.env.example

# ==============================================================================
# LYNKR CONFIGURATION - All Environment Variables
# ==============================================================================
# Copy this file to .env and fill in your values.
#
# FORMAT: Use plain KEY=VALUE syntax (no "export" prefix).
#   Good:  MODEL_PROVIDER=bedrock
#   Bad:   export MODEL_PROVIDER=bedrock
#
# Every variable Lynkr reads from the environment is documented below with:
#   - A one-line DESCRIPTION
#   - An example value
#   - Allowed values when finite (e.g. true|false, or a known provider set)
#
# Variables that need secrets are shown commented-out with a placeholder so the
# file is safe to commit.
# ==============================================================================


# ==============================================================================
# 1. TIER ROUTING (the main routing knob — REQUIRED)
# ==============================================================================
# Format: TIER_<LEVEL>=provider:model[:variant]
# Supported providers: ollama, openai, atlas, azure-openai, azure-anthropic,
#                      openrouter, edenai, databricks, bedrock, vertex, zai,
#                      moonshot, llamacpp, lmstudio
#
# When all 4 TIER_* are set, Lynkr enters "tier routing mode":
#   - MODEL_PROVIDER auto-detected from TIER_SIMPLE
#   - FALLBACK_PROVIDER auto-detected from TIER_REASONING
#   - FALLBACK_ENABLED becomes automatic
#   - Only validates credentials for providers actually used in tiers
#
# Setting MODEL_PROVIDER / FALLBACK_PROVIDER alongside tier routing is rejected.

# DESCRIPTION: Provider:model for trivial single-shot tasks (greetings, formatting)
TIER_SIMPLE=ollama:qwen2.5-coder:latest
# DESCRIPTION: Provider:model for moderate tasks (code edits, small refactors)
TIER_MEDIUM=ollama:qwen2.5-coder:latest
# DESCRIPTION: Provider:model for complex tasks (multi-file changes, design)
TIER_COMPLEX=moonshot:kimi-k2-thinking
# DESCRIPTION: Provider:model for hard reasoning (algorithms, debugging)
TIER_REASONING=moonshot:kimi-k2-thinking

# DESCRIPTION: Auto-fallback when the tier provider fails. Auto-true under tier routing.
# Values: true | false
FALLBACK_ENABLED=false
# DESCRIPTION: Fallback provider when tier provider fails (cannot be local).
# One of: databricks, azure-anthropic, azure-openai, openrouter, edenai, openai, atlas, bedrock
FALLBACK_PROVIDER=databricks
# DESCRIPTION: [DEPRECATED legacy knob] primary provider for credential validation.
# Auto-detected when TIER_* is set. Same allowed values as FALLBACK_PROVIDER plus ollama/llamacpp/lmstudio/vertex/zai/moonshot.
MODEL_PROVIDER=ollama
# DESCRIPTION: [DEPRECATED] legacy preference for Ollama. Use TIER_SIMPLE=ollama:<model> instead.
# Values: true | false
# PREFER_OLLAMA=false


# ==============================================================================
# 2. PER-PROVIDER CONFIG
# ==============================================================================

# ------------------------------------------------------------------------------
# Anthropic (direct + OAuth subscription mode)
# ------------------------------------------------------------------------------
# OAuth mode: when using `lynkr wrap claude`, the OAuth token from `claude login`
# is forwarded automatically — no API key needed.
# DESCRIPTION: Anthropic API key. Only needed when NOT using OAuth subscription.
# ANTHROPIC_API_KEY=sk-ant-your-key-here

# ------------------------------------------------------------------------------
# Azure Anthropic (Anthropic-format endpoint, OAuth-friendly)
# ------------------------------------------------------------------------------
# DESCRIPTION: Endpoint URL for Azure-hosted Anthropic (also used as OAuth passthrough target).
AZURE_ANTHROPIC_ENDPOINT=https://api.anthropic.com/v1/messages
# DESCRIPTION: API key for Azure Anthropic. Not needed if OAuth token is being forwarded.
# AZURE_ANTHROPIC_API_KEY=your-azure-anthropic-key
# DESCRIPTION: Anthropic API version header.
AZURE_ANTHROPIC_VERSION=2023-06-01

# ------------------------------------------------------------------------------
# Azure OpenAI
# ------------------------------------------------------------------------------
# DESCRIPTION: Azure OpenAI endpoint URL (standard or AI Foundry format).
# AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com
# DESCRIPTION: Azure OpenAI API key.
# AZURE_OPENAI_API_KEY=your-azure-openai-key
# DESCRIPTION: Deployment name (e.g. gpt-4o, gpt-5.2-chat).
AZURE_OPENAI_DEPLOYMENT=gpt-4o
# DESCRIPTION: API version to use.
AZURE_OPENAI_API_VERSION=2024-08-01-preview

# ------------------------------------------------------------------------------
# OpenAI (direct)
# ------------------------------------------------------------------------------
# DESCRIPTION: OpenAI API key.
# OPENAI_API_KEY=sk-your-openai-api-key
# DESCRIPTION: Default OpenAI model.
OPENAI_MODEL=gpt-4o
# DESCRIPTION: Chat completions endpoint (can point at any OpenAI-compatible host).
OPENAI_ENDPOINT=https://api.openai.com/v1/chat/completions
# DESCRIPTION: Optional OpenAI org id.
# OPENAI_ORGANIZATION=org-your-org-id

# ------------------------------------------------------------------------------
# Atlas Cloud (OpenAI-compatible)
# ------------------------------------------------------------------------------
# DESCRIPTION: Atlas Cloud API key from https://www.atlascloud.ai/console/api-keys.
# ATLASCLOUD_API_KEY=your-atlas-cloud-api-key
# DESCRIPTION: Default Atlas Cloud text model.
ATLASCLOUD_MODEL=qwen/qwen3.8-max
# DESCRIPTION: Atlas Cloud Chat Completions endpoint.
ATLASCLOUD_ENDPOINT=https://api.atlascloud.ai/v1/chat/completions

# ------------------------------------------------------------------------------
# Ollama (local models)
# ------------------------------------------------------------------------------
# DESCRIPTION: Default Ollama model id.
OLLAMA_MODEL=qwen2.5-coder:latest
# DESCRIPTION: Ollama server endpoint.
OLLAMA_ENDPOINT=http://localhost:11434
# DESCRIPTION: Per-request timeout in milliseconds.
OLLAMA_TIMEOUT_MS=120000
# DESCRIPTION: Ollama `keep_alive` parameter (e.g. "5m", "30m", "-1" for forever).
# OLLAMA_KEEP_ALIVE=5m
# DESCRIPTION: Cap on how many tools are injected when routing through Ollama.
OLLAMA_MAX_TOOLS_FOR_ROUTING=3
# DESCRIPTION: Embedding model (for semantic cache and Cursor @Codebase).
OLLAMA_EMBEDDINGS_MODEL=nomic-embed-text
# DESCRIPTION: Embeddings endpoint URL.
OLLAMA_EMBEDDINGS_ENDPOINT=http://localhost:11434/api/embeddings

