GthConfig

Defined in: packages/core/dist/config/types.d.ts:161

This is a processed Gaunt Sloth config ready to be passed down into components.

Default values can be found in DEFAULT_CONFIG

Properties

a2aAgents?

optional a2aAgents?: Record<string, unknown>

Defined in: packages/core/dist/config/types.d.ts:480

A2A (Agent-to-Agent) protocol agents configuration. Enables delegation of tasks to external AI agents. Each agent becomes available as a tool named a2a_agent_<agentId>. This feature is experimental and may change.

Experimental

This API should not be used in production and may be trimmed from a public release.

See

https://a2a-protocol.org/

---

acp?

optional acp?: object

Defined in: packages/core/dist/config/types.d.ts:546

EXT-117 — the ACP (Agent Client Protocol) surface, i.e. a session an editor such as Zed opens.

mode is the command an ACP session is resolved under, and it defaults to code: an editor asks for an agent that can do the job, and a read-only one is a worse default than an occasional prompt. Set acp: { mode: 'chat' } for a read-only editor agent.

The deliberate consequence, because it is what makes this overridable without a second parallel config surface: the mode names an EXISTING command, so an ACP session is configured by that command’s own block. At the default the session takes its filesystem, approvals, builtInTools and mode prompt from GthConfig.commands.code — the same settings gth code runs under — and at chat from commands.chat. There is no ACP-only copy of those knobs to keep in step.

Defaulted at the read site (not in DEFAULT_CONFIG) to avoid churning the effective-config snapshot, à la injectModelContext.

mode?

optional mode?: "chat" | "code"

---

agent?

optional agent?: object

Defined in: packages/core/dist/config/types.d.ts:176

Selects the agent backend.

lean — the plain LangChain agent (GthLangChainAgent), given gsloth’s full toolset (filesystem + hardened dev/shell + the gth_checklist planning tool) — is the only backend, and what runs when the key is omitted. Every command resolves the same one, so the key selects nothing today; it exists so a config can name what it runs on, and so a second backend has a name to arrive under.

The retired deep value is a hard config error, not a coercion — see RETIRED_AGENT_BACKENDS in config/schema.ts for why substituting a different agent for the one a config named is the wrong kindness.

backend?

optional backend?: "lean"

---

aiignore?

optional aiignore?: object

Defined in: packages/core/dist/config/types.d.ts:482

enabled?

optional enabled?: boolean

patterns?

optional patterns?: string[]

---

allowDirs?

optional allowDirs?: string[]

Defined in: packages/core/dist/config/types.d.ts:623

Transient (runtime-only) extra filesystem roots the agent may read/write for THIS run, in addition to the cwd sandbox. Populated by gth exec --allow-dir <path> (repeatable); never persisted to a config file.

Currently inert. The permission layer that widened the sandbox belonged to the deepagents backend; the filesystem toolkit the lean agent uses is rooted at the cwd and has no notion of extra roots. The field, the flag and the config merge policy are kept so a config and a script that already declare roots keep working when widening returns, and gth exec says plainly that the flag does nothing rather than announcing a widening that never happens — a warning about a guardrail that was not actually removed teaches the user to distrust the warnings.

---

allowedTools?

optional allowedTools?: string[]

Defined in: packages/core/dist/config/types.d.ts:307

Restrict the agent to this allow-list of tool names, applied after every tool source (filesystem, built-in, custom, MCP, A2A, and tools) is resolved. This is the only knob that can gate MCP and A2A tools, which have no per-source override of their own.

• omitted/undefined: no filtering, all resolved tools remain available.
• non-empty array: keep only tools whose name is in the list.
• empty array []: disable every tool. MCP servers are not even contacted (no OAuth), which is useful for agents that only need to reason over the prompt (e.g. the review agent).

Can be overridden per command via commands.<command>.allowedTools.

