cognitive-core 0.2.4 → 0.3.1
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/README.md +88 -143
- package/dist/agents/mock-provider.d.ts.map +1 -1
- package/dist/agents/mock-provider.js.map +1 -1
- package/dist/agents/types.d.ts.map +1 -1
- package/dist/agents/types.js +5 -4
- package/dist/agents/types.js.map +1 -1
- package/dist/atlas.d.ts +309 -97
- package/dist/atlas.d.ts.map +1 -1
- package/dist/atlas.js +743 -277
- package/dist/atlas.js.map +1 -1
- package/dist/bin/cli-utils.d.ts +25 -2
- package/dist/bin/cli-utils.d.ts.map +1 -1
- package/dist/bin/cli-utils.js +79 -9
- package/dist/bin/cli-utils.js.map +1 -1
- package/dist/bin/cognitive-core.js +32 -0
- package/dist/bin/cognitive-core.js.map +1 -1
- package/dist/bin/commands/dream.d.ts +17 -0
- package/dist/bin/commands/dream.d.ts.map +1 -0
- package/dist/bin/commands/dream.js +168 -0
- package/dist/bin/commands/dream.js.map +1 -0
- package/dist/bin/commands/eval.d.ts +181 -0
- package/dist/bin/commands/eval.d.ts.map +1 -0
- package/dist/bin/commands/eval.js +1582 -0
- package/dist/bin/commands/eval.js.map +1 -0
- package/dist/bin/commands/kb.js +0 -8
- package/dist/bin/commands/kb.js.map +1 -1
- package/dist/bin/commands/learn.js +12 -4
- package/dist/bin/commands/learn.js.map +1 -1
- package/dist/bin/commands/legacy.d.ts.map +1 -1
- package/dist/bin/commands/legacy.js +1 -1
- package/dist/bin/commands/legacy.js.map +1 -1
- package/dist/bin/commands/run.d.ts.map +1 -1
- package/dist/bin/commands/run.js +10 -9
- package/dist/bin/commands/run.js.map +1 -1
- package/dist/bin/commands/sessions.d.ts.map +1 -1
- package/dist/bin/commands/sessions.js.map +1 -1
- package/dist/bin/commands/skills.d.ts.map +1 -1
- package/dist/bin/commands/skills.js +283 -1
- package/dist/bin/commands/skills.js.map +1 -1
- package/dist/embeddings/bm25.d.ts.map +1 -1
- package/dist/embeddings/bm25.js +1 -2
- package/dist/embeddings/bm25.js.map +1 -1
- package/dist/embeddings/index.d.ts +3 -3
- package/dist/embeddings/index.d.ts.map +1 -1
- package/dist/embeddings/index.js +3 -3
- package/dist/embeddings/index.js.map +1 -1
- package/dist/embeddings/inverted-index.js +62 -7
- package/dist/embeddings/inverted-index.js.map +1 -1
- package/dist/embeddings/manager.d.ts.map +1 -1
- package/dist/embeddings/manager.js +6 -5
- package/dist/embeddings/manager.js.map +1 -1
- package/dist/embeddings/provider.d.ts +96 -5
- package/dist/embeddings/provider.d.ts.map +1 -1
- package/dist/embeddings/provider.js +171 -48
- package/dist/embeddings/provider.js.map +1 -1
- package/dist/embeddings/vector-store.d.ts.map +1 -1
- package/dist/embeddings/vector-store.js +3 -1
- package/dist/embeddings/vector-store.js.map +1 -1
- package/dist/eval/ablation-validate.d.ts +12 -0
- package/dist/eval/ablation-validate.d.ts.map +1 -0
- package/dist/eval/ablation-validate.js +78 -0
- package/dist/eval/ablation-validate.js.map +1 -0
- package/dist/eval/ablation.d.ts +94 -0
- package/dist/eval/ablation.d.ts.map +1 -0
- package/dist/eval/ablation.js +161 -0
- package/dist/eval/ablation.js.map +1 -0
- package/dist/eval/ci-smoke.d.ts +11 -0
- package/dist/eval/ci-smoke.d.ts.map +1 -0
- package/dist/eval/ci-smoke.js +49 -0
- package/dist/eval/ci-smoke.js.map +1 -0
- package/dist/eval/diagnostics.d.ts +214 -0
- package/dist/eval/diagnostics.d.ts.map +1 -0
- package/dist/eval/diagnostics.js +107 -0
- package/dist/eval/diagnostics.js.map +1 -0
- package/dist/eval/index.d.ts +27 -0
- package/dist/eval/index.d.ts.map +1 -0
- package/dist/eval/index.js +27 -0
- package/dist/eval/index.js.map +1 -0
- package/dist/eval/live/alfworld.d.ts +248 -0
- package/dist/eval/live/alfworld.d.ts.map +1 -0
- package/dist/eval/live/alfworld.js +637 -0
- package/dist/eval/live/alfworld.js.map +1 -0
- package/dist/eval/live/alfworld_h1.d.ts +16 -0
- package/dist/eval/live/alfworld_h1.d.ts.map +1 -0
- package/dist/eval/live/alfworld_h1.js +75 -0
- package/dist/eval/live/alfworld_h1.js.map +1 -0
- package/dist/eval/live/alfworld_memory.d.ts +38 -0
- package/dist/eval/live/alfworld_memory.d.ts.map +1 -0
- package/dist/eval/live/alfworld_memory.js +184 -0
- package/dist/eval/live/alfworld_memory.js.map +1 -0
- package/dist/eval/live/alfworld_strategy.d.ts +65 -0
- package/dist/eval/live/alfworld_strategy.d.ts.map +1 -0
- package/dist/eval/live/alfworld_strategy.js +89 -0
- package/dist/eval/live/alfworld_strategy.js.map +1 -0
- package/dist/eval/live/azure-openai-llm-provider.d.ts +41 -0
- package/dist/eval/live/azure-openai-llm-provider.d.ts.map +1 -0
- package/dist/eval/live/azure-openai-llm-provider.js +92 -0
- package/dist/eval/live/azure-openai-llm-provider.js.map +1 -0
- package/dist/eval/live/bridges/skillflow_e2b_bridge.d.ts +3 -0
- package/dist/eval/live/bridges/skillflow_e2b_bridge.d.ts.map +1 -0
- package/dist/eval/live/bridges/skillflow_e2b_bridge.js +12 -0
- package/dist/eval/live/bridges/skillflow_e2b_bridge.js.map +1 -0
- package/dist/eval/live/claude-cli-llm-provider.d.ts +53 -0
- package/dist/eval/live/claude-cli-llm-provider.d.ts.map +1 -0
- package/dist/eval/live/claude-cli-llm-provider.js +71 -0
- package/dist/eval/live/claude-cli-llm-provider.js.map +1 -0
- package/dist/eval/live/claude-cli.d.ts +96 -0
- package/dist/eval/live/claude-cli.d.ts.map +1 -0
- package/dist/eval/live/claude-cli.js +314 -0
- package/dist/eval/live/claude-cli.js.map +1 -0
- package/dist/eval/live/claude-delegate.d.ts +41 -0
- package/dist/eval/live/claude-delegate.d.ts.map +1 -0
- package/dist/eval/live/claude-delegate.js +59 -0
- package/dist/eval/live/claude-delegate.js.map +1 -0
- package/dist/eval/live/claude-workspace-backend.d.ts +24 -0
- package/dist/eval/live/claude-workspace-backend.d.ts.map +1 -0
- package/dist/eval/live/claude-workspace-backend.js +76 -0
- package/dist/eval/live/claude-workspace-backend.js.map +1 -0
- package/dist/eval/live/clbench.d.ts +64 -0
- package/dist/eval/live/clbench.d.ts.map +1 -0
- package/dist/eval/live/clbench.js +291 -0
- package/dist/eval/live/clbench.js.map +1 -0
- package/dist/eval/live/cold-reuse.d.ts +21 -0
- package/dist/eval/live/cold-reuse.d.ts.map +1 -0
- package/dist/eval/live/cold-reuse.js +102 -0
- package/dist/eval/live/cold-reuse.js.map +1 -0
- package/dist/eval/live/difficulty-registry.d.ts +68 -0
- package/dist/eval/live/difficulty-registry.d.ts.map +1 -0
- package/dist/eval/live/difficulty-registry.js +169 -0
- package/dist/eval/live/difficulty-registry.js.map +1 -0
- package/dist/eval/live/docker.d.ts +52 -0
- package/dist/eval/live/docker.d.ts.map +1 -0
- package/dist/eval/live/docker.js +164 -0
- package/dist/eval/live/docker.js.map +1 -0
- package/dist/eval/live/env-bridge.d.ts +68 -0
- package/dist/eval/live/env-bridge.d.ts.map +1 -0
- package/dist/eval/live/env-bridge.js +188 -0
- package/dist/eval/live/env-bridge.js.map +1 -0
- package/dist/eval/live/gsm8k.d.ts +44 -0
- package/dist/eval/live/gsm8k.d.ts.map +1 -0
- package/dist/eval/live/gsm8k.js +142 -0
- package/dist/eval/live/gsm8k.js.map +1 -0
- package/dist/eval/live/index.d.ts +38 -0
- package/dist/eval/live/index.d.ts.map +1 -0
- package/dist/eval/live/index.js +38 -0
- package/dist/eval/live/index.js.map +1 -0
- package/dist/eval/live/lifelongagent.d.ts +64 -0
- package/dist/eval/live/lifelongagent.d.ts.map +1 -0
- package/dist/eval/live/lifelongagent.js +216 -0
- package/dist/eval/live/lifelongagent.js.map +1 -0
- package/dist/eval/live/lift-registry.d.ts +128 -0
- package/dist/eval/live/lift-registry.d.ts.map +1 -0
- package/dist/eval/live/lift-registry.js +147 -0
- package/dist/eval/live/lift-registry.js.map +1 -0
- package/dist/eval/live/live-benchmark.d.ts +84 -0
- package/dist/eval/live/live-benchmark.d.ts.map +1 -0
- package/dist/eval/live/live-benchmark.js +2 -0
- package/dist/eval/live/live-benchmark.js.map +1 -0
- package/dist/eval/live/live-runner.d.ts +609 -0
- package/dist/eval/live/live-runner.d.ts.map +1 -0
- package/dist/eval/live/live-runner.js +1409 -0
- package/dist/eval/live/live-runner.js.map +1 -0
- package/dist/eval/live/refine.d.ts +88 -0
- package/dist/eval/live/refine.d.ts.map +1 -0
- package/dist/eval/live/refine.js +135 -0
- package/dist/eval/live/refine.js.map +1 -0
- package/dist/eval/live/registry-store.d.ts +109 -0
- package/dist/eval/live/registry-store.d.ts.map +1 -0
- package/dist/eval/live/registry-store.js +272 -0
- package/dist/eval/live/registry-store.js.map +1 -0
- package/dist/eval/live/run-checkpoint.d.ts +55 -0
- package/dist/eval/live/run-checkpoint.d.ts.map +1 -0
- package/dist/eval/live/run-checkpoint.js +86 -0
- package/dist/eval/live/run-checkpoint.js.map +1 -0
- package/dist/eval/live/run-resume.d.ts +22 -0
- package/dist/eval/live/run-resume.d.ts.map +1 -0
- package/dist/eval/live/run-resume.js +70 -0
- package/dist/eval/live/run-resume.js.map +1 -0
- package/dist/eval/live/skill-injector.d.ts +36 -0
- package/dist/eval/live/skill-injector.d.ts.map +1 -0
- package/dist/eval/live/skill-injector.js +81 -0
- package/dist/eval/live/skill-injector.js.map +1 -0
- package/dist/eval/live/skillflow-embedded-workbook-bundle.d.ts +3 -0
- package/dist/eval/live/skillflow-embedded-workbook-bundle.d.ts.map +1 -0
- package/dist/eval/live/skillflow-embedded-workbook-bundle.js +391 -0
- package/dist/eval/live/skillflow-embedded-workbook-bundle.js.map +1 -0
- package/dist/eval/live/skillflow.d.ts +154 -0
- package/dist/eval/live/skillflow.d.ts.map +1 -0
- package/dist/eval/live/skillflow.js +612 -0
- package/dist/eval/live/skillflow.js.map +1 -0
- package/dist/eval/live/skilllearnbench.d.ts +43 -0
- package/dist/eval/live/skilllearnbench.d.ts.map +1 -0
- package/dist/eval/live/skilllearnbench.js +154 -0
- package/dist/eval/live/skilllearnbench.js.map +1 -0
- package/dist/eval/live/skillopt_v1.d.ts +96 -0
- package/dist/eval/live/skillopt_v1.d.ts.map +1 -0
- package/dist/eval/live/skillopt_v1.js +0 -0
- package/dist/eval/live/skillopt_v1.js.map +1 -0
- package/dist/eval/live/skillopt_v1_causal.d.ts +26 -0
- package/dist/eval/live/skillopt_v1_causal.d.ts.map +1 -0
- package/dist/eval/live/skillopt_v1_causal.js +106 -0
- package/dist/eval/live/skillopt_v1_causal.js.map +1 -0
- package/dist/eval/live/skillopt_v1_memory.d.ts +53 -0
- package/dist/eval/live/skillopt_v1_memory.d.ts.map +1 -0
