# Reported issues for orangerail Pod holds 16 of 32 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user. Back to [orangerail](/mcp/orangerail). ## Most discussed ### ONT-112: Bring the npm page to the new front door, and cut 0.1.5 ## Problem #180 moved this project's front door from the approval gate to the generated tool surface, and published the register of claims that failed when they were measured. **None of it has reached npm.** npm renders a package page from the published tarball, so `npmjs.com/package/orangerail` still opens with the framing ten measured runs say is the weaker half: ```console $ npm view orangerail version 0.1.5-not-yet # today: 0.1.4, whose tarball still says "Decide once what your agent ma [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/181) · 2026-08-04 · closed · 0 comments ### ONT-110: List orangerail on the official MCP registry ## Problem `0.1.4` put `mcpName` on npm, which was the one thing that had to ship before a registry entry was possible: ```console $ npm view orangerail mcpName io.github.KimHyeongRae0/orangerail ``` npm versions are immutable, so that field could not be added to `0.1.3` after the fact and the listing was blocked on the release existing. It no longer is, and nothing else is stopping it. The registry at `registry.modelcontextprotocol.io` is where an agent host resolves a server by name rather [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/177) · 2026-08-04 · closed · 0 comments ### ONT-108: The README is 649 lines and the npm page is a different document with no images ## Problem `README.md` is 649 lines and `## Quickstart` alone is 226 of them. ```console $ awk '/^## /{if(h)print NR-1-s" lines\t"h; h=$0; s=NR} END{if(h)print NR-s" lines\t"h}' README.md 226 lines ## Quickstart 62 lines ## The run this is built for 47 lines ## Wire it into your agent host 43 lines ## Declaring a rule the generator cannot derive 26 lines ## v0 commands ``` The Quickstart's problem is not the eleven commands. It is the ~150 lines of prose between them explaining why each one i [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/171) · 2026-08-04 · closed · 0 comments ### ONT-101: The MCP server tells every client it is 0.1.0 ## Problem The MCP server reports a version it has not been since three releases ago. ```console $ node -p "require('orangerail/package.json').version" 0.1.3 $ printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}' \ | ./node_modules/.bin/orangerail mcp {"result":{"protocolVersion":"2024-11-05","capabilities":{"tools":{}},"serverInfo":{"name":"orangerail","version":"0.1.0"}},"js [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/163) · 2026-08-04 · closed · 0 comments ### ONT-102: The README nav line shows raw HTML entities off GitHub ## Problem The README's nav line separates its five links with ` · `. GitHub resolves the entity; a renderer that escapes ampersands first does not — and one is now showing it to strangers. ```console $ curl -sL https://glama.ai/mcp/servers/KimHyeongRae0/orangerail | grep -oE '.{40}nbsp.{30}' Quickstart  ·  What it does not govern $ curl -sL https://glama.ai/mcp/servers/KimHyeongRae0/orangerail | grep -c nbsp 10 ``` The reader sees this a [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/161) · 2026-08-04 · closed · 0 comments ### ONT-093: The Quickstart cannot reach the write it promises ## Problem #140 put the Quickstart on the first screen. #142 made every `examples/` README run as written. Nothing does that for the README's own Quickstart, and following it verbatim from a cold start does not work. **The promised payoff never happens.** The Quickstart closes (L191-196) with a human approving a staged delete and the agent's next `check_approval` executing it. Reproduced against `0.1.2` through a real MCP client: ``` --- call deleteCustomer {"id":1} --- Action staged for huma [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/152) · 2026-08-03 · closed · 0 comments ### ONT-088: The README's first paragraph promises an OpenAPI path that does not exist in v0 ## Problem `README.md:12`, the third sentence a visitor reads: > `orangerail init` turns a `prisma/schema.prisma` **or an OpenAPI spec** into an MCP server: a > `get` and a `list` per object, one action per write with a zod input schema, and nothing else. Both halves are false for the OpenAPI input. Against the shipped CLI, on a spec with `components.schemas`, a collection GET, an item GET, a POST and a DELETE: ```console $ orangerail init --yes --preset approval-for-writes --no-studio opena [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/143) · 2026-08-03 · closed · 0 comments ### ONT-081: Move the rules-file argument out of the README so the install comes first ## Problem `README.md` is 656 lines and the first command a reader can run is on line 281. ```console $ grep -n 'npx orangerail init --yes' README.md 281:$ npx orangerail init --yes --preset approval-for-writes --no-studio ``` Everything above it is argument. 