agentv-bench: automate keep/discard decision in iteration loop

[christso]

Objective

Automate the keep/discard decision inside the existing agentv-bench Step 5 iteration loop so obvious improvements do not require a human pause.

Design Latitude

Scope: Skill-only change — no CLI, schema, or core code changes required.

After each iteration in the bench skill's optimization loop (SKILL.md Step 5), automatically:

1. Run agentv compare baseline.jsonl candidate.jsonl --json
2. Parse the structured output: { summary: { wins, losses, ties, meanDelta } }
3. Apply keep/discard rules:
   • wins > losses → keep change, promote to new baseline
   • wins <= losses → discard change, revert, try different mutation
   • meanDelta == 0 but simpler prompt → keep (simplicity criterion)
4. Log the decision and rationale before proceeding to next iteration

Why this stays narrow

• This is the smallest useful improvement to the current bench loop.
• It should remain compatible with human checkpoints at iterations 3, 6, 9.
• It should remain complementary to #748 rather than expanding into full unattended autoresearch.

Acceptance Signals

• Clear-cut iterations no longer require manual keep/discard judgment
• Human checkpoints still fire at the existing intervals
• Decision logic uses existing agentv compare --json output only
• No new CLI flags, config fields, persistence layer, or runtime memory features are introduced

Non-Goals

• Not full autoresearch / overnight unattended mutation loops
• Not mutator generation logic (#746)
• Not eval bootstrapping (#747)
• Not core iteration metadata work (#335)
• Not persistent session search, personal memory, or self-improving runtime features

Related

• feat(bench): autoresearch mode — unattended eval-improve loop with hill-climbing ratchet #748 — full autoresearch mode
• feat(bench): mutator subagent — autonomous artifact rewriting from failure analysis #746 — mutator subagent
• feat(eval-writer): eval-generator subagent — bootstrap EVAL.yaml from existing artifacts #747 — eval-generator subagent
• feat(eval): iteration tracking, termination taxonomy, and cross-run regression detection #335 — iteration metadata / termination taxonomy
• tracking: optimization loop roadmap — bench automation, autoresearch, iteration state #1003 — tracking issue for optimization-loop roadmap coordination

[christso]

AutoAgent research update — validates approach + simplicity criterion detail

Updated AutoAgent research confirms this issue's approach is correct. Additional detail on the simplicity criterion:

AutoAgent's full decision rules:

1. passed count increases → keep
2. passed unchanged but code is simpler → keep (simplicity criterion)
3. Otherwise → discard (git revert)

The simplicity criterion maps directly to the meanDelta == 0 case in this issue's design. AutoAgent's benchmark results (96.5% SpreadsheetBench, 55.1% TerminalBench — both #1) validate that automated keep/discard decisions produce better outcomes than manual judgment.

No scope change needed — this issue is correctly scoped as the smallest useful win.