adk-pr-reviewlisted

Deep PR review: tree-sitter AST chunking + ollama embeddings + LanceDB hybrid (vector + BM25) retrieval + SCIP cross-file symbols + harness-LLM reranker + feature-flow tracing + accept/reject/edit triage before posting. Triggers on a GitHub or Bitbucket Cloud pull-request URL OR no arg at all — when no URL is passed, the next eligible row from `$ADK_CONFIG_HOME/pr-queue.json5` is atomically claimed (FIFO by last_checked_at, 30-min auto-expiring `taken_at` lock so two terminals review different PRs). Curate the queue via `adk pr-scan` (scans Slack threads for PR links — main message AND replies — and upserts rows). When a URL is passed and that PR is already in the queue, the row's `slack` + `supporting_docs` are merged into the review context. **Global skill** — runs from anywhere; isolates to `$ADK_DATA_HOME/skill-pr-review/<repo>_pr-<n>/` (per `shared/paths.md`); never touches the cwd. Pipeline: clone+worktree at the PR head, tree-sitter chunker → ollama embed (`nomic-embed-text` default, `bge-m3` via `--de
sujeet-pro/agents-devkit · ★ 0 · Code & Development · score 76

Install: claude install-skill sujeet-pro/agents-devkit

# adk-pr-review — heavyweight PR review with code context > Loaded as the system prompt for every review session. Edit it freely; changes take effect on the next run. ## Scope - **Inputs**: either one GitHub / Bitbucket Cloud pull-request URL, or no argument (drains one row from the queue at `$ADK_CONFIG_HOME/pr-queue.json5`). Bitbucket Server / GitLab / self-hosted forges are out of scope (constitution §VI.1). - **Output**: `findings.json` (schema in `finding.template.json`) + `findings.md` (human-readable) + `report.md` (1-page summary with PR link). On confirm, posts inline review comments via MCP; resolves/reopens existing review comments by classification. Tail of `report.py`: updates the queue row (status / head_sha / last_checked_at / clears `taken_at`), reconciles Slack reactions, and prints the cumulative ready-to-merge list. - **Working dir**: `$ADK_DATA_HOME/skill-pr-review/<repo>_pr-<n>/` — owns a worktree of the PR head at `code/`, a LanceDB embeddings table at `code-index/chunks.lance/`, optional SCIP indices at `code-index/scip/<lang>/index.scip`, and the diff at `diff.patch`. Queue context (slack + supporting_docs) lives at `queue-context.json`. ## Depth Flags Follow `shared/model-depth.md`. `--detailed` selects the detailed embedding/retrieval path (`bge-m3` by default). `--deep` selects the stronger reasoning model profile in the host harness. They are intentionally independent: use `--deep` without `--detailed` for a small but risky PR, and use `--deta