skillberry-ai
- 27 skills
- 0 followers
- 11 hours ago last updated
- ▌
- ▌ Primitive Skill · skillberry-ai bundleThis skill contains the basic set of primitive tools provided by Tau2.
- ▌ Skill Name · skillberry-ai bundle<One paragraph. WHAT this skill does and WHEN an agent should reach for it. This is the host's activation signal, so be concrete and self-contained — an agent decides whether to load the skill from this text alone.>
- ▌ Wiki · skillberry-ai bundleCross-iteration memory as a weakness graph (weakness nodes + solution cards), extracted from the deprecated `evograph` algorithm's run-dir format per its own maintainer note. Select via `memory_skill: wiki` in capevolve.yaml (default is `md-files`). Copied into every iteration's working dir as `./guidance/memory-wiki/` when selected; read this before writing anything under the run dir's `wiki/`.
- ▌ Gate · skillberry-ai bundleApply the acceptance decision that keeps optimization honest — always on the val split, by default requiring the improvement to exceed the significance bar (Δ > k·SE) so noise is not mistaken for progress. Use to inspect or reproduce a single accept/reject decision; the algorithms apply it internally every iteration.
- ▌
- ▌ Report · skillberry-ai bundleSummarize a run for a human — baseline val → best val → sealed test, the winning candidate, iterations spent, and pass^k. Use after finalize. Writes report.md and prints a compact JSON summary; the source of truth for "did this optimization actually work, and by how much".
- ▌ Gepa · skillberry-ai bundleRuns the GEPA optimization loop (arXiv:2507.19457) — sample-efficient reflective Pareto search. A cheap train-minibatch pre-gate decides whether a proposal is worth an expensive val evaluation, and parents are sampled from a per-instance frontier so specialists survive instead of being averaged away. Use when rollouts are expensive and the scorer returns informative per-task feedback, and you want the most quality per evaluation. Use hill-climb instead for a first baseline run or for feedback-poor binary pass/fail tasks.
- ▌ Baseline · skillberry-ai bundleEstablish the starting point. Use after implement-and-check and before any algorithm. Creates the run directory, freezes the seeded train/val/test split (written once), scores the unmodified seed capability on val, and records it as the candidate every algorithm must beat. Reports the remaining headroom so a saturated seed stops the run before it spends budget.
- ▌ Diagnose · skillberry-ai bundleExtract the learning signal from execution traces — the textual analogue of a gradient. Use between evaluation and proposing edits. Reads a candidate's rollouts and traces, separates good signals to keep from bad signals to fix, builds a reflective dataset (per failing task — Inputs, Generated Outputs, Feedback) and clusters the failures by a (failure-site, violated-expectation) signature, ranked by the score each cluster can recover, so the optimizer knows what to change and why.
- ▌ Evaluate · skillberry-ai bundleScore a candidate on a split with honest, variance-aware evaluation. Use whenever you need a number for a candidate (the algorithm calls it internally; you can also call it directly to inspect). Runs the target via the adapter for each task, scores each rollout, aggregates mean + standard error, and reports pass^k when trials > 1. Never touches the test split (that is finalize's sealed job).
- ▌ Finalize · skillberry-ai bundleScore the best candidate on the held-out TEST split exactly once and seal the run. Use as the last evaluation step, after optimization stops. The run dir enforces the seal — a second finalize raises an error — so the headline number is produced once on data the optimizer never saw, the way an honest benchmark result must be.
- ▌ Tools · skillberry-ai bundleOptimize an agent's OWN tool surface (tools it implements, not an external MCP server). Use when the agent mis-selects tools, fills arguments wrong, calls the same tool N times in a row, or has a confusing, redundant, or oversized toolset. Covers tool names and descriptions, parameter docs, tool schemas, handler code, function-calling accuracy, and adding or removing tools.
- ▌ Evograph · skillberry-ai bundleDeprecated agent-mode algorithm (evo-graph port): a weakness-graph search that dispatched one solver agent per failure cluster and reverted a whole round on regression. Do not start new runs with it — its per-weakness fan-out is already `agent-optimize`'s sibling fan-out, done behind the honest val significance gate that evograph never applied, and everything else it did (failure clustering, rejected-edit memory, budget-aware fan-out, free-text stop condition) lives in `agent-optimize` + `phases/diagnose`. Use when reading or repairing an existing evograph run dir, or when writing the run-dir `wiki/` format the dashboard's Weakness-graph tab reads — and to see what to select instead: `agent-optimize` for agent-mode search, `hill-climb`, `gepa`, or `skillopt` for a deterministic loop.
