Public, portable Codex plugins and skills derived from repeated agent-workflow evidence.
Every contribution must be useful without private environment context, use synthetic or sanitized examples, and pass the relevant skill and plugin validators. Raw sessions, personal paths, infrastructure topology, reachable endpoints, and credentials do not belong here.
The Codex marketplace manifest is .agents/plugins/marketplace.json; local plugin sources live under plugins/. Repo Graph is installed from its pinned canonical repository.
Install the local checkout as a marketplace, then install a plugin:
codex plugin marketplace add C:\path\to\agent-plugins
codex plugin add session-improvement-loop@fakoli-agent-plugins| Plugin | Purpose |
|---|---|
session-improvement-loop |
Research decisions and local inference candidates, evaluate scoped skill changes, and turn session friction into validated improvements. |
host-operations |
Diagnose Windows GPU ownership and environment credential loading safely. |
workflow-intake |
Pin and adapt external procedures before executing them locally. |
workspace-coordination |
Coordinate related repositories, worktrees, and ownership boundaries. |
feature-retro |
Prepare retrospective design reviews, user journeys, evidence registers and article briefs after a feature build, without publishing. |
codex-pi-coordination |
Delegate bounded implementation, research, review and validation to Pi in an owned OpenShell sandbox, with recorded checks and cleanup. |
repo-graph |
Build offline system diagrams and incremental local semantic/keyword search with bounded results. |
Install any catalog entry with
codex plugin add <plugin-name>@fakoli-agent-plugins.
$repo-graph /path/to/repo or $repo-graph https://github.com/owner/repo
writes a self-contained interactive view, Mermaid source, and JSON under the
user's cache. The clean-room scanner batches files, caches bounded import
extraction, and pages large directory maps. It makes no model calls by default.
The viewer includes a System tab, card/tree/radial diagrams, a file-count
treemap, a sortable data table, and a directed dependency matrix. Each view
shares directory drilldown and filters; export the diagram as SVG or the
filtered directory data as CSV.
Repo Graph's canonical source is fakoli/repo-graph.
The catalog pins the canonical v0.6.0 release and exact commit. Codex fetches that source during installation; this marketplace carries no copied runtime or skill.
Its generated SQLite index supports keyword search immediately and optional local
CPU embeddings. The Search tab in repo-graph serve OUTPUT opens results in the
diagram. The skill installs the optional semantic dependencies through the
committed uv lock. Setup, bounds, privacy and examples.
Repo Graph is one product with Pi, Codex and Claude support. Its CLI provides
repo-graph init --harness pi|codex|claude|all; see the canonical installation
guide for setup and scope. Product tests and ADRs live in that repository.
This catalog runs only its own entry and skill checks.
Use --jev to opt into one speculative TypeSafe System One request for
advisory component role labels; this sends top-level directory names only.
The session-improvement-loop plugin provides four entrypoints:
$research-synthesis: independent evidence gathering and one synthesis for a material unresolved decision, with native host agents and a shared budget.$local-inference-research: source-specific benchmark research, Chrome access when needed, and compact hardware-matched comparisons with explicit evidence limits. Reuses the project's decision tracker before searching, refreshes material changes, and preserves dated decisions and concrete reopening conditions. Workflow comparison and validation scope.$skill-improvement: one evidence-backed instruction proposal, frozen versions, independent checks, and scoped adoption or rejection.$review-session-improvements: bounded session inventory and evidence selection, followed by the shared methods when useful.
These skills work without Anvil Serving. The caller supplies the actual task, ranked objectives, tools, evidence location, success checks, authorized changes, and stop budget. Project-specific skills retain operational commands and gates. Astra-specific guidance is loaded only when relevant; the plugin preserves the host's configured model roles. No daemon, hooks, or model runtime is added.
Start a new task after installation to refresh native discovery. Resources resolve relative to the installed skill; no checkout or fixed cache path is required. A host without delegation can complete lead-only research, labeled as reduced coverage rather than a three-answer panel. Independent evaluation still requires a separate reviewer or executable acceptance evidence.