Issues

• E4 (#651): full Gemma profile — all four changes

  algorithmbenchmark-full-skillsbenchpaper-experiment
  skillberry-ai/cap-evolve#656 · Filter by author eranra opened on Oct 5, 2026
  2

• E3 (#651): short instructions — short_instructions

  algorithmpaper-experiment
  skillberry-ai/cap-evolve#655 · Filter by author eranra opened on Oct 5, 2026
  2

• E2 (#651): digested trajectories — digest_trajectories

  algorithmobservabilitypaper-experiment
  skillberry-ai/cap-evolve#654 · Filter by author eranra opened on Oct 5, 2026
  2

• E1 (#651): large-output guard — output_guard + no_subagents

  algorithmpaper-experiment
  skillberry-ai/cap-evolve#653 · Filter by author eranra opened on Oct 5, 2026
  2

• E0 (#651): control replicate — Gemma optimizer, no profile

  algorithmbenchmark-full-skillsbenchpaper-experiment
  skillberry-ai/cap-evolve#652 · Filter by author eranra opened on Oct 5, 2026
  2

• Hypothesis: a Gemma-specific optimizer profile makes Gemma-4-31B-It a better optimizer of its own skill

  algorithmenhancementpaper-experiment
  skillberry-ai/cap-evolve#651 · Filter by author eranra opened on Oct 5, 2026
  2

• paper-experiment: ablate null-control + growth (statistical rigor) and run agent-optimize vs. this repo's own GEPA/SkillOpt baselines on the same code-editing capability

  algorithmblocked-by-dependencydocumentationenhancementhonestypaper-experimentpriority-p2
  skillberry-ai/cap-evolve#640 · Filter by author OsherElhadad opened on Oct 2, 2026
  1

• agent-optimize: report optimizer self-compliance rate (screening/merge discipline) as a measured metric, not just an internal hard-block

  algorithmblocked-by-dependencydashboardenhancementhonestyobservabilitypaper-experimentpriority-p1
  skillberry-ai/cap-evolve#639 · Filter by author OsherElhadad opened on Oct 2, 2026
  1

• agent-optimize: adopt GEPA's per-instance Pareto-frontier tracking in branches.json instead of a weaker cluster-tag-only scheme

  algorithmblocked-by-dependencyenhancementhonestypaper-experimentpriority-p1
  skillberry-ai/cap-evolve#638 · Filter by author OsherElhadad opened on Oct 2, 2026
  1

• Experiment: a harder val so the gate can accept later rounds (follow-up to #606) — SpreadsheetBench, Gemma-4-31B-It, Opus 5.5

  algorithmpaper-experiment
  skillberry-ai/cap-evolve#622 · Filter by author eranra opened on Oct 1, 2026
  6

• Experiment: how the optimizer's reader profile (frontier vs Gemma-4-31B-It vs hidden) changes optimization — SpreadsheetBench, Claude Opus 5.5

  algorithmenhancementpaper-experimentpriority-p2
  skillberry-ai/cap-evolve#606 · Filter by author eranra opened on Sep 30, 2026
  25

• agent-optimize: search is a single linear lineage — no parallel branch exploration, no true graph/merge of independent candidates

  algorithmenhancementpriority-p1
  skillberry-ai/cap-evolve#586 · Filter by author OsherElhadad opened on Sep 29, 2026
  3

• Dashboard: run timeline + visual optimizer diagnosis view (POC attached)

  dashboardenhancementfix-with-bobjavascriptobservabilitypriority-p1
  skillberry-ai/cap-evolve#575 · Filter by author eranra opened on Sep 29, 2026
  18

• [Tau2] The blackbox arm requires customized Tau2 rather than using the same codebase for both direct and blackbox

  skillberry-ai/cap-evolve#565 · Filter by author aviweit opened on Sep 28, 2026
  1

  

• system-prompt on tau2 edits the domain POLICY, not the agent's system prompt

  skillberry-ai/cap-evolve#563 · Filter by author aviweit opened on Sep 28, 2026
  1

  

• [Praxis] Support Praxis as the blackbox proxy-agent

  skillberry-ai/cap-evolve#554 · Filter by author aviweit opened on Sep 27, 2026
  1

  

• SpreadsheetBench full_verified: first held-out result (76.4 vs WikiSkill 68.0) — what is still needed for a fair comparison

  documentationenhancement
  skillberry-ai/cap-evolve#538 · Filter by author eranra opened on Sep 26, 2026
  22

  

• Two packages named tau2 force a per-benchmark corner case in ci_setup.sh

  skillberry-ai/cap-evolve#514 · Filter by author aviweit opened on Sep 23, 2026
  1

  

• Benchmark sources are unpinned: results cannot be attributed to a benchmark version

  skillberry-ai/cap-evolve#513 · Filter by author aviweit opened on Sep 23, 2026
  1

  

• Add support for concurrent blackbox optimizations.

  skillberry-ai/cap-evolve#509 · Filter by author aviweit opened on Sep 23, 2026
  1

  

• cap-evolve run to be supported for tau2 airline spa arm (only run.sh supported)

  skillberry-ai/cap-evolve#461 · Filter by author aviweit opened on Sep 8, 2026
  1

  

• Make the SPA proxy port configurable instead of fixed at 7000

  skillberry-ai/cap-evolve#460 · Filter by author aviweit opened on Sep 8, 2026
  1

  

• target_model silently drifts out of sync with the model actually evaluated — derive it from the adapter

  skillberry-ai/cap-evolve#453 · Filter by author aviweit opened on Sep 5, 2026
  2

  

• agent-optimize Phase 2: TDD-style micro-tests (Harbor/Terminal-Bench-shaped)

  skillberry-ai/cap-evolve#436 · Filter by author OsherElhadad opened on Sep 3, 2026
  1

• Redesign agent-optimize: parallel graph search over candidate merges + adaptive task subsetting + TDD-style micro-tests

  skillberry-ai/cap-evolve#434 · Filter by author OsherElhadad opened on Sep 3, 2026
  3