- package/dist/eval/live/skillopt_v1_memory.js +147 -0
- package/dist/eval/live/skillopt_v1_memory.js.map +1 -0
- package/dist/eval/live/skillopt_v1_rules.d.ts +39 -0
- package/dist/eval/live/skillopt_v1_rules.d.ts.map +1 -0
- package/dist/eval/live/skillopt_v1_rules.js +91 -0
- package/dist/eval/live/skillopt_v1_rules.js.map +1 -0
- package/dist/eval/live/skillsbench.d.ts +138 -0
- package/dist/eval/live/skillsbench.d.ts.map +1 -0
- package/dist/eval/live/skillsbench.js +626 -0
- package/dist/eval/live/skillsbench.js.map +1 -0
- package/dist/eval/live/stopping.d.ts +69 -0
- package/dist/eval/live/stopping.d.ts.map +1 -0
- package/dist/eval/live/stopping.js +213 -0
- package/dist/eval/live/stopping.js.map +1 -0
- package/dist/eval/live/trace.d.ts +20 -0
- package/dist/eval/live/trace.d.ts.map +1 -0
- package/dist/eval/live/trace.js +41 -0
- package/dist/eval/live/trace.js.map +1 -0
- package/dist/eval/mechanism-trace-adapter.d.ts +8 -0
- package/dist/eval/mechanism-trace-adapter.d.ts.map +1 -0
- package/dist/eval/mechanism-trace-adapter.js +118 -0
- package/dist/eval/mechanism-trace-adapter.js.map +1 -0
- package/dist/eval/memory-ab.d.ts +61 -0
- package/dist/eval/memory-ab.d.ts.map +1 -0
- package/dist/eval/memory-ab.js +161 -0
- package/dist/eval/memory-ab.js.map +1 -0
- package/dist/eval/memory-mock.d.ts +54 -0
- package/dist/eval/memory-mock.d.ts.map +1 -0
- package/dist/eval/memory-mock.js +129 -0
- package/dist/eval/memory-mock.js.map +1 -0
- package/dist/eval/memory-qa/index.d.ts +13 -0
- package/dist/eval/memory-qa/index.d.ts.map +1 -0
- package/dist/eval/memory-qa/index.js +13 -0
- package/dist/eval/memory-qa/index.js.map +1 -0
- package/dist/eval/memory-qa/iterative.d.ts +80 -0
- package/dist/eval/memory-qa/iterative.d.ts.map +1 -0
- package/dist/eval/memory-qa/iterative.js +107 -0
- package/dist/eval/memory-qa/iterative.js.map +1 -0
- package/dist/eval/memory-qa/loaders.d.ts +11 -0
- package/dist/eval/memory-qa/loaders.d.ts.map +1 -0
- package/dist/eval/memory-qa/loaders.js +21 -0
- package/dist/eval/memory-qa/loaders.js.map +1 -0
- package/dist/eval/memory-qa/metrics.d.ts +3 -0
- package/dist/eval/memory-qa/metrics.d.ts.map +1 -0
- package/dist/eval/memory-qa/metrics.js +2 -0
- package/dist/eval/memory-qa/metrics.js.map +1 -0
- package/dist/eval/memory-qa/qa.d.ts +186 -0
- package/dist/eval/memory-qa/qa.d.ts.map +1 -0
- package/dist/eval/memory-qa/qa.js +424 -0
- package/dist/eval/memory-qa/qa.js.map +1 -0
- package/dist/eval/memory-qa/retrieval.d.ts +33 -0
- package/dist/eval/memory-qa/retrieval.d.ts.map +1 -0
- package/dist/eval/memory-qa/retrieval.js +44 -0
- package/dist/eval/memory-qa/retrieval.js.map +1 -0
- package/dist/eval/memory-qa/serving.d.ts +60 -0
- package/dist/eval/memory-qa/serving.d.ts.map +1 -0
- package/dist/eval/memory-qa/serving.js +267 -0
- package/dist/eval/memory-qa/serving.js.map +1 -0
- package/dist/eval/memory-qa/types.d.ts +2 -0
- package/dist/eval/memory-qa/types.d.ts.map +1 -0
- package/dist/eval/memory-qa/types.js +2 -0
- package/dist/eval/memory-qa/types.js.map +1 -0
- package/dist/eval/metrics.d.ts +120 -0
- package/dist/eval/metrics.d.ts.map +1 -0
- package/dist/eval/metrics.js +361 -0
- package/dist/eval/metrics.js.map +1 -0
- package/dist/eval/oracle-smoke.d.ts +75 -0
- package/dist/eval/oracle-smoke.d.ts.map +1 -0
- package/dist/eval/oracle-smoke.js +143 -0
- package/dist/eval/oracle-smoke.js.map +1 -0
- package/dist/eval/playbook-calibration-report.d.ts +40 -0
- package/dist/eval/playbook-calibration-report.d.ts.map +1 -0
- package/dist/eval/playbook-calibration-report.js +246 -0
- package/dist/eval/playbook-calibration-report.js.map +1 -0
- package/dist/eval/playbook-candidate-decisions-report.d.ts +36 -0
- package/dist/eval/playbook-candidate-decisions-report.d.ts.map +1 -0
- package/dist/eval/playbook-candidate-decisions-report.js +201 -0
- package/dist/eval/playbook-candidate-decisions-report.js.map +1 -0
- package/dist/eval/playbook-usage-report.d.ts +7 -0
- package/dist/eval/playbook-usage-report.d.ts.map +1 -0
- package/dist/eval/playbook-usage-report.js +222 -0
- package/dist/eval/playbook-usage-report.js.map +1 -0
- package/dist/eval/report.d.ts +28 -0
- package/dist/eval/report.d.ts.map +1 -0
- package/dist/eval/report.js +567 -0
- package/dist/eval/report.js.map +1 -0
- package/dist/eval/retrieval/corpus.d.ts +41 -0
- package/dist/eval/retrieval/corpus.d.ts.map +1 -0
- package/dist/eval/retrieval/corpus.js +158 -0
- package/dist/eval/retrieval/corpus.js.map +1 -0
- package/dist/eval/retrieval/index.d.ts +12 -0
- package/dist/eval/retrieval/index.d.ts.map +1 -0
- package/dist/eval/retrieval/index.js +14 -0
- package/dist/eval/retrieval/index.js.map +1 -0
- package/dist/eval/retrieval/knowledge-corpus.d.ts +91 -0
- package/dist/eval/retrieval/knowledge-corpus.d.ts.map +1 -0
- package/dist/eval/retrieval/knowledge-corpus.js +268 -0
- package/dist/eval/retrieval/knowledge-corpus.js.map +1 -0
- package/dist/eval/retrieval/knowledge-quality.d.ts +91 -0
- package/dist/eval/retrieval/knowledge-quality.d.ts.map +1 -0
- package/dist/eval/retrieval/knowledge-quality.js +247 -0
- package/dist/eval/retrieval/knowledge-quality.js.map +1 -0
- package/dist/eval/retrieval/ladder.d.ts +69 -0
- package/dist/eval/retrieval/ladder.d.ts.map +1 -0
- package/dist/eval/retrieval/ladder.js +80 -0
- package/dist/eval/retrieval/ladder.js.map +1 -0
- package/dist/eval/retrieval/quality.d.ts +39 -0
- package/dist/eval/retrieval/quality.d.ts.map +1 -0
- package/dist/eval/retrieval/quality.js +75 -0
- package/dist/eval/retrieval/quality.js.map +1 -0
- package/dist/eval/rng.d.ts +25 -0
- package/dist/eval/rng.d.ts.map +1 -0
- package/dist/eval/rng.js +43 -0
- package/dist/eval/rng.js.map +1 -0
- package/dist/eval/runner.d.ts +65 -0
- package/dist/eval/runner.d.ts.map +1 -0
- package/dist/eval/runner.js +216 -0
- package/dist/eval/runner.js.map +1 -0
- package/dist/eval/skill-bundle.d.ts +18 -0
- package/dist/eval/skill-bundle.d.ts.map +1 -0
- package/dist/eval/skill-bundle.js +38 -0
- package/dist/eval/skill-bundle.js.map +1 -0
- package/dist/eval/swarmkit-report.d.ts +21 -0
- package/dist/eval/swarmkit-report.d.ts.map +1 -0
- package/dist/eval/swarmkit-report.js +93 -0
- package/dist/eval/swarmkit-report.js.map +1 -0
- package/dist/eval/task-stream.d.ts +22 -0
- package/dist/eval/task-stream.d.ts.map +1 -0
- package/dist/eval/task-stream.js +99 -0
- package/dist/eval/task-stream.js.map +1 -0
- package/dist/eval/types.d.ts +173 -0
- package/dist/eval/types.d.ts.map +1 -0
- package/dist/eval/types.js +11 -0
- package/dist/eval/types.js.map +1 -0
- package/dist/experimental/causal-store.d.ts +97 -0
- package/dist/experimental/causal-store.d.ts.map +1 -0
- package/dist/experimental/causal-store.js +213 -0
- package/dist/experimental/causal-store.js.map +1 -0
- package/dist/experimental/graph-layers/base.d.ts +29 -0
- package/dist/experimental/graph-layers/base.d.ts.map +1 -0
- package/dist/experimental/graph-layers/base.js +143 -0
- package/dist/experimental/graph-layers/base.js.map +1 -0
- package/dist/experimental/graph-layers/causal.d.ts +14 -0
- package/dist/experimental/graph-layers/causal.d.ts.map +1 -0
- package/dist/experimental/graph-layers/causal.js +14 -0
- package/dist/experimental/graph-layers/causal.js.map +1 -0
- package/dist/experimental/graph-layers/entity.d.ts +14 -0
- package/dist/experimental/graph-layers/entity.d.ts.map +1 -0
- package/dist/experimental/graph-layers/entity.js +14 -0
- package/dist/experimental/graph-layers/entity.js.map +1 -0
- package/dist/experimental/graph-layers/index.d.ts +6 -0
- package/dist/experimental/graph-layers/index.d.ts.map +1 -0
- package/dist/experimental/graph-layers/index.js +6 -0
- package/dist/experimental/graph-layers/index.js.map +1 -0
- package/dist/experimental/graph-layers/semantic.d.ts +14 -0
- package/dist/experimental/graph-layers/semantic.d.ts.map +1 -0
- package/dist/experimental/graph-layers/semantic.js +14 -0
- package/dist/experimental/graph-layers/semantic.js.map +1 -0
- package/dist/experimental/graph-layers/temporal.d.ts +14 -0
- package/dist/experimental/graph-layers/temporal.d.ts.map +1 -0
- package/dist/experimental/graph-layers/temporal.js +14 -0
- package/dist/experimental/graph-layers/temporal.js.map +1 -0
- package/dist/experimental/healing-orchestrator.d.ts +182 -0
- package/dist/experimental/healing-orchestrator.d.ts.map +1 -0
- package/dist/experimental/healing-orchestrator.js +250 -0
- package/dist/experimental/healing-orchestrator.js.map +1 -0
- package/dist/experimental/knowledge-graph-types.d.ts +148 -0
- package/dist/experimental/knowledge-graph-types.d.ts.map +1 -0
- package/dist/experimental/knowledge-graph-types.js +40 -0
- package/dist/experimental/knowledge-graph-types.js.map +1 -0
- package/dist/experimental/knowledge-graph.d.ts +41 -0
- package/dist/experimental/knowledge-graph.d.ts.map +1 -0
- package/dist/experimental/knowledge-graph.js +270 -0
- package/dist/experimental/knowledge-graph.js.map +1 -0
- package/dist/experimental/meta-reflection.d.ts +47 -0
- package/dist/experimental/meta-reflection.d.ts.map +1 -0
- package/dist/experimental/meta-reflection.js +143 -0
- package/dist/experimental/meta-reflection.js.map +1 -0
- package/dist/experimental/moe-gate.d.ts +134 -0
- package/dist/experimental/moe-gate.d.ts.map +1 -0
- package/dist/experimental/moe-gate.js +240 -0
- package/dist/experimental/moe-gate.js.map +1 -0
- package/dist/experimental/reasoning-bank.d.ts +130 -0
- package/dist/experimental/reasoning-bank.d.ts.map +1 -0
- package/dist/experimental/reasoning-bank.js +427 -0
- package/dist/experimental/reasoning-bank.js.map +1 -0
- package/dist/experimental/team/team-config.d.ts +127 -0
- package/dist/experimental/team/team-config.d.ts.map +1 -0
- package/dist/experimental/team/team-config.js +48 -0
- package/dist/experimental/team/team-config.js.map +1 -0
- package/dist/experimental/team/team-experience.d.ts +300 -0
- package/dist/experimental/team/team-experience.d.ts.map +1 -0
- package/dist/experimental/team/team-experience.js +355 -0
- package/dist/experimental/team/team-experience.js.map +1 -0
