Skip to content

Agent Skills

The toolbox ships a suite of standard Agent Skills (the SKILL.md format) so you can drive the customs model in plain English, right inside your own project. They live in skills/ and are focused on one thing: using the model on your own codebase.

They are harness-agnostic — installable into Claude Code, Cursor, Codex, opencode and any other agent the skills CLI supports.

The suite

  •  customs-schema-setup

    Stand the model up in a database: create it, load schema + seed (+ optional e2e) in order, and report a clean/failed result.

    "Set up the customs sandbox in a local Postgres database."

  •  customs-query

    Generate correct SQL against the model — it knows the asycuda search path and the join paths — and verify it privacy-preservingly through the companion tester MCP (or a bundled script): metadata only, never row data.

    "Show me duty revenue by HS chapter last quarter — and test it."

  •  customs-seed

    Add reference/code values or generate realistic sample declarations and manifests that respect every foreign key.

    "Generate three more import declarations with 2–4 items each."

  •  customs-extend

    Add tables/columns while preserving the conventions and the provenance trail, and update the coverage docs in lock-step.

    "Add a container gate-movement table to the schema."

  •  customs-validate

    Re-run the done-condition checks: clean load, every table tagged, every cited source resolves. Your auditable green light.

    "Validate the customs schema still loads clean and is fully sourced."

Install into your codebase

Install the whole suite into your current project with the skills CLI — it auto-detects your agent (Claude Code, Cursor, Codex, opencode, …) and drops the skills where that harness discovers them:

npx skills add FrancoisChastel/sydonia-toolkit

Add a single skill with --skill:

npx skills add FrancoisChastel/sydonia-toolkit --skill customs-query

Also copy the model itself (the Sydonia/ folder, or at least Sydonia/schema/ and Sydonia/examples/) so the skills have SQL to load. The skills accept a --schema-dir argument if you put it elsewhere.

Verify your agent sees them — in Claude Code, for example:

/skills          # lists available skills; customs-* should appear

Use them

Just describe what you want — your agent matches the request to a skill:

> set up the customs model in a throwaway database and run the example
> write a query for assessed-vs-paid across all released declarations
> add a "carrier rating" column to trader, keep it sourced
> validate the schema is still clean and fully grounded

Each skill is a self-contained SKILL.md with bundled scripts and reference material, following Anthropic's progressive-disclosure guidance — small front-matter, details loaded only when needed.

The privacy-preserving query tester (optional MCP)

customs-query ships with a companion MCP server, customs-query-tester (pre-registered in the repo's .mcp.json), that closes the generate→verify loop: the assistant can prove a query is valid and runs — against a database that may contain real, sensitive customs declarations — without any row ever reaching the model.

The model gets an oracle, not a window. Four guarantees, enforced server-side on every call:

Guarantee How
Nothing can be written Sessions start with default_transaction_read_only=on — the PostgreSQL server itself refuses writes, even data-modifying CTEs
Only single SELECTs run A comment/string-aware scanner rejects multi-statements and every write/DDL/COPY/INTO keyword before the DB is touched
No row data in responses Result shape via \gdesc (describe without executing); volume via SELECT count(*) FROM (<query>) __q — one aggregate number
Bounded execution statement_timeout (default 5 s) + a hard subprocess cap

Three tools: describe_schema (structure), validate_query (EXPLAIN-only), test_query (read-only run → columns, types, row count, duration). No MCP connected? The skill falls back to scripts/test_query.sh with the identical guarantees through plain psql. For defence in depth on real data, run it as a SELECT-only database role — the server's README shows the three-line GRANT.

What they deliberately don't do

These skills are for using the model. They will not decompile ASYCUDA, touch a live customs system, or fabricate source citations — the same source policy that governs the model governs the skills.