feat: optionally split output arguments into their own file [ROBO-5900] - #159

Why

A consumer that only forwards the output arguments still has to materialize them. In Serverless the handler reads output.json, parses output into a JToken, stringifies it, converts to byte[], then uploads — roughly 4x the payload in the .NET heap for a byte-for-byte passthrough. At ~80 MB that is the difference between a job completing and an OOM.

This lets the runtime write the arguments to their own file and report only a path, so the consumer can stream it. Producer side only; the consumer change is in hdens.

What

One key, runtime.splitOutputArguments (bool, default false), in uipath.json. When set, __exit__ writes the payload next to the result file and replaces the envelope's output with an absolute outputArgumentsFilePath.

The host chooses neither the path nor the name. Both are derived from the result file, which it already named once via runtime.dir: the directory is the same, and the name inserts .args before the extension (output.json → output.args.json). So the knob has exactly one encoding, the two files cannot land in different directories, and they cannot collide — a result file named output.args.json yields output.args.args.json.

Contract

status / error / resume / resumeTriggers always stay inline, so faults and suspends still work. Consumers must treat the pointer as optional — absent by default and from older runtimes.

Worth reviewing closely:

• With the knob off the block is skipped and content is untouched; the only change on that path is hoisting output_payload above the write.
• The write is gated on job_id like the envelope write below it. --output-file is written job or no job because the caller named a path; this is a modifier on the envelope, not a request for a file, so it follows the envelope.
• A failed write propagates, exactly like the result-file and --output-file writes either side of it, and the existing handler turns it into a structured RUNTIME_SHUTDOWN_ERROR. An earlier revision degraded to inline instead — dropped, because it writes the same bytes to the same volume (so it cannot rescue a full disk, only defer it) and it would hand the consumer the payload this feature exists to keep out of its heap.

Testing

425 passed; ruff, ruff-format and mypy clean. Covers:

• knob off → result file byte-identical to the legacy envelope, no stray file
• knob on → arguments in the sibling file, envelope carries the absolute pointer, no output key
• the arguments file is a sibling of the result file and never in the CWD
• the name cannot collide with the result file, whatever outputFile is called
• nothing is written when there is no job_id
• --output-file still receives the real arguments when both are set
• resume / resumeTriggers stay inline
• a faulted run keeps status/error inline and the advertised file exists
• a failing write faults the run with RUNTIME_SHUTDOWN_ERROR

Five mutants that previously passed now fail: popping resume alongside output; skipping the write for an empty payload while still advertising the pointer; resolving the path against the CWD; reverting to a fixed filename; and dropping the job_id gate.

Also here: one unrelated CI fix

29b297c is the oldest commit and touches only .github/scripts/force-runtime-override.py, so it reads and reverts independently.

The script joined the existing override-dependencies items with the injected wheel entry but stripped only whitespace, so a multi-line array's trailing comma produced [.., , ..] and uv rejected the file. That is what failed langchain-cross here. The bug is latent on main — test-uipath.yml was skipped on every other recent branch, so this PR is just the first to run it.

Follow-ups (separate PRs, in order)

1. uipath-python raises its uipath-runtime floor to this version.
2. uipath-agents-python raises three ceilings that currently exclude it.
3. hdens consumes the pointer and streams the file into the blob upload (streaming from the POD filesystem through the HTTP Request all the way to Orchestrator)