Plannotator integration: Approve in browser never sends callback back to agterm

[basilboli]

Version: 0.12.1 (d29b64a)
The Plannotator plan-review integration doesn't seem to complete the approval round-trip. When the agent proposes a plan, the plan correctly opens in the browser via Plannotator. However, after clicking Approve in the Plannotator UI, nothing happens on the agterm side — no callback ever arrives, and the session just keeps waiting on the plan-approval step indefinitely.

Steps to reproduce:

Run agterm 0.12.1 with the Plannotator integration enabled
Have the agent propose a plan (plan opens in the browser as expected)
Click Approve in Plannotator

Expected: approval/feedback is sent back to agterm and the agent proceeds with the plan.
Actual: no callback is received; agterm stays blocked waiting for the approval.
Happy to provide logs or debug output if that helps.

Plannotator link: https://plannotator.ai/

P.S. And thanks for the great work! I love it

[umputun]

short answer: agterm isn't in that path, and there's nothing agterm-specific that would change how Plannotator behaves.

how the round-trip actually works (from Plannotator's README + apps/hook/server/index.ts): Plannotator runs as a hook inside your agent (Claude Code / Codex), starts a local HTTP server on a random localhost port, opens the browser to that localhost URL, and when you approve, the browser posts back to that local server, which prints the feedback to the hook's stdout for the agent to read. None of it goes through agterm; it's loopback between your browser and the hook process.

why it wouldn't behave differently in agterm: agterm is a plain, non-sandboxed terminal. It doesn't sandbox child processes, so nothing blocks the hook's localhost server. It doesn't intercept browser opens. And it injects only its own AGTERM_* vars plus TERM_PROGRAM=agterm, no SSH_*, so it can't trip Plannotator's SSH/remote-mode switch. A Plannotator hook runs the same here as in Terminal or iTerm.

this exact failure (approve, no feedback back) shows up in Plannotator's own tracker without any terminal angle, e.g. backnotprop/plannotator#678 (feedback lost when the localhost session auto-closes on a CLI timeout) and #154 (browser not opening on a non-TTY launch). That's the side to chase.

quickest way to confirm it's not us: run the same plan in Terminal or iTerm. If it fails there too (I'd expect it to), it's Plannotator/agent-side. If it somehow fails ONLY in agterm, send the exact repro (which agent, how the hook is launched, the port it picked) and I'll dig in, but nothing agterm does should touch that callback.

glad it's useful, and thanks for the kind words.