@prismatic-io/lux 0.0.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +21 -0
- package/lib/answerers/claude-code/index.d.ts +17 -0
- package/lib/answerers/claude-code/index.d.ts.map +1 -0
- package/lib/answerers/claude-code/index.js +280 -0
- package/lib/answerers/claude-code/index.js.map +1 -0
- package/lib/answerers/persona/index.d.ts +40 -0
- package/lib/answerers/persona/index.d.ts.map +1 -0
- package/lib/answerers/persona/index.js +265 -0
- package/lib/answerers/persona/index.js.map +1 -0
- package/lib/answerers/scripted/index.d.ts +8 -0
- package/lib/answerers/scripted/index.d.ts.map +1 -0
- package/lib/answerers/scripted/index.js +163 -0
- package/lib/answerers/scripted/index.js.map +1 -0
- package/lib/answerers/shared.d.ts +13 -0
- package/lib/answerers/shared.d.ts.map +1 -0
- package/lib/answerers/shared.js +17 -0
- package/lib/answerers/shared.js.map +1 -0
- package/lib/answerers/terminal/index.d.ts +9 -0
- package/lib/answerers/terminal/index.d.ts.map +1 -0
- package/lib/answerers/terminal/index.js +97 -0
- package/lib/answerers/terminal/index.js.map +1 -0
- package/lib/assertions/core/command-exits-zero.d.ts +21 -0
- package/lib/assertions/core/command-exits-zero.d.ts.map +1 -0
- package/lib/assertions/core/command-exits-zero.js +87 -0
- package/lib/assertions/core/command-exits-zero.js.map +1 -0
- package/lib/assertions/core/event-checks.d.ts +57 -0
- package/lib/assertions/core/event-checks.d.ts.map +1 -0
- package/lib/assertions/core/event-checks.js +123 -0
- package/lib/assertions/core/event-checks.js.map +1 -0
- package/lib/assertions/core/file-checks.d.ts +103 -0
- package/lib/assertions/core/file-checks.d.ts.map +1 -0
- package/lib/assertions/core/file-checks.js +288 -0
- package/lib/assertions/core/file-checks.js.map +1 -0
- package/lib/assertions/core/helpers.d.ts +66 -0
- package/lib/assertions/core/helpers.d.ts.map +1 -0
- package/lib/assertions/core/helpers.js +170 -0
- package/lib/assertions/core/helpers.js.map +1 -0
- package/lib/assertions/core/index.d.ts +15 -0
- package/lib/assertions/core/index.d.ts.map +1 -0
- package/lib/assertions/core/index.js +56 -0
- package/lib/assertions/core/index.js.map +1 -0
- package/lib/assertions/core/json-pointer-equals.d.ts +19 -0
- package/lib/assertions/core/json-pointer-equals.d.ts.map +1 -0
- package/lib/assertions/core/json-pointer-equals.js +71 -0
- package/lib/assertions/core/json-pointer-equals.js.map +1 -0
- package/lib/assertions/core/output-checks.d.ts +46 -0
- package/lib/assertions/core/output-checks.d.ts.map +1 -0
- package/lib/assertions/core/output-checks.js +121 -0
- package/lib/assertions/core/output-checks.js.map +1 -0
- package/lib/assertions/core/predicate.d.ts +14 -0
- package/lib/assertions/core/predicate.d.ts.map +1 -0
- package/lib/assertions/core/predicate.js +55 -0
- package/lib/assertions/core/predicate.js.map +1 -0
- package/lib/assertions/core/resource-checks.d.ts +71 -0
- package/lib/assertions/core/resource-checks.d.ts.map +1 -0
- package/lib/assertions/core/resource-checks.js +159 -0
- package/lib/assertions/core/resource-checks.js.map +1 -0
- package/lib/assertions/core/safe-regex.d.ts +23 -0
- package/lib/assertions/core/safe-regex.d.ts.map +1 -0
- package/lib/assertions/core/safe-regex.js +102 -0
- package/lib/assertions/core/safe-regex.js.map +1 -0
- package/lib/assertions/core/tool-called-with.d.ts +21 -0
- package/lib/assertions/core/tool-called-with.d.ts.map +1 -0
- package/lib/assertions/core/tool-called-with.js +69 -0
- package/lib/assertions/core/tool-called-with.js.map +1 -0
- package/lib/assertions/core/tool-calls.d.ts +22 -0
- package/lib/assertions/core/tool-calls.d.ts.map +1 -0
- package/lib/assertions/core/tool-calls.js +17 -0
- package/lib/assertions/core/tool-calls.js.map +1 -0
- package/lib/assertions/core/tool-invocation.d.ts +36 -0
- package/lib/assertions/core/tool-invocation.d.ts.map +1 -0
- package/lib/assertions/core/tool-invocation.js +170 -0
- package/lib/assertions/core/tool-invocation.js.map +1 -0
- package/lib/assertions/core/tool-result-checks.d.ts +59 -0
- package/lib/assertions/core/tool-result-checks.d.ts.map +1 -0
- package/lib/assertions/core/tool-result-checks.js +193 -0
- package/lib/assertions/core/tool-result-checks.js.map +1 -0
- package/lib/assertions/mcp/discovery-checks.d.ts +36 -0
- package/lib/assertions/mcp/discovery-checks.d.ts.map +1 -0
- package/lib/assertions/mcp/discovery-checks.js +122 -0
- package/lib/assertions/mcp/discovery-checks.js.map +1 -0
- package/lib/assertions/mcp/index.d.ts +10 -0
- package/lib/assertions/mcp/index.d.ts.map +1 -0
- package/lib/assertions/mcp/index.js +25 -0
- package/lib/assertions/mcp/index.js.map +1 -0
- package/lib/assertions/rubric/index.d.ts +104 -0
- package/lib/assertions/rubric/index.d.ts.map +1 -0
- package/lib/assertions/rubric/index.js +191 -0
- package/lib/assertions/rubric/index.js.map +1 -0
- package/lib/assertions/rubric/internal.d.ts +14 -0
- package/lib/assertions/rubric/internal.d.ts.map +1 -0
- package/lib/assertions/rubric/internal.js +335 -0
- package/lib/assertions/rubric/internal.js.map +1 -0
- package/lib/authoring.d.ts +3 -0
- package/lib/authoring.d.ts.map +1 -0
- package/lib/authoring.js +3 -0
- package/lib/authoring.js.map +1 -0
- package/lib/cli/bin.d.ts +3 -0
- package/lib/cli/bin.d.ts.map +1 -0
- package/lib/cli/bin.js +15 -0
- package/lib/cli/bin.js.map +1 -0
- package/lib/cli/command-runtime.d.ts +5 -0
- package/lib/cli/command-runtime.d.ts.map +1 -0
- package/lib/cli/command-runtime.js +30 -0
- package/lib/cli/command-runtime.js.map +1 -0
- package/lib/cli/init.d.ts +21 -0
- package/lib/cli/init.d.ts.map +1 -0
- package/lib/cli/init.js +340 -0
- package/lib/cli/init.js.map +1 -0
- package/lib/cli/program.d.ts +22 -0
- package/lib/cli/program.d.ts.map +1 -0
- package/lib/cli/program.js +757 -0
- package/lib/cli/program.js.map +1 -0
- package/lib/cli/render/format.d.ts +14 -0
- package/lib/cli/render/format.d.ts.map +1 -0
- package/lib/cli/render/format.js +71 -0
- package/lib/cli/render/format.js.map +1 -0
- package/lib/cli/render/reporter.d.ts +73 -0
- package/lib/cli/render/reporter.d.ts.map +1 -0
- package/lib/cli/render/reporter.js +519 -0
- package/lib/cli/render/reporter.js.map +1 -0
- package/lib/cli/render/tty.d.ts +77 -0
- package/lib/cli/render/tty.d.ts.map +1 -0
- package/lib/cli/render/tty.js +283 -0
- package/lib/cli/render/tty.js.map +1 -0
- package/lib/cli/run-options.d.ts +27 -0
- package/lib/cli/run-options.d.ts.map +1 -0
- package/lib/cli/run-options.js +58 -0
- package/lib/cli/run-options.js.map +1 -0
- package/lib/cli/skills.d.ts +38 -0
- package/lib/cli/skills.d.ts.map +1 -0
- package/lib/cli/skills.js +84 -0
- package/lib/cli/skills.js.map +1 -0
- package/lib/core/annotation.d.ts +56 -0
- package/lib/core/annotation.d.ts.map +1 -0
- package/lib/core/annotation.js +92 -0
- package/lib/core/annotation.js.map +1 -0
- package/lib/core/answerer.d.ts +68 -0
- package/lib/core/answerer.d.ts.map +1 -0
- package/lib/core/answerer.js +17 -0
- package/lib/core/answerer.js.map +1 -0
- package/lib/core/artifact-evidence.d.ts +50 -0
- package/lib/core/artifact-evidence.d.ts.map +1 -0
- package/lib/core/artifact-evidence.js +185 -0
- package/lib/core/artifact-evidence.js.map +1 -0
- package/lib/core/artifact-snapshot.d.ts +37 -0
- package/lib/core/artifact-snapshot.d.ts.map +1 -0
- package/lib/core/artifact-snapshot.js +339 -0
- package/lib/core/artifact-snapshot.js.map +1 -0
- package/lib/core/assertion.d.ts +129 -0
- package/lib/core/assertion.d.ts.map +1 -0
- package/lib/core/assertion.js +75 -0
- package/lib/core/assertion.js.map +1 -0
- package/lib/core/case.d.ts +79 -0
- package/lib/core/case.d.ts.map +1 -0
