ForgeRail / Health Research
Concrete repository changes and durable operating practices that reduce discovery, validation, and local startup time without sacrificing verification.
Internal
Read-only measurements from C:\Users\AdamK\Documents\forgeapps on 17 July 2026. The checkout contained concurrent work; no source, board, deployment, database, or external-system records were changed during research.
| Measurement | Result | Interpretation |
|---|---|---|
git status --short --untracked-files=normal | 147 ms · 32 lines | Fast enough; not a leading optimization target. |
git status --short --untracked-files=no | 70 ms · 21 lines | Only 77 ms saved; hiding untracked work is not worth the risk. |
git ls-files | 63 ms · 32,846 files | Git-aware inventory remains efficient. |
rg --files . | 66 ms · 22,925 files | Default source-visible breadth is still large. |
rg --files apps/forgerail/src | 19 ms · 150 files | Scoped discovery sees 153× fewer files. |
| Local filesystem inventory | 239,053 files | Generic walkers see dependencies, generated output, and local evidence. |
Root node_modules | 147,519 files · 1.60 GB | Must be excluded from recursive agent tooling. |
outputs/ | 1,600 files · 21.04 GB | Large media/output trees should not enter routine search or indexing. |
| Instruction surfaces | 2,057 AGENTS files on disk; 261 generated/dependency copies | Effective root-to-ForgeRail source chain is only 5,644 bytes; discovery duplication is the concern. |
| Skill library | 630 instruction files · 4.16 MB | Metadata-first discovery is appropriate; wholesale content loading is not. |
| Package topology | 92 named manifests · 80 workspace edges | 70 build, 71 test, 69 typecheck, and 61 lint scripts create a broad validation surface. |
ForgeRail tsc --noEmit | 6,877 ms · passed | Incremental checking has measurable potential. |
| pnpm package listing | 514 ms · failed | [ERR_SQLITE_ERROR] unable to open database file. |
| Repo context launcher | ~214 ms · failed | Error: spawn EPERM from the esbuild worker. |
Ranked Delivery Options
Impact ranks expected end-to-end Codex time savings. “Automate” means the change can be safely machine-enforced after its stated acceptance gate.
Route searches and context resolution to the closest workspace before considering the monorepo. Targets: root AGENTS.md, workspace resolver, health checklist.
Exclude .git, dependencies, builds, Vercel output, local media, browser evidence, and caches while retaining canonical source evidence.
Add content-addressed results around scripts/repo-check.ts for typecheck, test, lint, and build inputs.
Use an ignored, explicit tsBuildInfoFile and benchmark cold and warm --noEmit checks.
Skip unchanged requirements/meeting staging by hash, then evaluate background or lazy staging in apps/forgerail/vite.config.ts.
Resolve pnpm SQLite store access and esbuild child-process spawning without widening unrelated permissions.
pnpm list and repo context/dry-run commands succeed in Codex.Track search breadth, check latency, dev-ready latency, cache effectiveness, checkout breadth, and tool readiness with timestamps.
Choose one source for port 4225/4208 across package script, Vite config, README, and launch registry.
Extend changed-workspace checks with pnpm’s dependency and dependent selectors when shared packages change.
Set focused test.dir; benchmark threads and non-isolated projects only where cleanup is proven.
Select skills from validated metadata, then load only the chosen instruction and required references.
Build a solution graph for shared packages and true consumers—not a forced all-app project.
Use plugin-transform and CPU profiles to identify expensive hooks, modules, and resolution waterfalls.
Delivery Shape
Ship the bounded, observable changes first. Treat cache orchestration and project-graph redesign as experiments with baselines and rollback paths.
Guardrails
These ideas either save too little today, hide important evidence, or add structural risk before measurement justifies them.
Proposed Product Content
The page should describe evidence, not reassurance. Each card needs a command or collector, timestamp, commit, result, and explicit stale/unavailable state.
Codex Delivery Speed. Measures the local friction between a request and verified code. Missing evidence remains Unknown.
External
Current official OpenAI and primary tooling documentation supports smaller effective instruction scope, affected-package selection, incremental compilation, and evidence-led Vite/Vitest tuning.
Codex merges one instruction file per directory from repository root to the working directory and stops at a default 32 KiB. Model Fast mode is separate from repository/tool latency.
Filters can select exact workspaces, dependencies, dependents, directory sets, and projects changed since a Git reference. Recursive execution defaults to four concurrent tasks.
Project references and build mode avoid unnecessary recompilation; incremental compilation persists project-graph information for later runs.
Profile slow startup/plugin hooks before refactoring. Limit test-directory search and benchmark pools, isolation, caching, and sharding rather than applying them blindly.
Porcelain output is the stable automation surface. Current local timings do not justify making Git configuration a leading intervention.
Coverage And Caveats
Research used local source and health-page evidence plus current primary public documentation. Internal communications, client material, and unrelated business systems were intentionally excluded.
spawn EPERM.