Config

[user-repo]
url = "git@example.com:org/user-repo.git"
keyfile = "/path/to/deploy-key"
branch = "main"

[slurm]
# Optional. partition defaults to ""; omit account and time-limit, or leave them empty, to use Slurm defaults.
partition = ""
account = ""
time-limit = ""
extra = []
# "scheduler" passes --mem and --cpus-per-task; "metadata-only" records the requirements in the job environment only.
enforcement = "scheduler"

[agent]
# A turn uses two sandboxes. sbox-command starts both; sbox-args configure the
# tool sandbox. runner-command runs Pi in the separate runner sandbox.
sbox-command = "sbox"
sbox-args = ["--bind", "{home}", "{home}"]
runner-command = "pi"
runner-args = []
timeout-seconds = 1800
output-limit-bytes = 1048576
# Parsed for configuration compatibility; automatic age-based pruning is not currently implemented.
session-retention-days = 7
title-prompt = "Name the chat that opens with the request in this message. Reply with the title only: five words at most, no quotes, no punctuation, no explanation."

# AI providers. Each entry names the environment variable that holds its API key,
# so keys never live in this file. With no provider tables there is no model access.
[[providers]]
name = "neuralwatt"
base-url = "https://api.neuralwatt.com/v1"
api-key-env = "NEURALWATT_API_KEY"
timeout-seconds = 60

# The chat model the agent panel runs on. The backend writes it into Pi's
# models.json and settings.json before a turn and passes --model explicitly.
[models.chat]
provider = "neuralwatt"
protocol = "openai-completions"
model = "deepseek-v4.1-flash-speed"
thinking = "high"

[models.chat.pi]
name = "DeepSeek V4.1 Flash (Speed)"
contextWindow = 1000000
maxTokens = 384000
input = ["text", "image"]
reasoning = true
thinkingLevelMap = { minimal = "low", low = "low", medium = "high", high = "high", xhigh = "xhigh", max = "max" }
disabledThinkingLevels = ["off"]
# Prices per million tokens; all four are required and must be written as decimals.
cost = { input = 0.14, output = 0.28, cacheRead = 0.028, cacheWrite = 0.0 }
compat = { supportsStore = false, supportsDeveloperRole = false, supportsReasoningEffort = true, requiresReasoningContentOnAssistantMessages = true }

# Optional semantic step search. Without [search.embedding] the search box stays
# lexical; with [search.judge] the judge filters the embedding candidates.
# [search]
# limit = 10
# judge-candidates = 24
#
# [search.embedding]
# provider = "neuralwatt"
# protocol = "openai-embeddings"
# model = "qwen3-embedding-8b"
# query-template = "Instruct: Given a search typed into a research notebook, retrieve the analysis steps it refers to\nQuery: {query}"
# document-template = "{document}"
# batch-size = 128
# min-similarity = 0.4
#
# [search.judge]
# provider = "neuralwatt"
# protocol = "systemone"
# model = "clef-flash"
# instructions = "Is step {id} something the person searching for \"{query}\" wants to find?"
# min-relevance = 0.5