- package/lib/core/case.js +90 -0
- package/lib/core/case.js.map +1 -0
- package/lib/core/conversation-turn.d.ts +23 -0
- package/lib/core/conversation-turn.d.ts.map +1 -0
- package/lib/core/conversation-turn.js +31 -0
- package/lib/core/conversation-turn.js.map +1 -0
- package/lib/core/driver.d.ts +64 -0
- package/lib/core/driver.d.ts.map +1 -0
- package/lib/core/driver.js +18 -0
- package/lib/core/driver.js.map +1 -0
- package/lib/core/emitter.d.ts +20 -0
- package/lib/core/emitter.d.ts.map +1 -0
- package/lib/core/emitter.js +34 -0
- package/lib/core/emitter.js.map +1 -0
- package/lib/core/events.d.ts +202 -0
- package/lib/core/events.d.ts.map +1 -0
- package/lib/core/events.js +107 -0
- package/lib/core/events.js.map +1 -0
- package/lib/core/exec.d.ts +82 -0
- package/lib/core/exec.d.ts.map +1 -0
- package/lib/core/exec.js +168 -0
- package/lib/core/exec.js.map +1 -0
- package/lib/core/experiment.d.ts +285 -0
- package/lib/core/experiment.d.ts.map +1 -0
- package/lib/core/experiment.js +389 -0
- package/lib/core/experiment.js.map +1 -0
- package/lib/core/glob.d.ts +3 -0
- package/lib/core/glob.d.ts.map +1 -0
- package/lib/core/glob.js +32 -0
- package/lib/core/glob.js.map +1 -0
- package/lib/core/hash.d.ts +8 -0
- package/lib/core/hash.d.ts.map +1 -0
- package/lib/core/hash.js +214 -0
- package/lib/core/hash.js.map +1 -0
- package/lib/core/index.d.ts +29 -0
- package/lib/core/index.d.ts.map +1 -0
- package/lib/core/index.js +29 -0
- package/lib/core/index.js.map +1 -0
- package/lib/core/lifecycle-fixtures.d.ts +20 -0
- package/lib/core/lifecycle-fixtures.d.ts.map +1 -0
- package/lib/core/lifecycle-fixtures.js +109 -0
- package/lib/core/lifecycle-fixtures.js.map +1 -0
- package/lib/core/lifecycle-hooks.d.ts +52 -0
- package/lib/core/lifecycle-hooks.d.ts.map +1 -0
- package/lib/core/lifecycle-hooks.js +42 -0
- package/lib/core/lifecycle-hooks.js.map +1 -0
- package/lib/core/mcp-events.d.ts +44 -0
- package/lib/core/mcp-events.d.ts.map +1 -0
- package/lib/core/mcp-events.js +23 -0
- package/lib/core/mcp-events.js.map +1 -0
- package/lib/core/model-cli.d.ts +57 -0
- package/lib/core/model-cli.d.ts.map +1 -0
- package/lib/core/model-cli.js +237 -0
- package/lib/core/model-cli.js.map +1 -0
- package/lib/core/observer.d.ts +31 -0
- package/lib/core/observer.d.ts.map +1 -0
- package/lib/core/observer.js +25 -0
- package/lib/core/observer.js.map +1 -0
- package/lib/core/platform-process.d.ts +33 -0
- package/lib/core/platform-process.d.ts.map +1 -0
- package/lib/core/platform-process.js +214 -0
- package/lib/core/platform-process.js.map +1 -0
- package/lib/core/registry.d.ts +22 -0
- package/lib/core/registry.d.ts.map +1 -0
- package/lib/core/registry.js +45 -0
- package/lib/core/registry.js.map +1 -0
- package/lib/core/rubric.d.ts +18 -0
- package/lib/core/rubric.d.ts.map +1 -0
- package/lib/core/rubric.js +49 -0
- package/lib/core/rubric.js.map +1 -0
- package/lib/core/run.d.ts +419 -0
- package/lib/core/run.d.ts.map +1 -0
- package/lib/core/run.js +126 -0
- package/lib/core/run.js.map +1 -0
- package/lib/core/statistics.d.ts +39 -0
- package/lib/core/statistics.d.ts.map +1 -0
- package/lib/core/statistics.js +163 -0
- package/lib/core/statistics.js.map +1 -0
- package/lib/core/subject.d.ts +4 -0
- package/lib/core/subject.d.ts.map +1 -0
- package/lib/core/subject.js +46 -0
- package/lib/core/subject.js.map +1 -0
- package/lib/core/tool-events.d.ts +69 -0
- package/lib/core/tool-events.d.ts.map +1 -0
- package/lib/core/tool-events.js +242 -0
- package/lib/core/tool-events.js.map +1 -0
- package/lib/core/usage.d.ts +78 -0
- package/lib/core/usage.d.ts.map +1 -0
- package/lib/core/usage.js +104 -0
- package/lib/core/usage.js.map +1 -0
- package/lib/core/wire.d.ts +83 -0
- package/lib/core/wire.d.ts.map +1 -0
- package/lib/core/wire.js +101 -0
- package/lib/core/wire.js.map +1 -0
- package/lib/core/zod.d.ts +10 -0
- package/lib/core/zod.d.ts.map +1 -0
- package/lib/core/zod.js +18 -0
- package/lib/core/zod.js.map +1 -0
- package/lib/drivers/claude-code/index.d.ts +55 -0
- package/lib/drivers/claude-code/index.d.ts.map +1 -0
- package/lib/drivers/claude-code/index.js +779 -0
- package/lib/drivers/claude-code/index.js.map +1 -0
- package/lib/drivers/claude-code/parse-events.d.ts +46 -0
- package/lib/drivers/claude-code/parse-events.d.ts.map +1 -0
- package/lib/drivers/claude-code/parse-events.js +387 -0
- package/lib/drivers/claude-code/parse-events.js.map +1 -0
- package/lib/drivers/codex/app-events.d.ts +16 -0
- package/lib/drivers/codex/app-events.d.ts.map +1 -0
- package/lib/drivers/codex/app-events.js +184 -0
- package/lib/drivers/codex/app-events.js.map +1 -0
- package/lib/drivers/codex/config.d.ts +55 -0
- package/lib/drivers/codex/config.d.ts.map +1 -0
- package/lib/drivers/codex/config.js +125 -0
- package/lib/drivers/codex/config.js.map +1 -0
- package/lib/drivers/codex/index.d.ts +8 -0
- package/lib/drivers/codex/index.d.ts.map +1 -0
- package/lib/drivers/codex/index.js +904 -0
- package/lib/drivers/codex/index.js.map +1 -0
- package/lib/drivers/mcp/index.d.ts +52 -0
- package/lib/drivers/mcp/index.d.ts.map +1 -0
- package/lib/drivers/mcp/index.js +282 -0
- package/lib/drivers/mcp/index.js.map +1 -0
- package/lib/drivers/mcp/transport.d.ts +50 -0
- package/lib/drivers/mcp/transport.d.ts.map +1 -0
- package/lib/drivers/mcp/transport.js +190 -0
- package/lib/drivers/mcp/transport.js.map +1 -0
- package/lib/drivers/shared/artifacts.d.ts +4 -0
- package/lib/drivers/shared/artifacts.d.ts.map +1 -0
- package/lib/drivers/shared/artifacts.js +81 -0
- package/lib/drivers/shared/artifacts.js.map +1 -0
- package/lib/drivers/shared/secure-copy.d.ts +2 -0
- package/lib/drivers/shared/secure-copy.d.ts.map +1 -0
- package/lib/drivers/shared/secure-copy.js +35 -0
- package/lib/drivers/shared/secure-copy.js.map +1 -0
- package/lib/drivers/subprocess/index.d.ts +24 -0
- package/lib/drivers/subprocess/index.d.ts.map +1 -0
- package/lib/drivers/subprocess/index.js +451 -0
- package/lib/drivers/subprocess/index.js.map +1 -0
- package/lib/index.d.ts +40 -0
- package/lib/index.d.ts.map +1 -0
- package/lib/index.js +48 -0
- package/lib/index.js.map +1 -0
- package/lib/optimization/reflective.d.ts +27 -0
- package/lib/optimization/reflective.d.ts.map +1 -0
- package/lib/optimization/reflective.js +257 -0
- package/lib/optimization/reflective.js.map +1 -0
- package/lib/orchestrator/annotation-loader.d.ts +65 -0
- package/lib/orchestrator/annotation-loader.d.ts.map +1 -0
- package/lib/orchestrator/annotation-loader.js +73 -0
- package/lib/orchestrator/annotation-loader.js.map +1 -0
- package/lib/orchestrator/annotation-store.d.ts +3 -0
- package/lib/orchestrator/annotation-store.d.ts.map +1 -0
- package/lib/orchestrator/annotation-store.js +100 -0
- package/lib/orchestrator/annotation-store.js.map +1 -0
- package/lib/orchestrator/authored-dependencies.d.ts +3 -0
- package/lib/orchestrator/authored-dependencies.d.ts.map +1 -0
- package/lib/orchestrator/authored-dependencies.js +146 -0
- package/lib/orchestrator/authored-dependencies.js.map +1 -0
- package/lib/orchestrator/campaign-lifecycle.d.ts +105 -0
- package/lib/orchestrator/campaign-lifecycle.d.ts.map +1 -0
- package/lib/orchestrator/campaign-lifecycle.js +141 -0
- package/lib/orchestrator/campaign-lifecycle.js.map +1 -0
- package/lib/orchestrator/candidate-integrity.d.ts +4 -0
- package/lib/orchestrator/candidate-integrity.d.ts.map +1 -0
- package/lib/orchestrator/candidate-integrity.js +14 -0
- package/lib/orchestrator/candidate-integrity.js.map +1 -0
- package/lib/orchestrator/candidates.d.ts +13 -0
- package/lib/orchestrator/candidates.d.ts.map +1 -0
