chore: release main

[github-actions]

🤖 I have created a release beep boop

0.2.0

0.2.0 (2026-09-24)

⚠ BREAKING CHANGES

• demo: make Chinook the only demo dataset and test against it (#98)

Features

• api: expose plan, validation checks, row sample, status and timings on QueryResult (#88) (7a41a60)
• api: NL2SQL accepts env and env_file; embeddings fail with an actionable message (#90) (e8ca01c)
• api: take the role from an auth dependency, not the request body (#142) (f8c06fd)
• cli: pass an API key to nl2sql demo and persist it in .env.demo (#97) (b0305f8)
• demo: a link preview card for the hosted demo (#178) (927baf1)
• demo: add support and web-analytics sample databases linked to Chinook customers (#157) (659039f)
• demo: Dockerfile and Hugging Face Space configuration for the hosted demo (#168) (835687f)
• demo: hosted mode so a public demo can take each visitor's own key (#165) (cfb6cd3)
• demo: React playground with schema, plan, validation, SQL and rows, plus nl2sql demo (#93) (9f0a922)
• demo: vendor the Chinook database and add DemoManager.setup_chinook (#91) (22cd257)
• eval: --model and built-in presets for tier 2, with project-relative outputs and publish --from (#129) (1a8101d)
• eval: add the Chinook gold dataset with executed gold results (#114) (176a369)
• eval: alternative gold answers, confidence intervals and a paired regression gate (#156) (de42c40)
• eval: answer faithfulness and schema retrieval recall (#128) (571cd16)
• eval: keep each sub-query's plan and the answer text in tier 2 records (#146) (db73b92)
• eval: report a lenient accuracy that allows extra and reordered columns (#152) (37351c3)
• eval: tier 1 harness runs gold plans through the code nodes on every PR (#119) (6be6357)
• eval: tier 2 benchmark runs the real model across configs with a cost cap and a comparison scoreboard (#126) (e1dc47e)
• feedback: record answer feedback in the playground and report guardrail rates (#127) (4d24267)
• generator: break row-order ties with the selected columns so results are deterministic (#116) (88ed9f3)
• llm: default to gpt-5.4 and make temperature optional (#108) (1509153)
• llm: provider adapters per wire format, native Claude support, and a provider per node (#124) (2f39f9a)
• llm: replay store and recording proxy for key-free demo mode (#92) (a02c236)
• planner: cache validated plans by question and schema version, re-validating on every use (#122) (f5e34b3)
• playground: a Pipeline page showing what each step does (#171) (8535e21)
• playground: add a settings panel for the API key and a model per LLM node (#109) (0e60f4e)
• playground: choose a model per step in the hosted demo, with a key per provider (#173) (2220e74)
• playground: give Settings and Retrieval their own pages with a header nav (#154) (4e6c6f5)
• playground: let the demo show each of its databases (#172) (67f06bd)
• playground: redesign the demo page around one run, and add a Debug toggle (#102) (2cfe121)
• playground: tell a new visitor what the demo needs before they ask (#170) (4ff7dbd)
• resolver: skip vector search for a single datasource and refuse questions no datasource can answer (#121) (5430c9c)
• retrieval: pass the full schema to the planner for small schemas (#89) (88e4651)
• retrieval: record retrieval candidates, scores and MMR picks in traces, and add a Retrieval inspector to the playground (#123) (03f74c0)
• telemetry: report per-node and per-question LLM tokens, calls and time on every run (#101) (717758a)
• trace: record every node of a run to one file, show it, and replay it without the model (#104) (1e55710)

Bug Fixes

• adapters: return sqlglot dialect names from get_dialect (#136) (be25596)
• aggregator: resolve prefixed join keys when combining sub-query results (#162) (e03bab3)
• aggregator: use polars group_by, resolve prefixed join keys, apply every post-op field (#144) (4f052dc)
• cli: doctor checks the LLM key, errors name the env file, no traceback by default, add --version (#95) (5c6ae6e)
• cli: exit 1 when a run ends with pipeline errors (#115) (0ec4906)
• cli: load the --env file into the environment; plan ORDER BY keeps the dialect's NULL placement (#120) (efcd76f)
• config path options, explicit-path descriptions, and COUNT(DISTINCT) in plans (#125) (4d5e625)
• config: hold API keys and connection secrets as SecretStr so they never print (#113) (1cec9cf)
• decomposer: carry a sub-query's ranking and limit into its plan (#141) (46cdc43)
• decomposer: reject SQL syntax, not English words that happen to be SQL keywords (#105) (022a3b3)
• demo: be honest about replay mode, make --record replayable, and ship license notices (#117) (5430943)
• eval: identify a benchmark run by the dataset's datasources, not every registered one (#158) (3256dae)
• generator: attach the unjoined side of each join, render arithmetic, stop after a denial (#100) (fb618ab)
• generator: quote select aliases and column names that are not bare identifiers (#133) (460df7c)
• generator: render string concatenation as the dialect's own, never numeric + (#132) (e55c1dc)
• index: give the datasource entry its description and example questions, and refresh snapshot metadata on re-index (#111) (29dab2f)
• index: rebuild without ever leaving an empty index, and show index health with a Rebuild action (#110) (cf0cdcb)
• pipeline: derive sub-query status from its final attempt, not its error history (#106) (059f45a)
• pipeline: honor execute=False so plan-only runs never touch a database (#86) (cfa9ad0)
• pipeline: return a structured result when a sub-query fails instead of crashing or looping (#84) (2c9d6a0)
• rbac: refuse strictly without leaking forbidden data, and deny unknown or empty roles cleanly (#112) (ec18715)
• registry: one capability check that fails closed (#145) (f65e691)
• report global planner failures and pass the refiner the failed plan as JSON (#107) (e391d35)
• runtime: enforce the pipeline timeout and keep signal handlers on the main thread (#87) (4174e0b)
• validator: match FK joins when relationships carry qualified names (#94) (8575932)
• validator: only allow known function names in plans (#137) (948f1a3)
• validator: reject unjoinable plans before generation so the model can retry (#161) (906cf99)
• validator: wrap ORDER BY terms and stop treating sampled values as a domain (#99) (f443df1)

Performance Improvements

• prompts: compact the schema and put stable context first so prompts cache (#118) (4513d1a)

Documentation

• align README, docs and packaging metadata with the code that exists (#96) (3f35661)
• playground: describe the pages and several databases in the README and rail (#160) (d914978)
• readme: drop engine.results, removed with ResultAPI (#164) (072a6f4)
• readme: restructure around quickstart, results, CLI, REST API and screenshots (#131) (58df5c9)
• readme: update for the hosted demo and fix the diagram on dark backgrounds (#179) (6d47a14)
• sync documentation with recent changes (#103) (362d943)
• sync documentation with recent changes (#135) (b176f0f)
• sync documentation with recent changes (#167) (082f053)

Code Refactoring

• demo: make Chinook the only demo dataset and test against it (#98) (047751b)

---

This PR was generated with Release Please. See documentation.