config

type Config struct {
	// Endpoint is the llama-server base URL (e.g. http://127.0.0.1:11436).
	Endpoint string `json:"endpoint"`
	// CompletionPath is the native completion route used to pass a GBNF grammar.
	CompletionPath string `json:"completion_path"`
	// Model is the default workhorse (E4B) — used for summarize/extract and as
	// the fallback for any task without a specific route. Empty = dedicated server.
	Model string `json:"model"`
	// TriageModel serves the fast tier (triage/classify). Empty = use Model.
	// NOTE on 8GB: the tiers are swap-exclusive, so routing a few triages to a
	// cold E2B costs a model swap; for latency-critical mixed workloads set this
	// to "" (let E4B handle triage) or batch same-tier calls.
	TriageModel string `json:"triage_model"`
	// EscalationModel is tried once when the primary defers (validation fail / low
	// confidence) BEFORE deferring to Opus. Empty = no escalation (defer directly).
	EscalationModel string `json:"escalation_model"`
	// ReasoningModel is the terminal LOCAL tier (grammar tasks only): after the whole
	// cascade defers, a thinking model gets one shot under a think-wrapped grammar
	// (gbnf.WrapThinking) to reclaim the deferral before falling through to Opus. Run
	// it thinking-OFF — the grammar, not the chat template, supplies the <think> span.
	// Empty = no reasoning tier (defer straight to Opus, the prior behavior).
	ReasoningModel string `json:"reasoning_model,omitempty"`
	// VisionModel is the VLM alias used for the vqa task (multimodal). Empty = no
	// vision route (vqa defers).
	VisionModel string `json:"vision_model,omitempty"`
	// VisionMaxImageBytes caps a single decoded image before it is rejected
	// (guards context/VRAM blowups). 0 = use the loader default.
	VisionMaxImageBytes int `json:"vision_max_image_bytes,omitempty"`
	// VideoFPS is the frame-sampling rate for video_describe (frames/second).
	VideoFPS float64 `json:"video_fps,omitempty"`
	// VideoMaxFrames caps how many sampled frames are sent to the VLM (bounds the
	// 8GB activation budget — frames, not weights, are the VRAM pressure).
	VideoMaxFrames int `json:"video_max_frames,omitempty"`
	// VideoFrameWidth scales each sampled frame to this width (px), aspect kept.
	VideoFrameWidth int `json:"video_frame_width,omitempty"`
	// FFmpegPath is the ffmpeg executable used to sample frames. Default "ffmpeg".
	FFmpegPath string `json:"ffmpeg_path,omitempty"`
	// --- STT / transcribe (Phase A.2) ---
	// STTModel is the llama-swap alias for the default whisper-server upstream
	// (large-v3-turbo). Empty = no STT route (transcribe defers).
	STTModel string `json:"stt_model,omitempty"`
	// STTModelHQ is the quality-escalation whisper upstream (large-v3), used when
	// a transcribe request passes hq=true. Empty = hq falls back to STTModel.
	STTModelHQ string `json:"stt_model_hq,omitempty"`
	// STTLanguage forces the transcription language ("en"/"es"); "" = auto-detect.
	// Per-call params["language"] overrides this. Forcing is more reliable on
	// noisy/code-switching field audio than auto-detect.
	STTLanguage string `json:"stt_language,omitempty"`
	// STTVAD enables Silero VAD (server launched with -vm + --vad) — the biggest
	// hallucination reducer on noisy field audio. Default true.
	STTVAD bool `json:"stt_vad,omitempty"`
	// STTMaxInlineSegments caps how many timestamped segments are inlined in the
	// result (the rest live in the on-disk .segments.json pointer). Default 120.
	STTMaxInlineSegments int `json:"stt_max_inline_segments,omitempty"`
	// STTUnloadAfter force-unloads the whisper upstream after each transcription
	// (zero-always-warm). Default true; set false for a known batch loop.
	STTUnloadAfter bool `json:"stt_unload_after,omitempty"`
	// STTRequestTimeoutSec bounds one transcription HTTP call (long audio at
	// 5-8x realtime). Default 1800 (30 min). Separate from RequestTimeoutSec.
	STTRequestTimeoutSec int `json:"stt_request_timeout_sec,omitempty"`
	// MediaDir is where transcribe writes .srt/.txt/.segments.json. Default
	// <base>/media.
	MediaDir string `json:"media_dir,omitempty"`
	// SVGDir is where generate_svg writes rendered .svg files. Default <base>/svg.
	SVGDir string `json:"svg_dir,omitempty"`
	// --- image generation (generate_image) ---
	// ImageGenScript is the absolute path to render/comfy-generate.mjs (the Node
	// lifecycle wrapper around comfy-render.mjs). Empty = no image route (the task
	// defers), exactly like an empty STTModel/VisionModel.
	ImageGenScript string `json:"imagegen_script,omitempty"`
	// NodePath is the node executable used to run the image script. Default "node".
