ollama

Ollama Local Inference Skill

Use this skill when FORGE needs local LLM inference without external API calls. Ollama runs on the local GPU server (RTX 3090) and provides zero-cost text generation, classification, and analysis.

Default operating pattern:

1. Select the model based on task complexity (fast, balanced, or high).
2. Pipe the prompt to ollama run via stdin for multi-line prompts.
3. Capture stdout as the complete response.
4. Parse or post-process the text output as needed.

Ollama is not a coding agent. It does not edit files, manage sessions, or run tools. Use it for inference tasks: classification, summarisation, text generation, and structured analysis.

Model Selection

Choose the model based on the task:

Routing tier	Model	Use case
fast	gpt-oss:latest	Short answers, quick classification, low-latency checks
balanced	qwen3-coder:latest	Code analysis, structured classification, moderate reasoning
high	qwen3.5:27b	Complex reasoning, long-form analysis, nuanced generation

When the caller does not specify a model, default to qwen3-coder:latest.

Capabilities

generate(prompt, model)

General-purpose text generation.

Preferred command:

printf '%s\n' "$PROMPT" | ollama run "$MODEL"

For short single-line prompts:

ollama run "$MODEL" "$PROMPT"

Expected behavior:

• stdout contains the full generated text
• no side effects (no file writes, no network calls)
• exit code 0 on success

classify(text, categories, model)

Classify input text into one of the provided categories.

Preferred command:

printf 'Classify the following text into exactly one of these categories: %s\n\nText: %s\n\nRespond with only the category name.' "$CATEGORIES" "$TEXT" | ollama run "$MODEL"

Expected behavior:

• output is a single category name from the provided list
• use qwen3-coder:latest or gpt-oss:latest for fast classification
• useful for urgency detection, intent classification, and triage

analyze(prompt, model)

Longer-form analysis such as code review summaries, log analysis, or document review.

Preferred command:

printf '%s\n' "$PROMPT" | ollama run "$MODEL"

Expected behavior:

• output is structured prose (findings, recommendations, risks)
• prefer qwen3.5:27b for complex analysis tasks
• prefer qwen3-coder:latest for code-specific analysis

Session Adapter Mapping

FORGE session field	Ollama CLI
prompt	stdin pipe or positional arg
agent = "ollama"	ollama run
model	model name arg (e.g., qwen3-coder:latest)
tools	N/A
workdir	process working directory
timeout	managed by FORGE wrapper
budget	N/A (local inference, no cost)
resume	N/A
output_mode	N/A (always plain text)

AgentResult Mapping

Ollama output	AgentResult field
stdout (full text)	output
exit code 0	success (no error)
non-zero exit code	error
N/A	session_id (not supported)
N/A	tokens_in (not exposed by CLI)
N/A	tokens_out (not exposed by CLI)
0.00	cost_usd (local, always zero)

Limitation: ollama run does not expose token usage or timing metadata in its CLI output. If token tracking is needed, use the OpenAI-compatible API endpoint (http://<host>:11434/v1) instead of the CLI.

Safety Rules

• Ollama is inference-only. Never expect it to edit files or execute commands.
• Always validate that the target model is pulled before invoking (ollama list).
• Prefer stdin piping for prompts containing quotes, newlines, or special characters.
• Set FORGE-level timeouts to guard against slow generation on large models.
• Do not send sensitive data through Ollama unless the server is on a trusted network.
• Cap confidence at 0.80 for Ollama outputs (local models are less capable than frontier APIs).