Feat: TOON support

[christso]

Use package https://github.com/toon-format/toon for both golden dataset and output files.

This can reduce token usage by 60%.

[christso]

Agent-Ready Implementation Plan

Branch

git checkout -b feat/139-toon-support

Context

TOON is a token-efficient format (https://github.com/toon-format/toon). The claim is 60% token reduction for golden datasets and output files.

Reference files to read first

1. packages/core/src/evaluation/yaml-parser.ts — How eval files are currently loaded (YAML, JSONL)
2. packages/core/src/evaluation/types.ts — Dataset/case types
3. apps/cli/src/commands/eval/jsonl-writer.ts — How results are written
4. TOON package docs: https://github.com/toon-format/toon

Step-by-step implementation

Step 1: Research TOON format

cd /tmp && bun add toon

Read the TOON package API:

• How to parse TOON → JSON
• How to serialize JSON → TOON
• File extension conventions (.toon?)
• What makes it token-efficient

Step 2: Add TOON as input format

In packages/core/src/evaluation/yaml-parser.ts:

• Add .toon to detectFormat() function
• Add TOON parsing alongside YAML and JSONL:

  if (filePath.endsWith('.toon')) {
    const { parse } = await import('toon');
    const content = await readFile(filePath, 'utf-8');
    return parse(content);
  }

Step 3: Add TOON as output format

In the output writer system (alongside JSONL, JSON, XML from #232):

• Add .toon extension support
• Serialize results to TOON format

Step 4: Add dependency

cd packages/core && bun add toon

Step 5: Add tests

• Test: TOON file loads correctly as eval dataset
• Test: Results serialize to TOON format
• Test: Round-trip (TOON → parse → serialize → parse) preserves data

Verification checklist

bun run typecheck
bun run lint
bun run test
bun run build

Notes

• TOON support is additive — YAML and JSONL remain the primary formats
• Research the package thoroughly before implementing — the API may be different than expected
• Consider whether TOON is suitable for both input datasets AND output results, or just one
• This is a lower-priority item (P4) — keep implementation minimal

[christso]

Closing — none of the source repos (agentv, pi-mono, promptfoo) use the TOON format. No demand signal to justify the integration effort.