- package/dist/experimental/team/team-ingester.d.ts +152 -0
- package/dist/experimental/team/team-ingester.d.ts.map +1 -0
- package/dist/experimental/team/team-ingester.js +330 -0
- package/dist/experimental/team/team-ingester.js.map +1 -0
- package/dist/experimental/team/team-meta-learner.d.ts +51 -0
- package/dist/experimental/team/team-meta-learner.d.ts.map +1 -0
- package/dist/experimental/team/team-meta-learner.js +404 -0
- package/dist/experimental/team/team-meta-learner.js.map +1 -0
- package/dist/experimental/team/team-meta.d.ts +160 -0
- package/dist/experimental/team/team-meta.d.ts.map +1 -0
- package/dist/experimental/team/team-meta.js +42 -0
- package/dist/experimental/team/team-meta.js.map +1 -0
- package/dist/experimental/team/team-pipeline.d.ts +77 -0
- package/dist/experimental/team/team-pipeline.d.ts.map +1 -0
- package/dist/experimental/team/team-pipeline.js +262 -0
- package/dist/experimental/team/team-pipeline.js.map +1 -0
- package/dist/experimental/team/team-playbook-extraction.d.ts +45 -0
- package/dist/experimental/team/team-playbook-extraction.d.ts.map +1 -0
- package/dist/experimental/team/team-playbook-extraction.js +526 -0
- package/dist/experimental/team/team-playbook-extraction.js.map +1 -0
- package/dist/experimental/team/team-playbook.d.ts +276 -0
- package/dist/experimental/team/team-playbook.d.ts.map +1 -0
- package/dist/experimental/team/team-playbook.js +85 -0
- package/dist/experimental/team/team-playbook.js.map +1 -0
- package/dist/experimental/team/team-router.d.ts +98 -0
- package/dist/experimental/team/team-router.d.ts.map +1 -0
- package/dist/experimental/team/team-router.js +312 -0
- package/dist/experimental/team/team-router.js.map +1 -0
- package/dist/experimental/team/team-skill-library.d.ts +180 -0
- package/dist/experimental/team/team-skill-library.d.ts.map +1 -0
- package/dist/experimental/team/team-skill-library.js +385 -0
- package/dist/experimental/team/team-skill-library.js.map +1 -0
- package/dist/experimental/team/team-trajectory-analysis.d.ts +19 -0
- package/dist/experimental/team/team-trajectory-analysis.d.ts.map +1 -0
- package/dist/experimental/team/team-trajectory-analysis.js +442 -0
- package/dist/experimental/team/team-trajectory-analysis.js.map +1 -0
- package/dist/experimental/team/team-trajectory.d.ts +305 -0
- package/dist/experimental/team/team-trajectory.d.ts.map +1 -0
- package/dist/experimental/team/team-trajectory.js +315 -0
- package/dist/experimental/team/team-trajectory.js.map +1 -0
- package/dist/factory.d.ts +2 -3
- package/dist/factory.d.ts.map +1 -1
- package/dist/factory.js +12 -11
- package/dist/factory.js.map +1 -1
- package/dist/index.d.ts +9 -11
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +26 -32
- package/dist/index.js.map +1 -1
- package/dist/learning/analyzer.d.ts +26 -36
- package/dist/learning/analyzer.d.ts.map +1 -1
- package/dist/learning/analyzer.js +50 -11
- package/dist/learning/analyzer.js.map +1 -1
- package/dist/learning/compound-bootstrap/external-acquisition.d.ts +63 -0
- package/dist/learning/compound-bootstrap/external-acquisition.d.ts.map +1 -0
- package/dist/learning/compound-bootstrap/external-acquisition.js +255 -0
- package/dist/learning/compound-bootstrap/external-acquisition.js.map +1 -0
- package/dist/learning/compound-bootstrap/index.d.ts +3 -0
- package/dist/learning/compound-bootstrap/index.d.ts.map +1 -0
- package/dist/learning/compound-bootstrap/index.js +3 -0
- package/dist/learning/compound-bootstrap/index.js.map +1 -0
- package/dist/learning/compound-bootstrap/workspace-analyzer.d.ts +33 -0
- package/dist/learning/compound-bootstrap/workspace-analyzer.d.ts.map +1 -0
- package/dist/learning/compound-bootstrap/workspace-analyzer.js +379 -0
- package/dist/learning/compound-bootstrap/workspace-analyzer.js.map +1 -0
- package/dist/learning/consolidation-coordinator.d.ts +49 -0
- package/dist/learning/consolidation-coordinator.d.ts.map +1 -0
- package/dist/learning/consolidation-coordinator.js +75 -0
- package/dist/learning/consolidation-coordinator.js.map +1 -0
- package/dist/learning/convention-distiller.d.ts +84 -0
- package/dist/learning/convention-distiller.d.ts.map +1 -0
- package/dist/learning/convention-distiller.js +268 -0
- package/dist/learning/convention-distiller.js.map +1 -0
- package/dist/learning/distiller.d.ts +82 -0
- package/dist/learning/distiller.d.ts.map +1 -0
- package/dist/learning/distiller.js +353 -0
- package/dist/learning/distiller.js.map +1 -0
- package/dist/learning/effectiveness.d.ts +12 -2
- package/dist/learning/effectiveness.d.ts.map +1 -1
- package/dist/learning/effectiveness.js +18 -14
- package/dist/learning/effectiveness.js.map +1 -1
- package/dist/learning/escalation.d.ts +89 -0
- package/dist/learning/escalation.d.ts.map +1 -0
- package/dist/learning/escalation.js +168 -0
- package/dist/learning/escalation.js.map +1 -0
- package/dist/learning/external-exposure.d.ts +57 -0
- package/dist/learning/external-exposure.d.ts.map +1 -0
- package/dist/learning/external-exposure.js +95 -0
- package/dist/learning/external-exposure.js.map +1 -0
- package/dist/learning/gepa-proposer.d.ts +116 -0
- package/dist/learning/gepa-proposer.d.ts.map +1 -0
- package/dist/learning/gepa-proposer.js +258 -0
- package/dist/learning/gepa-proposer.js.map +1 -0
- package/dist/learning/index.d.ts +21 -8
- package/dist/learning/index.d.ts.map +1 -1
- package/dist/learning/index.js +28 -10
- package/dist/learning/index.js.map +1 -1
- package/dist/learning/instant-loop.d.ts +46 -16
- package/dist/learning/instant-loop.d.ts.map +1 -1
- package/dist/learning/instant-loop.js +114 -111
- package/dist/learning/instant-loop.js.map +1 -1
- package/dist/learning/knowledge-extractor.d.ts.map +1 -1
- package/dist/learning/knowledge-extractor.js +1 -2
- package/dist/learning/knowledge-extractor.js.map +1 -1
- package/dist/learning/llm-extractor.d.ts +88 -0
- package/dist/learning/llm-extractor.d.ts.map +1 -0
- package/dist/learning/llm-extractor.js +372 -0
- package/dist/learning/llm-extractor.js.map +1 -0
- package/dist/learning/loop-coordinator.d.ts +61 -0
- package/dist/learning/loop-coordinator.d.ts.map +1 -0
- package/dist/learning/loop-coordinator.js +96 -0
- package/dist/learning/loop-coordinator.js.map +1 -0
- package/dist/learning/maintenance-scheduler.d.ts +15 -4
- package/dist/learning/maintenance-scheduler.d.ts.map +1 -1
- package/dist/learning/maintenance-scheduler.js +19 -16
- package/dist/learning/maintenance-scheduler.js.map +1 -1
- package/dist/learning/mechanism-trace.d.ts +196 -0
- package/dist/learning/mechanism-trace.d.ts.map +1 -0
- package/dist/learning/mechanism-trace.js +81 -0
- package/dist/learning/mechanism-trace.js.map +1 -0
- package/dist/learning/merge-proposer.d.ts +102 -0
- package/dist/learning/merge-proposer.d.ts.map +1 -0
- package/dist/learning/merge-proposer.js +253 -0
- package/dist/learning/merge-proposer.js.map +1 -0
- package/dist/learning/meta-learner.d.ts +9 -7
- package/dist/learning/meta-learner.d.ts.map +1 -1
- package/dist/learning/meta-learner.js +16 -45
- package/dist/learning/meta-learner.js.map +1 -1
- package/dist/learning/mutation-gate.d.ts +218 -0
- package/dist/learning/mutation-gate.d.ts.map +1 -0
- package/dist/learning/mutation-gate.js +390 -0
- package/dist/learning/mutation-gate.js.map +1 -0
- package/dist/learning/oracle-signal.d.ts +67 -0
- package/dist/learning/oracle-signal.d.ts.map +1 -0
- package/dist/learning/oracle-signal.js +94 -0
- package/dist/learning/oracle-signal.js.map +1 -0
- package/dist/learning/pipeline.d.ts +31 -4
- package/dist/learning/pipeline.d.ts.map +1 -1
- package/dist/learning/pipeline.js +64 -12
- package/dist/learning/pipeline.js.map +1 -1
- package/dist/learning/playbook-extractor.d.ts +8 -0
- package/dist/learning/playbook-extractor.d.ts.map +1 -1
- package/dist/learning/playbook-extractor.js +119 -34
- package/dist/learning/playbook-extractor.js.map +1 -1
- package/dist/learning/playbook-optimizer.d.ts +37 -0
- package/dist/learning/playbook-optimizer.d.ts.map +1 -0
- package/dist/learning/playbook-optimizer.js +125 -0
- package/dist/learning/playbook-optimizer.js.map +1 -0
- package/dist/learning/reflective-proposer.d.ts +173 -0
- package/dist/learning/reflective-proposer.d.ts.map +1 -0
- package/dist/learning/reflective-proposer.js +354 -0
- package/dist/learning/reflective-proposer.js.map +1 -0
- package/dist/learning/reflexion-generator.d.ts.map +1 -1
- package/dist/learning/reflexion-generator.js +4 -3
- package/dist/learning/reflexion-generator.js.map +1 -1
- package/dist/learning/replay-executor.d.ts +64 -0
- package/dist/learning/replay-executor.d.ts.map +1 -0
- package/dist/learning/replay-executor.js +92 -0
- package/dist/learning/replay-executor.js.map +1 -0
- package/dist/learning/replay-types.d.ts +150 -0
- package/dist/learning/replay-types.d.ts.map +1 -0
- package/dist/learning/replay-types.js +207 -0
- package/dist/learning/replay-types.js.map +1 -0
- package/dist/learning/replay.d.ts +166 -0
- package/dist/learning/replay.d.ts.map +1 -0
- package/dist/learning/replay.js +228 -0
- package/dist/learning/replay.js.map +1 -0
- package/dist/learning/seed-report.d.ts +88 -0
- package/dist/learning/seed-report.d.ts.map +1 -0
- package/dist/learning/seed-report.js +120 -0
- package/dist/learning/seed-report.js.map +1 -0
- package/dist/learning/trajectory-sources/dataclaw.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/dataclaw.js +6 -18
- package/dist/learning/trajectory-sources/dataclaw.js.map +1 -1
- package/dist/learning/trajectory-sources/entire.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/entire.js +30 -9
- package/dist/learning/trajectory-sources/entire.js.map +1 -1
- package/dist/learning/trajectory-sources/file.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/file.js.map +1 -1
- package/dist/learning/trajectory-sources/huggingface.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/huggingface.js +7 -7
- package/dist/learning/trajectory-sources/huggingface.js.map +1 -1
- package/dist/learning/trajectory-sources/in-memory.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/in-memory.js.map +1 -1
- package/dist/learning/trajectory-sources/index.d.ts +3 -4