---

approvals?

optional approvals?: ApprovalsConfig

Defined in: packages/core/dist/config/types.d.ts:292

CFG-27 — the tool-approval ladder: one of the five rungs (manual · write · assisted · auto · bypass), written either as the bare rung name or as an object carrying the rater’s identity profile and the declared allow/deny lists. Each rung fully determines behaviour — there are no severity thresholds, no strictness levels and no independent rater switch.

Settable at the root or per command (commands.<command>.approvals). §9.1 — a per-command value overrides only the fields it NAMES. mode, rater and raterTimeoutMs replace the root’s; deny and escalate CONCATENATE across every scope, so a per-command rung can never discard the root’s prohibitions; allow is REPLACED when the command states its own and inherited when it does not, so a scope may narrow what runs unprompted and may never widen what is prohibited (§3.1: a too-broad allow entry runs unrated, a missed deny entry does not). Absent = assisted, resolved by resolveApprovals.

---

askWriteMode?

optional askWriteMode?: boolean

Defined in: packages/core/dist/config/types.d.ts:629

Transient (runtime-only) flag set by gth ask --write: opt ask into the same “do-the-job” filesystem + dev tools that exec/code get, so a question can act (read/write files, run commands) rather than only chat. Never persisted to a config file.

---

autocompact?

optional autocompact?: string | number | boolean | { enabled?: boolean; threshold?: string | number; }

Defined in: packages/core/dist/config/types.d.ts:734

EXT-161 — preventive conversation compaction: ON BY DEFAULT.

When the conversation is about to outgrow the model’s context window, the older part is folded into a summary before the request is sent, rather than after the provider has rejected it. The threshold is derived from the model’s real window — the models.dev catalog first, the provider package’s own model profile as a backstop — and when neither knows the window, nothing fires: a guessed threshold is worse than none, because the user cannot see it.

• false (or { enabled: false }) turns it off; the conversation is never folded preventively and state.messages is left exactly as it is.
• true / omitted → on, with the threshold derived from the window.
• a count (300000), a suffixed count ('300K', '0.9M') or a share of the window ('80%') sets the threshold explicitly. K is 1000 and M is 1,000,000 — decimal, because provider token counts are decimal everywhere.
• object → { enabled?, threshold? } spells both out.

A malformed threshold is a hard validation error naming the offending text; it never silently becomes NaN, 0 or a default. /autocompact moves the number for one session, and /status reports the number in force and where it came from.

The on-by-default is applied at the read site (resolveAutocompactConfig), not in DEFAULT_CONFIG, to avoid churning the effective-config snapshot.

---

binaryFormats?

optional binaryFormats?: false | BinaryFormatConfig[]

Defined in: packages/core/dist/config/types.d.ts:203

Binary format support configuration. Disabled by default unless explicitly configured.

---

builtInTools?

optional builtInTools?: BuiltInToolsSetting

Defined in: packages/core/dist/config/types.d.ts:276

Selects and configures the built-in tools the agent loads. Either a string[] of tool names (each enabled) or a registry keyed by tool name whose values enable (true), force-disable (false), or configure (BuiltInToolConfig) each tool. CFG-18 folded the former per-command devTools (the run_* commands + run_shell_command’s EXT-9/10/12 config) into this single registry: e.g. { "run_tests": { "command": "npm test" }, "run_shell_command": { "timeout": 300000 } }. Settable at the root or per command (commands.<command>.builtInTools); a per-command value replaces the top-level one.

---

builtInToolsConfig?

optional builtInToolsConfig?: BuiltInToolsConfig

Defined in: packages/core/dist/config/types.d.ts:481

---

canInterruptInferenceWithEsc

canInterruptInferenceWithEsc: boolean

Defined in: packages/core/dist/config/types.d.ts:427

Allow inference to be interrupted with esc. Only has an effect in TTY mode.

---

commandEnv?

optional commandEnv?: CommandEnvConfig

Defined in: packages/core/dist/config/types.d.ts:610

EXT-126 — what a command Gaunt Sloth spawns (the built-in shell tool and every custom tool) receives as its environment. See CommandEnvConfig. Top-level only: one setting governs every mode, so there is no per-command variant to diverge from it.

