Changelog

All notable changes to this project will be documented in this file.

The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

(no changes yet)

[0.6.3] - 2026-04-27

Documentation release. Library code is unchanged from 0.6.0; this release ships a Diátaxis-shaped doc set and hosted API references.

Fixed

• doc/cljdoc.edn tree shape: children are inline siblings, not wrapped in an extra vector. v0.6.2 published cleanly to Clojars but cljdoc rejected the malformed tree.
• build.clj writes <licenses> (MIT) into the pom; required by Clojars as of late 2025.
• :build deps alias includes slipset/deps-deploy, which the deploy task in build.clj requires.

Note on 0.6.1 and 0.6.2

v0.6.1 was tagged in git but never fully reached Clojars: the pom and jar uploaded but the metadata.xml step failed because the pom was missing the licenses block. Clojars locks coordinates on first attempted upload, so the version became unusable.

v0.6.2 published to Clojars cleanly, but the doc/cljdoc.edn tree was malformed (children wrapped in an extra vector), causing cljdoc's analysis step to throw "Don't know how to create ISeq from: clojure.lang.Keyword". The Clojars artifact at 0.6.2 is installable and equivalent to 0.6.3 — only the hosted API ref is broken.

0.6.3 is the first version that's both installable and renders on cljdoc. The v0.6.1 and v0.6.2 git tags are retained for traceability.

Added

• Doc set following the Diátaxis split: doc/TUTORIAL.md (45-minute walkthrough), doc/DESIGN.md (rationale consolidating ADRs), doc/WEB_UI.md (page-by-page tour), doc/MCP.md (16-tool reference), doc/FAQ.md and doc/TROUBLESHOOTING.md, plus six P0 how-to guides under doc/howto/ (SQLite→Postgres migration, Docker team deployment, MCP+Claude Code, Slack alerts, sklearn tracking, HTTP write API).
• Hosted API reference via cljdoc (doc/cljdoc.edn bundle config) and codox (:codox deps alias + bb docs task).
• gh-pages publishing via .github/workflows/docs.yml.
• README: documentation jump-table near the top with cljdoc and gh-pages badges.

[0.6.0] - 2026-04-20

Full backlog complete: S3 artifacts, Slack webhooks, run cleanup, Docker, HTTP write API.

Added

• M14 S3 artifact storage (chachaml.store.s3): ArtifactStore against S3/MinIO via Cognitect aws-api. Metadata in Postgres/SQLite, bytes in S3. :s3 deps alias.
• M16 Docker: Dockerfile (multi-stage JDK21) + docker-compose.yml (Postgres + chachaml UI). docker compose up for team deployment. UI server accepts env vars DB_TYPE, JDBC_URL, DB_USER, etc.
• M17 HTTP write API: POST endpoints at /api/w/runs, /params, /metrics, /end, /artifacts, /models. Full lifecycle test. Enables Python/Go/curl clients.
• M18 Slack webhook for alerts: webhook_url on alerts. When check-alerts! triggers, POSTs Slack-formatted JSON to the URL.
• M19 Run cleanup: archive-runs! marks old runs as :archived by age + experiment. delete-archived! permanently removes data. 2 tests.

Changed

• UI server start! now delegates to chachaml.store/open — accepts :type :postgres and all store options directly.
• Refactored: left-to-right style throughout (threading macros, let bindings, named predicates, DRY).
• Deterministic seeds per showcase use case — fixed flaky tests.

[0.5.0] - 2026-04-20

Team readiness: Postgres backend, user attribution, pipelines, alerts, chat-with-data.

Added

• M10 Pipelines (chachaml.pipeline): run-pipeline! executes named step sequences, each step tracked as a with-run. Results chain via :prev-result. defpipeline macro. pipelines + pipeline_steps tables.
• M11 Alerts (chachaml.alerts): set-alert! defines metric threshold rules, check-alerts! evaluates against latest runs, records triggered events. alert-history, deactivate-alert!.
• M12 Chat-with-data (chachaml.chat): ask sends questions to Anthropic/OpenAI API with chachaml tools injected. Tool-use loop. /chat UI page with provider selector and localStorage API key.
• M13 Postgres backend (chachaml.store.postgres): full protocol implementation with HikariCP pool. Same API as SQLite, team-shared database. chachaml.store/open dispatcher: :type :sqlite or :type :postgres. :postgres deps alias.
• M15 User attribution: created_by column on runs, auto-captured from system user, filterable in queries and visible in UI.
• Markdown notes with math: runs and models support markdown notes rendered via marked.js + KaTeX. set-note! stores markdown; the UI renders it with LaTeX math support. set-model-note! added to the registry for model descriptions.
• 16 MCP tools (was 6): search_runs, best_run, add_tag, set_note, get_tags, get_datasets, list_experiments, create_experiment, export_runs, diff_model_versions.
• 7 new API endpoints: tags CRUD, note POST, datasets GET, search, model note, version diff.
• Showcase demonstrates v0.4: UC01 uses log-table + log-dataset!, UC08 uses with-batched-metrics, UC17 uses create-experiment!, UC19 uses add-tag!, UC25 uses set-note! with markdown+math, -main uses best-run + export-runs.
• JS/CSS bundled locally: marked.min.js, katex.min.js, auto-render.min.js, katex.min.css.

