@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 +1,257 @@
|
|
|
1
|
+
import { readFile } from "node:fs/promises";
|
|
2
|
+
import { isAbsolute, relative, resolve, sep } from "node:path";
|
|
3
|
+
import { bindSubjectRoot, contentHash, valueHash } from "../core/index.js";
|
|
4
|
+
import { ReflectiveOptimizerConfigSchema } from "../optimization/reflective.js";
|
|
5
|
+
import { authoredDependencyGraphHash } from "./authored-dependencies.js";
|
|
6
|
+
import { assertValidationArgumentsBound } from "./candidates.js";
|
|
7
|
+
import { selectCases } from "./discover.js";
|
|
8
|
+
import { evaluatorImplementationHash } from "./evaluator-identity.js";
|
|
9
|
+
import { profileOrThrow } from "./experiment-context.js";
|
|
10
|
+
import { caseRubricCliReferences, selectorsBySplit } from "./experiment-corpus.js";
|
|
11
|
+
import { claudeConfigRoot, claudeUserBehaviorHash, cliIdentity, configuredClaudeSourceHashes, LUX_VERSION, RuntimeCliReferenceSchema, referencedAnswererCli, referencedCli, runtimeCliIdentityHashForReferences, } from "./experiment-runtime-identity.js";
|
|
12
|
+
import { loadEvalCase } from "./loader.js";
|
|
13
|
+
import { boundExperimentDriverConfig, resolveAnswererReference, resolveDriverReference, } from "./orchestrator.js";
|
|
14
|
+
import { runtimeIdentity } from "./provenance.js";
|
|
15
|
+
import { validatePluginConfig } from "./validate-plugin-config.js";
|
|
16
|
+
export const resolvedProfileSnapshot = async (execution) => {
|
|
17
|
+
const behaviorHashes = new Map();
|
|
18
|
+
const configuredSourceHashes = new Map();
|
|
19
|
+
const behaviorHash = (root) => {
|
|
20
|
+
const existing = behaviorHashes.get(root);
|
|
21
|
+
if (existing)
|
|
22
|
+
return existing;
|
|
23
|
+
const created = claudeUserBehaviorHash(root);
|
|
24
|
+
behaviorHashes.set(root, created);
|
|
25
|
+
return created;
|
|
26
|
+
};
|
|
27
|
+
return Object.fromEntries(await Promise.all(execution.loaded.campaign.profiles.map(async (profileId) => {
|
|
28
|
+
const profile = profileOrThrow(execution.config, profileId);
|
|
29
|
+
// Offline grader campaigns never launch a subject driver. Keep their
|
|
30
|
+
// identity independent of irrelevant local agent settings/plugins
|
|
31
|
+
// while still snapshotting the harness that performs regrading.
|
|
32
|
+
const driver = execution.loaded.campaign.evaluation.kind === "agent"
|
|
33
|
+
? (profile.driver ?? execution.config.defaultDriver)
|
|
34
|
+
: undefined;
|
|
35
|
+
const driverCli = referencedCli(driver);
|
|
36
|
+
const harness = profile.harness ?? execution.config.harness;
|
|
37
|
+
const answerer = execution.loaded.campaign.evaluation.kind === "agent"
|
|
38
|
+
? (profile.answerer ?? execution.config.defaultAnswerer)
|
|
39
|
+
: undefined;
|
|
40
|
+
const answererCli = referencedAnswererCli(answerer);
|
|
41
|
+
const driverUserBehaviorHash = driverCli?.name === "claude-code" ? await behaviorHash(claudeConfigRoot(driver)) : null;
|
|
42
|
+
const driverConfiguredSources = await configuredClaudeSourceHashes(driver, execution.config.rootDir, configuredSourceHashes);
|
|
43
|
+
let harnessCommand = harness?.command;
|
|
44
|
+
if (!harnessCommand) {
|
|
45
|
+
harnessCommand = harness?.provider === "codex" ? "codex" : "claude";
|
|
46
|
+
}
|
|
47
|
+
const capturesHarnessIdentity = Boolean(harness) || execution.loaded.campaign.evaluation.kind === "grader";
|
|
48
|
+
const harnessCliIdentity = capturesHarnessIdentity
|
|
49
|
+
? await cliIdentity(harnessCommand, execution.config.rootDir)
|
|
50
|
+
: null;
|
|
51
|
+
return [
|
|
52
|
+
profileId,
|
|
53
|
+
{
|
|
54
|
+
id: profileId,
|
|
55
|
+
driver: driver ?? null,
|
|
56
|
+
driverCliIdentity: driverCli
|
|
57
|
+
? await cliIdentity(driverCli.command, execution.config.rootDir)
|
|
58
|
+
: null,
|
|
59
|
+
...(driverCli?.name === "claude-code"
|
|
60
|
+
? {
|
|
61
|
+
driverClaudeConfigRoot: claudeConfigRoot(driver),
|
|
62
|
+
driverUserBehaviorHash,
|
|
63
|
+
driverConfiguredSources,
|
|
64
|
+
}
|
|
65
|
+
: {}),
|
|
66
|
+
answerer,
|
|
67
|
+
answererCliIdentity: answererCli
|
|
68
|
+
? await cliIdentity(answererCli.command, execution.config.rootDir)
|
|
69
|
+
: null,
|
|
70
|
+
harness: harness ?? null,
|
|
71
|
+
harnessCliIdentity,
|
|
72
|
+
},
|
|
73
|
+
];
|
|
74
|
+
})));
|
|
75
|
+
};
|
|
76
|
+
const trustedExecutableEntrypoints = async (execution, corpus) => {
|
|
77
|
+
const paths = new Set();
|
|
78
|
+
const add = (command) => {
|
|
79
|
+
if (!command || !isAbsolute(command))
|
|
80
|
+
return;
|
|
81
|
+
const fromRoot = relative(execution.config.rootDir, resolve(command));
|
|
82
|
+
if (fromRoot !== "" &&
|
|
83
|
+
fromRoot !== ".." &&
|
|
84
|
+
!fromRoot.startsWith(`..${sep}`) &&
|
|
85
|
+
!isAbsolute(fromRoot)) {
|
|
86
|
+
paths.add(fromRoot);
|
|
87
|
+
}
|
|
88
|
+
};
|
|
89
|
+
for (const validation of execution.loaded.campaign.subject.validate)
|
|
90
|
+
add(validation.command);
|
|
91
|
+
if (execution.loaded.campaign.optimizer) {
|
|
92
|
+
const optimizer = ReflectiveOptimizerConfigSchema.parse(execution.loaded.campaign.optimizer.config ?? {});
|
|
93
|
+
add(optimizer.command);
|
|
94
|
+
}
|
|
95
|
+
for (const profileId of execution.loaded.campaign.profiles) {
|
|
96
|
+
const profile = profileOrThrow(execution.config, profileId);
|
|
97
|
+
add(referencedCli(profile.driver ?? execution.config.defaultDriver)?.command);
|
|
98
|
+
add(referencedAnswererCli(profile.answerer ?? execution.config.defaultAnswerer)?.command);
|
|
99
|
+
add((profile.harness ?? execution.config.harness)?.command);
|
|
100
|
+
}
|
|
101
|
+
if (corpus.kind === "agent") {