---

commands?

optional commands?: object

Defined in: packages/core/dist/config/types.d.ts:486

api?

optional api?: object

api.approvals?

optional approvals?: ApprovalsConfig

§9.1 — per-command approvals posture. It overrides only the fields it NAMES: mode, rater, raterTimeoutMs and allow replace the root’s, while deny and escalate concatenate with it. See GthConfig.approvals.

api.builtInTools?

optional builtInTools?: BuiltInToolsSetting

api.cors?

optional cors?: object

api.cors.allowHeaders?

optional allowHeaders?: string

api.cors.allowMethods?

optional allowMethods?: string

api.cors.allowOrigin?

optional allowOrigin?: string

api.filesystem?

optional filesystem?: string[] | "none" | "all" | "read"

api.host?

optional host?: string

The interface the AG-UI server binds. Defaulted at the read site rather than in DEFAULT_CONFIG — see DEFAULT_AGUI_HOST in @gaunt-sloth/agent’s apiAgUiModule, which is IPv4 loopback. Set 0.0.0.0 (or :: for IPv6 as well) to accept connections from the network; the endpoint has no authentication, so that is a deliberate exposure.

One listen binds one address, so the default is IPv4 loopback and not both loopbacks: a client on this machine that resolves localhost to ::1 and does not retry over IPv4 is refused. Set ::1 for that one — IPv6 loopback, still this machine only.

api.port?

optional port?: number

ask?

optional ask?: CommandToolingConfig

chat?

optional chat?: CommandToolingConfig

code?

optional code?: CommandToolingConfig

gth code — interactive coding session. Carries the do-the-job tool/filesystem knobs; the dev/shell tools (CFG-18) are configured via CommandToolingConfig.builtInTools.

exec?

optional exec?: CommandToolingConfig

gth exec — prompt-as-script runtime. Like code, an exec run may need to actually do the job (read/write files, run commands), so it carries the same tool/filesystem knobs.

pr?

optional pr?: PrCommandConfig

review?

optional review?: ReviewCommandConfig

---

commit?

optional commit?: object

Defined in: packages/core/dist/config/types.d.ts:560

GS2-35/EXT-83 — identity for the Co-Authored-By trailer of agent-authored git commits. Gaunt Sloth has no dedicated commit tool (it commits via run_shell_command), so this identity is injected into the code-mode system prompt, which instructs the agent to co-author commits as this account. Optional and defaulted, each field independently: when the name is unset the default is DEFAULT_COMMIT_CO_AUTHOR_NAME decorated with the resolved active model — Gaunt Sloth (provider:model) — falling back to the bare DEFAULT_COMMIT_CO_AUTHOR_NAME when no model resolves or injectModelContext is false; when the email is unset it is DEFAULT_COMMIT_CO_AUTHOR_EMAIL. A CONFIGURED name is emitted verbatim — the model identity decorates only the default.

coAuthor?

optional coAuthor?: object

coAuthor.email?

optional email?: string

coAuthor.name?

optional name?: string

---

consoleLevel?

optional consoleLevel?: StatusLevel

Defined in: packages/core/dist/config/types.d.ts:445

Console logging level. Only messages at or above this level will be displayed. Valid values: ‘debug’, ‘info’, ‘display’, ‘success’, ‘warning’, ‘error’, ‘stream’ Default: ‘info’ (not debug)

---

contentSource

contentSource: string

Defined in: packages/core/dist/config/types.d.ts:209

Content source type. Source used to fetch content (usually diff) for review or pr command.

