Skip to content

feat(workflows): graph nodes that publish to feeds and make episodes - #547

Merged
adeelahmad merged 1 commit into
mainfrom
claude/project-thread-rovv80
Oct 18, 2026
Merged

adeelahmad merged 1 commit into
mainfrom
claude/project-thread-rovv80

Conversation

@adeelahmad

@adeelahmad adeelahmad commented Oct 18, 2026 •

Copy link
Copy Markdown
Owner

Description

Before: a feed was filled by hand, through POST /api/v1/feeds/{id}/items or the publish_to_feed MCP tool, and a podcast episode was asked for one at a time. Nothing ran on a schedule.

After: a graph workflow picks part of the graph and publishes it, so a routine fills a feed or makes an episode every day or week. Four nodes: Select assets passes on the resources in the routine's namespaces (or the namespaces named) that came in since the routine last ran, or in the past day, week or month, or all of them, matching words, topics, entities and status, newest first and at most a limit (podcast episodes left out unless asked, so an episode is never a source for the next one); Select activities does the same for the timeline's activities, by kind and a where match, and only what everyone who reads the namespace sees; Publish to feed adds what comes in (resources, episodes, activities, or a template node's text) to a feed, skipping what is there already and refusing what the feed's readers couldn't read; Podcast episode makes an episode about the resources that come in, placed where its readers may read every source, and lists it in the feed once it's made. Try on the canvas says what each node would do without publishing. A run's result counts what was published and the episodes made, and its log says what each node did.

The catalog marks the two new nodes as ones that keep their output, so a graph workflow with one of them is whole without Apply changes. This is the second of three pull requests for the ask; the Feeds page and the canvas nodes in the web app follow.

How: domain/feed_nodes.py holds the nodes' kits (config checks, the queries, the run), workflows.py registers them in the graph scope and in TERMINAL, and podcasts.create takes a space argument and podcasts.fits answers where an episode may go, so the node reuses the podcast feature as it is.

Motivation and Context

Adeel's ask in the project: RSS and iTunes podcast endpoints from routines, so a routine can pick part of the graph or all of it and publish a podcast, or timeline events, through a feed. #534 (merged) made the feeds; this makes routines fill them.

Screenshots (if appropriate):

n/a (workflow nodes; the canvas comes in the third PR)

Steps to reproduce (if appropriate):

Make a graph workflow select_assets → podcast_episode (feed: a podcast feed's id, length 10) with POST /api/v1/workflows, try it with POST /api/v1/workflows/test, put it on a routine that runs every Monday (POST /api/v1/routines), and follow the feed in a podcast app. Or select_activities (kinds asset.added) → publish_feed for the timeline in a feed reader.

Types of changes

  • Bug fix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to change; ! in the title)

Checklist:

  • The title is a Conventional Commit.
  • My code follows the code style of this project (make hooks checks it on commit).
  • I added or updated tests (tests/api/test_feed_nodes.py, 4 tests: a routine publishes new resources to a feed; publishes the timeline and text; makes an episode and the feed lists it once made; Try says what it would do and bad settings are refused).
  • My change requires documentation updates, and I have updated docs/ accordingly (docs/processing.md "Publishing from the graph", docs/feeds.md "From a routine", docs/podcasts.md, docs/requirements.md, README).
  • Users need more than the title to understand it, so I wrote it under ## Unreleased in CHANGELOG.md.

Done questions (AGENTS.md)

  • Goal: a routine now picks part of the graph and publishes a podcast or timeline events through a feed, which is the ask; the web app's page and canvas nodes are the third PR. Merged on green CI (local CI green: backend, frontend, pre-commit, title).
  • Agents: the nodes are ordinary workflow nodes, so an agent builds the workflow with POST /api/v1/workflows, tries it with POST /api/v1/workflows/test and schedules it with POST /api/v1/routines; the catalog (GET /api/v1/workflows) lists the nodes with their settings. Routines have no MCP tool yet (the feeds themselves do): filed as MCP tools for routines and graph workflows #548.
  • Zero setup: no new prerequisite, .env or step.
  • Fewest clicks: every setting has a default (since the last run, 50 at most, 10 minutes, deep dive, 10 sources); only Publish to feed needs a pick, the feed.
  • Configurable: everything is a node setting with a default, editable where workflows are; no new server setting, nothing in the wizard.
  • Light: queries over existing tables; an episode costs what Podcasts costs and only when a routine asks for one.
  • Private and secure: Publish to feed refuses what the feed's readers couldn't read and never a vault's; Select activities passes on only what every reader of the namespace sees; the episode goes where its readers may read every source (podcasts.fits). Threat model: P-FEEDS' voided-by names routines.
  • Cost: the select and publish nodes make no model calls; an episode's calls land in the activity ledger against it, and a namespace's budget holds the job, as Podcasts already does.
  • User decides: nothing publishes until someone puts these nodes on a routine; existing workflows and defaults are unchanged.
  • Upgrades: no change to stored data (node configs are JSON in the workflow row).
  • Honest: docs say what the nodes do and don't; Try says what a run would do; the run's log says what was refused and why.
  • Tests, docs updated; make openapi unchanged (no API change).
  • Open standards: builds on feat(feeds): private RSS and podcast feeds with API and MCP #534's RSS 2.0 and iTunes namespaces; no new dependency, MIT kept.
  • Result reported in the thread with the PR; "refine later" (picking passages, voices) stays in docs/podcasts.md.
  • Reuse: flow.Kit, events.check_filter/events.visible, library.STATUSES, syndication.publish, podcasts.create/fits, the routine run's started_at for "since the last run".

🤖 Generated with Claude Code

https://claude.ai/code/session_01CDEiKGoyQX8Vmnv9bi5JpD

Select assets and Select activities pick what came in since the routine last
ran (or in a period) in its namespaces; Publish to feed adds it to a feed and
Podcast episode makes an episode about it, added to a feed once made. A feed
must be in a namespace the run covers, items stay bounded by the feed's
readers, and an episode drops sources its namespace couldn't read.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CDEiKGoyQX8Vmnv9bi5JpD
@adeelahmad adeelahmad self-assigned this Oct 18, 2026
@adeelahmad
adeelahmad merged commit 2b62746 into main Oct 18, 2026
12 of 13 checks passed
@adeelahmad
adeelahmad deleted the claude/project-thread-rovv80 branch October 18, 2026 06:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants