yolop-config

name	yolop-config
description	View and change yolop's own configuration — default provider and model, per-provider API tokens and models, endpoint base URLs, attribution, and harness capabilities. Use when the user asks to configure yolop, set a default provider/model, store an API key, point at a custom endpoint, enable/disable capabilities, or asks "what is your config / what can you configure".
user-invocable	true

Yolop configuration

yolop stores its settings in a single TOML file (settings.toml in the yolop config dir). The file is loaded tolerantly — unknown keys are ignored, never fatal — and every known key carries semantics (title, description, type, default, examples) that you can read at runtime. This skill is the entry point for inspecting and editing that configuration the way a user describes it.

Do not hand-edit the TOML with the file tools. Use the schema-aware tools so values are validated and persisted atomically.

Inspect

1. Call get_config with no arguments to list every configuration key with its meaning, type, default, examples, and current value. Secrets (API tokens) are shown only as stored / unset, never echoed.
2. Call get_config with a single key (e.g. default_provider, models.anthropic, tokens.openai) to focus on one entry.
3. For harness capabilities, use get_config key=capabilities (registered catalog, stored overrides, effective harness) or get_config key=capabilities.<ref> (per-capability schema metadata from config_schema / config_ui_schema).

Lead with get_config whenever you are unsure of the exact key name or the accepted values — the returned schema is the source of truth, so you never have to guess.

Change

Call set_config with a key and a value for scalar settings:

• set_config key=default_provider value=anthropic — the default provider when neither --provider nor an env credential forces a choice.
• set_config key=default_model value="claude-sonnet-4-5" — global fallback model for the active provider. A per-provider pick wins over it.
• set_config key=models.openai value="gpt-5.5 high" — remember a model for one provider (survives provider switches). The spec is model [reasoning-effort].
• set_config key=tokens.anthropic value=… — store an API token (owner-only on disk). Environment variables still override stored tokens.
• set_config key=base_urls.custom value=http://localhost:8000/v1 — endpoint for the OpenAI-compatible custom provider.
• set_config key=attribution value=off — turn commit/PR attribution on/off.

Pass value=clear to unset an optional or secret key (e.g. set_config key=tokens.openai value=clear).

Harness capabilities

Overrides are an ordered [[capabilities]] list in the same file. Append entries with set_config key=capabilities and a json object (validated via each capability's validate_config). Pass value=clear to drop all stored overrides.

[[capabilities]]
ref = "message_metadata"
fields = ["timestamp"]

[[capabilities]]
ref = "duckduckgo"
enabled = false

Tool equivalents:

• set_config key=capabilities json={"ref":"message_metadata","fields":["timestamp"]}
• set_config key=capabilities json={"ref":"duckduckgo","enabled":false}
• set_config key=capabilities json={"ref":"web_fetch","enable_file_download":false}
• set_config key=capabilities json={"ref":"some_cap","append":true,...} — duplicate instance
• set_config key=capabilities value=clear — remove all stored overrides

Provider and model edits are persisted and take effect on the next run. To switch the live model in the current session, use the interactive /setup command instead.

Related surfaces

• Durable preferences / memory ("remember that I prefer terse answers"): these are not config keys. Use the remember / recall / forget tools (the global memory capability), not set_config. Memory tuning (disclosed_titles, recall_limit, soft_cap) is per-capability config exposed via the capability's config_schema, not a settings.toml key.
• Behavioral hooks (block/allow/audit tool calls): use the yolop-hooks skill and the hooks capability tools.
• Interactive provider/model setup: the /setup command runs a guided wizard and switches the live model immediately.