DEFAULT_CONFIG

---

contentSourceConfig?

optional contentSourceConfig?: Record<string, unknown>

Defined in: packages/core/dist/config/types.d.ts:448

---

customTools?

optional customTools?: CustomToolsConfig

Defined in: packages/core/dist/config/types.d.ts:446

---

debugDump?

optional debugDump?: object

Defined in: packages/core/dist/config/types.d.ts:602

GS2-47 — controls the shared secret-redaction pass applied to /debug-dump archives. Default ON (omitted = redact): secret-named env-var values, inline config secrets, provider-key/auth patterns and sensitive config fields are masked before any artifact hits disk. Set debugDump.redact: false (or run /debug-dump --unsafe-no-redact) to write a RAW archive, which the command flags with a loud “may contain secrets” warning. Defaulted at the read site (not in DEFAULT_CONFIG) to avoid churning the effective-config snapshot.

redact?

optional redact?: boolean

---

debugLog?

optional debugLog?: boolean

Defined in: packages/core/dist/config/types.d.ts:432

Log messages and events to gaunt-sloth.log, use llm.verbose or gth --verbose as more intrusive option, setting verbose to LangChain / LangGraph

---

evalToolCoverage?

optional evalToolCoverage?: EvalToolCoverageConfig

Defined in: packages/core/dist/config/types.d.ts:418

BATCH-48 — the run-level tool-coverage floor for gth eval.

Distinct from a suite’s own tool_coverage.min. That one is graded against the suite that declared it; this one is graded against the run’s aggregate, and the two are independent — neither is promoted into the other. A profile is the intended home: one profile declares the MCP server under test and the floor over that server’s surface.

Optional, with no default and deliberately absent from DEFAULT_CONFIG. A run that never set the key has no run floor, and a default written here would invent one for every config that never asked.

---

filesystem

filesystem: string[] | "none" | "all" | "read"

Defined in: packages/core/dist/config/types.d.ts:266

---

history?

optional history?: object

Defined in: packages/core/dist/config/types.d.ts:188

GS2-7 (B20) / GS2-20 — local session history store. DEFAULT ON (absent = enabled): each run is recorded to a local SQLite store (~/.gsloth/history.db by default, overridable via dbPath) for gth history search / gth insights, and interactive sessions checkpoint their graph state into the same store so a conversation can be resumed. The store is a directory of SQLite files (GS2-121; see history/historyLayout.ts), and a single-file store from 2.0.0 at that path is split into one on first open. enabled: false turns both off and restores the stateless identity. Local only — no telemetry leaves the machine.

dbPath?

optional dbPath?: string

enabled?

optional enabled?: boolean

---

identityProfile?

optional identityProfile?: string

Defined in: packages/core/dist/config/types.d.ts:245

Separate identity profile. May include separate identity, guidelines and command protocol, making gsloth behave as an agent different from default profile behaviour. for example, devops profile to detect changes such as properties and environment variables. Custom config can still win over this one. This setting requires .gsloth/.gsloth-settings directory to exist.

---

includeCurrentDateAfterGuidelines

includeCurrentDateAfterGuidelines: boolean

Defined in: packages/core/dist/config/types.d.ts:249

Whether to include the current date in the project review instructions or not.

---

injectModelContext?

optional injectModelContext?: boolean

Defined in: packages/core/dist/config/types.d.ts:593

GS2-34 — inject the resolved active provider:model identity into the assembled system prompt so the agent knows which model is serving it (to answer “what model are you?” and reason about its own capabilities/limits). Default ON (omitted = inject). Opt out with injectModelContext: false to keep reproducible / model-agnostic runs (e.g. review) blind to the identity — when off, the assembled prompt is exactly as it is without this feature.

EXT-83 — it governs the model identity EVERYWHERE in the prompt, not just the identity line: in code mode the same resolved identity decorates the default git commit co-author name (GthConfig.commit), so turning this off also removes the model from the commit trailer, which degrades to the plain default name. The identity LINE applies in all modes; the trailer it also feeds is code-mode-only, like the cwd/os-shell notes. Defaulted at the read site (not in DEFAULT_CONFIG) to avoid churning the effective-config snapshot.

