ucp-content 0.1.18

Python bindings for UCP (Unified Content Protocol) - Rust implementation

UCP Python bindings

Python bindings for the Rust UCP implementation, including both the generic UCP graph runtime and the specialized CodeGraph API.

For agent workflows, prefer the thin Python façade exposed by ucp.query(...) and ucp.run_python_query(...).

For model/runtime integration, also see:

• ucp.QueryLimits(...)
• ucp.PythonQueryTool(...)
• ucp.QueryBenchmarkCase(...) and ucp.run_query_benchmark_suite(...)

Installation

pip install ucp-content

Document usage

import ucp

doc = ucp.create("My Document")
root = doc.root_id
block = doc.add_block(root, "Hello, World!", role="paragraph")
doc.edit_block(block, "Updated content")
print(ucp.render(doc))

CodeGraph usage

import ucp

graph = ucp.CodeGraph.build("./repo")
session = graph.session()

session.seed_overview(max_depth=3)
session.expand("src/lib.rs", mode="file")

for node in graph.find_nodes(node_class="symbol", name_regex="Session|Context"):
    session.focus(node["logical_key"])
    session.apply_recommended(top=1, padding=2)

exported = session.export(compact=True, max_frontier_actions=6)
print(exported["summary"])
print(session.mutation_log()[-1]["operation"])

Agent-facing query façade

import ucp

graph = ucp.query(ucp.CodeGraph.build("./repo"))
session = graph.session()

for node in graph.find(node_class="symbol", name_regex="auth|login", limit=5):
    branch = session.fork()
    branch.add(node, detail="summary")
    branch.walk(node, mode="dependencies", depth=1, limit=8)
    if any("test" in (item.get("path") or "") for item in branch.export(compact=True)["nodes"]):
        session.add(node, detail="summary")
        session.walk(node, mode="dependencies", depth=1, limit=8)
        break

Core façade methods:

• graph.find(...), graph.describe(...), graph.explain_selector(...), graph.path(...)
• session.add(...), session.walk(...), session.focus(...), session.why(...)
• session.export(...), session.explain_export_omission(...), session.why_pruned(...)
• session.recommendations(...), session.estimate_expand(...), session.estimate_hydrate(...)
• session.mutation_log(), session.event_log()
• session.fork(), session.diff(...)
• session.hydrate(...) for CodeGraph

Python query runner

run = ucp.run_python_query(
    graph,
    """
candidates = graph.find(node_class="symbol", name_regex="auth|login", limit=5)
for node in candidates:
    session.add(node, detail="summary")
    session.walk(node, mode="dependencies", depth=1)
result = session.export(compact=True)
""",
    include_export=True,
)

print(run.ok)
print(run.summary)

The runner prebinds graph, session, re, json, math, and collections so the caller can use loops, conditionals, regex, and branching without writing a graph DSL.

It also accepts bindings={...} for parameterized queries and automatically dedents normal triple-quoted snippets before execution.

For CodeGraph, exported nodes also surface convenient top-level fields like logical_key, path, and symbol_name, so Python scoring/filtering code does not have to dig through nested coderef objects.

Guarded execution

run = ucp.run_python_query(
    graph,
    "result = graph.find(node_class='symbol', name_regex='auth', limit=5)",
    limits=ucp.QueryLimits(max_seconds=2.0, max_operations=40, max_trace_events=2000),
)

QueryLimits keeps model-authored queries bounded by wall-clock time, graph/session operations, traced Python events, and stdout size.

Provider-facing tool wrapper

tool = ucp.PythonQueryTool(
    graph,
    default_include_export=True,
    default_limits=ucp.QueryLimits(max_seconds=2.0, max_operations=40),
)

openai_tool = tool.openai_tool()
result = tool.execute({"code": "result = graph.find(node_class='symbol', name_regex='auth', limit=5)"})

The wrapper also supports:

• tool.execute_openai_tool_call(...)
• tool.execute_anthropic_tool_use(...)

