mirror of https://github.com/magnus919/agent-skills.git synced 2026-09-11 19:47:12 +03:00

Files

T

Magnus Hedemarkandfactory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com> 030e6dfd47 feat(skill): add LiteLLM AI gateway operations skill

Add litellm/, an operational tool skill for the LiteLLM AI gateway (proxy)
and Python SDK, in the same vein as the vllm and llama-cpp engine skills.

Contents:
- SKILL.md: operating contract, operating loop, verification boundaries,
  and hard boundaries; concise core sections routing depth to references
- README.md: human-facing install/use guide with required sections
- references/: nine dated, source-indexed references (source index,
  quickstart + SDK, config & routing, keys/teams/budgets/spend, caching &
  guardrails, observability & logging, deployment, security & public
  hosting, troubleshooting), researched against litellm 1.97.0
  (2026-08-22) including a live proxy probe of the health endpoints
- scripts/litellm-health: read-only GET-only probe (liveliness, readiness,
  /v1/models, /model/info); stdlib-only Python 3.9+, --json, --help
  without a server
- tests/test_litellm_health.py: 18 deterministic tests against a local
  stub HTTP server, including the observed-traffic GET-only contract
- templates/proxy-config-record.md and proxy-deployment.md: fillable
  records; the config record is the rollback unit
- evals/evals.json: schema_version 1, six output-quality cases

Also regenerates tracked catalog artifacts (.claude-plugin/marketplace.json,
.codex-plugin/plugin.json, llms.txt) and adds the root README catalog entry
plus the skill-triggers.md index row.

AI assistance: authored with AI assistance (Factory Droid) under human
direction; facts verified against litellm 1.97.0 and official docs dated
2026-08-22.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>

2026-08-22 15:31:16 -04:00

3.0 KiB

Raw Blame History

LiteLLM Proxy Configuration Record

Fill this record before changing a proxy configuration. It is the rollback unit: the previous record plus the previous pinned image is the rollback path.

Deployment identity

Requested outcome: [fill: what this gateway must do and for whom]
Deployment type: [fill: bare litellm / Docker / Compose / Helm / raw manifests]
Target and scope confirmed with: [fill: who confirmed, when]
Rollback path: [fill: previous image tag + previous config record]

Pinned artifacts

LiteLLM version or image tag/digest: [fill: ghcr.io/berriai/litellm:vX.Y.Z or pip litellm==...]
Cosign verification status: [fill: verified against pinned key commit / not verified]
fastapi pin (pip installs): [fill: e.g. fastapi==0.136.3 for 1.97.0, or n/a]
Config file source and path: [fill: repo path or S3/GCS bucket reference]
Config-in-DB state: [fill: store_model_in_db true/false; if true, note DB overlay wins]
Enterprise license in use: [fill: yes/no]

Model list summary

model_name	litellm_params.model	weight/rpm/tpm	order	notes
[fill]	[fill: provider/prefixed string]	[fill]	[fill]	[fill: base_model, access_groups, ...]

Routing and reliability

routing_strategy: [fill: simple-shuffle default unless deliberately changed]
num_retries (settings/deployment/request): [fill]
fallbacks / context_window_fallbacks / content_policy_fallbacks: [fill: alias targets only]
cooldown settings: [fill: allowed_fails, cooldown_time, policy overrides]
enable_pre_call_checks / optional_pre_call_checks: [fill: on/off and why]

Keys, budgets, limits

master_key source: [fill: secret manager reference — never the value]
salt_key set (never rotated after models added): [fill: yes/no]
global budget / duration: [fill]
team/key budget scheme summary: [fill: scopes and caps]
rate limits (tpm/rpm per scope; admin exemption acknowledged): [fill]

Caching, guardrails, observability

cache backend and ttl: [fill: redis/none; semantic caching excluded for agents?]
guardrails configured (names, modes): [fill]
callbacks wired: [fill: langfuse/otel/prometheus/...]
privacy posture: [fill: turn_off_message_logging, redact flags, store_prompts_in_spend_logs off?]

Data stores

Postgres endpoint and pool math: [fill: MAX_DB_CONNECTIONS / (instances x workers)]
Redis version and endpoints: [fill: >=7.0 required when >1 instance]
Backup schedule for Postgres: [fill]

Verification checklist

/health/liveliness returns 200 (litellm-health --check health)
/health/readiness returns 200 (DB reachable)
/v1/models lists expected aliases for a representative key
A representative chat request returns tokens; x-litellm-model-id matches intent
Budget enforcement proven once with a tiny test budget
Fallback path proven once by forcing a deployment failure

Changes from the previous record

[fill: what changed, why, which verification backs it]

3.0 KiB Raw Blame History