key itself in FLEET_LLM_API_KEY. It’s a credential, so Fleet reads it only from .env or the process environment. Without an explicit key, DSPy’s native provider authentication reads the provider’s own variables from the process environment, such as OPENAI_API_KEY. Fleet resolves api_key_env references from the process environment, then .env, when it loads settings. Restart Fleet to pick up a changed credential. Fleet rejects the unprefixed LLM_BASE_URL and LLM_API_KEY variables with Use FLEET_LLM_MODEL, FLEET_LLM_BASE_URL, and FLEET_LLM_API_KEY. Rename them with the FLEET_ prefix.

Default models

When you don’t set a shared model in [llm] or FLEET_LLM_MODEL, Root and Sub default to openai/system.ai.deepseek-v4-1-flash-1017, and Reflection and Judge use the built-in openai/gpt-4o-mini. An explicit shared model overrides every role that doesn’t set its own model.

`[llm.root]`, `[llm.sub]`, `[llm.reflection]`, and `[llm.judge]`

Role tables are optional overrides. Root runs the Turn program. Sub answers semantic calls from sandbox code and runs child agents. Reflection rewrites instructions during GEPA optimization. Judge is the DSPy correctness judge in fleet eval run and GEPA. A role without its own model inherits the shared [llm] model, endpoint, and key. A role with its own model starts an independent connection that uses native provider authentication, unless the role also sets api_base or api_key_env.

Key	Root	Sub	Reflection	Judge	Description
`model`	unset	unset	unset	unset	Role model route. Unset inherits `[llm].model`.
`api_base`	unset	unset	unset	unset	Role endpoint. Same rules as `[llm].api_base`.
`api_key_env`	unset	unset	unset	unset	Name of the environment variable that holds the role’s key.
`adapter`	unset	unset	unset	unset	`chat` or `json`. Unset inherits `[llm].adapter`.
`response_format`	unset	unset	unset	unset	`auto` or `json_object`. Unset inherits `[llm].response_format`.
`max_tokens`	`16384`	unset	`32000`	unset	Maximum output tokens, at least 1. Unset leaves the limit to the provider.
`temperature`	unset	unset	`1.0`	`0.0`	Sampling temperature, from 0 to 2. Unset leaves it to the provider.
`num_retries`	`3`	`3`	`3`	`3`	Provider retries, from 0 to 10. DSPy performs the retries.
`cache`	`false`	`false`	`false`	`true`	DSPy response cache.

A role value overrides the shared [llm] value. When neither sets a generation key, the role default in this table applies. The Root max_tokens = 16384 cap bounds runaway reasoning. When a capped reply is missing its output fields, DSPy’s native JSON recovery retries once with the same metered model, so the retry counts toward max_llm_calls_per_turn. Fleet doesn’t send its own corrective re-ask. Role generation settings have environment variables named FLEET_<ROLE>_<KEY>, for example FLEET_ROOT_MODEL, FLEET_SUB_MAX_TOKENS, FLEET_ROOT_ADAPTER, or FLEET_SUB_RESPONSE_FORMAT. FLEET_JUDGE_MODEL is an exception: it sets the MLflow judge in [evaluation], not the Judge role model. See Configure models.

`[gateway]`

Key	Shipped	Environment variable	Description
`host`	unset	`DATABRICKS_HOST`	Databricks workspace host for Databricks MLflow. Model inference doesn’t use it.

Configuration URLs can’t contain credentials, query strings, or fragments.

`[decision]`

Key	Shipped	Environment variable	Description
`model`	`"jev-1.14.0"`	`FLEET_TYPESAFE_MODEL`	TypeSafe Jev model for routing, Skill selection, `jev_decide_batched`, and the route-suitability judge in `fleet eval run`. Fleet uses it only when `FLEET_TYPESAFE_API_KEY` is set. Otherwise the Sub model makes these decisions, and `fleet eval run` judges route suitability with `[evaluation].judge_model`.