	NodePath string `json:"node_path,omitempty"`
	// ComfyDir is the local ComfyUI install dir (passed to the script as COMFY_DIR).
	// Default "C:/ComfyUI".
	ComfyDir string `json:"comfy_dir,omitempty"`
	// ImageGenTimeoutSec bounds one render: ComfyUI cold-start (~4min) + first SDXL
	// render (~6min) + margin. Default 720 (12min).
	ImageGenTimeoutSec int `json:"imagegen_timeout_sec,omitempty"`
	// --- video / audio generation (generate_video / generate_audio) ---
	// VideoGenScript is the path to render/comfy-video.mjs (the I2V lifecycle wrapper).
	// Empty = no video route (the task defers cleanly), like an empty ImageGenScript.
	VideoGenScript string `json:"videogen_script,omitempty"`
	// VoiceGenScript is the path to render/tts.mjs (Chatterbox TTS; comfyManaged:false).
	// It serves generate_audio kind=voice. Empty = voice defers.
	VoiceGenScript string `json:"voicegen_script,omitempty"`
	// MusicGenScript is the path to the ACE-Step music worker (render/comfy-music.mjs,
	// ComfyUI). It serves generate_audio kind=music. Default render/comfy-music.mjs (the
	// B3 worker); set "" to disable (music defers) if the ACE-Step checkpoint isn't present.
	MusicGenScript string `json:"musicgen_script,omitempty"`
	// VideoGenTimeoutSec bounds one video render: ComfyUI cold-start + a long I2V
	// render (Wan native two-stage is slow). Default 1500 (25min).
	VideoGenTimeoutSec int `json:"videogen_timeout_sec,omitempty"`
	// AudioGenTimeoutSec bounds one audio synthesis (TTS or ACE-Step). Default 720 (12min).
	AudioGenTimeoutSec int `json:"audiogen_timeout_sec,omitempty"`
	// VideoGenWaitMs is how long a queued video job waits for the single GPU slot before
	// deferring (passed to the runner as GPU_LOCK waitMs). Long because video is the hero
	// job. Default 1200000 (20min).
	VideoGenWaitMs int `json:"videogen_wait_ms,omitempty"`
	// AudioGenWaitMs is how long a queued audio job waits for the GPU slot before deferring.
	// Kept SHORTER than video so a cheap queued TTS isn't starved by a 20-min video job —
	// it defers cleanly after this window. Default 120000 (2min).
	AudioGenWaitMs int `json:"audiogen_wait_ms,omitempty"`
	// MemoryStack lists the CPU-only, zero-VRAM llama-swap models the GPU-free helper
	// must NEVER unload (the load-bearing mem0 stack). Sourced here (not a buried const)
	// so a renamed/added 3rd CPU member is honored. Threaded to the runner via the
	// MEMORY_STACK env. Default {embeddinggemma, bge-reranker-v2-m3}.
	MemoryStack []string `json:"memory_stack,omitempty"`
	// Temperature for deterministic structured output (default 0).
	Temperature float64 `json:"temperature"`
	// MaxRetries is how many correction re-prompts before deferring.
	MaxRetries int `json:"max_retries"`
	// ClassifyMinConfidence: classify results below this (self-reported) defer.
	ClassifyMinConfidence float64 `json:"classify_min_confidence"`
	// ConfidenceMarginThreshold: for triage/classify, if the logprob-derived
	// top-2 legal-class margin at the decision token is below this, escalate to a
	// larger tier (catches genuinely torn calls, e.g. eager-YES). 0 disables the
	// logprob gate. Default 0.35.
	ConfidenceMarginThreshold float64 `json:"confidence_margin_threshold"`
	// MaxInputChars caps input length before context-budget trimming.
	MaxInputChars int `json:"max_input_chars"`
	// CachePath / LedgerPath are bbolt files.
	CachePath  string `json:"cache_path"`
	LedgerPath string `json:"ledger_path"`
	// --- self-learning artifacts (written by the offline `calibrate`/`health`/
	// `train-router`/`optimize` jobs; read at pipeline startup). Empty = feature off. ---
	ThresholdsPath         string             `json:"thresholds_path,omitempty"`          // Phase 2: per-task conformal margin thresholds
	TierOverridesPath      string             `json:"tier_overrides_path,omitempty"`      // Phase 4: health-driven entry-tier bumps + P95 timeouts
	RouterWeightsPath      string             `json:"router_weights_path,omitempty"`      // Phase 5: logistic entry-tier router
	ConfHeadPath           string             `json:"confhead_path,omitempty"`            // Phase 2: logistic correctness head
	RouterLabelsPath       string             `json:"router_labels_path,omitempty"`       // Phase B: synthesized E2B-counterfactual router label sidecar
	ConfHeadLabelsPath     string             `json:"confhead_labels_path,omitempty"`     // Phase 2: cascade-agreement correctness-label sidecar (classify/triage)
	ConfHeadThresholdsPath string             `json:"confhead_thresholds_path,omitempty"` // Phase 2: per-task conformal p(correct) escalation thresholds (Task 3)
	ConfHeadEnabled        bool               `json:"confhead_enabled,omitempty"`         // Phase 2 Task 4: opt-in — gate ADOPT tasks on the head's p(correct). Default false.
