vercel/ai @ai-sdk/harness-pi@1.0.154

on GitHub

Patch Changes

• 90983d7: feat(harness-pi): add configurable writable sandbox roots for Pi's native file tools

• d10fa3e: feat(harness-pi): pass the harness session to extension factories

  Each extensionFactories entry now receives a second argument, a PiHarnessExtensionSession with the restricted sandboxSession, the sessionWorkDir and an instructions() getter for the current turn's instructions. An extension can now run work against the session's sandbox with the session's instructions, such as a tool that delegates to a child Pi session. Existing (pi) => void factories keep working unchanged.

• 9323fdb: feat(harness-pi): serve mcpServers with Pi's native MCP and tool_search

  mcpServers is now served by Pi's built-in MCP extension and its tool_search tool instead of pi-mcp-adapter, which is no longer a dependency. Each entry is Pi's McpServerConfig: stdio { command, args, env, cwd } or streamable HTTP { url, headers, timeout }, plus exposure, toolExposure, enabled and description. The setting stays typed as Record<string, unknown> like the other harnesses; type entries with the exported PiMcpServerConfig. Fields that only the adapter understood, such as lifecycle, are no longer supported.

  Tool names are always mcp__<server>__<tool>, with - in either name replaced by _, so a server named brand-ai now yields mcp__brand_ai__<tool>. exposure defaults to direct, which declares a server's tools to the model. deferred tools stay undeclared until the model loads them through tool_search and then calls them by name. Pi's own default, codemode, is not supported because the harness does not load Pi's codemode extension; an entry that asks for it fails session start, as does an entry Pi rejects.

  mcpSettings.toolPrefix and mcpSettings.outputGuard are deprecated. A toolPrefix other than 'mcp' or outputGuard: false now fails session start, since native MCP cannot honor either. Pi truncates text results over 20 KB and writes the full text to a file in the host temp directory.

  The harness has no sign-in flow, so an HTTP server that requires authentication must carry an Authorization header or auth.provider; with the header, OAuth never runs. The harness reads no mcp.json and neither reads nor writes mcp.log or mcp-auth.json, so a sign-in stored by the Pi CLI is never sent. @earendil-works/pi-ai is no longer a direct dependency.

• 286e0c7: fix(harness-pi): return images from the read tool as image blocks

  The read tool decoded every file as UTF-8, so the model received an image as unreadable text. The tool now runs Pi's own createReadToolDefinition over the sandbox: PNG, JPEG, GIF, WebP and BMP files come back as image blocks that Pi resizes, and text reads keep Pi's truncation and paging notices. The tool still takes file_path, and path mapping and fileToolPathPolicy apply as before.

• 8ae0804: fix(harness-pi): judge a turn by its last attempt and keep suspend errors out of the stream

  A turn that Pi recovered by retrying a failed request (rate limit, overload, context overflow after compaction) ended with an error part, because the first terminal error seen was kept for the whole turn. The last assistant message now decides the outcome, so a recovered turn ends with finish.

  Once doSuspendTurn has begun, every terminal or thrown error of the turn being cut belongs to the suspend and no longer emits an error part. Before, only errors that looked like an abort were kept silent, and a request Pi started on an already-aborted signal reports stopReason: 'error' ("This operation was aborted") while a provider can also fail during the cut. Reporting such an error ended the slice as failed with the unfinished turn nested in its resume state, and a session built from that state refused every later prompt.

• Updated dependencies [2e89fa4]

  • @ai-sdk/provider-utils@5.0.60
  • @ai-sdk/harness@1.0.152