- package/lib/orchestrator/candidates.js +577 -0
- package/lib/orchestrator/candidates.js.map +1 -0
- package/lib/orchestrator/compare.d.ts +80 -0
- package/lib/orchestrator/compare.d.ts.map +1 -0
- package/lib/orchestrator/compare.js +270 -0
- package/lib/orchestrator/compare.js.map +1 -0
- package/lib/orchestrator/comparison-identity.d.ts +10 -0
- package/lib/orchestrator/comparison-identity.d.ts.map +1 -0
- package/lib/orchestrator/comparison-identity.js +39 -0
- package/lib/orchestrator/comparison-identity.js.map +1 -0
- package/lib/orchestrator/config.d.ts +121 -0
- package/lib/orchestrator/config.d.ts.map +1 -0
- package/lib/orchestrator/config.js +176 -0
- package/lib/orchestrator/config.js.map +1 -0
- package/lib/orchestrator/content-bound-json.d.ts +5 -0
- package/lib/orchestrator/content-bound-json.d.ts.map +1 -0
- package/lib/orchestrator/content-bound-json.js +59 -0
- package/lib/orchestrator/content-bound-json.js.map +1 -0
- package/lib/orchestrator/discover.d.ts +45 -0
- package/lib/orchestrator/discover.d.ts.map +1 -0
- package/lib/orchestrator/discover.js +98 -0
- package/lib/orchestrator/discover.js.map +1 -0
- package/lib/orchestrator/doctor.d.ts +141 -0
- package/lib/orchestrator/doctor.d.ts.map +1 -0
- package/lib/orchestrator/doctor.js +525 -0
- package/lib/orchestrator/doctor.js.map +1 -0
- package/lib/orchestrator/driver-session.d.ts +177 -0
- package/lib/orchestrator/driver-session.d.ts.map +1 -0
- package/lib/orchestrator/driver-session.js +396 -0
- package/lib/orchestrator/driver-session.js.map +1 -0
- package/lib/orchestrator/evaluator-identity.d.ts +7 -0
- package/lib/orchestrator/evaluator-identity.d.ts.map +1 -0
- package/lib/orchestrator/evaluator-identity.js +42 -0
- package/lib/orchestrator/evaluator-identity.js.map +1 -0
- package/lib/orchestrator/experiment-budget.d.ts +11 -0
- package/lib/orchestrator/experiment-budget.d.ts.map +1 -0
- package/lib/orchestrator/experiment-budget.js +51 -0
- package/lib/orchestrator/experiment-budget.js.map +1 -0
- package/lib/orchestrator/experiment-context.d.ts +43 -0
- package/lib/orchestrator/experiment-context.d.ts.map +1 -0
- package/lib/orchestrator/experiment-context.js +9 -0
- package/lib/orchestrator/experiment-context.js.map +1 -0
- package/lib/orchestrator/experiment-contracts.d.ts +290 -0
- package/lib/orchestrator/experiment-contracts.d.ts.map +1 -0
- package/lib/orchestrator/experiment-contracts.js +38 -0
- package/lib/orchestrator/experiment-contracts.js.map +1 -0
- package/lib/orchestrator/experiment-corpus.d.ts +21 -0
- package/lib/orchestrator/experiment-corpus.d.ts.map +1 -0
- package/lib/orchestrator/experiment-corpus.js +276 -0
- package/lib/orchestrator/experiment-corpus.js.map +1 -0
- package/lib/orchestrator/experiment-evaluation.d.ts +44 -0
- package/lib/orchestrator/experiment-evaluation.d.ts.map +1 -0
- package/lib/orchestrator/experiment-evaluation.js +542 -0
- package/lib/orchestrator/experiment-evaluation.js.map +1 -0
- package/lib/orchestrator/experiment-finish.d.ts +6 -0
- package/lib/orchestrator/experiment-finish.d.ts.map +1 -0
- package/lib/orchestrator/experiment-finish.js +19 -0
- package/lib/orchestrator/experiment-finish.js.map +1 -0
- package/lib/orchestrator/experiment-identity.d.ts +6 -0
- package/lib/orchestrator/experiment-identity.d.ts.map +1 -0
- package/lib/orchestrator/experiment-identity.js +257 -0
- package/lib/orchestrator/experiment-identity.js.map +1 -0
- package/lib/orchestrator/experiment-loader.d.ts +175 -0
- package/lib/orchestrator/experiment-loader.d.ts.map +1 -0
- package/lib/orchestrator/experiment-loader.js +77 -0
- package/lib/orchestrator/experiment-loader.js.map +1 -0
- package/lib/orchestrator/experiment-report.d.ts +450 -0
- package/lib/orchestrator/experiment-report.d.ts.map +1 -0
- package/lib/orchestrator/experiment-report.js +570 -0
- package/lib/orchestrator/experiment-report.js.map +1 -0
- package/lib/orchestrator/experiment-runs.d.ts +379 -0
- package/lib/orchestrator/experiment-runs.d.ts.map +1 -0
- package/lib/orchestrator/experiment-runs.js +796 -0
- package/lib/orchestrator/experiment-runs.js.map +1 -0
- package/lib/orchestrator/experiment-runtime-identity.d.ts +41 -0
- package/lib/orchestrator/experiment-runtime-identity.d.ts.map +1 -0
- package/lib/orchestrator/experiment-runtime-identity.js +411 -0
- package/lib/orchestrator/experiment-runtime-identity.js.map +1 -0
- package/lib/orchestrator/experiment-selection.d.ts +25 -0
- package/lib/orchestrator/experiment-selection.d.ts.map +1 -0
- package/lib/orchestrator/experiment-selection.js +286 -0
- package/lib/orchestrator/experiment-selection.js.map +1 -0
- package/lib/orchestrator/experiment.d.ts +11 -0
- package/lib/orchestrator/experiment.d.ts.map +1 -0
- package/lib/orchestrator/experiment.js +639 -0
- package/lib/orchestrator/experiment.js.map +1 -0
- package/lib/orchestrator/fixture-identity.d.ts +7 -0
- package/lib/orchestrator/fixture-identity.d.ts.map +1 -0
- package/lib/orchestrator/fixture-identity.js +55 -0
- package/lib/orchestrator/fixture-identity.js.map +1 -0
- package/lib/orchestrator/fs-read.d.ts +19 -0
- package/lib/orchestrator/fs-read.d.ts.map +1 -0
- package/lib/orchestrator/fs-read.js +113 -0
- package/lib/orchestrator/fs-read.js.map +1 -0
- package/lib/orchestrator/grader-experiment.d.ts +190 -0
- package/lib/orchestrator/grader-experiment.d.ts.map +1 -0
- package/lib/orchestrator/grader-experiment.js +579 -0
- package/lib/orchestrator/grader-experiment.js.map +1 -0
- package/lib/orchestrator/grading.d.ts +31 -0
- package/lib/orchestrator/grading.d.ts.map +1 -0
- package/lib/orchestrator/grading.js +186 -0
- package/lib/orchestrator/grading.js.map +1 -0
- package/lib/orchestrator/index.d.ts +22 -0
- package/lib/orchestrator/index.d.ts.map +1 -0
- package/lib/orchestrator/index.js +30 -0
- package/lib/orchestrator/index.js.map +1 -0
- package/lib/orchestrator/lifecycle-hook-execution.d.ts +4 -0
- package/lib/orchestrator/lifecycle-hook-execution.d.ts.map +1 -0
- package/lib/orchestrator/lifecycle-hook-execution.js +67 -0
- package/lib/orchestrator/lifecycle-hook-execution.js.map +1 -0
- package/lib/orchestrator/list-runs.d.ts +33 -0
- package/lib/orchestrator/list-runs.d.ts.map +1 -0
- package/lib/orchestrator/list-runs.js +308 -0
- package/lib/orchestrator/list-runs.js.map +1 -0
- package/lib/orchestrator/loader.d.ts +3 -0
- package/lib/orchestrator/loader.d.ts.map +1 -0
- package/lib/orchestrator/loader.js +17 -0
- package/lib/orchestrator/loader.js.map +1 -0
- package/lib/orchestrator/optimization-lifecycle.d.ts +12 -0
- package/lib/orchestrator/optimization-lifecycle.d.ts.map +1 -0
- package/lib/orchestrator/optimization-lifecycle.js +92 -0
- package/lib/orchestrator/optimization-lifecycle.js.map +1 -0
- package/lib/orchestrator/orchestrator.d.ts +144 -0
- package/lib/orchestrator/orchestrator.d.ts.map +1 -0
- package/lib/orchestrator/orchestrator.js +632 -0
- package/lib/orchestrator/orchestrator.js.map +1 -0
- package/lib/orchestrator/owned-lock.d.ts +17 -0
- package/lib/orchestrator/owned-lock.d.ts.map +1 -0
- package/lib/orchestrator/owned-lock.js +149 -0
- package/lib/orchestrator/owned-lock.js.map +1 -0
- package/lib/orchestrator/package-version.d.ts +2 -0
- package/lib/orchestrator/package-version.d.ts.map +1 -0
- package/lib/orchestrator/package-version.js +4 -0
- package/lib/orchestrator/package-version.js.map +1 -0
- package/lib/orchestrator/provenance.d.ts +5 -0
- package/lib/orchestrator/provenance.d.ts.map +1 -0
- package/lib/orchestrator/provenance.js +137 -0
- package/lib/orchestrator/provenance.js.map +1 -0
- package/lib/orchestrator/regrade.d.ts +80 -0
- package/lib/orchestrator/regrade.d.ts.map +1 -0
