feat(standard): give each unit a spending cap of its own

app/nullius/runs/classify.py

@@ -81,6 +81,7 @@
 # one can place it here.
 _AGENT_EXCEPTIONS = {
     "agenttimeouterror",            # the agent spent its budget — that is a result
+    "unitspendingcaperror",         # the unit spent its own spending cap (the standard's relay): the same, in money
     "agentsafetyrefusalerror",      # the agent (or its provider) chose to refuse
     "contextwindowexceedederror",   # the agent filled its own context
     "contextlengthexceedederror",   # same, from Harbor's LLM layer
@@ -131,7 +132,7 @@
     NO_PATCH: "The agent finished without producing a patch at all.",
     AGENT_ERROR: (
         "The agent itself failed — it timed out, filled its context, truncated its "
-        "own output, refused, or its process exited non-zero. Counted against the "
+        "own output, refused, spent its own spending cap, or its process exited non-zero. Counted against the "
         "agent, because each of those is a result rather than an interruption."
     ),
     RATE_LIMITED: (

app/nullius/runs/parse.py

@@ -30,7 +30,7 @@
 # The shape this module produces. A workspace caches parsed Jobs keyed on it,
 # so bump it when a Trial or Job field is added, renamed or means something
 # new — otherwise a cache written by the old parse is read back as current.
-SCHEMA = "14"
+SCHEMA = "15"
 
 # What a rate-limit record in `agent/gateway.log` looks like. The agent's SDK
 # retries a 429 below the harness (an agent's `max_retries` reaching its
@@ -284,7 +284,8 @@ class Job:
     # recipe from when the run kept none. Empty when config.json is absent.
     launch: dict = field(default_factory=dict)
     # A run with a spending cap (a Nullius standard job's ``result.json``, ``budget``): the cap and what was
-    # spent in USD, whether the cap stopped the run, the units it cut off and those never begun. {} without one.
+    # spent in USD, whether the cap stopped the run, the units it cut off and those never begun; a cap per unit
+    # (``unit_max_usd``) and the units that reached it (``over_unit``). {} without one.
     budget: dict = field(default_factory=dict)
     # A Harbor job run with several agents or models (`-m a -m b`, or a config
     # listing several): each one's {name, model, effort}, which split_agents
@@ -837,11 +838,16 @@ def _budget(b) -> dict:
     """A job's recorded spending cap, read leniently: {} unless it names a cap; an amount that is no number is
     None (not reported), never zero."""
     num = lambda v: v if isinstance(v, (int, float)) and not isinstance(v, bool) else None
-    if not isinstance(b, dict) or not num(b.get("max_usd")):
+    if not isinstance(b, dict) or not (num(b.get("max_usd")) or num(b.get("unit_max_usd"))):
         return {}
     names = lambda v: [str(x) for x in v] if isinstance(v, list) else []
-    return {"max_usd": b["max_usd"], "spent_usd": num(b.get("spent_usd")), "unsettled_usd": num(b.get("unsettled_usd")),
-            "stopped": b.get("stopped") is True, "cut_off": names(b.get("cut_off")), "not_run": names(b.get("not_run"))}
+    return {"max_usd": num(b.get("max_usd")), "spent_usd": num(b.get("spent_usd")), "unsettled_usd": num(b.get("unsettled_usd")),
+            # only the run's cap stops a run: without one there is no stop to say
+            "stopped": b.get("stopped") is True and num(b.get("max_usd")) is not None,
+            "cut_off": names(b.get("cut_off")), "not_run": names(b.get("not_run")),
+            "unit_max_usd": num(b.get("unit_max_usd")), "over_unit": names(b.get("over_unit")),
+            # a merged job: its shards, and how many stopped at their share
+            "shards": num(b.get("shards")) or 0, "shards_stopped": num(b.get("shards_stopped")) or 0}
 
 
 def parse_job(job_dir: str, reuse=None) -> Job:

app/nullius/standard/cli.py

@@ -6,7 +6,7 @@
     PYTHONPATH=app python3 -m nullius.standard.cli new benchmark|agent|capability KEY [--family F] [--port P]
     PYTHONPATH=app python3 -m nullius.standard.cli new version KEY VERSION
     PYTHONPATH=app python3 -m nullius.standard.cli conform AGENT|CAPABILITY --model M [--provider P] [--port PORT ...] [--repeat K] [--option NAME=VALUE ...] [--unsandboxed | --backend docker]
-    PYTHONPATH=app python3 -m nullius.standard.cli run BENCHMARK[@VERSION] AGENT|CAPABILITY --model M [--provider P] [--limit N] [-j N] [--option NAME=VALUE ...] [--bench-option NAME=VALUE ...] [--price MODEL=INPUT,OUTPUT[,CACHED] ...] [--max-usd USD] [--unsandboxed | --backend docker]
+    PYTHONPATH=app python3 -m nullius.standard.cli run BENCHMARK[@VERSION] AGENT|CAPABILITY --model M [--provider P] [--limit N] [-j N] [--option NAME=VALUE ...] [--bench-option NAME=VALUE ...] [--price MODEL=INPUT,OUTPUT[,CACHED] ...] [--max-usd USD] [--max-usd-per-unit USD] [--unsandboxed | --backend docker]
     PYTHONPATH=app python3 -m nullius.standard.cli run BENCHMARK[@VERSION] nop|oracle [--dataset D] [--limit N]
     PYTHONPATH=app python3 -m nullius.standard.cli prove BENCHMARK[@VERSION] [--dataset D] [--limit N]
     PYTHONPATH=app python3 -m nullius.standard.cli run ... --shard I/N [--launch ID]
@@ -469,7 +469,8 @@ def cmd_run(a):
                       concurrency=a.concurrency, jobs_root=a.jobs, name=a.name, unsandboxed=a.unsandboxed, judge_model=a.judge_model, backend=a.backend,
                       simulator_model=a.simulator_model, options=options(a.option), shard=_shard(a.shard), launch=a.launch,
                       licences=held, bench_options=options(a.bench_option), repeat=a.repeat, remove_images=a.remove_images,
-                      prices=prices(a.price), max_usd=a.max_usd, **({"env": run_env(a.provider)} if a.from_run else {}))
+                      prices=prices(a.price), max_usd=a.max_usd, max_usd_per_unit=a.max_usd_per_unit,
+                      **({"env": run_env(a.provider)} if a.from_run else {}))
     except run.Refused as e:
         print("refused:", e)
         return 2
@@ -626,6 +627,8 @@ def main(argv=None):
                            help="this run's price for a model in USD per million tokens, over your price list (repeatable; nullius prices)")
             p.add_argument("--max-usd", type=float, metavar="USD",
                            help="the run's spending cap: no call goes out once the spend plus the calls in flight would pass it")
+            p.add_argument("--max-usd-per-unit", type=float, metavar="USD",
+                           help="each unit's own spending cap: a unit that reaches it is cut off there, and the run goes on")
     v = sub.add_parser("prove", help="prove a benchmark: its controls, its units, no leak; record what passed")
     v.add_argument("benchmark")
     v.add_argument("--dataset", help="one dataset (default: each)")