| name | discover |
| description | Initialize gepa-research for the current repository by exploring the codebase, proposing unexplored optimization dimensions, constructing the benchmark inside a baseline worktree, and running the first experiment. Use when the user invokes /gepa-research:discover, mentions setting up gepa-research, wants to instrument a codebase for autonomous optimization, or asks to start a new gepa-research run on a project. |
| argument-hint | <optional context about what to optimize> |
Discover
Internal procedure for gepa-research:discover. The user only sees the user-facing prompts, the dashboard URL, and the baseline score -- everything else is the agent's choreography.
Host conventions
This skill runs on any host that implements the Agent Skills spec. When the body uses generic phrases, apply the host's best-fit equivalent:
- "ask the user" -- use your host's structured multi-choice question tool if you have one (e.g.
AskUserQuestion, request_user_input). If the host has none, phrase the question as plain text in your next reply and wait for the user's answer.
- File paths like
references/... -- relative to this SKILL.md; resolve from the skill directory.
- Slash commands shown in user-facing copy (e.g.
/gepa-research:discover) -- translate to your host's mention syntax when speaking to the user (e.g. $gepa-research discover on Codex -- plugin namespace then skill name, separated by a space).
0. Verify the gepa-research CLI is available and in sync with the plugin
Before anything else, run:
gepa-research-version-check
This wraps gepa-research --version and additionally asserts the installed CLI matches the plugin manifest version (hosts refetch the plugin on version bumps, but do not reinstall the globally-installed CLI -- drift between the two breaks skills silently).
Four outcomes to handle:
- Exit 0,
gepa-research-version-check: OK (plugin=X, cli=X) -- continue to step 1.
- Exit 1, "plugin manifest and installed CLI disagree" -- stop and show the user the script's stderr verbatim; it tells them the
uv tool install --force "git+https://github.com/CyrusNuevoDia/gepa-research@v<version>#subdirectory=plugins/gepa-research" command to run. Then re-invoke this skill.
- Exit 2, "gepa-research CLI not on PATH" -- stop and tell the user:
gepa-research-cli isn't on your PATH. Install it once from GitHub: uv tool install "git+https://github.com/CyrusNuevoDia/gepa-research#subdirectory=plugins/gepa-research" (or substitute pipx install for uv tool install). Then re-invoke this skill.
gepa-research-version-check: command not found -- the host's plugin install is incomplete (missing the bin/ wrapper). Fall back to running gepa-research --version directly and check for gepa-research-cli in the output; if it's a different package, tell the user to uninstall it and install gepa-research-cli from GitHub in its place (uv tool install "git+https://github.com/CyrusNuevoDia/gepa-research#subdirectory=plugins/gepa-research").
Do not try to auto-install. Host sandbox + network policy may block it; leaving the install as a user action keeps failure modes clear.
Guiding principles
- Main stays clean. Never commit gepa-research-specific artifacts (benchmark harness, instrumentation, SDK imports) to main. Main should contain only what existed before gepa-research plus anything the user already had. All gepa-research-specific work happens inside worktree 0 (the baseline experiment).
- Baseline is a worktree, not a main commit.
gepa-research init creates .gepa-research/ but nothing in main changes. The first real experiment (exp_0000, created by gepa-research new --parent root) is where the benchmark and instrumentation live.
- Ask the user as little as possible. Every question is a beat of friction. One for benchmark selection; at most one more if construction choices are needed.
- Relay the dashboard URL verbatim when it prints. This is the user's window into the run.
1. Explore the repo
Understand what the codebase does. Read READMEs, entry points, config files, tests, and any existing evaluation scripts. Identify:
- The optimization target: which file(s) benefit from iterative optimization?
- Metric direction for each candidate: is higher better (
max) or lower better (min)?
- Critical behaviors worth gating: invariants that must never break regardless of score (e.g., "refund flow works", "core tests pass", "output is valid JSON"). Gates are commands that exit 0 on success, non-zero on failure.
