> ## Documentation Index > Fetch the complete documentation index at: https://gladlabs.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Grafana dashboard links # Grafana dashboard links — why the host is rendered, not written A dashboard link that points at a **sibling service** (Prefect `:4200`, Langfuse `:3010`, GlitchTip `:8080`, pgAdmin `:18443`, Prometheus `:9091`, worker API `:8002`) needs a hostname, and for a long time it had the wrong one. ## The three link classes Not every link has the same problem, and conflating them is how the fix stayed partial for months. | Class | Correct form | Why | | - | - | - | | **Grafana's own pages** | relative — `/d/`, `/explore?...` | Inherits whatever origin the reader already reached Grafana on. Needs no host and no render. | | **Sibling service UIs** | `http://__POINDEXTER_SERVICE_HOST__:` | Different port, same machine — so it needs a host the reader's *browser* can resolve. | | **External sites** | the literal URL | `github.com`, `grafana.com`, `vercel.com` are public and fine as-is. | ## Why a render step exists at all Grafana **does not interpolate environment variables inside dashboard JSON**. `${__env.POINDEXTER_SERVICE_HOST}` is not a thing there — verified on `grafana-oss:13.0.1`, where the anchor came back with the literal `http://${__env.POINDEXTER_SERVICE_HOST}:4200` as its `href`. Dashboard *variables* (`$var`) don't help either: provisioned dashboards are file-backed, so a `constant` variable's value reverts on every re-provision. So the host cannot be configuration unless something rewrites the file before Grafana reads it. The two alternatives were both worse: * **`http://localhost:`** — what shipped. Resolves only for a browser on the Docker host. Dead from a phone, which is the normal way this system is operated. * **`http://:`** — works for exactly one operator, and puts a private hostname into a publicly-mirrored repo. The mirror leak guard refuses it, and the last time it went in anyway, the remedy was to strip the entire dashboard out of the OSS mirror (`mission-control.json` lived in `_STRIP_FILES` for that reason until 2026-09-23). ## How it works 1. Dashboards in `infrastructure/grafana/dashboards/*.json` carry the literal `__POINDEXTER_SERVICE_HOST__` placeholder. No `$`, so Grafana never tries to interpolate it, and it is obviously not a hostname if it ever leaks through. 2. Compose mounts that directory **read-only** at `/etc/grafana/dashboards-src`. 3. The `grafana` service's **entrypoint** renders src → `/var/lib/grafana/dashboards`, substituting `${POINDEXTER_SERVICE_HOST:-localhost}`, then `exec`s Grafana's own `/run.sh`. 4. The file provisioning provider reads the **rendered** directory. 5. `scripts/start-stack.sh` resolves `app_settings.operator_service_host` (via `scripts/_grafana_service_host.py`) and writes it into `.poindexter-grafana.env`, which the grafana service loads as an `env_file` — so a plain `docker compose up -d` outside `start-stack.sh` still picks it up. Same mechanism, and same rationale, as `GRAFANA_WEBHOOK_TOKEN`. ### Three details that are load-bearing * **The render target is inside `grafana-data`**, not a second named volume. A fresh named volume is created **root-owned** and Grafana runs as uid 472, so the first render dies with `Permission denied` and the container exits 1. * **It renders in the existing service's entrypoint, not an init container.** Adding a compose service changes the set that brain's `compose_drift_probe` audits, and an init container that exits reads as "down". * **A 30-second re-sync loop runs in the background.** The bind mount used to give hot-reload (edit JSON → provider rescans in 30s → board updates). Rendering once at start would have quietly removed that. Files are compared and only replaced when they differ, so an unchanged board is never rewritten — a needless mtime bump would make Grafana re-provision every cycle. Rendered copies whose source disappears are pruned. ## Setting it ```bash theme={null} poindexter settings set operator_service_host my-workstation.example.ts.net bash scripts/start-stack.sh up -d ``` Host **only** — no scheme, no port, no path. `_grafana_service_host.py` refuses anything else and falls back, because a malformed value would be substituted into every dashboard URL and produce silently broken links rather than an error. Default is `localhost`, which reproduces the historical behaviour exactly: right for a browser on the Docker host, and the value a fresh install wants. ## Guards | Guard | Catches | | - | - | | `scripts/ci/grafana_dashboard_host_lint.py` | A link hardcoded back to `localhost`/`127.0.0.1`. Loopback **only** — operator-host leaks belong to `check_public_mirror_safety.py`, and a lint that also reddened on `github.com` would get ignored then deleted. | | `tests/unit/scripts/test_grafana_dashboard_render_wiring.py` | The mount, the entrypoint, the fallback, the provider path, and bare placeholders in prose. | | `tests/unit/scripts/test_check_public_mirror_safety_strip_list.py` | `mission-control.json` re-acquiring an operator hostname or the voice URL. | | `scripts/system-health-check.sh` | An **unrendered placeholder surviving in the container** — i.e. the render silently no-opping, which would otherwise look healthy while every cross-service link is a dead hostname. | ## Related * The **Explore** links (Tempo, Loki) are a different bug with the same symptom: they are `/explore` deep-links, and Grafana gates Explore behind `datasources:explore`, which the Viewer role does not carry by default. Fixed with `GF_USERS_VIEWERS_CAN_EDIT=true`. The legacy `left={...}` URL format is *not* deprecated to the point of breaking — it still loads the query on 13.0.1. * `mission-control.json` now **ships to the OSS mirror** for the first time, a side effect of having no operator literal left in it. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.