Jevbridge Decision Sidecar

CI explicitly grants read-only contents access, the package claims no runtime npm dependencies, and remote access is described as limited to user-configured TypeSafe or LLM services. However, no implementation files are supplied, so actual runtime privilege minimization cannot be confirmed and least_privilege receives only thin credit. The README describes confirm, escalate, abort, and destructive-action gating, but supplies neither gate thresholds nor implementation evidence, so user_confirmation is partial. SECURITY.md clearly says supplied state leaves the machine for configured services and that the heuristic backend stays local; the environment table also identifies provider selection, although provider retention, logging, and secondary use are not covered. Sensitive-data guidance warns against publishing credentials or customer source and provides private vulnerability reporting, but omits redaction, minimization, key rotation, and leak-response procedures. Zero runtime dependencies and a fixed temporary typechecker version reduce supply-chain exposure, while movable major-version Action tags and the absence of a lockfile, audit process, or update policy prevent full dependency-security credit. External effects are presented as decisions and gates rather than direct execution, with confirmation advised before destructive calls, but the actual boundary is not verifiable from these files. No transactional recovery, undo, or compensating-action design is documented, so rollback is absent. The license, acknowledgments, project links, and temporary publishing account are disclosed, but the repository account, intended organization, and copyright entity differ, and publisher identity is unverified, preventing full attribution credit.

README, package.json, SECURITY.md, and CI are broadly consistent about Node 22, stdio protocols, environment variables, and the lack of runtime dependencies. The discrepancy between the current account and intended organization is disclosed, supporting good but not exhaustive self-consistency. Node 22 is the sole required platform, remote backends are optional, and a local heuristic fallback exists when keys are absent, so dependency availability is handled thoroughly. No exception types, exit codes, JSON-RPC errors, timeout or retry behavior, or user-facing diagnostic text is shown, so failure_messages scores zero; this deduction concerns missing static failure-message evidence, not absent execution results.

The material clearly covers support routing, destructive-command gating, context retention, GUI next-action selection, and MCP and ACP host scenarios. Capability boundaries are unusually explicit: Jev does not generate text, Jevbridge is a decision sidecar, GUI choices use a closed action set, and the heuristic is expressly unsuitable as a safety signal. The backend:auto selection order and environment triggers are documented, but exact thresholds and conflict handling for execute, confirm, escalate, and abort are omitted, limiting trigger precision. Configuration examples span Node, CLI, library use, Claude Desktop, Codex, Cursor, OpenCode, and Zed, with an offline trial path, providing strong environment-fit evidence.

The README has a useful contents structure, distinct sections, capability tables, and protocol explanations; prerequisites and installation steps are directly usable. Names are mostly consistent across the CLI, package, and tools, but the package is unpublished and the repository remains under gamesonrblx while targeting tacticocc, creating transitional naming and distribution instability. Examples cover library, CLI, MCP, ACP, and computer use, but there is no dedicated FAQ or systematic troubleshooting guide. The heuristic bias and unfinished session resume, upstream proxy, publication, and transfer work are disclosed, though security and protocol-compatibility limits are not consolidated. MIT metadata matches a complete LICENSE, justifying full license credit. Version 0.1.0 exists, but there is no changelog, release history, or compatibility policy. Maintenance paths include Tactico contact details, Issues, and Security Advisories, yet no verified maintainer, response commitment, or explicit update owner is established, so maintenance responsibility remains thin.

Typed noul, choice, and score answers, gate actions, and named protocol tools make the intended outputs reasonably consumable, and examples show branching usage. Full response and error schemas and compatibility guarantees are absent, preventing maximum output-usability credit. A typed fast-decision sidecar has a plausible and clearly articulated marginal role, especially for routing and closed-set GUI actions, but no benchmark, comparison, or outcome data substantiates superiority. The files only suggest reducing frontier-model token use and provide an offline fallback; they do not quantify latency, API expense, extra calls, accuracy tradeoffs, or operational cost, so cost_benefit is weakly supported.

Basic metadata is traceable across package.json, LICENSE, SECURITY.md, and CI, but the core capability, gating, protocol-compatibility, and testing claims lack the referenced implementation and test files, leaving claim traceability low. Node requirements, licensing, repository location, dependency shape, and outbound-data behavior receive useful cross-file corroboration. The README distinguishes completed roadmap items, future work, heuristic limitations, and the planned repository transfer, but broad language such as “any LLM” exceeds the specific adapters evidenced, so fact/inference separation is good rather than complete.