Local LLM Skill (AGILAB)

Use this skill when working on local/offline engines or prompts.

Correctness-First Defaults

• Prefer deterministic settings for code edits (lower temperature, explicit constraints).
• In WORKFLOW, prefer validated generated-action JSON contracts over raw Python

snippets for normal dataframe transformations. This keeps AGILAB lightweight and power-efficient without requiring containers or VMs for the default path.

• Require the model to return:

- file list to edit - exact patch intent - tests/commands to validate

Generated Workflow Actions

• Treat raw model Python as an advanced/manual mode, not the default UX.
• For dataframe transformations, ask the model for a versioned action contract,

validate it against the loaded dataframe schema, then let AGILAB convert the approved contract into deterministic pandas code.

• If the request cannot be represented by the safe action registry, fail closed

with an actionable message instead of executing or repairing arbitrary code.

• Keep container/VM/process sandbox guidance for explicitly untrusted apps or

advanced raw-Python execution. Do not make it the primary recommendation for AGILAB's normal local, energy-efficient workflow.

No Silent Fallbacks

• Detect missing local endpoints/models up-front and surface an actionable error.
• Do not auto-switch APIs or rewrite parameters silently.
• Probe the requested backend before claiming validation:

- GPT-OSS Responses-compatible endpoint, usually http://127.0.0.1:8000/v1/responses - Ollama tags endpoint, usually http://127.0.0.1:11434/api/tags

• If the endpoint is unavailable or the selected model is absent, report that

condition directly. Do not silently replace it with a fixture, cloud API, or different local model.

Standalone Engine Validation

• For local contract tests, a temporary Responses-compatible HTTP server is

acceptable when the goal is to prove AGILAB's request/response path, schema validation, and artifact generation without downloading a model.

• Label that result as a local contract engine, not as real model validation.

Real model validation requires the actual requested engine to be running and selected.

• Validate the generated payload against the app schema and downstream artifacts,

not only the raw model response. For educational diagnostics, assert fields such as student_score are persisted in CSV/JSON summaries when the UI or README claims them.

• Keep generated examples deterministic where possible by constraining prompts,

seeds, and output schemas. Reject invalid JSON with an actionable error rather than repairing it silently.

Ollama Notes

• Let users select:

- model name - temperature/topp/topk - max tokens - seed (if supported) - “auto-run/auto-fix” loop guardrails (max iterations, stop on failure)