OpenCode Swarm

The evidence shows strong least-privilege engineering: read-only review roles, scoped writes, containment and symlink guards, external-skill curation and optimization disabled by default, and oversight for higher-risk actions. This justifies full credit for least_privilege. Confirmation and external-effect controls are substantial—destructive reset requires --confirm and Full-Auto routes ambiguous or risky actions through oversight—but installation changes global configuration and disables native agents, while auto-proceed can remove phase prompts; user_confirmation and external_effects therefore each lose one point. The README identifies .swarm state, telemetry, providers, PR monitoring, and network-sensitive actions, but does not provide a complete field inventory, retention policy, destination map, or privacy treatment, so data_flow_transparency loses one point. Secrets scanning, dependency audits, quarantine, and hash/provenance validation are present, but credential storage, log redaction, and incident handling are not documented; sensitive_data_handling loses one point. Dependency auditing is asserted and the shown GitHub Action is commit-pinned, but runtime dependencies mostly use ranges and the supplied evidence includes no lockfile, vulnerability-response policy, or update automation; dependency_security loses one point. Tested configuration backups and the optimizer's atomic activation/rollback support full credit for rollback. Repository and npm metadata identify the project and an MIT license is supplied, but the license has no named copyright holder and publisher identity remains unknown, so source_attribution loses one point.

The README, package metadata, and limited CLI tests generally describe one product, but the installation documentation directly conflicts: the first-run note says a missing project override is created, while Quick Start says the installer does not create project configuration. self_consistency is therefore limited to 1. Bun/Node requirements, OpenCode dependencies, an optional peer, model fallback, and cache-upgrade paths are documented, but no lockfile or comprehensive offline/degraded behavior for external services is shown, so dependency_availability loses one point. Tests demonstrate clear handling for missing flag values, malformed JSON, missing environment variables, and hangs, but they cover only the swarm-model helper rather than systematic failures across the main orchestration pipeline; failure_messages loses one point.

The source clearly addresses everyday development, rapid iteration, parallel execution, unattended operation, PR feedback, and multilingual projects, earning full credit for audience_and_scenarios. Core, optional, and conditional agents and the strict/balanced/fast, Turbo, and Full-Auto modes provide meaningful boundaries, but broad claims such as compatibility with every OpenCode-supported provider are not qualified in detail, so capability_boundaries loses one point. Slash commands, issue labels, configuration switches, conditional agents, and the active-architect requirement make activation conditions precise, earning full credit for trigger_precision. Bun and Node requirements, Linux/macOS/Windows cache layouts, 13 language profiles, 20 grammars, and provider formats support full credit for environment_fit.

The README has clear navigation and dedicated architecture, mode, command, agent, upgrade, and configuration sections, earning full credit for information_architecture. Installation requirements, npm fallback, first-run effects, and cache-aware upgrades are concrete, earning full credit for install_notes. A stated command registry as the source of truth plus canonical names and deprecated aliases supports full credit for naming_stability. Examples, a recording script, and configuration snippets are useful, but the supplied evidence lacks a complete FAQ and the illustrated demo outcome is asserted rather than statically demonstrated, so examples_and_faq loses one point. Cache behavior, mode safety differences, command conflicts, and pause conditions are disclosed, but there is no consolidated, comprehensive limitation register, so known_limitations loses one point. LICENSE and package metadata consistently specify MIT, earning full credit for license. A semantic package version, upgrade instructions, and deprecation compatibility are present, but no changelog or complete version-to-change history is supplied, so versioning_changelog loses one point. The repository is identified, but no named rights holder, maintainers, support path, security contact, or maintenance commitment is shown; maintenance_responsibility is limited to 1.

Status, plan, evidence, diagnosis, resumability, and structured feedback are designed for usable operational output, and configuration tests show preservation of nested fields and backups. The main workflow was not run, however, and the supplied static evidence does not validate every claimed gate artifact, so output_usability loses one point. Role separation, independent review, scoped mutation, resumable state, and provider routing offer plausible added value over a single-agent flow, but the competitor comparison is unsupported by accompanying evidence; marginal_value loses one point. Free-model support, configurable agents, serial fallback, and an explicit cost burn-in show cost awareness, but there are no latency, token, resource, or measured quality-benefit results, so cost_benefit loses one point.

Several capabilities are tied to concrete commands, configuration keys, source-file names, issue identifiers, and test labels, supporting good claim_traceability. Major promotional claims such as “6000+” tests, review/testing of every task, and competitor comparisons are not individually substantiated by the supplied files, so one point is deducted. Version, engines, license, scripts, and some helper behavior are corroborated across README, package metadata, and tests, but implementation and tests for the central orchestration, security gates, and much of the agent roster are absent from the evidence set; cross_source_corroboration loses one point. The source distinguishes defaults, opt-in features, and conditional roles and identifies live registries as sources of truth, but the storyboard and comparison table still present expected behavior as established fact; fact_inference_separation loses one point.