flowscript-ldp

First implementation of LDP Mode 3 (Semantic Graphs) using FlowScript IR.

Reference implementation for Mode 3 of the LLM Delegate Protocol (Prakash, 2026). Mode 3 — "structured relationship representations for planning and formal reasoning" — is specified but not yet evaluated in the paper. This package provides the first working implementation: a queryable semantic graph format for inter-agent communication.

What is FlowScript?

FlowScript is a semantic notation that compiles to a typed intermediate representation (IR). The IR is a graph with three collections:

• Nodes (12 types): statements, questions, thoughts, decisions, blockers, insights, actions, completions, alternatives, exploring, parking, blocks
• Relationships (10 types): causes, temporal, derives_from, bidirectional, tension, equivalent, different, alternative, alternative_worse, alternative_better
• States (4 types): blocked, decided, exploring, parking

Every element has provenance metadata (source file, line number, timestamp) and SHA-256 content-addressed deduplication. This gives you a queryable graph where you can trace causal chains, map tradeoffs, find blockers, and reconstruct decisions computationally — without parsing natural language.

Quick Start

Load a pre-compiled IR graph and start querying — no external tools needed:

import json
from flowscript_ldp import FlowScriptPayload

with open("examples/sample_ir.json") as f:
    ir_data = json.load(f)

payload = FlowScriptPayload.from_dict(ir_data)

# Find all tradeoffs in the graph
tensions = payload.query.tensions()
# → 3 tensions: "cost vs control", "latency vs cost", "performance vs freshness"

# Track blockers with impact scores
blocked = payload.query.blocked()
# → 1 blocker: "add cache hit/miss monitoring to Datadog"

# Trace causal ancestry
node_id = "fdc98c25..."  # "centralized cache invalidation"
why = payload.query.why(node_id, format="minimal")
# → root_cause: "Redis cache layer"

# Encode for LDP transport
envelope = payload.encode()
# → {"ldp_version": "1.0", "payload_mode": 3, "payload_format": "flowscript-ir", ...}

If you have the FlowScript CLI installed, you can also parse .fs files directly:

from flowscript_ldp import ParserBridge, FlowScriptPayload
bridge = ParserBridge()
ir = bridge.parse_file("thinking.fs")
payload = FlowScriptPayload(ir)

Why Mode 3?

The LLM Delegate Protocol defines 6 payload modes (0–5) for inter-agent communication:

Mode	Name	Status
0	Text	Evaluated in paper
1	Semantic Frames	Evaluated in paper
2	Embedding Hints	Specified, unimplemented
3	Semantic Graphs	Specified, first implementation here
4	Latent Capsules	Future work
5	Cache Slices	Future work

Modes 0–1 pass text or structured JSON between agents. Mode 3 passes queryable graphs — agents can trace causality, find tradeoffs, and reconstruct decisions computationally instead of inferring them from prose. Five operations make the structure computable:

Query	What it does	Example
why(node_id)	Trace causal ancestry backward	root_cause: "Redis cache layer"
what_if(node_id)	Trace downstream impact forward	"affects 4 downstream considerations"
tensions()	Extract all tradeoffs	"cost vs control", "latency vs cost"
blocked()	Find blockers with impact scores	"Datadog trial expired" (impact: 0)
alternatives(question_id)	Reconstruct decisions	3 options considered, chosen: "Redis"

Each query supports multiple output formats (chain/tree/minimal for why, tree/list/summary for what_if, axis/node/flat for tensions, comparison/tree/simple for alternatives).

Fallback Chain

Per LDP spec, when Mode 3 fails or the receiver doesn't support it, the protocol degrades gracefully:

Mode 3 (Semantic Graph) → Mode 1 (Semantic Frame) → Mode 0 (Natural Language)

from flowscript_ldp import FallbackChain

fallback = FallbackChain(ir)

# Mode 3 → Mode 1: Structured semantic frame
mode1 = fallback.to_mode1()
# → {"task_type": "decision_analysis", "instruction": "caching strategy for...", ...}

# Mode 3 → Mode 0: Natural language prose
mode0 = fallback.to_mode0()
# → "Question: caching strategy for read-heavy API endpoints\n  Option: Redis cache layer\n  ..."

Provenance and Quality

The LDP paper's key finding: noisy provenance degrades synthesis quality below the no-provenance baseline. FlowScript IR's temporal graduation model — observations must survive quality gates to persist — acts as a provenance noise filter. Mode 3 payloads carrying pre-filtered relational structure sidestep the degradation the paper identifies.

JamJet + LDP Integration

As of v0.2.0, flowscript-ldp integrates with both JamJet v0.2.0's ProtocolAdapter interface and the standalone ldp-protocol SDK.

