Pulse Documentation

name: tool-process-chain
kind: tool
description: Linear chain of stages; stage N+1 consumes stage N's output rows.
type: reference
applies_to: process-chain, process, mcp

When to use

CALL WHEN EACH STAGE SHOULD READ THE PREVIOUS STAGE’S OUTPUT ROWS. Collapses N round-trips into one open + N stage validations. Common: “group-then-regress”, “filter-then-aggregate-then-attribute”. Each stage’s synthesized output schema is the next stage’s cohort.

Input

request (string): JSON-encoded pulse.ChainRequest. Fields: cohort.filename (source for stage 0), stages ([]ChainStage with name + request). Stage 0 supplies the source cohort; later stages ignore their inner cohort field. A stage’s request.return shapes that stage only after the whole chain ran (the next stage still reads full rows); final follows the last stage’s.

Output

descriptor.Envelope wrapping per-stage Responses plus a whole-chain summary. Chain-host overlays fold at the post-chain barrier.

Gotchas

• MERGEABLE-ONLY at v1. Permitted: exactly the manifest process_chain block’s mergeable_aggregators, mergeable_groupers and row_local_attributes. Windows, features, statistical tests, regressions, two-pass attributes and non-mergeable aggregators/groupers → PULSE_CHAIN_NOT_MERGEABLE. Fall back to per-stage requests (batched by pulse_compose).
• Synthesized schema between stages: grouper keys → categorical_u32 columns; aggregator outputs → f64.

See

• process-chain — full mergeability matrix and recipe library.
• response-components — per-stage emission contract.
• overlay-system — CHAIN-host overlays.