2. Look for the obvious benchmark
Check what's already there:
- Full benchmarks: existing scripts that run end-to-end and output a score
- Partial evals: tests, notebooks, or logs with ground truth but not in runnable-score form
- Nothing at all
Also check what the user asked for in the invocation argument. If they named a specific metric or target, that's intent.
If one benchmark is obviously the right one — a runnable eval that measures what the user clearly cares about, or what the repo is plainly built to do — use it. Skip step 3, go to step 4 with that benchmark as the only candidate.
If it's not obvious — multiple candidate surfaces, no existing eval, user didn't specify intent, or the existing eval covers a narrow slice while the interesting optimization sits elsewhere — run step 3.
3. Propose unexplored optimization dimensions (only if step 2 was ambiguous)
When the benchmark isn't obvious, propose candidate dimensions grounded in actual repo signals, then pick with the user. See references/proposing-dimensions.md for the full rubric, project-type examples, and presentation format. Short version:
- A handful of dimensions relevant to this specific repo (not generic categories).
- Ground each in repo signals: already-instrumented code, stated goals in READMEs, TODO/FIXME patterns, domain defaults.
- Rank by signal × slack × cost answered in prose (no numeric scores — they're vibes).
4. Ask the user to pick the benchmark
If step 2 produced one obvious benchmark, confirm it in one sentence and move on — no ranked list needed.
Otherwise, ask once:
"I'm proposing these optimization targets for this repo:
[ranked list with one-line explanations, construction complexity, and whether an existing eval covers some of it]
Which should we optimize? Recommended: [default pick with reasoning]."
Record the selection. If step 3 ran, save non-picked dimensions to .gepa-research/project.md under "Future experiment candidates" after init.
5. Ask the user for instrumentation mode
Three cases, in order of how to handle them:
-
Selected benchmark already exists AND is already instrumented for gepa-research (you can see from gepa_research_agent import Run, an import { Run } from 'gepa-research', or the inline log_task / logTask helpers in the benchmark source). No wiring needed. Skip this question entirely. Detect the instrumentation style from the source and pass the matching --instrumentation-mode <sdk|inline> value to gepa-research init in step 7.
-
Selected benchmark already exists but is NOT instrumented (it just prints a score JSON, or it's a test runner that doesn't yet write per-task traces). Wiring is needed. Ask the question.
-
Selected benchmark needs to be constructed from scratch (case B or C from step 4). Wiring is needed. Ask the question.
For cases 2 and 3, ask once:
"I can wire up the benchmark in one of two ways:
- SDK mode -- install the SDK directly from this GitHub repo. Python:
pip install "git+https://github.com/CyrusNuevoDia/gepa-research#subdirectory=sdk/python". Node: git clone https://github.com/CyrusNuevoDia/gepa-research /tmp/gepa-research && npm install /tmp/gepa-research/sdk/node (npm has no native subdirectory-from-git install, so clone + local-path install is the canonical recipe). Richer per-task logs, ~5 lines of user code.
- Inline mode -- paste a ~30-line helper directly into the benchmark. Zero new dependencies. Same data contract."
Pass the answer to gepa-research init via --instrumentation-mode <sdk|inline> in step 7. Never install packages without this confirmation. If you skip the question (case 1), still pass the detected mode to gepa-research init so optimize/subagent runs see a consistent value.
6. Prepare main (without committing to it)
The agent never creates commits on main. Main stays byte-identical to what the user committed before gepa-research ran. Two things to set up, both local-only.
Order matters: do 6a (audit) before 6b (excludes). The excludes in 6b will hide files inside node_modules/, dist/, build/, etc. from git status. If you run the audit after adding excludes, you'll be blind to anything missing inside those directories -- and benchmark dependencies often live exactly there.
