# =============================================================================
# Decepticon — Environment Variables
# =============================================================================
# The installer copies this file to ~/.decepticon/.env
# Run `decepticon onboard` to (re)configure interactively.

# --- LLM API Keys (configure at least one credential below) ---
# Each entry corresponds to one AuthMethod. The onboard wizard rewrites
# the keys you select; unselected keys keep their placeholder values and
# are ignored at runtime.
ANTHROPIC_API_KEY=your-anthropic-key-here
OPENAI_API_KEY=your-openai-key-here
GEMINI_API_KEY=your-gemini-key-here
MINIMAX_API_KEY=your-minimax-key-here
OPENROUTER_API_KEY=your-openrouter-key-here
NVIDIA_API_KEY=your-nvidia-key-here
DEEPSEEK_API_KEY=your-deepseek-key-here
XAI_API_KEY=your-xai-key-here
MISTRAL_API_KEY=your-mistral-key-here
# Voyage AI — embeddings for skillogy hybrid retrieval (ADR-0011). Anthropic's
# official embedding partner, so an Anthropic-key deployment can power skillogy
# semantics with this alone (set DECEPTICON_SKILLOGY_EMBED_MODEL=voyage/voyage-3).
VOYAGE_API_KEY=your-voyage-key-here

# --- Cloud Gateways (added in the OpenClaude provider migration) ---
# All routed through LiteLLM's native provider implementations. The
# onboard wizard rewrites only the keys you explicitly select; placeholder
# values are skipped at credentials-resolution time.
GROQ_API_KEY=your-groq-key-here
TOGETHER_API_KEY=your-together-key-here
FIREWORKS_API_KEY=your-fireworks-key-here
COHERE_API_KEY=your-cohere-key-here
MOONSHOT_API_KEY=your-moonshot-key-here
# Kimi for Coding (api.kimi.com/coding), K3 / K2.7. Separate platform
# from the legacy Moonshot API above; keys start with sk-kimi-.
KIMI_API_KEY=your-kimi-key-here
ZAI_API_KEY=your-zai-key-here
DASHSCOPE_API_KEY=your-dashscope-key-here

# Optional current-model routes mirrored from the production LiteLLM catalog.
# The proxy registers each route only when its provider key is configured.
CEREBRAS_API_KEY=your-cerebras-key-here
CHUTES_API_KEY=your-chutes-key-here
DEEPINFRA_API_KEY=your-deepinfra-key-here
NEBIUS_API_KEY=your-nebius-key-here
NOVITA_API_KEY=your-novita-key-here
SAMBANOVA_API_KEY=your-sambanova-key-here
SILICONFLOW_API_KEY=your-siliconflow-key-here
UPSTAGE_API_KEY=your-upstage-key-here
REQUESTY_API_KEY=your-requesty-key-here
KILO_API_KEY=your-kilo-key-here
OLLAMA_CLOUD_API_KEY=your-ollama-cloud-key-here
# RUNYOUR_API_BASE=https://your-runyour-endpoint/v1
RUNYOUR_API_KEY=your-runyour-key-here

# GitHub Models — fine-grained PAT with the 'models' permission scope.
# Generate at github.com/settings/personal-access-tokens.
GITHUB_TOKEN=your-github-token-here

# --- OpenAI-compatible gateways / aggregators (oh-my-pi parity) ---
# Each routes through LiteLLM's openai/ provider with a fixed api_base
# override (see config/litellm.yaml). Set only the gateways you use; a
# placeholder key is skipped at credentials-resolution time. Order in the
# default fallback chain follows _DEFAULT_AUTH_PRIORITY (factory.py).
OPENCODE_API_KEY=your-opencode-zen-key-here          # opencode.ai/zen
# Vercel names this AI_GATEWAY_API_KEY; the Decepticon-namespaced alias is
# passed explicitly to LiteLLM, so set THIS var (not Vercel's bare name).
VERCEL_AI_GATEWAY_API_KEY=your-vercel-ai-gateway-key-here  # vercel.com/ai-gateway
HF_TOKEN=your-hf-token-here                           # huggingface.co/settings/tokens
VENICE_API_KEY=your-venice-key-here                   # venice.ai/settings/api
NANOGPT_API_KEY=your-nanogpt-key-here                 # nano-gpt.com
SYNTHETIC_API_KEY=your-synthetic-key-here             # synthetic.new
ZENMUX_API_KEY=your-zenmux-key-here                   # zenmux.ai
QIANFAN_API_KEY=your-qianfan-key-here                 # qianfan.baidubce.com (Baidu ERNIE)
# Cloudflare AI Gateway is per-account: set BOTH the key and the
# OpenAI-compat base URL (the gateway's `…/compat` path). A key without
# the base URL 404s at call time.
CLOUDFLARE_AI_GATEWAY_API_KEY=your-cloudflare-ai-gateway-key-here
# CLOUDFLARE_AI_GATEWAY_API_BASE=https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/compat
CHEAPER_INFERENCE_API_KEY=your-cheaper-inference-key-here  # cheaperinference.com (keys start ci_live_)