`[daytona]`

Key	Shipped	Range	Environment variable	Description
`org_id`	unset		`FLEET_DAYTONA_ORG_ID`	Daytona organization ID. Required.
`volume_id`	unset		`FLEET_DAYTONA_VOLUME_ID`	Daytona Volume ID for Session workspaces. Required. A Volume name isn’t a Volume ID.
`auto_stop_minutes`	`30`	1 to 1440	`FLEET_DAYTONA_AUTO_STOP_MINUTES`	Idle minutes before Daytona stops a Sandbox. Fleet restarts a stopped Sandbox before the next Turn.
`shared_notes_subpath`	`"shared/learned-notes"`		`FLEET_SHARED_NOTES_SUBPATH`	Volume subpath mounted at `/learned` in root Sandboxes. An empty string disables shared learned notes.

shared_notes_subpath must be relative. It can’t contain .. or //, or end with /. See Shared learned notes.

`[storage]`

Key	Shipped	Range	Environment variable	Description
`max_upload_bytes`	`25000000`	1 to 25000000	`FLEET_MAX_UPLOAD_BYTES`	Maximum size of one uploaded file.
`compiled_program_dir`	`".fleet/compiled"`		`FLEET_COMPILED_PROGRAM_DIR`	Directory for compiled (optimized) programs.
`local_skills_dir`	`".fleet/skills"`		`FLEET_LOCAL_SKILLS_DIR`	Local Skills that Fleet syncs into each Session workspace.
`lakebase_endpoint`	`""`		`FLEET_LAKEBASE_ENDPOINT`	Lakebase endpoint path (`projects/<p>/branches/<b>/endpoints/<e>`). Enables per-connection Databricks OAuth for PostgreSQL.
`database_schema`	`"fleet_rlm_lite"`		`FLEET_DATABASE_SCHEMA`	PostgreSQL schema for Fleet tables. Must be a lowercase identifier.

With lakebase_endpoint set, FLEET_DATABASE_URL must include a host, a user, and sslmode=require, and must not embed a password.

`[mlflow]`, `[mlflow.local]`, and `[mlflow.databricks]`

Fleet keeps two separate MLflow destinations. [mlflow] mode sets the default backend: "local" (shipped and built-in default) or "databricks". The command can override it: fleet cli databricks selects Databricks, and every command accepts --mlflow local|databricks. Fleet never borrows a value from the other destination.

`[mlflow.local]` key	Shipped	Description
`port`	`5001`	Port of the local tracking server at `http://127.0.0.1:<port>`.
`data_dir`	`".fleet/mlflow"`	Local MLflow data directory.
`experiment_name`	`"fleet-rlm-dev"`	Runtime traces.
`eval_experiment_name`	`"fleet-rlm-eval"`	Evaluation runs.
`optimization_experiment_name`	`"fleet-rlm-gepa"`	GEPA optimization runs.

`[mlflow.databricks]` key	Shipped	Description
`tracking_uri`	`"databricks"`	`databricks` or `databricks://<profile>`.
`experiment_name`	`"/Shared/fleet-rlm-runtime"`	Runtime traces. Must be an absolute workspace path.
`eval_experiment_name`	`"/Shared/fleet-rlm-eval"`	Evaluation runs. Must be an absolute workspace path.
`optimization_experiment_name`	`"/Shared/fleet-rlm-gepa"`	GEPA runs. Must be an absolute workspace path.
`trace_catalog`	`"example_catalog"`	Unity Catalog catalog for traces. Replace the placeholder with your own catalog.
`trace_schema`	`"fleet_observability"`	Schema for runtime and evaluation traces.
`trace_table_prefix`	`"fleet_rlm_runtime"`	Table prefix for runtime traces.
`eval_trace_table_prefix`	`"fleet_rlm_eval"`	Table prefix for evaluation traces.
`tracing_sql_warehouse_id`	`""`	SQL warehouse for trace storage. Set your warehouse ID before you select Databricks.
`optimization_trace_schema`	`"fleet_gepa"`	Schema for optimization traces.
`optimization_trace_table_prefix`	`"fleet_rlm_gepa"`	Table prefix for optimization traces.

Each key has an environment variable named FLEET_MLFLOW_LOCAL_<KEY> or FLEET_MLFLOW_DATABRICKS_<KEY>. With Databricks selected, every trace storage key must be set. Fleet rejects the retired variables FLEET_MLFLOW_MODE, MLFLOW_TRACKING_URI, MLFLOW_EXPERIMENT_NAME, and unprefixed FLEET_MLFLOW_<KEY> names. Move their values to the FLEET_MLFLOW_LOCAL_* or FLEET_MLFLOW_DATABRICKS_* names, and select the backend with [mlflow] mode or on the command line.