- package/lib/orchestrator/regrade.js +433 -0
- package/lib/orchestrator/regrade.js.map +1 -0
- package/lib/orchestrator/rubric-loader.d.ts +7 -0
- package/lib/orchestrator/rubric-loader.d.ts.map +1 -0
- package/lib/orchestrator/rubric-loader.js +65 -0
- package/lib/orchestrator/rubric-loader.js.map +1 -0
- package/lib/orchestrator/run-dir.d.ts +40 -0
- package/lib/orchestrator/run-dir.d.ts.map +1 -0
- package/lib/orchestrator/run-dir.js +161 -0
- package/lib/orchestrator/run-dir.js.map +1 -0
- package/lib/orchestrator/run-execution.d.ts +30 -0
- package/lib/orchestrator/run-execution.d.ts.map +1 -0
- package/lib/orchestrator/run-execution.js +531 -0
- package/lib/orchestrator/run-execution.js.map +1 -0
- package/lib/orchestrator/run-lifecycle.d.ts +115 -0
- package/lib/orchestrator/run-lifecycle.d.ts.map +1 -0
- package/lib/orchestrator/run-lifecycle.js +136 -0
- package/lib/orchestrator/run-lifecycle.js.map +1 -0
- package/lib/orchestrator/suite-compare.d.ts +76 -0
- package/lib/orchestrator/suite-compare.d.ts.map +1 -0
- package/lib/orchestrator/suite-compare.js +159 -0
- package/lib/orchestrator/suite-compare.js.map +1 -0
- package/lib/orchestrator/suite-lifecycle.d.ts +157 -0
- package/lib/orchestrator/suite-lifecycle.d.ts.map +1 -0
- package/lib/orchestrator/suite-lifecycle.js +253 -0
- package/lib/orchestrator/suite-lifecycle.js.map +1 -0
- package/lib/orchestrator/suite-runs.d.ts +91 -0
- package/lib/orchestrator/suite-runs.d.ts.map +1 -0
- package/lib/orchestrator/suite-runs.js +157 -0
- package/lib/orchestrator/suite-runs.js.map +1 -0
- package/lib/orchestrator/suite-summary.d.ts +109 -0
- package/lib/orchestrator/suite-summary.d.ts.map +1 -0
- package/lib/orchestrator/suite-summary.js +213 -0
- package/lib/orchestrator/suite-summary.js.map +1 -0
- package/lib/orchestrator/trace-limits.d.ts +20 -0
- package/lib/orchestrator/trace-limits.d.ts.map +1 -0
- package/lib/orchestrator/trace-limits.js +28 -0
- package/lib/orchestrator/trace-limits.js.map +1 -0
- package/lib/orchestrator/validate-plugin-config.d.ts +10 -0
- package/lib/orchestrator/validate-plugin-config.d.ts.map +1 -0
- package/lib/orchestrator/validate-plugin-config.js +26 -0
- package/lib/orchestrator/validate-plugin-config.js.map +1 -0
- package/package.json +84 -0
- package/skills/lux/SKILL.md +85 -0
- package/skills/lux/references/cli.md +102 -0
- package/skills/lux/references/eval-authoring.md +105 -0
- package/skills/lux-answerer/SKILL.md +169 -0
- package/src/answerers/claude-code/index.ts +342 -0
- package/src/answerers/persona/index.ts +327 -0
- package/src/answerers/scripted/index.ts +186 -0
- package/src/answerers/shared.ts +23 -0
- package/src/answerers/terminal/index.ts +117 -0
- package/src/assertions/core/command-exits-zero.ts +94 -0
- package/src/assertions/core/event-checks.ts +138 -0
- package/src/assertions/core/file-checks.ts +336 -0
- package/src/assertions/core/helpers.ts +203 -0
- package/src/assertions/core/index.ts +80 -0
- package/src/assertions/core/json-pointer-equals.ts +84 -0
- package/src/assertions/core/output-checks.ts +134 -0
- package/src/assertions/core/predicate.ts +67 -0
- package/src/assertions/core/resource-checks.ts +180 -0
- package/src/assertions/core/safe-regex.ts +196 -0
- package/src/assertions/core/tool-called-with.ts +68 -0
- package/src/assertions/core/tool-calls.ts +28 -0
- package/src/assertions/core/tool-invocation.ts +198 -0
- package/src/assertions/core/tool-result-checks.ts +211 -0
- package/src/assertions/mcp/discovery-checks.ts +139 -0
- package/src/assertions/mcp/index.ts +37 -0
- package/src/assertions/rubric/index.ts +258 -0
- package/src/assertions/rubric/internal.ts +367 -0
- package/src/authoring.ts +2 -0
- package/src/cli/bin.ts +14 -0
- package/src/cli/command-runtime.ts +34 -0
- package/src/cli/init.ts +410 -0
- package/src/cli/program.ts +992 -0
- package/src/cli/render/format.ts +69 -0
- package/src/cli/render/reporter.ts +643 -0
- package/src/cli/render/tty.ts +314 -0
- package/src/cli/run-options.ts +82 -0
- package/src/cli/skills.ts +129 -0
- package/src/core/annotation.ts +104 -0
- package/src/core/answerer.ts +75 -0
- package/src/core/artifact-evidence.ts +249 -0
- package/src/core/artifact-snapshot.ts +429 -0
- package/src/core/assertion.ts +156 -0
- package/src/core/case.ts +103 -0
- package/src/core/conversation-turn.ts +47 -0
- package/src/core/driver.ts +75 -0
- package/src/core/emitter.ts +38 -0
- package/src/core/events.ts +135 -0
- package/src/core/exec.ts +263 -0
- package/src/core/experiment.ts +467 -0
- package/src/core/glob.ts +31 -0
- package/src/core/hash.ts +237 -0
- package/src/core/index.ts +28 -0
- package/src/core/lifecycle-fixtures.ts +145 -0
- package/src/core/lifecycle-hooks.ts +101 -0
- package/src/core/mcp-events.ts +57 -0
- package/src/core/model-cli.ts +271 -0
- package/src/core/observer.ts +50 -0
- package/src/core/platform-process.ts +284 -0
- package/src/core/registry.ts +57 -0
- package/src/core/rubric.ts +52 -0
- package/src/core/run.ts +144 -0
- package/src/core/statistics.ts +194 -0
- package/src/core/subject.ts +50 -0
- package/src/core/tool-events.ts +292 -0
- package/src/core/usage.ts +123 -0
- package/src/core/wire.ts +135 -0
- package/src/core/zod.ts +20 -0
- package/src/drivers/claude-code/index.ts +866 -0
- package/src/drivers/claude-code/parse-events.ts +464 -0
- package/src/drivers/codex/app-events.ts +218 -0
- package/src/drivers/codex/config.ts +128 -0
- package/src/drivers/codex/index.ts +995 -0
- package/src/drivers/mcp/index.ts +351 -0
- package/src/drivers/mcp/transport.ts +222 -0
- package/src/drivers/shared/artifacts.ts +86 -0
- package/src/drivers/shared/secure-copy.ts +40 -0
- package/src/drivers/subprocess/index.ts +490 -0
- package/src/index.ts +594 -0
- package/src/optimization/reflective.ts +319 -0
- package/src/orchestrator/annotation-loader.ts +97 -0
- package/src/orchestrator/annotation-store.ts +112 -0
- package/src/orchestrator/authored-dependencies.ts +193 -0
- package/src/orchestrator/campaign-lifecycle.ts +181 -0
- package/src/orchestrator/candidate-integrity.ts +18 -0
- package/src/orchestrator/candidates.ts +683 -0
- package/src/orchestrator/compare.ts +358 -0
- package/src/orchestrator/comparison-identity.ts +45 -0
- package/src/orchestrator/config.ts +205 -0
- package/src/orchestrator/content-bound-json.ts +71 -0
- package/src/orchestrator/discover.ts +132 -0
- package/src/orchestrator/doctor.ts +781 -0
- package/src/orchestrator/driver-session.ts +507 -0
- package/src/orchestrator/evaluator-identity.ts +45 -0
- package/src/orchestrator/experiment-budget.ts +61 -0
- package/src/orchestrator/experiment-context.ts +41 -0
- package/src/orchestrator/experiment-contracts.ts +43 -0
- package/src/orchestrator/experiment-corpus.ts +416 -0
- package/src/orchestrator/experiment-evaluation.ts +749 -0
- package/src/orchestrator/experiment-finish.ts +33 -0
- package/src/orchestrator/experiment-identity.ts +356 -0
- package/src/orchestrator/experiment-loader.ts +105 -0
- package/src/orchestrator/experiment-report.ts +788 -0
- package/src/orchestrator/experiment-runs.ts +960 -0
- package/src/orchestrator/experiment-runtime-identity.ts +465 -0
- package/src/orchestrator/experiment-selection.ts +467 -0
- package/src/orchestrator/experiment.ts +991 -0
- package/src/orchestrator/fixture-identity.ts +67 -0
- package/src/orchestrator/fs-read.ts +126 -0
- package/src/orchestrator/grader-experiment.ts +856 -0
- package/src/orchestrator/grading.ts +260 -0
- package/src/orchestrator/index.ts +186 -0
- package/src/orchestrator/lifecycle-hook-execution.ts +90 -0
- package/src/orchestrator/list-runs.ts +384 -0
- package/src/orchestrator/loader.ts +19 -0