6a. Detect (don't auto-commit) dirty or untracked dependencies
gepa-research new forks a worktree from the current branch's HEAD commit, not from your dirty working tree. Any uncommitted edits to the target, benchmark, or gate dependencies are silently absent from exp_0000, and the whole optimization tree gets built against stale code while you think gepa-research is running on what you see locally.
Run three checks, in this order:
-
Tracked-but-modified files -- run git diff --name-only and git diff --cached --name-only. If any output line is the optimization target, an existing benchmark file, a gate-referenced script, or any of their import-graph dependencies, stop and ask the user to commit or stash before continuing. Do not commit on their behalf -- the user might be in the middle of an unrelated change.
-
Untracked files visible to git -- run git status --short --untracked-files=all and look for ?? entries that the target or gates will reference. Classify each:
- Part of the user's project (e.g., a smoke test they wrote but hadn't committed) -- stop and ask the user to commit it to main themselves.
- GEPAResearch-specific new files (a new gate script you're about to write, a new test fixture) -- do not create these in main. Defer to step 10; they go into the baseline worktree and commit to experiment 0's branch. Every descendant experiment inherits via git branching.
-
Explicit paths inside soon-to-be-ignored directories -- inspect the benchmark command and every gate command for path references (e.g., ./dist/eval-helper, node_modules/some-tool/cli.js, build/golden_outputs/). For each such path, run git ls-files --error-unmatch <path> to confirm it's tracked. If any aren't, stop and ask the user to commit them. This catches dependencies that step 6b is about to hide from git status.
Any one of these three checks failing is a hard stop. Do not proceed to 6b or beyond until the working tree is clean with respect to anything gepa-research will read.
Anything else (benchmark harness, instrumentation) always gets constructed inside the baseline worktree, never in main.
6b. Add local-only git excludes
After the audit passes, append to .git/info/exclude (not .gitignore -- we do not commit to main):
.gepa-research/
__pycache__/
*.pyc
.pytest_cache/
node_modules/
dist/
build/
.git/info/exclude is git's per-clone ignore file -- same effect as .gitignore, but never committed, never shared, invisible to history. Right tool for per-machine tooling state.
7. Initialize the workspace
gepa-research init --target <file> --benchmark "<command using {worktree} and {target}>" --metric <max|min> \
--instrumentation-mode <sdk|inline> [--gate "<gate command>"] \
[--objective "<one-sentence optimization goal>"] \
[--background "<domain knowledge / constraints for the LLM>"] \
[--reflection-lm "<provider/model, e.g. anthropic/claude-opus-4-7>"]
The last three flags are stored in config.json and forwarded to
gepa.optimize_anything during /optimize. They are optional but
strongly recommended — GEPA's reflection prompts are materially better
with a clear objective than with a generic "improve the score".
Good --objective values are one sentence naming the behavior, not the
artifact. "Maximize the pass rate on the tau3 benchmark while keeping
latency under 200ms per task." is better than "Improve solve.py".
--background is for facts the reflection LLM couldn't infer from code
alone: domain invariants, constraints, why prior approaches failed.
Placeholder semantics. Benchmark and gate commands support two placeholders, resolved lazily at run time by gepa-research run / gate evaluation:
{worktree} resolves to the absolute path of the experiment's worktree directory (e.g. /path/to/repo/.gepa-research/run_0000/worktrees/exp_0000). Use this to reference files that live on the experiment branch, not on main.{target} resolves to the absolute path of the target file inside that worktree (e.g. {worktree}/agent/solve.py). Use this when your benchmark needs to load or exec the target dynamically.
Critical rule: gepa-research run executes from the main repo root. When the benchmark script is constructed inside the worktree (the default in this flow), the command must reference it via {worktree} or the path won't resolve.
