Configuration

Skryba has no configuration file. The CLI reads flags and environment variables; flags backed by clap environment bindings win over the environment. Skryba Desktop optionally supplies missing environment variables from ~/.config/skryba/env before spawning the CLI.

Model inference

Variable / flag	Default	Reader	Meaning
BRAMA_URL / --brama-url	unset	src/cli/args/mod.rs	Brama base URL; Skryba appends /v1/chat/completions for generation or /v1/decisions for typed choices
SKRYBA_MODEL_SERVICE_URL / --model-service-url	unset	src/cli/args/mod.rs	OpenAI-compatible service base URL; conflicts with --brama-url
BRAMA_TOKEN_ROLE	unset	src/app/brama/credentials.rs	ROLE#FIELD resolved by stado credentials get --role; returned value is the bearer
SKRYBA_MODEL_SERVICE_KEY_FILE	unset	src/app/brama/credentials.rs	owner-only regular file holding a bearer; required with the direct-service URL and refused with the Brama URL
SKRYBA_MODEL / --model	required for chat generation and direct-service decisions	src/cli/args/mod.rs	chat model sent to the selected route; Brama's folder and held-out voice decisions use decision-model instead
SKRYBA_JUDGE_MODEL / --judge-model	generator model	src/cli/args/mod.rs	quality-judge model
SKRYBA_AGENT_ID	unset	src/app/brama/mod.rs	optional agent identity; pair with the secret
SKRYBA_AGENT_AUTH_SECRET_ROLE	unset	src/app/brama/credentials.rs	optional ROLE#FIELD for the HMAC secret; pair with agent id
STADO_BIN	stado	src/app/brama/credentials.rs	executable resolving credential roles
SKRYBA_CREDENTIALS_FILE	unset	src/app/brama/credentials.rs	explicitly selected owner-only JSON role → field → value file for use without Stado

The URL flag determines the credential route. --brama-url (or BRAMA_URL) requires BRAMA_TOKEN_ROLE and refuses a service key-file declaration. --model-service-url (or SKRYBA_MODEL_SERVICE_URL) requires SKRYBA_MODEL_SERVICE_KEY_FILE and refuses Brama token and signing declarations; it sends no agent signature. The URL flags conflict. On the Brama route, agent id and secret are all-or-nothing; a present-but-blank declaration is refused. The generator and judge must support the requested output language. A style or language selection does not remove credential requirements. Calls carry no token budget: the routed model writes until it ends.

Role references never name vault items. Missing fields, empty secrets, invalid references and failed credential commands are errors; the explicit credentials file is not an automatic fallback. BRAMA_TOKEN and SKRYBA_AGENT_AUTH_SECRET are retired and are not read. See credential boundary for the file shape.

Detector selection

Variable / flag	Default	Reader	Meaning
GPTZERO_API_KEY	unset	src/text/detectors/ensemble.rs, src/text/detectors/http.rs	enables GPTZero
WINSTON_API_KEY	unset	src/text/detectors/ensemble.rs, src/text/detectors/http.rs	enables Winston
ORIGINALITY_API_KEY	unset	src/text/detectors/ensemble.rs, src/text/detectors/http.rs	enables Originality.ai
SAPLING_API_KEY	unset	src/text/detectors/ensemble.rs, src/text/detectors/http.rs	enables Sapling
GPTZERO_API_URL	https://api.gptzero.me/v2/predict/text	src/text/detectors/http.rs	endpoint override
WINSTON_API_URL	https://api.gowinston.ai/v2/ai-content-detection	src/text/detectors/http.rs	endpoint override
ORIGINALITY_API_URL	https://api.originality.ai/api/v3/scan	src/text/detectors/http.rs	endpoint override
SAPLING_API_URL	https://api.sapling.ai/api/v1/aidetect	src/text/detectors/http.rs	endpoint override
--detectors	infer from non-empty API keys	src/cli/args/mod.rs, src/text/detectors/ensemble.rs	comma-separated provider names
--detector-interval-ms	none; requests are not paced	src/cli/args/mod.rs	pause between two requests to one detector

Endpoint overrides are primarily for controlled deployments and offline fixtures. The endpoint is part of the score cache namespace, so an override cannot reuse a different endpoint's scores. Fixture responses do not qualify real style or translation behavior.

