ReskillSQL skill
ReskillSQL analyses failed agent SQL runs and ships query-writing skill versions only after checks on held-out examples.
- 184 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Schema and questions are sent to the chosen LLM provider.
- The skill is rewritten automatically from run results, so patches deserve review.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/reskillsqlDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Schema and questions are sent to the chosen LLM provider.
- The skill is rewritten automatically from run results, so patches deserve review.
This is third-party code. Review the repository files before installing.
What it does
ReskillSQL, a Text-to-SQL agent with updatable skills, traces a wrong query to its cause: extra projected columns, the wrong aggregation, value formatting, a broken join path. From those causes it drafts a patch for the `query-writing` skill, which passes a SQL-aware gate before it goes live. The repository holds two skills, `query-writing` and `schema-exploration`, their history under `skills/query-writing/versions` and reproducible benchmarks. Model weights stay untouched: only versioned instructions and tool-use policies change. SQLite runs read-only with a timeout. Optional experience memory is off by default.
Who it is for. Researchers and engineers tuning a Text-to-SQL agent who want instruction updates driven by logged failures instead of manual prompt edits.
Good fit when
- An agent writes SQL and the same mistakes recur across queries.
- A skill needs a version history with the reason behind each change.
- A team compares skill updates against experience memory on a held-out split.
- The `query-writing` and `schema-exploration` skills are attached to an agent for SQLite.
Not a fit when
- You need a ready plugin that installs into Claude Code with one command.
- The database is not SQLite but another engine.
- You require published accuracy figures: the repository has none.
Example request
Review the failures of this Text-to-SQL run, name the category of each error and propose a patch for the `query-writing` skill.Limitations
This is a research project: there is no official accuracy table and reports stay local under `reports/`. The main protocol targets a BIRD-derived set and Gemini 2.5 Flash. Evaluation needs access to an LLM provider, and experience memory also needs an embeddings-compatible endpoint. A service mode is documented, but the project states it has not been load-tested. Skills have to be copied into the agent directory by hand because the repository has no ready installer for Claude Code. Review accepted patches yourself, since the skill rewrites itself automatically.
How to disable. Remove the skill folders from the agent configuration and delete the project directory. Clean local reports, memory databases and provider keys separately.
FAQ
Does ReskillSQL fine-tune the model?
No. Only versioned skills and tool-use policies change; model weights stay as they were.
How can I try it without keys?
Run `uv run text2sql-demo`. The demo is deterministic and makes no LLM calls, and `uv run text2sql-render-report` builds the report.
What does the gate decide?
It compares accuracy before and after a patch and checks the target error category and regressions. An accepted patch becomes a new version; a rejected one stays as evidence.
Is experience memory enabled?
No, it is off by default. It needs an embeddings endpoint and learns only from the update and validation splits.
Is SQL execution safe?
SQLite opens read-only and each query has a timeout, so bad SQL cannot hang a run.
Discussion
No questions or reviews yet.
Related
Skillssqlsure checks SQL against declared schema facts and catches join fan-out double-counting before the query runs.
A Neon skill with Postgres 14 to 18 guidance: schema design, indexes, query tuning, migrations and common pitfalls.
31 skills for Claude Code and Codex for the analyst workflow: data checks, cohorts, funnels, A/B tests, dashboards, summaries
Probabl skills for machine learning experiments: a scikit-learn and skrub pipeline, evaluation, audit and a link to Skore Hub
Skill and CLI for tables: cleaning with recorded decisions, level-based statistics, a PDF report and a separate audit.
The CLI imports CSV and JSON into SQLite, runs SQL and reshapes tables for agent-assisted local data preparation.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| ReskillSQL (this card) | no | 184 | |
| Self-Improving Agent | no | 790 | records an agent's lessons in general form; ReskillSQL is SQL-specific and validates each patch with a gate |
| SQL Query Analyzer | no | 0 | analyses existing queries; ReskillSQL teaches the agent how to write them |
README in short
A skill update runs in three steps: error diagnosis, a candidate patch, a gate check. Update, validation and holdout splits are kept apart to measure transfer to new tasks.