ariadne-sensemaking 0.1.1

Sensemaking harness for nonatomic entities: an agentic analytic interface over heterogeneous, multimodal data

Ariadne

Sensemaking for nonatomic entities using AI coding agents.

Docs: https://ajbarea.github.io/ariadne/

Ariadne gave Theseus the thread that let him traverse the Labyrinth and find his way back out. This project is that thread: a single line of reasoning that traces one entity's evidence through a maze of heterogeneous data systems and back into a coherent analytic picture.

[!NOTE] Every architectural decision is grounded in current best-practice research, captured in docs/research/: new runtime dependencies and tools aren't adopted until grounded there with a # research(YYYY-MM): note. See ROADMAP.md.

The problem

Intelligence about a single entity (a person, a unit, an organizational node) is scattered across heterogeneous systems (graph, relational, and unstructured stores) and modalities (metadata, text, imagery, video). No single interface spans that range, so analysts pivot between systems by hand and lose the thread at every step. The hard part isn't access; it's multi-hop reasoning across representations, where the decisive link between two facts often exists only through an implicit organizational relationship buried in another store and another modality.

The approach

Use an agentic AI harness — the Claude Agent SDK — as a unifying analytic interface, not a replacement for existing infrastructure. The harness is an orchestration layer that dispatches specialized tools and skills to retrieve, interpret, and synthesize across graph, structured, and unstructured sources in one coordinated workflow.

Central research question: given the harness and its UI, what specific tools, skills, and hooks are necessary to support a rigorous end-to-end analytic workflow targeting entities within an organizational hierarchy, and what is the minimum viable toolset that demonstrates real analytic value?

Deliverable

A working prototype that takes a target entity or organizational node as input and, through a coordinated sequence of tool invocations, surfaces relevant evidence from across all available data structures and modalities, synthesizing findings into a coherent analytic product. Success is measured by the harness's ability to:

• traverse organizational relationships,
• reconcile information across modalities,
• reduce the analyst's manual-pivot burden, and
• surface non-obvious connections impractical to find with conventional tooling.

Umbrella role

Ariadne is conceived as an umbrella effort. It does not duplicate companion-project work; it defines integration interfaces so that external capabilities — graph-extraction pipelines, entity-resolution models, multimodal indexing schemes — can be surfaced as callable tools within the harness. It is both a standalone research contribution and a unifying demonstration layer for a wider analytic toolchain.

Install

pip install ariadne-sensemaking      # or: uv add ariadne-sensemaking

This installs the ariadne CLI and the ariadne-sensemaking MCP server entry point.

Use from any AI CLI

The ariadne MCP server exposes the workup tool to any MCP client: Claude Code, Copilot, Gemini CLI, Cursor, Windsurf, and others. Two install paths:

Claude Code (one-click): add this repo's marketplace, then install the plugin:

/plugin marketplace add ajbarea/ariadne
# then: /plugin install ariadne

This bundles the Ariadne MCP server + the analyst-workup skill into Claude Code in a single step. A workup is then available via the ariadne server tool.

Any other MCP client: point the client's MCP config at the published package, and uvx will fetch and run it with no clone or manual install:

{
  "mcpServers": {
    "ariadne": { "command": "uvx", "args": ["ariadne-sensemaking"] }
  }
}

To track an unreleased commit instead, point uvx at the repo with "args": ["--from", "git+https://github.com/ajbarea/ariadne", "ariadne-sensemaking"]; to run from a local checkout, use "command": "python", "args": ["-m", "ariadne.mcp_server"].

Config caveat: the tool is portable; the data is not. Point the server at your stores (NEO4J_* / DATABASE_URI) and set ANTHROPIC_API_KEY per install. See ADR-0009.

Observability

Ariadne emits OpenTelemetry traces and metrics via the optional otel extra. The base install (api-only) is a no-op — nothing is emitted unless you configure an OTLP endpoint.

# 1. install the otel extra
uv sync --extra otel

# 2. point at your OTLP collector (Jaeger, Grafana, Datadog, etc.)
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318

# 3. optionally add the Claude Agent SDK's per-LLM-call spans
export CLAUDE_CODE_ENABLE_TELEMETRY=1

What is emitted per workup:

• invoke_agent span — covers the full workup; duration = task time.
• ariadne.workups counter and ariadne.workup.duration histogram — latency and throughput.
• ariadne.evidence.calls counter — evidence-tool calls (from the provenance ledger); surfaces existing provenance.jsonl data.
• ariadne.citation.failures counter — uncited + unsupported claims; surfaces existing citations.json gate output.
• Span attributes for citation compliance (citation.ok, citation.uncited, citation.unsupported) and tradecraft compliance (tradecraft.estimative_terms, tradecraft.has_confidence) — surfaces existing tradecraft.json output.

