Dashboards as code

Every dashboard in the stack is generated by grafana/build.py. Grafana provisions the generated JSON read-only, so the Python file is the source of truth and the Grafana UI holds no edits of its own.

Workflow

Edit grafana/build.py.
Run python3 grafana/build.py. It writes one JSON file per dashboard to grafana/dashboards/ and prints each panel count.
Commit the Python and the JSON together.
Copy the repository to the home server and run docker compose up -d --build.

On 2026-10-09 a rebuild from the committed build.py in a clean directory produced JSON identical to the seven committed files.

Building blocks

Helper	What it produces
`Layout`	Places panels left to right on Grafana's 24-column grid and starts rows
`stat`, `timeseries`, `heatmap`, `logs`, `table`	Panels with fixed defaults; a Loki stat runs an instant query
`loki_series`	A time series on Loki, drawn as lines or bars
`request_sum`	LogQL sum of one field of Claude Code's `api_request` events, with `keep` before `unwrap`
`dashboard`	Shared settings: tag agent-observability, not editable, a dashboards dropdown link
`query_var`	A multi-value template variable with All as `.*`

Generated dashboards

uid	Title	Panels with rows	Variables	Default range
local-llm	Local LLM (Ollama)	19	model	6 h
local-llm-public	Local LLM (Ollama), public	19	none	6 h
host	Host (Ollama box)	41	none	6 h
host-public	Host (Ollama box), public	36	none	6 h
claude-code	Claude Code agents	20	none	24 h
claude-code-public	Claude Code agents, public	10	none	24 h
site-deploys	Website deploys	18	none	7 days

Public cuts

Each public cut is a function that takes the internal dashboard and transforms it:

local_llm_public replaces every $model filter with .* and empties the variable list.
host_public swaps per-device queries for sums or maxima, wraps every expression in max without (...) over name labels, and drops the containers row.
claude_code_public drops the log, trace and skill panels and wraps every expression in sum without (...) over identifying labels.
site_deploys has no cut: it has no variables and no private details.

A public cut gets its own uid with a -public suffix, so the internal and public versions sit side by side in the folder "Agent Observability". Which cuts are shared: Public dashboards.

Provisioning

The provider in grafana/provisioning/dashboards/dashboards.yml sets disableDeletion: true and allowUiUpdates: false, and every dashboard sets editable: false. A change made in the UI cannot be saved over a provisioned file, so every change goes through build.py.

The now() helper in site_deploys() and the request_sum helper exist because of two traps: Stat panel shows No data and Counter resets from parallel sessions.