# ─── Proxy Server ──────────────────────────────────────────────────────────────

# Port the proxy listens on
MOEPT_PORT=8080

# ─── Lemonade / Backend Server ────────────────────────────────────────────────

# Base URL of the OpenAI-compatible chat-completions server API
MOEPT_SERVER__URL=http://localhost:13305/api/v1

# API key forwarded to the LLM server.
MOEPT_SERVER__LLM_API_KEY=lemonade

# Base URL of the OpenAI-compatible embedding server.
# Defaults to the same API as the chat server; set separately if embeddings are hosted elsewhere.
MOEPT_SERVER__EMBED_URL=http://localhost:13305/api/v1

# API key forwarded to the embeddings server.
MOEPT_SERVER__EMBED_API_KEY=lemonade

# LLM model identifier used for chat completions
MOEPT_SERVER__LLM_MODEL=Qwen3.6-35B-A3B-MTP-GGUF

# Embedding model identifier used for semantic search
MOEPT_SERVER__EMBED_MODEL=embed-gemma-300m-FLM

# Tokenizer used for budget/counting. 'auto' tries a LOCAL Qwen HF tokenizer and
# only falls back to tiktoken cl100k_base (GPT-4 BPE, approximate for Qwen) if
# none is present. For exact Qwen counts, set this to a HuggingFace repo id or a
# local path to the model's tokenizer.json, e.g.:
#   MOEPT_SERVER__TOKENIZER=Qwen/Qwen2.5-Coder-32B-Instruct
#   MOEPT_SERVER__TOKENIZER=/models/qwen3.6-35b-a3b-mtp/tokenizer.json
# Runtime calibration from the backend's real prompt_tokens still corrects the
# residual ratio each turn regardless of this setting.
MOEPT_SERVER__TOKENIZER=auto

# Request timeout in seconds for long context conversations
MOEPT_SERVER__TIMEOUT=300.0

# ─── Agentic Loop Tuning ──────────────────────────────────────────────────────

# Last N user-assistant pairs kept in full detail for immediate reasoning context.
# Older complete turns are evicted from the top; middle-history content is never mutated.
MOEPT_AGENTIC__KEEP_FULL_STEPS=3

# Steps before this index are archived/compressed
MOEPT_AGENTIC__ARCHIVE_THRESHOLD=3

# Character fallback cap for optimized context; token budget takes precedence when positive.
# Should be a meaningful fraction of the backend window, not ~1% of it.
MOEPT_AGENTIC__MAX_OPTIMIZED_CHARS=400000

# Token budget cap. With the adaptive budget ON this is the BASE / last-resort
# floor; the effective budget grows with the conversation horizon up to a fraction
# of the backend window (see adaptive flags below). With adaptive OFF it is the
# fixed hard cap on optimized context.
MOEPT_AGENTIC__MAX_OPTIMIZED_TOKENS=100000

# Adaptive budget (review §4.2.2): replace the fixed cap with a GROWING ceiling so
# the immutable zones always fit (reserved < budget), the eviction governors stop
# fighting, and the rolling-summary fold fires far less often — raising prefix-cache
# reuse and retaining more verbatim context. Disable to restore the legacy fixed cap.
MOEPT_AGENTIC__ADAPTIVE_BUDGET_ENABLED=true

# Tokens added to the adaptive budget per conversation turn (horizon term).
MOEPT_AGENTIC__ADAPTIVE_HORIZON_GROWTH_TOKENS=800

# Ceiling on the adaptive budget as a fraction of the live backend context window
# (0.15 = 15%). Lets a long session grow toward a sane fraction of the window
# instead of over-compacting against an idle one.
MOEPT_AGENTIC__ADAPTIVE_WINDOW_FRACTION=0.15

# Task-complexity term of the adaptive budget: the budget grows by this fraction of
# the code tokens in the current request, so a code-heavy turn keeps more verbatim
# code (and folds less often). 0 disables the term.
MOEPT_AGENTIC__ADAPTIVE_CODE_DENSITY_FACTOR=0.25

# Hard ceiling on per-turn OPTIMIZED-context GROWTH (tokens). Bounds one turn from
# jumping to the full budget and forcing a mid-body rewrite that breaks prefix-cache
# reuse. 0 disables the cap.
MOEPT_AGENTIC__MAX_CONTEXT_GROWTH_PER_TURN=1500