When CLAUDE_CODE_ENABLE_TELEMETRY=1 is set, the Agent SDK's per-LLM-call spans nest under Ariadne's invoke_agent span — one trace covers the full workup without any extra config.

See ADR-0010 for the design rationale and considered alternatives.

Quickstart

Work up a target entity against a graph store, open an interactive report, and score the run. This first run is graph-only (one store, ~90s); add --sql --semantic once you have Postgres up for the full cross-store workup.

# 1a. start a local Neo4j
docker compose -f infra/neo4j/docker-compose.yml up -d
# 1b. seed the synthetic org graph
docker exec -i ariadne-neo4j cypher-shell -u neo4j -p password < infra/neo4j/seed.cypher

# 2. the live agent loop needs an API key
export ANTHROPIC_API_KEY=sk-...

# 3. work up a target entity. The output is an immutable run dir
#    runs/synthetic/halberd/<run-id>/ (note.md, provenance.jsonl, citations.json,
#    manifest.json, ...), with runs/synthetic/halberd/latest pointing at the newest run.
uv run ariadne workup Halberd

# 4. open the self-contained interactive report: entity-network view, clickable
#    provenance + evidence drawer, reconciliation panel, eval dashboard. The fastest
#    way to see what the workup found.
uv run ariadne report runs/synthetic/halberd/latest
#    -> writes runs/synthetic/halberd/latest/report.html; open it in a browser

# 5. score the run: planted-needle grounding (+ optional cross-store reconciliation)
uv run ariadne eval runs/synthetic/halberd/latest --reconcile synthetic
#    and the ICD-203 analytic-quality rubric (LLM judge; needs the 'rubric' extra)
uv run ariadne rubric runs/synthetic/halberd/latest

# 6. enforce the read-only contract offline (CI gate, no API key; exit 3 on a write attempt)
uv run ariadne governance runs/synthetic/halberd/latest

Every fact in the note carries a [cite:gN] id that resolves to a recorded graph query in provenance.jsonl; the run fails if any citation is uncited, dangling, or (with the optional eval extra) unsupported by its cited evidence. ariadne eval then scores whether the run surfaced and actually traversed the seed's planted non-obvious bridge; grounded=True means it reasoned, not guessed.

The analytic loop is read-only by construction (the graph and relational MCP connectors run in read-only / restricted mode); ariadne governance verifies that on a persisted run by auditing the provenance ledger for any mutating verb, and gates by default (exit 3 on a write attempt). Pass --strict to ariadne workup to apply the same gate to the live run.

Dev gates (requires uv):

make setup        # install dependencies (uv sync)
make lint         # ruff format --check + ruff check + ty
make test-unit    # fast hermetic tests
make validate     # lint + unit tests (pre-push gate)

Archived, auditable runs (reconciled by /techne:audit):

./scripts/dev-runner.sh lint   # writes logs/dev-<UTC>-lint.log + logs/dev-latest.log

Repository layout

Path	Purpose
src/ariadne/	The package: CLI + MCP server, read-only store connectors (graph, relational, unstructured), the adapt layer (introspect, mapping), bounded self-improvement (learning), and the eval / provenance / report modules.
plugins/ariadne/	The Claude Code plugin: the bundled MCP server config + the analyst-workup skill.
.claude/skills/entity-workup/	The analytic procedure (gather, act, verify, synthesize) the agent loop follows.
infra/	docker-compose for a local Neo4j + Postgres.
docs/architecture/decisions/	MADR ADRs: every contestable choice and its rejected alternatives.
docs/research/	The June-2026 best-practice research backing each design decision.
tests/	Unit + integration suite.
scripts/	dev-runner.sh and other repo scripts.

Documentation site

Built with Zensical and deployed to GitHub Pages on every push to main touching docs/. Serve locally:

make docs        # uv run --with zensical zensical serve

Status

Ariadne works up an entity end to end: graph, relational, full-text, and semantic retrieval; a citation-gated analytic note; a self-contained interactive report; planted-needle and ICD-203 evaluation; and a read-only governance gate.

Beyond the built-in datasets it adapts to your own Postgres (introspect, propose a mapping, ratify, then run the existing pipeline unchanged) and learns from experience (distil a skill from a strong run, reflect on a weak one, deepen a skill, and use ariadne compare to measure whether a learned change actually helps before adopting it).

See IMPL.md for what's in flight and ROADMAP.md for the build order.

License

MIT © 2026 AJ Barea. See LICENSE.