# --- Generic OpenAI-compatible gateway (opt-in, DEFAULT OFF) ---
# Point at ANY OpenAI-compatible /v1 endpoint. No vendor is hardcoded.
# Requires base URL + real key + at least one model. Uncomment to enable.
# See issue #989 (vendor-neutral alternative to a named third-party entry).
# OPENAI_COMPAT_GATEWAY_API_BASE=https://gateway.example.com/v1
# OPENAI_COMPAT_GATEWAY_API_KEY=
# OPENAI_COMPAT_GATEWAY_MODEL=gpt-4o-mini
# Optional per-tier overrides (fall back to OPENAI_COMPAT_GATEWAY_MODEL):
# OPENAI_COMPAT_GATEWAY_MODEL_HIGH=
# OPENAI_COMPAT_GATEWAY_MODEL_MID=
# OPENAI_COMPAT_GATEWAY_MODEL_LOW=

# --- AWS Bedrock ---
# IAM access key + secret + region. The region must have your chosen
# Anthropic models enabled (us-east-1 / us-west-2 are typical).
# AWS_ACCESS_KEY_ID=AKIA...
# AWS_SECRET_ACCESS_KEY=...
# AWS_REGION=us-east-1

# --- GCP Vertex AI ---
# Service-account JSON path + project + region. Download a key from
# IAM & Admin → Service Accounts. Project + location must match where
# Anthropic / Gemini models are enabled.
# GOOGLE_APPLICATION_CREDENTIALS=/abs/path/to/service-account.json
# VERTEXAI_PROJECT=my-gcp-project-id
# VERTEXAI_LOCATION=us-central1

# --- Azure OpenAI Service ---
# Deployment endpoint + key + API version. Deployment names default to
# the model id (gpt-5.5/gpt-5.4/gpt-5-nano); override per role via
# DECEPTICON_MODEL_<ROLE> if your Azure deployment names differ.
# AZURE_API_KEY=...
# AZURE_API_BASE=https://<resource>.openai.azure.com
# AZURE_API_VERSION=2024-08-01-preview

# --- LM Studio (local OpenAI-compatible) ---
# LM Studio runs a local OpenAI-compatible server on port 1234.
# Launch the LM Studio app → Developer tab → Start Server.
# LMSTUDIO_API_BASE=http://host.docker.internal:1234/v1
# LMSTUDIO_MODEL=qwen2.5-coder-7b-instruct
# LMSTUDIO_API_KEY=lm-studio  # accepts any string; sentinel for the shim

# --- llama.cpp (local OpenAI-compatible) ---
# llama.cpp's `llama-server` exposes an OpenAI-compatible REST API
# (default http://localhost:8080/v1). Run a GGUF model with:
#   llama-server -m /path/to/model.gguf --port 8080
# The route prefix is `llamacpp/` and remaps to LiteLLM's openai-
# compatible path with LLAMACPP_API_BASE. See issue #151.
# LLAMACPP_API_BASE=http://host.docker.internal:8080/v1
# LLAMACPP_MODEL=qwen2.5-coder-7b-instruct-q4_k_m
# LLAMACPP_API_KEY=llama-cpp  # accepts any string; sentinel for the shim

# --- Custom OpenAI-compatible Endpoint ---
# Bring-your-own gateway (vLLM, internal LiteLLM, third-party hub, etc.).
# CUSTOM_OPENAI_API_BASE=https://gateway.example.com/v1
# CUSTOM_OPENAI_API_KEY=...
# CUSTOM_OPENAI_MODEL=gpt-4o-mini