# ------------------------------------------------------------------------------
# OpenRouter (100+ models via single API)
# ------------------------------------------------------------------------------
# DESCRIPTION: OpenRouter API key.
# OPENROUTER_API_KEY=your-openrouter-key
# DESCRIPTION: Default OpenRouter model.
OPENROUTER_MODEL=openai/gpt-4o-mini
# DESCRIPTION: Embedding model used through OpenRouter.
OPENROUTER_EMBEDDINGS_MODEL=openai/text-embedding-ada-002
# DESCRIPTION: OpenRouter chat completions endpoint.
OPENROUTER_ENDPOINT=https://openrouter.ai/api/v1/chat/completions
# DESCRIPTION: Cap on tool count sent during routing.
OPENROUTER_MAX_TOOLS_FOR_ROUTING=15

# Eden AI (600+ models via single API, OpenAI-compatible, EU/GDPR)
# DESCRIPTION: Eden AI API key.
# EDENAI_API_KEY=your-edenai-key
# DESCRIPTION: Default Eden AI model (provider/model naming).
EDENAI_MODEL=openai/gpt-4o-mini
# DESCRIPTION: Embedding model used through Eden AI.
EDENAI_EMBEDDINGS_MODEL=openai/text-embedding-ada-002
# DESCRIPTION: Eden AI chat completions endpoint (OpenAI-compatible v3).
EDENAI_ENDPOINT=https://api.edenai.run/v3/chat/completions

# ------------------------------------------------------------------------------
# Databricks
# ------------------------------------------------------------------------------
# DESCRIPTION: Databricks workspace base URL.
# DATABRICKS_API_BASE=https://your-workspace.cloud.databricks.com
# DESCRIPTION: Databricks personal access token.
# DATABRICKS_API_KEY=dapi1234567890abcdef
# DESCRIPTION: Path to the serving endpoint to invoke.
# DATABRICKS_ENDPOINT_PATH=/serving-endpoints/databricks-claude-sonnet-4-5/invocations

# ------------------------------------------------------------------------------
# AWS Bedrock (uses Bedrock API Key, NOT IAM)
# ------------------------------------------------------------------------------
# DESCRIPTION: Bedrock bearer API key (starts with ABSK, generated in AWS console).
# AWS_BEDROCK_API_KEY=ABSK...your-bedrock-api-key
# DESCRIPTION: AWS region.
AWS_BEDROCK_REGION=us-east-1
# DESCRIPTION: Fallback region if AWS_BEDROCK_REGION is unset.
# AWS_REGION=us-east-1
# DESCRIPTION: Default Bedrock model id (often a US inference profile).
AWS_BEDROCK_MODEL_ID=us.anthropic.claude-3-5-sonnet-20241022-v2:0

# ------------------------------------------------------------------------------
# Moonshot AI (Kimi)
# ------------------------------------------------------------------------------
# DESCRIPTION: Moonshot API key.
# MOONSHOT_API_KEY=your-moonshot-api-key
# DESCRIPTION: Moonshot chat completions endpoint.
MOONSHOT_ENDPOINT=https://api.moonshot.ai/v1/chat/completions
# DESCRIPTION: Default Moonshot model.
MOONSHOT_MODEL=kimi-k2-thinking

# ------------------------------------------------------------------------------
# Google Vertex AI / Gemini
# ------------------------------------------------------------------------------
# DESCRIPTION: Vertex/Gemini API key (preferred name).
# VERTEX_API_KEY=your-google-api-key
# DESCRIPTION: Fallback Google API key if VERTEX_API_KEY is unset.
# GOOGLE_API_KEY=your-google-api-key
# DESCRIPTION: Default Gemini model.
VERTEX_MODEL=gemini-2.0-flash

# ------------------------------------------------------------------------------
# llama.cpp (local GGUF server)
# ------------------------------------------------------------------------------
# DESCRIPTION: llama.cpp server base URL.
LLAMACPP_ENDPOINT=http://localhost:8080
# DESCRIPTION: llama.cpp model name to request.
LLAMACPP_MODEL=default
# DESCRIPTION: Per-request timeout in ms.
LLAMACPP_TIMEOUT_MS=120000
# DESCRIPTION: Optional bearer token if your llama.cpp server requires auth.
# LLAMACPP_API_KEY=your-optional-api-key
# DESCRIPTION: Embeddings endpoint (defaults to ENDPOINT + /embeddings).
LLAMACPP_EMBEDDINGS_ENDPOINT=http://localhost:8080/embeddings

# ------------------------------------------------------------------------------
# LM Studio
# ------------------------------------------------------------------------------
# DESCRIPTION: LM Studio OpenAI-compatible endpoint.
LMSTUDIO_ENDPOINT=http://localhost:1234
# DESCRIPTION: LM Studio model id to request.
LMSTUDIO_MODEL=default
# DESCRIPTION: Per-request timeout in ms.
LMSTUDIO_TIMEOUT_MS=120000
# DESCRIPTION: Optional bearer token.
# LMSTUDIO_API_KEY=your-optional-api-key

# ------------------------------------------------------------------------------
# Z.AI (Zhipu AI)
# ------------------------------------------------------------------------------
# DESCRIPTION: Z.AI API key.
# ZAI_API_KEY=your-zai-api-key
# DESCRIPTION: Z.AI Anthropic-compatible messages endpoint.
ZAI_ENDPOINT=https://api.z.ai/api/anthropic/v1/messages
# DESCRIPTION: Default Z.AI model.
ZAI_MODEL=GLM-4.7
# DESCRIPTION: Max concurrent in-flight Z.AI requests.
ZAI_MAX_CONCURRENT=2

# ------------------------------------------------------------------------------
# Baidu Qianfan (ERNIE)
# ------------------------------------------------------------------------------
# DESCRIPTION: Baidu Qianfan API key (format: bce-v3/ALTAK-...).
# BAIDU_API_KEY=your-baidu-qianfan-api-key
# DESCRIPTION: Qianfan v2 OpenAI-compatible chat completions endpoint.
BAIDU_ENDPOINT=https://qianfan.baidubce.com/v2/chat/completions
# DESCRIPTION: Default Baidu ERNIE model.
BAIDU_MODEL=glm-5.2