`[evaluation]`

Key	Shipped	Environment variable	Description
`judge_model`	`"openai:/gpt-4o-mini"`	`FLEET_JUDGE_MODEL`	MLflow diagnostic judge URI in `<provider>:/<model>` form. MLflow judges use their own provider credentials, separate from Fleet’s `[llm]` connection.

`[server]`

Key	Shipped	Environment variable	Description
`host`	`"127.0.0.1"`	`FLEET_HOST`	Bind address for `fleet serve`.
`port`	`8000`	`FLEET_PORT`	Bind port, from 1 to 65535.
`allow_non_loopback_bind`	`false`	`FLEET_ALLOW_NON_LOOPBACK_BIND`	Allow binding to a non-loopback address.

Credentials

Set credentials in .env or the process environment. Fleet never reads them from config/fleet.toml and never includes them in the configuration view.

Variable	Required	Description
`FLEET_DATABASE_URL`	Yes	PostgreSQL DSN.
`FLEET_DATABASE_MIGRATION_URL`	No	DSN for schema migrations. Must target the same database as `FLEET_DATABASE_URL`.
`FLEET_DAYTONA_API_KEY`	Yes	Daytona API key.
`FLEET_LLM_API_KEY`	No	API key for the shared `[llm]` connection. Without it, DSPy uses the provider’s own environment variables.
`DATABRICKS_TOKEN`	With Databricks MLflow	Databricks token for tracking. Model inference uses it only when you reference it with `api_key_env`, as for Unity Gateway.
`FLEET_TYPESAFE_API_KEY`	No	Enables the TypeSafe Jev decision model.

Retired settings

Schema version 5 accepts only the tables above. Fleet rejects any other table with Unsupported Fleet TOML table. The [models] and [connections] tables fail with [models] and [connections] were removed; configure [llm] directly, and [gateway].base_url is gone. The old [application], [logging], and [capture] tables, per-role base_url_env and timeout_seconds, and the old Daytona lease and child-snapshot settings are gone. Fleet RLM 0.8.1 also removed daytona.snapshot and FLEET_DAYTONA_SNAPSHOT. Fleet rejects a snapshot key in [daytona] and ignores the environment variable. Every Sandbox now starts from the snapshot that Fleet pins. See Snapshots. Removing the setting changes the configuration digest, so revalidate any GEPA approval evidence recorded under the old digest.

`[rlm]`

The [rlm] table is retired. Fleet doesn’t map its keys silently.

The shared model-call limit moved to runtime.max_llm_calls_per_turn.
rlm.max_provider_attempts and rlm.recursion_enabled no longer exist. DSPy owns provider retries through num_retries, and child agents are controlled by runtime.max_agent_depth.
Iteration, wrap-up, finalization, output, and memory knobs are retired. The remaining limits are listed in Turn budgets.

`[posthog]`

The [posthog] table is retired. Fleet has no PostHog configuration.

Fleet-RLM

Fleet Prime Agent

Fleet Pi

GEPA Omni

Qredence Skills

Fleet Reasoner

Qredence Plugins

Configuration reference

`[config]`

`[runtime]`

Turn deadlines

`[llm]`

Default models

`[llm.root]`, `[llm.sub]`, `[llm.reflection]`, and `[llm.judge]`

`[gateway]`

`[decision]`

`[daytona]`

`[storage]`

`[mlflow]`, `[mlflow.local]`, and `[mlflow.databricks]`

`[evaluation]`

`[server]`

Credentials

Retired settings

`[rlm]`

`[posthog]`

​[config]

​[runtime]

​Turn deadlines

​[llm]

​Default models

​[llm.root], [llm.sub], [llm.reflection], and [llm.judge]

​[gateway]

​[decision]

​[daytona]

​[storage]

​[mlflow], [mlflow.local], and [mlflow.databricks]

​[evaluation]

​[server]

​Credentials

​Retired settings

​[rlm]

​[posthog]

`[config]`

`[runtime]`

Turn deadlines

`[llm]`

Default models

`[llm.root]`, `[llm.sub]`, `[llm.reflection]`, and `[llm.judge]`

`[gateway]`

`[decision]`

`[daytona]`

`[storage]`

`[mlflow]`, `[mlflow.local]`, and `[mlflow.databricks]`

`[evaluation]`

`[server]`

Credentials

Retired settings

`[rlm]`

`[posthog]`