Oh My Subagents Is a Real 130-Star Python Project, and Its Own Docs Cap the Interruption Recovery at "Local Recovery, Not Distributed Failover or Blind Replay"

Every product claim in this video traces back to a line in the repo's README, and the one claim that does not, the promise that work resumes after the internet blips, is narrowed by the project's own recovery documentation. Github Signals posted the 40-second clip on September 22, 2026. There is no host and no voice on camera. All 20 sampled frames show the same thing: a browser window sitting on github.com/ringlochid/oh-my-subagents, scrolled to the README, with the README's embedded demo GIF autoplaying inside it. The page creeps up a few pixels per second across the whole runtime. A synthetic voiceover reads copy adapted from the README while karaoke-style captions burn in over the GIF, one phrase at a time, in white with the stressed word in yellow: OH MY SUB-AGENTS, FROM BABYSITTING, TEMPORARY TOOLS IT, DURABLE ACCOUNTABLE, PERSISTENT STATE, INTERRUPTION RECOVERY, OR THE INTERNET, START OVER FROM, ONE-OFF PROMPTS, FINISHES THE JOB. A translucent @GithubSignals watermark drifts around the lower third. Frame 20 is a dark end card with a cartoon-cat logo and the line "Follow for more open-source projects."

The repo is real and the description is copied nearly word for word

The repository exists at github.com/ringlochid/oh-my-subagents. GitHub's API returned 130 stars, 10 forks, 1 watcher, 0 open issues, 144 commits, and 1 contributor when I queried it on September 23, 2026. It was created on July 20, 2026 and last pushed on September 14, 2026. The primary language is Python and it requires Python 3.12 or newer.

The video's opening line, that this is "the local orchestrator that finally stops you from babysitting AI agents," lifts directly from the README's first section heading, "Stop babysitting subagents." The phrase "durable, accountable team with persistent state" is a recombination of the README's tagline, "Turn ad-hoc subagents into durable, accountable AI teams," and its one-line description, "Local subagent orchestration for Codex and Claude, with persistent task state, reusable teams, and interruption recovery." The narration is closer to paraphrased marketing copy than to an independent assessment.

"Local" checks out as a technical claim rather than a slogan. Installation is pipx install oh-my-subagents, the console serves on http://127.0.0.1:18125/, SQLite is the default store so no database server is needed, and oms service install registers a per-user background service: a systemd user service on Linux, a LaunchAgent on macOS, a Scheduled Task on Windows 11. PostgreSQL is available as an optional extra. Nothing here routes through a vendor's cloud.

The interruption recovery claim is the one the docs walk back

This is the video's headline feature. The narration says that if "a session crashes or the internet blips it doesn't start over from scratch, it remembers exactly where the work stopped and picks right back up."

The project's own operations guide, docs/guides/run-and-operate.md, has a section titled "Recover after interruption." It reads: "Controller records survive a provider process interruption or Oh My Subagents restart. On startup, Oh My Subagents audits durable state and converges work that can legally continue. Refresh the run to read current truth before taking another action."

The next paragraph is the qualifier: "This is local recovery, not distributed failover or blind replay. A provider's terminal success also does not prove assignment or run success; the controller must accept the corresponding state transition."

Three gaps open up between that and the video. The documented failure cases are a provider process interruption and a controller restart. Network loss, the "internet blips" scenario the voiceover names, is not one of the cases the docs address. The recovery is scoped to work that "can legally continue," which is narrower than resuming from the exact point of interruption. And the guide instructs the operator to refresh the run and read current state before acting, which is a human step the video's "picks right back up" removes.

The README's own phrasing on the same point is more careful than the video's. Its comparison table says the product will "Recover from committed controller records without discarding accepted history," and its accountability section promises "Honest recovery. Browser closure, provider interruption, and controller restart do not silently fabricate completion or discard accepted history." Not fabricating completion is a different guarantee from resuming mid-task. The README makes the weaker, defensible claim. The video makes the stronger one.

The license is not simply MIT, and the video's own footage shows the tab that says so

Several frames show the README tab bar reading README | Contributing | MIT license. A viewer would reasonably conclude the whole thing is MIT.

The README's license section splits it. The core is MIT, "except for the visual Console in console/, which contains material derived from n8n and is distributed under the Sustainable Use License." I fetched console/LICENSE and console/NOTICE directly. The NOTICE file states: "The Oh My Subagents Console contains source code, stylesheets, and design tokens derived from n8n, modified by the Oh My Subagents authors." The Sustainable Use License text carried over from n8n also excludes files with .ee. in the filename or .ee in the dirname, which it reserves for n8n Enterprise License holders.

