deterministic-agent-tools
Use this skill when designing mine-rs deterministic tools, JSON schemas, VFS artifacts, task tools, subagents, verifier agents, LangGraph-style orchestration, agentic workflows, or AI-native automatio
SKILL.md.github/skills/deterministic-agent-tools/SKILL.md
GitHubDeterministic agent tools skill
Use this skill when a task involves tools for agents, structured contracts, VFS artifacts, task orchestration, subagents or verifier logic.
Source of truth
Read these files first:
docs/references/agentic-layer.mddocs/references/repository-strategy.mddocs/references/domain-capabilities.mdAGENTS.md
Central rule
Agents reason and orchestrate.
Deterministic tools calculate and validate.
Agents must not invent mining calculations. They must call SDK-backed tools and explain results from structured outputs.
Ownership
mine-toolsowns deterministic tool contracts and SDK-backed execution.python/mine-agentsowns Python-first orchestration, prompts, task tools, subagents and graph runtime.- Optional Rust crates may own deterministic VFS or verifier components if needed.
- The SDK must not depend on the agentic layer.
Tool contract checklist
Every tool should define:
- Name.
- Purpose.
- Input schema.
- Output schema.
- Error schema.
- Assumptions.
- Required model/data references.
- Produced artifacts.
- Version.
Initial tools
Prioritize:
inspect_model.validate_model.query_blocks.aggregate_blocks.grade_tonnage.create_scenario.evaluate_scenario.compare_scenarios.
Workflow procedure
- Identify the user objective.
- Translate the objective into deterministic tool calls.
- Validate model/data before downstream calculations when quality matters.
- Store intermediate outputs as VFS artifacts.
- Run verifier checks before summarizing.
- Produce natural-language explanations only from structured outputs.
- If a required SDK capability does not exist, report the limitation instead of simulating it.
VFS artifact guidance
Artifacts should be named and typed:
- Model profiles.
- Validation reports.
- Grade-tonnage curves.
- Aggregation tables.
- Scenario definitions.
- Scenario evaluations.
- Comparison reports.
- Decision logs.
Gotchas
- Do not put LangGraph/LangChain dependencies in
mine-sdkorpython/miners. - Do not let tool schemas drift from SDK behavior.
- Do not make validation optional in workflows where invalid inputs could invalidate results.
- Do not summarize unsupported claims.