Codex app as glue
Codex is the interface that reads the repo, follows the local instructions, runs scripts, checks git state, opens files, and turns a question into a repeatable workflow.
Project / Dynasty Codex
How the fantasy research system works.
Dynasty Codex is a local-first fantasy football research workspace where the Codex app acts as the operating layer across APIs, local data, knowledgebase notes, skills, plugins, and publishable reports.
The important design choice is that the repo is the application state. Codex reads it, updates it, runs its scripts, calls external tools when needed, and turns fresh context into decisions, notes, reports, or public writing.
Codex app orchestrationSkills and pluginsAPI snapshotsDuckDB analysisKnowledgebase and reports
System Map
The pieces.
It is not one model or one dashboard. It is a set of durable parts that make each new fantasy question start with memory instead of a blank chat box.
Skills and plugins
Project skills route common work like trade evaluation, Sleeper research, daily refresh, analyst ingest, KB lint, and SQL maintenance. Plugins add outside capabilities when needed, such as Gmail, GitHub, browser QA, or deploys.
Knowledgebase
Markdown files store source cards, player pages, concepts, league notes, and preserved research answers. Facts, analyst opinions, and my takes stay separate so the system can challenge itself later.
Local data layer
Raw API captures and processed tables stay local by default. DuckDB makes those snapshots easy to query, join, rebuild, and back up without running a database server.
Report generation
Scripts can produce Markdown, CSV, charts, HTML, and PDF-style report packages. The public site gets only cleaned, original synthesis that clears privacy and freshness checks.
Automations
Scheduled Codex runs can preview data refreshes, check advisory warnings, mine source queues, build action lists, and keep local backups without making every day a manual maintenance session.
Data Flow
From source to decision.
The workflow is intentionally boring in the right places: capture raw inputs, normalize them, query them locally, then promote only useful conclusions.
APIs, exports, and notes
Sleeper provides league reality: rosters, trades, drafts, picks, transactions, matchups, and league history. nflverse, DynastyProcess, ADP sources, college data, betting context, and analyst notes add public and subscription-safe context around those leagues.
Raw snapshots stay local
The repo tracks source registries, docs, scripts, and safe metadata. Bulky raw data, local databases, private league detail, and paid raw text stay in ignored local folders unless there is an explicit personal-data branch.
DuckDB becomes the workbench
Builder scripts turn CSV, JSON, Parquet, and release assets into schemas for Sleeper, market, best-ball, college, betting, and identity work. Codex can ask narrow questions against those views instead of inventing answers from memory.
Skills keep the workflow honest
A trade question starts from live roster and pick reality. A player page checks identity, production, exposure, market context, and source notes. A research claim needs a baseline, freshness label, validation check, and action threshold.
Knowledge compounds
Good answers become player notes, concept pages, daily digests, action queues, report packages, or public articles. Weak signals stay as watch items or get discarded, which is as important as publishing the winners.
Why This Shape
Local-first is the feature.
Fantasy research mixes private leagues, paid sources, stale data, public APIs, and personal judgment. The architecture has to make those boundaries visible.
Reproducible
Snapshot IDs, catalog metadata, source cards, and queryable tables make it possible to ask where a conclusion came from and whether it is still current.
Private by default
Paid newsletters, Gmail bodies, raw local exports, and league-specific details stay out of public publishing. The public layer is paraphrased, original, and stripped of private identifiers.
Agent-friendly
Instructions, skills, scripts, docs, lint checks, and generated manifests make the project navigable for future Codex sessions. A new agent can start warm.
Where It Goes
From research desk to product surface.
The long-term path is not just more pages. It is better decision support: sharper reports, interactive tools, and trusted workflows that know the league context before analysis starts.
Trade and roster cockpit
Combine manager tendencies, pick liquidity, roster windows, market values, historical baselines, and scoring settings into focused offer construction.
Edge watch dashboard
Track market dislocations, opportunity without points, analyst disagreement, format-specific value gaps, and stale blockers in one reviewable surface.
Public research pipeline
Promote only the best sanitized outputs into this site: technical notes, fantasy articles, report previews, draft boards, and small interactive tools.