- package/dist/learning/trajectory-sources/index.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/index.js +3 -4
- package/dist/learning/trajectory-sources/index.js.map +1 -1
- package/dist/learning/trajectory-sources/pipeline.d.ts +4 -3
- package/dist/learning/trajectory-sources/pipeline.d.ts.map +1 -1
- package/dist/learning/trajectory-sources/pipeline.js +4 -3
- package/dist/learning/trajectory-sources/pipeline.js.map +1 -1
- package/dist/learning/unified-pipeline.d.ts +486 -61
- package/dist/learning/unified-pipeline.d.ts.map +1 -1
- package/dist/learning/unified-pipeline.js +1501 -232
- package/dist/learning/unified-pipeline.js.map +1 -1
- package/dist/learning/usage-inference.d.ts.map +1 -1
- package/dist/learning/usage-inference.js +4 -9
- package/dist/learning/usage-inference.js.map +1 -1
- package/dist/learning/workspace-snapshot.d.ts +146 -0
- package/dist/learning/workspace-snapshot.d.ts.map +1 -0
- package/dist/learning/workspace-snapshot.js +269 -0
- package/dist/learning/workspace-snapshot.js.map +1 -0
- package/dist/mcp/playbook-server.d.ts +61 -2
- package/dist/mcp/playbook-server.d.ts.map +1 -1
- package/dist/mcp/playbook-server.js +179 -9
- package/dist/mcp/playbook-server.js.map +1 -1
- package/dist/memory/coherence.d.ts.map +1 -1
- package/dist/memory/coherence.js +5 -7
- package/dist/memory/coherence.js.map +1 -1
- package/dist/memory/curated-loader.d.ts +24 -0
- package/dist/memory/curated-loader.d.ts.map +1 -1
- package/dist/memory/curated-loader.js +22 -5
- package/dist/memory/curated-loader.js.map +1 -1
- package/dist/memory/evolution/llm-evolver.d.ts +36 -0
- package/dist/memory/evolution/llm-evolver.d.ts.map +1 -0
- package/dist/memory/evolution/llm-evolver.js +106 -0
- package/dist/memory/evolution/llm-evolver.js.map +1 -0
- package/dist/memory/experience.d.ts +34 -13
- package/dist/memory/experience.d.ts.map +1 -1
- package/dist/memory/experience.js +59 -48
- package/dist/memory/experience.js.map +1 -1
- package/dist/memory/extraction/conversation-extractor.d.ts +62 -0
- package/dist/memory/extraction/conversation-extractor.d.ts.map +1 -0
- package/dist/memory/extraction/conversation-extractor.js +169 -0
- package/dist/memory/extraction/conversation-extractor.js.map +1 -0
- package/dist/memory/filesystem-cascade.d.ts +116 -0
- package/dist/memory/filesystem-cascade.d.ts.map +1 -0
- package/dist/memory/filesystem-cascade.js +221 -0
- package/dist/memory/filesystem-cascade.js.map +1 -0
- package/dist/memory/index.d.ts +14 -10
- package/dist/memory/index.d.ts.map +1 -1
- package/dist/memory/index.js +15 -12
- package/dist/memory/index.js.map +1 -1
- package/dist/memory/knowledge-bank.d.ts +145 -29
- package/dist/memory/knowledge-bank.d.ts.map +1 -1
- package/dist/memory/knowledge-bank.js +511 -200
- package/dist/memory/knowledge-bank.js.map +1 -1
- package/dist/memory/knowledge-defrag.d.ts +32 -9
- package/dist/memory/knowledge-defrag.d.ts.map +1 -1
- package/dist/memory/knowledge-defrag.js +57 -49
- package/dist/memory/knowledge-defrag.js.map +1 -1
- package/dist/memory/meta-playbook-bank.d.ts +24 -0
- package/dist/memory/meta-playbook-bank.d.ts.map +1 -0
- package/dist/memory/meta-playbook-bank.js +135 -0
- package/dist/memory/meta-playbook-bank.js.map +1 -0
- package/dist/memory/meta.d.ts.map +1 -1
- package/dist/memory/meta.js +33 -14
- package/dist/memory/meta.js.map +1 -1
- package/dist/memory/playbook-applicability.d.ts +43 -0
- package/dist/memory/playbook-applicability.d.ts.map +1 -0
- package/dist/memory/playbook-applicability.js +155 -0
- package/dist/memory/playbook-applicability.js.map +1 -0
- package/dist/memory/playbook-incorporation.d.ts +24 -0
- package/dist/memory/playbook-incorporation.d.ts.map +1 -0
- package/dist/memory/playbook-incorporation.js +41 -0
- package/dist/memory/playbook-incorporation.js.map +1 -0
- package/dist/memory/playbook.d.ts +444 -7
- package/dist/memory/playbook.d.ts.map +1 -1
- package/dist/memory/playbook.js +1231 -50
- package/dist/memory/playbook.js.map +1 -1
- package/dist/memory/procedural-skill.d.ts +142 -0
- package/dist/memory/procedural-skill.d.ts.map +1 -0
- package/dist/memory/procedural-skill.js +846 -0
- package/dist/memory/procedural-skill.js.map +1 -0
- package/dist/memory/search-provider.d.ts +8 -3
- package/dist/memory/search-provider.d.ts.map +1 -1
- package/dist/memory/search-providers/embedding.d.ts +64 -0
- package/dist/memory/search-providers/embedding.d.ts.map +1 -0
- package/dist/memory/search-providers/embedding.js +96 -0
- package/dist/memory/search-providers/embedding.js.map +1 -0
- package/dist/memory/search-providers/hybrid.d.ts +74 -0
- package/dist/memory/search-providers/hybrid.d.ts.map +1 -0
- package/dist/memory/search-providers/hybrid.js +94 -0
- package/dist/memory/search-providers/hybrid.js.map +1 -0
- package/dist/memory/search-providers/index.d.ts +3 -0
- package/dist/memory/search-providers/index.d.ts.map +1 -1
- package/dist/memory/search-providers/index.js +3 -0
- package/dist/memory/search-providers/index.js.map +1 -1
- package/dist/memory/search-providers/keyword-expanding.d.ts +22 -0
- package/dist/memory/search-providers/keyword-expanding.d.ts.map +1 -0
- package/dist/memory/search-providers/keyword-expanding.js +35 -0
- package/dist/memory/search-providers/keyword-expanding.js.map +1 -0
- package/dist/memory/search-providers/minimem.d.ts +16 -4
- package/dist/memory/search-providers/minimem.d.ts.map +1 -1
- package/dist/memory/search-providers/minimem.js +14 -3
- package/dist/memory/search-providers/minimem.js.map +1 -1
- package/dist/memory/search-providers/text-similarity.d.ts +2 -4
- package/dist/memory/search-providers/text-similarity.d.ts.map +1 -1
- package/dist/memory/search-providers/text-similarity.js.map +1 -1
- package/dist/memory/seeds/meta-playbook-seeds.d.ts +27 -0
- package/dist/memory/seeds/meta-playbook-seeds.d.ts.map +1 -0
- package/dist/memory/seeds/meta-playbook-seeds.js +560 -0
- package/dist/memory/seeds/meta-playbook-seeds.js.map +1 -0
- package/dist/memory/seeds/meta-strategy-seeds.d.ts +23 -0
- package/dist/memory/seeds/meta-strategy-seeds.d.ts.map +1 -0
- package/dist/memory/seeds/meta-strategy-seeds.js +82 -0
- package/dist/memory/seeds/meta-strategy-seeds.js.map +1 -0
- package/dist/memory/seeds/seed-loader.d.ts +34 -0
- package/dist/memory/seeds/seed-loader.d.ts.map +1 -0
- package/dist/memory/seeds/seed-loader.js +60 -0
- package/dist/memory/seeds/seed-loader.js.map +1 -0
- package/dist/memory/skill-exporter.d.ts.map +1 -1
- package/dist/memory/skill-exporter.js +1 -2
- package/dist/memory/skill-exporter.js.map +1 -1
- package/dist/memory/source-resolver.d.ts.map +1 -1
- package/dist/memory/source-resolver.js +7 -8
- package/dist/memory/source-resolver.js.map +1 -1
- package/dist/memory/system.d.ts +119 -14
- package/dist/memory/system.d.ts.map +1 -1
- package/dist/memory/system.js +416 -60
- package/dist/memory/system.js.map +1 -1
- package/dist/memory/temporal-compressor.d.ts +1 -1
- package/dist/memory/temporal-compressor.d.ts.map +1 -1
- package/dist/memory/temporal-compressor.js +26 -16
- package/dist/memory/temporal-compressor.js.map +1 -1
- package/dist/persistence/index.d.ts +2 -2
- package/dist/persistence/index.d.ts.map +1 -1
- package/dist/persistence/index.js +1 -1
- package/dist/persistence/index.js.map +1 -1
- package/dist/persistence/migrator.d.ts +34 -1
- package/dist/persistence/migrator.d.ts.map +1 -1
- package/dist/persistence/migrator.js +163 -33
- package/dist/persistence/migrator.js.map +1 -1
- package/dist/persistence/sqlite-persistence.d.ts +412 -1
- package/dist/persistence/sqlite-persistence.d.ts.map +1 -1
- package/dist/persistence/sqlite-persistence.js +1038 -47
- package/dist/persistence/sqlite-persistence.js.map +1 -1
- package/dist/runtime/backends/acp-protocol.d.ts.map +1 -1
- package/dist/runtime/backends/acp-protocol.js +2 -1
- package/dist/runtime/backends/acp-protocol.js.map +1 -1
- package/dist/runtime/backends/acp.d.ts.map +1 -1
- package/dist/runtime/backends/acp.js +5 -4
- package/dist/runtime/backends/acp.js.map +1 -1
- package/dist/runtime/backends/index.d.ts +2 -2
- package/dist/runtime/backends/index.d.ts.map +1 -1
- package/dist/runtime/backends/index.js +2 -2
- package/dist/runtime/backends/index.js.map +1 -1
- package/dist/runtime/backends/macro-agent.d.ts +17 -1
- package/dist/runtime/backends/macro-agent.d.ts.map +1 -1
- package/dist/runtime/backends/macro-agent.js +1 -1
- package/dist/runtime/backends/macro-agent.js.map +1 -1
- package/dist/runtime/backends/mock.d.ts.map +1 -1
- package/dist/runtime/backends/mock.js.map +1 -1
- package/dist/runtime/backends/subprocess.d.ts.map +1 -1
- package/dist/runtime/backends/subprocess.js +5 -4
- package/dist/runtime/backends/subprocess.js.map +1 -1
- package/dist/runtime/compute-provider.d.ts +2 -18
- package/dist/runtime/compute-provider.d.ts.map +1 -1
- package/dist/runtime/compute-provider.js +8 -8
- package/dist/runtime/compute-provider.js.map +1 -1
- package/dist/runtime/flows/learning.d.ts +2 -1
- package/dist/runtime/flows/learning.d.ts.map +1 -1
- package/dist/runtime/flows/learning.js +4 -5
- package/dist/runtime/flows/learning.js.map +1 -1
- package/dist/runtime/flows/validation.d.ts.map +1 -1
- package/dist/runtime/flows/validation.js +2 -5
- package/dist/runtime/flows/validation.js.map +1 -1
- package/dist/runtime/index.d.ts +1 -1
- package/dist/runtime/index.d.ts.map +1 -1
- package/dist/runtime/index.js.map +1 -1
- package/dist/runtime/manager.d.ts +128 -0
- package/dist/runtime/manager.d.ts.map +1 -1
- package/dist/runtime/manager.js +375 -64
- package/dist/runtime/manager.js.map +1 -1
- package/dist/runtime/types.d.ts +92 -1
- package/dist/runtime/types.d.ts.map +1 -1
- package/dist/search/evaluator.d.ts +71 -0
- package/dist/search/evaluator.d.ts.map +1 -1
- package/dist/search/evaluator.js +40 -5
- package/dist/search/evaluator.js.map +1 -1
- package/dist/search/index.d.ts +3 -5
- package/dist/search/index.d.ts.map +1 -1
- package/dist/search/index.js +4 -7
- package/dist/search/index.js.map +1 -1
- package/dist/search/moe-gate.d.ts +48 -2
- package/dist/search/moe-gate.d.ts.map +1 -1
- package/dist/search/moe-gate.js +117 -4
