BatchTicket (CLI: adt) is a deterministic structured-data transducer for Agent workflows. Like the small ticket that travels with a production or shipping batch, each run carries a versioned plan and an explicit account of what execution changed. That account does not claim the mapping was semantically correct.
It is not another jq or SQL dialect. DuckDB, JSON Schema, PyYAML, and standard parsers own established execution work. This project owns the Agent contract around them.
Version 0.2.0 is an unreleased source release candidate. The project is licensed under the
Apache License 2.0. The stable technical identifiers remain
agent-data-transformer, data-transformer, adt, and the four data_* MCP tools.
- Inspect JSON, JSONL, CSV, TSV, YAML, and Parquet without returning full payloads.
- Discover bounded nested record sets in JSON/YAML envelopes and profile their logical fields.
- Compare a source record shape with a target JSON Schema, surface only structural mapping candidates, and emit an executable draft plan only after explicit compatible mappings.
- Safe Transformation Plan v1 with no raw code, SQL, shell, jq, regex, or template execution.
- Filter, project, drop, rename, sort, limit, deduplicate, cast, derive, explode, join, group, pivot, unpivot, flatten, unflatten, and tree mutation.
- JSON Schema plus non-null, unique, row-count, field, and type assertions.
- Schema-aware keyed or unkeyed diff.
- Dry-run with real destination preflight, staged atomic publication, and overwrite protection.
- Whole-call worker isolation with cumulative source/byte/row/item/depth/time/RSS/temp limits.
- A byte ceiling over the complete serialized response, including samples, shapes, execution effects, diffs, validation failures, and errors.
- One strict typed Plan v1 model generates runtime validation, the published JSON Schema, and the live MCP schema.
- One shared core with CLI and four task-level MCP tools.
Development checkout:
git clone https://github.com/tetracoralla/BatchTicket.git
cd BatchTicket
uv sync --frozen --extra dev
uv run adt inspect examples/users.json --select 'data.users[*]'
uv run adt inspect examples/users.json --select 'data.users[*]' \
--target-schema target.schema.json --mappings mappings.json
uv run adt transform examples/adults.plan.yaml
uv run adt transform examples/adults.plan.yaml --dry-run
uv run adt validate examples/users.json --select 'data.users[*]' \
--assertions examples/users.assertions.json
ADT_WORKSPACE_ROOT=/absolute/granted/workspace uv run adt mcpBuild a platform-specific plugin that does not need the repository, uv, a Python
installation, or a network connection at runtime:
uv run python scripts/build_plugin.py
uv run python scripts/probe_plugin.py \
dist/plugin/data-transformer-0.2.0-darwin-arm64
codex plugin marketplace add dist/plugin
codex plugin add data-transformer@data-transformer-localThe build produces a plugin directory, a .tar.gz archive, and a SHA-256 checksum in
dist/plugin/. Install or copy the complete generated directory; the repository-root
.mcp.json is development configuration, while the generated plugin's .mcp.json invokes
its bundled executable directly. The generated local marketplace points Codex at that
self-contained directory. Rebuilding with --replace refuses symlinked output roots or
generated targets before deleting or overwriting anything.
Each generated bundle contains legal/THIRD_PARTY_NOTICES.md, the copied license texts under
legal/licenses/, and legal/sbom.cdx.json. The inventory is built from the locked runtime
dependency closure plus the incorporated CPython runtime and PyInstaller bootloader; it is checked
again from the archive by the plugin test path. The archive is stable for the same source, lock,
and build environment, so its companion SHA-256 file identifies the exact generated bundle.
This local macOS arm64 artifact is not a publication action. A GitHub Release upload, code-signing or notarization decision still requires separate authorization; see the release checklist.
The example transformation returns two records inline and reports through execution_effects that one input row was removed. File output is opt-in through output.path; existing files are never replaced unless output.overwrite is explicitly true.
CLI paths are explicit user paths and may be absolute. MCP file paths are a narrower
capability: the server first uses workspaces granted through the MCP roots protocol. A
single root is selected automatically; when several roots are granted, pass their exact
root name in the tool's optional workspace field. Hosts without roots support may pass
an explicit ADT_WORKSPACE_ROOT as a compatibility grant. Tool paths remain relative to
the selected root, and absolute, parent, URI, and symlink escapes are rejected. Inline-only
MCP calls do not require a workspace grant.
Codex CLI 0.148 does not currently pass MCP roots to local plugin servers. For a cold CLI
host test, launch Codex with ADT_WORKSPACE_ROOT set to the exact test workspace. Read-only
tools can run under approval=never; data_transform has conditional file-output
capability and therefore requires an approval-capable host policy even when a particular
call returns data inline. A blocked call is an authorization result, not permission to
fall back to shell or model-side data rewriting.
data_inspect accepts an optional target_schema and mappings object. It can select a
unique nested record set by structural evidence, suggest exact or normalized-name matches,
and report incompatible, missing, duplicated, omitted, or dropped fields. It does not use
fuzzy synonyms, casts, or defaults. The v1 adapter maps top-level record fields and supports
an object schema or an array whose items is one object schema. Composed/reference schemas
and ambiguous record sets remain explicitly unresolved.
When every required field has an explicit compatible mapping,
adaptation.status is ready and adaptation.draft_plan is a normal Transformation Plan
v1. Run that plan through data_transform; the adapter does not create a separate execution
path or fifth public tool.
capabilities/provider.json binds the read-only data_inspect and
data_validate core paths to
org.openadam.structured-data.analyze@0.1.0. The JSONL adapter validates the
portable request, preserves the same restricted workspace and isolated worker,
and projects provider results into the canonical shape and validation
contracts. A failed content constraint remains a successful result with
valid=false; malformed constraints and execution failures use stable
Capability errors.
The v0.2 Provider Manifest declares canonical JSONL adapter targets for
inspect and validate separately from the public FastMCP tool targets. Its
executable transport schema probe introspects the current data_inspect and
data_validate tool schemas, keeping canonical adapter and live transport
conformance as separate evidence lanes.
Installed source wheels expose adt-capability and
adt-transport-schema-probe. Their canonical schemas are package resources,
so neither command depends on a repository checkout or a sibling Procedure.
The self-contained plugin exposes the same two interfaces through the
runtime/adt-capability and runtime/adt-transport-schema-probe launchers,
which reuse one frozen runtime, and ships a release-bound Provider Manifest plus canonical
schemas. Installed Hosts can therefore use the Capability without a source
checkout, Python environment, package manager, or network access.
from data_transformer import DataTransformer
result = DataTransformer().transform(
{
"version": "1",
"sources": {"rows": {"inline": [{"x": 2}, {"x": 4}]}},
"steps": [
{
"op": "derive",
"field": "doubled",
"expr": {"multiply": [{"field": "x"}, {"value": 2}]},
}
],
}
)Public calls return status: ok, status: dry_run, or status: error with a stable error.code. Successful transformations expose runtime-observed changes under execution_effects; they do not leak stack traces or DuckDB internals.
See the product model, Transformation Plan v1, and the review contract. Contributions are described in CONTRIBUTING.md, security reports in SECURITY.md, and notable changes in CHANGELOG.md.