Description
Every session in apps/dev-workstation/bin/dev-session is one claude -p call at --effort low with a $0.30 budget, so every conversation has exactly one user message and a similar length. Real sessions include follow-ups, corrections (“no, the other file”), a change of mind, abandoned work and the occasional long task. Gateway spend caps must keep bounding cost.
Acceptance Criteria
- #1 A share of sessions continue for one to three follow-up user turns on the same Claude Code session, with follow-ups written per scenario
- #2 Some sessions end early as abandoned, and effort and budget vary by scenario within the existing gateway caps
- #3 Agent Observability shows multi-turn conversations for the developers, and the dev_session result line records the turn count
Definition of Done
- #1 just check
Implementation Plan
- Prompt files carry ‘— followup’ turns; dev-session resumes the same session for 1-3 follow-ups on a share of sessions. 2. Abandoned sessions and per-scenario effort/budget within gateway caps. 3. dev_session result line records user turn count.
Implementation Notes
Harness-verified (fake claude): follow-ups resume the same session id, budget shrinks per turn, abandonment and short-timeout paths, user_turns in the result line. Live-verified 2026-09-29: claude -p –resume reports total_cost_usd cumulatively and –max-budget-usd per run, so dev-session now takes the latest total instead of summing. AC3 not yet observed: the first five live sessions were single-turn by chance (45% follow-up share); Rob declined a wait, so it stays open until a multi-turn dev conversation shows in Agent Observability.