@assemblyline-agents/pi
v10.4.0
Published
Canonical Pi agent loop for Assembly Line behind the internal AgentHarness runtime contract.
Downloads
12,249
Maintainers
Readme
@assemblyline-agents/pi
The canonical Pi agent loop for Assembly Line, packaged behind the framework's internal AgentHarness runtime contract.
It adapts the Pi agent engine (@earendil-works/pi-agent-core /
@earendil-works/pi-ai) to Assembly Line's harness interface: tool
descriptors, continuations, retries, transcript management, media inputs,
message phases, and normalized events. It starts with no model providers and
loads only the packages selected and frozen in the agent manifest.
The runtime owns tool-output limits and durable full-result retrieval. Pi
preserves its model-facing result envelopes, including read_tool_result
references, without applying a second 8,000-character cap. The obsolete
boundModelOutput export has been removed from both Pi and core.
Before every provider request, the harness retains only the two newest image-bearing tool results (subject to an 8 MiB history budget) and replaces older images with a small marker. Stale image references are also removed from tool-detail metadata. User-provided images remain intact. Raw tool outputs remain in durable runtime records, while continuations store the same bounded model-facing transcript. This keeps long computer-use runs from resending every prior screenshot and growing input cost and request size on each action.
Model plugins own provider registration, discovery, authentication, request
payload semantics, embeddings, response accounting, reconciliation, billing
mode, and deployment requirements. Pi contains no providers/all catalog and
no OpenRouter, OpenAI, or Codex branches. The Assembly Line runtime uses this
provider-neutral harness to drive agent turns; it is included automatically
with the umbrella @assemblyline-agents/sdk package.
Install
npm install @assemblyline-agents/piPart of Assembly Line
This package is part of Assembly Line, a vendor-neutral, filesystem-first framework for durable AI agents. See the developer documentation in docs/developers/ for setup, agent authoring, and deployment guides.
License
MIT