Example for a benchmark written at {worktree}/benchmark.py that will be committed to exp_0000:
gepa-research init \
--target agent/solve.py \
--benchmark "python3 {worktree}/benchmark.py --target {target}" \
--metric max
If the project uses a specific interpreter (poetry, pipenv, a venv), qualify it: "poetry run python {worktree}/benchmark.py ...", ".venv/bin/python {worktree}/benchmark.py ...", etc.
gepa-research init creates .gepa-research/, the synthetic root node, and auto-starts the dashboard. It prints a line like:
Dashboard live: http://127.0.0.1:8080 (pid 12345)
Relay that line back to the user verbatim. If port 8080 is busy, gepa-research auto-increments -- show whatever port prints. The URL is how the user watches the run.
8. Set up gates
Gates inherit down the experiment tree -- children automatically get all ancestor gates.
Gate semantics (read this first). gepa-research run decides "gate passed" purely from the command's exit code: 0 = pass, non-zero = fail. A benchmark-style command that just prints {"score": 0.0} and exits 0 passes the gate. That defeats the purpose. Every gate command must be wired to exit non-zero when the protected behavior regresses. Two ways to do that:
- Test-suite gates --
pytest, cargo test, npm test, etc. already exit non-zero on failure. Use them as-is.
- Score-threshold gates -- gate the benchmark on a minimum acceptable score. The benchmark script needs a flag like
--min-score <float> that exits 1 when the computed score falls below the threshold. The inline_instrumentation.{py,js} helpers in references/ show the pattern: write_result() returns the final score; the script can then compare and sys.exit(1).
Examples:
gepa-research gate add root --name core_tests --command "uv run --with pytest pytest tests/core/ -x"
gepa-research gate add root --name refund_flow --command "python3 {worktree}/benchmark.py --target {target} --task-ids 5 --min-score 0.9"
gepa-research gate add root --name no_crash --command "python3 smoke_test.py --target {target}"
If a benchmark you constructed doesn't yet have a --min-score mode, add it now (a few lines: parse the threshold flag, compute the score, sys.exit(1) if below). Without it the gate is decorative.
Gate commands support {target} and {worktree} placeholders with the same semantics as benchmark commands (resolved at run time, not at registration). Registering a gate that references {worktree}/benchmark.py before the benchmark exists is safe -- the placeholder resolves only when the gate is evaluated, which happens during gepa-research run after the benchmark is committed.
Verify registered gates:
gepa-research gate list root
Gate pairing rule based on benchmark provenance:
- If the selected benchmark already existed in the repo (not constructed from scratch): gates are optional at this step, but if you register any benchmark-derived gate, it must use a score-threshold (
--min-score or equivalent) -- not a bare invocation. Subagents can add more during optimization.
- If the benchmark was constructed from scratch (case B or C from the A/B/C classification): a Goodhart-mitigation gate is mandatory before the baseline can run, AND that gate must be a real pass/fail check (score-threshold or correctness assertion that exits non-zero on regression), not a bare benchmark rerun. See
references/constructing-benchmark.md section 6 on "Required gate pairing." Do not proceed to gepa-research new or gepa-research run without it. This is the safety against metric gaming -- it is not optional.
9. Create the baseline worktree
gepa-research new --parent root -m "baseline: instrument + score"
This returns experiment id (typically exp_0000) and its worktree path. All subsequent construction work happens inside that worktree -- never in main.
10. Work inside the baseline worktree
Cd into the worktree path returned by gepa-research new. Then:
10a. Construct the benchmark (if needed)
If the selected benchmark is new, build it in the worktree. See references/constructing-benchmark.md for the full procedure:
- Design the scoring function (range, direction, meaningful-improvement threshold)
- Assemble test cases (10-20 for programmatic, 15-30 for fuzzy, realistic workload for perf)
- Write the runnable harness (stdout = single JSON with
score, stderr = everything else)
- Goodhart check (document gaming strategies, mitigate each with a gate or held-out slice)
- Held-out validation slice (60/70 training, 30/40 held-out) if the benchmark is hand-written
Do not run separate determinism checks during setup. Note the benchmark's determinism property in project.md (step 12) and move on. Variance surfaces during optimization itself, where it can be handled with real evidence rather than guessed at during setup.