# --- Self-hosted LLM (vLLM) ---
# For running your own model on your own GPU. Set both and the proxy
# registers hosted_vllm/<model> at startup — the same shape as OLLAMA_MODEL
# below, so you don't need to know the LiteLLM provider prefix.
#
# VLLM_MODEL must match the server's --served-model-name, NOT the HuggingFace
# repo. vLLM answers on the served name only, and a mismatch 404s in a way
# that reads like a routing bug rather than a typo.
# VLLM_API_BASE=http://host.docker.internal:8000/v1
# VLLM_MODEL=qwen3-coder-30b
# VLLM_API_KEY=  # only if you started vLLM with --api-key
#
# Serving notes, learned the hard way on a rented Ampere GPU:
#
#   --enable-auto-tool-choice --tool-call-parser <parser>  is mandatory.
#     Without the right parser vLLM returns tool calls as plain text, the
#     agent calls nothing, and the run fails with no error anywhere. Pick the
#     parser by reading the model's chat_template: a <tool_call>{"name":...}
#     JSON body is `hermes`, while <function=name><parameter=k> XML is
#     `qwen3_coder` / `qwen3_xml`. Guessing costs a full model load to find out.
#
#   Prefer a pure-transformer model. vLLM disables automatic prefix caching on
#     Mamba/linear-attention hybrids, and this agent resends a large prompt
#     every turn — losing the cache costs far more than the architecture wins.
#     Check the startup log for `enable_prefix_caching=True`.
#
#   MoE beats dense at equal VRAM. A 30B-A3B (3B active) measured ~4x the
#     generation speed of a 27B dense on the same card, and its smaller weights
#     leave several times more room for KV cache.
#
#   VLLM_USE_FLASHINFER_SAMPLER=0 on hosts with the NVIDIA driver but no CUDA
#     toolkit. The flashinfer sampler JIT-compiles at startup and dies with
#     "Could not find nvcc"; attention is unaffected, so disabling just the
#     sampler is cheaper than installing the toolkit.

# --- Local LLM (Ollama) ---
# Set both to run Decepticon against a local Ollama model. The
# launcher's onboard wizard writes these when you pick "Local LLM".
# OLLAMA_MODEL is any tag you pulled (`ollama pull qwen3-coder:30b`);
# the proxy registers ollama_chat/<model> at startup, routing through
# Ollama's /api/chat endpoint so tool calling keeps working.
#
# OLLAMA_API_BASE points the LiteLLM container at your Ollama server.
# The right URL depends on where Ollama runs:
#   - macOS / Docker Desktop on Windows / Linux Docker:
#       http://host.docker.internal:11434
#       (the litellm service has extra_hosts: host.docker.internal:host-gateway)
#   - WSL2 with Ollama installed on the Windows host:
#       http://host.docker.internal:11434  (Docker Desktop wires this up)
#   - WSL2 with Ollama installed inside the WSL distro itself:
#       http://localhost:11434
#       (or http://<wsl-ip>:11434 from `ip addr` inside WSL if Docker
#        runs natively in WSL without Docker Desktop)
#   - Remote Ollama on another host:
#       http://<host-or-ip>:11434
# OLLAMA_API_BASE=http://host.docker.internal:11434
# OLLAMA_MODEL=qwen3-coder:30b

# --- LiteLLM Proxy ---
# WARNING: the values below are dev fallbacks. `decepticon onboard` replaces
# them with strong random values; `decepticon start` refuses to use them.
# To generate values manually:
#   openssl rand -hex 32 | sed 's/^/sk-/'   # LITELLM_MASTER_KEY
#   openssl rand -hex 24                    # POSTGRES_PASSWORD
LITELLM_MASTER_KEY=sk-decepticon-master
LITELLM_SALT_KEY=sk-decepticon-salt-change-me

# --- PostgreSQL ---
POSTGRES_PASSWORD=decepticon

# --- Neo4j (Attack Chain Graph DB) ---
# Used when DECEPTICON_KG_BACKEND=neo4j.
#
# SECURITY: Neo4j is dual-homed across sandbox-net and decepticon-net.
# The default password below is PUBLIC KNOWLEDGE and MUST be changed
# for any deployment that isn't a single-operator dev box on loopback.
# Generate a strong one with `openssl rand -base64 32`. See
# docs/security/neo4j-hardening.md for the full hardening posture.
NEO4J_PASSWORD=decepticon-graph