---

llm

llm: BaseChatModel

Defined in: packages/core/dist/config/types.d.ts:162

---

mcpServers?

optional mcpServers?: Record<string, unknown>

Defined in: packages/core/dist/config/types.d.ts:454

MCP (Model Context Protocol) server connections. Allows connecting to external MCP servers including those requiring OAuth.

See

https://modelcontextprotocol.io/

---

memory?

optional memory?: object

Defined in: packages/core/dist/config/types.d.ts:196

GS2-7 (B21) — opt-in file-backed memory (MEMORY.md / USER.md). DEFAULT OFF. Forward-compat toggle only; the feature is a deferred follow-up.

enabled?

optional enabled?: boolean

---

middleware?

optional middleware?: unknown[]

Defined in: packages/core/dist/config/types.d.ts:340

Middleware configuration for LangChain v1. Middleware provides hooks to intercept and control agent execution at critical points.

Middleware can be:

• Predefined middleware (string or config object) - works in both JSON and JS configs
• Custom middleware objects - only available in JS configs

Example (JSON config):

{
  "middleware": [
    "summarization",
    { "name": "anthropic-prompt-caching", "ttl": "5m" }
  ]
}

Example (JS config):

{
  middleware: [
    "summarization",
    { beforeModel: (state) => { /* custom logic */ return state; } }
  ]
}

Available predefined middleware:

• anthropic-prompt-caching: Reduces API costs by caching prompts (Anthropic only)
• summarization: Condenses conversation history when approaching token limits

---

modelDisplayName?

optional modelDisplayName?: string

Defined in: packages/core/dist/config/types.d.ts:566

---

modelProviderType?

optional modelProviderType?: string

Defined in: packages/core/dist/config/types.d.ts:578

GS2-53 — the configured provider type string (openrouter/deepseek/xai/anthropic/…), stashed by the loader from the raw llm.type before the built BaseChatModel replaces the raw spec. INTERNAL (loader-set, never user-supplied), so it is deliberately absent from the config schema. resolveModelIdentity PREFERS this over the live model’s _llmType() for the injected identity, because that is the model class’s own label rather than the gth provider namespace (huggingface reports openai, and both Gemini providers report google) and would otherwise mislabel the provider half. Absent for module configs (which hand us an already-built LLM with no raw type), where resolution falls back to the guarded _llmType().

---

noDefaultPrompts?

optional noDefaultPrompts?: boolean

Defined in: packages/core/dist/config/types.d.ts:265

If true, only use user-provided system prompts. Do not fall back to the bundled .gsloth.*.md prompt files shipped with the installation. This applies to all .gsloth.*.md files (backstory, system, chat, code, guidelines, review).

---

organization?

optional organization?: object

Defined in: packages/core/dist/config/types.d.ts:255

Organisation name, locale and timezone. Only used with includeCurrentDateAfterGuidelines. timeZone and locale should be in format supported by Intl.DateTimeFormat

locale?

optional locale?: string

name?

optional name?: string

timezone?

optional timezone?: string

---

output?

optional output?: object

Defined in: packages/core/dist/config/types.d.ts:659

GS2-93 — output surface controls.

output.header grades what a NON-TUI text run (--no-tui, ask, exec, eval, pr, review, piped/CI) opens with, across three rungs:

• debug — the full technical preamble: the Workdir/Model/Tools/Middleware status block, the Press Escape or Q to interrupt hint, and their surrounding blank lines.
• compact (the DEFAULT, so an unset key opens with this) — no preamble, and one attribution line naming the command and the model that served it. review/pr render their existing attribution block instead of this line.
• none — nothing at all, including the review attribution block, for a byte-clean stream a caller can diff or template.

The interactive TUI ignores the setting and always shows the full header. Only the opening is graded — never model/tool output, errors, or config-validation warnings, and never the live Thinking… indicator. Defaulted at the read site, not in DEFAULT_CONFIG, to avoid churning the effective-config snapshot.

