CC fireworks-inference
Run inference on Fireworks AI — direct, through the Leanmcp AI Gateway proxy (aigateway.leanmcp.com) for observability, or via LiteLLM / the OpenAI SDK / raw HTTP. Covers the endpoint-and-key matrix for each route, serverless vs dedicated deployments and their cold starts, browsing and filtering the model catalogue, reasoning models that return content:null with the answer in reasoning_content, function calling, concurrency and load testing, and runner scripts where many models share one protocol with a --use-proxy toggle. Use this skill whenever the user mentions Fireworks, FIREWORKS_API_KEY, api.fireworks.ai, fireworks_ai/ model ids, accounts/fireworks/models/..., a Fireworks deployment id, the Leanmcp gateway or LEANMCP_API_KEY, app.leanmcp.com observability, or gpt-oss / kimi / glm / qwen / deepseek served on Fireworks; wants to route LLM calls through a proxy for logging or cost tracking; is debugging empty responses, 404s, scope errors, cold starts, or "the gateway isn't working"; wants to pick a Fireworks model with tool calling; or wants to sweep several models over one benchmark. Reach for it even on vague asks like "why is this model returning nothing" or "set up inference for these models" when Fireworks is in play.
Promises to cover the full Fireworks AI stack: routing options (direct, proxy, LiteLLM), model catalogue filtering, reasoning models with content:null quirks, load testing. The skill contains 14 files describing endpoint-key matrices, cold starts for serverless vs dedicated, and runner scripts. Graded C with quality 68, process 51, safety 93. Three medium-low findings; linter complains about description length. No model runs. In practice, you get a reference on routing and scripts, but without real-world error handling or execution examples. Install if you already have a Fireworks account and need an API cheat sheet—reference material, not a solution.
Run inference on Fireworks AI — direct, through the Leanmcp AI Gateway proxy (aigateway.leanmcp.com) for observability, or via LiteLLM / the OpenAI SDK / raw…
As a process C 51/100 · Has gaps — weak spots: result and completion, inputs and preconditions, failures and branches
What is at stake
Medium-severity findings: the skill is probably honest, but read what alarmed the scanner.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The instructions or scripts send data out: environment variables, keys, file contents, chat history. You may never notice, because the agent performs the upload as "part of the task".
If the upload is not needed for the task, remove it: catalogs flag such skills and delist them. If it is needed, name the destination explicitly, say what leaves the machine, and give the user a switch.
How to improve
- Shorten the description to 1024 characters.
- Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
- A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.
Guard findings · 3
✓ No critical or high findings
Medium and low: 3
-
medium Exfiltration
net-credential-usereferences/models.md:124Credential used in a network call (verify the destination is the intended service)curl -s -H "Authorization: Bearer $FIREWORKS_API_KEY" \
-
low Secrets in code
secret-password-literalassets/run_fireworks.sh:69Hard-coded password / key literal (may be an example) (placeholder value)MASKED="${API_KEY:0:6}…${API_KEY: -4}"placeholder -
low Secrets in code
secret-password-literalreferences/runner-scripts.md:50Hard-coded password / key literal (may be an example) (placeholder value)`${API_KEY:0:6}…${API_KEY: -4}`. Note that litellm's debug logging prints fullplaceholder
Files scanned: 14. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
- error
description-longdescription is 1246 chars, limit 1024
Process rating: all ten parameters 51/100
- 0Result and completion. Does not say what the result is
- 0Inputs and preconditions. Does not say what the process needs to start
- 0Failures and branches. Linear process with no failure handling
- 30Running it twice. 9 mutating operations with no state check
- 60Tools and files. Uses tools (bash, python) that frontmatter does not declare
- 70When it triggers. States when to use, but not when not to
- 100Steps. 5 steps
- 100Consistency. Name and required fields are in place
- 100Execution cost. Instruction body is 1640 tokens
- 100Progress reporting. Reports progress
Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.
Quality signals
- +4Description does not say when NOT to use the skill (false activations)
- +3Description length 1246: 120–800 characters recommended
- +3Output format is not stated: the model decides each time
- +1No license
- +2Single-language instructions
- +5Description quotes 3 example trigger phrases
- +4Structure: 8 headings
- +3Step-by-step instructions: 5 items
- +4Has examples (4 code blocks)
- +4Reference files are cited in the instructions (8 of 8)
- +3All 4 scripts are documented
Quality base 70; lint remarks subtract, signals add up to 100. Result: 68.