hillclimber
Optimization ledger / no account required

Every climb needs a trail.

Record experiments, compare canonical metrics, connect pull requests, and keep blends and ablations legible across agents. Live boards stay public; a local project key protects every write.

curl -fsSL https://hillclimber.experiment.sh/install.sh | bash
One evolving record

See the climb, not a pile of runs.

Every result stays attached to its code, PR, merge base, and outcome. When head moves, stale work is visible before it gets merged.

vision-r34head 9fe21d870a
live
Current head9fe21d870a
Active candidates2
Need rebase1
Accepted7
ExperimentIntegrationCommit pathAccuracyLatency
blend reranker + long contexte81fa902 · blend · PR #184
merged 72c641fa10 → 9fe21d870a91.7%42 ms
remove auxiliary headb937cc10 · ablation · PR #187
candidate 9fe21d870a91.2%36 ms
parallel optimizer sweep112dac8e · single · PR #181
needs rebase 72c641fa1090.4%47 ms
Reports that travel

From terminal output to an update worth sharing.

Turn the full experiment history into an executive summary, direction-aware metric charts, a decision record, and a ready-made social image.

01Optimization-level winsSee the percentage lift contributed by each accepted PR.
02Interaction reconciliationCompare the sum of isolated wins with the result actually observed on main.
03PNG and print exportPost the climb or send a clean PDF upstairs.
Hillclimber / evolution report

vision-r34

12 evaluated  ·  7 accepted  ·  3 metrics improved
accuracy · max91.7%
+7.7 pts
wider head reranker blend aux loss ablation main benchmark +3.7% +3.1% +2.1% mix −1.3%
84.0% baselineBest observed 91.9%Head 9fe21d870a
Agent-native workflow

Keep every result anchored to main.

01Define the objective hillclimber init --project vision-r34 --metric accuracy:max --metric latency_ms:min:ms
02Record the evaluated PR hillclimber --project vision-r34 add --description "wider prediction head" --pr https://github.com/org/repo/pull/42 --metric accuracy=.917 --metric latency_ms=42
03Record a merged optimization hillclimber --project vision-r34 merge --pr 42 --commit $(git rev-parse HEAD)
04Record the rebase and rerun together hillclimber --project vision-r34 rebase --pr 43 --onto $(git rev-parse origin/main) --metric accuracy=.921 --metric latency_ms=40
05Record the next merge hillclimber --project vision-r34 merge --pr 43 --commit $(git rev-parse HEAD)
06Benchmark main hillclimber --project vision-r34 benchmark --description "post-merge main" --metric accuracy=.924 --metric latency_ms=44
07Revert a conflicting PR hillclimber --project vision-r34 revert --pr 43 --commit $(git rev-parse HEAD) --reason "negative interaction"
08Share the climb hillclimber --project vision-r34 share --output vision-r34-hillclimb.png

Give the workflow to an agent.

A concise operating contract for experiments, metrics, and PR state.

Open skill.md