@projectpac/executor-cli

An executor that is a command line -- the shape every coding harness already has. pi and claude are two rows of this in executors.yml, not two packages: what differs between them is argv, not machinery.

What the seam needs from a subprocess is small: a prompt in, text out, a way to stop it, and a record of what happened. Everything here is about the ways a subprocess fails to be that.

• The prompt goes in on stdin by default, because that is the only route with no length limit and no quoting. file and args are there for commands that insist.
• stdout is the answer; stderr is kept beside it in the turn's artifact directory, along with the prompt, because a human reading a trace afterwards needs what the model was told and not only what it said.
• Nothing is inherited that was not named. A model's process is the least trusted thing in the node, and the parent's environment is where a token lives.
• Output past a bound is a runaway process rather than an answer, and the command is killed.
• The command runs in its own process group, so stopping it reaches the whole tree rather than a shell that already forked the real work.

What a turn cost

Every harness reports it differently, so this is per command rather than per provider -- and unconfigured means unknown, which a budget treats as free. There are two shapes, and naming both is refused because a turn has one cost:

• costPattern -- a regular expression over the command's output whose first capture is a decimal amount. claude -p --output-format json prints "total_cost_usd": 0.0123, so "total_cost_usd"\s*:\s*([0-9.]+) reads it.
• costFile -- a path inside the turn's own artifact directory, a dotted field, and how several amounts become one. pi writes message.usage.cost.total for each assistant message into the --session file it was already told to keep, and prints nothing about cost at all, so reading the file is the only way to know:

"costFile": {
  "path": "session.jsonl",
  "field": "message.usage.cost.total",
  "reduce": "sum"
}

sum because each message's usage is its own; last is for a log that records a running total instead. Either way the amount lands as minor units, rounded up, so many small turns cannot slide under a budget -- and an unreadable file leaves the cost unknown rather than failing the turn it only describes.