dsh-migrate-on-429

When you hand a large codebase or an extremely long document to an Agent, the context window balloons with every turn. Each request carries megabytes of tokens, and TPM rate-limiting (HTTP 429) kicks in almost immediately. The instinctive response is to keep sending continue — but that just resubmits the same massive context over and over, burning through your quota with zero progress.

dsh-migrate-on-429 changes the strategy. Below the failure threshold (default: 3 consecutive 429s) it still auto-continues as before. Once the threshold is hit, the plugin cancels the stale session, waits until it is truly idle, extracts a structured handoff summary (original task, user instructions, recent progress, touched files, working directory, model, and presets), and then spawns a brand-new session with that distilled context so the task resumes where it left off. The entire handoff is strictly serial — cancel, whenIdle, create, followup — so the old and new sessions never run in parallel and never double up token consumption. For multi-sub-agent concurrency it adds cross-session global coordination: burst detection, a migration mutex, and a post-migration cooldown window, preventing N sessions from all triggering simultaneous handoffs. The handoff payload itself is deduplicated and pruned to stop "handoff stacking" from re-triggering a 429 on the very first message of the new session.

Best for developers whose Agents routinely work through large repositories, lengthy documents, or multi-step workflows and keep hitting TPM limits. It supersedes the older dsh-auto-continue-429 pure-retry approach, turning stuck, keep retrying into stuck, start fresh from a clean slate and pick up where you left off.