Various AI/LLM/harness-related tools and utilities. Each top-level directory is a self-contained component with its own docs; they may reference each other but each stands on its own.
| Directory | What it is |
|---|---|
vendor-clients/ |
A generic toolset for calling LLM vendors from the CLI and getting back structured, typed output -- one uniform interface across vendors, a routing dispatcher, vendor discovery, per-vendor rate-limit budgeting, and a local call trail. Ships a Gemini client; adding more is a small, defined task. See its README (concepts, adding a vendor) and USAGE (how to invoke). |
cross-vendor-review/ |
A skill/playbook for getting a plan, test spec, or code change reviewed by a non-Anthropic model (via vendor-clients/) -- a second opinion whose cognition is the other vendor's, to catch the blind spots a same-family reviewer would share. See its SKILL.md. |
agent-session-analysis/ |
Measures what a Claude Code agent actually spends its time doing over a date window -- working time, tokens, per-session work-type, and knowledge-base read/write usage -- from session transcripts. Built on three strictly-separated epistemic tiers (MEASURED / CLASSIFIED / INDICATIVE) so an honest usage picture never gets conflated into a vanity number. See its README (overview), RUNBOOK (how to run), and METHODOLOGY (the why). Run data stays local; only the methodology + tools are shared. |
project-knowledgebase/ |
A file-based project knowledge base pattern for working with AI agents, plus a ready-to-use directory template that implements it: a small, predictable set of plain files an agent loads selectively (tiered loading) so each conversation starts oriented instead of cold. Ships the blank project-template/ and a downloadable project-template.zip of it. See its README (the concept) and the template's own README (setup + agent protocol). |
Apache-2.0 (see LICENSE), unless a specific directory states otherwise in its own README/LICENSE.