This matters because the README calls the visual Console "the primary experience." The part of the product the demo GIF is actually showing is the part not under MIT. The TikTok's description says "Open source: ringlochid/oh-my-subagents" with no qualifier, and the Sustainable Use License permits internal business, non-commercial, and personal use while limiting redistribution to free non-commercial distribution. That is source-available with restrictions rather than an OSI-style open source grant.

The project was called Banksia five weeks ago, and you can read it in the frames

The console sidebar inside the embedded demo GIF reads Banksia in the top-left corner of nearly every frame from 1 through 19. The audio never mentions it and the README text visible on screen never mentions it above the fold.

PyPI confirms the history. A package named banksia shipped eight releases from version 0.1.0 on July 28, 2026 through 0.1.7 on August 19, 2026. The package oh-my-subagents starts at 0.2.0 on August 21, 2026 and reaches 0.3.2 on August 30, 2026, which is the version the README's PyPI badge displays in every frame. GitHub's release list matches, with the v0.1.7 release still titled "Banksia 0.1.7." The README carries a full migration section, including oms migrate-from-banksia and a warning not to run oms init first.

So the video is presenting a project that has existed under its current name for roughly one month, whose latest release at the time of posting was 23 days old, and whose demo footage still shows the previous brand. None of that makes it bad software. It does undercut the settled, ready-for-your-workflow tone of the narration.

"Spotted on Hacker News" means the author posted it himself, repeatedly

The TikTok description ends with "Spotted on Hacker News." I searched the Hacker News Algolia API for the repo. It returned 13 hits. Every one of the eight submissions I retrieved was posted by the account ringlochid, which is the GitHub owner's handle.

The scores are small. "Show HN: Oh My Subagents" on August 21 drew 2 points and 1 comment. "Show HN: I let an subagent workflow refactor my codebase for three days" on August 22 drew 7 points and 2 comments, the best of the set. "Show HN: Visual workspace to design and operate daily multi-agent workflows" on August 29 drew 4 points and 1 comment. "Show HN: Open-sourced runtime for better Codex and Claude subagent experience" on September 1 drew 3 points and 1 comment. "Show HN: Durable subagents for Codex and Claude users" on September 11 drew 3 points and 0 comments. "Turn ad-hoc subagents into durable, accountable AI teams" on September 22, the same day this TikTok went up, drew 4 points and 0 comments.

A repeated self-submission that tops out at 7 points is not the same signal as a project surfacing organically on the front page. The account name @github.signals and the "Spotted on Hacker News" line both imply discovery. What happened is closer to the author marketing his own work across two platforms on the same day.

The baseline the video attacks has moved

The video's contrast is with "manually spawning and polling temporary tools" that leave you with "a chat log that dies when the tab closes." Claude Code's own subagents documentation, which I loaded, shows that baseline is now more capable than the video implies.

Claude Code subagents support a memory frontmatter field with three scopes, user, project, and local, writing to directories such as .claude/agent-memory/<name>/, with the first 200 lines of MEMORY.md preloaded on each invocation. Completed subagents return an agent ID that allows resumption via the SendMessage tool, and a resumed subagent "retain[s] their full conversation history including all previous tool calls and results." Subagents can spawn their own subagents up to a default depth of three layers below the main conversation, adjustable through CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH, with a concurrent cap of 20.

Persistent memory, resumption, and recursive delegation are already in the tool the video positions against. The distinction that survives is where the state lives. Claude Code's subagent state is tied to the CLI session and the agent-memory files under the project. Oh My Subagents puts assignments, waits, waves, and checkpoints in a controller database that a background service owns, outside any provider transcript, which is what lets it survive the controller process dying. That is a genuine architectural difference. It is a smaller one than "the difference between playing with AI and actually managing it."

What the model underneath actually is

The README defines a vocabulary the video never touches, and it is the substance of the project. A Workflow is "a reusable team of responsibilities. It answers who is responsible," and explicitly does not prescribe a DAG or a schedule. A Task pins one published Workflow revision at start. A Manager is "any current Member with direct children." A Wave "always belongs to one immediate Manager and may contain one or more current direct children." Children return Checkpoints, and only "the Task lead's accepted completed or blocked Checkpoint becomes the Result shown to you."

oms init publishes eight provider-neutral starter workflows, named in the README as production-feature-delivery, incident-investigation-and-recovery, migration-and-modernisation, deep-research-and-decision-brief, decision-through-competing-prototypes, idea-to-validated-demo, experiment-and-replication-program, and security-audit-and-hardening. Frames 2 through 5 show those exact strings in the Console's workflow list inside the demo GIF, which is the one place the video accidentally conveys real product detail.