- package/dist/search/moe-gate.js.map +1 -1
- package/dist/search/refinement-loop.d.ts +63 -3
- package/dist/search/refinement-loop.d.ts.map +1 -1
- package/dist/search/refinement-loop.js +278 -23
- package/dist/search/refinement-loop.js.map +1 -1
- package/dist/search/refinement-types.d.ts +4 -2
- package/dist/search/refinement-types.d.ts.map +1 -1
- package/dist/search/refinement-types.js +2 -11
- package/dist/search/refinement-types.js.map +1 -1
- package/dist/search/router.d.ts +41 -34
- package/dist/search/router.d.ts.map +1 -1
- package/dist/search/router.js +156 -90
- package/dist/search/router.js.map +1 -1
- package/dist/search/solver.js +1 -1
- package/dist/search/solver.js.map +1 -1
- package/dist/search/task-difficulty.d.ts +54 -0
- package/dist/search/task-difficulty.d.ts.map +1 -0
- package/dist/search/task-difficulty.js +116 -0
- package/dist/search/task-difficulty.js.map +1 -0
- package/dist/search/tau-bench-verifier.d.ts +74 -0
- package/dist/search/tau-bench-verifier.d.ts.map +1 -0
- package/dist/search/tau-bench-verifier.js +96 -0
- package/dist/search/tau-bench-verifier.js.map +1 -0
- package/dist/search/verification-runner.d.ts.map +1 -1
- package/dist/search/verification-runner.js +6 -5
- package/dist/search/verification-runner.js.map +1 -1
- package/dist/session-bank/git-reader.d.ts.map +1 -1
- package/dist/session-bank/git-reader.js +5 -19
- package/dist/session-bank/git-reader.js.map +1 -1
- package/dist/session-bank/parser.d.ts.map +1 -1
- package/dist/session-bank/parser.js +26 -0
- package/dist/session-bank/parser.js.map +1 -1
- package/dist/session-bank/session-bank.d.ts.map +1 -1
- package/dist/session-bank/session-bank.js +2 -5
- package/dist/session-bank/session-bank.js.map +1 -1
- package/dist/session-bank/types.d.ts +24 -0
- package/dist/session-bank/types.d.ts.map +1 -1
- package/dist/surfacing/index.d.ts +3 -3
- package/dist/surfacing/index.d.ts.map +1 -1
- package/dist/surfacing/index.js +2 -2
- package/dist/surfacing/index.js.map +1 -1
- package/dist/surfacing/shared-bank.d.ts +43 -0
- package/dist/surfacing/shared-bank.d.ts.map +1 -0
- package/dist/surfacing/shared-bank.js +104 -0
- package/dist/surfacing/shared-bank.js.map +1 -0
- package/dist/surfacing/skill-library.d.ts +98 -2
- package/dist/surfacing/skill-library.d.ts.map +1 -1
- package/dist/surfacing/skill-library.js +192 -17
- package/dist/surfacing/skill-library.js.map +1 -1
- package/dist/surfacing/skill-publisher.d.ts +3 -1
- package/dist/surfacing/skill-publisher.d.ts.map +1 -1
- package/dist/surfacing/skill-publisher.js +28 -28
- package/dist/surfacing/skill-publisher.js.map +1 -1
- package/dist/surfacing/sqlite-storage-adapter.d.ts +2 -2
- package/dist/surfacing/sqlite-storage-adapter.d.ts.map +1 -1
- package/dist/surfacing/sqlite-storage-adapter.js +27 -18
- package/dist/surfacing/sqlite-storage-adapter.js.map +1 -1
- package/dist/types/analysis.d.ts +43 -0
- package/dist/types/analysis.d.ts.map +1 -0
- package/dist/types/analysis.js +10 -0
- package/dist/types/analysis.js.map +1 -0
- package/dist/types/config.d.ts +2178 -425
- package/dist/types/config.d.ts.map +1 -1
- package/dist/types/config.js +600 -88
- package/dist/types/config.js.map +1 -1
- package/dist/types/dataclaw.d.ts +3 -3
- package/dist/types/index.d.ts +4 -7
- package/dist/types/index.d.ts.map +1 -1
- package/dist/types/index.js +2 -10
- package/dist/types/index.js.map +1 -1
- package/dist/types/knowledge-graph.d.ts +1 -1
- package/dist/types/knowledge-graph.d.ts.map +1 -1
- package/dist/types/knowledge.d.ts +219 -11
- package/dist/types/knowledge.d.ts.map +1 -1
- package/dist/types/knowledge.js +90 -30
- package/dist/types/knowledge.js.map +1 -1
- package/dist/types/memory.d.ts +8 -0
- package/dist/types/memory.d.ts.map +1 -1
- package/dist/types/memory.js +12 -4
- package/dist/types/memory.js.map +1 -1
- package/dist/types/outcome.d.ts +11 -0
- package/dist/types/outcome.d.ts.map +1 -1
- package/dist/types/outcome.js +9 -0
- package/dist/types/outcome.js.map +1 -1
- package/dist/types/playbook.d.ts +63 -0
- package/dist/types/playbook.d.ts.map +1 -1
- package/dist/types/playbook.js +51 -0
- package/dist/types/playbook.js.map +1 -1
- package/dist/types/trajectory-source.d.ts +1 -1
- package/dist/types/trajectory-source.d.ts.map +1 -1
- package/dist/types/trajectory.d.ts +61 -0
- package/dist/types/trajectory.d.ts.map +1 -1
- package/dist/types/trajectory.js +25 -0
- package/dist/types/trajectory.js.map +1 -1
- package/dist/utils/clock.d.ts +36 -0
- package/dist/utils/clock.d.ts.map +1 -0
- package/dist/utils/clock.js +48 -0
- package/dist/utils/clock.js.map +1 -0
- package/dist/utils/frontmatter.d.ts.map +1 -1
- package/dist/utils/frontmatter.js +10 -2
- package/dist/utils/frontmatter.js.map +1 -1
- package/dist/utils/index.d.ts +4 -1
- package/dist/utils/index.d.ts.map +1 -1
- package/dist/utils/index.js +4 -1
- package/dist/utils/index.js.map +1 -1
- package/dist/utils/logger.d.ts +31 -0
- package/dist/utils/logger.d.ts.map +1 -0
- package/dist/utils/logger.js +50 -0
- package/dist/utils/logger.js.map +1 -0
- package/dist/utils/objectives.d.ts +45 -0
- package/dist/utils/objectives.d.ts.map +1 -0
- package/dist/utils/objectives.js +74 -0
- package/dist/utils/objectives.js.map +1 -0
- package/dist/utils/partitioned-store.d.ts.map +1 -1
- package/dist/utils/partitioned-store.js.map +1 -1
- package/dist/utils/playbook-hash.d.ts +19 -0
- package/dist/utils/playbook-hash.d.ts.map +1 -0
- package/dist/utils/playbook-hash.js +51 -0
- package/dist/utils/playbook-hash.js.map +1 -0
- package/dist/utils/similarity.d.ts.map +1 -1
- package/dist/utils/similarity.js.map +1 -1
- package/dist/utils/solver-canonicalizer.d.ts +22 -0
- package/dist/utils/solver-canonicalizer.d.ts.map +1 -0
- package/dist/utils/solver-canonicalizer.js +41 -0
- package/dist/utils/solver-canonicalizer.js.map +1 -0
- package/dist/utils/storage.d.ts.map +1 -1
- package/dist/utils/storage.js +1 -3
- package/dist/utils/storage.js.map +1 -1
- package/dist/utils/task-signature.d.ts +34 -0
- package/dist/utils/task-signature.d.ts.map +1 -0
- package/dist/utils/task-signature.js +112 -0
- package/dist/utils/task-signature.js.map +1 -0
- package/dist/utils/trajectory-helpers.d.ts.map +1 -1
- package/dist/utils/trajectory-helpers.js.map +1 -1
- package/dist/utils/validation.d.ts.map +1 -1
- package/dist/utils/validation.js.map +1 -1
- package/dist/workspace/efficacy-toolkit.d.ts.map +1 -1
- package/dist/workspace/efficacy-toolkit.js +18 -26
- package/dist/workspace/efficacy-toolkit.js.map +1 -1
- package/dist/workspace/index.d.ts +4 -3
- package/dist/workspace/index.d.ts.map +1 -1
- package/dist/workspace/index.js +5 -3
- package/dist/workspace/index.js.map +1 -1
- package/dist/workspace/runner.d.ts.map +1 -1
- package/dist/workspace/runner.js +95 -21
- package/dist/workspace/runner.js.map +1 -1
- package/dist/workspace/skill-converter.d.ts +66 -1
- package/dist/workspace/skill-converter.d.ts.map +1 -1
- package/dist/workspace/skill-converter.js +238 -19
- package/dist/workspace/skill-converter.js.map +1 -1
- package/dist/workspace/skill-import.d.ts +43 -0
- package/dist/workspace/skill-import.d.ts.map +1 -0
- package/dist/workspace/skill-import.js +40 -0
- package/dist/workspace/skill-import.js.map +1 -0
- package/dist/workspace/templates/index.d.ts +6 -8
- package/dist/workspace/templates/index.d.ts.map +1 -1
- package/dist/workspace/templates/index.js +7 -11
- package/dist/workspace/templates/index.js.map +1 -1
- package/dist/workspace/templates/knowledge-defrag.d.ts.map +1 -1
- package/dist/workspace/templates/knowledge-defrag.js +0 -2
- package/dist/workspace/templates/knowledge-defrag.js.map +1 -1
- package/dist/workspace/templates/knowledge-extraction.d.ts +11 -0
- package/dist/workspace/templates/knowledge-extraction.d.ts.map +1 -1
- package/dist/workspace/templates/knowledge-extraction.js +11 -6
- package/dist/workspace/templates/knowledge-extraction.js.map +1 -1
- package/dist/workspace/templates/playbook-conditioning.d.ts +38 -0
- package/dist/workspace/templates/playbook-conditioning.d.ts.map +1 -0
- package/dist/workspace/templates/playbook-conditioning.js +147 -0
- package/dist/workspace/templates/playbook-conditioning.js.map +1 -0
- package/dist/workspace/templates/playbook-decay-detection.d.ts.map +1 -1
- package/dist/workspace/templates/playbook-decay-detection.js +24 -18
- package/dist/workspace/templates/playbook-decay-detection.js.map +1 -1
- package/dist/workspace/templates/playbook-efficacy-audit.d.ts.map +1 -1
- package/dist/workspace/templates/playbook-efficacy-audit.js +26 -11
- package/dist/workspace/templates/playbook-efficacy-audit.js.map +1 -1
- package/dist/workspace/templates/playbook-extraction.d.ts.map +1 -1
- package/dist/workspace/templates/playbook-extraction.js +75 -15
- package/dist/workspace/templates/playbook-extraction.js.map +1 -1
- package/dist/workspace/templates/playbook-lifecycle-review.d.ts.map +1 -1
- package/dist/workspace/templates/playbook-lifecycle-review.js +12 -8
- package/dist/workspace/templates/playbook-lifecycle-review.js.map +1 -1
- package/dist/workspace/templates/refinement-analysis.d.ts.map +1 -1
- package/dist/workspace/templates/refinement-analysis.js +11 -5
- package/dist/workspace/templates/refinement-analysis.js.map +1 -1
- package/dist/workspace/templates/skill-enrichment.d.ts.map +1 -1
- package/dist/workspace/templates/skill-enrichment.js +7 -11
- package/dist/workspace/templates/skill-enrichment.js.map +1 -1
- package/dist/workspace/templates/solution-evaluation.d.ts.map +1 -1
- package/dist/workspace/templates/solution-evaluation.js +15 -8
- package/dist/workspace/templates/solution-evaluation.js.map +1 -1
- package/dist/workspace/templates/trajectory-analysis.d.ts.map +1 -1
- package/dist/workspace/templates/trajectory-analysis.js +1 -3
- package/dist/workspace/templates/trajectory-analysis.js.map +1 -1
- package/dist/workspace/templates/usage-inference.d.ts.map +1 -1
- package/dist/workspace/templates/usage-inference.js +9 -7
- package/dist/workspace/templates/usage-inference.js.map +1 -1
- package/dist/workspace/types.d.ts +32 -2
- package/dist/workspace/types.d.ts.map +1 -1