GS2-106 — output.resumeHint grades the line gth ask and gth exec print on stderr after a recorded run, naming the command that continues the conversation, on the same three rungs:

• compact (the DEFAULT) — the one line, e.g. gth ask --resume <run id> "…".
• debug — that line plus the conversation’s integer id and the history file it lives in; and when a run gets no hint, a line saying why.
• none — nothing.

Batch, eval and workflow cells never print it, whatever the rung. Defaulted at the read site.

header?

optional header?: "none" | "compact" | "debug"

resumeHint?

optional resumeHint?: "none" | "compact" | "debug"

---

prompts?

optional prompts?: PromptsConfig

Defined in: packages/core/dist/config/types.d.ts:220

GS2-43 — the unified prompt-segment config (see PromptsConfig). Each of the seven segments (backstory | guidelines | system | chat | code | exec | review) accepts a string path (e.g. "guidelines": "AGENTS.md") or an object ({ path?, enabled?, mode? }). When a segment is omitted its default-named file / bundled default applies unchanged.

---

recap?

optional recap?: "off" | "outstanding" | "always"

Defined in: packages/core/dist/config/types.d.ts:688

[[EXT-178]] — the end-of-run recap: a short paragraph on the one stop that says nothing today.

Every ending except completed and suspended already announces itself, so the run that looks like success — text, no tool calls, no error — is the silent one, whether or not it finished the job. This grades what that stop says, across three rungs:

• off (the DEFAULT, so an unset key behaves exactly as before this existed) — nothing. The deterministic unfinished-checklist notice still speaks, unchanged.
• outstanding — a recap only on a clean stop whose own checklist still lists work. The cheapest rung that buys the feature’s point, and the one to reach for first.
• always — a recap on every clean stop, including the ones where nothing is left. A recap of a run that went well is useful in its own right, which is what makes this a feature rather than a diagnostic.

It costs a model call, a small non-agentic one made after the turn has already ended, which is why the default is off — see DEFAULT_RUN_RECAP_RUNG in core/runRecap.ts for the whole argument. Defaulted at the read site, not in DEFAULT_CONFIG, to avoid churning the effective-config snapshot. When a recap renders, it replaces the unfinished-checklist notice for that stop rather than appearing beside it, and carries the same counts.

Honoured on chat/code (both the TUI and --no-tui), ask and exec. Deliberately not on batch, eval, workflow, review, pr, ACP or AG-UI — see SingleShotOptions and the node for why each is excluded.

---

recursionLimit?

optional recursionLimit?: number

Defined in: packages/core/dist/config/types.d.ts:439

LangGraph recursion limit for an agent run — the maximum number of super-steps (model ↔ tool round-trips) before the graph throws. Defaults to 1000, which suits long coding chains; embodied / tight-loop consumers can lower it so a stuck run fails fast and visibly instead of grinding.

---

reporters?

optional reporters?: Record<string, string>

Defined in: packages/core/dist/config/types.d.ts:761

BATCH-19 — custom gth eval reporters, keyed by the NAME they are selected under (gth eval … --reporter <name>). Each value is a MODULE PATH, resolved relative to the project dir, whose default export is an EvalReporterFactory (() => EvalReporter). Loaded and registered through the SAME seam the bundled reporters (text, junit) use, so a config reporter can also override a built-in of the same name (config wins). A missing file, a failed import, or a non-function default export is a hard error (the eval harness exits 2). Trusted: it is the user’s own config, which already executes arbitrary JS.

---

requirementSource

requirementSource: string

Defined in: packages/core/dist/config/types.d.ts:213

Requirement source type. Source used to fetch requirements for review or pr command.

---

requirementSourceConfig?

optional requirementSourceConfig?: Record<string, unknown>

Defined in: packages/core/dist/config/types.d.ts:447

---

scopedPrompts?

optional scopedPrompts?: ScopedPromptsEntry[]

Defined in: packages/core/dist/config/types.d.ts:236

CFG-70 — runtime only: the prompts.paths entries this run’s diff actually selected.