The README also includes a caution the video inverts: "A one-Member Workflow is valid, and usually wiser, when delegation would add ceremony without independent evidence, useful specialization, or real integration." The author is telling you not to reach for a multi-agent team by default. The video sells the team as the upgrade.

Key Takeaways

The repo is real and matches the video's description. ringlochid/oh-my-subagents had 130 stars, 10 forks, 1 watcher, 144 commits, and 1 contributor on September 23, 2026, is written in Python, requires 3.12 or newer, installs via pipx, serves a local console on port 18125, and defaults to SQLite.
The interruption recovery claim is the video's weakest point. The project's own guide says "This is local recovery, not distributed failover or blind replay," scopes recovery to work that "can legally continue," names provider process interruption and controller restart rather than network loss, and tells the operator to refresh the run before acting.
The license is split. Core code is MIT, but the visual Console, which the README calls "the primary experience" and which is what the demo GIF shows, is under the n8n-derived Sustainable Use License that limits redistribution to free non-commercial distribution.
The project was published on PyPI as banksia through eight releases ending August 19, 2026, and the name Banksia is still visible in the console sidebar of the demo GIF in most frames of this video.
"Spotted on Hacker News" describes eight submissions of the same repo by the repo owner's own account, the highest scoring 7 points with 2 comments.
Claude Code subagents already offer persistent memory scopes, resumption with full conversation history via agent ID, and three-layer recursive delegation, so the "temporary tools" baseline the video attacks is narrower than presented. The real difference is controller-owned state outside the provider session.
Whisper transcribed the audio accurately across the full 40.7 seconds with one substantive error. It rendered "polling" as "pulling" in "instead of manually spawning and pulling temporary tools," which destroys the reference to the README's line "A parent spawns children, polls them." The transcript captures none of the visuals, so the URL, the MIT license tab, the Banksia sidebar label, the PyPI v0.3.2 badge, the eight starter workflow names, and the fact that the entire video is a slow scroll of one GitHub page are invisible to anyone reading the transcript alone.
Unverified: the maintainer's real-world identity. Both the banksia and oh-my-subagents PyPI packages list an author email of Leo Zhang <yzha0012@uni.sydney.edu.au>, but I did not load any page where the GitHub account ringlochid states that name, so the link rests on package metadata alone. Also unverified: whether a network interruption specifically recovers, since the docs name only provider process interruption and controller restart; whether anyone other than the author runs this in production, given 1 watcher and 1 contributor; the YouTube tutorial at -prDEZYpx9M linked from the README, which I did not load; and the star count at the moment the video was recorded on September 22, since my 130-star reading is from September 23.

Resources

This tool stops your AI agents from forgetting everything when a sess... - the source video, 2,793 views, 146 likes, 0 comments, and 33 reposts at capture, 40 seconds at 720x1280
ringlochid/oh-my-subagents on GitHub - establishes the repo exists, 130 stars, 10 forks, 1 watcher, 0 open issues, 144 commits, 1 contributor, Python, created July 20 2026, last pushed September 14 2026
oh-my-subagents README (raw) - source of the tagline, the ad-hoc comparison table, the install commands, the eight starter workflow names, the Banksia migration instructions, and the split MIT / Sustainable Use License notice
Run and operate work (docs/guides/run-and-operate.md) - the "Recover after interruption" section containing "local recovery, not distributed failover or blind replay," which narrows the video's central claim
Workflows and teams (docs/concepts/workflows-and-teams.md) - definitions of Workflow, Member, Task, Assignment, Wave, Checkpoint, Manager, and Task lead
Console NOTICE - confirms the Console is derived from n8n and modified, and is not an n8n product
Console LICENSE - the n8n-derived Sustainable Use License text, including the .ee. file carve-out reserved for n8n Enterprise License holders
oh-my-subagents on PyPI - version 0.3.2 uploaded August 30 2026, first release 0.2.0 on August 21 2026, requires Python 3.12 or newer, author email Leo Zhang
banksia on PyPI - the prior name, eight releases from 0.1.0 on July 28 2026 to 0.1.7 on August 19 2026, same author email
Show HN: Oh My Subagents - the August 21 2026 submission by ringlochid, 2 points and 1 comment, the first of eight self-submissions returned by the Hacker News Algolia search
Claude Code subagents documentation - establishes that Claude Code subagents support memory scopes, resumption with full history via agent ID, and three-layer recursive spawning, which narrows the video's contrast