|
|
102
|
+
const selected = [
|
|
103
|
+
...new Map(selectorsBySplit(execution.loaded.campaign)
|
|
104
|
+
.flatMap((split) => selectCases(corpus.cases, split.selector))
|
|
105
|
+
.map((evalCase) => [evalCase.path, evalCase])).values(),
|
|
106
|
+
];
|
|
107
|
+
for (const evalCase of selected) {
|
|
108
|
+
const authored = bindSubjectRoot(await loadEvalCase(evalCase.path), execution.loaded.subjectRoot);
|
|
109
|
+
add(referencedCli(authored.driver)?.command);
|
|
110
|
+
add(referencedAnswererCli(authored.answerer)?.command);
|
|
111
|
+
for (const reference of caseRubricCliReferences(execution, authored)) {
|
|
112
|
+
add(reference.command);
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
}
|
|
116
|
+
return paths;
|
|
117
|
+
};
|
|
118
|
+
export const runtimeCliReferences = async (execution, corpus, mode) => {
|
|
119
|
+
const references = new Map();
|
|
120
|
+
const add = (command, baseRoot) => {
|
|
121
|
+
if (!command)
|
|
122
|
+
return;
|
|
123
|
+
const reference = RuntimeCliReferenceSchema.parse({ command, baseRoot: resolve(baseRoot) });
|
|
124
|
+
references.set(`${reference.command}\u0000${reference.baseRoot}`, reference);
|
|
125
|
+
};
|
|
126
|
+
for (const validation of execution.loaded.campaign.subject.validate) {
|
|
127
|
+
add(validation.command, execution.loaded.subjectRoot);
|
|
128
|
+
}
|
|
129
|
+
if (mode === "optimize" && execution.loaded.campaign.optimizer) {
|
|
130
|
+
const optimizer = ReflectiveOptimizerConfigSchema.parse(execution.loaded.campaign.optimizer.config ?? {});
|
|
131
|
+
add(optimizer.command ?? (optimizer.provider === "codex" ? "codex" : "claude"), execution.loaded.directory);
|
|
132
|
+
}
|
|
133
|
+
if (corpus.kind === "grader") {
|
|
134
|
+
for (const profileId of execution.loaded.campaign.profiles) {
|
|
135
|
+
const profile = profileOrThrow(execution.config, profileId);
|
|
136
|
+
const harness = profile.harness ?? execution.config.harness;
|
|
137
|
+
add(harness?.command ?? (harness?.provider === "codex" ? "codex" : "claude"), execution.config.rootDir);
|
|
138
|
+
}
|
|
139
|
+
}
|
|
140
|
+
if (corpus.kind === "agent") {
|
|
141
|
+
const selected = [
|
|
142
|
+
...new Map(selectorsBySplit(execution.loaded.campaign)
|
|
143
|
+
.flatMap((split) => selectCases(corpus.cases, split.selector))
|
|
144
|
+
.map((evalCase) => [evalCase.path, evalCase])).values(),
|
|
145
|
+
];
|
|
146
|
+
for (const evalCase of selected) {
|
|
147
|
+
const authored = bindSubjectRoot(await loadEvalCase(evalCase.path), execution.loaded.subjectRoot);
|
|
148
|
+
for (const profileId of execution.loaded.campaign.profiles) {
|
|
149
|
+
const profile = profileOrThrow(execution.config, profileId);
|
|
150
|
+
const harness = profile.harness ?? execution.config.harness;
|
|
151
|
+
const driver = resolveDriverReference(authored, profile.driver ?? execution.config.defaultDriver);
|
|
152
|
+
const answerer = resolveAnswererReference(authored, profile.answerer ?? execution.config.defaultAnswerer, harness);
|
|
153
|
+
add(referencedCli(driver)?.command, execution.config.rootDir);
|
|
154
|
+
add(referencedAnswererCli(answerer)?.command, execution.config.rootDir);
|
|
155
|
+
}
|
|
156
|
+
for (const reference of caseRubricCliReferences(execution, authored)) {
|
|
157
|
+
add(reference.command, execution.config.rootDir);
|
|
158
|
+
}
|
|
159
|
+
}
|
|
160
|
+
}
|
|
161
|
+
return [...references.values()].sort((left, right) => {
|
|
162
|
+
const commandOrder = left.command.localeCompare(right.command);
|
|
163
|
+
return commandOrder || left.baseRoot.localeCompare(right.baseRoot);
|
|
164
|
+
});
|
|
165
|
+
};
|
|
166
|
+
const runtimeCliIdentityHash = async (execution, corpus, mode) => runtimeCliIdentityHashForReferences(await runtimeCliReferences(execution, corpus, mode));
|
|
167
|
+
const authoredConfigDependencyHash = async (execution, corpus) => {
|
|
168
|
+
const configRoot = resolve(execution.config.rootDir);
|
|
169
|
+
const runsRoot = resolve(execution.runsRoot ?? execution.config.runsRoot);
|
|
170
|
+
const runsFromConfig = relative(configRoot, runsRoot);
|
|
171
|
+
if (runsFromConfig === "") {
|
|
172
|
+
throw new Error("experiment runsRoot cannot be the Lux config root");
|
|
173
|
+
}
|
|
174
|
+
const entrypoints = await trustedExecutableEntrypoints(execution, corpus);
|
|
175
|
+
const add = (path) => {
|
|
176
|
+
if (path)
|
|
177
|
+
entrypoints.add(relative(configRoot, resolve(path)));
|
|
178
|
+
};
|
|
179
|
+
add(execution.config.configPath);
|
|
180
|
+
add(execution.loaded.path);
|
|
181
|
+
if (corpus.kind === "agent") {
|
|
182
|
+
for (const split of selectorsBySplit(execution.loaded.campaign)) {
|
|
183
|
+
for (const evalCase of selectCases(corpus.cases, split.selector))
|
|
184
|
+
add(evalCase.path);
|
|
185
|
+
}
|
|
186
|
+
}
|
|
187
|
+
else {
|
|
188
|
+
add(execution.loaded.graderAnnotations?.path);
|
|
189
|
+
}
|
|
190
|
+
return authoredDependencyGraphHash(configRoot, [...entrypoints]);
|
|
191
|
+
};
|
|
192
|
+
const validateAgentCaseProfileDrivers = async (execution, corpus) => {
|
|
193
|
+
const selected = [
|
|
194
|
+
...new Map(selectorsBySplit(execution.loaded.campaign)
|
|
195
|
+
.flatMap((split) => selectCases(corpus.cases, split.selector))
|
|
196
|
+
.map((evalCase) => [evalCase.path, evalCase])).values(),
|
|
197
|
+
];
|
|
198
|
+
for (const evalCase of selected) {
|
|
199
|
+
const authored = bindSubjectRoot(await loadEvalCase(evalCase.path), execution.loaded.subjectRoot);
|
|
200
|
+
for (const profileId of execution.loaded.campaign.profiles) {
|
|
201
|
+
const profile = profileOrThrow(execution.config, profileId);
