Verified the direct_* providers end-to-end after billing was enabled.
- OpenAI direct_gpt-4o / direct_gpt-4o-mini: working.
- Gemini: gemini-2.0-flash 404s (LiteLLM 1.55.10 rewrites it to a retired
experimental name) and gemini-1.5-pro is retired -> switch to the current GA
gemini-2.5-flash / gemini-2.5-pro (both verified).
- Drop direct_o3-mini: o-series needs max_completion_tokens, which 1.55.10 won't
translate from the max_tokens clients (Open WebUI) send -> 400. Re-add after a
LiteLLM bump.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Backend-prefix taxonomy so the Open WebUI picker is self-documenting and a
model name can't lie about where it routes:
local_* -> Anvil/Ollama (free) e.g. local_qwen2.5-72b
proxy_* -> Claude via Meridian/Max e.g. proxy_claude-sonnet-4-6
direct_* -> metered OpenAI/Gemini e.g. direct_gpt-4o, direct_gemini-2.0-flash
Drops the redundant -max suffix (proxy_ already implies Max). api_base is now
emitted only when a model defines it, so direct_* hit the provider default
endpoint instead of Meridian. direct_* are SCAFFOLDED (no live keys): litellm.env
writes a placeholder so the proxy boots; deploy.sh pulls OPENAI_API_KEY/
GEMINI_API_KEY from Infisical /meridian if present (non-fatal). They 401 until
real keys land.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Honest model names only — local picks up real Ollama names (qwen2.5-72b,
llama-3.3-70b, llama-3.1-8b, nomic-embed-text), Claude via *-max only.
The shadows were briefly useful (paperless-ai wizard probe quirk) and
then briefly used to make the ALL-LOCAL cutover transparent to clients,
but having "claude-sonnet-4-6" silently route to llama3.3:70b in the
Open WebUI picker was a constant foot-gun. Pulse re-pointed to a clean
alias in its UI prior to this push; paperless-ai was already on
qwen2.5-72b. Trade-off captured in [[litellm-openai-alias-shadowing]].
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the short-lived mistral-large alias. Backed by ollama_chat/qwen2.5:72b
on Anvil. Consumers (paperless-ai, RAG chat, HA, morning-report) target this.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Per-model api_base/api_key overrides in the template (default stays
Meridian's local port). All standard aliases (claude-*, gpt-*) now point
at Anvil's Ollama (mini/haiku-class -> llama3.1:8b, rest -> llama3.3:70b).
Claude/Max reachable only via new *-max escape-hatch aliases.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Earlier Pass 2 cleanup sed only matched bare '- node_exporter' lines;
these used '{ role: ..., tags: [...] }' syntax which fell through.
Deploys via Semaphore were erroring with 'role node_exporter not found'.
Embedded prometheus.exporter.unix in Alloy has been pushing identical
metrics fleet-wide since 2026-05-21 cutover; the standalone binary
and systemd service have been removed from each host. Drops the
role + site.yml entry so future deploys don't reinstate them.
See homelab-docs/docs/audit/alloy-consolidation-2026-05-21.md.
Drops the _canary suffix on alloy_prom_job. Prometheus retired its
static node_* scrape jobs in the same release; Alloy's remote_write
fills the gap with identical job/instance/group/hostname labels.
Embeds node_exporter inside Alloy alongside Loki shipping; pushes
metrics via remote_write to observe Prom with job=node_lxc_canary
to run side-by-side with the existing node_exporter scrape until
cutover. See homelab-docs/docs/audit/alloy-consolidation-2026-05-21.md.
Meridian + LiteLLM both run as systemd services on this LXC (no docker)
so the Docker-container Alloy pattern from other repos doesn't apply.
Apt-install grafana/alloy via apt.grafana.com, journald-only scrape,
ships to Loki on observe.lan.balders.ca.
Side benefit: Meridian.service + LiteLLM.service logs (including the
gpt-* alias shadowing requests from paperless-ai) now searchable in
Loki, not just journalctl on the LXC.
paperless-ai's setup wizard validates the OpenAI provider by hardcoding
model=gpt-4o-mini in the probe, regardless of the OPENAI_MODEL env. Without
the alias LiteLLM 400s ("Invalid model name") and the wizard rejects the
key. Shadow common OpenAI names onto our Claude backends so any client that
probes gpt-* gets a healthy response (and routes to the Max sub).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
LiteLLM sits in front of Meridian for clients that can't talk Anthropic's
/v1/messages format (Pulse OpenAI provider, paperless-ai, etc.). Routes
OpenAI-shaped requests to localhost:3456 (Meridian) which forwards to the
Max sub.
- New roles/litellm/ — Python venv, pip install litellm[proxy], systemd
- vars/main.yml — model map (haiku/sonnet/opus) + LITELLM_MASTER_KEY env lookup
- site.yml — adds litellm role + sanity-check assert
- deploy.sh — pulls LITELLM_MASTER_KEY from Infisical (/meridian/) on the
controller and exports it for the playbook
- New Infisical secret /meridian/vault_litellm_master_key
Smoke: Pulse → LiteLLM /v1/chat/completions → Meridian /v1/messages → Max sub
returns "pong" through both the LiteLLM master key auth and the Claude Code
SDK OAuth.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Deploys @rynfar/meridian on a Debian 12 LXC, bound to 0.0.0.0:3456.
OAuth credentials transferred manually after first deploy (claude login on
Mac, scp ~/.claude to /opt/meridian/.claude). systemd unit is enabled but
gated on credentials.json existence so the first deploy doesn't crash-loop.
LXC has no auth layer — security model is LAN-only reachability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Your Name
cbalders