# ------------------------------------------------------------------------------
# Codex (uses your ChatGPT subscription via local codex CLI)
# ------------------------------------------------------------------------------
# DESCRIPTION: Enable the Codex local provider (requires `codex` CLI installed).
# Values: true | false
# CODEX_ENABLED=true
# DESCRIPTION: Codex model id.
# CODEX_MODEL=gpt-5.3-codex
# DESCRIPTION: Path to the `codex` binary; auto-detected if unset.
# CODEX_BINARY_PATH=codex
# DESCRIPTION: Per-request timeout in ms.
# CODEX_TIMEOUT=120000

# ------------------------------------------------------------------------------
# Embeddings provider override
# ------------------------------------------------------------------------------
# DESCRIPTION: Force a specific embeddings provider (otherwise inferred from MODEL_PROVIDER).
# One of: ollama, llamacpp, openrouter, openai
# EMBEDDINGS_PROVIDER=ollama


# ==============================================================================
# 3. SERVER
# ==============================================================================

# DESCRIPTION: HTTP port the proxy listens on.
PORT=8080
# DESCRIPTION: Pino log level.
# Values: trace | debug | info | warn | error | fatal | silent
LOG_LEVEL=info
# DESCRIPTION: Node runtime mode. "development" enables pino-pretty (requires install).
# Values: development | production | test
NODE_ENV=production
# DESCRIPTION: Max JSON request body size (express bodyParser units, e.g. "1gb").
REQUEST_JSON_LIMIT=1gb
# DESCRIPTION: SQLite path for session storage.
SESSION_DB_PATH=./data/sessions.db
# DESCRIPTION: Absolute path to the workspace Lynkr operates on.
WORKSPACE_ROOT=/path/to/your/workspace
# DESCRIPTION: Pretty-print SQL statements to stdout (very verbose).
# Values: 1 | unset
# DEBUG_SQL=1
# DESCRIPTION: Print per-stage timing breakdowns to stdout.
# Values: true | false
# PERF_TIMER=false

# DESCRIPTION: Persistent file logging via pino-roll.
# Values: true | false
# LOG_FILE_ENABLED=true
# DESCRIPTION: Log file path.
# LOG_FILE_PATH=./logs/lynkr.log
# DESCRIPTION: Log file verbosity.
# LOG_FILE_LEVEL=debug
# DESCRIPTION: Roll frequency.
# Values: daily | hourly | <bytes>
# LOG_FILE_FREQUENCY=daily
# DESCRIPTION: Max rotated files to retain.
# LOG_FILE_MAX_FILES=14


# ==============================================================================
# 4. ROUTING INTELLIGENCE
# ==============================================================================

# DESCRIPTION: Include `lynkr_interaction` block in every successful response showing where it routed.
# Values: true | false
LYNKR_VISIBLE_ROUTING=false
# DESCRIPTION: Cost-optimized routing (downgrade tier when safe).
# Values: true | false
LYNKR_COST_OPTIMIZE=true
# DESCRIPTION: Enable cascading retry/escalation between tiers.
# Values: true | false
LYNKR_CASCADE_ENABLED=false
# DESCRIPTION: For OAuth/subscription requests, score the last N user messages
# instead of just the latest one. Catches "this conversation HAD a complex
# turn earlier" (e.g. an "audit credentials" ask 4 turns back) without
# inflating short follow-ups ("yes", "continue"). Combined with
# LYNKR_INTENT_DECAY as exponential recency weighting; the message with
# the highest decayed score wins.
# Values: positive integer (default 5). Set 1 to disable (latest-only).
LYNKR_INTENT_WINDOW_N=5
# DESCRIPTION: Per-turn exponential decay applied during window scoring.
# weighted_score = raw_score * decay^age, where age=0 is the latest user
# message. Higher (~0.9) = old turns linger longer; lower (~0.5) = old
# turns forgotten faster. 0.7 means a complex turn from 4 messages back
# contributes ~24% of its raw score to the max comparison.
# Values: float in (0, 1] (default 0.7).
LYNKR_INTENT_DECAY=0.7
# DESCRIPTION: Shadow-mode policy name (records what an alt policy would do without applying it).
# LYNKR_SHADOW_POLICY=
# DESCRIPTION: Master switch for the budget enforcer middleware.
# Values: true | false (set "false" to disable)
LYNKR_BUDGET_ENFORCER=true
# DESCRIPTION: Enable the regret-estimator post-hoc routing critic.
# Values: true | false
LYNKR_REGRET_ESTIMATOR=false

# ---- Sticky sessions (WS1) -----------------------------------------
# DESCRIPTION: Cache-aware sticky session pinning. Once a session is
# routed, subsequent turns reuse the pinned provider/model instead of
# re-deciding — avoids a cold-cache re-read (~10× that turn's input cost)
# on every provider switch. Re-decides on compaction, guard escalation
# (risk/context/vision), or economic downgrade below the switch-max
# threshold. Vision requests are pinExempt (per-turn upgrade, no re-pin).
# Values: true | false (default true — set "false" to disable)
LYNKR_STICKY_SESSIONS=true
# DESCRIPTION: TTL for a session pin (ms). Older pins are evicted by the
# 5-minute cleanup tick in src/sessions/cleanup.js.
# Default: 6 hours (21_600_000 ms).
LYNKR_STICKY_TTL_MS=21600000
# DESCRIPTION: Prompt-token cap under which an economic downgrade may
# override the pin. Above this the cold-cache re-read dominates the
# per-token savings, so the pin wins even if the fresh routing decision
# picks a cheaper model. Default 20 000.
LYNKR_SWITCH_MAX_PROMPT_TOKENS=20000
# DESCRIPTION: WS1.5 upward-drift margin. A pinned session re-decides when
# the latest user message's heuristic score exceeds the pinned tier's
# calibrated ceiling + this margin. Smaller = pins escape sooner (more
# re-routes, better tier fidelity); larger = stickier pins (fewer cold
# cache reads, risk of under-tiering). Live guidance: with tool-baseline
# subtraction active, trivial turns score 5-15 and mid-weight asks 22-35,
# so 5 splits them cleanly; the old default of 15 let 32-scoring
# architecture questions ride a SIMPLE pin.
LYNKR_PIN_DRIFT_MARGIN=5
# DESCRIPTION: WS6 cascade verification (EXPERIMENTAL). After a cheap-tier
# (SIMPLE/MEDIUM) response, run structural checks (language drift,
# degeneration, truncation, malformed tool calls, empty/echo) plus a coarse
# content score. A failing answer is DISCARDED and the request escalates up
# the tier ladder — a routing mistake costs seconds, not a garbage answer.
# Failed attempts feed the learning loop as hard negatives. Expensive-tier
# answers are never re-verified. Values: true | false (default off).
# LYNKR_CASCADE_VERIFY=true

