CastForge

Open-source Python framework for repeatable, source-transparent podcast pipelines.

CastForge separates reusable production mechanics from show-owned editorial policy. It can normalize cited sources, record an auditable episode manifest, generate NotebookLM audio, publish MP3s to Cloudflare R2, update RSS atomically, and validate the result. Each show keeps its own sources, prompts, identity, schedule, and feed.

CastForge powers Nitan Podcast, a production Chinese podcast generated from USCardForum discussions.

Install

CastForge 0.1.3 adds schema-v2 editorial metadata, minimum-story no-episode gating, measured MP3 duration, a fail-closed NotebookLM short retry above a configured duration ceiling, and portable CA validation for public audio. Releases are published on PyPI with trusted-publishing attestations:

pip install castforge==0.1.1

# Optional production integrations
pip install "castforge[notebooklm,r2]==0.1.3"

Python 3.10 or newer is supported. Gemini, NotebookLM, and R2 dependencies remain optional.

First episode in under 20 minutes

mkdir my-show && cd my-show
castforge init
castforge run --config podcast.yaml --date 2026-08-11
castforge validate --config podcast.yaml --date 2026-08-11

The fixture-backed starter creates:

• a cited NotebookLM source document;
• a deterministic episode manifest;
• an RSS feed with a positive audio enclosure contract.

Fixture mode does not call an LLM, audio service, or public endpoint. Replace the example sources and public values before production use.

Core contracts

from castforge import EpisodeManifest, SourceItem, StoryCluster
from castforge.contracts import AudioProvider, Publisher, SourceAdapter

• SourceItem normalizes one primary, independent, or trend-signal source.
• StoryCluster groups duplicate coverage and qualifies only a primary source or two independent reports.
• EpisodeManifest preserves the selected stories, citations, source document, pipeline version, and public audio identity.
• SourceAdapter, AudioProvider, and Publisher are intentionally small protocols implemented by show repositories or integrations.

The generic CLI consumes podcast.yaml; show-specific collectors and ranking remain in the show repository.

Commands

castforge init [directory]
castforge run --config podcast.yaml --date YYYY-MM-DD [--shadow]
castforge validate --config podcast.yaml [--date YYYY-MM-DD] [--check-public]

--shadow creates the source and manifest artifacts but does not mutate RSS or R2. Same-date production reruns replace the date-keyed RSS item instead of duplicating it.

Production configuration

See examples/podcast.yaml for the complete schema.

For NotebookLM:

audio:
  provider: notebooklm
  output_dir: build/audio
  duration: 00:06:00
  public_url_template: https://audio.example.com/episodes/{filename}
  language: en
  audio_length: short

Install and authenticate the integration once on the runner:

pip install "castforge[notebooklm]"
playwright install chromium
notebooklm login

Set NOTEBOOKLM_NOTEBOOK_ID. Authentication state and notebook ownership stay outside the show repository.

For Cloudflare R2:

publication:
  provider: r2
  bucket: podcast-audio
  endpoint_url: https://ACCOUNT_ID.r2.cloudflarestorage.com
  public_base_url: https://audio.example.com
  access_key_env: R2_ACCESS_KEY_ID
  secret_key_env: R2_SECRET_ACCESS_KEY
  download_url_prefix: https://op3.dev/e/
  max_bucket_bytes: 9000000000

CastForge totals the bucket before each upload and refuses a projected size above max_bucket_bytes; overwrites subtract the prior object and listings paginate safely. It then uploads MP3s as audio/mpeg and sends a public HEAD request to verify status, MIME type, and byte length before updating RSS. A show may apply a privacy-respecting download redirect such as OP3 after the R2 origin passes validation. A failed capacity check, generation, upload, or public check leaves the feed unchanged.

Existing hook pipeline

Production shows can continue wiring show-specific extraction and publishing through PipelineHooks while migrating to config-driven artifacts:

from castforge.pipeline import PipelineHooks, main as castforge_main

def main(argv=None):
    hooks = PipelineHooks(
        extract_weekly_key_info=extract,
        fetch_thread_details=fetch_details,
        list_mcp_tools=list_tools,
        select_threads=select,
        threads_to_source_markdown=to_markdown,
        write_forum_post=write_post,
        generate_rss_feed=generate_feed,
        episode_file_prefix="weekly_episode",
        week_episode_filename=episode_filename,
        week_episode_url=episode_url,
    )
    return castforge_main(argv, hooks=hooks)

CastForge owns execution and provider integrations. The show owns sources, editorial policy, branding, secrets, scheduling, feeds, episodes, and public compatibility.

Development

python -m pip install -e ".[test]"
python -m pytest
python -m build

Tests are offline and use fake provider clients. Live NotebookLM and R2 checks require explicit credentials and are not part of the ordinary suite.

License

MIT — see LICENSE.