agent.toml

# A Salvor agent whose entire tool layer is a Python MCP server (server.py).
# Nothing here imports Salvor or knows Rust: the agent is data (model, prompt,
# budgets, one MCP server), and the tools are ordinary Python functions the
# runtime reaches over stdio. This is the v0.1 polyglot story:
# a Python developer extends a Salvor agent by writing an MCP server, with no
# bindings and no SDK.
#
# Run from the repository root: the paths below (the venv python, server.py,
# and the ledger) are all relative to the directory salvor is invoked from.
# See README.md for setup, the run, and the kill/resume walkthrough.

model = "claude-opus-4-8"

system_prompt = """
You are an expense assistant backed by three tools: `add_expense`,
`list_expenses`, and `total_by_category`.

Your input is a JSON object with one field, `entries`: a list of
natural-language expenses to log. Work in this order:

1. For each entry in `entries`, call `add_expense` exactly once, parsing the
   amount, a short category (one lowercase word like `food`, `transit`,
   `lodging`), and an optional note. Log each entry once and only once; do not
   re-log an entry you have already recorded.
2. After every entry is logged, call `total_by_category` once to get the
   per-category totals.
3. Reply with a short plain-text summary: how many expenses you logged, the
   total spent, and the per-category breakdown. Then stop. Do not call any tool
   after the summary.
"""

[llm]
# The key is read from this variable at run time and never stored in any file.
api_key_env = "DEMO_ANTHROPIC_API_KEY"

[budgets]
# Logging a handful of expenses is a short run: one model call per entry, one
# for the totals, one for the summary. These rails leave headroom while still
# parking a run that has clearly run away rather than letting it spend on.
steps = 20
tokens = 200000
cost_usd = 0.50
wall_time_seconds = 300

[pricing]
# claude-opus-4-8 list price, dollars per million tokens. Required because a
# cost budget is declared; the two together let the runtime stop the run on
# spend, not just on step count.
input_per_mtok = 5.0
output_per_mtok = 25.0

[[mcp_servers]]
# The Python server, run through the example's virtualenv interpreter so the
# `mcp` package it imports is on the path. EXPENSE_LEDGER tells the server where
# to append; that file is the durable state and the duplicate-witness the
# README's kill/resume check inspects.
command = "examples/python-tools/.venv/bin/python"
args = ["examples/python-tools/server.py"]
env = { EXPENSE_LEDGER = "examples/python-tools/ledger.jsonl" }
# The server annotates `add_expense` with `idempotentHint: true`, which Salvor's
# default mapping would read as Idempotent (auto-retry an interrupted call under
# the same key). That is wrong for this tool: it APPENDS a line, so a retry
# writes a second, duplicate expense. We know better than the hint and pin it to
# `write`. On resume, an interrupted append then surfaces for a human instead of
# being retried blind. The two read tools annotate `readOnlyHint: true` and map
# to Read on their own, so they need no override here.
effect_overrides = { add_expense = "write" }