# ---- Learning loop (WS5) -------------------------------------------
# DESCRIPTION: Minimum kNN index size before the router will use its
# advice. Below this size the router returns null (heuristic wins).
# Post-WS5 the confidence returned by query() is also damped by
# min(1, size/1000) so a small index advises weakly rather than not
# at all. Default 100 (WS5 lowered from the pre-WS5 default of 1000).
LYNKR_KNN_MIN_INDEX_SIZE=100
# DESCRIPTION: HIGH-confidence threshold above which a kNN suggestion
# overrides the tier-config model directly. Values in (0, 1].
LYNKR_KNN_CONFIDENCE_HIGH=0.7
# DESCRIPTION: LOW-confidence threshold. Between LOW and HIGH the
# suggestion is treated as "ambiguous" and — subject to the WS2 leash
# (only escalate when telemetry.underProvisionedPct ≥ 2%) — may bump
# the tier one level for safety.
LYNKR_KNN_CONFIDENCE_LOW=0.4
# NOTE: LYNKR_AUTO_CALIBRATE and LYNKR_TELEMETRY_DB_PATH used to live
# here. Both are now hardcoded — auto-calibration is always armed (it
# self-gates on telemetry sample count) and the telemetry DB lives at
# `<cwd>/.lynkr/telemetry.db`. Tests use `telemetry._setDbPathForTests()`
# / `_disableForTests()` to isolate their state.

# DESCRIPTION: Forward incoming OAuth Bearer tokens straight to the upstream Anthropic endpoint.
# Set automatically by `lynkr wrap claude`.
# Values: true | false
# LYNKR_OAUTH_PASSTHROUGH=true
# DESCRIPTION: Upstream URL used for OAuth passthrough (defaults to AZURE_ANTHROPIC_ENDPOINT).
# LYNKR_OAUTH_PASSTHROUGH_URL=https://api.anthropic.com/v1/messages
# DESCRIPTION: Inject long-term memory into OAuth-passthrough requests.
# Values: true | false
# LYNKR_OAUTH_MEMORY_INJECTION=false

# DESCRIPTION: Run client-supplied preflight commands (cwd = workspace) and short-circuit if they all pass.
# Values: true | false
LYNKR_PREFLIGHT_ENABLED=false
# DESCRIPTION: Per-command timeout for preflight checks, in ms.
LYNKR_PREFLIGHT_TIMEOUT_MS=120000

# DESCRIPTION: Show stats summary on exit when running `lynkr wrap claude`.
# Values: true | false
LYNKR_WRAP_SHOW_STATS=true

# DESCRIPTION: OpenClaw mode — rewrites response `model` field with actual provider/model used.
# Values: true | false
# OPENCLAW_MODE=false

# DESCRIPTION: Default fallback model name when no tier/provider model is known.
# MODEL_DEFAULT=claude-3-5-sonnet
# DESCRIPTION: JSON of per-model price overrides for the cost registry.
# Format: {"model-name":{"input":0.5,"output":1.5}}
# MODEL_PRICE_OVERRIDES={}

# DESCRIPTION: Suggestion-mode model override.
# Values: default (same as MODEL_PROVIDER) | none | <model-name>
SUGGESTION_MODE_MODEL=default


# ==============================================================================
# 5. TOOL EXECUTION
# ==============================================================================

# DESCRIPTION: Inject native tool definitions into Ollama requests (for models without tool-calling).
# Values: true | false (default true)
INJECT_TOOLS_OLLAMA=true
# DESCRIPTION: Inject native tool definitions into llama.cpp requests.
# Values: true | false (default true)
INJECT_TOOLS_LLAMACPP=true

# ==============================================================================
# 5b. UPSTREAM STREAMING
# ==============================================================================

# DESCRIPTION: Kill switch for native-format streaming passthrough (Anthropic
# client + Anthropic upstream pipes SSE bytes straight through, skipping the
# buffered orchestrator).
# Values: true | false (default true)
LYNKR_NATIVE_PASSTHROUGH=true
# DESCRIPTION: Kill switch for cross-format streaming (OpenAI upstream SSE
# reshaped into Anthropic events in flight instead of buffering).
# Values: true | false (default true)
LYNKR_STREAM_TRANSFORM=true
# DESCRIPTION: Buffer Ollama responses instead of streaming. Buffering repairs
# leaked <think> tags from thinking models (MiniMax) but delivers all at once.
# Values: true | false (default true — set false to stream)
LYNKR_OLLAMA_BUFFER_RESPONSES=true
# DESCRIPTION: Providers eligible for the cross-format stream transform.
# Values: comma-separated (default openai,atlas,azure-openai,openrouter,databricks,lmstudio,llamacpp,moonshot)
#LYNKR_STREAM_TRANSFORM_PROVIDERS=openai,azure-openai
# DESCRIPTION: Abort a passthrough stream when the upstream goes silent this long.
LYNKR_STREAM_IDLE_TIMEOUT_MS=60000

# DESCRIPTION: Smart tool selection strategy (also feeds the routing complexity analyzer).
# Values: heuristic | aggressive | conservative | disabled
SMART_TOOL_SELECTION_MODE=heuristic
# DESCRIPTION: Token budget the smart-selector tries to stay under.
SMART_TOOL_SELECTION_TOKEN_BUDGET=2500