|
|
202
|
+
const reference = resolveDriverReference(authored, profile.driver ?? execution.config.defaultDriver);
|
|
203
|
+
const name = typeof reference === "string" ? reference : reference.name;
|
|
204
|
+
const rawConfig = typeof reference === "object" ? reference.config : undefined;
|
|
205
|
+
boundExperimentDriverConfig(name, rawConfig, "identity-check");
|
|
206
|
+
const plugin = execution.registry.getDriver(name);
|
|
207
|
+
if (!plugin) {
|
|
208
|
+
throw new Error(`case '${evalCase.id}' profile '${profileId}' references unknown driver '${name}'`);
|
|
209
|
+
}
|
|
210
|
+
const validated = validatePluginConfig(plugin.configSchema, rawConfig, `case '${evalCase.id}' profile '${profileId}' driver '${name}'`);
|
|
211
|
+
boundExperimentDriverConfig(name, validated, "identity-check");
|
|
212
|
+
}
|
|
213
|
+
}
|
|
214
|
+
};
|
|
215
|
+
export const evaluationIdentity = async (execution, corpus, mode, boundRuntimeCliIdentityHash) => {
|
|
216
|
+
const profiles = await resolvedProfileSnapshot(execution);
|
|
217
|
+
if (corpus.kind === "agent") {
|
|
218
|
+
await validateAgentCaseProfileDrivers(execution, corpus);
|
|
219
|
+
}
|
|
220
|
+
const configSourceHash = execution.config.configPath
|
|
221
|
+
? contentHash(await readFile(execution.config.configPath))
|
|
222
|
+
: null;
|
|
223
|
+
const configDependencyHash = await authoredConfigDependencyHash(execution, corpus);
|
|
224
|
+
const optimizerConfig = mode === "optimize" && execution.loaded.campaign.optimizer
|
|
225
|
+
? ReflectiveOptimizerConfigSchema.parse(execution.loaded.campaign.optimizer.config ?? {})
|
|
226
|
+
: undefined;
|
|
227
|
+
let optimizerCommand;
|
|
228
|
+
if (optimizerConfig) {
|
|
229
|
+
optimizerCommand = optimizerConfig.command;
|
|
230
|
+
if (!optimizerCommand) {
|
|
231
|
+
optimizerCommand = optimizerConfig.provider === "codex" ? "codex" : "claude";
|
|
232
|
+
}
|
|
233
|
+
}
|
|
234
|
+
return valueHash({
|
|
235
|
+
evaluationKind: execution.loaded.campaign.evaluation.kind,
|
|
236
|
+
evaluatorVersion: 1,
|
|
237
|
+
luxVersion: LUX_VERSION,
|
|
238
|
+
evaluatorImplementationHash: await evaluatorImplementationHash(),
|
|
239
|
+
runtime: runtimeIdentity(),
|
|
240
|
+
corpusHash: corpus.hash,
|
|
241
|
+
profiles,
|
|
242
|
+
configSourceHash,
|
|
243
|
+
configDependencyHash,
|
|
244
|
+
validationCliIdentities: await Promise.all(execution.loaded.campaign.subject.validate.map(async (validation) => {
|
|
245
|
+
assertValidationArgumentsBound(execution.loaded.subjectRoot, validation.args);
|
|
246
|
+
return {
|
|
247
|
+
args: validation.args,
|
|
248
|
+
cli: await cliIdentity(validation.command, execution.loaded.subjectRoot),
|
|
249
|
+
};
|
|
250
|
+
})),
|
|
251
|
+
optimizerCliIdentity: optimizerCommand
|
|
252
|
+
? await cliIdentity(optimizerCommand, execution.loaded.directory)
|
|
253
|
+
: null,
|
|
254
|
+
runtimeCliIdentityHash: boundRuntimeCliIdentityHash ?? (await runtimeCliIdentityHash(execution, corpus, mode)),
|
|
255
|
+
});
|
|
256
|
+
};
|
|
257
|
+
//# sourceMappingURL=experiment-identity.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-identity.js","sourceRoot":"","sources":["../../src/orchestrator/experiment-identity.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAC5C,OAAO,EAAE,UAAU,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,EAAE,MAAM,WAAW,CAAC;AAC/D,OAAO,EAAE,eAAe,EAAE,WAAW,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AAC3E,OAAO,EAAE,+BAA+B,EAAE,MAAM,+BAA+B,CAAC;AAChF,OAAO,EAAE,2BAA2B,EAAE,MAAM,4BAA4B,CAAC;AACzE,OAAO,EAAE,8BAA8B,EAAE,MAAM,iBAAiB,CAAC;AACjE,OAAO,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAC5C,OAAO,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AAEtE,OAAO,EAAE,cAAc,EAAE,MAAM,yBAAyB,CAAC;AACzD,OAAO,EAAE,uBAAuB,EAAE,gBAAgB,EAAE,MAAM,wBAAwB,CAAC;AACnF,OAAO,EACL,gBAAgB,EAChB,sBAAsB,EACtB,WAAW,EACX,4BAA4B,EAC5B,WAAW,EAEX,yBAAyB,EACzB,qBAAqB,EACrB,aAAa,EACb,mCAAmC,GACpC,MAAM,kCAAkC,CAAC;AAC1C,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAC3C,OAAO,EACL,2BAA2B,EAC3B,wBAAwB,EACxB,sBAAsB,GACvB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,EAAE,eAAe,EAAE,MAAM,iBAAiB,CAAC;AAClD,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AAEnE,MAAM,CAAC,MAAM,uBAAuB,GAAG,KAAK,EAC1C,SAAmC,EACD,EAAE;IACpC,MAAM,cAAc,GAAG,IAAI,GAAG,EAAkC,CAAC;IACjE,MAAM,sBAAsB,GAAG,IAAI,GAAG,EAA2B,CAAC;IAClE,MAAM,YAAY,GAAG,CAAC,IAAY,EAA0B,EAAE;QAC5D,MAAM,QAAQ,GAAG,cAAc,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAC1C,IAAI,QAAQ;YAAE,OAAO,QAAQ,CAAC;QAC9B,MAAM,OAAO,GAAG,sBAAsB,CAAC,IAAI,CAAC,CAAC;QAC7C,cAAc,CAAC,GAAG,CAAC,IAAI,EAAE,OAAO,CAAC,CAAC;QAClC,OAAO,OAAO,CAAC;IACjB,CAAC,CAAC;IACF,OAAO,MAAM,CAAC,WAAW,CACvB,MAAM,OAAO,CAAC,GAAG,CACf,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,SAAS,EAAE,EAAE;QACzD,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;QAC5D,qEAAqE;QACrE,kEAAkE;QAClE,gEAAgE;QAChE,MAAM,MAAM,GACV,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,KAAK,OAAO;YACnD,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,IAAI,SAAS,CAAC,MAAM,CAAC,aAAa,CAAC;YACpD,CAAC,CAAC,SAAS,CAAC;QAChB,MAAM,SAAS,GAAG,aAAa,CAAC,MAAM,CAAC,CAAC;QACxC,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;QAC5D,MAAM,QAAQ,GACZ,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,KAAK,OAAO;YACnD,CAAC,CAAC,CAAC,OAAO,CAAC,QAAQ,IAAI,SAAS,CAAC,MAAM,CAAC,eAAe,CAAC;YACxD,CAAC,CAAC,SAAS,CAAC;QAChB,MAAM,WAAW,GAAG,qBAAqB,CAAC,QAAQ,CAAC,CAAC;QACpD,MAAM,sBAAsB,GAC1B,SAAS,EAAE,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,MAAM,YAAY,CAAC,gBAAgB,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;QAC1F,MAAM,uBAAuB,GAAG,MAAM,4BAA4B,CAChE,MAAM,EACN,SAAS,CAAC,MAAM,CAAC,OAAO,EACxB,sBAAsB,CACvB,CAAC;QACF,IAAI,cAAc,GAAG,OAAO,EAAE,OAAO,CAAC;QACtC,IAAI,CAAC,cAAc,EAAE,CAAC;YACpB,cAAc,GAAG,OAAO,EAAE,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC;QACtE,CAAC;QACD,MAAM,uBAAuB,GAC3B,OAAO,CAAC,OAAO,CAAC,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,KAAK,QAAQ,CAAC;QAC7E,MAAM,kBAAkB,GAAG,uBAAuB;YAChD,CAAC,CAAC,MAAM,WAAW,CAAC,cAAc,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;YAC7D,CAAC,CAAC,IAAI,CAAC;QACT,OAAO;YACL,SAAS;YACT;gBACE,EAAE,EAAE,SAAS;gBACb,MAAM,EAAE,MAAM,IAAI,IAAI;gBACtB,iBAAiB,EAAE,SAAS;oBAC1B,CAAC,CAAC,MAAM,WAAW,CAAC,SAAS,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;oBAChE,CAAC,CAAC,IAAI;gBACR,GAAG,CAAC,SAAS,EAAE,IAAI,KAAK,aAAa;oBACnC,CAAC,CAAC;wBACE,sBAAsB,EAAE,gBAAgB,CAAC,MAAM,CAAC;wBAChD,sBAAsB;wBACtB,uBAAuB;qBACxB;oBACH,CAAC,CAAC,EAAE,CAAC;gBACP,QAAQ;gBACR,mBAAmB,EAAE,WAAW;oBAC9B,CAAC,CAAC,MAAM,WAAW,CAAC,WAAW,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;oBAClE,CAAC,CAAC,IAAI;gBACR,OAAO,EAAE,OAAO,IAAI,IAAI;gBACxB,kBAAkB;aACnB;SACF,CAAC;IACJ,CAAC,CAAC,CACH,CACF,CAAC;AACJ,CAAC,CAAC;AAEF,MAAM,4BAA4B,GAAG,KAAK,EACxC,SAAmC,EACnC,MAAwB,EACF,EAAE;IACxB,MAAM,KAAK,GAAG,IAAI,GAAG,EAAU,CAAC;IAChC,MAAM,GAAG,GAAG,CAAC,OAA2B,EAAQ,EAAE;QAChD,IAAI,CAAC,OAAO,IAAI,CAAC,UAAU,CAAC,OAAO,CAAC;YAAE,OAAO;QAC7C,MAAM,QAAQ,GAAG,QAAQ,CAAC,SAAS,CAAC,MAAM,CAAC,OAAO,EAAE,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC;QACtE,IACE,QAAQ,KAAK,EAAE;YACf,QAAQ,KAAK,IAAI;YACjB,CAAC,QAAQ,CAAC,UAAU,CAAC,KAAK,GAAG,EAAE,CAAC;YAChC,CAAC,UAAU,CAAC,QAAQ,CAAC,EACrB,CAAC;YACD,KAAK,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QACtB,CAAC;IACH,CAAC,CAAC;IACF,KAAK,MAAM,UAAU,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,QAAQ;QAAE,GAAG,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC;IAC7F,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,EAAE,CAAC;QACxC,MAAM,SAAS,GAAG,+BAA+B,CAAC,KAAK,CACrD,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,MAAM,IAAI,EAAE,CACjD,CAAC;QACF,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC;IACzB,CAAC;IACD,KAAK,MAAM,SAAS,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;QAC3D,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;QAC5D,GAAG,CAAC,aAAa,CAAC,OAAO,CAAC,MAAM,IAAI,SAAS,CAAC,MAAM,CAAC,aAAa,CAAC,EAAE,OAAO,CAAC,CAAC;QAC9E,GAAG,CAAC,qBAAqB,CAAC,OAAO,CAAC,QAAQ,IAAI,SAAS,CAAC,MAAM,CAAC,eAAe,CAAC,EAAE,OAAO,CAAC,CAAC;QAC1F,GAAG,CAAC,CAAC,OAAO,CAAC,OAAO,IAAI,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC,CAAC;IAC9D,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC5B,MAAM,QAAQ,GAAG;YACf,GAAG,IAAI,GAAG,CACR,gBAAgB,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC;iBACxC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,WAAW,CAAC,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;iBAC7D,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC,CAChD,CAAC,MAAM,EAAE;SACX,CAAC;QACF,KAAK,MAAM,QAAQ,IAAI,QAAQ,EAAE,CAAC;YAChC,MAAM,QAAQ,GAAG,eAAe,CAC9B,MAAM,YAAY,CAAC,QAAQ,CAAC,IAAI,CAAC,EACjC,SAAS,CAAC,MAAM,CAAC,WAAW,CAC7B,CAAC;YACF,GAAG,CAAC,aAAa,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC,CAAC;YAC7C,GAAG,CAAC,qBAAqB,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,OAAO,CAAC,CAAC;YACvD,KAAK,MAAM,SAAS,IAAI,uBAAuB,CAAC,SAAS,EAAE,QAAQ,CAAC,EAAE,CAAC;gBACrE,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC;YACzB,CAAC;QACH,CAAC;IACH,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,oBAAoB,GAAG,KAAK,EACvC,SAAmC,EACnC,MAAwB,EACxB,IAA+B,EACC,EAAE;IAClC,MAAM,UAAU,GAAG,IAAI,GAAG,EAA+B,CAAC;IAC1D,MAAM,GAAG,GAAG,CAAC,OAA2B,EAAE,QAAgB,EAAQ,EAAE;QAClE,IAAI,CAAC,OAAO;YAAE,OAAO;QACrB,MAAM,SAAS,GAAG,yBAAyB,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;QAC5F,UAAU,CAAC,GAAG,CAAC,GAAG,SAAS,CAAC,OAAO,SAAS,SAAS,CAAC,QAAQ,EAAE,EAAE,SAAS,CAAC,CAAC;IAC/E,CAAC,CAAC;IAEF,KAAK,MAAM,UAAU,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,QAAQ,EAAE,CAAC;QACpE,GAAG,CAAC,UAAU,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,WAAW,CAAC,CAAC;IACxD,CAAC;IACD,IAAI,IAAI,KAAK,UAAU,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,EAAE,CAAC;QAC/D,MAAM,SAAS,GAAG,+BAA+B,CAAC,KAAK,CACrD,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,MAAM,IAAI,EAAE,CACjD,CAAC;QACF,GAAG,CACD,SAAS,CAAC,OAAO,IAAI,CAAC,SAAS,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,EAC1E,SAAS,CAAC,MAAM,CAAC,SAAS,CAC3B,CAAC;IACJ,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAC7B,KAAK,MAAM,SAAS,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;YAC3D,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;YAC5D,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;YAC5D,GAAG,CACD,OAAO,EAAE,OAAO,IAAI,CAAC,OAAO,EAAE,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,EACxE,SAAS,CAAC,MAAM,CAAC,OAAO,CACzB,CAAC;QACJ,CAAC;IACH,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC5B,MAAM,QAAQ,GAAG;YACf,GAAG,IAAI,GAAG,CACR,gBAAgB,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC;iBACxC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,WAAW,CAAC,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;iBAC7D,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC,CAChD,CAAC,MAAM,EAAE;SACX,CAAC;QACF,KAAK,MAAM,QAAQ,IAAI,QAAQ,EAAE,CAAC;YAChC,MAAM,QAAQ,GAAG,eAAe,CAC9B,MAAM,YAAY,CAAC,QAAQ,CAAC,IAAI,CAAC,EACjC,SAAS,CAAC,MAAM,CAAC,WAAW,CAC7B,CAAC;YACF,KAAK,MAAM,SAAS,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;gBAC3D,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;gBAC5D,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;gBAC5D,MAAM,MAAM,GAAG,sBAAsB,CACnC,QAAQ,EACR,OAAO,CAAC,MAAM,IAAI,SAAS,CAAC,MAAM,CAAC,aAAa,CACjD,CAAC;gBACF,MAAM,QAAQ,GAAG,wBAAwB,CACvC,QAAQ,EACR,OAAO,CAAC,QAAQ,IAAI,SAAS,CAAC,MAAM,CAAC,eAAe,EACpD,OAAO,CACR,CAAC;gBACF,GAAG,CAAC,aAAa,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;gBAC9D,GAAG,CAAC,qBAAqB,CAAC,QAAQ,CAAC,EAAE,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;YAC1E,CAAC;YACD,KAAK,MAAM,SAAS,IAAI,uBAAuB,CAAC,SAAS,EAAE,QAAQ,CAAC,EAAE,CAAC;gBACrE,GAAG,CAAC,SAAS,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;YACnD,CAAC;QACH,CAAC;IACH,CAAC;IACD,OAAO,CAAC,GAAG,UAAU,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE;QACnD,MAAM,YAAY,GAAG,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;QAC/D,OAAO,YAAY,IAAI,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC;IACrE,CAAC,CAAC,CAAC;AACL,CAAC,CAAC;AAEF,MAAM,sBAAsB,GAAG,KAAK,EAClC,SAAmC,EACnC,MAAwB,EACxB,IAA+B,EACd,EAAE,CACnB,mCAAmC,CAAC,MAAM,oBAAoB,CAAC,SAAS,EAAE,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC;AAE3F,MAAM,4BAA4B,GAAG,KAAK,EACxC,SAAmC,EACnC,MAAwB,EACP,EAAE;IACnB,MAAM,UAAU,GAAG,OAAO,CAAC,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;IACrD,MAAM,QAAQ,GAAG,OAAO,CAAC,SAAS,CAAC,QAAQ,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;IAC1E,MAAM,cAAc,GAAG,QAAQ,CAAC,UAAU,EAAE,QAAQ,CAAC,CAAC;IACtD,IAAI,cAAc,KAAK,EAAE,EAAE,CAAC;QAC1B,MAAM,IAAI,KAAK,CAAC,mDAAmD,CAAC,CAAC;IACvE,CAAC;IACD,MAAM,WAAW,GAAG,MAAM,4BAA4B,CAAC,SAAS,EAAE,MAAM,CAAC,CAAC;IAC1E,MAAM,GAAG,GAAG,CAAC,IAA+B,EAAQ,EAAE;QACpD,IAAI,IAAI;YAAE,WAAW,CAAC,GAAG,CAAC,QAAQ,CAAC,UAAU,EAAE,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACjE,CAAC,CAAC;IACF,GAAG,CAAC,SAAS,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC;IACjC,GAAG,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IAC3B,IAAI,MAAM,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC5B,KAAK,MAAM,KAAK,IAAI,gBAAgB,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,EAAE,CAAC;YAChE,KAAK,MAAM,QAAQ,IAAI,WAAW,CAAC,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC;gBAAE,GAAG,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACvF,CAAC;IACH,CAAC;SAAM,CAAC;QACN,GAAG,CAAC,SAAS,CAAC,MAAM,CAAC,iBAAiB,EAAE,IAAI,CAAC,CAAC;IAChD,CAAC;IACD,OAAO,2BAA2B,CAAC,UAAU,EAAE,CAAC,GAAG,WAAW,CAAC,CAAC,CAAC;AACnE,CAAC,CAAC;AACF,MAAM,+BAA+B,GAAG,KAAK,EAC3C,SAAmC,EACnC,MAAoD,EACrC,EAAE;IACjB,MAAM,QAAQ,GAAG;QACf,GAAG,IAAI,GAAG,CACR,gBAAgB,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC;aACxC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,WAAW,CAAC,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;aAC7D,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC,CAChD,CAAC,MAAM,EAAE;KACX,CAAC;IACF,KAAK,MAAM,QAAQ,IAAI,QAAQ,EAAE,CAAC;QAChC,MAAM,QAAQ,GAAG,eAAe,CAC9B,MAAM,YAAY,CAAC,QAAQ,CAAC,IAAI,CAAC,EACjC,SAAS,CAAC,MAAM,CAAC,WAAW,CAC7B,CAAC;QACF,KAAK,MAAM,SAAS,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;YAC3D,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;YAC5D,MAAM,SAAS,GAAG,sBAAsB,CACtC,QAAQ,EACR,OAAO,CAAC,MAAM,IAAI,SAAS,CAAC,MAAM,CAAC,aAAa,CACjD,CAAC;YACF,MAAM,IAAI,GAAG,OAAO,SAAS,KAAK,QAAQ,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,CAAC;YACxE,MAAM,SAAS,GAAG,OAAO,SAAS,KAAK,QAAQ,CAAC,CAAC,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;YAC/E,2BAA2B,CAAC,IAAI,EAAE,SAAS,EAAE,gBAAgB,CAAC,CAAC;YAC/D,MAAM,MAAM,GAAG,SAAS,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC;YAClD,IAAI,CAAC,MAAM,EAAE,CAAC;gBACZ,MAAM,IAAI,KAAK,CACb,SAAS,QAAQ,CAAC,EAAE,cAAc,SAAS,gCAAgC,IAAI,GAAG,CACnF,CAAC;YACJ,CAAC;YACD,MAAM,SAAS,GAAG,oBAAoB,CACpC,MAAM,CAAC,YAAY,EACnB,SAAS,EACT,SAAS,QAAQ,CAAC,EAAE,cAAc,SAAS,aAAa,IAAI,GAAG,CAChE,CAAC;YACF,2BAA2B,CAAC,IAAI,EAAE,SAAS,EAAE,gBAAgB,CAAC,CAAC;QACjE,CAAC;IACH,CAAC;AACH,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,kBAAkB,GAAG,KAAK,EACrC,SAAmC,EACnC,MAAwB,EACxB,IAA+B,EAC/B,2BAAoC,EACnB,EAAE;IACnB,MAAM,QAAQ,GAAG,MAAM,uBAAuB,CAAC,SAAS,CAAC,CAAC;IAC1D,IAAI,MAAM,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC5B,MAAM,+BAA+B,CAAC,SAAS,EAAE,MAAM,CAAC,CAAC;IAC3D,CAAC;IACD,MAAM,gBAAgB,GAAG,SAAS,CAAC,MAAM,CAAC,UAAU;QAClD,CAAC,CAAC,WAAW,CAAC,MAAM,QAAQ,CAAC,SAAS,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC;QAC1D,CAAC,CAAC,IAAI,CAAC;IACT,MAAM,oBAAoB,GAAG,MAAM,4BAA4B,CAAC,SAAS,EAAE,MAAM,CAAC,CAAC;IACnF,MAAM,eAAe,GACnB,IAAI,KAAK,UAAU,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS;QACxD,CAAC,CAAC,+BAA+B,CAAC,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,MAAM,IAAI,EAAE,CAAC;QACzF,CAAC,CAAC,SAAS,CAAC;IAChB,IAAI,gBAAoC,CAAC;IACzC,IAAI,eAAe,EAAE,CAAC;QACpB,gBAAgB,GAAG,eAAe,CAAC,OAAO,CAAC;QAC3C,IAAI,CAAC,gBAAgB,EAAE,CAAC;YACtB,gBAAgB,GAAG,eAAe,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC;QAC/E,CAAC;IACH,CAAC;IACD,OAAO,SAAS,CAAC;QACf,cAAc,EAAE,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI;QACzD,gBAAgB,EAAE,CAAC;QACnB,UAAU,EAAE,WAAW;QACvB,2BAA2B,EAAE,MAAM,2BAA2B,EAAE;QAChE,OAAO,EAAE,eAAe,EAAE;QAC1B,UAAU,EAAE,MAAM,CAAC,IAAI;QACvB,QAAQ;QACR,gBAAgB;QAChB,oBAAoB;QACpB,uBAAuB,EAAE,MAAM,OAAO,CAAC,GAAG,CACxC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,UAAU,EAAE,EAAE;YAClE,8BAA8B,CAAC,SAAS,CAAC,MAAM,CAAC,WAAW,EAAE,UAAU,CAAC,IAAI,CAAC,CAAC;YAC9E,OAAO;gBACL,IAAI,EAAE,UAAU,CAAC,IAAI;gBACrB,GAAG,EAAE,MAAM,WAAW,CAAC,UAAU,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,WAAW,CAAC;aACzE,CAAC;QACJ,CAAC,CAAC,CACH;QACD,oBAAoB,EAAE,gBAAgB;YACpC,CAAC,CAAC,MAAM,WAAW,CAAC,gBAAgB,EAAE,SAAS,CAAC,MAAM,CAAC,SAAS,CAAC;YACjE,CAAC,CAAC,IAAI;QACR,sBAAsB,EACpB,2BAA2B,IAAI,CAAC,MAAM,sBAAsB,CAAC,SAAS,EAAE,MAAM,EAAE,IAAI,CAAC,CAAC;KACzF,CAAC,CAAC;AACL,CAAC,CAAC"}
|
|
@@ -0,0 +1,175 @@
|
|
|
1
|
+
import { z } from "zod";
|
|
2
|
+
import { type ExperimentCampaign } from "../core/index.js";
|
|
3
|
+
export declare const LoadedExperimentCampaignSchema: z.ZodObject<{
|
|
4
|
+
path: z.ZodString;
|
|
5
|
+
directory: z.ZodString;