Changed

• Refactoring: extracted chachaml.format namespace with shared short-id, size-str, fmt-instant, fmt-duration, pad, metric-summary. Eliminated 4 duplicated helper fns from chachaml.repl and chachaml.ui.views. Extracted ->experiment-map in sqlite.clj (was copy-pasted 3 times). Consolidated sklearn interop require/resolve pattern into core-fn helper.
• Mutable tags + run notes: add-tag!, set-note!, get-tags — annotate runs after they complete. Tags stored in a dedicated table with upsert semantics; merged with inline tags in (run id).
• Dataset tracking: log-dataset!, get-datasets — capture row/col counts, feature names, content hash, data source per run. New datasets table. Included in (run id) results.
• Metric-based run search: search-runs with :metric-key, :op, :metric-value filters. best-run shortcut for {:metric :accuracy :direction :max}.
• Batch metric logging: with-batched-metrics macro — buffers metrics in an atom, flushes to store in one batch on exit (or on exception). log-metric/log-metrics respect *metric-buffer*.
• Structured table artifacts: log-table stores tables as EDN with content-type application/x-chachaml-table. Added to the serialize content-type→format mapping.
• Image artifact preview: run detail UI detects image/* content-type and renders inline <img> tags. Other artifacts get a download link via /api/artifacts/:id/download.
• Export runs to CSV: export-runs returns flat maps (params + final metrics per run). /api/export?format=csv endpoint.
• Experiment metadata: create-experiment!, experiment, experiments — name, description, owner. New experiments table with upsert.
• Model version diff: chachaml.registry/diff-versions compares the runs behind two versions of the same model.
• 18 tests covering all new features.

[0.3.0] - 2026-04-17

sklearn interop and 25 ML use case showcase.

Added

• M9 libpython-clj2 sklearn interop (chachaml.interop.sklearn):
  • tracked-fit! — fit a sklearn-compatible model, logging training time + model hyperparameters.
  • tracked-predict — predict with timing.
  • evaluate! — compute + log accuracy/R² metrics.
  • train-and-evaluate! — full pipeline: fit, predict, evaluate, save artifact, optionally register model.
  • extract-params — pull sklearn's .get_params().
  • All Python calls go through a *bridge* dynamic var so tests run without Python (mock bridge replaces libpython-clj2).
  • requiring-resolve for the Python bridge — namespace compiles without libpython-clj2 on the classpath.
  • :python alias in deps.edn for users with Python + sklearn.
  • examples/sklearn_iris.clj — runnable with clojure -M:python:examples -m sklearn-iris.
  • 8 mock-based tests covering fit, predict, evaluate, the full pipeline, and param extraction.

[0.2.0] - 2026-04-17

MCP server, web UI with local JS deps, and SQLite WAL mode for concurrent multi-agent access.

Added

• M7 MCP server (chachaml.mcp):

  • JSON-RPC 2.0 over stdio; run via clojure -M:mcp [db-path].
  • 6 tools: list_runs, get_run, compare_runs, list_models, get_model, get_model_version.
  • Configurable in .claude/mcp.json or any MCP-compatible client.
  • org.clojure/data.json added as an optional dep (:mcp alias for runtime, test/coverage aliases for CI).
  • 18 tests covering handshake, all tools, error paths, and a ByteArrayStream integration test.

• SQLite WAL mode + busy_timeout=5000 enabled on every open — concurrent readers (UI) never block writers (training agents), and simultaneous writers wait instead of failing with SQLITE_BUSY. Includes a concurrent-write integration test.

• M8 Web UI (chachaml.ui.*):

  • Ring + Reitit + Hiccup + HTMX + Vega-Lite + Tailwind CDN.
  • Start with clojure -M:ui [db-path] [port] (default 8080).
  • 5 screens: runs dashboard (auto-refreshes every 10s via HTMX), run detail (params, scalar metrics, time-series Vega-Lite charts, artifacts), run comparison (param diff + overlaid metric charts), model registry, model version detail.
  • JSON API at /api/{runs,runs/:id,compare,models,models/:name, experiments} — foundation for a future chat-with-data layer.
  • 22 UI tests via ring-mock (HTML + JSON routes, 404s, filters).

Changed

• Cleanup pass after the v0.1.0 review:
  • Stale namespace docstrings in chachaml.core and chachaml.store.sqlite updated to reflect shipped milestones.
  • Collapsed duplicated status->db/stage->db (and inverse) helpers into a single kw->db/db->kw pair in chachaml.store.sqlite.
  • chachaml.serialize/auto-format no longer has a dead :nippy branch for strings; behavior is unchanged (strings still go to :nippy) but documented as intentional.
  • chachaml.repl/pad reduced to a one-line format call.
  • chachaml.schema/validate docstring tightened (no behavior change).