# DESCRIPTION: Master switch for the MCP sandbox.
# Values: true | false
MCP_SANDBOX_ENABLED=true
# DESCRIPTION: Container image used when sandboxing MCP servers.
# MCP_SANDBOX_IMAGE=node:20-alpine
# DESCRIPTION: Sandbox runtime.
# Values: docker | podman
MCP_SANDBOX_RUNTIME=docker
# DESCRIPTION: Workspace mount point inside the container.
MCP_SANDBOX_CONTAINER_WORKSPACE=/workspace
# DESCRIPTION: Mount the host workspace into the container.
# Values: true | false
MCP_SANDBOX_MOUNT_WORKSPACE=true
# DESCRIPTION: Allow network access from inside the sandbox.
# Values: true | false
MCP_SANDBOX_ALLOW_NETWORKING=false
# DESCRIPTION: Docker network mode.
# Values: none | bridge | host | <custom>
MCP_SANDBOX_NETWORK_MODE=none
# DESCRIPTION: Comma-separated env vars to forward into the sandbox.
MCP_SANDBOX_PASSTHROUGH_ENV=PATH,LANG,LC_ALL,TERM,HOME
# DESCRIPTION: Extra bind mounts (HOST:CONTAINER[:ro], comma-separated).
# MCP_SANDBOX_EXTRA_MOUNTS=/host/path:/container/path:ro
# DESCRIPTION: Timeout for a single MCP tool call (ms).
MCP_SANDBOX_TIMEOUT_MS=20000
# DESCRIPTION: Run as this user inside the container.
# MCP_SANDBOX_USER=node
# DESCRIPTION: Override container entrypoint.
# MCP_SANDBOX_ENTRYPOINT=/bin/sh
# DESCRIPTION: Reuse the same sandbox container across calls within a session.
# Values: true | false
MCP_SANDBOX_REUSE_SESSION=true
# DESCRIPTION: Mount the container root filesystem read-only.
# Values: true | false
MCP_SANDBOX_READ_ONLY_ROOT=false
# DESCRIPTION: Set --security-opt no-new-privileges.
# Values: true | false
MCP_SANDBOX_NO_NEW_PRIVILEGES=true
# DESCRIPTION: Linux capabilities to drop (comma-separated, or "ALL").
MCP_SANDBOX_DROP_CAPABILITIES=ALL
# DESCRIPTION: Linux capabilities to add back.
# MCP_SANDBOX_ADD_CAPABILITIES=NET_BIND_SERVICE
# DESCRIPTION: Container memory limit.
MCP_SANDBOX_MEMORY_LIMIT=512m
# DESCRIPTION: Container CPU limit (cores).
MCP_SANDBOX_CPU_LIMIT=1.0
# DESCRIPTION: Max PIDs inside the container.
MCP_SANDBOX_PIDS_LIMIT=100
# DESCRIPTION: How tool permissions are decided.
# Values: auto | allowlist | denylist | prompt
MCP_SANDBOX_PERMISSION_MODE=auto
# DESCRIPTION: Comma-separated tool names always allowed.
# MCP_SANDBOX_PERMISSION_ALLOW=tool1,tool2
# DESCRIPTION: Comma-separated tool names always denied.
# MCP_SANDBOX_PERMISSION_DENY=tool3,tool4
# DESCRIPTION: Single MCP servers.json manifest path.
# MCP_SERVER_MANIFEST=~/.claude/mcp/servers.json
# DESCRIPTION: Comma-separated directories scanned for MCP manifests.
MCP_MANIFEST_DIRS=~/.claude/mcp

# ==============================================================================
# 6. COMPRESSION & CACHING
# ==============================================================================

# DESCRIPTION: Master switch for the in-memory prompt cache.
# Values: true | false
PROMPT_CACHE_ENABLED=true
# DESCRIPTION: Max number of prompts to cache.
PROMPT_CACHE_MAX_ENTRIES=1000
# DESCRIPTION: Cache entry TTL (ms).
PROMPT_CACHE_TTL_MS=300000

# DESCRIPTION: Master switch for semantic (embedding-based) response cache.
# Values: true | false
SEMANTIC_CACHE_ENABLED=true
# DESCRIPTION: Cosine-similarity threshold for a cache hit.
SEMANTIC_CACHE_THRESHOLD=0.95
# DESCRIPTION: Max number of cached entries.
SEMANTIC_CACHE_MAX_ENTRIES=50
# DESCRIPTION: Cache entry TTL (ms).
SEMANTIC_CACHE_TTL_MS=300000

# DESCRIPTION: Enable TOON (token-optimized object notation) encoding for large structured payloads.
# Values: true | false
TOON_ENABLED=true
# DESCRIPTION: Minimum byte size before TOON encoding kicks in.
TOON_MIN_BYTES=4096
# DESCRIPTION: Continue without TOON on encoder failure instead of erroring.
# Values: true | false
TOON_FAIL_OPEN=true
# DESCRIPTION: Log per-request TOON savings stats.
# Values: true | false
TOON_LOG_STATS=true

# DESCRIPTION: Enable GCF (Graph Compact Format) encoding for large structured payloads.
# Drop-in alternative to TOON; takes precedence over TOON when enabled. Opt-in.
# Values: true | false
GCF_ENABLED=false
# DESCRIPTION: Minimum byte size before GCF encoding kicks in.
GCF_MIN_BYTES=4096
# DESCRIPTION: Continue without GCF on encoder failure instead of erroring.
# Values: true | false
GCF_FAIL_OPEN=true
# DESCRIPTION: Log per-request GCF savings stats.
# Values: true | false
GCF_LOG_STATS=true
# DESCRIPTION: Round-trip verify each encoding and keep the original JSON on any mismatch
# (makes the compression provably lossless per payload; adds one decode per converted blob).
# Values: true | false
GCF_VERIFY=true