# Hard floor on per-turn OPTIMIZED-context SHRINK (tokens). Symmetric to the growth
# ceiling: bounds one-shot front-eviction so the backend's cached KV for the body
# head stays valid. 0 = AUTO (derive from current_size * shrink_context_fraction,
# floored by max_context_growth_per_turn and shrink_min_tokens).
MOEPT_AGENTIC__MAX_CONTEXT_SHRINK_PER_TURN=0

# Fraction of the CURRENT lean context size used to derive the AUTO per-turn shrink
# cap when max_context_shrink_per_turn=0. The cap scales with what we are actually
# carrying (a lean context), not the model's full window.
MOEPT_AGENTIC__SHRINK_CONTEXT_FRACTION=0.15

# Absolute floor on the AUTO per-turn shrink cap (tokens). Guarantees even a tiny
# lean context still has a bounded, non-trivial shrink rate.
MOEPT_AGENTIC__SHRINK_MIN_TOKENS=800

# Append the quality anchor (original request + accumulated constraints) to the
# trailing volatile turn on over-budget turns. The anchor is a redundant 3rd copy
# of the original request (also in the frozen prefix + rolling-summary head); set
# false to drop it (review §4.2.6) once benchmark quality confirms no regression.
MOEPT_AGENTIC__VOLATILE_QUALITY_ANCHOR_ENABLED=true

# Ratio of max tokens where proactive top-only trimming starts (base default;
# the balanced profile raises this to 0.6)
MOEPT_AGENTIC__PROACTIVE_TRIM_RATIO=0.45

# Ratio of max tokens where compaction/compression starts (base default;
# the balanced profile raises this to 0.85)
MOEPT_AGENTIC__COMPACTION_TRIGGER_RATIO=0.75

# Keep full thinking content for last N steps
MOEPT_AGENTIC__THINKING_PROTECT_RECENT=2

# Session inactivity timeout in seconds before cleanup
MOEPT_AGENTIC__SESSION_TIMEOUT=3600

# Hard cap on concurrently tracked sessions (LRU eviction over the cap; 0 disables)
MOEPT_AGENTIC__MAX_SESSIONS=256

# Use token-based budget enforcement
MOEPT_AGENTIC__USE_TOKEN_BUDGET=true

# Bypass expensive transformations for contexts already under budget
MOEPT_AGENTIC__FAST_PATH_ENABLED=true

# Enable state-based RAG injection for long/over-budget sessions
MOEPT_AGENTIC__RAG_ENABLED=true

# Run tree-sitter code-block optimization (chunk dedup); budget-gated so it
# only fires when the context exceeds the proactive trim threshold.
MOEPT_AGENTIC__OPTIMIZE_CODE_BLOCKS=true

# Compress large code blocks to skeletons under context pressure. Base default is
# true, but the balanced quality profile (the default) sets it false so the model
# keeps full code to edit; only the aggressive profile enables skeletonization.
MOEPT_AGENTIC__CODE_SKELETON_ENABLED=false

# Max code-signature lines carried forward into the evicted-turn code ledger. When
# front-eviction drops a code-bearing turn, its signatures accumulate into a compact
# [Evicted-turn code index] system message so the model keeps code awareness.
MOEPT_AGENTIC__CODE_LEDGER_MAX_SIGS=40

# Apply task-template specialization to rewrite the user prompt (OFF: risks
# changing the client's exact wording in agentic coding)
MOEPT_AGENTIC__PROMPT_TEMPLATE_ENABLED=false

# Replace full re-read file bodies with a unified diff vs the prior snapshot.
# ON by default, but injection is decided dynamically per re-read: it only fires
# when the prior version is already in context (so the model can apply the diff);
# on first read or when the prior was evicted, the full code is kept verbatim.
MOEPT_AGENTIC__DELTA_ENCODE_INJECT=true

# Inject reasoning scaffolding into user messages
MOEPT_AGENTIC__REASONING_PRESEED_ENABLED=false

# Boundary-compress large user code pastes (file dumps pasted into a user turn).
# Same cheap transforms as tool-output compression; frozen into the stable prefix
# so the backend prefix cache stays valid. Disable only if you need verbatim pastes.
MOEPT_AGENTIC__USER_PASTE_COMPRESSION_ENABLED=true
MOEPT_AGENTIC__USER_PASTE_COMPRESSION_MAX_CHARS=4000