10b. Apply instrumentation
Based on the instrumentation mode passed to gepa-research init:
Paths below are relative to this SKILL.md file (resolve them against the skill directory).
- SDK mode: add
from gepa_research_agent import Run (Python) or import { Run } from 'gepa-research' (Node) to the benchmark script. Wrap the eval loop per references/sdk_python.py or references/sdk_node.js.
- Inline mode: copy the helper from
references/inline_instrumentation.py (or .js) into the benchmark. Use log_task / logTask per task and write_result / writeResult once at the end.
The wire protocol is the same either way: task_<id>.json written to $GEPA_RESEARCH_TRACES_DIR, single {"score": ...} JSON on stdout at the end, everything else on stderr.
If the underlying tool prints noisy stdout (progress bars, rich formatting, logging frameworks), redirect it to stderr.
10c. Cheap validation run
Before the full baseline, validate the toolchain with the cheapest possible end-to-end run (single task, smallest split, dry-run flag -- whatever is fastest).
Important: run this from the main repo root, not from inside the worktree. The validation writes traces to .gepa-research/validate/, which must resolve to the workspace's .gepa-research/ at the main repo root. If you run from the worktree, the relative path creates <worktree>/.gepa-research/validate/ and those artifacts get staged into the experiment commit when you run git add later.
Resolve {worktree}, {target}, and the validator script path yourself before running. GEPAResearch substitutes {worktree} / {target} only inside gepa-research run, not in a plain shell. The validation here is a plain shell call, so build the command with concrete absolute paths:
WORKTREE = the worktree path returned by gepa-research newTARGET = $WORKTREE/<relative target path, e.g. agent/solve.py>VALIDATOR = <absolute path to this skill dir>/scripts/validate_stdout.py -- resolve by taking the absolute path of this SKILL.md's directory and appending scripts/validate_stdout.py
WORKTREE="<...>"
TARGET="$WORKTREE/<...>"
VALIDATOR="<...>/scripts/validate_stdout.py"
mkdir -p .gepa-research/validate
GEPA_RESEARCH_TRACES_DIR=.gepa-research/validate \
python3 "$WORKTREE/benchmark.py" --target "$TARGET" \
2>.gepa-research/validate/stderr.log \
| python3 "$VALIDATOR"
Adapt the benchmark invocation (interpreter, args) to whatever you stored with gepa-research init. The non-negotiable part is that the resulting bash command contains no literal {worktree}, {target}, or relative-script paths -- expand all of them to absolute paths before the shell runs the line.
Notes:
- Traces go to
.gepa-research/validate/ at the main repo root (already locally ignored via step 6a). Avoids /tmp collisions and cross-process conflicts, keeps artifacts scoped to the project.
The validator checks:
- stdout is only valid JSON (no progress bars, tables, or logging mixed in)
- JSON contains a
score field with a numeric value
If validation fails, the script prints a diagnostic. Fix the benchmark wrapper and re-validate before proceeding. Also verify:
- All dependencies resolve and the command completes.
- Traces appear in
$GEPA_RESEARCH_TRACES_DIR (if applicable).
- Each gate script runs cleanly on the unmodified target.
Fix any issues and re-validate before proceeding. This catches environment problems, import errors, missing data, and stdout pollution cheaply.
10d. Commit inside the worktree
Logical commits are ideal but not required. Minimal acceptable:
add: benchmark harness + test casesadd: instrumentation (only in SDK mode -- inline mode keeps the harness and instrumentation in one file, so this commit collapses into the previous one)
Use git from inside the worktree directory. These commits are on the experiment's branch, not main.