# DESCRIPTION: Master switch for Headroom sidecar context compression.
# Values: true | false
HEADROOM_ENABLED=true
# DESCRIPTION: Headroom sidecar endpoint.
HEADROOM_ENDPOINT=http://localhost:8787
# DESCRIPTION: Sidecar request timeout in ms.
HEADROOM_TIMEOUT_MS=5000
# DESCRIPTION: Skip compression below this estimated token count.
HEADROOM_MIN_TOKENS=100
# DESCRIPTION: Operating mode.
# Values: audit (observe only) | optimize (apply)
HEADROOM_MODE=optimize
# DESCRIPTION: Provider hint that selects which cache markers to emit.
# Values: anthropic | openai | google
HEADROOM_PROVIDER=anthropic
# DESCRIPTION: Sidecar log level.
# Values: debug | info | warning | error
HEADROOM_LOG_LEVEL=info
# DESCRIPTION: Auto-manage a Docker container for the sidecar.
# Values: true | false
HEADROOM_DOCKER_ENABLED=true
# DESCRIPTION: Sidecar image name.
HEADROOM_DOCKER_IMAGE=lynkr/headroom-sidecar:latest
# DESCRIPTION: Sidecar container name.
HEADROOM_DOCKER_CONTAINER_NAME=lynkr-headroom
# DESCRIPTION: Host port the sidecar publishes on.
HEADROOM_DOCKER_PORT=8787
# DESCRIPTION: Sidecar memory limit.
HEADROOM_DOCKER_MEMORY_LIMIT=512m
# DESCRIPTION: Sidecar CPU limit.
HEADROOM_DOCKER_CPU_LIMIT=1.0
# DESCRIPTION: Docker restart policy for the sidecar.
HEADROOM_DOCKER_RESTART_POLICY=unless-stopped
# DESCRIPTION: Optional Docker network the sidecar joins.
# HEADROOM_DOCKER_NETWORK=lynkr-network
# DESCRIPTION: Build context path when auto-building the sidecar image.
HEADROOM_DOCKER_BUILD_CONTEXT=./headroom-sidecar
# DESCRIPTION: Auto-build the image if not found locally.
# Values: true | false
HEADROOM_DOCKER_AUTO_BUILD=true
# DESCRIPTION: Smart-crusher transform (compress large text blocks).
# Values: true | false
HEADROOM_SMART_CRUSHER=true
# DESCRIPTION: Min token count before smart-crusher engages.
HEADROOM_SMART_CRUSHER_MIN_TOKENS=200
# DESCRIPTION: Max items the smart-crusher processes per request.
HEADROOM_SMART_CRUSHER_MAX_ITEMS=15
# DESCRIPTION: Tool-crusher transform (compress large tool results).
# Values: true | false
HEADROOM_TOOL_CRUSHER=true
# DESCRIPTION: Cache-aligner transform (align cache breakpoints).
# Values: true | false
HEADROOM_CACHE_ALIGNER=true
# DESCRIPTION: Rolling-window transform (keep newest N turns intact).
# Values: true | false
HEADROOM_ROLLING_WINDOW=true
# DESCRIPTION: How many recent turns the rolling window keeps verbatim.
HEADROOM_KEEP_TURNS=10
# DESCRIPTION: CCR (Compress-Cache-Retrieve) mode.
# Values: true | false
HEADROOM_CCR=true
# DESCRIPTION: TTL for CCR cached chunks (seconds).
HEADROOM_CCR_TTL=300
# DESCRIPTION: LLMLingua ML compression (requires GPU).
# Values: true | false
HEADROOM_LLMLINGUA=false
# DESCRIPTION: Device for LLMLingua model.
# Values: auto | cpu | cuda | mps
HEADROOM_LLMLINGUA_DEVICE=auto

# DESCRIPTION: TencentDB-Agent-Memory sidecar (team memory hub: L0-L3 chat
# memory, Skills, Wiki, CodeGraph). When enabled, `lynkr start` launches the
# memory-core + memory-hub containers from Docker Hub (agentmemory/*) — same
# pattern as the Headroom sidecar. Panel UI: http://localhost:8125
# Requires Docker (skips gracefully if unavailable). Containers persist
# across Lynkr restarts (remove with: docker rm -f tdai-memory-core tdai-memory-hub).
# Values: true | false
TENCENTDB_MEMORY_ENABLED=true
# DESCRIPTION: Let Lynkr manage the containers. Set false if you run the
# stack yourself via the project's deploy scripts.
# Values: true | false
# TENCENTDB_MEMORY_DOCKER_ENABLED=true
# DESCRIPTION: LLM endpoint the memory services use for extraction and wiki
# ingest. Defaults to Lynkr's own OpenAI-compatible endpoint (tier routing
# picks the model), so no extra API key is needed. Override to point at a
# provider directly.
# TENCENTDB_MEMORY_LLM_BASE_URL=http://host.docker.internal:8081/v1
# TENCENTDB_MEMORY_LLM_API_KEY=lynkr-local
# TENCENTDB_MEMORY_LLM_MODEL=auto
# DESCRIPTION: Memory extraction style. `code` extracts changes/issues/tool
# usage (coding agents); `chat` extracts general conversational facts.
# Values: code | chat
# TENCENTDB_MEMORY_PROMPT_MODE=code
# DESCRIPTION: Host port overrides (defaults shown).
# TENCENTDB_MEMORY_CORE_PORT=8420
# TENCENTDB_MEMORY_PANEL_PORT=8125
# TENCENTDB_MEMORY_KNOWLEDGE_PORT=8424

# DESCRIPTION: Master switch for the long-term Titans-inspired memory system.
# Values: true | false
MEMORY_ENABLED=true
# DESCRIPTION: Max memories retrieved per request.
MEMORY_RETRIEVAL_LIMIT=5
# DESCRIPTION: Surprise threshold above which a turn is written to memory.
MEMORY_SURPRISE_THRESHOLD=0.3
# DESCRIPTION: Hard cap on memory age (days).
MEMORY_MAX_AGE_DAYS=90
# DESCRIPTION: Hard cap on total memory count.
MEMORY_MAX_COUNT=10000
# DESCRIPTION: Include cross-session global memories.
# Values: true | false
MEMORY_INCLUDE_GLOBAL=true
# DESCRIPTION: How retrieved memories are injected.
# Values: system | user | assistant
MEMORY_INJECTION_FORMAT=system
# DESCRIPTION: Auto-extract memories from turns.
# Values: true | false
MEMORY_EXTRACTION_ENABLED=true
# DESCRIPTION: Apply time-based decay to memory scores.
# Values: true | false
MEMORY_DECAY_ENABLED=true
# DESCRIPTION: Half-life for memory decay (days).
MEMORY_DECAY_HALF_LIFE=30
# DESCRIPTION: Memory rendering format.
# Values: compact | verbose | json
MEMORY_FORMAT=compact
# DESCRIPTION: Dedupe memories before injection.
# Values: true | false
MEMORY_DEDUP_ENABLED=true
# DESCRIPTION: Turns of history to scan for dedup.
MEMORY_DEDUP_LOOKBACK=5

# DESCRIPTION: Track input/output tokens per request.
# Values: true | false
TOKEN_TRACKING_ENABLED=true
# DESCRIPTION: System prompt rendering strategy.
# Values: dynamic | static | minimal
SYSTEM_PROMPT_MODE=dynamic
# DESCRIPTION: How verbose tool descriptions are.
# Values: minimal | normal | verbose
TOOL_DESCRIPTIONS=minimal
# DESCRIPTION: Summarize older history turns instead of dropping them.
# Values: true | false
HISTORY_COMPRESSION_ENABLED=true
# DESCRIPTION: How many recent turns to keep verbatim.
HISTORY_KEEP_RECENT_TURNS=10
# DESCRIPTION: Summarize history older than the recent window.
# Values: true | false
HISTORY_SUMMARIZE_OLDER=true
# DESCRIPTION: Token budget that triggers a "you're approaching the limit" warning.
TOKEN_BUDGET_WARNING=100000
# DESCRIPTION: Hard token budget ceiling.
TOKEN_BUDGET_MAX=180000
# DESCRIPTION: Refuse requests over TOKEN_BUDGET_MAX instead of warning.
# Values: true | false
TOKEN_BUDGET_ENFORCEMENT=true