- package/dist/workspace/types.js +29 -0
- package/dist/workspace/types.js.map +1 -1
- package/package.json +43 -19
- package/playbooks/agent-ops/long-task-checkpointing.json +54 -0
- package/playbooks/agent-ops/progressive-context-gathering.json +56 -0
- package/playbooks/agent-ops/safe-destructive-operations.json +59 -0
- package/playbooks/agent-ops/scope-discipline.json +56 -0
- package/playbooks/agent-ops/stall-detection-strategy-switch.json +55 -0
- package/playbooks/agent-ops/verification-before-completion.json +54 -0
- package/playbooks/alfworld/clean-and-place.json +35 -0
- package/playbooks/alfworld/cool-and-place.json +35 -0
- package/playbooks/alfworld/examine-in-light.json +32 -0
- package/playbooks/alfworld/heat-and-place.json +35 -0
- package/playbooks/alfworld/pick-and-place.json +37 -0
- package/playbooks/alfworld/pick-two-and-place.json +35 -0
- package/.claude/settings.json +0 -197
- package/.entire/settings.json +0 -4
- package/.sessionlog/settings.json +0 -4
- package/CLAUDE.md +0 -233
- package/SKILL.md +0 -193
- package/docs/DESIGN-agentic-workspace.md +0 -2057
- package/docs/DESIGN-semantic-memory-knowledge-bank.md +0 -1789
- package/docs/DESIGN-session-bank.md +0 -1134
- package/docs/DESIGN-team-extraction-pipeline.md +0 -1369
- package/docs/PROPOSAL-team-aware-learning.md +0 -1080
- package/docs/RESEARCH-semantic-memory-knowledge-structures.md +0 -517
- package/docs/RESEARCH-team-trajectory-learning.md +0 -553
- package/gaps.md +0 -204
- package/scripts/migrate-to-playbooks.ts +0 -307
- package/src/agents/index.ts +0 -14
- package/src/agents/mock-provider.ts +0 -93
- package/src/agents/types.ts +0 -137
- package/src/atlas.ts +0 -1198
- package/src/bin/cli-utils.ts +0 -220
- package/src/bin/cognitive-core.ts +0 -162
- package/src/bin/commands/kb.ts +0 -266
- package/src/bin/commands/learn.ts +0 -100
- package/src/bin/commands/legacy.ts +0 -182
- package/src/bin/commands/run.ts +0 -113
- package/src/bin/commands/sessions.ts +0 -221
- package/src/bin/commands/skills.ts +0 -146
- package/src/embeddings/bm25.ts +0 -337
- package/src/embeddings/index.ts +0 -42
- package/src/embeddings/inverted-index.ts +0 -134
- package/src/embeddings/manager.ts +0 -293
- package/src/embeddings/provider.ts +0 -318
- package/src/embeddings/vector-store.ts +0 -365
- package/src/factory.ts +0 -280
- package/src/index.ts +0 -495
- package/src/learning/analyzer.ts +0 -319
- package/src/learning/effectiveness.ts +0 -428
- package/src/learning/energy-evaluator.ts +0 -282
- package/src/learning/healing-orchestrator.ts +0 -383
- package/src/learning/index.ts +0 -150
- package/src/learning/instant-loop.ts +0 -357
- package/src/learning/knowledge-extractor.ts +0 -470
- package/src/learning/maintenance-scheduler.ts +0 -271
- package/src/learning/meta-learner.ts +0 -552
- package/src/learning/playbook-extractor.ts +0 -702
- package/src/learning/reflexion-generator.ts +0 -273
- package/src/learning/team-ingester.ts +0 -499
- package/src/learning/team-meta-learner.ts +0 -558
- package/src/learning/team-pipeline.ts +0 -364
- package/src/learning/trajectory-sources/dataclaw.ts +0 -403
- package/src/learning/trajectory-sources/entire.ts +0 -251
- package/src/learning/trajectory-sources/file.ts +0 -136
- package/src/learning/trajectory-sources/huggingface.ts +0 -248
- package/src/learning/trajectory-sources/in-memory.ts +0 -61
- package/src/learning/trajectory-sources/index.ts +0 -15
- package/src/learning/trajectory-sources/pipeline.ts +0 -69
- package/src/learning/unified-pipeline.ts +0 -1191
- package/src/learning/usage-inference.ts +0 -372
- package/src/mcp/index.ts +0 -12
- package/src/mcp/playbook-server.ts +0 -565
- package/src/memory/candidate-retrieval.ts +0 -72
- package/src/memory/causal-store.ts +0 -273
- package/src/memory/coherence.ts +0 -252
- package/src/memory/curated-loader.ts +0 -213
- package/src/memory/experience.ts +0 -682
- package/src/memory/graph-layers/base.ts +0 -184
- package/src/memory/graph-layers/causal.ts +0 -16
- package/src/memory/graph-layers/entity.ts +0 -16
- package/src/memory/graph-layers/index.ts +0 -5
- package/src/memory/graph-layers/semantic.ts +0 -16
- package/src/memory/graph-layers/temporal.ts +0 -16
- package/src/memory/index.ts +0 -131
- package/src/memory/knowledge-bank.ts +0 -1317
- package/src/memory/knowledge-defrag.ts +0 -329
- package/src/memory/knowledge-graph.ts +0 -361
- package/src/memory/meta.ts +0 -488
- package/src/memory/playbook.ts +0 -676
- package/src/memory/reasoning-bank.ts +0 -458
- package/src/memory/reflexion.ts +0 -122
- package/src/memory/search-provider.ts +0 -35
- package/src/memory/search-providers/index.ts +0 -3
- package/src/memory/search-providers/minimem.ts +0 -84
- package/src/memory/search-providers/text-similarity.ts +0 -35
- package/src/memory/skill-exporter.ts +0 -357
- package/src/memory/source-resolver.ts +0 -422
- package/src/memory/system.ts +0 -620
- package/src/memory/team-experience.ts +0 -604
- package/src/memory/temporal-compressor.ts +0 -409
- package/src/persistence/index.ts +0 -37
- package/src/persistence/migrator.ts +0 -298
- package/src/persistence/serializers.ts +0 -79
- package/src/persistence/sqlite-persistence.ts +0 -925
- package/src/runtime/backends/acp-protocol.ts +0 -231
- package/src/runtime/backends/acp.ts +0 -239
- package/src/runtime/backends/index.ts +0 -38
- package/src/runtime/backends/macro-agent.ts +0 -177
- package/src/runtime/backends/mock.ts +0 -218
- package/src/runtime/backends/subprocess.ts +0 -356
- package/src/runtime/compute-provider.ts +0 -206
- package/src/runtime/flows/learning.ts +0 -166
- package/src/runtime/flows/validation.ts +0 -381
- package/src/runtime/index.ts +0 -75
- package/src/runtime/manager.ts +0 -685
- package/src/runtime/types.ts +0 -198
- package/src/search/evaluator.ts +0 -499
- package/src/search/index.ts +0 -78
- package/src/search/moe-gate.ts +0 -304
- package/src/search/refinement-loop.ts +0 -457
- package/src/search/refinement-types.ts +0 -159
- package/src/search/router.ts +0 -368
- package/src/search/solver.ts +0 -303
- package/src/search/team-router.ts +0 -459
- package/src/search/verification-runner.ts +0 -570
- package/src/session-bank/git-reader.ts +0 -200
- package/src/session-bank/index.ts +0 -26
- package/src/session-bank/parser.ts +0 -548
- package/src/session-bank/session-bank.ts +0 -485
- package/src/session-bank/types.ts +0 -176
- package/src/surfacing/index.ts +0 -29
- package/src/surfacing/publisher.ts +0 -23
- package/src/surfacing/skill-library.ts +0 -625
- package/src/surfacing/skill-publisher.ts +0 -257
- package/src/surfacing/sqlite-storage-adapter.ts +0 -285
- package/src/surfacing/team-skill-library.ts +0 -610
- package/src/types/config.ts +0 -661
- package/src/types/dataclaw.ts +0 -99
- package/src/types/huggingface-transformers.d.ts +0 -12
- package/src/types/index.ts +0 -293
- package/src/types/knowledge-graph.ts +0 -246
- package/src/types/knowledge.ts +0 -388
- package/src/types/memory.ts +0 -300
- package/src/types/meta.ts +0 -218
- package/src/types/outcome.ts +0 -66
- package/src/types/playbook.ts +0 -215
- package/src/types/step.ts +0 -40
- package/src/types/task.ts +0 -52
- package/src/types/team-meta.ts +0 -212
- package/src/types/team-playbook.ts +0 -384
- package/src/types/team-trajectory.ts +0 -673
- package/src/types/trajectory-source.ts +0 -47
- package/src/types/trajectory.ts +0 -80
- package/src/utils/error-classifier.ts +0 -113
- package/src/utils/frontmatter.ts +0 -118
- package/src/utils/index.ts +0 -63
- package/src/utils/partitioned-store.ts +0 -299
- package/src/utils/similarity.ts +0 -139
- package/src/utils/storage.ts +0 -249
- package/src/utils/trajectory-helpers.ts +0 -79
- package/src/utils/validation.ts +0 -286
- package/src/workspace/efficacy-toolkit.ts +0 -496
- package/src/workspace/index.ts +0 -77
- package/src/workspace/runner.ts +0 -281
- package/src/workspace/skill-converter.ts +0 -288
- package/src/workspace/templates/index.ts +0 -95
- package/src/workspace/templates/knowledge-defrag.ts +0 -223
- package/src/workspace/templates/knowledge-extraction.ts +0 -337
- package/src/workspace/templates/meta-reflection.ts +0 -208
- package/src/workspace/templates/playbook-decay-detection.ts +0 -272
- package/src/workspace/templates/playbook-efficacy-audit.ts +0 -246
- package/src/workspace/templates/playbook-extraction.ts +0 -240
- package/src/workspace/templates/playbook-lifecycle-review.ts +0 -274
- package/src/workspace/templates/refinement-analysis.ts +0 -162
- package/src/workspace/templates/skill-enrichment.ts +0 -275
- package/src/workspace/templates/solution-evaluation.ts +0 -199
- package/src/workspace/templates/team-playbook-extraction.ts +0 -631
- package/src/workspace/templates/team-trajectory-analysis.ts +0 -564
- package/src/workspace/templates/trajectory-analysis.ts +0 -234
- package/src/workspace/templates/usage-inference.ts +0 -163
- package/src/workspace/types.ts +0 -185
- package/tests/atlas-knowledge.test.ts +0 -165
- package/tests/embeddings/bm25.test.ts +0 -130
- package/tests/embeddings/inverted-index.test.ts +0 -138
- package/tests/embeddings/manager.test.ts +0 -205
- package/tests/feature-toggles.test.ts +0 -275
- package/tests/fixtures/behavioral-trajectories.ts +0 -210
- package/tests/gap-fixes.test.ts +0 -1116
- package/tests/integration/atlas.test.ts +0 -266
- package/tests/integration/cli-e2e.test.ts +0 -621
- package/tests/integration/curated-sources-e2e.test.ts +0 -502
- package/tests/integration/dataclaw-e2e.test.ts +0 -559
- package/tests/integration/e2e.test.ts +0 -1337
- package/tests/integration/entire-e2e.test.ts +0 -376
- package/tests/integration/huggingface-e2e.test.ts +0 -627
- package/tests/integration/persistence-e2e.test.ts +0 -741
- package/tests/integration/phase-e2e.test.ts +0 -1143
- package/tests/integration/pipeline-data-correctness.test.ts +0 -794
- package/tests/integration/session-bank.test.ts +0 -231
- package/tests/integration/sessionlog-e2e.test.ts +0 -329
- package/tests/integration/skill-tree-wiring.test.ts +0 -152
- package/tests/integration/unified-pipeline-e2e.test.ts +0 -634
- package/tests/learning/analyzer.test.ts +0 -426
