.coderabbit.yaml

# CodeRabbit configuration for sqlite-predict.
# Reference: https://docs.coderabbit.ai/reference/yaml-template

language: "en-US"

tone_instructions: >-
  Be direct and specific. Cite the file and line. Skip praise and
  restatements of the diff; comment only where something should change,
  and say exactly what and why.

reviews:
  profile: "chill"
  # Request changes while comments are open, approve once they resolve:
  # this is what lets CodeRabbit act as an approving reviewer.
  request_changes_workflow: true
  high_level_summary: true
  poem: false
  collapse_walkthrough: true

  auto_review:
    enabled: true
    drafts: false
    base_branches:
      - main

  path_filters:
    - "!dist/**"
    - "!vendor/**"
    - "!npm/**"
    - "!website/node_modules/**"
    - "!bindings/rust/target/**"
    - "!benchmarks/results/*.jsonl"

  path_instructions:
    - path: "**/*.{c,h}"
      instructions: >-
        This is a zero-dependency C99 SQLite extension in the sqlite-vec
        mold. Hold changes to these rules: (1) fail loudly, never
        silently — errors use the closed PREDICT_ERR_* set formatted
        "CODE: detail", unknown option keys are rejected, and there are
        no silent fallback paths; (2) memory discipline is
        sqlite3_malloc/sqlite3_free with goto-cleanup, every error path
        must free what it allocated — trace them; (3) all deserialization
        of registry blobs is bounds-checked, and inline weights are
        verified against content_hash before use; (4) serving functions
        (forecast/detect_anomalies aggregates, predict) must stay pure:
        no writes, no table creation, safe under PRAGMA query_only;
        (5) determinism: same input, same output — no clock, randomness,
        or float formatting that breaks %.17g round-tripping; (6) new
        cross-TU symbols need the predict0_ prefix (the amalgamation
        makes unprefixed externs collision bait); (7) no new
        dependencies in the core build, ever. Flag any comment that
        references internal design documents (RFC section numbers): the
        public docs site is the contract.

    - path: "tests/**"
      instructions: >-
        Tests here should try to break the code, not confirm the happy
        path: flag new features that lack adversarial cases (malformed
        blobs, unknown options, imbalanced or degenerate data, error
        paths). Error-message assertions should match PREDICT_ERR_*
        codes, not prose. Registered test model ids must not reuse
        bundled ids (auto, tsb, theta-classic, stub-seasonal-naive,
        sub-pca, knn5-incontext). Soak drivers must propagate statement
        failures to the exit code — flag any discarded result.

    - path: "benchmarks/**"
      instructions: >-
        Benchmark claims must be measured and reproducible: flag numbers
        stated without a script or results file behind them, and any
        methodology change that isn't reflected in the results doc.
        Model licensing is load-bearing here: TabFM and TabPFN-2.5/2.6/3
        are non-commercial teachers that must never be distilled for
        anything published as more than evaluation, and no teacher
        weights or teacher-derived artifacts may be committed to the
        repo (caches live under ~/.cache). TabPFN-3 outputs must not be
        used to train or distill any model. Permissive teachers (TabPFN-2,
        TabICL, Mitra, Chronos) are the only distillation sources whose
        students may be presented as shippable.

    - path: "{website/src/content/**,*.md,.github/**/*.md}"
      instructions: >-
        Docs must read standalone for a first-time public reader. Flag:
        em or en dashes (house style forbids them; use periods, colons,
        commas, or parentheses), "an SQL" (house style is "a SQL", the
        sequel pronunciation; "an SQLITE_" for the spelled-out C
        identifier prefix is correct and exempt), references to internal
        design docs or RFC section numbers, receipts-era terminology
        (the feature was removed), title-case headings (use sentence
        case), and any
        distillation framing that suggests distilling a non-commercial
        model for production use. Claims of accuracy or speed must match
        the benchmarks pages.

    - path: ".github/workflows/**"
      instructions: >-
        Every workflow must declare an explicit permissions block (the
        repo default token is read-only). Release automation is
        tag-driven (v* fires release/wheels/npm/crate): flag anything
        that would publish outside that path or duplicate the version
        source of truth (make sync-version owns version stamps).

    - path: "bindings/**"
      instructions: >-
        Binding metadata (descriptions, READMEs, __version__) rots
        independently of the docs site: flag stale references to removed
        features and version strings not managed by scripts/sync-version.mjs.

  pre_merge_checks:
    docstrings:
      mode: "off"
    title:
      mode: "warning"
      requirements: >-
        Conventional-commit style prefix (feat/fix/docs/bench/site/chore/
        refactor/test/ci) with a concise, specific subject.
    custom_checks:
      - name: "Changelog entry for user-visible changes"
        mode: "warning"
        instructions: >-
          If the PR changes user-visible behavior (SQL surface, option
          keys, error codes, defaults, output formats, packaging), it
          should add an entry under CHANGELOG.md's [Unreleased] section
          in the same PR. Refactors, docs, tests, benchmarks, and CI
          changes do not need one. Release-time promotion of
          [Unreleased] into a dated section is owned by make
          sync-version; PRs should never do that themselves.

chat:
  auto_reply: true