Research

Exploratory notes. Not-yet-normative findings — read as drafts, not contracts. (Back to docs home.)

Note	Investigates
Agency Agents Evaluation	Eval spike (#1234): what the msitarzewski/agency-agents persona collection offers BoB’s agent framework — catalog, structural comparison, candidates, and verdict.
AI Website Cloner Evaluation	Eval (#1239): ai-website-cloner-template’s /clone-website skill read against BoB — its phases mapped to BoB primitives, its parallel builders against bob fork and bob swarm, its retired AGENTS.md sync scripts as input to #500, and a borrow-patterns verdict.
Claude Code Ecosystem Triage 2026-10	Eval (#1052): the ‘Claude Code Resource Bible’ list triaged against BoB — every link resolved and checked live, adopt / reference / skip per tool, multiplexers compared with warp-drive and bob fork, agent frameworks against BoB’s orchestration model.
claude-video Evaluation	Eval (#1024): bradautomates/claude-video’s /watch skill — how it downloads, samples frames and transcribes, its two engines and where each sends the video, its dependencies, and a reference-only verdict: install the upstream plugin when a need appears, never vendor it.
CLI Agent Orchestrator Evaluation	Eval (#1048): awslabs/cli-agent-orchestrator (CAO) read against warp-drive, bob fork and bob swarm — architecture, overlaps and gaps, the patterns worth keeping, and a reference-only verdict.
Ecosystem Positioning	Where BoB/warp-drive sits vs. the wider Claude Code community; weak points and over-engineering candidates through a harness-agnostic lens.
Jev Model Router Evaluation	Eval (#2667): whether the jev-model-router Claude Code mod should route BoB’s subagent models and main-loop effort — source review, fit against the model-assignment policy, and a no-go verdict with re-evaluation conditions.
Ondrej Skills Library Evaluation	Eval (#1045): David Ondrej’s open-source skills repo (davidondrej/skills, 56 skills) mapped against BoB — what each category overlaps, what conflicts with BoB’s conventions, the one pattern taken (a tested catastrophic-command guard, #3563), and a reference-only verdict.
Parallel Agent Orchestration	Patterns for fanning out work across multiple agents.
PlayCanvas Tooling Evaluation	Eval (#2312): PlayCanvas’s declarative 3D tooling spiked in throwaway projects — @playcanvas/react and web-components built, rendered and measured against the plain engine, create-playcanvas compared with BoB’s provisioning, playcanvas-sync’s local↔Editor sync read for patterns, and an archive verdict: use the upstream scaffolder when a project needs 3D, add no template.
PM Skills Evaluation	Eval (#1029): phuryn/pm-skills, a marketplace of 69 PM skills and 42 commands, inventoried against BoB — what overlaps, the three candidates (intent-vs-code auditing, an unsteered release review, pre-mortem), the article’s evals and shipping sections, and a mostly reference-only verdict.
SkillClaw Evaluation	Eval (#1240): AMAP-ML/SkillClaw read against BoB — source and paper confirmed, how its proxy rewires Claude Code, its SKILL.md format against the orchestrator frontmatter, its validated-publish gate, and a pass verdict against the #2387 skill-evolution design.
Spec-Driven Development Skill Evaluation	Eval (#1046): the FredAntB/Spec-Driven-Development Claude skill compared with BoB’s scope-control surfaces — how it enforces a spec, where BoB already does more, the one idea taken on, and a pass verdict.
SSG Evaluation	Whether a different static-site generator would fix the Quartz flat-serving link bug without post-build hacks (#487).
Agent-Tool Evaluation 2026-07	Fit assessments of seven trending agent tools against BoB’s stack (#544) — verdicts, trial notes, pilot recommendation.
Tracer Classifier Evaluation	Eval (#2738): the Jev noul that decides whether a requirement is greenfield or crosses more than one layer — the labelled set, the offline baseline, and the live threshold tune still owed against a pinned model version.
vuln-hunt Evaluation	Eval (#2668): Cloudflare’s security-audit skill, vendored as registry/skills/vuln-hunt — how it fits BoB’s agent and orchestration model, how its sandbox requirement is met on the fleet and where it is not, and the orchestration patterns worth reusing in other BoB skills.
Warp-Drive Model Routing	Cheap-tier delegation of mechanical chunks and the cost model behind it.
Hosted X MCP Evaluation	Eval (#837): X’s hosted MCP server (api.x.com/mcp) read from its docs — the two auth routes, pay-per-use pricing and rate limits, how Claude Code would connect, data and privacy, and a reference-only verdict that leaves wiring to the first consumer (#825), opt-in per project.