Cost
A preference a reasonable engineer might hold the other way, so it is a switch, not a rule. Set stances.cost in
config.json, or override one session with HARNESS_STANCE_COST.
Cost stance: balanced
Session effort runs medium, fan-out stays at six subagents or fewer, fast mode is off unless asked, and a task
ends with /clear, not /compact. Past 120,000 then 160,000 tokens of context the feed says to hand off and start
fresh — starting points chosen against a 200k window, not measured figures. Spend the effort dial on judgment work
and keep gathering cheap. The table also prices each role and band: class, effort, soft budget. Tier comes from the
delegation stance; this one sets how much. Cache rules: cache-hygiene.md.
Cost stance: frugal
Session effort runs low unless the task is design or adversarial review. Subagents are gatherers only, fan-out no
wider than three, and fast mode is never used. Agent teams are off. End a task with /clear, never /compact; past
80,000 then 120,000 tokens of context the feed says to hand off and start fresh — starting points chosen against a
200k window, not measured figures. Before a model call, try a deterministic command — the test, the linter, the exit
code. The table prices each role and band: class, effort, budget. Tier comes from delegation. Cache: cache-hygiene.md.
Cost stance: max
Session effort stays at the model’s default, fan-out is as wide as the task needs, fast mode is allowed, compaction
is allowed when a long task needs it, and the feed never nudges you to start a fresh session. Optimise for the
answer, not the bill. The table prices each role and band: class, effort, budget. Tier comes from delegation; this
one sets how much. cache-hygiene.md costs apply; you choose to pay them.
Other dimensions: autonomy · build-vs-buy · commits · delegation · licensing · plan-ceremony · testing · voice