# Reversible compression (review §4.1.2 / Forward plan B1). When a tool output is
# compressed, keep the original in a per-session content-addressed store and embed
# its handle in the placeholder, so it is retrievable (GET
# /v1/agent/sessions/{id}/content/{handle}) instead of lost forever. OFF by default:
# the model-facing expand(id) retrieval tool is not yet wired into the protocol.
MOEPT_AGENTIC__REVERSIBLE_COMPRESSION_ENABLED=false
MOEPT_AGENTIC__REVERSIBLE_COMPRESSION_MIN_CHARS=2000

# Whole-file tool-output dedup (review §4.1.3/§4.5.2 / Forward plan B2). Keep the
# FIRST occurrence of a repeated tool output (e.g. a file read) full and collapse
# later identical occurrences to a handle reference (original retained in the
# per-session content store). Stateless + idempotent => cache-stable. OFF by default:
# it rewrites context and should be benchmarked before enabling.
MOEPT_AGENTIC__TOOL_OUTPUT_DEDUP_ENABLED=false
MOEPT_AGENTIC__TOOL_OUTPUT_DEDUP_MIN_CHARS=200

# Query-aware syntactic code slicing (review §4.5.3 / Forward plan B4). For a message
# containing fenced code, keep the imports header plus the top-level definitions named
# in the session's original request, collapsing the rest (tree-sitter). Strictly
# fail-open + never-expands + idempotent => cache-stable. OFF by default: it is
# context-dependent and should be benchmarked before enabling.
MOEPT_AGENTIC__CODE_SLICING_ENABLED=false

# Live config reload without restart (review §11.5 / C9). SIGUSR2 or
# POST /v1/config/reload re-reads AppConfig from the environment. New sessions
# pick up the new config; existing sessions keep their optimizer (no mid-turn race).
MOEPT_AGENTIC__CONFIG_HOT_RELOAD_ENABLED=true

# ─── Code Chunking ─────────────────────────────────────────────────────────────

# Maximum characters per code chunk for semantic indexing
MOEPT_CODE_CHUNKING__CHUNK_MAX_CHARS=1500

# Number of top relevant chunks to retrieve during optimization
MOEPT_CODE_CHUNKING__TOP_K_CHUNKS=5

# Minimum embedding similarity score to include a chunk (0.2 drops low-relevance chunks)
MOEPT_CODE_CHUNKING__MIN_CHUNK_SCORE=0.2

# Embedding vector dimension
MOEPT_CODE_CHUNKING__EMBEDDING_DIM=384

# ─── Cache ─────────────────────────────────────────────────────────────────────

# Maximum number of embeddings to keep in memory cache
MOEPT_CACHE__EMBED_CACHE_MAX=512

# Path to LanceDB vector database data directory
MOEPT_CACHE__LANCEDB_PATH=~/.moeptimizer/lancedb

# ─── Speculative Decoding ─────────────────────────────────────────────────────
# NOTE: client-proxy speculative decoding is non-functional (review03.md §2.1);
# the only effective path is a backend with native MTP support, enabled via
# MOEPT_V050__NATIVE_MTP_PASSTHROUGH (auto-detected by MOEPT_V050__NATIVE_MTP_AUTODETECT).
# The old MOEPT_SPECULATIVE__* variables were removed.

# ─── v0.5.x Optimizations ─────────────────────────────────────────────────────

# Static Prefix KV-Cache Reuse
MOEPT_V050__STATIC_PREFIX_KV_ENABLED=true
MOEPT_V050__STATIC_PREFIX_KV_MAX_ENTRIES=64

# Token-Aware Truncation
MOEPT_V050__TOKEN_AWARE_TRUNCATION_ENABLED=true

# Chunk Fingerprinting & Reuse
MOEPT_V050__CHUNK_FINGERPRINT_ENABLED=true
MOEPT_V050__CHUNK_FINGERPRINT_MAX_ENTRIES=2048

# Embedding batching
MOEPT_V050__EMBEDDING_BATCH_SIZE=32

# Lightweight Hit-Prediction Model
MOEPT_V050__HIT_PREDICTION_ENABLED=true
MOEPT_V050__HIT_PREDICTION_RETRAIN_THRESHOLD=50

# Hierarchical Summarization
# The cache-stable rolling-summary path (cache_stable_summary_enabled) is the SAFE
# summarization mode: older dynamic turns are folded into an append-only block
# placed right after the frozen prefix and protected from front-eviction, so the
# backend's prefix cache stays valid. It only fires under budget pressure with
# cache_stable_mode on. Enabled by default (review §1/§3/§5, #7). The legacy
# hierarchical_summary_enabled flag is an alias for the same safe path.
MOEPT_V050__CACHE_STABLE_SUMMARY_ENABLED=true
MOEPT_V050__HIERARCHICAL_SUMMARY_ENABLED=false
MOEPT_V050__HIERARCHICAL_SUMMARY_MAX_FULL_TURNS=5