Fleet database

The catalog, drafts, writing-style rows and score cache are tables of Skryba's fleet database. Nothing here names a file.

Variable / file	Default	Reader	Meaning
SKRYBA_DATABASE_SCHEMA	skryba	src/app/fleet/mod.rs	schema every table lives in; another name gives a test rows of its own
~/.stado/skryba-database-client-skarbiec-token	—	src/app/fleet/mod.rs, through Stado's stado-database crate	the bearer Skarbiec answers pooler_url and ca_certificate of skryba-database to

The database itself is named by Stado (stado database resolve skryba --consumer skryba) and the route to Skarbiec by Stado's service directory (stado service directory connect skarbiec --consumer skryba); stado must be on PATH. A new schema's catalog starts with the styles preserve, academic, message and the languages preserve, pl, en; after that it holds exactly what skryba style and skryba language put in it. Changing an entry changes what the next rewrite is held to, on the CLI and in Skryba Desktop alike.

Writing-style database operations do not need model or detector credentials. Folder selection and held-out comparison use Brama's typed decision route or the explicitly selected direct service's chat model. Samples and new writing assignments use chat generation. Each model operation reads only its route's credential. train also reaches outside the database: the ster process it spawns resolves a checkpoint through Hugging Face's own environment — HF_HOME for the cache and token location, default ~/.cache/huggingface — which Skryba neither reads nor sets, and Ster inherits it as it stands.

Other CLI controls

These are flags rather than environment keys. Every number a run uses is the caller's: Skryba assumes none, and a missing required one is refused before anything runs.

Flag	Default	Scope
--minimum-quality	required by rewrite, benchmark, train	global judge threshold
rewrite --style	preserve	required style, by name, from the catalog
rewrite --language	preserve	required output language, by name, from the catalog
rewrite --candidates	required	candidates per chunk
rewrite --max-words-per-chunk	required	chunk budget
rewrite/train/benchmark --temperature	required	generator temperature
rewrite/train/benchmark/corpus --seed	required	deterministic local sampling/order
benchmark --samples / --candidates / --min-words / --max-words	required	passages, candidates and passage bounds
train --epochs / --generations / --learning-rate / --beta / --clip-epsilon / --inner-iterations	required	GRPO optimizer
train --detector-weight / --quality-weight / --quality-failure-penalty	required	reward
writing-style add --min-words / --max-words	required	passage word bounds
writing-style export/train --name	writing-style	name in an SFT set, trait_name in a pair set
writing-style export/train --lead-words	required for SFT	words of the writer's own opening used as the SFT prompt
writing-style export/train --limit	every row	oldest rows first
writing-style train --method	sft	ster tune sft on passages, ster tune dpo on pairs
writing-style train --device / --precision	cpu / f32	passed through to ster tune
writing-style train --rank / --alpha / --epochs / --learning-rate / --accumulation / --max-sequence / --batch-size / --seed	required	passed through to ster tune, which assumes none of them
writing-style train --beta	required with --method dpo, refused with sft	passed through to ster tune dpo
writing-style train --export	beside --output	where the export Ster reads is written
rewrite --generator	required	brama or local, which spawns ster generate per candidate
rewrite --local-device / --local-precision	cpu / f32	passed through to ster generate
rewrite --local-chat-template	auto	passed through to ster generate
rewrite --local-max-new-tokens	required with --generator local	passed through to ster generate, which assumes no token budget
database drop-schema --schema	—	schema to drop; skryba is refused

The style and language choices are independent of the model, personal adapter, and sampled policy. train and benchmark use the catalog's preserve entries for both choices. Every command's flags are in CLI reference and train and benchmark.

Desktop environment file

~/.config/skryba/env syntax:

# comments and blank lines are ignored
BRAMA_TOKEN_ROLE=skryba-inference#token
SAPLING_API_KEY=value

Each non-comment line is split at its first =. Key and value are trimmed. Empty keys and malformed lines are ignored. A variable already present in the app's launch environment wins; the file fills only missing keys. The CLI itself does not read this file.