config_manifest

Default wall-clock budget for one Auto classifier decision batch.

Single source of truth shared by the manifest option, the middleware default, and the resolver, so the three cannot drift (pinned by test).

Smallest accepted Auto classifier timeout.

A sanity bound, not a workable budget: at least a second is required for any provider round trip to have a chance, and below that every gated batch would be denied as classifier_unavailable. A resolved value under the floor is rejected and falls through to the next layer / default.

Largest accepted Auto classifier timeout.

The deadline is what stops a stalled classifier from hanging every gated tool call indefinitely, so it stays bounded: a mistyped or hostile override cannot effectively remove it. A resolved value above the ceiling is rejected and falls through to the next layer / default.

Default LangGraph recursion_limit for the main agent.

Single source of truth shared by the runtime.recursion_limit option, the config.config runnable-config default, and resolve_recursion_limit. Raised above the LangGraph/SDK default (25) to accommodate deeply nested agent graphs in long-running sessions without hitting GRAPH_RECURSION_LIMIT.

Smallest accepted recursion_limit; matches the LangGraph default ceiling.

A value below this would break otherwise-valid runs, so a resolved value under the floor is rejected and falls through to the next layer / default.

Largest accepted recursion_limit.

Bounds the graph step budget so a mistyped or hostile override cannot request effectively unbounded traversal. A resolved value above the ceiling is rejected and falls through to the next layer / default.

Context size above which a resumed thread is offered compaction.

Zero or negative disables the suggestion.

Project agent traces fall back to when no project env var is set.

Single source of truth shared by the tracing.langsmith_project option and config.get_langsmith_project_name.

Visual style for the chat input cursor (a block cell or an underline).

Allowlist derived from CursorStyle so the two never drift.

_env_vars constants intentionally excluded from the option catalog.

Classify a raw env-var string as a truthy, falsy, or unrecognized token.

The single source of truth for which strings count as boolean on/off values; is_env_truthy and the config resolver both build on it so they agree on what "recognizably boolean" means.

Load ~/.deepagents/config.toml.

Resolve an option against the environment then config.toml.

Resolve the [interpreter] options into Settings constructor kwargs.

Only the interpreter group is resolved through the manifest. Credentials, the shell allow-list, and the LangSmith project keep their dedicated loaders in config.py (their empty-string-to-None and reload semantics do not fit the generic resolver), so this stays scoped to the section whose defaults this module owns.

Resolve the Auto classifier decision-batch budget and its source.

Resolve the wall-clock budget for one Auto classifier decision batch.

Resolves models.auto_classifier_timeout through the standard env → config.toml → default precedence. An out-of-range value (below AUTO_CLASSIFIER_TIMEOUT_FLOOR or above AUTO_CLASSIFIER_TIMEOUT_CEILING) is discarded with a logged warning and the next lower-precedence layer is tried, so a bad higher-precedence override cannot mask a valid TOML setting (or the default) and can never remove the deadline that keeps a stalled classifier from hanging every gated tool call.

Resolve the effective main-agent recursion_limit.

Resolves runtime.recursion_limit through the standard env → config.toml → default precedence. An out-of-range value (below RECURSION_LIMIT_FLOOR or above RECURSION_LIMIT_CEILING) is discarded with a logged warning and the next lower-precedence layer is tried, so a bad higher-precedence override cannot mask a valid TOML setting (or the default).

Return the deepagents-code extra that installs provider, if known.

Return the PyPI distribution that provides provider, if known.

Derived from the provider's integration import module by replacing underscores with hyphens (e.g. langchain_google_genai -> langchain-google-genai), which matches the distribution name for every curated entry -- see the _PROVIDER_DEPENDENCIES docstring. This is not PEP 503 normalization, and the result is not validated against PyPI.

Return whether provider's integration package is importable.

Providers without a curated extra (no _PROVIDER_DEPENDENCIES entry) are reported as installed — they manage their own dependencies, so the app should never prompt to install an extra for them.

Return every option, credentials-first then by domain group.

Cached: provider credentials are generated once from PROVIDER_API_KEY_ENV on first call (which lazily imports model_config). The cache assumes that registry is an immutable module constant; a test that monkeypatches it must call get_config_options.cache_clear() (and _options_by_key.cache_clear()).

Return the manifest entry for key, or None when unknown.

Return every manifest key in definition order.

Return every option whose key sits under the dotted prefix section.

Matching is exact on segment boundaries: credentials matches credentials.openai but credential matches nothing, so config get can accept a section name without also accepting truncated guesses.

Matching is case-insensitive, and key prefixes are the only section namespace: display group titles (Credentials, Tools) are not accepted, since several headings (Models, Tools) name a different set of options than the same word as a prefix — one namespace keeps a section unambiguous.

Return group names from options in first-seen order.

How an option's raw env/TOML value is coerced to a typed value.

All kinds flow through resolve_scalar. The scalar kinds (BOOL, BOOL_PRESENCE, INT, FLOAT, STR) are coerced inline by _coerce_env/_coerce_toml. LOG_LEVEL_DELEGATE, SHELL_LIST_DELEGATE, SKILLS_DIRS_DELEGATE, PTC_DELEGATE, and STARTUP_MODE_DELEGATE defer to bespoke parsers (their semantics — dynamic debug fallback, colon-split Path resolution, comma + recommended/all sentinels, and the PTC/startup-mode allowlists — do not compress into a generic coercion). THEME_DELEGATE is resolved separately at the top of resolve_scalar and never reaches the inline coercers. STRUCTURED marks user-defined tables that the scalar resolver only passes through for display.

One user-tunable configuration option and where it can be set.