.env.example

# Environment
# The agent CLI (`pnpm start`), `pnpm test:e2e` and `pnpm benchmark:case:add` auto-load this file:
# generic keys enter the process environment, already-exported values win.

# GitHub (required only for commands that read GitHub issues or post comments)
GITHUB_TOKEN=your_github_token_here
# Repository identity — nothing is defaulted. The explicit pair wins, then GITHUB_REPOSITORY,
# then the origin remote of the REPOSITORY_PATH checkout below. GitHub-reading commands fail
# naming these variables when no source provides the identity.
# GITHUB_REPO_OWNER=your_repo_owner_here
# GITHUB_REPO_NAME=your_repo_name_here
# GITHUB_REPOSITORY=owner/repo

# LLM provider: OpenAI chat-completions or Anthropic Messages API. The base URL, key, reasoning
# model and token limits are required; a missing one fails the run naming the variable. What each
# value must be, and how to run the benchmark with it: docs/guides/benchmark.md.
# Optional API for both model roles: openai-completions (default) | anthropic-messages.
# LLM_API=openai-completions
# For Anthropic, use these settings (the base URL has no /v1):
# LLM_API=anthropic-messages
# LLM_BASE_URL=https://api.anthropic.com
# Keep the key on the host; it is never written to issue or run artifacts.
LLM_BASE_URL=https://provider.example/v1
LLM_API_KEY=your_provider_key_here
LLM_MODEL=your_reasoning_model_here
# Optional: defaults to the effective reasoning model, including --model overrides.
# When unset, that reasoning model must also accept screenshot images.
# LLM_VISION_MODEL=your_vision_model_here
# The models' own limits: a value above the provider's limit fails every request with HTTP 400.
# The context window is registered for both models; LLM_VISION_MAX_OUTPUT_TOKENS may be left out
# when LLM_VISION_MODEL is unset or both model variables name the same model.
LLM_CONTEXT_WINDOW_TOKENS=131072
LLM_MAX_OUTPUT_TOKENS=32768
# LLM_VISION_MAX_OUTPUT_TOKENS=8192
# Optional: a JSON price file (format: docs/guides/benchmark.md#count-the-cost). Unset, runs
# proceed and their cost is reported as not computed.
# LLM_PRICING_FILE=prices.json

# The AdGuard CLI module (`--blocker-module adguard-cli/blocker-module.json`): the official AdGuard
# CLI release (https://github.com/AdguardTeam/AdGuardCLI/releases) and the licence it is activated
# with for each run. Without the licence a run that plugs the module in stops before it starts. A value here
# is the plain key: an `op://` reference in this file is not resolved and would be sent to
# `activate` as is. To keep the key out of files, leave it empty and pass it on the command line
# instead, for example through the 1Password CLI:
#   ADGUARD_LICENSE_KEY='op://<vault>/<item>/<field>' op run -- pnpm start ...
ADGUARD_CLI_PATH=/path/to/adguard-cli
ADGUARD_LICENSE_KEY=
# Optional: a seeded CLI home archive (scripts/seed-adguard-cli-home.sh) the run restores instead of
# activating; the licence key stays the fallback when the seed's licence is no longer active.
# ADGUARD_CLI_HOME_ARCHIVE=/path/to/adguard-cli-home.tar.gz

# Optional LLM request tuning.
# LLM_REQUEST_TIMEOUT_MS=300000
# LLM_REQUEST_MAX_ATTEMPTS=2
# Thinking effort for loop requests on either API: off | minimal | low | medium | high.
# On Anthropic, loop off registers a non-reasoning entry and omits thinking; a model that
# thinks by default keeps thinking. Other levels map to the configured API.
# LLM_REASONING_EFFORT=high
# Thinking effort of single-shot calls (issue extraction, vision verdicts), same values.
# On Anthropic with loop effort on, single-shot off sends thinking.type=disabled on reasoning
# entries. Claude Opus 5.5 and Sonnet 5.5 reject it with 400; Haiku 5.5 accepts it.
# LLM_SINGLE_SHOT_REASONING_EFFORT=low

# Browser
HEADLESS=true
# Pass --no-sandbox to Chromium, for a container that cannot provide the sandbox.
# NO_SANDBOX=false
# CloakBrowser binary auto-update. The CLI and benchmark default to false for one Chromium build.
# Explicit values from the environment or .env are preserved.
# CLOAKBROWSER_AUTO_UPDATE=false

# Local AdguardFilters checkout (used by repo-context slices)
# REPOSITORY_PATH=/path/to/AdguardFilters