New app apps/mcp (NestJS + @nestjs-mcp/server 2): Streamable HTTP at
POST /mcp, stateless.
- 24 tools: projects (mine, public, details with settings, board, stats,
analytics, activity), tickets (my work, search, details with comments and
history, create, update, move, assign, release, comment, link), releases
(list, upcoming, details with progress and burnup, create, publish) and
users (whoami, search). Admin operations stay in the web.
- No credentials of its own: each call's Authorization header (a Pireon
token) is forwarded to the API; without it only public projects are read.
- Typed REST client generated from apps/api/openapi.json (openapi-fetch +
openapi-typescript via `nx run mcp:openapi-types`); API refusals become
readable tool errors; parameters validated with zod.
- Tests: the real MCP against a simulated API (token forwarding, errors,
validation, writes).
- Docker target `mcp` (Node 24) and an optional compose service under the
`mcp` profile; zod declared as a root dependency.
- Docs: chapter 9 «Servidor MCP», references in 01/06/07 and the index;
Word/PDF regenerated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pireon OAuth clients allow a single grant type. With one PIREON_CLIENT_ID of
type password, email + password worked but "Sign in with Pireon" failed with
`unauthorized_client`.
- PIREON_CLIENT_ID/SECRET: authorization_code client (Sign in with Pireon,
Swagger UI). New PIREON_PASSWORD_CLIENT_ID/SECRET: password client, falling
back to the former so existing setups keep working.
- Refresh tries both clients: a refresh token only works with its issuer.
- PIREON_PASSWORD_GRANT=false now also rejects password grants in the API
(before it only hid the form).
- /api/auth/config returns clientId: null when there is no authorization_code
client, and the web hides "Sign in with Pireon".
- docker-compose passes the new variables; docs 05/07, .env.example,
openapi.json; unit tests for the client selection.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In Dokploy, services with a domain join the shared dokploy-network, where
another project also has an `api`. nginx proxied /api to `api:3000`, so
Docker DNS returned both containers and half of the requests went to the
other app: 404 "Cannot POST /api/auth/token" on sign-in and "sign-in not
configured" when /api/auth/config was answered by it.
Services now get task-stream-* aliases on the project network and nginx and
the API use them (task-stream-api, -postgres, -mongo, -redis). The API's
OPENAPI_FILE is empty in the container; docs 07 explain the aliases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docker-compose.yml is the production definition (deployed by Dokploy) and no
longer publishes any port: postgres, mongo, redis and the API are reachable
only on the compose network, and the web goes out through the platform's
proxy (domain → service web, port 80). The ports moved to
docker-compose.override.yml, which Compose loads automatically locally and
Dokploy (-f docker-compose.yml) does not.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dokploy builds failed with "failed to stat /root/.npmrc": the compose build
secret pointed at ${HOME}/.npmrc, which only exists on a developer machine.
The token now comes from the NEXUS_TOKEN environment variable (Dokploy's
Environment tab or .env), is written to a temporary npmrc and removed in the
same RUN, so it never lands in an image layer.
Also removes a temporary UI dev-server spec that was committed by mistake,
and documents the variable (.env.example, docs 06/07, Word/PDF export).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
