PanzerLlama/youtrack-aitrack v0.2.0b0

v0.2.0b0 — Vendor-agnostic AgentRunner backend

on GitHub

This release pivots the AI-agent backend from "single Anthropic SDK
client" to a vendor-agnostic AgentRunner Protocol. The same engine
now drives both the SDK path and a new subprocess-based path that
spawns the local Claude Code CLI per action. Workflow YAML gains a
per-action agent: field; instance config gains a registry-driven
default_agent. No changes are required to existing workflow YAML
to keep using the SDK backend — default_agent defaults to
anthropic_api.

Highlights

• 🆕 AgentRunner Protocol — one contract, two shipping backends:
  • claude_code_cli (recommended for production): spawns local
    claude -p / claude --bare -p per action; the agent inspects the
    working tree directly via its file-reading tools. Scales well on
    large PRs and on frameworks where routing is decoupled from file
    paths.
  • anthropic_api (existing path, unchanged): single SDK call per
    action with the full diff embedded inline.
• 🆕 Per-action agent: field in workflow YAML.
• 🆕 cli_agent_mode: bare | oauth config — bare mode skips local
  CLAUDE.md/hooks/plugins so daemon spawns are deterministic and
  cost-predictable.
• 🆕 yta --version / -V flag (sourced from importlib.metadata).
• 🆕 AgentRunnerError.stderr surfaces in ActionResult.error; CLI
  failures are diagnosable from the run report.
• 🆕 New CLI-style reference prompt: prompts/security_audit_cli.md.
• 📝 README, installation, configuration, workflows, and architecture
  docs reframed around the two-backend story.

Migration from 0.1.0b0

Zero-change migration if you stay on the SDK backend:

defaults:
  default_agent: anthropic_api    # this is the default — omit entirely

To switch to the CLI backend (recommended for production daemons):

defaults:
  default_agent: claude_code_cli
  cli_agent_mode: bare            # daemon-friendly

Requires claude on PATH and ANTHROPIC_API_KEY set in .env.

Verified

• 351 unit tests green
• mypy --strict clean on src/
• ruff clean
• End-to-end against YouTrack Cloud 2026.1 on a real PR: 3 parallel
  reports (security audit, pages changed, QA plan) populated three
  custom fields and fired the Audit Status: done hook
• Smoke-tested with the bundled CLI backend in bare mode against a
  real claude binary

Install

uv tool install --from git+https://github.com/PanzerLlama/youtrack-aitrack@v0.2.0b0 youtrack-aitrack

Full changelog: https://github.com/PanzerLlama/youtrack-aitrack/blob/v0.2.0b0/CHANGELOG.md