# DRIFT fold margin: the live zone may drift this many turns past the keep window
# before a batch fold pulls it back. 0 = use the keep window. Larger => rarer folds
# (better prefix-cache reuse) but a larger context between folds. Only used when
# FOLD_WINDOW_FRACTION is 0.
MOEPT_V050__FOLD_MARGIN_TURNS=0

# Space-based folding (compaction-geometry redesign, review §4.7 / Forward plan
# A1-lite). When > 0, the turn-count DRIFT trigger is disabled and the rolling summary
# folds only once the context exceeds this fraction of the live backend window — so on
# a huge window the conversation is near append-only and cache-breaking folds become
# rare/absent (at low utilization append-and-reuse beats compact-and-break for TTFT).
# Default 0.25 = zero prefix breaks in the opencode/30 dry-run at ~5% of a 262K window,
# with folds still kicking in near ~65K for long sessions. 0 = keep turn-count folding.
MOEPT_V050__FOLD_WINDOW_FRACTION=0.25

# Delta-Encoding of Code
MOEPT_V050__DELTA_ENCODING_ENABLED=true
MOEPT_V050__DELTA_ENCODING_MAX_SNAPSHOTS=100

# Async I/O for Heavy Stages
MOEPT_V050__ASYNC_IO_ENABLED=true
MOEPT_V050__ASYNC_IO_MAX_THREAD_WORKERS=4
MOEPT_V050__ASYNC_IO_MAX_CONCURRENCY=16

# ─── Prefix-Cache Reuse & Native MTP (target: llama.cpp / Qwen3-MTP) ──────────
# These are the settings that make the backend actually reuse the conversation
# prefix across turns (the "cached_tokens" signal) and use the model's own MTP
# speculative decoding. By default the proxy AUTO-DETECTS what the live backend
# supports (per active device, NPU<->GPU aware) and enables the right features
# without manual flags; the explicit flags below are force-on overrides.

# Live, device-aware capability auto-detection. When on (default), the proxy
# probes the backend's own metadata (/health device + backend_url, native /slots,
# native /tokenize, model checkpoint/labels) on a TTL and drives slot pinning,
# MTP passthrough, and tokenizer selection from what the ACTIVE device actually
# supports. Best-effort: probe failures never block startup or requests.
MOEPT_V050__CAPABILITY_AUTODETECT=true

# How long (seconds) a capability snapshot is trusted before a live re-check.
# Re-checks are cheap (cached between TTLs) and let device hot-swaps be picked up
# without restarting the proxy.
MOEPT_V050__CAPABILITY_PROBE_TTL_SECONDS=30.0

# Prefer the backend's exact native /tokenize for authoritative token counts when
# the active device supports it (gated behind fingerprint caches; never on the
# per-fragment hot path). Falls back to the local/tiktoken counter otherwise.
MOEPT_V050__REMOTE_TOKENIZE_ENABLED=true

# Force-on override for slot pinning. Leave FALSE to let auto-detect enable it
# only on devices that expose llama.cpp `/slots` (e.g. GPU); setting it true pins
# a session `id_slot` unconditionally, which is wrong for a device (e.g. NPU) that
# has no slots. id_slot is sent via extra_body so non-llama.cpp backends ignore it.
MOEPT_V050__SLOT_PINNING_ENABLED=false

# Freeze a stable prefix block (system + first user + early turns) verbatim and
# never front-evict from it, so the frozen prefix stays byte-identical across
# turns and the backend's automatic prefix cache reuses it.
MOEPT_V050__CACHE_STABLE_MODE=true

# Number of early complete user-assistant turns (after the first user message)
# to freeze verbatim as part of the stable prefix when cache_stable_mode is on.
MOEPT_V050__FROZEN_PREFIX_TURNS=2

# Force-on override for MTP passthrough. Leave FALSE to let auto-detect enable it
# from backend metadata (labels contains "mtp", a speculative slot, or
# --spec-type ...mtp launch args); the chat-probe below is only a last-resort
# fallback when metadata is inconclusive.
MOEPT_V050__NATIVE_MTP_PASSTHROUGH=false

# Last-resort chat-probe for native MTP when metadata was inconclusive. Best-effort:
# a probe failure never blocks startup and leaves non-MTP backends untouched.
MOEPT_V050__NATIVE_MTP_AUTODETECT=true