feat(registry): open native library projects at schema 29; repoint ex-mcp #1649

Merged

binjovi-bot merged 2 commits from feat/ex-mcp-project into trunk 2026-09-12 14:40:56 +00:00

Conversation 0 Commits 2 Files changed 6 +136 -4

sean commented 2026-09-12 14:03:16 +00:00

Owner

Copy link

Why

sean/ex-mcp is the fork Binjovi builds its own MCP server from, and it is not a Binjovi project. It cannot become one today: schema 21 pins the project to four names — bytey, hello-lib, libseanfarm-operator, zot-src — so a new library cannot be added by a plans-only pull request at all.

Schema 29

Widening v21 would change what a durable generation means for records already written against it, which its own comment forbids. So 29 is a new generation: v21 with the allowlist replaced by the two properties the allowlist was standing in for.

defp parse(%{"version" => 29} = values) do
  with true <- valid_dns_name?(values["project"]),
       true <- valid_repo?(values["repo"]),
       true <- values["repo"] == "sean/" <> values["project"] do
    parse_native_project(values, @v21_keys, nil)
  else
    _ -> {:error, :invalid_project_fields}
  end
end

Everything else is @v21_keys and parse_native_project unchanged: no taxonomy key (a library delivers nothing, so it names no delivery shape), executor native, recipe rooted at recipes/<project>.

valid_dns_name? is what rejects ex_mcp, and why the repository was renamed to sean/ex-mcp.

29, not 28 — the suite caught this

I first wrote it as 28. 28 is already taken by the binjovi-plans catalog clause, whose head also pins "project" => "binjovi-plans". A generic clause at 28 silently swallowed every other project at that version and turned :invalid_project_version into :invalid_project_fields.

object_store_project_test.exs:105 — "no other project may declare an object-store target" — failed, which is exactly its job.

Worth recording, so the next person does not repeat it: my version survey grepped defp parse(%{"version" => N, and 28's literal lives inside a multi-line with pattern at line 1229, not a clause head. The comment now says to grep every "version" => N literal in the file.

Dependency repoint

apps/binjovi_api, apps/binjovi_events and mix.lock now name code.sean.farm/sean/ex-mcp.git. The pin is by SHA (2bbb22ba…), so only the URL moved. Forgejo's post-rename redirect kept the old URL resolving — this stops depending on it. Verified with a clean deps.get against the new URL.

Also

One assertion in my new tests was wrong and is removed rather than worked around: parse_native_project legitimately allows deployment_targets, and bytey is the same generation shipping configmap/workflows/bytey-floorplan. ex-mcp simply has none.

Verification

891 tests pass, full umbrella suite, no failures. Postgres-tagged tests are excluded locally and run in CI.

Sequencing: this must be released before the binjovi-plans catalog record, because the decoder has to exist in the running image before a bundle naming version 29 can activate. The seanfarm side is PR #674.

https://claude.ai/code/session_01XdBRc9CHgFhAxCyHbdVtt6

## Why `sean/ex-mcp` is the fork Binjovi builds its own MCP server from, and it is not a Binjovi project. It cannot become one today: **schema 21 pins the project to four names** — `bytey`, `hello-lib`, `libseanfarm-operator`, `zot-src` — so a new library cannot be added by a plans-only pull request at all. ## Schema 29 Widening v21 would change what a durable generation means for records already written against it, which its own comment forbids. So 29 is a new generation: v21 with the allowlist replaced by the two properties the allowlist was standing in for. ```elixir defp parse(%{"version" => 29} = values) do with true <- valid_dns_name?(values["project"]), true <- valid_repo?(values["repo"]), true <- values["repo"] == "sean/" <> values["project"] do parse_native_project(values, @v21_keys, nil) else _ -> {:error, :invalid_project_fields} end end ``` Everything else is `@v21_keys` and `parse_native_project` unchanged: no taxonomy key (a library delivers nothing, so it names no delivery shape), executor `native`, recipe rooted at `recipes/<project>`. `valid_dns_name?` is what rejects `ex_mcp`, and why the repository was renamed to `sean/ex-mcp`. ## 29, not 28 — the suite caught this I first wrote it as 28. **28 is already taken** by the `binjovi-plans` catalog clause, whose head also pins `"project" => "binjovi-plans"`. A generic clause at 28 silently swallowed every *other* project at that version and turned `:invalid_project_version` into `:invalid_project_fields`. `object_store_project_test.exs:105` — "no other project may declare an object-store target" — failed, which is exactly its job. Worth recording, so the next person does not repeat it: my version survey grepped `defp parse(%{"version" => N`, and 28's literal lives inside a multi-line `with` pattern at line 1229, not a clause head. The comment now says to grep **every** `"version" => N` literal in the file. ## Dependency repoint `apps/binjovi_api`, `apps/binjovi_events` and `mix.lock` now name `code.sean.farm/sean/ex-mcp.git`. The pin is by SHA (`2bbb22ba…`), so only the URL moved. Forgejo's post-rename redirect kept the old URL resolving — this stops depending on it. Verified with a clean `deps.get` against the new URL. ## Also One assertion in my new tests was wrong and is removed rather than worked around: `parse_native_project` legitimately allows `deployment_targets`, and `bytey` is the same generation shipping `configmap/workflows/bytey-floorplan`. `ex-mcp` simply has none. ## Verification `891 tests pass`, full umbrella suite, no failures. Postgres-tagged tests are excluded locally and run in CI. Sequencing: this must be **released before** the `binjovi-plans` catalog record, because the decoder has to exist in the running image before a bundle naming version 29 can activate. The seanfarm side is [PR #674](https://code.sean.farm/sean/seanfarm/pulls/674). https://claude.ai/code/session_01XdBRc9CHgFhAxCyHbdVtt6

