vibe: one ongoing Claude Code session takes the prompts in turn

[rossry]

What

SessionCoder now keeps one Claude Code session open (the Agent SDK's ClaudeSDKClient) and feeds it the crew's prompts one after another, instead of spawning a session per prompt.

• It remembers the night. What it shipped and what people asked stay in context, so "like foundation/light-layout/_: add comprehensive planning documentation #3 but slower" works. A generation the session shipped itself is referred to by number in the brief (# you shipped this one earlier … var/vibe/vibe-0003.py) rather than re-sent; a repo pattern's source still travels. The system-prompt addendum tells it every generation is on disk to Read when the crew names a number it doesn't remember (e.g. after a restart).
• Faster turns. No CLI startup and no re-reading the craft notes per prompt. Live in the sandbox: prompt initial readme #1 took 30 s, prompt foundation/_ #2 in the same session 17 s — and its note shows it built on initial readme #1 ("Tide's breath now takes about two and a half minutes, and a sparse handful of cool-white stars…").
• Lifecycle. The session lives on its own event-loop thread; the vibe worker blocks on each prompt (serial by construction). A prompt past its wall-clock cutoff (240 s) or a broken session (CLI died, stream error) is dropped and the next prompt lazily starts a fresh one, with the CLI's stderr in the error text. A session idle for 20 minutes is closed. The model switches in place (set_model) when a submitter picks a different one. VibeCore.stop() closes the session, so a server shutdown takes the CLI process with it (verified: clean exit, no orphan).
• Repair is a follow-up in the same conversation ("The module you just shipped failed on the server: …"), not a fresh brief.
• The ship tool is bound once per session and dispatches to the prompt in progress.
• The thread now shows how long each generation took (seconds on each entry, · 17 s beside the model).

Verified

• Tests: the session backend is driven by a fake ClaudeSDKClient-shaped session via the injectable session_factory — one session across prompts (model switch, self-shipped pointer vs. repo source), broken session dropped and restarted with stderr in the error, runaway prompt cut off, idle close and lazy reopen, spoken-code fallback / error result / repair follow-up, end-to-end on the stage with core.stop() closing the session. Full suite green; black + mypy on CI's lists (with and without the SDK installed locally).
• Live: serve --stage-key k with the real CLI — two prompts in a row served by the same claude process (same PID), both validated and cut onto the stage, then a clean shutdown.

Docs

README "Vibe mode" backend paragraph, docs/deploy.md backend bullet, plan/spec/implementation-notes.md row.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N97HVk3sAr5jRGYodXcDoC

---

Generated by Claude Code