Diagnose — why did this run learn nothing?

show gives the evidence: worst rollouts (samples), refused candidates (rejected) and the outcome histogram. A user still has to decide whether the agent, the reflector, the gate or the scorer is the problem. diagnose makes that decision explicit and names the next step.

agentdescent diagnose <run_id> [--json]

The verdicts are deliberately few and ordered so the strongest evidence wins:

Verdict	Meaning	First thing to try
solved	the target was reached	review and apply
partial-improvement	candidates committed, reward still below target	more rounds/workers
running	the run has not finished	status / watch, then diagnose again
cancelled	the run was cancelled	resume if the answer is still wanted
run-failed	the run reported an error	log.txt / status.json; include propose_error when present
reflector-error	reflector calls failed and nothing committed	model/endpoint/key (doctor)
scoring-mismatch	untruncated samples' output equals gold but reward is low	inspect the scorer
gate-rejected	candidates were proposed and refused, nothing committed	read rejected diffs and p_improve
agent-failures	most listed samples answer wrongly	fix the skill/prompt/data
no-improvement	partial reward, no candidates and no gate rejections	raise rounds/workers
unknown	not enough evidence	show --json

• show lists the worst rollouts, so "every sample failed" is not enough to call the agent broken. agent-failures requires most listed samples to fail; rejected candidates and committed-but-partial runs are classified first.
• scoring-mismatch only compares samples whose output/gold were not truncated.
• A run that committed a candidate is reported as partial-improvement, never as "no candidate was adopted".
• Running and cancelled runs are named instead of being diagnosed as finished.
• diagnose_payload calls show_payload(..., diff=False), so it does not build an artifact diff it would throw away.

from agentdescent import diagnose_payload
report = diagnose_payload(run_id)
print(report.verdict, report.summary, report.recommendations)

The same read model is exposed as the MCP diagnose tool and the CLI verb, so the answer is bounded and consistent. It does not re-run the agent or reselect a model.