Internal

• New shared test fixture chachaml.test-helpers/with-fresh-store, used via (use-fixtures :each h/with-fresh-store) across core_test, registry_test, repl_test, tracking_test, and examples_test. Drops ~30 lines of inline (with-fresh-store (fn [] …)) boilerplate per file.
• New use-store!-returns-resolved-store test for the store-value path.

[0.1.0] - 2026-04-16

First feature-complete release covering the full minimal MLOps loop: run tracking, artifacts, model registry, tracking macro, REPL helpers.

Added

• M0 project skeleton: dual build (deps.edn + project.clj), Babashka task runner (bb.edn), tools.build entry points (build.clj).
• Quality tooling: clj-kondo config (warning-level fail), cljfmt config, kaocha runner (tests.edn), cloverage 85% line gate.
• Namespace skeletons for chachaml.{core,context,env,serialize, registry,tracking,repl,schema} and chachaml.store.{protocol,sqlite}.
• Baseline load test asserting every namespace compiles.
• GitHub Actions CI: lint + format + 4-job test matrix ({deps,lein} × JDK {17,21}) + coverage gate.
• Process artifacts: CONTRIBUTING.md, LICENSE (MIT), .github/PULL_REQUEST_TEMPLATE.md, ADRs 0001–0004.
• M1 storage: chachaml.store.sqlite implements RunStore and Lifecycle against either a file-backed or in-memory SQLite database. Schema migrates idempotently on every open. Tests cover CRUD, EDN value round-tripping, query filters, and tempfile persistence across opens.
• M2.1 chachaml.env: best-effort capture of git SHA/branch/dirty state, JVM, OS, user, and Clojure version.
• M2.2 chachaml.context: *store* and *run* dynamic vars plus a delay-initialized default SQLite store at ./chachaml.db.
• M2.3–2.5 chachaml.core public API: use-store!, with-store, start-run!, end-run!, with-run, log-params, log-param, log-metrics, log-metric, runs, run, last-run. Exceptions inside with-run mark the run :failed and re-throw; completed runs are tap>-ed.
• chachaml.schema: Malli schemas for run lifecycle and logging inputs (Status, Run, StartRunOpts, Params, Metrics, MetricRow, QueryFilters) plus a validate helper.
• doc/USING-LOCALLY.md: how to depend on a local checkout from another Clojure project (:local/root, lein install, lein checkouts/).
• Runnable ML examples under examples/ (added to the test classpath via the :examples alias):
  • linear-regression — gradient-descent univariate fit on synthetic data; logs per-epoch loss and final weight errors.
  • kmeans — Lloyd's algorithm on 2D Gaussian blobs; logs per-iteration inertia and final cluster sizes. Each has a run-experiment! fn (called by tests against an in-memory store) and a -main (called by clojure -M:examples -m <name>).
• M3 artifacts:
  • chachaml.serialize: multimethod-dispatched encoding for :nippy (default), :edn, :bytes, and :file formats. Auto-detects byte[]/File and falls back to :nippy.
  • SQLite ArtifactStore impl with filesystem-backed bytes under <db>-artifacts/<run-id>/<name> (default ./chachaml-artifacts/). SHA-256 hash + size captured. Unique (run_id, name) index rejects duplicates.
  • chachaml.core/log-artifact, log-file, load-artifact, list-artifacts. (run id) now also includes :artifacts.
  • Artifact and ArtifactFormat Malli schemas.
  • In-memory mode now creates a per-store temp artifact directory that is removed on close!.
  • Examples now persist their trained models as nippy artifacts; tests round-trip and assert on the loaded values.
  • Property-based serialize round-trip (test.check) covers arbitrary nested EDN values for :nippy and :edn.
• ADR-0005: filesystem-backed artifacts (rationale for keeping bytes outside SQLite).
• M4 chachaml.tracking/deftracked: defn-shaped macro that wraps the body in with-run. Supports docstring + opts map (:experiment, :name, :tags); a :tracked-fn tag carrying the qualified symbol is auto-added. Nests under existing runs. clj-kondo :lint-as config registers it as a defn-like macro.
• M5 model registry:
  • models and model_versions tables added to the SQLite schema.
  • SQLite ModelRegistry impl with atomic stage transitions in a transaction.
  • chachaml.registry public API: register-model, models, model, model-versions, promote!, load-model (with :version/:stage selectors; default = latest production).
  • Model, Version, Stage, RegisterOpts Malli schemas.
  • Property-based test asserting at most one :production version per model survives arbitrary add/promote sequences.
  • ADR-0006 records production-stage exclusivity.
  • Both ML examples now register their trained model as a :staging version of linear-regression-baseline / kmeans-baseline. Tests round-trip through the registry too.
• M6 REPL helpers in chachaml.repl:
  • runs-table — compact one-row-per-run table of recent runs
  • inspect — polymorphic pretty-printer for runs (by id or map), models (by name or map), and model versions
  • compare-runs — diff params + latest metric values across N runs, returns {:runs … :params {…} :metrics {…}} with :same, :differ, :partial partitions
  • print-comparison — render the compare-runs map for humans
  • 10 tests covering empty store, multi-run diff, partial-key detection, version inspection, model fallback by string id