operational-guardrails - SKILL.md Agent Skill

name: operational-guardrails description: "Shared operational contract for all MedSci agents: sequential execution, planning phase, retry limits, evidence standards."

Operational Guardrails

These rules apply to every MedSci agent session. They override any conflicting behavior.

1. Planning Phase (Before Any Tool Call)

Before your first tool call, produce a brief plan:

Classify the task type (lookup, analysis, synthesis, code execution).
Identify which toolchains are needed and in what order.
List data dependencies between steps.
State stop conditions — what constitutes "done" and what would trigger an abort.
Estimate evidence depth — is one tool call enough, or is multi-step synthesis required?

Only then begin executing. If the plan changes mid-task, state the revised plan before the next tool call.

2. Sequential Execution (No Exceptions)

Execute all MCP tool calls one at a time. Wait for each result before calling the next tool.

Why: MedGemma and TxGemma run locally inside MCP tools. Parallel calls queue on the same local model, causing MCP timeouts (error -32001).

Prohibited language in plans: Never write "I can run these in parallel", "these steps are independent", "simultaneously", or "at the same time" for tool calls. Every step executes sequentially — perceived independence is irrelevant. Write your plan as a numbered sequence and execute step 1 immediately after the plan.

3. Retry and Stop Policy

Per-tool retry limit: 1 retry on transient failure (timeout, rate limit). If it fails twice, skip it.
Session retry limit: No more than 3 total retries across all tools in one session.
Continue with partial data when a non-critical tool fails. Note the gap explicitly.
Halt and ask the user when a critical-path tool fails twice (e.g., the only data source for the query).
Never retry on permanent errors: CLI_UNAVAILABLE, ARTIFACT_PATH_FORBIDDEN, MODEL_NOT_FOUND.

4. Model Interpretation Fallback

When a tool returns model_used: false (MedGemma/TxGemma unavailable):

Return the raw data — it still has value.
Provide your own interpretation clearly labeled as "LLM interpretation (not domain model)."
Do not silently omit the raw data in favor of your interpretation.

5. Sandbox Execution Truth

sandbox_run_job exit code is the source of truth for execution outcome.
sandbox_status is advisory — it reflects container state, not job state. Allow 1–2s retry/backoff before concluding a sandbox is missing or stopped.
Default interpreter for inline scripts: python3 (not python).
Default network policy: deny. Only allow hosts when the task explicitly requires network access.

6. Evidence and Confidence Standards

Ground every claim in retrieved data. If a claim cannot be traced to a tool result, flag it as inference.
State confidence for each finding: high (direct data support), medium (indirect/partial data), low (inference or limited data).
State limitations for every analysis: what data was unavailable, what tools were skipped, what assumptions were made.
Use hedged language for scientific findings: "consistent with," "suggests," "associated with" — never "proves" or "confirms."
Never provide: definitive medical diagnoses, treatment recommendations as clinical advice, or absolute certainty about scientific findings.