# Skillogy hybrid retrieval (ADR-0011) embedding model, routed through litellm.
# Default is openai/text-embedding-3-small (needs OPENAI_API_KEY). Switch to
# voyage/voyage-3 (VOYAGE_API_KEY) or gemini/gemini-embedding-001 (GEMINI_API_KEY)
# for non-OpenAI deployments. With the default OpenAI model and no
# OPENAI_API_KEY, Skillogy skips vector ingest and uses lexical search
# immediately; it no longer retries unusable embedding requests. A valid
# OpenAI key retains semantic search. Explicit Voyage/Gemini/custom model
# settings keep their existing proxy routing and fallback behavior.
# DECEPTICON_SKILLOGY_EMBED_MODEL=openai/text-embedding-3-small

# --- Model Profile ---
# eco  = per-agent tier (production default)
# max  = every agent on HIGH (high-value targets)
# test = every agent on LOW (development / CI)
DECEPTICON_MODEL_PROFILE=eco

# --- Auth Method Priority ---
# Comma-separated list of AuthMethods, ordered by user preference. The
# first method supplies primary models for every tier; the rest are
# fallbacks. Methods whose credential isn't configured (placeholder key
# or DECEPTICON_AUTH_CLAUDE_CODE=false) are skipped at runtime.
#
# Valid methods: anthropic_api, anthropic_oauth, openai_api, openai_oauth,
#   google_api, google_oauth, minimax_api, deepseek_api, xai_api,
#   mistral_api, openrouter_api, nvidia_api, copilot_oauth, grok_oauth,
#   perplexity_oauth, bedrock_api, vertex_api, azure_api, groq_api,
#   together_api, fireworks_api, cohere_api, moonshot_api, zai_api,
#   dashscope_api, github_models_api, lmstudio_local, ollama_local,
#   ollama_cloud, custom_openai_api
# Empty value → factory walks _DEFAULT_AUTH_PRIORITY in priority order,
# keeping only methods whose credential is actually configured.
DECEPTICON_AUTH_PRIORITY=

# --- Claude Code OAuth (subscription) ---
# Set true to route Anthropic models through the Claude Code OAuth handler
# (auth/* in LiteLLM) instead of the x-api-key path. Requires a valid
# ~/.claude/.credentials.json from `claude` CLI auth.
DECEPTICON_AUTH_CLAUDE_CODE=false

# --- ChatGPT OAuth (subscription) ---
# Set true to route OpenAI models through the ChatGPT subscription handler
# (auth/gpt-* in LiteLLM) instead of the API-key path. Requires one of:
#   - CHATGPT_SESSION_TOKEN below (browser cookie), OR
#   - CHATGPT_ACCESS_TOKEN env var (pre-extracted Bearer), OR
#   - ~/.config/chatgpt/tokens.json file
DECEPTICON_AUTH_CHATGPT=false

# Session token from a signed-in chatgpt.com browser session:
#   1. Open chatgpt.com and sign in
#   2. DevTools → Application → Cookies → chatgpt.com
#   3. Copy the value of `__Secure-next-auth.session-token`
# The onboard wizard prompts for this when ChatGPT OAuth is selected.
CHATGPT_SESSION_TOKEN=

# --- Other OAuth subscriptions (set true + provide one credential each) ---
# All are selectable from the onboard wizard. Each handler walks the same
# resolution order at runtime: <X>_ACCESS_TOKEN → <X>_SESSION_TOKEN /
# _SESSION_COOKIES / _REFRESH_TOKEN → ~/.config/<x>/tokens.json.

# Gemini Advanced (Google One AI Premium) — gemini-sub/* models
# DECEPTICON_AUTH_GEMINI=false
# GEMINI_SESSION_COOKIES=                # NID + __Secure-1PSID + __Secure-1PSIDTS
# GEMINI_ACCESS_TOKEN=                   # alternative: pre-extracted Bearer

# SuperGrok (X Premium+) — grok-sub/* models
# DECEPTICON_AUTH_GROK=false
# GROK_SESSION_TOKEN=                    # auth_token cookie from grok.com
# GROK_ACCESS_TOKEN=                     # alternative: pre-extracted Bearer