- package/src/orchestrator/optimization-lifecycle.ts +126 -0
- package/src/orchestrator/orchestrator.ts +919 -0
- package/src/orchestrator/owned-lock.ts +147 -0
- package/src/orchestrator/package-version.ts +5 -0
- package/src/orchestrator/provenance.ts +140 -0
- package/src/orchestrator/regrade.ts +580 -0
- package/src/orchestrator/rubric-loader.ts +74 -0
- package/src/orchestrator/run-dir.ts +209 -0
- package/src/orchestrator/run-execution.ts +694 -0
- package/src/orchestrator/run-lifecycle.ts +174 -0
- package/src/orchestrator/suite-compare.ts +195 -0
- package/src/orchestrator/suite-lifecycle.ts +335 -0
- package/src/orchestrator/suite-runs.ts +202 -0
- package/src/orchestrator/suite-summary.ts +276 -0
- package/src/orchestrator/trace-limits.ts +65 -0
- package/src/orchestrator/validate-plugin-config.ts +34 -0
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
import { type ExperimentCampaign } from "../core/index.js";
|
|
2
|
+
import type { CampaignExecutionOptions } from "./experiment-context.js";
|
|
3
|
+
import { type OpenedExperimentRun } from "./experiment-runs.js";
|
|
4
|
+
export declare const throwIfAborted: (signal: AbortSignal | undefined) => void;
|
|
5
|
+
export declare const boundedExecutionSignal: (execution: CampaignExecutionOptions, experiment: OpenedExperimentRun) => AbortSignal;
|
|
6
|
+
export declare const ensureBudget: (campaign: ExperimentCampaign, experiment: OpenedExperimentRun, additional?: {
|
|
7
|
+
metricCalls?: number;
|
|
8
|
+
judgeCalls?: number;
|
|
9
|
+
proposalCalls?: number;
|
|
10
|
+
}) => void;
|
|
11
|
+
//# sourceMappingURL=experiment-budget.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-budget.d.ts","sourceRoot":"","sources":["../../src/orchestrator/experiment-budget.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,kBAAkB,EAAe,MAAM,kBAAkB,CAAC;AACxE,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,yBAAyB,CAAC;AACxE,OAAO,EAA6B,KAAK,mBAAmB,EAAE,MAAM,sBAAsB,CAAC;AAE3F,eAAO,MAAM,cAAc,WAAY,WAAW,GAAG,SAAS,KAAG,IAGhE,CAAC;AAEF,eAAO,MAAM,sBAAsB,cACtB,wBAAwB,cACvB,mBAAmB,KAC9B,WAUF,CAAC;AAEF,eAAO,MAAM,YAAY,aACb,kBAAkB,cAChB,mBAAmB,eACnB;IAAE,WAAW,CAAC,EAAE,MAAM,CAAC;IAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAAC,aAAa,CAAC,EAAE,MAAM,CAAA;CAAE,KAChF,IAgCF,CAAC"}
|
|
@@ -0,0 +1,51 @@
|
|
|
1
|
+
import { totalTokens } from "../core/index.js";
|
|
2
|
+
import { experimentActiveElapsedMs } from "./experiment-runs.js";
|
|
3
|
+
export const throwIfAborted = (signal) => {
|
|
4
|
+
if (!signal?.aborted)
|
|
5
|
+
return;
|
|
6
|
+
throw signal.reason instanceof Error ? signal.reason : new Error("experiment aborted");
|
|
7
|
+
};
|
|
8
|
+
export const boundedExecutionSignal = (execution, experiment) => {
|
|
9
|
+
const elapsed = experimentActiveElapsedMs(experiment.manifest);
|
|
10
|
+
const remaining = execution.loaded.campaign.budget.maxWallTimeMs - elapsed;
|
|
11
|
+
if (remaining <= 0) {
|
|
12
|
+
throw new Error(`experiment exceeded its wall-time budget ${execution.loaded.campaign.budget.maxWallTimeMs}ms`);
|
|
13
|
+
}
|
|
14
|
+
const deadline = AbortSignal.timeout(Math.max(1, remaining));
|
|
15
|
+
return execution.abortSignal ? AbortSignal.any([execution.abortSignal, deadline]) : deadline;
|
|
16
|
+
};
|
|
17
|
+
export const ensureBudget = (campaign, experiment, additional = {}) => {
|
|
18
|
+
const usage = experiment.manifest.budgetUsage;
|
|
19
|
+
const elapsed = experimentActiveElapsedMs(experiment.manifest);
|
|
20
|
+
const checks = [
|
|
21
|
+
[
|
|
22
|
+
usage.metricCalls + (additional.metricCalls ?? 0) > campaign.budget.maxMetricCalls,
|
|
23
|
+
`metric-call budget ${campaign.budget.maxMetricCalls}`,
|
|
24
|
+
],
|
|
25
|
+
[
|
|
26
|
+
usage.judgeCalls + (additional.judgeCalls ?? 0) > campaign.budget.maxJudgeCalls,
|
|
27
|
+
`judge-call budget ${campaign.budget.maxJudgeCalls}`,
|
|
28
|
+
],
|
|
29
|
+
[
|
|
30
|
+
usage.proposalCalls + (additional.proposalCalls ?? 0) > campaign.budget.maxProposalCalls,
|
|
31
|
+
`proposal-call budget ${campaign.budget.maxProposalCalls}`,
|
|
32
|
+
],
|
|
33
|
+
[
|
|
34
|
+
totalTokens(usage.usage) > campaign.budget.maxTokens,
|
|
35
|
+
`token budget ${campaign.budget.maxTokens}`,
|
|
36
|
+
],
|
|
37
|
+
[usage.costUsd > campaign.budget.maxCostUsd, `cost budget $${campaign.budget.maxCostUsd}`],
|
|
38
|
+
[
|
|
39
|
+
(usage.unpricedModelCalls ?? 0) > campaign.budget.maxUnpricedModelCalls,
|
|
40
|
+
`unpriced-model-call budget ${campaign.budget.maxUnpricedModelCalls}`,
|
|
41
|
+
],
|
|
42
|
+
[
|
|
43
|
+
elapsed > campaign.budget.maxWallTimeMs,
|
|
44
|
+
`wall-time budget ${campaign.budget.maxWallTimeMs}ms`,
|
|
45
|
+
],
|
|
46
|
+
];
|
|
47
|
+
const exceeded = checks.find(([failed]) => failed);
|
|
48
|
+
if (exceeded)
|
|
49
|
+
throw new Error(`experiment exceeded its ${exceeded[1]}`);
|
|
50
|
+
};
|
|
51
|
+
//# sourceMappingURL=experiment-budget.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-budget.js","sourceRoot":"","sources":["../../src/orchestrator/experiment-budget.ts"],"names":[],"mappings":"AAAA,OAAO,EAA2B,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAExE,OAAO,EAAE,yBAAyB,EAA4B,MAAM,sBAAsB,CAAC;AAE3F,MAAM,CAAC,MAAM,cAAc,GAAG,CAAC,MAA+B,EAAQ,EAAE;IACtE,IAAI,CAAC,MAAM,EAAE,OAAO;QAAE,OAAO;IAC7B,MAAM,MAAM,CAAC,MAAM,YAAY,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,oBAAoB,CAAC,CAAC;AACzF,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,sBAAsB,GAAG,CACpC,SAAmC,EACnC,UAA+B,EAClB,EAAE;IACf,MAAM,OAAO,GAAG,yBAAyB,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC;IAC/D,MAAM,SAAS,GAAG,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,aAAa,GAAG,OAAO,CAAC;IAC3E,IAAI,SAAS,IAAI,CAAC,EAAE,CAAC;QACnB,MAAM,IAAI,KAAK,CACb,4CAA4C,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,aAAa,IAAI,CAC/F,CAAC;IACJ,CAAC;IACD,MAAM,QAAQ,GAAG,WAAW,CAAC,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC,CAAC;IAC7D,OAAO,SAAS,CAAC,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,SAAS,CAAC,WAAW,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC;AAC/F,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,YAAY,GAAG,CAC1B,QAA4B,EAC5B,UAA+B,EAC/B,UAAU,GAA0E,EAAE,EAChF,EAAE;IACR,MAAM,KAAK,GAAG,UAAU,CAAC,QAAQ,CAAC,WAAW,CAAC;IAC9C,MAAM,OAAO,GAAG,yBAAyB,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC;IAC/D,MAAM,MAAM,GAAwB;QAClC;YACE,KAAK,CAAC,WAAW,GAAG,CAAC,UAAU,CAAC,WAAW,IAAI,CAAC,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,cAAc;YAClF,sBAAsB,QAAQ,CAAC,MAAM,CAAC,cAAc,EAAE;SACvD;QACD;YACE,KAAK,CAAC,UAAU,GAAG,CAAC,UAAU,CAAC,UAAU,IAAI,CAAC,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,aAAa;YAC/E,qBAAqB,QAAQ,CAAC,MAAM,CAAC,aAAa,EAAE;SACrD;QACD;YACE,KAAK,CAAC,aAAa,GAAG,CAAC,UAAU,CAAC,aAAa,IAAI,CAAC,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,gBAAgB;YACxF,wBAAwB,QAAQ,CAAC,MAAM,CAAC,gBAAgB,EAAE;SAC3D;QACD;YACE,WAAW,CAAC,KAAK,CAAC,KAAK,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,SAAS;YACpD,gBAAgB,QAAQ,CAAC,MAAM,CAAC,SAAS,EAAE;SAC5C;QACD,CAAC,KAAK,CAAC,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,UAAU,EAAE,gBAAgB,QAAQ,CAAC,MAAM,CAAC,UAAU,EAAE,CAAC;QAC1F;YACE,CAAC,KAAK,CAAC,kBAAkB,IAAI,CAAC,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,qBAAqB;YACvE,8BAA8B,QAAQ,CAAC,MAAM,CAAC,qBAAqB,EAAE;SACtE;QACD;YACE,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,aAAa;YACvC,oBAAoB,QAAQ,CAAC,MAAM,CAAC,aAAa,IAAI;SACtD;KACF,CAAC;IACF,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,EAAE,EAAE,CAAC,MAAM,CAAC,CAAC;IACnD,IAAI,QAAQ;QAAE,MAAM,IAAI,KAAK,CAAC,2BAA2B,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AAC1E,CAAC,CAAC"}
|
|
@@ -0,0 +1,43 @@
|
|
|
1
|
+
import type { Registry } from "../core/index.js";
