Reliable Tool Context

Build reliable tool context from command output using artifacts and compact reproducible code queries.

Use this skill to build reliable tool context whenever command output will be analyzed by the model and correctness matters. Capture once, query from artifacts, and return compact answers without copying raw payloads into model context.

CLI Decision Rule

Use Sift to establish reliable tool context when any of these are true:

• Output will be consumed by the model (analysis, transformation, or follow-up querying).
• Pagination exists or may exist (pagination.next.kind=="command").
• JSON schema/root confidence is low, or rows may be heterogeneous (even for small payloads).
• You need reproducibility, redaction discipline, or auditability.

Use direct CLI only when all of these are true:

• Output is clearly small.
• Schema/root path is obvious.
• It is a one-off human inspection with no follow-up model reasoning.

Required workflow

1. Capture output as an artifact:

sift-gateway run --json -- <command>

2. Keep only artifact_id plus a short summary in prompt context.
3. Handle pagination explicitly only when present. If pagination.next.kind=="command", continue with:

sift-gateway run --json --continue-from <artifact_id> -- <next-command-with-next_params-applied>

4. Query artifacts with explicit root path and compact output:

sift-gateway code --json <artifact_id> '$' --code "def run(data, schema, params): return [{'id': row.get('id'), 'status': row.get('status')} for row in data[:20]]"

5. Resolve root path from response hints, not guesswork:

• Current run behavior uses canonical root path $; use $ for follow-up code queries.
• If response_mode=="schema_ref" and schemas are present, use schema root_path as the source of truth.
• Treat sample_item as a preview row only.

Schema discovery protocol

• Do not use jq '.[0]' (or equivalent "first-item" shortcuts) to infer schema or root path. Many payloads are object-wrapped, have multiple candidate roots, or include heterogeneous rows where first-item heuristics are misleading.
• In Sift responses, sample_item is emitted only when Sift can verify consistent item shape across the resolved list. If sample_item is absent, inspect schemas; for current run captures, schema root_path should be $.
• If sample_item_text_truncated is true, treat long text fields as truncated previews and confirm details with a focused code query.

Guardrails

• sift-gateway code defaults to --scope all_related; start with --scope single for anchor-only analysis and widen scope only when needed.
• Start with run(data, schema, params); move to run(artifacts, schemas, params) only when cross-artifact joins are required.
• Use pure Python first; do not assume optional packages are available.
• Return aggregates or top <= 20 rows; avoid full-record dumps.
• Never paste raw captured payloads back into context.
• Do not claim completeness until pagination.retrieval_status == COMPLETE.