llmff v0.1.5

Agent harness contract release. This release turns the agent-workflow direction
from guidance into a concrete subprocess contract: agents can now hand bounded
pipeline work to llmff run --run-dir, then supervise stable artifacts,
process status, and machine-readable results.

Supported Install

cargo install --git https://github.com/syndicalt/llmff --tag v0.1.5 llmff

Included Since v0.1.4

• llmff run --run-dir <dir> creates a run-scoped artifact directory with
  inspect.json, trace.jsonl, events.jsonl, checkpoint.json, and
  result.json.
• result.json records schema version, final status, exit code, manifest hash,
  artifact names, failure kind, safe failure message, and retry
  recommendation.
• run-result-v1.schema.json freezes the result artifact shape for downstream
  supervisors and conformance tooling.
• Agent harness conformance now validates successful runs and stage-execution
  failures through scripts/check-agent-harness-conformance.sh.
• docs/agent-harness-contract.md documents process semantics, stdout
  ownership, exit-code authority, artifact separation, failure kinds, retry
  posture, safe metadata handling, and adapter expectations.
• OpenAI Agents SDK and LangGraph examples show how to wrap llmff as a bounded
  subprocess tool or graph node without turning llmff into an agent framework.
• The adapters materialize run-scoped manifests and inputs from a
  {{LLMFF_INPUT_PATH}} placeholder, so concurrent agent calls do not share or
  overwrite input files.
• The Python subprocess supervisor, batch supervisor, Node.js streaming supervisor,
  agent runner adoption guide, OpenTelemetry bridge, and ecosystem
  readiness gates remain part of the release contract.
• Release preflight continues to validate the documented ecosystem readiness
  surface before tagging; release preflight is still the local release gate.

Packaged Artifacts

Release-tag CI is expected to publish:

• Linux x86_64 .tar.gz archive, .deb package, and Arch PKGBUILD plus
  llmff-0.1.5-arch.SRCINFO metadata.
• macOS Apple Silicon and Intel .tar.gz archives and unsigned .pkg
  installers.
• Windows x86_64 unsigned .zip archive and unsigned MSI installer.
• Adjacent SHA-256 checksums and llmff-0.1.5-release-trust.json.

Manual workflow dispatch keeps generated files as Actions artifacts instead of
GitHub Release assets.

Known Limitations

• This is a pipeline runner, not a native inference kernel, model converter,
  serving platform, or full agent framework.
• Windows and macOS native artifacts are unsigned in this release.
• Package-manager distribution through Homebrew, winget, Scoop, apt
  repositories, or an official AUR package remains parked until maintainers
  decide those channels are support-ready.
• Authenticode signing, Apple Developer ID signing, and notarization remain
  deferred paid distribution tracks.
• Live provider smoke jobs remain opt-in and require maintainers to configure
  secrets and runner expectations explicitly.
• --run-dir currently owns trace, events, and checkpoint paths and is not yet
  combined with batch mode.

Verification

Release verification should pass before tagging:

scripts/release-preflight.sh v0.1.5
scripts/smoke-install.sh --git https://github.com/syndicalt/llmff --tag v0.1.5