* feat: add runtime vm cache sync
* feat: configure runtime vm cache sync
* feat: prefetch vm cache on startup
* feat: await vm cache before vm startup
* fix: recheck vm cache after prefetch wait
* fix: address vm cache review feedback
* build(server): require VM cache manifest env
* refactor: remove redundant context-overflow middleware
The middleware caught provider overflow errors and re-tried with a
naive prompt truncation, but its `nonSystem.slice()` had no awareness
of tool_use/tool_result pairing — a cut between an assistant tool_use
and the matching tool_result produces an orphaned tool_use that
providers reject with a different error.
Compaction (`createCompactionPrepareStep`) already handles this safely:
`findSafeSplitPoint` walks past tool messages to preserve pair
integrity, and the pipeline (strip binary → prune → reduce outputs →
LLM summarize → sliding window) handles every overflow path before
the request leaves the agent.
Drops 426 lines: the middleware itself, its wiring in ai-sdk-agent,
and the matching test block + helpers in compaction.test.ts.
* docs: document BROWSEROS_AI_SDK_DEVTOOLS in .env.example
Surfaces the opt-in dev flag so contributors know it exists. Captures
every LLM call to .devtools/generations.json for post-hoc inspection.