Before the first commit in the worktree, add a .gitignore for build artifacts and any stray gepa-research workspace writes that shouldn't land on the experiment branch. At minimum:
.gepa-research/
__pycache__/
*.pyc
.pytest_cache/
node_modules/
dist/
build/
Otherwise, running the benchmark once before committing will drag bytecode caches, .pytest_cache/, or stray .gepa-research/ writes into the experiment's tree and pollute every descendant branch. Belt-and-suspenders with step 10c's "run from main repo root" rule: even if cwd slips, the ignore catches it.
11. Run the baseline
First, cd back to main repo root. If the previous step left the shell inside the worktree, gepa-research run will fail with "workspace not initialized" because .gepa-research/ only lives at the main repo root.
cd <main-repo-root>
gepa-research run exp_0000
gepa-research run executes the benchmark, captures the score, runs all inherited gates, and marks the experiment committed in a single step. Its output line ends with something like COMMITTED exp_0000 0.4286.
Do NOT call gepa-research done afterward. In the current CLI, gepa-research run is terminal: the experiment is already committed when it returns successfully, and calling gepa-research done exp_0000 --score <n> errors with "exp_0000 has status 'committed' -- cannot record again". The gepa-research done command exists for cases where a human recorded a score outside of gepa-research run, which is not the discover flow.
If gates failed, gepa-research run exits non-zero and leaves the experiment in a failed state. Fix the benchmark or target inside the worktree, commit, then gepa-research run exp_0000 again.
If gepa-research run fails with a path error (typically: benchmark.py not found), the stored benchmark command in .gepa-research/run_0000/config.json is missing the {worktree} placeholder. Fix by re-initializing: gepa-research discard exp_0000 --reason "benchmark command missing {worktree}" then re-run step 7 with the correct --benchmark string. Hand-editing config.json works but is technical debt.
12. Write .gepa-research/project.md
Lives at the top level of .gepa-research/ (run-agnostic, stable path regardless of active run). gepa-research init creates an empty stub; overwrite it.
Document:
- What the target does
- What can be changed by optimization vs what must stay stable
- How to interpret benchmark output (score meaning, direction)
- Benchmark determinism -- one line, pick what fits:
deterministic by construction -- pure code, no randomness, no networkuses LLMs with temp=0 -- expected to be deterministic in practice; flag if it isn'tsampling-based, variance expected -- inherent noise; optimize will need multi-run strategies
- Environment requirements discovered during validation
- What each gate protects
- Benchmark gaming risks identified during the Goodhart check
- Future experiment candidates (the non-picked dimensions from step 3)
13. Report to the user
End the skill by reporting in chat:
- The dashboard URL (if not already mentioned)
- The baseline experiment ID and score
- The chosen optimization dimension and why
- A one-liner on next steps: "Run
/gepa-research:optimize to start the optimization loop."
- Resume after crash: if the host, the shell, or the machine restarts mid-flow, re-invoke
gepa-research:optimize. GEPAResearch reads .gepa-research/ and resumes from the last committed experiment -- no special restore procedure.
- State is local to this machine: experiment commits on branches like
gepa-research/run_0000/exp_* survive git push --all, but orchestration state (graph, annotations, project notes) lives only in .gepa-research/. If that history matters to you, back up .gepa-research/ separately (e.g., tar -czf gepa-research-state-$(date +%F).tar.gz .gepa-research/).
Inspection commands (for debugging, reference only)
gepa-research get <id>
gepa-research traces <id> <task>
gepa-research annotate <id> "analysis" --task <task>
gepa-research scratchpad
gepa-research gate list <id>
Rules
- Do NOT modify main after
gepa-research init unless the user explicitly asks. All new artifacts live in worktree 0.
- Do NOT install packages without the user's confirmation from step 5.
- Do NOT skip the held-out gate pairing when the benchmark was constructed from scratch. The gate is the safety net against Goodhart gaming, regardless of whether the benchmark is deterministic.
- Do NOT skip the Goodhart check when the benchmark was constructed from scratch. Gate pairing is mandatory, not optional.