|
|
6
|
+
campaign: z.ZodObject<{
|
|
7
|
+
id: z.ZodString;
|
|
8
|
+
subject: z.ZodObject<{
|
|
9
|
+
root: z.ZodString;
|
|
10
|
+
mutable: z.ZodArray<z.ZodString>;
|
|
11
|
+
exclude: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
12
|
+
validate: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
13
|
+
command: z.ZodString;
|
|
14
|
+
args: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
15
|
+
timeoutMs: z.ZodDefault<z.ZodNumber>;
|
|
16
|
+
unsafeHostExecution: z.ZodLiteral<true>;
|
|
17
|
+
}, z.core.$strict>>>;
|
|
18
|
+
}, z.core.$strict>;
|
|
19
|
+
evaluation: z.ZodDefault<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
20
|
+
kind: z.ZodLiteral<"agent">;
|
|
21
|
+
}, z.core.$strict>, z.ZodObject<{
|
|
22
|
+
kind: z.ZodLiteral<"grader">;
|
|
23
|
+
annotations: z.ZodString;
|
|
24
|
+
allowedAssertionAdditions: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodArray<z.ZodString>>>;
|
|
25
|
+
scoreTolerance: z.ZodDefault<z.ZodNumber>;
|
|
26
|
+
}, z.core.$strict>], "kind">>;
|
|
27
|
+
profiles: z.ZodArray<z.ZodString>;
|
|
28
|
+
splits: z.ZodObject<{
|
|
29
|
+
train: z.ZodOptional<z.ZodObject<{
|
|
30
|
+
patterns: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
31
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
32
|
+
}, z.core.$strict>>;
|
|
33
|
+
validation: z.ZodObject<{
|
|
34
|
+
patterns: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
35
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
36
|
+
}, z.core.$strict>;
|
|
37
|
+
test: z.ZodOptional<z.ZodObject<{
|
|
38
|
+
patterns: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
39
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
40
|
+
}, z.core.$strict>>;
|
|
41
|
+
}, z.core.$strict>;
|
|
42
|
+
repetitions: z.ZodDefault<z.ZodObject<{
|
|
43
|
+
search: z.ZodDefault<z.ZodNumber>;
|
|
44
|
+
promotion: z.ZodDefault<z.ZodNumber>;
|
|
45
|
+
}, z.core.$strict>>;
|
|
46
|
+
variants: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
47
|
+
id: z.ZodString;
|
|
48
|
+
name: z.ZodOptional<z.ZodString>;
|
|
49
|
+
parentIds: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
50
|
+
components: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodString>>;
|
|
51
|
+
}, z.core.$strict>>>;
|
|
52
|
+
objectives: z.ZodDefault<z.ZodObject<{
|
|
53
|
+
constraints: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
54
|
+
metric: z.ZodString;
|
|
55
|
+
min: z.ZodOptional<z.ZodNumber>;
|
|
56
|
+
max: z.ZodOptional<z.ZodNumber>;
|
|
57
|
+
profile: z.ZodOptional<z.ZodString>;
|
|
58
|
+
split: z.ZodDefault<z.ZodEnum<{
|
|
59
|
+
test: "test";
|
|
60
|
+
train: "train";
|
|
61
|
+
validation: "validation";
|
|
62
|
+
}>>;
|
|
63
|
+
confidence: z.ZodOptional<z.ZodObject<{
|
|
64
|
+
method: z.ZodEnum<{
|
|
65
|
+
hoeffding: "hoeffding";
|
|
66
|
+
"percentile-bootstrap": "percentile-bootstrap";
|
|
67
|
+
wilson: "wilson";
|
|
68
|
+
}>;
|
|
69
|
+
level: z.ZodDefault<z.ZodNumber>;
|
|
70
|
+
resamples: z.ZodDefault<z.ZodNumber>;
|
|
71
|
+
seed: z.ZodDefault<z.ZodNumber>;
|
|
72
|
+
minSamples: z.ZodDefault<z.ZodNumber>;
|
|
73
|
+
}, z.core.$strict>>;
|
|
74
|
+
}, z.core.$strict>>>;
|
|
75
|
+
minimize: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
76
|
+
metric: z.ZodString;
|
|
77
|
+
profile: z.ZodOptional<z.ZodString>;
|
|
78
|
+
split: z.ZodDefault<z.ZodEnum<{
|
|
79
|
+
test: "test";
|
|
80
|
+
train: "train";
|
|
81
|
+
validation: "validation";
|
|
82
|
+
}>>;
|
|
83
|
+
}, z.core.$strict>>>;
|
|
84
|
+
}, z.core.$strict>>;
|
|
85
|
+
optimizer: z.ZodOptional<z.ZodObject<{
|
|
86
|
+
algorithm: z.ZodLiteral<"reflective">;
|
|
87
|
+
config: z.ZodOptional<z.ZodUnknown>;
|
|
88
|
+
candidateSelection: z.ZodDefault<z.ZodEnum<{
|
|
89
|
+
"current-best": "current-best";
|
|
90
|
+
pareto: "pareto";
|
|
91
|
+
}>>;
|
|
92
|
+
mergeFrontier: z.ZodDefault<z.ZodBoolean>;
|
|
93
|
+
promotionTopK: z.ZodDefault<z.ZodNumber>;
|
|
94
|
+
}, z.core.$strict>>;
|
|
95
|
+
budget: z.ZodDefault<z.ZodObject<{
|
|
96
|
+
maxCandidates: z.ZodDefault<z.ZodNumber>;
|
|
97
|
+
patience: z.ZodDefault<z.ZodNumber>;
|
|
98
|
+
maxMetricCalls: z.ZodDefault<z.ZodNumber>;
|
|
99
|
+
maxJudgeCalls: z.ZodDefault<z.ZodNumber>;
|
|
100
|
+
maxProposalCalls: z.ZodDefault<z.ZodNumber>;
|
|
101
|
+
maxTokens: z.ZodDefault<z.ZodNumber>;
|
|
102
|
+
maxCostUsd: z.ZodDefault<z.ZodNumber>;
|
|
103
|
+
maxUnpricedModelCalls: z.ZodDefault<z.ZodNumber>;
|
|
104
|
+
maxWallTimeMs: z.ZodDefault<z.ZodNumber>;
|
|
105
|
+
}, z.core.$strict>>;
|
|
106
|
+
meta: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
107
|
+
}, z.core.$strict>;
|
|
108
|
+
subjectRoot: z.ZodString;
|
|
109
|
+
campaignHash: z.ZodString;
|
|
110
|
+
graderAnnotations: z.ZodOptional<z.ZodObject<{
|
|
111
|
+
path: z.ZodString;
|
|
112
|
+
directory: z.ZodString;
|
|
113
|
+
dataset: z.ZodObject<{
|
|
114
|
+
schemaVersion: z.ZodLiteral<1>;
|
|
115
|
+
annotations: z.ZodArray<z.ZodObject<{
|
|
116
|
+
id: z.ZodString;
|
|
117
|
+
runDir: z.ZodString;
|
|
118
|
+
rubricPath: z.ZodPipe<z.ZodString, z.ZodTransform<string, string>>;
|
|
119
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
120
|
+
expected: z.ZodObject<{
|
|
121
|
+
casePassed: z.ZodOptional<z.ZodBoolean>;
|
|
122
|
+
caseScore: z.ZodOptional<z.ZodNumber>;
|
|
123
|
+
assertions: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodObject<{
|
|
124
|
+
passed: z.ZodBoolean;
|
|
125
|
+
score: z.ZodOptional<z.ZodNumber>;
|
|
126
|
+
}, z.core.$strict>>>;
|
|
127
|
+
}, z.core.$strict>;
|
|
128
|
+
source: z.ZodEnum<{
|
|
129
|
+
human: "human";
|
|
130
|
+
synthetic: "synthetic";
|
|
131
|