LDP Delegate (server-side)

Run a Mode 3 delegate as an HTTP service that any LDP client can discover, negotiate with, and submit tasks to:

from flowscript_ldp.delegate import FlowScriptMode3Delegate

delegate = FlowScriptMode3Delegate()
delegate.run(port=8090)  # HTTP server with /ldp/identity, /ldp/capabilities, /ldp/messages

The delegate advertises 6 skills (flowscript.tensions, flowscript.blocked, flowscript.why, flowscript.what_if, flowscript.alternatives, flowscript.degrade) and negotiates Mode 3 (Semantic Graph) payloads during session establishment.

from ldp_protocol import LdpClient

async with LdpClient() as client:
    # Discover delegate capabilities
    identity = await client.discover("http://localhost:8090")

    # Submit a query task
    result = await client.submit_task(
        "http://localhost:8090",
        skill="flowscript.tensions",
        input_data={"ir": ir_json},
    )
    # → {"output": {"tensions": [...], "metadata": {...}}, "provenance": {...}}

JamJet ProtocolAdapter (client-side)

Register flowscript-ldp with JamJet's protocol registry so workflows can route to LDP delegates via ldp:// URLs:

from flowscript_ldp.adapter import FlowScriptLdpAdapter

FlowScriptLdpAdapter.register()  # registers for ldp:// and ldp+flowscript:// URL prefixes

JamJet @tool integration

FlowScript query operations are also available as JamJet-compatible @tool functions for use in agent workflows:

from jamjet import Agent
from flowscript_ldp import get_jamjet_tools

agent = Agent(
    "analyst",
    model="claude-haiku-4-5-20251001",
    tools=get_jamjet_tools(),  # 6 async tools: tensions, blocked, why, what_if, alternatives, degrade
    instructions="Analyze the semantic graph for tradeoffs and blockers.",
)
result = await agent.run(f"Analyze this: {ir_json}")

Standalone use

The sync query functions and adapter work without JamJet or ldp-protocol:

from flowscript_ldp.adapter import flowscript_tensions, FlowScriptMode3Adapter

# Direct function call
result = flowscript_tensions(ir_data)

# Adapter with query dispatch
adapter = FlowScriptMode3Adapter()
result = adapter.invoke(envelope, query="tensions", fallback_mode=1)

See examples/standalone_demo.py for a runnable demo of all 5 queries.

CLI

flowscript-ldp info graph.json                              # IR statistics
flowscript-ldp query tensions graph.json                    # Find tradeoffs
flowscript-ldp query blocked graph.json                     # Find blockers
flowscript-ldp query why <node_id> graph.json               # Trace causes
flowscript-ldp query what-if <node_id> graph.json           # Impact analysis
flowscript-ldp query alternatives <question_id> graph.json  # Decision reconstruction
flowscript-ldp encode graph.json                            # Wrap in Mode 3 envelope
flowscript-ldp degrade graph.json --mode 0                  # Degrade to natural language

Installation

pip install flowscript-ldp              # core (IR, queries, payload, fallback, CLI)
pip install flowscript-ldp[ldp]         # + LDP delegate (ldp-protocol SDK)
pip install flowscript-ldp[jamjet]      # + JamJet ProtocolAdapter
pip install flowscript-ldp[all]         # everything

From source:

pip install git+https://github.com/phillipclapham/flowscript-ldp.git

Core dependency: pydantic>=2.0. JamJet and ldp-protocol are optional — the core package (IR models, query engine, payload, fallback, adapter, CLI) works standalone. The ParserBridge optionally requires the FlowScript CLI for parsing .fs text files into IR.

Architecture

flowscript_ldp/
├── ir.py              # Pydantic models for FlowScript IR schema
├── parser_bridge.py   # Subprocess bridge to FlowScript CLI (optional)
├── query.py           # 5 query operations, 3 formats each (Python port of TypeScript engine)
├── payload.py         # Mode 3 payload encode/decode/envelope
├── fallback.py        # Mode 3 → Mode 1 → Mode 0 degradation
├── adapter.py         # Sync tools + get_jamjet_tools() + FlowScriptLdpAdapter(ProtocolAdapter)
├── delegate.py        # FlowScriptMode3Delegate(LdpDelegate) — LDP server
├── round_trip.py      # Round-trip verification utilities
└── cli.py             # Command-line interface

168 tests covering IR models, all 5 query operations with all format variants (edge cases: cycles, diamond graphs, empty graphs, depth limiting), payload round-trips, fallback chain, adapter dispatch, JamJet tool integration, LDP delegate skills/identity/negotiation, ProtocolAdapter client-side bridge, and full HTTP integration round-trips (delegate server + LdpClient + adapter).

References

• LDP Paper: arXiv:2603.08852 — Sunil Prakash, March 2026
• ldp-protocol: pypi.org/project/ldp-protocol — Standalone Python SDK for LDP (LdpDelegate, LdpClient, LdpRouter)
• FlowScript: github.com/phillipclapham/flowscript — Semantic notation for cognitive graphs
• JamJet: github.com/jamjet-labs/jamjet — Agent-native runtime (Rust core, Python SDK)

License

MIT