	ExemplarsDir           string             `json:"exemplars_dir,omitempty"`            // Phase 6: few-shot exemplar sidecar + selected pool
	ExemplarShots          int                `json:"exemplar_shots,omitempty"`           // Phase 6: 0 = disabled
	AutoHeal               bool               `json:"auto_heal,omitempty"`                // Phase 7: opt-in autonomous tier reload
	TargetErrorRate        map[string]float64 `json:"target_error_rate,omitempty"`        // Phase 2: per-task α for calibration
	// OpusInputPricePerMTok estimates dollar savings ($ per 1M input tokens).
	OpusInputPricePerMTok float64 `json:"opus_input_price_per_mtok"`
	// RequestTimeoutSec for a single model call.
	RequestTimeoutSec int `json:"request_timeout_sec"`
	// --- shadow-labeling flywheel (Phase A.3) ---
	// ShadowEnabled gates sampled capture of non-escalated classify/triage/extract
	// entry-tier rows to the shadow queue for nightly counterfactual labeling.
	// Default false (off by default; never affects request latency or behavior).
	ShadowEnabled bool `json:"shadow_enabled,omitempty"`
	// ShadowRate is the fraction of eligible rows to capture (0.0–1.0).
	// Default 0.10 (10 %). A value of 0 disables capture even when ShadowEnabled.
	ShadowRate float64 `json:"shadow_rate,omitempty"`
	// ShadowQueuePath is the append-only JSONL file for captured shadow items.
	// Default <DataDir>/shadow-queue.jsonl.
	ShadowQueuePath string `json:"shadow_queue_path,omitempty"`
	// SummarizeSimThreshold is the minimum cosine similarity (0..1) between the
	// entry-tier and escalation-tier summary embeddings to count as "agreed".
	// Used by the B2 summarize judge in shadow-label. Default 0.80.
	SummarizeSimThreshold float64 `json:"summarize_sim_threshold,omitempty"`
	// --- zero-training kNN entry-tier pre-filter (meta-router v2) ---
	// KNNPreFilterEnabled gates the kNN pre-filter — a bridge before the LR
	// router (router_weights) is trained. When on, classify/triage inputs are
	// embedded and matched against KNNIndexPath to decide whether to skip the
	// E2B tier; it yields to the LR router once that task is trained. Default
	// false; with it off the request path is byte-identical to the prior build.
	KNNPreFilterEnabled bool `json:"knn_prefilter_enabled,omitempty"`
	// KNNIndexPath is the JSONL substrate (task, vec, accept) appended by the
	// nightly shadow-label drain. Default <base>/knn-index.jsonl.
	KNNIndexPath string `json:"knn_index_path,omitempty"`
	// KNNPreFilterK is how many nearest neighbors to poll. Default 7.
	KNNPreFilterK int `json:"knn_prefilter_k,omitempty"`
	// KNNMinNeighbors is the minimum usable rows-per-task before the kNN acts
	// (guards against a too-thin substrate). Default 20.
	KNNMinNeighbors int `json:"knn_min_neighbors,omitempty"`
	// KNNPreFilterThreshold: skip E2B when the fraction of the k nearest
	// neighbors that accepted at E2B is below this. Default 0.5.
	KNNPreFilterThreshold float64 `json:"knn_prefilter_threshold,omitempty"`
	// KNNEmbedTimeoutMs bounds the request-path embedding call (fail-open on
	// timeout). Default 2000.
	KNNEmbedTimeoutMs int `json:"knn_embed_timeout_ms,omitempty"`
	// --- explicit remote NIM tool (`nim` subcommand / offload_nim) ---
	// An opt-in path to an OpenAI-compatible NVIDIA NIM endpoint, separate from the
	// local cascade: the GBNF grammar path and the savings ledger are untouched.
	// NIMEndpoint is the base (with the /v1 segment): NVIDIA's hosted API by
	// default, or a self-hosted NIM container (e.g. http://127.0.0.1:8000/v1).
	NIMEndpoint string `json:"nim_endpoint,omitempty"`
	// NIMModel is the default model id (override per call with --model). The hosted
	// catalog has dozens of free endpoints — browse with `nim --list-models`.
	NIMModel string `json:"nim_model,omitempty"`
	// NIMMaxTokens caps a nim completion when the caller sets none. Reasoning models
	// need headroom (they spend tokens thinking before answering). Default 1024.
	NIMMaxTokens int `json:"nim_max_tokens,omitempty"`
	// NIMTimeoutSec bounds one nim call (large hosted models can be slow). Default 120.
	NIMTimeoutSec int `json:"nim_timeout_sec,omitempty"`
}