[alexa-sift]

Background: Every list_* MCP tool builds its response through list_body. A caller can pass fields: ["name", "asset_id"] to keep only those keys on each row. If a requested name matches nothing on any row (a typo like nmae), list_body adds unmatched_fields: ["nmae"]. Without that key, a typo would quietly return rows with that field missing, and the agent could report the data as empty when the field name was wrong. unmatched_fields lets the agent catch the mistake and retry, or at least tell the user.

The problem: serde_json sorts keys alphabetically, so unmatched_fields comes after the items array. In 7 of the 14 list tools (channels, assets, annotations, and others), the items key also sorts before count and has_more. The agent output guard cuts large responses from the end, so these keys disappear on exactly the big responses where they matter most.

The fix: list_body now returns a ListBody that carries the items key, and its into_result method writes the text block agents read with every metadata key first and the items array last. All 14 list tools return through it, so callers name the items key once and can't let it drift. Tools that add keys after the fact (artifact_url, next_step) do so through ListBody::insert. count, has_more, unmatched_fields, artifact_url, and next_step now survive head truncation regardless of how the items key sorts or whether serde_json's preserve_order is ever enabled. structuredContent is unchanged, and the key keeps its existing name. sift_mcp/CLAUDE.md now teaches the same pattern.

Verification

• New into_result tests in common/test.rs:
  • All 14 items keys write count, has_more, and unmatched_fields before the items.
  • With 500 rows, all three sit in the first 200 characters.
  • The text block parses to the same JSON as structuredContent.
  • next_step gets its own leading block.
• Handler-level assertions on the text block for list_channels, list_artifact_versions (artifact_url first), and list_calculated_channel_versions (next_step block first).
• With the text block reverted to alphabetical output, 6 of the new tests fail.
• cargo fmt --check --all, cargo check --all-features, and cargo clippy --all-features pass with no new warnings. cargo test -p sift_mcp passes 605 tests.