cpsat-logutils

Parse CP-SAT (OR-Tools) solver logs into line-anchored, JSON-serializable pydantic models.

pip install cpsat-logutils

from cpsat_logutils import parse_log

log = parse_log(open("solver.log").read())

log.solver.version.value  # "9.15.6755"
log.solver.parameters.value  # {"max_time_in_seconds": 5, "num_workers": 8, ...}
log.response.status.value  # "OPTIMAL"
log.response.status.line  # 346  <- every value knows its line
log.search.events[0].objective  # first solution value
log.stats.search_stats.row("core").values["Conflicts"]
log.block_at(line=42)  # -> BlockRef(kind="presolve", span=..., path="/presolve")
log.model_dump_json(indent=2)  # plain JSON

Sections that do not occur in a log are simply None. Nothing raises on a log the parser does not fully understand: unknown sections are kept verbatim in log.unparsed.

Getting a log

from ortools.sat.python import cp_model

solver = cp_model.CpSolver()
solver.parameters.log_search_progress = True
solver.parameters.log_to_stdout = False
lines: list[str] = []
solver.log_callback = lines.append
solver.solve(model)

log = parse_log("\n".join(lines))

A log copied out of a terminal works just as well, including one that was clipped, prefixed by a logging framework, or has your own prints mixed in.

Why line anchors?

Log analyzers want to show the raw text next to the parsed data. Every scalar is a Loc[T] (value + line), table rows carry their line, every section has a span, and CpSatLog.blocks is an ordered index (kind, span, JSON pointer path) that maps any line back to the parsed section.

Supported versions

OR-Tools 9.3 to 9.15. Known format traps (older [Name] ... time= presolve lines, Solutions found per subsolver: lists, progress lines inside the presolve block, LRAT_status, renamed table columns) are handled. Every release is tested against the example logs in example_logs/ and against 295 real logs from public instance libraries in corpus/.

Extending

splitter.py cuts text into chunks; add forced cut points there when a new version glues sections together.
parsers/ holds one BlockParser per section; register new ones in parsers/__init__.py (first match wins).
parsers/tables.py::TABLE_IDS maps table titles to stable ids; add a line for a new table and give it a field in schema/tables.py::FinalStats.
assemble.py places blocks into the root model and merges multi-chunk sections.

Development

uv sync
uv run pytest
uv run ruff check . && uv run ty check

The CP-SAT Log Analyzer is built on this library: it explains a parsed log section by section in a web UI, and is hosted at https://cpsat-loganalyzer.krupke-algorithms.de/.

History

Version 1.0 is a rewrite. Up to 0.0.4 this package exposed a block-based API (LogParser, SolverBlock, SearchProgressBlock, ...) that returned loosely typed dictionaries; it is gone. The parser now produces one validated pydantic model per log, in which every value carries the line it was read from. Pin cpsat-logutils<1 if you depend on the old API.