|
|
2
|
+
import type { ResolvedLuxConfig } from "./config.js";
|
|
3
|
+
import type { discoverCases } from "./discover.js";
|
|
4
|
+
import type { LoadedExperimentCampaign } from "./experiment-loader.js";
|
|
5
|
+
export type EvaluationCorpus = {
|
|
6
|
+
kind: "agent";
|
|
7
|
+
cases: Awaited<ReturnType<typeof discoverCases>>["cases"];
|
|
8
|
+
hash: string;
|
|
9
|
+
} | {
|
|
10
|
+
kind: "grader";
|
|
11
|
+
hash: string;
|
|
12
|
+
};
|
|
13
|
+
export type CampaignExecutionOptions = {
|
|
14
|
+
config: ResolvedLuxConfig;
|
|
15
|
+
registry: Registry;
|
|
16
|
+
loaded: LoadedExperimentCampaign;
|
|
17
|
+
runsRoot?: string;
|
|
18
|
+
resumeFrom?: string;
|
|
19
|
+
abortSignal?: AbortSignal;
|
|
20
|
+
onEvaluationUnit?: (unit: {
|
|
21
|
+
unitId: string;
|
|
22
|
+
candidateId: string;
|
|
23
|
+
profileId: string;
|
|
24
|
+
split: "train" | "validation" | "test";
|
|
25
|
+
}) => void | Promise<void>;
|
|
26
|
+
};
|
|
27
|
+
export declare const profileOrThrow: (config: ResolvedLuxConfig, profileId: string) => {
|
|
28
|
+
driver?: string | {
|
|
29
|
+
name: string;
|
|
30
|
+
config?: unknown;
|
|
31
|
+
} | undefined;
|
|
32
|
+
answerer?: string | {
|
|
33
|
+
name: string;
|
|
34
|
+
config?: unknown;
|
|
35
|
+
} | undefined;
|
|
36
|
+
harness?: {
|
|
37
|
+
provider: "claude-code" | "codex";
|
|
38
|
+
model: string;
|
|
39
|
+
command?: string | undefined;
|
|
40
|
+
reasoningEffort?: "high" | "low" | "medium" | "minimal" | "none" | "xhigh" | undefined;
|
|
41
|
+
} | undefined;
|
|
42
|
+
};
|
|
43
|
+
//# sourceMappingURL=experiment-context.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-context.d.ts","sourceRoot":"","sources":["../../src/orchestrator/experiment-context.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AACjD,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AACrD,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AACnD,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,wBAAwB,CAAC;AAEvE,MAAM,MAAM,gBAAgB,GACxB;IACE,IAAI,EAAE,OAAO,CAAC;IACd,KAAK,EAAE,OAAO,CAAC,UAAU,CAAC,OAAO,aAAa,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC;IAC1D,IAAI,EAAE,MAAM,CAAC;CACd,GACD;IACE,IAAI,EAAE,QAAQ,CAAC;IACf,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAEN,MAAM,MAAM,wBAAwB,GAAG;IACrC,MAAM,EAAE,iBAAiB,CAAC;IAC1B,QAAQ,EAAE,QAAQ,CAAC;IACnB,MAAM,EAAE,wBAAwB,CAAC;IACjC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,WAAW,CAAC,EAAE,WAAW,CAAC;IAC1B,gBAAgB,CAAC,EAAE,CAAC,IAAI,EAAE;QACxB,MAAM,EAAE,MAAM,CAAC;QACf,WAAW,EAAE,MAAM,CAAC;QACpB,SAAS,EAAE,MAAM,CAAC;QAClB,KAAK,EAAE,OAAO,GAAG,YAAY,GAAG,MAAM,CAAC;KACxC,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CAC5B,CAAC;AAEF,eAAO,MAAM,cAAc,WAAY,iBAAiB,aAAa,MAAM;;;;;;;;;;;;;;;CAS1E,CAAC"}
|
|
@@ -0,0 +1,9 @@
|
|
|
1
|
+
export const profileOrThrow = (config, profileId) => {
|
|
2
|
+
const profile = config.profiles?.[profileId];
|
|
3
|
+
if (!profile) {
|
|
4
|
+
const available = Object.keys(config.profiles ?? {}).sort();
|
|
5
|
+
throw new Error(`experiment profile '${profileId}' is not configured${available.length > 0 ? `; available: ${available.join(", ")}` : ""}`);
|
|
6
|
+
}
|
|
7
|
+
return profile;
|
|
8
|
+
};
|
|
9
|
+
//# sourceMappingURL=experiment-context.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-context.js","sourceRoot":"","sources":["../../src/orchestrator/experiment-context.ts"],"names":[],"mappings":"AA+BA,MAAM,CAAC,MAAM,cAAc,GAAG,CAAC,MAAyB,EAAE,SAAiB,EAAE,EAAE;IAC7E,MAAM,OAAO,GAAG,MAAM,CAAC,QAAQ,EAAE,CAAC,SAAS,CAAC,CAAC;IAC7C,IAAI,CAAC,OAAO,EAAE,CAAC;QACb,MAAM,SAAS,GAAG,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;QAC5D,MAAM,IAAI,KAAK,CACb,uBAAuB,SAAS,sBAAsB,SAAS,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,gBAAgB,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CAC3H,CAAC;IACJ,CAAC;IACD,OAAO,OAAO,CAAC;AACjB,CAAC,CAAC"}
|
|
@@ -0,0 +1,290 @@
|
|
|
1
|
+
import { z } from "zod";
|
|
2
|
+
export declare const CampaignExecutionResultSchema: z.ZodObject<{
|
|
3
|
+
experimentDir: z.ZodString;
|
|
4
|
+
champion: z.ZodNullable<z.ZodObject<{
|
|
5
|
+
schemaVersion: z.ZodLiteral<1>;
|
|
6
|
+
id: z.ZodString;
|
|
7
|
+
parentIds: z.ZodArray<z.ZodString>;
|
|
8
|
+
createdAt: z.ZodString;
|
|
9
|
+
proposer: z.ZodString;
|
|
10
|
+
baseSourceHash: z.ZodString;
|
|
11
|
+
candidateHash: z.ZodString;
|
|
12
|
+
components: z.ZodRecord<z.ZodString, z.ZodString>;
|
|
13
|
+
sourceBytes: z.ZodNumber;
|
|
14
|
+
rationale: z.ZodExactOptional<z.ZodString>;
|
|
15
|
+
proposalUsage: z.ZodExactOptional<z.ZodObject<{
|
|
16
|
+
input: z.ZodNumber;
|
|
17
|
+
output: z.ZodNumber;
|
|
18
|
+
cacheRead: z.ZodNumber;
|
|
19
|
+
cacheCreation: z.ZodNumber;
|
|
20
|
+
}, z.core.$strip>>;
|
|
21
|
+
}, z.core.$strip>>;
|
|
22
|
+
manifest: z.ZodObject<{
|
|
23
|
+
schemaVersion: z.ZodLiteral<1>;
|
|
24
|
+
experimentId: z.ZodString;
|
|
25
|
+
campaignId: z.ZodString;
|
|
26
|
+
campaignPath: z.ZodString;
|
|
27
|
+
campaignHash: z.ZodString;
|
|
28
|
+
campaignSnapshot: z.ZodObject<{
|
|
29
|
+
id: z.ZodString;
|
|
30
|
+
subject: z.ZodObject<{
|
|
31
|
+
root: z.ZodString;
|
|
32
|
+
mutable: z.ZodArray<z.ZodString>;
|
|
33
|
+
exclude: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
34
|
+
validate: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
35
|
+
command: z.ZodString;
|
|
36
|
+
args: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
37
|
+
timeoutMs: z.ZodDefault<z.ZodNumber>;
|
|
38
|
+
unsafeHostExecution: z.ZodLiteral<true>;
|
|
39
|
+
}, z.core.$strict>>>;
|
|
40
|
+
}, z.core.$strict>;
|
|
41
|
+
evaluation: z.ZodDefault<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
42
|
+
kind: z.ZodLiteral<"agent">;
|
|
43
|
+
}, z.core.$strict>, z.ZodObject<{
|
|
44
|
+
kind: z.ZodLiteral<"grader">;
|
|
45
|
+
annotations: z.ZodString;
|
|
46
|
+
allowedAssertionAdditions: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodArray<z.ZodString>>>;
|
|
47
|
+
scoreTolerance: z.ZodDefault<z.ZodNumber>;
|
|
48
|
+
}, z.core.$strict>], "kind">>;
|
|
49
|
+
profiles: z.ZodArray<z.ZodString>;
|
|
50
|
+
splits: z.ZodObject<{
|
|
51
|
+
train: z.ZodOptional<z.ZodObject<{
|
|
52
|
+
patterns: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
53
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
54
|
+
}, z.core.$strict>>;
|
|
55
|
+
validation: z.ZodObject<{
|
|
56
|
+
patterns: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
57
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
58
|
+
}, z.core.$strict>;
|
|
59
|
+
test: z.ZodOptional<z.ZodObject<{
|
|
60
|
+
patterns: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
61
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
62
|
+
}, z.core.$strict>>;
|
|
63
|
+
}, z.core.$strict>;
|
|
64
|
+
repetitions: z.ZodDefault<z.ZodObject<{
|
|
65
|
+
search: z.ZodDefault<z.ZodNumber>;
|
|
66
|
+
promotion: z.ZodDefault<z.ZodNumber>;
|
|
67
|
+
}, z.core.$strict>>;
|
|
68
|
+
variants: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
69
|
+
id: z.ZodString;
|
|
70
|
+
name: z.ZodOptional<z.ZodString>;
|
|
71
|
+
parentIds: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
72
|
+
components: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodString>>;
|
|
73
|
+
}, z.core.$strict>>>;
|
|
74
|
+
objectives: z.ZodDefault<z.ZodObject<{
|
|
75
|
+
constraints: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
76
|
+
metric: z.ZodString;
|
|
77
|
+
min: z.ZodOptional<z.ZodNumber>;
|
|
78
|
+
max: z.ZodOptional<z.ZodNumber>;
|
|
79
|
+
profile: z.ZodOptional<z.ZodString>;
|
|
80
|
+
split: z.ZodDefault<z.ZodEnum<{
|
|
81
|
+
test: "test";
|
|
82
|
+
train: "train";
|
|
83
|
+
