Bellows Ships Coding-Agent Session Handoff in Six Days

A feature request filed on Anthropic's tracker on October 2 shipped on October 8 — in someone else's product.

The issue, anthropics/claude-code#98937, asked for a one-command way to hand a coding-agent session to a teammate. Six days later Bellows, a third-party desktop app, shipped it in version 1.19.0. The details below are the vendor's own description of its release; none of it has been independently tested.

What "hand over" claims to do

Bellows describes itself as a free desktop app for Windows, macOS and Linux that runs coding agents — Claude, Codex, Gemini CLI, Qwen Code, any ACP agent, and local models. Hand over, per the vendor, writes a single .bellows-session file containing four things:

• the whole conversation, with every tool call and its result;
• a handover note the agent writes as you hand over — where things stand, what's next, what to watch out for;
• keys and passwords redacted before the file is written;
• a digest, so a file changed after it was made says so when it's opened.

The teammate opens the file, picks their own checkout of the repo and their own agent, and continues. The vendor's claim worth noting: it does not have to be the same agent. "You worked in Claude; they can carry on in Codex."

Why this matters past one product

The underlying problem is not vendor-specific. An hour into a debugging session the agent holds context that exists nowhere else — which files were read, which hypotheses died, which tool calls failed and why. Forwarding a chat log does not transfer that to someone else's agent, and neither does a handwritten summary.

That the same request is open on more than one tracker is the signal here. Bellows points to anthropics/claude-code#60082 and openai/codex#46016 as public asks for the related live-sharing feature. Session portability is turning into a category expectation rather than one vendor's differentiator — and the file format that wins it becomes the interoperability layer between agent vendors.

The parts to be skeptical about

Two of the four claimed file contents are security claims, and both are the kind that need verification before you email one of these files to a colleague. "Keys and passwords redacted before the file is written" depends entirely on the redaction patterns; a session transcript that includes tool output is a broad surface, and anything the redactor does not recognize ships with the file. The digest detects post-hoc modification but says nothing about whether redaction was complete.

The rest of 1.19.0, again as described by the vendor: a Review pane showing each edit as a highlighted diff at the moment the agent asks for it; phone pairing so permission prompts from any session reach your phone on any network to answer Allow once or Deny; and mid-task agent switching, which shipped earlier in 1.17.0. Live session sharing gives a colleague a browser link with one of three roles — watch, suggest or type — with suggestions queued until accepted.

What to watch

Bellows attributes the six-day turnaround to release cadence, saying it shipped eight releases between October 5 and 8. Treat that as a statement about a small team's shipping speed, not about the difficulty of the problem.

The question for anyone evaluating this is whether the first-party agent vendors ship their own handoff format. If they do, a third-party file format is a stopgap; if they do not, the apps that wrap multiple agents inherit the portability layer by default. The open issues on both the Anthropic and OpenAI trackers are the place that gets answered.

Source

• "Asked for on 2 October, shipped on 8 October: handing a coding-agent session to a teammate" — Bellows AI, Dev.to (vendor post)