+
}>;
|
|
132
|
+
annotator: z.ZodOptional<z.ZodString>;
|
|
133
|
+
labeledAt: z.ZodOptional<z.ZodString>;
|
|
134
|
+
reviewer: z.ZodOptional<z.ZodString>;
|
|
135
|
+
feedback: z.ZodOptional<z.ZodString>;
|
|
136
|
+
meta: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
137
|
+
}, z.core.$strict>>;
|
|
138
|
+
meta: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
139
|
+
}, z.core.$strict>;
|
|
140
|
+
annotations: z.ZodArray<z.ZodObject<{
|
|
141
|
+
id: z.ZodString;
|
|
142
|
+
runDir: z.ZodString;
|
|
143
|
+
rubricPath: z.ZodPipe<z.ZodString, z.ZodTransform<string, string>>;
|
|
144
|
+
tags: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
145
|
+
expected: z.ZodObject<{
|
|
146
|
+
casePassed: z.ZodOptional<z.ZodBoolean>;
|
|
147
|
+
caseScore: z.ZodOptional<z.ZodNumber>;
|
|
148
|
+
assertions: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodObject<{
|
|
149
|
+
passed: z.ZodBoolean;
|
|
150
|
+
score: z.ZodOptional<z.ZodNumber>;
|
|
151
|
+
}, z.core.$strict>>>;
|
|
152
|
+
}, z.core.$strict>;
|
|
153
|
+
source: z.ZodEnum<{
|
|
154
|
+
human: "human";
|
|
155
|
+
synthetic: "synthetic";
|
|
156
|
+
}>;
|
|
157
|
+
annotator: z.ZodOptional<z.ZodString>;
|
|
158
|
+
labeledAt: z.ZodOptional<z.ZodString>;
|
|
159
|
+
reviewer: z.ZodOptional<z.ZodString>;
|
|
160
|
+
feedback: z.ZodOptional<z.ZodString>;
|
|
161
|
+
meta: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
162
|
+
}, z.core.$strict>>;
|
|
163
|
+
hash: z.ZodString;
|
|
164
|
+
}, z.core.$strict>>;
|
|
165
|
+
}, z.core.$strict>;
|
|
166
|
+
export type LoadedExperimentCampaign = z.infer<typeof LoadedExperimentCampaignSchema>;
|
|
167
|
+
export declare const ExperimentCommandModeSchema: z.ZodEnum<{
|
|
168
|
+
experiment: "experiment";
|
|
169
|
+
optimize: "optimize";
|
|
170
|
+
}>;
|
|
171
|
+
export type ExperimentCommandMode = z.infer<typeof ExperimentCommandModeSchema>;
|
|
172
|
+
export declare const experimentCampaignMode: (path: string, campaign: ExperimentCampaign, requestedMode?: ExperimentCommandMode) => ExperimentCommandMode;
|
|
173
|
+
export declare const resolveExperimentCampaignPath: (projectRoot: string, filePath: string) => Promise<string>;
|
|
174
|
+
export declare const loadExperimentCampaign: (filePath: string) => Promise<LoadedExperimentCampaign>;
|
|
175
|
+
//# sourceMappingURL=experiment-loader.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-loader.d.ts","sourceRoot":"","sources":["../../src/orchestrator/experiment-loader.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EACL,KAAK,kBAAkB,EAIxB,MAAM,kBAAkB,CAAC;AAG1B,eAAO,MAAM,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAShC,CAAC;AACZ,MAAM,MAAM,wBAAwB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,8BAA8B,CAAC,CAAC;AAEtF,eAAO,MAAM,2BAA2B;;;EAAqC,CAAC;AAC9E,MAAM,MAAM,qBAAqB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,2BAA2B,CAAC,CAAC;AAEhF,eAAO,MAAM,sBAAsB,SAC3B,MAAM,YACF,kBAAkB,kBACZ,qBAAqB,KACpC,qBAqBF,CAAC;AAEF,eAAO,MAAM,6BAA6B,gBAC3B,MAAM,YACT,MAAM,KACf,OAAO,CAAC,MAAM,CAgBhB,CAAC;AAEF,eAAO,MAAM,sBAAsB,aACvB,MAAM,KACf,OAAO,CAAC,wBAAwB,CA2BlC,CAAC"}
|
|
@@ -0,0 +1,77 @@
|
|
|
1
|
+
import { lstat, realpath } from "node:fs/promises";
|
|
2
|
+
import { dirname, isAbsolute, relative, resolve, sep } from "node:path";
|
|
3
|
+
import { pathToFileURL } from "node:url";
|
|
4
|
+
import { z } from "zod";
|
|
5
|
+
import { ExperimentCampaignSchema, formatZodIssues, valueHash, } from "../core/index.js";
|
|
6
|
+
import { LoadedGraderAnnotationsSchema, loadGraderAnnotations } from "./annotation-loader.js";
|
|
7
|
+
export const LoadedExperimentCampaignSchema = z
|
|
8
|
+
.object({
|
|
9
|
+
path: z.string(),
|
|
10
|
+
directory: z.string(),
|
|
11
|
+
campaign: ExperimentCampaignSchema,
|
|
12
|
+
subjectRoot: z.string(),
|
|
13
|
+
campaignHash: z.string(),
|
|
14
|
+
graderAnnotations: LoadedGraderAnnotationsSchema.optional(),
|
|
15
|
+
})
|
|
16
|
+
.strict();
|
|
17
|
+
export const ExperimentCommandModeSchema = z.enum(["experiment", "optimize"]);
|
|
18
|
+
export const experimentCampaignMode = (path, campaign, requestedMode) => {
|
|
19
|
+
const experimentSuffix = path.endsWith(".experiment.ts");
|
|
20
|
+
const optimizeSuffix = path.endsWith(".optimize.ts");
|
|
21
|
+
if (experimentSuffix && campaign.optimizer) {
|
|
22
|
+
throw new Error("a .experiment.ts campaign must not define an optimizer; rename it to .optimize.ts or remove the optimizer");
|
|
23
|
+
}
|
|
24
|
+
if (optimizeSuffix && !campaign.optimizer) {
|
|
25
|
+
throw new Error("a .optimize.ts campaign must define an optimizer; rename it to .experiment.ts or add one");
|
|
26
|
+
}
|
|
27
|
+
let authoredMode = "experiment";
|
|
28
|
+
if (campaign.optimizer)
|
|
29
|
+
authoredMode = "optimize";
|
|
30
|
+
if (requestedMode && requestedMode !== authoredMode) {
|
|
31
|
+
throw new Error(`lux ${requestedMode} cannot run this ${authoredMode} campaign; use \`lux ${authoredMode}\` or change its optimizer`);
|
|
32
|
+
}
|
|
33
|
+
return authoredMode;
|
|
34
|
+
};
|
|
35
|
+
export const resolveExperimentCampaignPath = async (projectRoot, filePath) => {
|
|
36
|
+
const root = resolve(projectRoot);
|
|
37
|
+
const path = resolve(root, filePath);
|
|
38
|
+
const fromRoot = relative(root, path);
|
|
39
|
+
if (isAbsolute(fromRoot) || fromRoot === ".." || fromRoot.startsWith(`..${sep}`)) {
|
|
40
|
+
throw new Error(`experiment campaign must be contained by the Lux project: ${path}`);
|
|
41
|
+
}
|
|
42
|
+
const info = await lstat(path);
|
|
43
|
+
if (info.isSymbolicLink() || !info.isFile()) {
|
|
44
|
+
throw new Error(`experiment campaign must be a regular, non-symlink file: ${path}`);