- package/tests/learning/dataclaw-trajectory-source.test.ts +0 -642
- package/tests/learning/effectiveness.test.ts +0 -542
- package/tests/learning/energy-evaluator.test.ts +0 -180
- package/tests/learning/entire-trajectory-source.test.ts +0 -380
- package/tests/learning/healing-orchestrator.test.ts +0 -269
- package/tests/learning/huggingface-trajectory-source.test.ts +0 -817
- package/tests/learning/instant-loop.test.ts +0 -243
- package/tests/learning/knowledge-extractor.test.ts +0 -491
- package/tests/learning/maintenance-scheduler.test.ts +0 -191
- package/tests/learning/meta-learner.test.ts +0 -418
- package/tests/learning/pipeline-memory-updates.test.ts +0 -721
- package/tests/learning/playbook-extractor-provenance.test.ts +0 -114
- package/tests/learning/reflexion-generator.test.ts +0 -411
- package/tests/learning/team-ingester.test.ts +0 -349
- package/tests/learning/team-meta-learner.test.ts +0 -618
- package/tests/learning/team-pipeline.test.ts +0 -334
- package/tests/learning/trajectory-sources.test.ts +0 -320
- package/tests/learning/unified-pipeline-efficacy.test.ts +0 -232
- package/tests/learning/unified-pipeline.test.ts +0 -322
- package/tests/learning/usage-inference.test.ts +0 -254
- package/tests/mcp/playbook-server.test.ts +0 -257
- package/tests/memory/candidate-retrieval.test.ts +0 -167
- package/tests/memory/causal-store.test.ts +0 -276
- package/tests/memory/coherence.test.ts +0 -232
- package/tests/memory/compound-engineering-seed.test.ts +0 -338
- package/tests/memory/curated-loader-extended.test.ts +0 -225
- package/tests/memory/experience.test.ts +0 -203
- package/tests/memory/graph-layers/semantic.test.ts +0 -219
- package/tests/memory/knowledge-bank-extraction.test.ts +0 -558
- package/tests/memory/knowledge-bank.test.ts +0 -705
- package/tests/memory/knowledge-defrag.test.ts +0 -366
- package/tests/memory/knowledge-evolution.test.ts +0 -563
- package/tests/memory/knowledge-graph.test.ts +0 -492
- package/tests/memory/knowledge-inbox.test.ts +0 -258
- package/tests/memory/knowledge-minimem.test.ts +0 -251
- package/tests/memory/meta.test.ts +0 -399
- package/tests/memory/playbook-quality-validation.test.ts +0 -430
- package/tests/memory/playbook.test.ts +0 -644
- package/tests/memory/provenance.test.ts +0 -648
- package/tests/memory/reasoning-bank.test.ts +0 -239
- package/tests/memory/reflexion.test.ts +0 -166
- package/tests/memory/skill-exporter.test.ts +0 -475
- package/tests/memory/source-resolver.test.ts +0 -700
- package/tests/memory/system.test.ts +0 -330
- package/tests/memory/team-experience.test.ts +0 -411
- package/tests/memory/temporal-compressor.test.ts +0 -318
- package/tests/persistence/migrator.test.ts +0 -1009
- package/tests/persistence/sqlite-persistence.test.ts +0 -635
- package/tests/runtime/agent-manager.test.ts +0 -517
- package/tests/runtime/compute-provider.test.ts +0 -288
- package/tests/runtime/delegate.test.ts +0 -354
- package/tests/runtime/macro-agent-backend.test.ts +0 -266
- package/tests/runtime/mock-backend.test.ts +0 -248
- package/tests/search/evaluator.test.ts +0 -257
- package/tests/search/moe-gate.test.ts +0 -250
- package/tests/search/refinement-loop.test.ts +0 -477
- package/tests/search/refinement.test.ts +0 -267
- package/tests/search/router.test.ts +0 -506
- package/tests/search/team-router.test.ts +0 -376
- package/tests/search/verification-runner.test.ts +0 -357
- package/tests/session-bank/fixtures/multi-tool-session/full.jsonl +0 -12
- package/tests/session-bank/fixtures/multi-tool-session/metadata.json +0 -28
- package/tests/session-bank/fixtures/root-metadata.json +0 -18
- package/tests/session-bank/fixtures/sessionlog-root-metadata.json +0 -16
- package/tests/session-bank/fixtures/sessionlog-session/full.jsonl +0 -6
- package/tests/session-bank/fixtures/sessionlog-session/metadata.json +0 -55
- package/tests/session-bank/fixtures/simple-session/full.jsonl +0 -6
- package/tests/session-bank/fixtures/simple-session/metadata.json +0 -38
- package/tests/session-bank/git-reader.test.ts +0 -232
- package/tests/session-bank/parser.test.ts +0 -453
- package/tests/session-bank/session-bank.test.ts +0 -546
- package/tests/surfacing/skill-library.test.ts +0 -297
- package/tests/surfacing/skill-publisher.test.ts +0 -437
- package/tests/surfacing/sqlite-storage-adapter.test.ts +0 -206
- package/tests/surfacing/team-skill-library.test.ts +0 -444
- package/tests/types/outcome.test.ts +0 -147
- package/tests/types/step.test.ts +0 -133
- package/tests/types/task.test.ts +0 -158
- package/tests/types/team-meta.test.ts +0 -147
- package/tests/types/team-playbook.test.ts +0 -246
- package/tests/types/team-trajectory.test.ts +0 -557
- package/tests/types/trajectory.test.ts +0 -253
- package/tests/utils/error-classifier.test.ts +0 -149
- package/tests/utils/frontmatter.test.ts +0 -208
- package/tests/utils/partitioned-store.test.ts +0 -230
- package/tests/utils/similarity.test.ts +0 -188
- package/tests/utils/trajectory-helpers.test.ts +0 -163
- package/tests/utils/validation.test.ts +0 -252
- package/tests/workspace/efficacy-toolkit.test.ts +0 -404
- package/tests/workspace/full-flow.test.ts +0 -845
- package/tests/workspace/manager.test.ts +0 -215
- package/tests/workspace/runner.test.ts +0 -336
- package/tests/workspace/skill-converter.test.ts +0 -205
- package/tests/workspace/templates/knowledge-extraction.test.ts +0 -235
- package/tests/workspace/templates/playbook-efficacy.test.ts +0 -377
- package/tests/workspace/templates/team-playbook-extraction.test.ts +0 -341
- package/tests/workspace/templates/team-trajectory-analysis.test.ts +0 -417
- package/tsconfig.json +0 -25
- package/vitest.config.ts +0 -22
|
@@ -0,0 +1,314 @@
|
|
|
1
|
+
import { spawn } from 'node:child_process';
|
|
2
|
+
import { existsSync } from 'node:fs';
|
|
3
|
+
import { delimiter, dirname } from 'node:path';
|
|
4
|
+
import { generateResponsesText, isResponsesProvider } from 'swarmkit-eval';
|
|
5
|
+
/** Tools disabled when `disallowTools` is set — forces a single-inference step. */
|
|
6
|
+
const DISALLOWED_TOOLS = [
|
|
7
|
+
'Bash',
|
|
8
|
+
'Read',
|
|
9
|
+
'Write',
|
|
10
|
+
'Edit',
|
|
11
|
+
'NotebookEdit',
|
|
12
|
+
'Glob',
|
|
13
|
+
'Grep',
|
|
14
|
+
'WebFetch',
|
|
15
|
+
'WebSearch',
|
|
16
|
+
'Task',
|
|
17
|
+
'TodoWrite',
|
|
18
|
+
];
|
|
19
|
+
const CLAUDE_COMMAND_FALLBACKS = [
|
|
20
|
+
'/usr/local/bin/claude',
|
|
21
|
+
'/opt/homebrew/bin/claude',
|
|
22
|
+
'/Applications/cmux.app/Contents/Resources/bin/claude',
|
|
23
|
+
];
|
|
24
|
+
export function resolveClaudeCommand() {
|
|
25
|
+
const configured = process.env.CLAUDE_CLI_PATH;
|
|
26
|
+
if (configured && existsSync(configured))
|
|
27
|
+
return configured;
|
|
28
|
+
for (const candidate of CLAUDE_COMMAND_FALLBACKS) {
|
|
29
|
+
if (existsSync(candidate))
|
|
30
|
+
return candidate;
|
|
31
|
+
}
|
|
32
|
+
return 'claude';
|
|
33
|
+
}
|
|
34
|
+
export function claudePathEntries(command = resolveClaudeCommand()) {
|
|
35
|
+
const entries = new Set();
|
|
36
|
+
if (command !== 'claude')
|
|
37
|
+
entries.add(dirname(command));
|
|
38
|
+
for (const candidate of CLAUDE_COMMAND_FALLBACKS) {
|
|
39
|
+
if (existsSync(candidate))
|
|
40
|
+
entries.add(dirname(candidate));
|
|
41
|
+
}
|
|
42
|
+
return [...entries];
|
|
43
|
+
}
|
|
44
|
+
export function buildClaudeChildEnv(base, command = resolveClaudeCommand()) {
|
|
45
|
+
const pathEntries = claudePathEntries(command);
|
|
46
|
+
const existingPath = base.PATH ? [base.PATH] : [];
|
|
47
|
+
return {
|
|
48
|
+
...base,
|
|
49
|
+
CLAUDE_CLI_PATH: command !== 'claude' ? command : base.CLAUDE_CLI_PATH,
|
|
50
|
+
PATH: [...pathEntries, ...existingPath].filter(Boolean).join(delimiter),
|
|
51
|
+
};
|
|
52
|
+
}
|
|
53
|
+
/**
|
|
54
|
+
* Friendly model name → Bedrock cross-region inference-profile id. Used for
|
|
55
|
+
* ANTHROPIC_MODEL when `bedrock` is set. Unknown names pass through (so a full
|
|
56
|
+
* profile id can be supplied directly). Ids verified against the account's
|
|
57
|
+
* enabled profiles (us-east-1).
|
|
58
|
+
*/
|
|
59
|
+
export const BEDROCK_MODELS = {
|
|
60
|
+
haiku: 'us.anthropic.claude-haiku-4-5-20251001-v1:0',
|
|
61
|
+
sonnet: 'us.anthropic.claude-sonnet-4-6',
|
|
62
|
+
opus: 'us.anthropic.claude-opus-4-1-20250805-v1:0',
|
|
63
|
+
};
|
|
64
|
+
/**
|
|
65
|
+
* Cross-region inference profiles are geo-prefixed (`us.`, `eu.`, `apac.`) and
|
|
66
|
+
* a profile only resolves from a region in its own geo. Daily token quotas are
|
|
67
|
+
* per calling region, so being able to point runs at eu-* regions matters when
|
|
68
|
+
* the us-* quotas are exhausted.
|
|
69
|
+
*/
|
|
70
|
+
function bedrockGeoPrefix(region) {
|
|
71
|
+
if (region.startsWith('eu-'))
|
|
72
|
+
return 'eu';
|
|
73
|
+
if (region.startsWith('ap-'))
|
|
74
|
+
return 'apac';
|
|
75
|
+
return 'us';
|
|
76
|
+
}
|
|
77
|
+
export function bedrockModelId(name, region) {
|
|
78
|
+
const id = BEDROCK_MODELS[name] ?? name;
|
|
79
|
+
const prefix = bedrockGeoPrefix(region ?? 'us-east-1');
|
|
80
|
+
return prefix === 'us' ? id : id.replace(/^us\./, `${prefix}.`);
|
|
81
|
+
}
|
|
82
|
+
/**
|
|
83
|
+
* Run the Claude CLI with retry-on-failure. A single hung/throttled `claude -p`
|
|
84
|
+
* call should not abort an entire multi-hour eval; we retry a couple of times
|
|
85
|
+
* with backoff before giving up.
|
|
86
|
+
*/
|
|
87
|
+
export async function runClaude(prompt, opts = {}) {
|
|
88
|
+
const retries = opts.retries ?? 2;
|
|
89
|
+
let lastErr;
|
|
90
|
+
let nextCommandOverride = opts.commandOverride;
|
|
91
|
+
for (let attempt = 0; attempt <= retries; attempt++) {
|
|
92
|
+
try {
|
|
93
|
+
const attemptOpts = nextCommandOverride
|
|
94
|
+
? { ...opts, commandOverride: nextCommandOverride }
|
|
95
|
+
: opts;
|
|
96
|
+
return await (isResponsesProvider(opts.provider)
|
|
97
|
+
? runResponsesOnce(prompt, attemptOpts)
|
|
98
|
+
: runClaudeOnce(prompt, attemptOpts));
|
|
99
|
+
}
|
|
100
|
+
catch (err) {
|
|
101
|
+
lastErr = err;
|
|
102
|
+
// Surface the failure cause so Bedrock throttling/timeouts are visible
|
|
103
|
+
// instead of being silently retried away.
