Pi Jev Semantic Router

Automatic routing, model selection, compaction, and automatic orchestration are documented as off by default; tool activation is additive, and enable, disable, and runtime toggles are provided, supporting ordinary least-privilege, confirmation, and recovery expectations. The README discloses API-key sources, per-prompt request cost, evaluator state examples, and tool activation effects, but it does not fully describe everything transmitted to TypeSafe, retention, third-party processing, or a privacy policy. Sensitive-data guidance is limited to an environment variable and a plaintext secret-file example, with no file-permission, redaction, rotation, or leak-prevention guidance, so that criterion is thin. The dependency set is small, CI covers Node 20 and 22, and npm publishing uses provenance; wildcard peer dependencies and the absence of supplied lockfile or audit/vulnerability evidence limit dependency-security credit. Toggles and graceful fallbacks are clear, but no transactional reversal is described for tools already activated or subagents already dispatched. Author, repository, license, issue tracker, and security-contact attribution are consistent and comprehensive, justifying full source-attribution credit. Unknown enterprise-registry identity is treated as unverified rather than suspicious.

The README, configuration names, and tests agree on the shared JEV_THRESHOLD, opt-in defaults, exact subcommands, unconfigured skips, and non-throwing failure behavior. Tests explicitly cover concurrent skips and threshold consistency, supporting full self-consistency credit. Node requirements, installation, the sole runtime dependency, and fallback behavior are documented, but wildcard peer versions and the lack of a demonstrated compatibility matrix or detailed offline behavior reduce dependency-availability credit. Tests cover unknown commands, invalid arguments, evaluation-design failures, configuration-origin status, and error severity; the gate also distinguishes pass, rejection, and error exit codes, providing strong failure-message evidence.

The documentation covers interactive coding, skill discovery, tool routing, CI gates, classification, compaction, model selection, and multi-agent workflows across several task types, giving strong audience-and-scenario coverage. Boundaries include thresholds, skip conditions, missing configuration, low confidence, model compatibility, and fallbacks, but the source does not systematically identify high-risk decisions unsuitable for Jev, probability-error risks, or input-size limits, so capability boundaries are not complete. Exact subcommand matching, a shared cutoff, slash-command skipping, concurrency protection, and corresponding tests provide thorough trigger-precision evidence. Pi and Node 20+, environment variables, and secret-file setup are specified, but wildcard peers and absent OS, Pi-version, and network-requirement matrices limit environment-fit credit.

The README is clearly organized by features, installation, setup, modes, commands, tools, development/testing, and licensing. npm and GitHub installation, key configuration, status checking, and development commands are actionable. The jev_* tool names, /jev commands, environment variables, and gate binary aliases remain consistent across documentation, package metadata, and tests. Examples cover JSON tool calls, workflows, subagent gates, pipelines, and commands; the lack of a separate FAQ does not materially reduce their usefulness. Defaults, costs, fallbacks, and some skip conditions are documented, but there is no consolidated limitations section covering privacy, misclassification, or orchestration risks. MIT metadata matches the complete LICENSE, earning full license credit. Version 0.4.0 is declared and CHANGELOG.md is listed for publication, but its contents are absent from the supplied evidence, preventing assessment of history quality. A security email, issue URL, and author are present, while maintenance commitments, response times, governance, and support lifecycle are unspecified.

Typed judgments produce probabilities or choices, status reports expose configuration and usage, and the gate offers explicit exit codes plus JSON output, making results readily usable. Semantic discovery, typed low-overhead decisions, and workflow gates provide clear incremental value over permanently loading every tool or launching full LLM subprocesses. Cost controls are disclosed through opt-in execution, one Jev request per automatic prompt, local fallback shortlists, and request/token counters. However, claims such as “sub-second” and “calibrated,” and the magnitude of context or cost savings, lack benchmarks, measurements, or pricing analysis, so cost-benefit is not fully demonstrated.

Core behavior is traceable to concrete configuration, scripts, and tests: shared thresholds, opt-in defaults, exact command parsing, missing-configuration handling, concurrent skipping, failure reporting, agent RPC, CI, and publishing all have static evidence. The README, package metadata, workflows, and tests corroborate installation and core routing behavior, but automatic model selection, compaction, orchestration generation, gate implementation, and security-sensitive data handling lack corresponding supplied implementation or test evidence, reducing cross-source corroboration. The documentation generally separates configuration facts from behavior descriptions, but strong phrases such as “calibrated,” “safe,” “sub-second,” and “never silently truncates” are not statically substantiated or labeled as expected claims, so fact-versus-inference separation is incomplete.