158 of those lines — `### What a rules file already gets you`, `### What the tie does not survive`, `### The one that survived on its own terms` — are one sustained answer to a single question: *why not just write the rules in a `CLAUDE. [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/129) · 2026-08-02 · closed · 0 comments ## Most recent ### ONT-079: The Quickstart's prerequisites omit the adapter Prisma 7 requires ## Problem A first-time reader on Prisma 7 who follows the README Quickstart exactly as written meets `exit 1` at the first command it asks them to run. Reproduced cold on `2f5d1e3`, from `pnpm pack`ed `0.1.1` tarballs, in a throwaway project holding a Prisma 7 SQLite schema and nothing else: ```console $ npx orangerail init --yes --preset approval-for-writes --no-studio orangerail init: this repo is on Prisma 7+ (@prisma/client 7.9.1 (installed)) and no driver adapter for `sqlite` is installe [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/125) · 2026-08-02 · closed · 0 comments ### ONT-074: A row that drifts from its declared schema opens the gate ## Version `0.1.1` (`062e527`) ## What happened `packages/core/src/policy/where.ts:23-25` documents the `where` gate as fail-closed: > a `null`/`undefined` object (target not found) fails closed — we cannot confirm the condition > holds, so the action is refused (§4.5 secure-by-default). It fails **closed** on *no row* and **open** on a *wrong row*. Measured on `062e527` against the repo's own flagship fixture policy (`packages/core/test/fixtures.ts:64`, `where: { field: 'status', op: 'neq' [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/115) · 2026-08-02 · closed · 0 comments ### ONT-071: Two surfaces still render a stored value unguarded, and one tells the agent the wrong thing ## Version `0.1.1`, on the merge of #104 (ONT-068) and #105 (ONT-069) ## What happened Two surfaces still render a stored value unguarded. Each was found by a sibling ticket that stopped at a file it did not own, and named the gap rather than reaching across. **1. `audit_unrecorded` reaches the agent as `"Unexpected execute result."`** #105 added an `ExecuteResult` variant at `packages/core/src/lifecycle/engine.ts:105-113` for the honest outcome *"the side effect landed and nothing about it [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/107) · 2026-08-01 · closed · 0 comments ### ONT-067: The Prisma 7 default generator produces a green `init` and an ontology where every tool fails ## Version `0.1.1` (`752ff7b`) ## What happened `npx prisma init` on prisma 7.9.1 writes this generator block: ```prisma generator client { provider = "prisma-client" output = "../generated/prisma" } ``` `prisma generate` emits the client into `./generated/prisma` and puts nothing into `node_modules/@prisma/client`. `orangerail init` then prints its full success banner and exits 0: ``` ✓ scanned your sources — 2 object(s), 6 action(s) ✓ --gate delete: 2 of 6 write action(s) gated ✓ r [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/99) · 2026-08-01 · closed · 0 comments ### ONT-061: mcp: publish the action input contract, and name what it refuses ## Version `main` @ `8af6078` (the unreleased successor to `0.1.0`), on a project generated by `orangerail init` against a local Postgres. ## What happened Two things, and they compound into one failure. **1. The published `inputSchema` erases the type of every OPTIONAL field.** Verbatim from `tools/list` on a generated project: ```jsonc createorder_item → {"id":{"type":"string"},"order_id":{"type":"string"},"product_id":{"type":"string"},"qty":{"type":"number"},"unit_cents":{"type":"number [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/89) · 2026-07-30 · closed · 0 comments ### ONT-060: Report the MCP servers mounted next to this project ## Problem Nothing in this product reads any host agent configuration. ```console $ grep -rn "\.mcp\.json\|claude_desktop_config\|cursor/mcp" packages/*/src/ $ # (nothing) ``` `orangerail init` scans data sources — `schema.prisma`, an OpenAPI spec — and nothing else. So a user can run ```console $ orangerail init --models customer,order,order_item,product ``` get exactly the narrow governed surface they asked for, and have their agent's exposure change by **exactly zero**, because a wide SQ [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/85) · 2026-07-30 · closed · 0 comments ### ONT-056: `orangerail init` picks a gating posture instead of gating everything ## Problem `orangerail init` emits `policy: { approval: 'required' }` on every write action it generates. That reads as the safe default. It is the wrong one, and the reason is not theoretical — it is what an operator does next. On a six-action schema, nothing an agent calls ever completes. Every `create`, every `update`, every `delete` returns `approval_pending` and waits for a human. So the first thing anyone does after `init` is open the generated files and delete the gate from most of them [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/82) · 2026-07-30 · closed · 0 comments ### ONT-057: core: record what a governed write changed *from* ## Problem An orangerail audit record for a successful write carries the **input** and the **resulting row**, and never the **prior** one. Verbatim from a live run against a real Postgres: ```jsonc {"phase":"succeeded","action":"updateproduct","input":{"id":"p3","stock":25}, "result":{"id":"p3","sku":"SKU-CABLE","title":"USB-C cable","price_cents":1500,"stock":25}} ``` `p3.stock` was `0` before that call, and nothing on the chain says so. Two consequences, and the second is worse than the f [Read the thread](https://github.com/KimHyeongRae0/orangerail/issues/80) · 2026-07-30 · closed · 0 comments The remaining reports are on [the project's issue tracker](https://github.com/KimHyeongRae0/orangerail/issues).