Skill llm gemini
Skill joeylu/runtime-free-llm-integration-skills/LLM/skill-llm-gemini
Runtime-free skills for coding agents to build fail-fast LLM provider integrations across OpenAI, Aliyun Bailian / DashScope, DeepSeek, and Gemini.
npx -y skills add joeylu/runtime-free-llm-integration-skills --skill skill-llm-geminiAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Runtime contract for direct Gemini Developer API integration. Use it to resolve exact Gemini model IDs, Interactions or GenerateContent URLs, thinking controls, tools, structured output, multimodal input, image generation/editing, state, streaming, caching, response mapping, pricing references, and logging. Model choice remains with the user or host project.
SKILL.md
4.8 KB, as published. Nobody here has run it
Skill LLM Gemini
Purpose
Define how an agent integrates a Gemini model already named by the user or host project. This skill does not make model recommendations and does not cover Vertex AI.
Read Order
Read shared contracts first:
../_shared/route-key-schema.md../_shared/model-catalog-schema.md../_shared/pricing-matrix-schema.md../_shared/capability-matrix-schema.md../_shared/connection-profile-schema.md../_shared/request-url-matrix-schema.md../_shared/request-envelope.md../_shared/response-envelope.md../_shared/error-contract.md../_shared/progress-contract.md../_shared/ui-binding.md../_shared/sync-policy.md../_shared/logging-fields.md
Then read:
references/connection-profiles.mdreferences/request-urls.mdreferences/model-catalog.mdreferences/pricing-matrix.mdreferences/capability-matrix.md- the transport for the requested kind
references/logging-contract.mdwhen logging is implementedreferences/model-sync.mdonly for an explicit update or verification task
Read references/hosted-tools.md when provider-hosted tools are requested.
Runtime Rules
- Send the exact documented model ID from
model-catalog.md. - Resolve connection profile, request kind, API surface, API version, endpoint kind, full route key, base URL, and request URL before capability lookup.
- Match capabilities by the exact full
RouteKeyrow, including API version and endpoint kind. - Interactions
v1, Interactionsv1beta,generate-content, andstream-generate-contentare separate surfaces. Do not migrate or fall back between them silently. - Use stable Interactions
v1when the project requests Interactions without an explicit beta requirement. Keepv1betaexplicit. - Reject model IDs documented as shut down. Do not rewrite them to another model.
- For Gemini 3.x text rows in this skill, caller overrides for
temperature,top_p, andtop_kare blocked where the capability matrix records provider-default-only behavior. - Map verified thinking levels through
ReasoningEffort; do not combine them with legacythinking_budget. - Preserve thought signatures and function-call identity exactly when replay is required. Do not display signatures or raw model thoughts.
- Keep caller-defined
Toolsseparate fromHostedTools. - Interactions custom safety settings and explicit caching are unsupported in the documented rows. GenerateContent explicit caching requires an existing cache resource and its exact transport mapping.
- A continued Interaction must resend tools, system instruction, and generation configuration when needed; the continuation ID does not preserve them.
- When
store=false, do not treat the returned Interaction ID as reusable state. - Reject requested fields marked
unsupportedorunknown.
Request Flow
- Read the requested model, request kind, and API surface.
- Resolve the connection profile and exact request URL row.
- Resolve the exact catalog and capability rows.
- Validate thinking, tools, structured output, media input, cache, state, and stream fields.
- Apply the matching transport and hosted-tool rules.
- Normalize the response, errors, progress, and logs through the shared contracts.
Gemini Field Mapping
- Interactions
ContinuationIdmaps toprevious_interaction_id;StoreResponsemaps tostore; streaming usesstream: trueon the same endpoint. - GenerateContent multi-turn state replays the complete content history, including required thought signatures and function-call identity.
- Thinking-level values and defaults are model-specific and come from the capability matrix.
- Structured-output payloads differ between Interactions and GenerateContent.
- Interactions reasoning summaries map through
generation_config.thinking_summariesonly where verified. visionmeans image understanding through typed media parts. Audio, video, and PDF require their own typed host inputs and must not be disguised asInputs.Images.imagingmeans image generation or editing. Reference images are imaging inputs;Inputs.ImageCountis an output-count field and requires separate verification.
Out of Scope
- Vertex AI endpoints, service accounts, regions, or provisioned throughput
- Gemini consumer application settings
- automatic model discovery during ordinary implementation
- raw chain-of-thought display
- model ranking or permission policy