vercel/ai @ai-sdk/harness-pi@1.0.145 on GitHub

c371c00: feat(harness-pi): add a cacheRetention setting that is passed to every Pi model request, so Anthropic models can use the 1-hour prompt cache
5901911: fix(harness-pi): stream extension-registered tools as provider-executed dynamic tools, and close a compaction that arrives after the turn's last step in its own step
d7dedf2: fix(harness-pi): shut down MCP servers on dispose without reloading the Pi session
10bc612: fix(harness-pi): report real step and turn usage
finish-step usage was always zero, and finish.totalUsage was the running total of the whole Pi session, not the usage of the current turn. Each step now reports the usage of its assistant message. The turn reports the change in Pi's session stats across the prompt. inputTokens.total now includes cached input. noCache is set, and outputTokens.text and reasoning are set when the provider reports reasoning tokens.
cadd7c1: fix(harness-pi): fail the turn when the requested model is not in Pi's model catalog instead of silently running on the previous model
21148df: feat(harness-pi): add mcpSettings to configure the MCP adapter's tool prefix and output guard
d45e044: feat(harness-pi): add fileToolPathPolicy setting with readable and denied roots for native file tools
Updated dependencies [c4a5c8d]
Updated dependencies [50ac48b]
Updated dependencies [e8af189]
- @ai-sdk/harness@1.0.143