Translation pipeline¶

The translation pipeline converts technical texts — cells (CodeManifests, usage files) and standalone documents — from one language to another while preserving structure, semantics, and requirements.

How to invoke¶

Run /goga:tool scriba with a request that signals translation intent. The dispatcher routes to the goga-tool-scriba-trans skill, which runs the pipeline.

Examples¶

/goga:tool scriba translate cell path/to/cell to French
/goga:tool scriba переведи cell path/to/cell на английский
/goga:tool scriba translate documents path/to/docs to German

The pipeline detects the source language automatically. You provide the target cell or document path and the target language.

How it works¶

The pipeline runs eight stages in sequence. Each stage builds on the results of the previous one and writes its output into a shared pipeline_context.

1. Terminology¶

Extracts technical terms, entities, and aliases from the source. Builds a canonical glossary that all subsequent stages use for consistency.

2. Semantic Model¶

Analyzes the source structure — sections, entities, actors, actions, constraints, and workflows — without translating. This model guides the generation stage.

3. Generation¶

Produces three independent translation variants:

Literal — closest to the original, maximum semantic preservation
Technical — architecture-grade language with active voice and industry terminology
AI-optimized — explicit subjects, entities, and actions for LLM consumers

4. Context Enrichment¶

Merges the glossary, semantic model, and all three variants into a unified context — no new information is added.

5. Synthesis¶

Selects the best variant for each segment using a priority matrix: semantics first, then terminology accuracy, then readability, then AI clarity.

6. Validation¶

Runs an independent quality audit covering structure integrity, semantic fidelity, terminology consistency, AI readability, and formatting integrity. The pipeline retries on failure up to 2 times.

7. Apply Results¶

Applies synthesized documents to files at the paths defined in the cell manifest.

8. Finalize¶

Runs linting across all cells, fixes errors, and produces a human-readable report summarizing the entire pipeline output — languages, glossary, semantic model, validation results, translated files, and final status.

What stays unchanged¶

Throughout all stages, the pipeline guarantees:

Document structure
Section and list order
Algorithm order
Requirements and constraints
Prompt logic
Instruction priority

Forbidden operations: reordering sections or steps, merging or splitting steps, removing or adding requirements, inferring new logic.

For cells, only annotations and usage files are translated — the contract structure remains untouched.