sean added 1 commit 2026-09-12 14:03:16 +00:00

feat(registry): open native library projects at schema 29; repoint ex-mcp ... a01ba24ea0

Two changes that together let sean/ex-mcp become an ordinary Binjovi project.

SCHEMA 29. Version 21 pins the project to four names -- bytey, hello-lib,
libseanfarm-operator, zot-src -- so a new library could not be added by a
plans-only pull request at all. Widening v21 would change what a durable
generation means for records already written against it, which its own comment
forbids. So 29 is a new generation: v21 with the allowlist replaced by the two
properties the allowlist stood in for.

  * the project is a DNS label -- this is what rejects `ex_mcp`, and why
    sean/ex_mcp was renamed to sean/ex-mcp;
  * the repository is exactly sean/<project>.

Everything else is @v21_keys and parse_native_project unchanged: no taxonomy
key, executor native, recipe rooted at recipes/<project>.

29, NOT 28. 28 is already the binjovi-plans catalog clause, whose head also
pins "project" => "binjovi-plans". A generic clause at 28 swallowed every other
project at that version and turned :invalid_project_version into
:invalid_project_fields. object_store_project_test.exs:105 caught it. The
version survey that missed it grepped clause heads; the literal lives inside a
multi-line `with` pattern. A comment now says to grep every `"version" => N`.

DEPENDENCY. apps/binjovi_api, apps/binjovi_events and mix.lock now name
code.sean.farm/sean/ex-mcp.git. The pin is by SHA, so only the URL moved.
Forgejo's post-rename redirect kept the old URL resolving, and this stops
depending on it. Verified with a clean deps.get against the new URL.

A deployment_targets assertion in the new tests was wrong and was removed, not
worked around: parse_native_project allows targets, and bytey is the same
generation with "configmap/workflows/bytey-floorplan". ex-mcp simply has none.

891 tests pass.

Claude-Session: https://claude.ai/code/session_01XdBRc9CHgFhAxCyHbdVtt6

sean added 1 commit 2026-09-12 14:17:04 +00:00

fix(contract): pin the ExMCP dependency at its renamed URL ... 22e988e97a

The build failed at `make contract-check`, in tests/mcp-dependency-contract.sh.
It asserts the locked ex_mcp URL by hand, and still named
code.sean.farm/sean/ex_mcp.git after the repository was renamed to ex-mcp.

The local suite could not have caught this: `mix test` does not run
contract-check, and `mix deps.get --locked` succeeds either way because Forgejo
redirects the old URL. Pinning the URL by hand is exactly the right design --
a redirect is what would otherwise let the pin rot in silence -- so the guard
is corrected rather than loosened, and now says why the name changed.

Found with `binjovictl attempt-log`, which is the tool for this and which I
should have reached for before reasoning about the failure.

Claude-Session: https://claude.ai/code/session_01XdBRc9CHgFhAxCyHbdVtt6

binjovi-bot merged commit 32210f03c4 into trunk 2026-09-12 14:40:56 +00:00

binjovi-bot deleted branch feat/ex-mcp-project 2026-09-12 14:40:57 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

sean/binjovi!1649

No description provided.