# GitHub Copilot Pro — copilot/* models
# DECEPTICON_AUTH_COPILOT=false
# COPILOT_REFRESH_TOKEN=                 # ghu_... or ghr_... refresh token
# COPILOT_ACCESS_TOKEN=                  # alternative: pre-extracted Bearer

# Perplexity Pro — pplx-sub/* models
# DECEPTICON_AUTH_PERPLEXITY=false
# PERPLEXITY_SESSION_TOKEN=              # next-auth.session-token cookie
# PERPLEXITY_ACCESS_TOKEN=               # alternative: pre-extracted Bearer

# --- Benchmark Mode ---
# Set to 1 to switch the LangGraph orchestrator into BENCHMARK MODE for the
# XBOW / CTF benchmark harness. Activates Rule-8/9 suspension and accepts
# challenge context (target_url, vulnerability tags, flag format, mission
# brief) via state — used by `make benchmark`. Leave empty for normal
# engagement work. Restart the LangGraph container after toggling.
BENCHMARK_MODE=

# --- LangSmith Tracing (optional) ---
# Set LANGSMITH_TRACING=true AND fill LANGSMITH_API_KEY to enable.
# The onboard wizard handles both when you opt in.
#
# LANGSMITH_API_KEY is intentionally left unset (commented out) — leaving
# a placeholder value here causes the LangChain SDK to attempt metadata
# uploads to api.smith.langchain.com and spam the langgraph container
# logs with 403 Forbidden, even when LANGSMITH_TRACING=false.
LANGSMITH_TRACING=false
# LANGSMITH_API_KEY=
LANGSMITH_PROJECT=decepticon

# --- Ports (optional) ---
# On a confirmed bind conflict, start tries port + 1 and saves the free port.
# Override host-side ports when running multiple Decepticon stacks
# side by side, or when one of the defaults clashes with another
# service on your host. Container-internal ports are unchanged.
# LANGGRAPH_PORT=2024
# LITELLM_PORT=4000
# POSTGRES_PORT=5432
# WEB_PORT=3000
# TERMINAL_PORT=3003
# NEO4J_HTTP_PORT=7474
# NEO4J_BOLT_PORT=7687
# SKILLOGY_REST_PORT=9100

# --- LangGraph dev-mode strictness ---
# By default langgraph dev runs with --allow-blocking so a sync I/O
# call in a third-party dep (e.g. deepagents' subagents.py atask path)
# downgrades blockbuster's BlockingError to a non-fatal warning.
# Set this to 1 to remove --allow-blocking and reinstate the fatal-
# on-sync behavior (useful when debugging async issues in your own code).
# LANGGRAPH_STRICT_ASYNC=1

# --- C2 Framework / Specialist workloads ---
# ADR-0006 (v1.1.8+): specialist workloads (c2-sliver, c2-havoc, ad,
# reversing, …) come up on demand through the opscontrol daemon —
# the orchestrator agent calls `ops_start("c2-sliver")` after gaining
# a foothold, `ops_start("ad")` once recon identifies a Windows
# domain, and so on. Default `decepticon start` keeps every
# specialist plane cold to minimize idle cost + attack surface.
#
# Override only when you need every profile up by hand (e.g. CI
# regression of every workload at once):
#   COMPOSE_PROFILES=cli,c2-sliver,ad,reversing

# --- Paths ---
# Install directory (used for workspace bind mount into containers).
# - Production: The installer sets this to ~/.decepticon automatically.
# - Development: `decepticon` (Python entry point) sets this to the repo root automatically.
# Do not use ~ (tilde) — Docker Compose cannot expand it.
# DECEPTICON_HOME=/home/youruser/.decepticon

# --- Advanced (optional) ---
# Startup checks for a newer release and offers Update now / Skip on a TTY.
# Skip is selected by default. Non-interactive launches only show a notice.
# Legacy AUTO_UPDATE=true/1/yes/on also require approval; no silent updates.
#   AUTO_UPDATE=false   # air-gapped / version-pinned deploys
#   AUTO_UPDATE=prompt  # same as the default
# AUTO_UPDATE=false
#
# Update channel — which release stream `update` / startup checks track.
# Defaults to latest. Both channels are final-only (Claude-Code-style soak
# model); the difference is a bake delay. See docs/update-channels.md.
#   stable  — newest final release baked >= the soak window (default 7d).
#             Opt in to the delayed channel (GHCR :stable).
#   latest  — newest final release, immediately (GHCR :latest). Default.
# DECEPTICON_CHANNEL=stable
# Soak window (days) the stable channel waits before adopting a release:
# DECEPTICON_STABLE_SOAK_DAYS=7
# DECEPTICON_DEBUG=true

