Synapsor
Log inGet access
All docs
Docs
Core Concepts

Agent syntax glossary

A concise glossary for Synapsor-specific SQL extensions that are not standard SQL.

Overview

Synapsor adds agent-native DDL around standard relational data. The new words describe trust boundaries, model-visible arguments, hidden session values, evidence capture, write-proposal targets, deterministic plans, and replayable resources.

The pattern is deliberate: ARG is caller/model-visible input, HIDDEN and BINDING are trusted runtime values, TENANT and LOOKUP constrain write targets, OPERATION limits what can change, PAYLOAD/EVIDENCE separate response data from audit data, and AS OF AGENT RUN creates a historical read boundary.

Replay itself is API/SDK-first through POST /v1/agent/runs/replay. Use SQL for historical reads and for creating an investigation branch from an agent-run snapshot.

Names such as support.ticket_context, billing.propose_late_fee_waiver, and billing.green_late_fee_settle are developer-defined. Handles such as wrp://1, evidence://bundle/1, ctx://..., and agent-run://1 are returned by Synapsor and can be found again in system views.

Use this page as a glossary when asking an LLM to write Synapsor examples. Pair it with /llms.txt and the quickstart so generated code keeps tenant scope, evidence, and writes inside reviewed capabilities.

Syntax roles

Synapsor syntax separates caller input, trusted scope, capability execution, and durable evidence.

Inputs and scope
  • ARG: visible caller input
  • HIDDEN: value read from SESSION
  • BIND: trusted context slot
  • BINDING: trusted slot used later by proposals
  • TENANT: tenant guard for a write target
Plan and output
  • PLAN: deterministic Synapsor execution block
  • DEFAULT DECISION: fallback if rules do not terminate
  • RULE: named reason-code transition
  • PAYLOAD: immediate response fields
  • EVIDENCE: replayable why trail
  • AS OF AGENT RUN: historical read boundary

Developer notes

  • Use ARG for values the caller may choose; use HIDDEN/BINDING for values the caller must not control.
  • Use TENANT and LOOKUP in every proposal target so writes cannot drift across tenants or rows.
  • Use PAYLOAD for structured decisions and EVIDENCE for the durable explanation trail.
  • Use the replay API for run replay and AS OF AGENT RUN SQL for historical reads or branch-from-run investigation.
  • When asking an LLM to produce Synapsor SQL, provide /llms.txt plus this glossary.