Skip to content

v0.10.0

Choose a tag to compare

@github-actions github-actions released this 28 Oct 03:24
· 29 commits to main since this release

Added

  • A builder sees what the last check said (#6). When the previous iteration's check failed, the builder's prompt carries a PREVIOUS CHECK RESULT section: the exit code, the failing slots, and an excerpt capped at 6,000 characters, taken from checkride's digest (volley now runs pnpm exec checkride --json --digest and archives check/digest.md), else the start of each failing slot's raw output, else the end of a command's output. A resume rebuilds it from the archive, and a check the user rejected under volley acp reads as never having run, not as failed.
  • A tool-loop builder's last steps count down (#6). The last five steps of a session, or a tenth of a longer budget, end with a note of how many are left and the command that runs the check, and the last says that only finish runs on it. When a session hits its step limit, the next prompt says so and asks for the check first.
  • --builder-max-output-tokens <n> caps a tool-loop builder's output per turn (#6). Or VOLLEY_BUILDER_MAX_OUTPUT_TOKENS. Reasoning counts, and Ollama gets it as num_predict. An ollama or lmstudio builder is capped at 16,384 by default, an openrouter one only when you set a cap. A turn cut off before it calls a tool costs that step instead of ending the session, and each iteration's summary counts those turns as builder.truncated_turns.
  • builder_max_steps_hits counts the builder sessions that ran out of steps (#6). It's on RunResult, so --json and summary.json carry it, since the warning each one prints goes to stderr, which --json silences.

Changed

  • Reads don't count against --builder-max-tool-calls (#6). read_file, search_files, and list_files are read-only, so a builder that batches them gets them all, and past the cap the next request names the calls that were dropped. Above a cap of one, step 2 of the builder's prompt no longer names the cap.
  • The builder is asked to check as it goes (#6). In a checkride workspace it's told to run pnpm exec checkride --digest after each batch of edits and to read .check/digest.md before any slot's raw output, and both harness prompts say to verify after each batch of edits, not only before finishing.
  • RunResult has builder_max_steps_hits and ResolvedConfig has builder_max_output_tokens. Both are required, so SDK code that builds either by hand needs them.

Fixed

  • A contained --worktree run keeps its work. The sandbox image has no git identity, so volley's own commits failed inside it: the salvage commit that keeps a converged run's work on volley/<run id>, the --git checkpoints, and the squash merge. They now fall back to volley <volley@volley.invalid> for whatever git lacks, and an identity you configured still wins.
  • A failing slot's raw output is read only from inside the check directory. Slot names and output files come from a summary the builder could have written, and a ../ in one could put a file from outside it in front of the critic or the next builder.

Internal

  • Built on fascicle 0.17.1, which names a turn's dropped tool calls to the model and can go on past a turn cut off at its output cap.
  • The publish workflow keeps its OIDC identity out of the job that runs dependencies. A build job verifies the tag, installs, runs the gate, builds, and packs with read-only permissions, and a separate publish job holds id-token: write behind the npm-publish approval and publishes that tarball with --ignore-scripts.
  • The tests drive the step countdown, the per-turn cap's read-only reads, and the output cap through fascicle's real tool loop, against a stand-in OpenAI server that now takes several calls a turn and can end one on length.