|
|
45
|
+
}
|
|
46
|
+
const [canonicalRoot, canonicalPath] = await Promise.all([realpath(root), realpath(path)]);
|
|
47
|
+
if (canonicalPath !== resolve(canonicalRoot, fromRoot)) {
|
|
48
|
+
throw new Error(`experiment campaign path must not contain symlinks: ${path}`);
|
|
49
|
+
}
|
|
50
|
+
return canonicalPath;
|
|
51
|
+
};
|
|
52
|
+
export const loadExperimentCampaign = async (filePath) => {
|
|
53
|
+
const path = resolve(filePath);
|
|
54
|
+
const mod = (await import(pathToFileURL(path).href));
|
|
55
|
+
if (mod.default === undefined) {
|
|
56
|
+
throw new Error(`Lux experiment at ${path} has no default export. Use \`export default defineExperiment({ ... })\`.`);
|
|
57
|
+
}
|
|
58
|
+
const parsed = ExperimentCampaignSchema.safeParse(mod.default);
|
|
59
|
+
if (!parsed.success) {
|
|
60
|
+
throw new Error(`Invalid Lux experiment at ${path}: ${formatZodIssues(parsed.error)}`);
|
|
61
|
+
}
|
|
62
|
+
const directory = dirname(path);
|
|
63
|
+
const graderAnnotations = parsed.data.evaluation.kind === "grader"
|
|
64
|
+
? await loadGraderAnnotations(resolve(directory, parsed.data.evaluation.annotations))
|
|
65
|
+
: undefined;
|
|
66
|
+
return LoadedExperimentCampaignSchema.parse({
|
|
67
|
+
path,
|
|
68
|
+
directory,
|
|
69
|
+
campaign: parsed.data,
|
|
70
|
+
subjectRoot: resolve(directory, parsed.data.subject.root),
|
|
71
|
+
campaignHash: graderAnnotations
|
|
72
|
+
? valueHash({ campaign: parsed.data, annotationsHash: graderAnnotations.hash })
|
|
73
|
+
: valueHash(parsed.data),
|
|
74
|
+
...(graderAnnotations ? { graderAnnotations } : {}),
|
|
75
|
+
});
|
|
76
|
+
};
|
|
77
|
+
//# sourceMappingURL=experiment-loader.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"experiment-loader.js","sourceRoot":"","sources":["../../src/orchestrator/experiment-loader.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AACnD,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,EAAE,MAAM,WAAW,CAAC;AACxE,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AACzC,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAEL,wBAAwB,EACxB,eAAe,EACf,SAAS,GACV,MAAM,kBAAkB,CAAC;AAC1B,OAAO,EAAE,6BAA6B,EAAE,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AAE9F,MAAM,CAAC,MAAM,8BAA8B,GAAG,CAAC;KAC5C,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE;IAChB,SAAS,EAAE,CAAC,CAAC,MAAM,EAAE;IACrB,QAAQ,EAAE,wBAAwB;IAClC,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE;IACvB,YAAY,EAAE,CAAC,CAAC,MAAM,EAAE;IACxB,iBAAiB,EAAE,6BAA6B,CAAC,QAAQ,EAAE;CAC5D,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC,CAAC;AAG9E,MAAM,CAAC,MAAM,sBAAsB,GAAG,CACpC,IAAY,EACZ,QAA4B,EAC5B,aAAqC,EACd,EAAE;IACzB,MAAM,gBAAgB,GAAG,IAAI,CAAC,QAAQ,CAAC,gBAAgB,CAAC,CAAC;IACzD,MAAM,cAAc,GAAG,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;IACrD,IAAI,gBAAgB,IAAI,QAAQ,CAAC,SAAS,EAAE,CAAC;QAC3C,MAAM,IAAI,KAAK,CACb,2GAA2G,CAC5G,CAAC;IACJ,CAAC;IACD,IAAI,cAAc,IAAI,CAAC,QAAQ,CAAC,SAAS,EAAE,CAAC;QAC1C,MAAM,IAAI,KAAK,CACb,0FAA0F,CAC3F,CAAC;IACJ,CAAC;IACD,IAAI,YAAY,GAA0B,YAAY,CAAC;IACvD,IAAI,QAAQ,CAAC,SAAS;QAAE,YAAY,GAAG,UAAU,CAAC;IAClD,IAAI,aAAa,IAAI,aAAa,KAAK,YAAY,EAAE,CAAC;QACpD,MAAM,IAAI,KAAK,CACb,OAAO,aAAa,oBAAoB,YAAY,wBAAwB,YAAY,4BAA4B,CACrH,CAAC;IACJ,CAAC;IACD,OAAO,YAAY,CAAC;AACtB,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,6BAA6B,GAAG,KAAK,EAChD,WAAmB,EACnB,QAAgB,EACC,EAAE;IACnB,MAAM,IAAI,GAAG,OAAO,CAAC,WAAW,CAAC,CAAC;IAClC,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;IACrC,MAAM,QAAQ,GAAG,QAAQ,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;IACtC,IAAI,UAAU,CAAC,QAAQ,CAAC,IAAI,QAAQ,KAAK,IAAI,IAAI,QAAQ,CAAC,UAAU,CAAC,KAAK,GAAG,EAAE,CAAC,EAAE,CAAC;QACjF,MAAM,IAAI,KAAK,CAAC,6DAA6D,IAAI,EAAE,CAAC,CAAC;IACvF,CAAC;IACD,MAAM,IAAI,GAAG,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC;IAC/B,IAAI,IAAI,CAAC,cAAc,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC;QAC5C,MAAM,IAAI,KAAK,CAAC,4DAA4D,IAAI,EAAE,CAAC,CAAC;IACtF,CAAC;IACD,MAAM,CAAC,aAAa,EAAE,aAAa,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC3F,IAAI,aAAa,KAAK,OAAO,CAAC,aAAa,EAAE,QAAQ,CAAC,EAAE,CAAC;QACvD,MAAM,IAAI,KAAK,CAAC,uDAAuD,IAAI,EAAE,CAAC,CAAC;IACjF,CAAC;IACD,OAAO,aAAa,CAAC;AACvB,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,sBAAsB,GAAG,KAAK,EACzC,QAAgB,EACmB,EAAE;IACrC,MAAM,IAAI,GAAG,OAAO,CAAC,QAAQ,CAAC,CAAC;IAC/B,MAAM,GAAG,GAAG,CAAC,MAAM,MAAM,CAAC,aAAa,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,CAA0B,CAAC;IAC9E,IAAI,GAAG,CAAC,OAAO,KAAK,SAAS,EAAE,CAAC;QAC9B,MAAM,IAAI,KAAK,CACb,qBAAqB,IAAI,2EAA2E,CACrG,CAAC;IACJ,CAAC;IACD,MAAM,MAAM,GAAG,wBAAwB,CAAC,SAAS,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;IAC/D,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;QACpB,MAAM,IAAI,KAAK,CAAC,6BAA6B,IAAI,KAAK,eAAe,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IACzF,CAAC;IACD,MAAM,SAAS,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAChC,MAAM,iBAAiB,GACrB,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,IAAI,KAAK,QAAQ;QACtC,CAAC,CAAC,MAAM,qBAAqB,CAAC,OAAO,CAAC,SAAS,EAAE,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,WAAW,CAAC,CAAC;QACrF,CAAC,CAAC,SAAS,CAAC;IAChB,OAAO,8BAA8B,CAAC,KAAK,CAAC;QAC1C,IAAI;QACJ,SAAS;QACT,QAAQ,EAAE,MAAM,CAAC,IAAI;QACrB,WAAW,EAAE,OAAO,CAAC,SAAS,EAAE,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC;QACzD,YAAY,EAAE,iBAAiB;YAC7B,CAAC,CAAC,SAAS,CAAC,EAAE,QAAQ,EAAE,MAAM,CAAC,IAAI,EAAE,eAAe,EAAE,iBAAiB,CAAC,IAAI,EAAE,CAAC;YAC/E,CAAC,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC;QAC1B,GAAG,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,iBAAiB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACpD,CAAC,CAAC;AACL,CAAC,CAAC"}
|