It is deliberately absent from the zod schema, from gsloth-config.schema.json and from every config layer. A user never writes it; review() sets it on the already-resolved config after matching the diff’s paths against prompts.paths, and the prompt-reading layer (readPromptSegment) is its only reader. Writing it in a config file would therefore do nothing useful, which is why the schema does not offer it — the schema describes what a user can say, and this is what the run worked out.

It carries the selected entries, not composed text. File resolution needs the config dir, the identity profile and noDefaultPrompts, all of which live in core’s prompt-reading layer; keeping this a list of entries is what lets the review package select without knowing any of that.

---

streamOutput

streamOutput: boolean

Defined in: packages/core/dist/config/types.d.ts:346

Stream output. Some models do not support streaming. Set value to false for them.

DEFAULT_CONFIG

---

streamSessionInferenceLog

streamSessionInferenceLog: boolean

Defined in: packages/core/dist/config/types.d.ts:423

Stream session log instead of writing it when inference streaming is complete. (only works when streamOutput is true)

---

subagents?

optional subagents?: SubagentProfileSpec[]

Defined in: packages/core/dist/config/types.d.ts:771

GS2-33 — profile-backed subagents. Each entry names a subagent and the named config profile the CHILD resolves when the parent spawns it, so a subagent can run under a different model/tools/prompt than the parent (a cheap profile for recall/search while the parent runs on a strong model). Currently honored by nothing: its only consumer was the deep (deepagents) backend, whose task tool gained one selectable subagent per entry, and that backend has been removed. The key is still accepted and warns on use; the lean backend’s own subagent primitive lands in GS2-25.

---

tls?

optional tls?: object

Defined in: packages/core/dist/config/types.d.ts:461

TLS trust for outbound HTTPS. Primarily so an http-transport MCP server behind a private/corporate CA can be reached without prepending NODE_EXTRA_CA_CERTS on every invocation. The mechanism is a process-global undici dispatcher, so it applies to ALL outbound fetch this process makes (LLM provider calls included), not only MCP.

extraCaCerts?

optional extraCaCerts?: string[]

Extra CA certificate file(s) to trust IN ADDITION to Node’s built-in roots. Paths resolve relative to the project dir (or ~/absolute). Additive — never removes a default root.

rejectUnauthorized?

optional rejectUnauthorized?: boolean

DANGER — false disables TLS certificate verification for ALL outbound HTTPS this process makes, not just MCP. Escape hatch only; a loud security warning is emitted every session.

---

toolLoopGuard?

optional toolLoopGuard?: boolean | { halt?: boolean; threshold?: number; warn?: boolean; }

Defined in: packages/core/dist/config/types.d.ts:705

EXT-36 — the tool-loop guard: a repeated-identical-(tool, args) / no-progress detector that runs as a lean-backend beforeModel middleware, the orthogonal sibling of GS2-36’s consecutive-tool-ERROR budget. It catches the case GS2-36 leaves open — a model re-issuing the SAME call verbatim, whether it keeps erroring or keeps “succeeding” with the same result.

• false disables it entirely.
• true / omitted → WARN on, HALT off, default threshold (DEFAULT_TOOL_LOOP_THRESHOLD).
• object → per-field: warn (default ON) injects a control-flow-free nudge at the threshold; halt (default OFF, opt-in) ends the run cleanly (jumpTo:'end', never a throw) at the threshold; threshold is the number of consecutive identical calls that trip it.

WARN is provably harmless (no routing effect, one nudge per signature per streak). The WARN-on default is applied at the read site (not in DEFAULT_CONFIG) to avoid churning the effective-config snapshot.

---

toolOutputPreviewLines?

optional toolOutputPreviewLines?: number

Defined in: packages/core/dist/config/types.d.ts:390

TUI-C105 — the default number of output lines a COLLAPSED tool call previews, on both the Ink TUI and the plain surface, before the … (+N more lines) overflow marker. 0 reduces every tool call to its one-line summary. Absent, the built-in cap applies (TOOL_OUTPUT_PREVIEW_LINES in core/toolDisplay); override it for one tool with builtInTools.<tool>.previewLines, which ranks above this.

