medsci-agent: Command for Claude Code

Use ACE MCP to perform controlled post-run learning for this completed task.

Task context (JSON-encoded): $ARGUMENTS

Parse $ARGUMENTS as a JSON object. It contains exactly three fields:

• question — the user's task description
• answer — the medsci agent's final response
• telemetry — execution telemetry (tool_calls, failures, last_tool, etc.)

Example: {"question":"Find KRAS inhibitors","answer":"Found 3 hits...","telemetry":"tool_calls=8 | failures=0 | ..."}

Extract these three values before calling any tools.

Workflow (strictly sequential — wait for each result before the next step):

Step 1 — Load prior skillbook Call ace.skillbook.load with:

• session_id: "medsci:multidomain"
• path: ".opencode/ace/skillbooks/medsci_multidomain.json"

If the file does not exist yet, the load will fail — that is expected on first run. Continue to Step 2.

Step 2 — Assess evidence quality Before calling ace.learn.feedback, assess:

• Is there a real question and answer (not the "(unknown)" placeholder)?
• Were at least 2 domain tool calls made (from telemetry)?
• Is the outcome meaningful (partial success, failure, or clear success)?

If evidence is too weak (e.g. question is placeholder AND no failures), skip Steps 3–4 and explain why.

Step 3 — Learn from feedback Call ace.learn.feedback with:

• session_id: "medsci:multidomain"
• question: <extracted question — the user's actual scientific task>
• answer: <extracted answer — what the medsci agent actually did and concluded>
• context: "MedSci multi-domain scientific research agent: drug discovery, protein structure, literature synthesis, omics analysis"
• feedback: <telemetry field — failures, tool outcomes, what succeeded or failed>

The Reflector will analyze what worked and what failed. The SkillManager will generate atomic imperative strategy updates (ADD/UPDATE/TAG/REMOVE). Skills must be concrete and actionable — e.g. "Use search_type='molecule' not 'target' for ChEMBL inhibitor queries."

Step 4 — Persist Call ace.skillbook.save with:

• session_id: "medsci:multidomain"
• path: ".opencode/ace/skillbooks/medsci_multidomain.json"

Step 5 — Report delta Call ace.skillbook.get with session_id: "medsci:multidomain", limit: 5. Report:

• Skills before / after counts
• Top new or updated skills
• Any stale/harmful skills flagged for future removal

Rules:

• Do not invent outcomes or feedback.
• Skip Steps 3–4 if evidence quality is too weak; explain why.
• Keep feedback specific, testable, and tied to observed tool behavior.
• Strategies must be imperative commands under 15 words (e.g. "Retry ChEMBL with synonym if target search returns no compounds").