validation: "validation";
|
|
84
|
+
}>>;
|
|
85
|
+
confidence: z.ZodOptional<z.ZodObject<{
|
|
86
|
+
method: z.ZodEnum<{
|
|
87
|
+
hoeffding: "hoeffding";
|
|
88
|
+
"percentile-bootstrap": "percentile-bootstrap";
|
|
89
|
+
wilson: "wilson";
|
|
90
|
+
}>;
|
|
91
|
+
level: z.ZodDefault<z.ZodNumber>;
|
|
92
|
+
resamples: z.ZodDefault<z.ZodNumber>;
|
|
93
|
+
seed: z.ZodDefault<z.ZodNumber>;
|
|
94
|
+
minSamples: z.ZodDefault<z.ZodNumber>;
|
|
95
|
+
}, z.core.$strict>>;
|
|
96
|
+
}, z.core.$strict>>>;
|
|
97
|
+
minimize: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
98
|
+
metric: z.ZodString;
|
|
99
|
+
profile: z.ZodOptional<z.ZodString>;
|
|
100
|
+
split: z.ZodDefault<z.ZodEnum<{
|
|
101
|
+
test: "test";
|
|
102
|
+
train: "train";
|
|
103
|
+
validation: "validation";
|
|
104
|
+
}>>;
|
|
105
|
+
}, z.core.$strict>>>;
|
|
106
|
+
}, z.core.$strict>>;
|
|
107
|
+
optimizer: z.ZodOptional<z.ZodObject<{
|
|
108
|
+
algorithm: z.ZodLiteral<"reflective">;
|
|
109
|
+
config: z.ZodOptional<z.ZodUnknown>;
|
|
110
|
+
candidateSelection: z.ZodDefault<z.ZodEnum<{
|
|
111
|
+
"current-best": "current-best";
|
|
112
|
+
pareto: "pareto";
|
|
113
|
+
}>>;
|
|
114
|
+
mergeFrontier: z.ZodDefault<z.ZodBoolean>;
|
|
115
|
+
promotionTopK: z.ZodDefault<z.ZodNumber>;
|
|
116
|
+
}, z.core.$strict>>;
|
|
117
|
+
budget: z.ZodDefault<z.ZodObject<{
|
|
118
|
+
maxCandidates: z.ZodDefault<z.ZodNumber>;
|
|
119
|
+
patience: z.ZodDefault<z.ZodNumber>;
|
|
120
|
+
maxMetricCalls: z.ZodDefault<z.ZodNumber>;
|
|
121
|
+
maxJudgeCalls: z.ZodDefault<z.ZodNumber>;
|
|
122
|
+
maxProposalCalls: z.ZodDefault<z.ZodNumber>;
|
|
123
|
+
maxTokens: z.ZodDefault<z.ZodNumber>;
|
|
124
|
+
maxCostUsd: z.ZodDefault<z.ZodNumber>;
|
|
125
|
+
maxUnpricedModelCalls: z.ZodDefault<z.ZodNumber>;
|
|
126
|
+
maxWallTimeMs: z.ZodDefault<z.ZodNumber>;
|
|
127
|
+
}, z.core.$strict>>;
|
|
128
|
+
meta: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
129
|
+
}, z.core.$strict>;
|
|
130
|
+
resolvedProfiles: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
131
|
+
evaluatorVersion: z.ZodLiteral<1>;
|
|
132
|
+
luxVersion: z.ZodString;
|
|
133
|
+
evaluationKind: z.ZodDefault<z.ZodEnum<{
|
|
134
|
+
agent: "agent";
|
|
135
|
+
grader: "grader";
|
|
136
|
+
}>>;
|
|
137
|
+
evaluationHash: z.ZodExactOptional<z.ZodString>;
|
|
138
|
+
mode: z.ZodEnum<{
|
|
139
|
+
experiment: "experiment";
|
|
140
|
+
optimize: "optimize";
|
|
141
|
+
}>;
|
|
142
|
+
status: z.ZodEnum<{
|
|
143
|
+
aborted: "aborted";
|
|
144
|
+
failed: "failed";
|
|
145
|
+
passed: "passed";
|
|
146
|
+
running: "running";
|
|
147
|
+
}>;
|
|
148
|
+
startedAt: z.ZodString;
|
|
149
|
+
endedAt: z.ZodNullable<z.ZodString>;
|
|
150
|
+
activeTimeMs: z.ZodDefault<z.ZodNumber>;
|
|
151
|
+
activeSegmentStartedAt: z.ZodExactOptional<z.ZodString>;
|
|
152
|
+
subjectRoot: z.ZodString;
|
|
153
|
+
baseSourceHash: z.ZodString;
|
|
154
|
+
candidateIntegrityVersion: z.ZodExactOptional<z.ZodLiteral<1>>;
|
|
155
|
+
candidates: z.ZodArray<z.ZodObject<{
|
|
156
|
+
id: z.ZodString;
|
|
157
|
+
path: z.ZodString;
|
|
158
|
+
candidateHash: z.ZodString;
|
|
159
|
+
sourceBytes: z.ZodNumber;
|
|
160
|
+
candidateEnvelopeHash: z.ZodOptional<z.ZodString>;
|
|
161
|
+
}, z.core.$strip>>;
|
|
162
|
+
evaluations: z.ZodArray<z.ZodObject<{
|
|
163
|
+
profileId: z.ZodString;
|
|
164
|
+
split: z.ZodEnum<{
|
|
165
|
+
test: "test";
|
|
166
|
+
train: "train";
|
|
167
|
+
validation: "validation";
|
|
168
|
+
}>;
|
|
169
|
+
suiteDir: z.ZodExactOptional<z.ZodString>;
|
|
170
|
+
passRate: z.ZodNumber;
|
|
171
|
+
casePassRates: z.ZodRecord<z.ZodString, z.ZodNumber>;
|
|
172
|
+
metrics: z.ZodRecord<z.ZodString, z.ZodNumber>;
|
|
173
|
+
metricCalls: z.ZodExactOptional<z.ZodNumber>;
|
|
174
|
+
judgeCalls: z.ZodExactOptional<z.ZodNumber>;
|
|
175
|
+
unpricedModelCalls: z.ZodExactOptional<z.ZodNumber>;
|
|
176
|
+
usage: z.ZodExactOptional<z.ZodObject<{
|
|
177
|
+
input: z.ZodNumber;
|
|
178
|
+
output: z.ZodNumber;
|
|
179
|
+
cacheRead: z.ZodNumber;
|
|
180
|
+
cacheCreation: z.ZodNumber;
|
|
181
|
+
}, z.core.$strip>>;
|
|
182
|
+
costUsd: z.ZodExactOptional<z.ZodNumber>;
|
|
183
|
+
samples: z.ZodExactOptional<z.ZodRecord<z.ZodString, z.ZodArray<z.ZodNumber>>>;
|
|
184
|
+
failedCases: z.ZodArray<z.ZodObject<{
|
|
185
|
+
caseId: z.ZodString;
|
|
186
|
+
score: z.ZodNullable<z.ZodNumber>;
|
|
187
|
+
feedback: z.ZodArray<z.ZodString>;
|
|
188
|
+
}, z.core.$strip>>;
|
|
189
|
+
candidateId: z.ZodString;
|
|
190
|
+
repetitions: z.ZodNumber;
|
|
191
|
+
summaryPath: z.ZodExactOptional<z.ZodString>;
|
|
192
|
+
evidencePath: z.ZodExactOptional<z.ZodString>;
|
|
193
|
+
}, z.core.$strip>>;
|
|
194
|
+
accountedEvaluationUnits: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
195
|
+
accountingJournalEntries: z.ZodDefault<z.ZodNumber>;
|
|
196
|
+
budgetUsage: z.ZodObject<{
|
|
197
|
+
metricCalls: z.ZodNumber;
|
|
198
|
+
judgeCalls: z.ZodNumber;
|
|
199
|
+
unpricedModelCalls: z.ZodOptional<z.ZodNumber>;
|
|
200
|
+
proposalCalls: z.ZodNumber;
|
|
201
|
+
usage: z.ZodObject<{
|
|
202
|
+
input: z.ZodNumber;
|
|
203
|
+
output: z.ZodNumber;
|
|
204
|
+
cacheRead: z.ZodNumber;
|
|
205
|
+
cacheCreation: z.ZodNumber;
|
|
206
|
+
}, z.core.$strip>;
|
|
207
|
+
costUsd: z.ZodNumber;
|
|
208
|
+
}, z.core.$strip>;
|
|
209
|
+
championCandidateId: z.ZodExactOptional<z.ZodString>;
|
|
210
|
+
selection: z.ZodExactOptional<z.ZodObject<{
|
|
211
|
+
decidedAt: z.ZodString;
|
|
212
|
+
decisionCandidateIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
213
|
+
promotionFinalistCandidateIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
214
|
+
rankedCandidateIds: z.ZodArray<z.ZodString>;
|
|
215
|
+
eligibleCandidateIds: z.ZodArray<z.ZodString>;
|
|
216
|
+
championCandidateId: z.ZodExactOptional<z.ZodString>;
|
|
217
|
+
constraints: z.ZodArray<z.ZodObject<{
|
|
218
|
+
candidateId: z.ZodString;
|
|
219
|
+
metric: z.ZodString;
|
|
220
|
+
profile: z.ZodExactOptional<z.ZodString>;
|
|
221
|
+
split: z.ZodEnum<{
|
|
222
|
+
test: "test";
|
|
223
|
+
train: "train";
|
|
224
|
+
validation: "validation";
|
|
225
|
+
}>;
|
|
226
|
+
values: z.ZodArray<z.ZodNumber>;
|
|
227
|
+
passed: z.ZodBoolean;
|
|
228
|
+
interval: z.ZodExactOptional<z.ZodObject<{
|
|
229
|
+
method: z.ZodEnum<{
|
|
230
|
+
hoeffding: "hoeffding";
|
|
231
|
+
"percentile-bootstrap": "percentile-bootstrap";
|
|
232
|
+
wilson: "wilson";
|
|
233
|
+
}>;
|
|
234
|
+
level: z.ZodNumber;
|
|
235
|
+
estimate: z.ZodNumber;
|
|
236
|
+
lower: z.ZodNumber;
|
|
237
|
+
upper: z.ZodNumber;
|
|
238
|
+
n: z.ZodNumber;
|
|
239
|
+
resamples: z.ZodNumber;
|
|
240
|
+
}, z.core.$strict>>;
|
|
241
|
+
note: z.ZodExactOptional<z.ZodString>;
|
|
242
|
+
}, z.core.$strip>>;
|
|
243
|
+
}, z.core.$strip>>;
|
|
244
|
+
errorMessage: z.ZodExactOptional<z.ZodString>;
|
|
245
|
+
}, z.core.$strip>;
|
|
246
|
+
}, z.core.$strict>;
|
|
247
|
+
export type CampaignExecutionResult = z.infer<typeof CampaignExecutionResultSchema>;
|
|
248
|
+
export declare const ExperimentCampaignPlanSchema: z.ZodObject<{
|
|
249
|
+
campaignId: z.ZodString;
|
|
250
|
+
mode: z.ZodEnum<{
|
|
251
|
+
experiment: "experiment";
|
|
252
|
+
optimize: "optimize";
|
|
253
|
+
}>;
|
|
254
|
+
evaluationKind: z.ZodEnum<{
|
|
255
|
+
agent: "agent";
|
|
256
|
+
grader: "grader";
|
|
257
|
+
}>;
|
|
258
|
+
subjectRoot: z.ZodString;
|
|
259
|
+
mutableComponents: z.ZodArray<z.ZodString>;
|
|
260
|
+
profiles: z.ZodArray<z.ZodString>;
|
|
261
|
+
splits: z.ZodArray<z.ZodObject<{
|
|
262
|
+
split: z.ZodEnum<{
|
|
263
|
+
test: "test";
|
|
264
|
+
train: "train";
|
|
265
|
+
validation: "validation";