- ▌ Skillopt · skillberry-ai bundleRuns the SkillOpt single-lineage optimization loop, which organizes a hill-climb into epochs over mini-batches of train tasks under a textual learning rate — an integer edit budget that decays on a constant|linear|cosine schedule — and ends each epoch with one extra gated consolidation step. Parent is always the current best; acceptance is the val significance gate. Use when a run should anneal from broad early edits to small late ones and consolidate once per epoch, rather than hill-climb's one-shot whole-trainset proposals or gepa's Pareto frontier.
- ▌ Hill Climb · skillberry-ai bundleRuns a global hill-climb optimization loop where the parent is always the current best candidate and the val significance gate decides acceptance. Use as the algorithm for most runs — the first run on a new project, binary pass/fail scorers, and small task sets. Pick how each iteration's reflection is focused with --focus all (every failing val task), cyclic (one task at a time), or hardest-first (lowest-scoring first). Switch to gepa when rollouts are expensive and per-task feedback is rich, or skillopt when you want an annealed edit budget.
- ▌
- ▌ Orchestrate · skillberry-ai bundleDrive the entire cap-evolve pipeline end to end, autonomously. Use when the user wants the whole optimization run with minimal hand-holding. Sequences intake → implement-and-check → baseline → the chosen algorithm loop → finalize → report, enforces the cap-evolve-check hard gate before spending budget, decides when to stop (budget/stall), and surfaces the honest test number at the end. Reads capevolve.yaml; respects the ask-user-if-missing rule for inputs.
- ▌ Run Optimizer · skillberry-ai bundleDrives any shell-invokable coding agent (Claude Code, Codex, Gemini CLI, opencode, Cursor, Factory Droid, GitHub Copilot CLI, Kimi, Pi, Antigravity, OpenClaw, IBM Bob, or a fully custom command) as the edit proposer in a cap-evolve run, resolving the named optimizer from optimizers/registry.yaml. Use this as the optimizer for every run; pick the concrete agent with --name (or optimizer_skill in the spec). Use --name mock for a deterministic, zero-API proposer in tests and CI.
- ▌ Agent Optimize · skillberry-ai bundleFree-form optimization algorithm for agent orchestration mode: the conversational agent owns the whole search — proposing capability edits itself, screening them cheaply, gating each on full val, and sealing test once. Use when orchestration_mode is agent and algorithm_skill is agent-optimize. For a deterministic loop use hill-climb, gepa or skillopt instead.
- ▌ Skill Package · skillberry-ai bundleOptimize an Agent Skill package itself — its SKILL.md (frontmatter + body), its references, and its bundled scripts. Use when the capability under optimization IS a skill, you want the downstream agent to trigger it correctly and follow it without wasted steps, or you want a step the agent keeps skipping turned into deterministic bundled code. Checks every edit against the skill-creator authoring rules (valid frontmatter, progressive disclosure, one-level references, body budget, scripts that compile and self-check) so a candidate stays a valid, runnable skill.
- ▌
- ▌ Implement And Check · skillberry-ai bundleRuns the hard gate that has to pass before any optimization budget is spent. Use right after intake. Walks the agent through implementing the 3 required adapter methods plus any defaulted hooks that need overriding (and any selected skill's abstract methods), then runs `cap-evolve check` on the project plus each involved skill's check.py, listing exactly what is still stubbed or non-deterministic and what to do about each kind of failure.
- ▌ Arithmetic Answers · skillberry-aiAnswers a short arithmetic question with just the number. Use when the user asks to compute a sum, difference, or product and wants only the result back.
- ▌ Using Cap Evolve · skillberry-ai bundleFront door for cap-evolve: routes an optimization request to the right pipeline phase. Use when someone wants an agent, skill, system prompt, tool surface, or MCP toolset to score higher on an eval, benchmark, or task suite — "optimize my skill", "raise the pass rate on these tasks", "my agent keeps failing these cases", "get this prompt's accuracy up on my evals" — even when they never say "optimize", and whenever a .capevolve/ project or an unfinished run is in the tree. Routes to intake, the check gate, a resumed run, or the report; optimizes nothing itself. Not for making code or a query faster, and not for rewording one prompt with no eval to score it against. When the user names a phase or algorithm outright (baseline, gate, hill-climb, gepa), use that skill directly.
- ▌
- ▌ Swebench Solver · skillberry-ai bundleUse when fixing a bug in an open-source repository given a GitHub issue description. Analyzes the problem, locates the relevant code, and produces a minimal unified diff patch.