Super Translate

Verification: Target file list confirmed by user; all 3 required files exist.

Step 2: Terminology Preflight (Fail-Closed)

uv run python scripts/validate_glossary.py
uv run python scripts/term_read.py --fail-on-missing --fail-on-forbidden

If fails → stop and fix terminology first.

Verification: Both commands exit 0 with no missing/forbidden terms.

Step 3: Resolve Translation Mode

Read style-decisions.json.translation_mode.mode. If missing, ask user:

• 完整翻譯：完整翻譯所有內容，保留原始結構與細節
• 摘要翻譯：精簡翻譯重點規則，省略範例與冗長說明

Verification: translation_mode.mode is persisted in style-decisions.json.

Step 4: Pipeline Execution

Pre-read shared context once per batch:

• GLOSSARY_CONTENT = glossary.json
• STYLE_CONTENT = style-decisions.json (includes translation_notes as hard constraints)

For each target file, run the pipeline:

1. Update task → in_progress; update translation-progress.json → in_progress
2. Read source content; resolve draft path:

   uv run python scripts/draft.py --skill super-translate path <TARGET_FILE>
   

3. Dispatch translator (Agent tool, general-purpose) using ./translator-prompt.md
   • Inline all context: source, glossary, style, draft path
   • Translator must not read files; all context is pre-inlined
   • Translator must complete a block-shape self-check before returning: frontmatter, heading levels, list structure, blank-line boundaries, tables, code fences, admonitions, images, and MDX/import blocks must still align with the source
4. Read draft content after translator returns
5. Dispatch reviewer (Agent tool, general-purpose) using ./reviewer-prompt.md
   • Inline: source, draft, glossary, style
6. Dispatch Markdown reviewer by invoking the md-review skill using ../md-review/reviewer-prompt.md
   • Inline: source, draft, glossary, style, project conventions from AGENTS.md
   • This gate checks Markdown structure, frontmatter, heading hierarchy, block boundaries, lists, tables, links, image syntax, Starlight syntax, and zh-TW style rules
7. If either reviewer fails → dispatch refiner using ./refiner-prompt.md
   • Inline: source, draft, translation review JSON, md review JSON, glossary, style
   • Refiner must repair structure before wording polish when Markdown findings exist
   • Re-read draft → re-run reviewer → re-run Markdown reviewer. Cap at 2 total iterations.
8. If 2 iterations still fail, ask user:
   • 保留草稿，稍後手動修正
   • 停止此檔案，先處理術語、格式或規則歧義

Unknown terms: Run term_edit.py --set-zh workflow, then rerun file.

Parallel dispatch: When batch has 2+ independent files and no shared terminology conflicts, dispatch multiple translator agents concurrently using Agent tool. Reviewer/refiner remain sequential per file.

Verification: Per file: reviewer JSON and md-review JSON both return "pass": true, and no block-shape mismatch remains; otherwise iteration cap reached and user consulted.

Step 5: Controlled Writeback

Only if reviewer passes:

uv run python scripts/draft.py --skill super-translate writeback <TARGET_FILE>

Immediately update translation-progress.json: status → completed, recalculate _meta.completed, update _meta.updated. Update task → completed.

If blocked: keep source unchanged, status stays in_progress, mark task blocked.

Verification: Writeback script exits 0; translation-progress.json shows file as completed with updated _meta; task marked completed.

Step 6: Batch Checkpoint

After each batch:

1. Report: completed/blocked count, iteration counts, 已完成 X / Y 個章節
2. Stage only batch-touched files and commit:

   git commit -m "progress: X/Y"
   

3. If remaining files exist → ask user to continue (re-run Step 1) or proceed to verification.

Verification: git log -1 shows checkpoint commit with progress: X/Y message; report displayed to user.

Step 7: Final Verification

uv run python scripts/validate_glossary.py
uv run python scripts/term_read.py --fail-on-missing --fail-on-forbidden

Invoke check-consistency skill. Resolve violations before marking run complete.

Verification: Both validation commands exit 0; check-consistency reports no violations; all tasks marked completed.

Prompt Templates

Colocated with this skill. Orchestrator inlines all placeholders before dispatch:

• ./translator-prompt.md — draft generation
• ./reviewer-prompt.md — source fidelity + translation quality check
• ../md-review/reviewer-prompt.md — markdown structure + style compliance check
• ./refiner-prompt.md — apply both review streams

Flowchart

digraph super_translate {
    rankdir=TB;
    start [label="Resolve scope\n& preconditions", shape=box];
    preflight [label="Terminology\npreflight", shape=box];
    mode [label="Resolve\ntranslation mode", shape=box];
    translate [label="Dispatch\ntranslator", shape=box];
    review [label="Dispatch\nreviewer", shape=box];
    mdreview [label="Dispatch\nmd reviewer", shape=box];
    pass [label="Both pass?", shape=diamond];
    refine [label="Dispatch\nrefiner", shape=box];
    cap [label="Iteration\ncap?", shape=diamond];
    writeback [label="Writeback +\nupdate progress", shape=box];
    ask [label="Ask user", shape=box];
    checkpoint [label="Batch checkpoint\n& commit", shape=box];
    more [label="More files?", shape=diamond];
    verify [label="Final\nverification", shape=box];

    start -> preflight -> mode -> translate;
    translate -> review;
    review -> mdreview;
    mdreview -> pass;
    pass -> writeback [label="yes"];
    pass -> cap [label="no"];
    cap -> refine [label="< 2"];
    cap -> ask [label="= 2"];
    refine -> review;
    writeback -> checkpoint;
    checkpoint -> more;
    more -> start [label="yes"];
    more -> verify [label="no"];
}

Progress Sync Contract

1. Sync tasks and translation-progress.json at file start, every review loop, and file close.
2. NEVER defer sync until end-of-run.
3. Create batch checkpoint commit immediately after each completed batch.

Red Flags

When to Stop and Ask

• Repeated critical findings remain after iteration cap
• Subagent output is malformed and not safely recoverable
• Unknown term requires user decision (rare characters, puns, culturally nuanced)

References

See ./translator-prompt.md, ./reviewer-prompt.md, ../md-review/reviewer-prompt.md, and ./refiner-prompt.md for full dispatch context and placeholder specifications.