|
|
266
|
+
}>;
|
|
267
|
+
examples: z.ZodNumber;
|
|
268
|
+
}, z.core.$strict>>;
|
|
269
|
+
projected: z.ZodObject<{
|
|
270
|
+
candidates: z.ZodNumber;
|
|
271
|
+
metricCalls: z.ZodNumber;
|
|
272
|
+
judgeCalls: z.ZodNumber;
|
|
273
|
+
proposalCalls: z.ZodNumber;
|
|
274
|
+
}, z.core.$strict>;
|
|
275
|
+
budget: z.ZodObject<{
|
|
276
|
+
maxCandidates: z.ZodDefault<z.ZodNumber>;
|
|
277
|
+
patience: z.ZodDefault<z.ZodNumber>;
|
|
278
|
+
maxMetricCalls: z.ZodDefault<z.ZodNumber>;
|
|
279
|
+
maxJudgeCalls: z.ZodDefault<z.ZodNumber>;
|
|
280
|
+
maxProposalCalls: z.ZodDefault<z.ZodNumber>;
|
|
281
|
+
maxTokens: z.ZodDefault<z.ZodNumber>;
|
|
282
|
+
maxCostUsd: z.ZodDefault<z.ZodNumber>;
|
|
283
|
+
maxUnpricedModelCalls: z.ZodDefault<z.ZodNumber>;
|
|
284
|
+
maxWallTimeMs: z.ZodDefault<z.ZodNumber>;
|
|
285
|
+
}, z.core.$strict>;
|
|
286
|
+
fitsProjectedCallBudgets: z.ZodBoolean;
|
|
287
|
+
warnings: z.ZodArray<z.ZodString>;
|
|
288
|
+
}, z.core.$strict>;
|
|
289
|
+
export type ExperimentCampaignPlan = z.infer<typeof ExperimentCampaignPlanSchema>;
|
|
290
|
+
//# sourceMappingURL=experiment-contracts.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-contracts.d.ts","sourceRoot":"","sources":["../../src/orchestrator/experiment-contracts.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAIxB,eAAO,MAAM,6BAA6B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAM/B,CAAC;AACZ,MAAM,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,6BAA6B,CAAC,CAAC;AAEpF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBA4B9B,CAAC;AACZ,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC"}
|
|
@@ -0,0 +1,38 @@
|
|
|
1
|
+
import { z } from "zod";
|
|
2
|
+
import { CandidateSchema, ExperimentBudgetSchema } from "../core/index.js";
|
|
3
|
+
import { ExperimentRunSchema } from "./experiment-runs.js";
|
|
4
|
+
export const CampaignExecutionResultSchema = z
|
|
5
|
+
.object({
|
|
6
|
+
experimentDir: z.string(),
|
|
7
|
+
champion: CandidateSchema.nullable(),
|
|
8
|
+
manifest: ExperimentRunSchema,
|
|
9
|
+
})
|
|
10
|
+
.strict();
|
|
11
|
+
export const ExperimentCampaignPlanSchema = z
|
|
12
|
+
.object({
|
|
13
|
+
campaignId: z.string(),
|
|
14
|
+
mode: z.enum(["experiment", "optimize"]),
|
|
15
|
+
evaluationKind: z.enum(["agent", "grader"]),
|
|
16
|
+
subjectRoot: z.string(),
|
|
17
|
+
mutableComponents: z.array(z.string()),
|
|
18
|
+
profiles: z.array(z.string()),
|
|
19
|
+
splits: z.array(z
|
|
20
|
+
.object({
|
|
21
|
+
split: z.enum(["train", "validation", "test"]),
|
|
22
|
+
examples: z.number().int().nonnegative(),
|
|
23
|
+
})
|
|
24
|
+
.strict()),
|
|
25
|
+
projected: z
|
|
26
|
+
.object({
|
|
27
|
+
candidates: z.number().int().nonnegative(),
|
|
28
|
+
metricCalls: z.number().int().nonnegative(),
|
|
29
|
+
judgeCalls: z.number().int().nonnegative(),
|
|
30
|
+
proposalCalls: z.number().int().nonnegative(),
|
|
31
|
+
})
|
|
32
|
+
.strict(),
|
|
33
|
+
budget: ExperimentBudgetSchema,
|
|
34
|
+
fitsProjectedCallBudgets: z.boolean(),
|
|
35
|
+
warnings: z.array(z.string()),
|
|
36
|
+
})
|
|
37
|
+
.strict();
|
|
38
|
+
//# sourceMappingURL=experiment-contracts.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-contracts.js","sourceRoot":"","sources":["../../src/orchestrator/experiment-contracts.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAAE,eAAe,EAAE,sBAAsB,EAAE,MAAM,kBAAkB,CAAC;AAC3E,OAAO,EAAE,mBAAmB,EAAE,MAAM,sBAAsB,CAAC;AAE3D,MAAM,CAAC,MAAM,6BAA6B,GAAG,CAAC;KAC3C,MAAM,CAAC;IACN,aAAa,EAAE,CAAC,CAAC,MAAM,EAAE;IACzB,QAAQ,EAAE,eAAe,CAAC,QAAQ,EAAE;IACpC,QAAQ,EAAE,mBAAmB;CAC9B,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,CAAC,MAAM,4BAA4B,GAAG,CAAC;KAC1C,MAAM,CAAC;IACN,UAAU,EAAE,CAAC,CAAC,MAAM,EAAE;IACtB,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC;IACxC,cAAc,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAC;IAC3C,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE;IACvB,iBAAiB,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC;IACtC,QAAQ,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC;IAC7B,MAAM,EAAE,CAAC,CAAC,KAAK,CACb,CAAC;SACE,MAAM,CAAC;QACN,KAAK,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,YAAY,EAAE,MAAM,CAAC,CAAC;QAC9C,QAAQ,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE;KACzC,CAAC;SACD,MAAM,EAAE,CACZ;IACD,SAAS,EAAE,CAAC;SACT,MAAM,CAAC;QACN,UAAU,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE;QAC1C,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE;QAC3C,UAAU,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE;QAC1C,aAAa,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE;KAC9C,CAAC;SACD,MAAM,EAAE;IACX,MAAM,EAAE,sBAAsB;IAC9B,wBAAwB,EAAE,CAAC,CAAC,OAAO,EAAE;IACrC,QAAQ,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC;CAC9B,CAAC;KACD,MAAM,EAAE,CAAC"}
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
import { type Candidate, type ExperimentCampaign, type ExperimentSelector } from "../core/index.js";
|
|
2
|
+
import { discoverCases } from "./discover.js";
|
|
3
|
+
import type { CampaignExecutionOptions, EvaluationCorpus } from "./experiment-context.js";
|
|
4
|
+
import { loadEvalCase } from "./loader.js";
|
|
5
|
+
type ExperimentSplit = "train" | "validation" | "test";
|
|
6
|
+
export declare const selectorsBySplit: (campaign: ExperimentCampaign) => Array<{
|
|
7
|
+
name: ExperimentSplit;
|
|
8
|
+
selector: ExperimentSelector;
|
|
9
|
+
}>;
|
|
10
|
+
export declare const caseRubricCliReferences: (execution: CampaignExecutionOptions, authored: Awaited<ReturnType<typeof loadEvalCase>>) => Array<{
|
|
11
|
+
assertionId: string;
|
|
12
|
+
profileId: string;
|
|
13
|
+
provider: "claude-code" | "codex";
|
|
14
|
+
command: string;
|
|
15
|
+
}>;
|
|
16
|
+
export declare const agentCorpusHash: (execution: CampaignExecutionOptions, cases: Awaited<ReturnType<typeof discoverCases>>["cases"], splits: ReturnType<typeof selectorsBySplit>) => Promise<string>;
|
|
17
|
+
export declare const graderCorpusHash: (execution: CampaignExecutionOptions, splits: ReturnType<typeof selectorsBySplit>) => Promise<string>;
|
|
18
|
+
export declare const prepareCorpus: (execution: CampaignExecutionOptions) => Promise<EvaluationCorpus>;
|
|
19
|
+
export declare const assertGraderSubject: (execution: CampaignExecutionOptions, baseline: Candidate) => Promise<void>;
|
|
20
|
+
export {};
|
|
21
|
+
//# sourceMappingURL=experiment-corpus.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-corpus.d.ts","sourceRoot":"","sources":["../../src/orchestrator/experiment-corpus.ts"],"names":[],"mappings":"AAGA,OAAO,EAEL,KAAK,SAAS,EAEd,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EAExB,MAAM,kBAAkB,CAAC;AAE1B,OAAO,EAAE,aAAa,EAAe,MAAM,eAAe,CAAC;AAC3D,OAAO,KAAK,EAAE,wBAAwB,EAAE,gBAAgB,EAAE,MAAM,yBAAyB,CAAC;AAW1F,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAK3C,KAAK,eAAe,GAAG,OAAO,GAAG,YAAY,GAAG,MAAM,CAAC;AAEvD,eAAO,MAAM,gBAAgB,aACjB,kBAAkB,KAC3B,KAAK,CAAC;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,QAAQ,EAAE,kBAAkB,CAAA;CAAE,CAI/D,CAAC;AA0BF,eAAO,MAAM,uBAAuB,cACvB,wBAAwB,YACzB,OAAO,CAAC,UAAU,CAAC,OAAO,YAAY,CAAC,CAAC,KACjD,KAAK,CAAC;IACP,WAAW,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,MAAM,CAAC;IAClB,QAAQ,EAAE,aAAa,GAAG,OAAO,CAAC;IAClC,OAAO,EAAE,MAAM,CAAC;CACjB,CAsBA,CAAC;AAoBF,eAAO,MAAM,eAAe,cACf,wBAAwB,SAC5B,OAAO,CAAC,UAAU,CAAC,OAAO,aAAa,CAAC,CAAC,CAAC,OAAO,CAAC,UACjD,UAAU,CAAC,OAAO,gBAAgB,CAAC,KAC1C,OAAO,CAAC,MAAM,CAsEhB,CAAC;AAEF,eAAO,MAAM,gBAAgB,cAChB,wBAAwB,UAC3B,UAAU,CAAC,OAAO,gBAAgB,CAAC,KAC1C,OAAO,CAAC,MAAM,CAgFhB,CAAC;AAEF,eAAO,MAAM,aAAa,cACb,wBAAwB,KAClC,OAAO,CAAC,gBAAgB,CA6E1B,CAAC;AAEF,eAAO,MAAM,mBAAmB,cACnB,wBAAwB,YACzB,SAAS,KAClB,OAAO,CAAC,IAAI,CAyDd,CAAC"}
|