# Multi-stack container naming (PR #216). Unset/empty keeps the default
# `decepticon-*` names. Set to e.g. `stack2` to get `decepticon-stack2-*`
# so two stacks can coexist on the same Docker host. The declaration here
# (even when blank) silences the compose "variable is not set" WARN.
DECEPTICON_STACK_NAME=

# --- Language ---
# Pin agent output language to an ISO 639-1 code. All agent prose output
# (menus, questions, summaries, errors) renders in this language.
# Default: en (English). Presets:
#
#   en    English           no    Norwegian         da    Danish
#   sv    Swedish           fi    Finnish           is    Icelandic
#   ko    Korean            ja    Japanese          zh    Chinese
#   zh-cn Simplified Chinese                        zh-tw Traditional Chinese
#   es    Spanish           pt    Portuguese        pt-br Brazilian Portuguese
#   de    German            fr    French            nl    Dutch
#   it    Italian           ca    Catalan           el    Greek
#   pl    Polish            cs    Czech             sk    Slovak
#   uk    Ukrainian         ro    Romanian          hr    Croatian
#   bg    Bulgarian         sr    Serbian           sl    Slovenian
#   hu    Hungarian         ru    Russian           et    Estonian
#   lv    Latvian           lt    Lithuanian
#   tr    Turkish           ar    Arabic            fa    Persian
#   he    Hebrew            hi    Hindi             bn    Bengali
#   ta    Tamil             te    Telugu            th    Thai
#   vi    Vietnamese        id    Indonesian        ms    Malay
#   tl    Filipino          sw    Swahili           af    Afrikaans
#
# Country-code aliases (auto-resolved):
#   dk → Danish    se → Swedish    jp → Japanese    cn → Chinese
#   br → Brazilian Portuguese      tw → Traditional Chinese
#
# Special:
#   wenyan   文言文 Classical Chinese prose + English technical terms
# --- Reverse Engineering ---
# REA 6.3.0 and Ghidra 12.1.4 run in the optional reversing sidecar.
# The orchestrator starts it when native analysis is needed; for manual startup,
# use COMPOSE_PROFILES=reversing.

DECEPTICON_LANGUAGE=en

# --- LLM client timeout ---
# Seconds to wait for a single LLM completion before raising
# APITimeoutError / LLMTimeoutError. Default 120s (httpx transport)
# and 600s (whole-coroutine guard in decepticon.llm.factory) are too
# short for long Opus generations with max_tokens=128000 — generations
# routinely exceed 120s mid-stream while LiteLLM proxy keeps returning
# 200 OK upstream. 600s (10min) covers worst-case extended-thinking +
# large tool_use payloads.
#
# This single knob drives both the httpx transport timeout
# (``LLMConfig.timeout`` via pydantic-settings nested env) AND the
# whole-coroutine guard (``_resolve_llm_timeout_seconds`` alias). For
# tests or one-off overrides, ``DECEPTICON_LLM_TIMEOUT_SECONDS`` takes
# precedence over this for the coroutine guard only.
DECEPTICON_LLM__TIMEOUT=600

# --- Telemetry (opt-in) ---
# Off by default. Set to `research` (masked reasoning corpus) or `basic`
# (structural stats only) to share anonymous usage with the maintainers.
# See TELEMETRY.md for exactly what is collected. DO_NOT_TRACK=1 forces off.
DECEPTICON_TELEMETRY=off

# Terminal presentation (interactive TTYs; /tui and /mouse change the current session).
# --no-alt-screen and --no-mouse override these settings. Non-TTY/dumb terminals use inline.
DECEPTICON_TUI=fullscreen
DECEPTICON_MOUSE=1

# Optional per-session CLI runtime controls (flags override environment/.env).
# Persisted in the selected thread; omitted settings survive resume.
# DECEPTICON_LLM__FALLBACK_MODELS=["openai/gpt-5.5"]  # auto/configured chain; none disables
# DECEPTICON_LLM__EFFORT=auto                       # low|medium|high|max; provider support required
# DECEPTICON_LLM__AUTOCOMPACT=auto                  # or 1024–10000000 token threshold