vercel/ai @ai-sdk/harness-pi@1.0.170

on GitHub

Patch Changes

• c85285b: refactor(harness-pi): resolve sandbox paths on first use so session start makes no sandbox calls

  Starting a Pi session no longer resolves the sandbox HOME or the fileToolPathPolicy denied roots. The harness resolves them the first time a native file tool runs or a turn has skills to write, and retries the resolution on the next use if it fails. Skills that an earlier process wrote and the current turn no longer passes are removed the first time a native file tool runs. A turn without skills or native file tool calls makes no sandbox calls.

  A denied root that cannot be resolved now fails the native file tool call that needs it instead of failing session start.

• c85285b: refactor(harness-pi): drop the workspace mirror and list sandbox skills from settings

  The harness no longer copies the sandbox workspace to the host or patches Node's fs module to serve that copy at the sandbox path. Pi used the copy only to discover AGENTS.md, CLAUDE.md and project skills, so Pi no longer reads those files from the sandbox workspace. Native file tools resolve relative paths against the sandbox work directory directly.

  Put project context in instructions instead. For skills that already exist in the sandbox, the new resources.skills setting lists them to the model; each must exist at filePath, and the model reads it with the read tool, which can read the skill's directory. Pi reads no project settings, context files or skills from the host.

  Inline extensions no longer see a host copy of the workspace, including when the sandbox shares the host filesystem. Read project files through session.sandboxSession.

• c85285b: feat(harness-pi): keep the Pi session in lifecycle data

  The data of both lifecycle states (continue-turn and resume-session) is now { entries, pendingToolBatch? }: the Pi session header followed by its entries, plus the tool batch that was waiting on host input. Pi keeps the session in memory and restores it from those entries, migrating entries recorded under an older Pi session version. Nothing is written to or read from the sandbox for session state, so stopping, suspending and resuming a session make no sandbox calls. sessionFileName is gone from lifecycle data.

  A rebuild of the Pi session for a changed tool set now keeps the conversation. It used to start the rebuilt session with an empty history.

  Breaking: a state recorded before this version ({ sessionFileName }) still passes lifecycleStateSchema but resumes as a fresh session, so conversations suspended across the upgrade lose their history. A pending tool batch in such a state is dropped, because there is no journal to replay it against.

  Deferred MCP tools that tool_search loaded in an earlier turn are not re-activated when the session is rebuilt from entries, as before.

• c85285b: feat(harness-pi): keep the Pi session in an optional caller-provided append-only sessionStorage, so lifecycle state holds only the id of the entry the session ends at

  The harness appends each new Pi session entry with its entryId, parentEntryId and Pi session version, through the optional appendMany when the storage has it. A storage can ignore a retried append, and read can return only the entries that lead to leafId.

  • @ai-sdk/harness@1.0.167