Pi Subagents

The docs say children start only when Pi invokes them, with no automatic background reviewer; tool activation requires authorization, and configuration can limit features and spawn counts or disable Orca, supporting a reasonably clear least-privilege and confirmation story (2). Session, run-artifact, mirror-file, and child-output flows are described in detail, including a mirror cap and private file modes (3); sensitive content is still written to sessions and temporary mirrors, with cleanup described as delayed rather than immediate (2). Dependencies are pinned and the release workflow uses npm provenance, but the supplied material shows no vulnerability audit or dependency-update policy (1). Orca is an optional best-effort observer whose failure does not affect the authoritative run; release publication requires GitHub environment approval and some stop/recovery controls exist, though recovery coverage is incomplete (2). Repository and author fields are present, but the publisher is unverified by the curated registry, so attribution is not fully established (2).

README, configuration docs, and package metadata are broadly consistent and describe multiple execution modes and boundaries; the supplied material does not cover all code, so full consistency cannot be established (2). Pi requirements, dependencies, and platform limits are described, and CI spans several platforms, though some environments and external programs remain conditional dependencies (2). Invalid configuration, child failures, and unavailable tools have specific error-handling descriptions, but the supplied files do not provide a complete error reference (2).

The project covers review, scouting, implementation, research, parallel workflows, and background jobs with natural-language examples (3). Boundaries are explicit, including authorization, feature switches, spawn budgets, platform limits, and the observer's non-authoritative role (3). Activation timing, when tools appear, and invocation patterns are clearly documented (3). npm Pi and the official Linux x64 standalone release are covered, with Orca system limits stated; use still depends on Pi version, model tool capabilities, and external integrations (2).

README, configuration, agents, models, workflows, observability, and other docs have clear separation and an accessible install path (3). Installation is one command with a minimum Pi version, but the supplied material does not establish a detailed stepwise troubleshooting path (2). Package and export names are identifiable, but naming compatibility policy is not shown (2). Examples, workflow guidance, doctor, and topic-based help are provided (3). Platform, model tool activation, Orca availability, and configuration limitations are stated (3). The MIT license text is complete (3). A package version and release workflow are present; CHANGELOG contents were not supplied for assessing the quality of change history (2). Author metadata and release approval are visible, but maintenance commitments, support response, and responsibility boundaries are unclear (2).

Results are available in chat and through FleetView, inspectors, logs, and machine-readable artifacts; truncation, ordering, and failure results are documented (3). Focused agents, parallel review, background execution, and reusable workflows add clear delegation and oversight capabilities to Pi (3). Installation is simple and features can be trimmed, but multi-agent operation, run records, and configuration add maintenance cost (2).

Docs give concrete descriptions of configuration keys, parameters, platform boundaries, and behavior that make claims traceable (3). README, package metadata, license, and CI workflows provide complementary evidence, but implementation files and test results are not fully supplied (2). Docs distinguish user configuration assertions, verification behavior, and best-effort integrations; static evidence cannot independently establish runtime behavior (2).