A display cap, not a content cap. It changes only what is drawn for the human; the model receives every tool result in full regardless. The knobs that change what a tool returns to the model are separate and per-tool (builtInTools.gth_gh_read_file.maxBytes, builtInTools.run_shell_command.maxOutputBytes).

MUST stay optional: absence is what distinguishes “the user chose a depth” from the built-in cap, and the fallback in getToolPreviewLines collapses without that distinction.

---

toolResultCaptureMaxBytes?

optional toolResultCaptureMaxBytes?: number

Defined in: packages/core/dist/config/types.d.ts:405

BATCH-49 — how many UTF-8 bytes of a tool result are recorded into run stats and eval results.

The recorded stage, not the other two. builtInTools.<tool>.maxOutputBytes / maxBytes cap what a tool returns to the model, and toolOutputPreviewLines caps what is drawn; this caps what is stored, after the model has already seen the whole result. A longer payload is cut on a character boundary, and the record says so (contentTruncated).

A positive integer. 0 is rejected rather than read as “unlimited”: toolOutputPreviewLines uses 0 to mean the minimum, and the same number meaning the opposite here is the confusion the rejection exists to prevent. Absent, the read site applies TOOL_RESULT_CONTENT_CAP; the default is deliberately not in DEFAULT_CONFIG, so the effective-config snapshot does not grow a key nobody set.

---

tools?

optional tools?: StructuredToolInterface<ToolInputSchemaBase, any, any>[] | BaseToolkit[] | ServerTool[]

Defined in: packages/core/dist/config/types.d.ts:293

---

tui?

optional tui?: boolean

Defined in: packages/core/dist/config/types.d.ts:751

CFG-37 — persistent surface preference for the chat/code interactive sessions, settable in both the global and the project config (the project layer wins). true asks for the Ink TUI, false for the plain readline session, absent leaves the terminal auto-detect in charge — which is why it must stay optional, exactly as useColour/useMouse do.

This is the CONFIG-FILE preference, one rung of a chain rather than the answer: the --tui/--no-tui flags and the GTH_NO_TUI escape hatch both outrank it, and the capability gates (no TTY, TERM=dumb, ink not installed) outrank everything because they are checks and not preferences — so true degrades to readline rather than forcing a crash. Deliberately NOT in DEFAULT_CONFIG. The flag arrives separately as CommandLineConfigOverrides.tui; the two meet only in shouldUseTui.

---

useColour

useColour: boolean

Defined in: packages/core/dist/config/types.d.ts:367

Use colour in output

---

useMouse

useMouse: boolean

Defined in: packages/core/dist/config/types.d.ts:374

Enable terminal mouse reporting in the Ink TUI, making its affordances clickable. On by default in an interactive terminal. While it is on the terminal’s own text selection needs a modifier (Shift, or Option in some macOS terminals) — set this to false, or set GTH_NO_MOUSE, to get unmodified selection back.

---

writeBinaryOutputsToFile

writeBinaryOutputsToFile: boolean

Defined in: packages/core/dist/config/types.d.ts:363

Whether binary model outputs should be written to files instead of printed inline. When enabled, supported binary content blocks are materialized as gth_*.<ext> files.

---

writeOutputToFile

writeOutputToFile: string | boolean

Defined in: packages/core/dist/config/types.d.ts:358

Should the output be written to md file. (e.g. gth_2025-07-26_22-59-06_REVIEW.md). Defaults to false (no file is written); set to true for the standard gth_<timestamp>_<COMMAND>.md name. Can be set to false with -wn or -w0 Can be set to a specific filename or path by passing a string:

• Bare filenames (e.g. "review.md") are placed in .gsloth/ when it exists, otherwise project root
• Paths with separators (e.g. "./review.md" or "reviews/last.md") are always relative to project root Please note the string does not accept absolute path, but allows to exit project with .. if necessary.