Guild

README states local-first operation, on-machine default embedding, and that optional Ollama sends embedding text to a configured endpoint, giving decent data-flow transparency (2). least_privilege scores 1: `guild hooks install` rewrites editor settings files and injects lifecycle hooks; it claims to only rewrite guild-prefixed command groups, but this is write access to user config with no explicit least-privilege constraint. user_confirmation scores 1: init is guided with prompts, but confirmation boundaries for hook installation and MCP client auto-registration are vague. sensitive_data_handling scores 1: SECURITY.md points to a README Privacy section that is absent from the evidence, so usage.log contents and opt-out cannot be verified. external_effects scores 1: writes to settings files, PATH, and ~/.guild state exist, with rollback only referenced via a link (docs/runtime-and-upgrades.md not provided). rollback scores 1: only a claim of rollback docs, no concrete mechanism. dependency_security scores 1: go.mod pins recent versions but there is no checksum verification, SBOM, or dependency audit evidence; onnxruntime-purego is a pseudo-version commit. source_attribution scores 2: LICENSE carries a copyright line and the model workflow emits a MANIFEST recording BAAI provenance and tool versions, but publisher identity is unverified.

README is internally consistent: install paths, retrieval degradation, and Windows limitations align (2). dependency_availability scores 1: requires the unusual Go 1.26.9 and an untagged pseudo-version of onnxruntime-purego, posing availability risk; no vendoring or checksum notes. failure_messages scores 2: agent mode defines an ok/error/hint envelope with exit codes, and embedder failure degrades with a reported reason, but no actual output samples are verified.

Targets coding agents across multiple MCP clients (Claude Code, Codex, Cursor) with concrete scenarios (2). capability_boundaries scores 2: clearly separates journal/lore/brief write surfaces and notes Windows lacks semantic retrieval and go install lacks embedding. trigger_precision scores 1: session start relies on the agent interpreting natural-language instructions, and hook triggering depends on environment-variable detection, so precision depends on external harness behavior with insufficient evidence. environment_fit scores 2: covers macOS/Linux/Windows/Docker, but Windows is feature-limited and the Go version requirement is strict, hence the deduction.

README is well-structured with clear sections pointing to docs/ and examples/ (2). install_notes scores 3: installation is thorough, covering curl, brew, PowerShell, make, go install, and Docker with trade-offs for each. naming_stability scores 1: heavy bespoke terminology (quest/lore/oath/brief/guildmaster), and the agent-mode envelope claims append-only but carries no version number, so stability evidence is weak. examples_and_faq scores 1: only links to an examples/ directory not shown in evidence, and no FAQ. known_limitations scores 2: explicitly lists Windows lacking semantic retrieval, go install limits, and Codex lacking compaction lifecycle. license scores 3: full Apache-2.0 text with copyright line. versioning_changelog scores 1: SECURITY.md says pre-1.0, no CHANGELOG, no versioning policy detail. maintenance_responsibility scores 1: publisher unverified, only SECURITY.md mentions response timelines, ownership unclear.

Agent mode emits a single JSON envelope with clear fields, easy for programs to consume (2). marginal_value scores 2: cross-agent shared context and atomic task claiming have real value, but many similar MCP memory solutions exist and differentiation evidence is limited. cost_benefit scores 1: requires an embedded ONNX runtime, ~127MB model download, and a specific Go version, so deployment cost is high while benefit is unquantified.

README claims CI badges, Docker tests, and retrieval evaluation docs, but evidence shows only workflow files, no test results or evaluation data (1). cross_source_corroboration scores 1: README partially aligns with SECURITY.md and go.mod, but the Privacy section is missing and examples/docs content is not provided, limiting corroboration. fact_inference_separation scores 1: docs blend factual description with mythic narrative, and performance/retrieval-quality claims lack supporting data, blurring fact and inference.