CodePot
Theme
Python packageSupported stable generatorAvailable now

Infer a normalized generation model from OpenAPI and apply bundled or project-owned Jinja template packs.

GitHubPyPIAll links
python -m pip install codepotg
On this page

Performance, JSONL, and memory tracing

CodepotG has separate input paths for JSON and YAML.

JSON input

JSON is streamed into an indexed JSONL cache. Later stages can retrieve source records lazily rather than holding multiple complete decoded copies.

The cache is visible under .codepotg/cache so reuse and invalidation can be inspected.

YAML input

YAML must first be parsed. CodepotG writes a canonical JSON conversion incrementally and persists it as:

An unchanged YAML file reuses that conversion without parsing YAML again.

For very large contracts, JSON is the preferred source format.

Build or inspect JSONL

Use the JSONL command to compile a visible indexed cache and inspect source-index behavior independently from full generation.

Profile a pipeline

From packages/python/codepotg:

Profile stages

StageMeaning
startImported process baseline
jsonl_readyIndexed cache compiled or reused
document_loadedCompatibility OpenAPI object is in memory
graph_inferredSchema, operation, resource, and dependency graph built
contract_builtCompatibility and normalized contracts built
template_contract_builtLanguage adapter and template plan built
emission_completeOptional templates rendered and written
releasedLarge locals released and garbage collection completed

Metrics

MetricMeaning
rssCurrent process working set
rss_peakOperating-system high-water working set
privatePrivate committed/data bytes where available
pythonCurrent traced Python allocations
python_peakPeak traced Python allocations

rss_peak never decreases. Python and the system allocator may retain freed arenas, so current RSS does not always return to startup levels.

A likely retained-object issue is indicated when both traced Python memory and private memory remain near their peaks after released, or repeated runs continue growing instead of stabilizing.

Trace normal generation

  • 1 records process metrics with lower overhead.
  • full enables tracemalloc and should be used for diagnosis.

Write JSONL snapshots:

Memory summaries are included in diagnostics while tracing is enabled.

Template performance

  • Prefer graph selections over repeated global scans.
  • Declare providers and dependencies explicitly.
  • Avoid copying api.raw into derived contexts.
  • Use grouped emissions for intentional aggregation.
  • Keep barrel planning based on emission facts, not filesystem scans.
  • Avoid quadratic Jinja loops over large global collections.

Cache behavior

Use --refresh for project cleanup, not as a general source-cache invalidation flag. Source caches use content and metadata checks and should be reused when inputs are unchanged.

Fixture strategy

The package test suite separates production-scale streaming fixtures from small targeted unit fixtures. Custom pack tests should follow the same principle: one representative full contract plus focused edge cases.