Benchmark helpers

cases = [
    ucp.QueryBenchmarkCase(
        name="rank-tests",
        description="Rank likely tests for a symbol",
        code="result = graph.find(node_class='symbol', path_regex=r'tests/.*', limit=10)",
    )
]
results = ucp.run_query_benchmark_suite(graph, cases)
summary = ucp.summarize_query_benchmark_suite(results)

Example with parameterized regexes:

run = ucp.run_python_query(
    graph,
    """
        hits = graph.find(node_class="symbol", path_regex=path_rx, name_regex=name_rx, limit=6)
        best = next(node for node in hits if "context_show" in node["logical_key"])
        session.add(best, detail="summary")
        result = session.export(compact=True)
    """,
    bindings={
        "path_rx": r"crates/ucp-cli/src/commands/(agent|codegraph)\.rs",
        "name_rx": r"context_show|get_session_mut",
    },
)

Example: rank likely tests for a symbol with plain Python heuristics:

target = graph.find(node_class="symbol", path_regex=r"crates/ucp-python/python/ucp/query\.py", name_regex=r"^run_python_query$", limit=1)[0]
tests = graph.find(node_class="symbol", path_regex=r"crates/ucp-python/tests/.*\.py", name_regex=r"test_.*query.*", limit=80)
target_words = set(re.findall(r"[A-Za-z]+", target["logical_key"].lower()))
ranked = []
for node in tests:
    words = set(re.findall(r"[A-Za-z]+", (node.get("logical_key") or "").lower()))
    score = len((target_words & words) - {"symbol", "py", "python"})
    if score:
        ranked.append((score, node["logical_key"]))
print(sorted(ranked, reverse=True)[:5])

Generic graph usage

import ucp

doc = ucp.create("Graph demo")
section = doc.add_block(doc.root_id, "Section", role="section", label="section")
note = doc.add_block(section, "Important note", role="paragraph", label="note")
helper = doc.add_code(section, "rust", "fn helper() {}", label="helper")
doc.add_edge(note, ucp.EdgeType.References, helper)

graph = ucp.Graph.from_document(doc)
sqlite = graph.persist_sqlite("graph.db", "demo")
session = sqlite.session()
session.seed_overview(max_depth=1)
session.expand("note", mode="outgoing", depth=1)
print(session.export())

Generic graph features

• Graph.from_document(...)
• Graph.from_json(...), Graph.load(...), Graph.save(...)
• Graph.persist_sqlite(...), Graph.from_sqlite(...)
• find_nodes(...), describe(...), path_between(...)
• GraphSession with seed_overview, select, focus, expand, collapse, pin, prune, why_selected, and diff

CodeGraph features

• CodeGraph.build(...) from a repository
• find_nodes(...) with regex filters
• resolve(...) and describe(...)
• explain_selector(...) for selector provenance / ambiguity
• path_between(...) for short graph explanations
• CodeGraphSession for stateful exploration
• why_selected(...) provenance/explainability with provenance chains
• apply_recommended(...) and structured recommendations(...)
• estimate_expand(...) / estimate_hydrate(...) for budget-aware traversal
• mutation_log() / event_log() for observability
• explain_export_omission(...) and why_pruned(...)
• fork() and diff(...) for branch-and-compare workflows
• to_json(), from_json(...), save(...), and load(...)
• session to_json(), save(...), graph.load_session(...), and graph.load_session_json(...)

General features

• Document operations: create, edit, move, delete blocks
• Traversal: children, parent, ancestors, descendants, siblings
• Finding: by tag, label, role, content type
• Edges: create relationships between blocks
• LLM utilities: IdMapper, PromptBuilder, prompt presets
• Snapshots: snapshot and rollback helpers
• UCL execution: execute UCL commands on documents

Related docs

• docs/ucp-api/codegraph-programmatic.md
• docs/ucp-api/python-query-tools.md
• docs/ucp-api/graph-runtime.md
• docs/ucp-cli/codegraph.md
• scripts/demo_ucp_python_query.py
• scripts/demo_codegraph_python_query.py
• scripts/demo_codegraph_query_tool_wrapper.py
• scripts/demo_codegraph_query_benchmarks.py
• scripts/demo_codegraph_query_recipes.py
• scripts/demo_codegraph_query_edge_cases.py
• scripts/demo_codegraph_context_walk.py
• scripts/demo_codegraph_session_observability.py