PanzerLlama/youtrack-aitrack v0.3.0b0

v0.3.0b0 — plan-implementation workflow, CLI-only backend

on GitHub

This release adds a second reference workflow that runs before code
exists: plan-implementation branches off, has the agent study the issue
and the code read-only, and hands you a plan to discuss. To make that
possible the engine now carries the issue's own text, workflows can create
branches and persist files or beads issues, and the CLI agent can run in
plan mode. The anthropic_api SDK backend is removed; claude_code_cli
is the sole execution path (see Changed for the one breaking default).

Added

• plan-implementation workflow (workflows/plan-implementation.yaml,
  manual). yta run <id> --workflow=plan-implementation creates the branch
  <id>-<slug-of-summary> from defaults.git_base_branch and switches to
  it, runs the agent read-only in plan mode against the issue summary and
  description, posts the plan as a YouTrack comment, then persists it in the
  project: as a beads issue (external-ref = the YouTrack id) when the repo
  has bd + .beads/, otherwise committed to docs/plans/<id>.md on the
  new branch. Prompt: prompts/implementation_plan.md.
• Issue text in the context. IssueDetails (summary, description,
  state) is fetched from YouTrack and exposed as ctx.issue_details.
  yta run now makes one GET instead of two; a failed fetch marks
  task_meta unavailable instead of failing the run.
• New action types: git_branch (idempotent create-or-switch from a
  base branch; refuses to switch over uncommitted tracked changes;
  deterministic {task_id}-{slug} naming that matches
  defaults.branch_pattern), write_file (persist an upstream action's
  text inside the repo, optionally committing just that file), bd_issue
  (create a beads issue from an upstream text, with fallback_path or a
  self-skip when beads is unavailable). New adapters:
  GitWorkspaceAdapter (write-side git) and BeadsCliClient.
• ai_report.mode: plan runs the CLI agent with --permission-mode plan (read-only; the final message is the deliverable). AgentMode is
  part of the AgentRunner Protocol.
• yta run --workflow=NAME bypasses trigger matching, so a
  trigger: manual workflow can finally be fired from the CLI. The poll /
  daemon path is unchanged. Idempotency still applies.
• yta run --show-output prints each action's text output after the
  summary table; the NOTE column shows a one-line output.note for
  successful actions ("created PROJ-12-… from dev; switched to it").
• Live progress for yta run. On an interactive terminal the command now
  shows a live region with each action's state (pending / running / ok / fail /
  skipped) and a per-second elapsed timer, so a multi-minute CLI-agent run no
  longer looks hung. Non-TTY callers (pipes, tests, daemons) are unaffected.
• Per-action wall-time is measured for every action and recorded on
  ActionResult.duration_ms; the final yta run summary gains a TIME column.
• The engine emits ProgressEvents through an optional on_progress callback
  threaded from the runtime/CLI; engine/ stays free of any renderer import.
  (Per-action tokens/cost is tracked separately as a follow-up.)

Removed

• anthropic_api SDK backend. The direct-API AgentRunner
  (AnthropicAgentRunner + AnthropicLLMClient, adapters/llm/anthropic.py)
  is dropped; claude_code_cli is now the sole execution path. The SDK path
  only inlined the git diff into the prompt, so the model had no way to
  explore the repository — the CLI agent does, via its own file/git tools.
  This completes the CLI pivot rather than reversing it (Phase 2 keeps adding
  CLI backends under the same Protocol).
• anthropic.default_model config field. Now dead — the CLI takes its
  model from the per-action model: field, falling back to the claude
  CLI's own default when omitted.
• The anthropic runtime dependency, now unused.

Changed

• BREAKING: defaults.default_agent now defaults to claude_code_cli
  (was anthropic_api). Configs that relied on the old default, or that set
  agent: anthropic_api on any action, must switch to claude_code_cli.
• anthropic.api_key is retained but is now only consumed by
  claude_code_cli in cli_agent_mode: bare; it is unused in oauth mode.
• task_meta input semantics: available when the issue text could be
  fetched (previously "always").
• Report prompts emit bold section labels instead of ## headings, so a
  report inside a YouTrack custom field no longer renders larger than the
  field's own label.
• --dry-run also disables the new git and beads writes.

Fixed

• yta run no longer reports "No matching workflows." when the real reason
  was idempotency dedup. Every skipped workflow is listed with its reason
  (trigger mismatch vs. already dispatched, with the idempotency key) and
  the closing line says what to do ("Re-run with --force").
• Skip notes for unavailable inputs carry the cause: the branch glob that
  matched nothing, the ambiguous branch list, or the git error including
  the base branch name — instead of a bare missing inputs: ['git_diff'].