# DESCRIPTION: Caveman terse-output injection (cuts output tokens at the cost of style).
# Values: true | false
CAVEMAN_ENABLED=false
# DESCRIPTION: Aggressiveness of the brevity instruction.
# Values: lite | full | ultra
CAVEMAN_LEVEL=lite

# DESCRIPTION: Render markdown to ANSI for CLIs without a markdown renderer.
# Leave false for Claude Code (it renders markdown itself).
# Values: true | false
MARKDOWN_RENDER_ANSI=false


# ==============================================================================
# 7. POLICY & SAFETY
# ==============================================================================

# DESCRIPTION: Hard cap on routing/tool-call steps per request.
POLICY_MAX_STEPS=20
# DESCRIPTION: Hard cap on tool calls per request.
POLICY_MAX_TOOL_CALLS=12
# DESCRIPTION: Force-terminate after this many same-tool calls in a row (loop guard).
POLICY_TOOL_LOOP_THRESHOLD=10
# DESCRIPTION: Comma-separated tool names that are never allowed.
# POLICY_DISALLOWED_TOOLS=dangerous_tool1,dangerous_tool2

# DESCRIPTION: Allow `git push`.
# Values: true | false
POLICY_GIT_ALLOW_PUSH=false
# DESCRIPTION: Allow `git pull`.
# Values: true | false
POLICY_GIT_ALLOW_PULL=true
# DESCRIPTION: Allow `git commit`.
# Values: true | false
POLICY_GIT_ALLOW_COMMIT=true
# DESCRIPTION: Test command run before allowing a commit (when REQUIRE_TESTS=true).
# POLICY_GIT_TEST_COMMAND=npm test
# DESCRIPTION: Refuse commits unless POLICY_GIT_TEST_COMMAND passes.
# Values: true | false
POLICY_GIT_REQUIRE_TESTS=false
# DESCRIPTION: Regex that commit messages must match.
# POLICY_GIT_COMMIT_REGEX=^(feat|fix|docs|style|refactor|test|chore):
# DESCRIPTION: Auto-stash uncommitted changes before risky git operations.
# Values: true | false
POLICY_GIT_AUTOSTASH=false

# DESCRIPTION: Comma-separated paths that file tools may touch (allowlist).
# POLICY_FILE_ALLOWED_PATHS=/path1,/path2
# DESCRIPTION: Comma-separated paths that file tools may NOT touch.
POLICY_FILE_BLOCKED_PATHS=/.env,.env,/etc/passwd,/etc/shadow

# DESCRIPTION: Apply the safe-commands allowlist to bash tool calls.
# Values: true | false
POLICY_SAFE_COMMANDS_ENABLED=true
# DESCRIPTION: JSON config for the safe-commands allowlist.
# POLICY_SAFE_COMMANDS_CONFIG={"allowed":["ls","cat","grep"]}

# DESCRIPTION: Master switch for the security content filter.
# Values: true | false
SECURITY_CONTENT_FILTER_ENABLED=true
# DESCRIPTION: Block requests when the filter triggers (vs. just log).
# Values: true | false
SECURITY_BLOCK_ON_DETECTION=true
# DESCRIPTION: Master switch for the security rate limiter.
# Values: true | false
SECURITY_RATE_LIMIT_ENABLED=true
# DESCRIPTION: Per-IP request cap per minute.
SECURITY_PER_IP_LIMIT=100
# DESCRIPTION: Per-endpoint request cap per minute.
SECURITY_PER_ENDPOINT_LIMIT=1000
# DESCRIPTION: Persist security events to disk.
# Values: true | false
SECURITY_AUDIT_LOG_ENABLED=true
# DESCRIPTION: Directory for the security audit log.
SECURITY_AUDIT_LOG_DIR=./logs


# ==============================================================================
# 8. AGENTS
# ==============================================================================

# DESCRIPTION: Inject agent-delegation instructions into the system prompt
# when the client declares a Task tool (tools always execute on the client).
# Values: true | false
AGENTS_ENABLED=true

# ==============================================================================
# 9. RATE LIMITING & BUDGETS
# ==============================================================================

# DESCRIPTION: Master switch for the per-session rate limiter.
# Values: true | false
RATE_LIMIT_ENABLED=true
# DESCRIPTION: Sliding window length (ms).
RATE_LIMIT_WINDOW_MS=60000
# DESCRIPTION: Max requests per window.
RATE_LIMIT_MAX=100
# DESCRIPTION: How requests are bucketed.
# Values: session | ip | both
RATE_LIMIT_KEY_BY=session


# ==============================================================================
# 10. WEB TOOLS
# ==============================================================================

# DESCRIPTION: Endpoint for the WebSearch tool (often a local SearXNG instance).
WEB_SEARCH_ENDPOINT=http://localhost:8888/search
# DESCRIPTION: Bearer token for the search endpoint.
# WEB_SEARCH_API_KEY=your-search-key
# DESCRIPTION: Allow searching any host (overrides the allowlist).
# Values: true | false
WEB_SEARCH_ALLOW_ALL=true
# DESCRIPTION: Comma-separated allowlist of search hosts.
# WEB_SEARCH_ALLOWED_HOSTS=localhost,127.0.0.1
# DESCRIPTION: Per-search timeout (ms).
WEB_SEARCH_TIMEOUT_MS=10000
# DESCRIPTION: Retry failed searches.
# Values: true | false
WEB_SEARCH_RETRY_ENABLED=true
# DESCRIPTION: Max retry attempts on failure.
WEB_SEARCH_MAX_RETRIES=2
# DESCRIPTION: Max bytes of page body shown in WebFetch results.
WEB_FETCH_BODY_PREVIEW_MAX=10000

# DESCRIPTION: TinyFish.ai API key (enables the WebAgent browser-automation tool).
# TINYFISH_API_KEY=sk-tinyfish-your-key
# DESCRIPTION: TinyFish automation endpoint.
TINYFISH_ENDPOINT=https://agent.tinyfish.ai/v1/automation/run-sse
# DESCRIPTION: Browser profile preset.
# Values: lite | standard | stealth
TINYFISH_BROWSER_PROFILE=lite
# DESCRIPTION: Per-run timeout (ms).
TINYFISH_TIMEOUT_MS=120000
# DESCRIPTION: Route the browser through a residential proxy.
# Values: true | false
TINYFISH_PROXY_ENABLED=false
# DESCRIPTION: Proxy egress country (ISO 3166 alpha-2).
TINYFISH_PROXY_COUNTRY=US


# ==============================================================================
# 11. WORKSPACE / TEST RUNNER
# ==============================================================================

# DESCRIPTION: Test command the workspace runner invokes.
# WORKSPACE_TEST_COMMAND=npm test
# DESCRIPTION: Extra args appended to the test command.
# WORKSPACE_TEST_ARGS=--coverage
# DESCRIPTION: Test timeout (ms).
WORKSPACE_TEST_TIMEOUT_MS=600000
# DESCRIPTION: How tests are sandboxed.
# Values: auto | docker | none
WORKSPACE_TEST_SANDBOX=auto
# DESCRIPTION: Coverage report paths the runner picks up.
WORKSPACE_TEST_COVERAGE_FILES=coverage/coverage-summary.json
# DESCRIPTION: JSON array of named test profiles.
# WORKSPACE_TEST_PROFILES=[{"name":"unit","command":"npm test"}]

# DESCRIPTION: Storage path for the Files tool.
FILES_STORAGE_PATH=./data/files
# DESCRIPTION: Max stored files (LRU cap).
FILES_MAX_COUNT=1000
# DESCRIPTION: Max per-file size in MB.
FILES_MAX_SIZE_MB=100


# ==============================================================================
# 12. OBSERVABILITY (audit, error logs)
# ==============================================================================

# DESCRIPTION: Master switch for the LLM audit log.
# Values: true | false
LLM_AUDIT_ENABLED=false
# DESCRIPTION: Audit log file path.
# LLM_AUDIT_LOG_FILE=./logs/llm-audit.log
# DESCRIPTION: Include audit annotations on each entry.
# Values: true | false
LLM_AUDIT_ANNOTATIONS=true
# DESCRIPTION: Max system-prompt chars retained per audit entry.
LLM_AUDIT_MAX_SYSTEM_LENGTH=2000
# DESCRIPTION: Max user-message chars retained per audit entry.
LLM_AUDIT_MAX_USER_LENGTH=3000
# DESCRIPTION: Max response chars retained per audit entry.
LLM_AUDIT_MAX_RESPONSE_LENGTH=3000
# DESCRIPTION: Legacy fallback for max content length.
LLM_AUDIT_MAX_CONTENT_LENGTH=5000
# DESCRIPTION: Rotated audit file count.
LLM_AUDIT_MAX_FILES=30
# DESCRIPTION: Per-file rotation size.
LLM_AUDIT_MAX_SIZE=100M
# DESCRIPTION: Deduplicate repeated audit payloads via dictionary compression.
# Values: true | false
LLM_AUDIT_DEDUP_ENABLED=true
# DESCRIPTION: Path to the dedup dictionary file.
# LLM_AUDIT_DEDUP_DICT_PATH=./logs/llm-audit-dictionary.jsonl
# DESCRIPTION: LRU cache size for the dedup dictionary.
LLM_AUDIT_DEDUP_CACHE_SIZE=100
# DESCRIPTION: Smallest payload size (bytes) eligible for dedup.
LLM_AUDIT_DEDUP_MIN_SIZE=500
# DESCRIPTION: Sanitize secrets out of dedup-eligible payloads.
# Values: true | false
LLM_AUDIT_DEDUP_SANITIZE=true
# DESCRIPTION: Cache dedup state per session.
# Values: true | false
LLM_AUDIT_DEDUP_SESSION_CACHE=true

# DESCRIPTION: Persist oversized-payload errors to disk.
# Values: true | false
OVERSIZED_ERROR_LOGGING_ENABLED=true
# DESCRIPTION: Bytes above which a payload is considered "oversized".
OVERSIZED_ERROR_THRESHOLD=200
# DESCRIPTION: Where oversized-error dumps go.
OVERSIZED_ERROR_LOG_DIR=./logs/oversized-errors
# DESCRIPTION: Max retained oversized-error dump files.
OVERSIZED_ERROR_MAX_FILES=100


# ==============================================================================
# 13. HOT RELOAD
# ==============================================================================

# DESCRIPTION: Hot-reload config when .env changes.
# Values: true | false
HOT_RELOAD_ENABLED=true
# DESCRIPTION: Debounce window for the reload watcher (ms).
HOT_RELOAD_DEBOUNCE_MS=1000


# ==============================================================================
# 14. CLUSTERING & LOAD SHEDDING
# ==============================================================================

# DESCRIPTION: Run the proxy in multi-worker cluster mode.
# Values: true | false
CLUSTER_ENABLED=false
# DESCRIPTION: Worker count.
# Values: auto | <integer>
CLUSTER_WORKERS=auto

# DESCRIPTION: Heap-utilization fraction above which load shedding triggers.
LOAD_SHEDDING_HEAP_THRESHOLD=0.95
# DESCRIPTION: RSS-memory fraction above which load shedding triggers.
LOAD_SHEDDING_MEMORY_THRESHOLD=0.85
# DESCRIPTION: In-flight request count above which load shedding triggers.
LOAD_SHEDDING_ACTIVE_REQUESTS_THRESHOLD=1000

# DESCRIPTION: Master switch for the worker thread pool (heavy parsing, embeddings).
# Values: true | false
WORKER_POOL_ENABLED=true
# DESCRIPTION: Worker pool size. 0 = auto (CPU cores - 1).
WORKER_POOL_SIZE=0
# DESCRIPTION: Per-task timeout (ms).
WORKER_TASK_TIMEOUT_MS=5000
# DESCRIPTION: Payload size (bytes) above which work is offloaded to the pool.
WORKER_OFFLOAD_THRESHOLD_BYTES=10000
# DESCRIPTION: Large-payload optimization (chunked encoding, streaming).
# Values: true | false
LARGE_PAYLOAD_OPTIMIZATION=true
# DESCRIPTION: Bytes that count as a "large" payload.
LARGE_PAYLOAD_THRESHOLD=1048576


# ==============================================================================
# 15. OPTIONAL FEATURES
# ==============================================================================

# DESCRIPTION: Master switch for the code-graph indexer.
# Values: true | false
CODE_GRAPH_ENABLED=false
# DESCRIPTION: External indexer binary the code-graph feature shells out to.
CODE_GRAPH_COMMAND=graphify
# DESCRIPTION: Workspace path the indexer operates on (defaults to cwd).
# CODE_GRAPH_WORKSPACE=/path/to/repo
# DESCRIPTION: Per-call timeout (ms).
CODE_GRAPH_TIMEOUT=10000