|
|
104
|
+
console.error(`[claude-cli] attempt ${attempt + 1}/${retries + 1} failed: ${lastErr.message.slice(0, 300)}`);
|
|
105
|
+
if (isClaudePathLookupError(lastErr)) {
|
|
106
|
+
const fallback = alternateClaudeCommand(nextCommandOverride ?? resolveClaudeCommand());
|
|
107
|
+
if (fallback) {
|
|
108
|
+
nextCommandOverride = fallback;
|
|
109
|
+
console.error(`[claude-cli] retrying with alternate Claude binary: ${fallback}`);
|
|
110
|
+
}
|
|
111
|
+
}
|
|
112
|
+
if (attempt < retries) {
|
|
113
|
+
// Linear backoff: 5s, 10s — gives a throttled CLI room to recover.
|
|
114
|
+
await new Promise((r) => setTimeout(r, 5000 * (attempt + 1)));
|
|
115
|
+
}
|
|
116
|
+
}
|
|
117
|
+
}
|
|
118
|
+
throw lastErr ?? new Error('claude CLI failed');
|
|
119
|
+
}
|
|
120
|
+
/**
|
|
121
|
+
* Bridge the Azure credential naming gap for the Responses solver/learning path.
|
|
122
|
+
*
|
|
123
|
+
* swarmkit-eval's `generateResponsesText` only reads the canonical
|
|
124
|
+
* `AZURE_OPENAI_*` env names, but this environment commonly exports the shorter
|
|
125
|
+
* `AZURE_API_*` aliases (endpoint/version/key). Resolve both, canonical names
|
|
126
|
+
* winning, and pass them explicitly so `--provider azure` works with whichever
|
|
127
|
+
* credential set is present. Deployment stays `AZURE_OPENAI_DEPLOYMENT` only;
|
|
128
|
+
* when unset the caller's `--model` becomes the deployment name downstream.
|
|
129
|
+
*/
|
|
130
|
+
export function resolveAzureResponsesEnv(env = process.env) {
|
|
131
|
+
const apiKey = env.AZURE_OPENAI_API_KEY ?? env.AZURE_OPENAI_KEY ?? env.AZURE_API_KEY;
|
|
132
|
+
const azureEndpoint = env.AZURE_OPENAI_ENDPOINT ?? env.AZURE_OPENAI_BASE_URL ?? env.AZURE_API_BASE;
|
|
133
|
+
const azureApiVersion = env.AZURE_OPENAI_API_VERSION ?? env.AZURE_API_VERSION;
|
|
134
|
+
const azureDeployment = env.AZURE_OPENAI_DEPLOYMENT;
|
|
135
|
+
const out = {};
|
|
136
|
+
if (apiKey)
|
|
137
|
+
out.apiKey = apiKey;
|
|
138
|
+
if (azureEndpoint)
|
|
139
|
+
out.azureEndpoint = azureEndpoint;
|
|
140
|
+
if (azureApiVersion)
|
|
141
|
+
out.azureApiVersion = azureApiVersion;
|
|
142
|
+
if (azureDeployment)
|
|
143
|
+
out.azureDeployment = azureDeployment;
|
|
144
|
+
return out;
|
|
145
|
+
}
|
|
146
|
+
async function runResponsesOnce(prompt, opts = {}) {
|
|
147
|
+
const azure = opts.provider === 'azure' || opts.provider === 'azure-openai' ? resolveAzureResponsesEnv() : {};
|
|
148
|
+
const result = await generateResponsesText({
|
|
149
|
+
provider: opts.provider,
|
|
150
|
+
model: opts.model ?? 'gpt-5.5',
|
|
151
|
+
input: prompt,
|
|
152
|
+
maxOutputTokens: Number(process.env.OPENAI_MAX_OUTPUT_TOKENS) || undefined,
|
|
153
|
+
reasoningEffort: parseReasoningEffort(process.env.OPENAI_REASONING_EFFORT),
|
|
154
|
+
timeoutMs: opts.timeoutMs,
|
|
155
|
+
...azure,
|
|
156
|
+
});
|
|
157
|
+
return {
|
|
158
|
+
text: result.text,
|
|
159
|
+
ok: true,
|
|
160
|
+
};
|
|
161
|
+
}
|
|
162
|
+
function parseReasoningEffort(value) {
|
|
163
|
+
if (value === 'minimal' || value === 'low' || value === 'medium' || value === 'high') {
|
|
164
|
+
return value;
|
|
165
|
+
}
|
|
166
|
+
return undefined;
|
|
167
|
+
}
|
|
168
|
+
/**
|
|
169
|
+
* Invoke `claude -p --output-format json` with `prompt` on stdin and parse the
|
|
170
|
+
* single JSON result. Rejects on spawn error, non-zero exit, timeout, or
|
|
171
|
+
* unparseable output.
|
|
172
|
+
*/
|
|
173
|
+
function runClaudeOnce(prompt, opts = {}) {
|
|
174
|
+
// Always isolate from host plugins/settings. Otherwise every `claude -p` loads
|
|
175
|
+
// the user's global plugins (omc/swarm), each spawning ~5 MCP subprocesses that
|
|
176
|
+
// leak and hang at MCP-init — the dominant cause of multi-minute eval stalls.
|
|
177
|
+
// --strict-mcp-config (no host MCP) + empty --setting-sources keeps eval calls
|
|
178
|
+
// clean and fast (auth via ~/.claude / Bedrock env is unaffected).
|
|
179
|
+
const args = ['-p', '--output-format', 'json', '--strict-mcp-config', '--setting-sources', ''];
|
|
180
|
+
const command = opts.commandOverride ?? resolveClaudeCommand();
|
|
181
|
+
const env = buildClaudeChildEnv(process.env, command);
|
|
182
|
+
// `provider: 'bedrock'` must take the SAME env-var routing as `bedrock: true`.
|
|
183
|
+
// The old behavior passed `--settings '{"apiProvider":"bedrock"}'` — which is
|
|
184
|
+
// not a Claude Code setting, so the CLI silently routed to the user's
|
|
185
|
+
// Anthropic subscription instead of AWS. Found 2026-07-02 when "bedrock"
|
|
186
|
+
// SkillFlow runs died with a subscription 429 ("session limit") — every
|
|
187
|
+
// adapter that passed provider without bedrock (skillflow, clbench,
|
|
188
|
+
// skillsbench) had been burning the subscription.
|
|
189
|
+
const useBedrock = opts.bedrock || opts.provider === 'bedrock';
|
|
190
|
+
if (useBedrock) {
|
|
191
|
+
// Env-var Bedrock routing; DROP --model (CLI aliases are invalid Bedrock ids,
|
|
192
|
+
// → 400). Route the model via ANTHROPIC_MODEL = inference-profile id.
|
|
193
|
+
env.CLAUDE_CODE_USE_BEDROCK = '1';
|
|
194
|
+
env.AWS_REGION = opts.awsRegion ?? process.env.AWS_REGION ?? 'us-east-1';
|
|
195
|
+
if (opts.model)
|
|
196
|
+
env.ANTHROPIC_MODEL = bedrockModelId(opts.model, env.AWS_REGION);
|
|
197
|
+
}
|
|
198
|
+
else if (opts.model) {
|
|
199
|
+
args.push('--model', opts.model);
|
|
200
|
+
}
|
|
201
|
+
if (opts.resumeSessionId)
|
|
202
|
+
args.push('--resume', opts.resumeSessionId);
|
|
203
|
+
if (opts.permissionMode) {
|
|
204
|
+
args.push('--permission-mode', opts.permissionMode);
|
|
205
|
+
}
|
|
206
|
+
if (opts.allowedTools && opts.allowedTools.length > 0) {
|
|
207
|
+
args.push('--allowedTools', ...opts.allowedTools);
|
|
208
|
+
}
|
|
209
|
+
if (opts.provider && !useBedrock) {
|
|
210
|
+
// Legacy settings pass-through for non-bedrock providers. NOTE: this is
|
|
211
|
+
// dubious — `apiProvider` is not a documented Claude Code setting — but
|
|
212
|
+
// responses-providers (openai/azure) never reach runClaudeOnce, so this
|
|
213
|
+
// only fires for exotic values. Bedrock is env-routed above.
|
|
214
|
+
const settings = { apiProvider: opts.provider };
|
|
215
|
+
args.push('--settings', JSON.stringify(settings));
|
|
216
|
+
}
|
|
217
|
+
if (opts.disallowTools) {
|
|
218
|
+
args.push('--disallowedTools', ...DISALLOWED_TOOLS);
|
|
219
|
+
}
|
|
220
|
+
const timeoutMs = opts.timeoutMs ?? 90_000;
|
|
221
|
+
return new Promise((resolve, reject) => {
|
|
222
|
+
const child = spawn(command, args, {
|
|
223
|
+
stdio: ['pipe', 'pipe', 'pipe'],
|
|
224
|
+
env,
|
|
225
|
+
cwd: opts.cwd,
|
|
226
|
+
});
|
|
227
|
+
let stdout = '';
|
|
228
|
+
let stderr = '';
|
|
229
|
+
let settled = false;
|
|
230
|
+
const timer = setTimeout(() => {
|
|
231
|
+
if (settled)
|
|
232
|
+
return;
|
|
233
|
+
settled = true;
|
|
234
|
+
child.kill('SIGKILL');
|
|
235
|
+
reject(new Error(`claude CLI timed out after ${timeoutMs}ms`));
|
|
236
|
+
}, timeoutMs);
|
|
237
|
+
child.stdout.on('data', (d) => (stdout += d.toString()));
|
|
238
|
+
child.stderr.on('data', (d) => (stderr += d.toString()));
|
|
239
|
+
child.on('error', (err) => {
|
|
240
|
+
if (settled)
|
|
241
|
+
return;
|
|
242
|
+
settled = true;
|
|
243
|
+
clearTimeout(timer);
|
|
244
|
+
reject(new Error(`Failed to spawn claude CLI: ${err.message}`));
|
|
245
|
+
});
|
|
246
|
+
child.on('close', (code) => {
|
|
247
|
+
if (settled)
|
|
248
|
+
return;
|
|
249
|
+
settled = true;
|
|
250
|
+
clearTimeout(timer);
|
|
251
|
+
if (code !== 0) {
|
|
252
|
+
const errMsg = stderr.slice(0, 500);
|
|
253
|
+
// Nonzero exit with no stdout is an infra failure (CLI crash, auth
|
|
254
|
+
// error, rate-limit). Use a message pattern isEnvError() can match.
|
|
255
|
+
if (!stdout.trim()) {
|
|
256
|
+
reject(new Error(`claude CLI exited ${code} with no output: ${errMsg}`));
|
|
257
|
+
return;
|
|
258
|
+
}
|
|
259
|
+
reject(new Error(`claude CLI exited ${code}: ${errMsg}`));
|
|
260
|
+
return;
|
|
261
|
+
}
|
|
262
|
+
try {
|
|
263
|
+
const json = JSON.parse(stdout);
|
|
264
|
+
// Bedrock throttling/errors surface as api_error_status or is_error even
|
|
265
|
+
// on a clean exit. Log them so rate-limiting is diagnosable.
|
|
266
|
+
if (json.api_error_status || json.is_error === true) {
|
|
267
|
+
const status = String(json.api_error_status ?? '');
|
|
268
|
+
const subtype = String(json.subtype ?? '');
|
|
269
|
+
console.error(`[claude-cli] API error (provider=${opts.provider ?? 'default'}): ` +
|
|
270
|
+
`status=${status} subtype=${subtype} ` +
|
|
271
|
+
`result=${String(json.result ?? '').slice(0, 200)}`);
|
|
272
|
+
// 429 / 529 rate-limit or overload — surface as env_error-classifiable.
|
|
273
|
+
if (status === '429' ||
|
|
274
|
+
status === '529' ||
|
|
275
|
+
/rate.?limit|too many requests|overloaded/i.test(subtype)) {
|
|
276
|
+
reject(new Error(`claude CLI API ${status || subtype}: rate-limit/overloaded`));
|
|
277
|
+
return;
|
|
278
|
+
}
|
|
279
|
+
}
|
|
280
|
+
resolve({
|
|
281
|
+
text: typeof json.result === 'string' ? json.result : '',
|
|
282
|
+
sessionId: json.session_id,
|
|
283
|
+
ok: json.is_error !== true && !json.api_error_status,
|
|
284
|
+
costUsd: typeof json.total_cost_usd === 'number' ? json.total_cost_usd : undefined,
|
|
285
|
+
});
|
|
286
|
+
}
|
|
287
|
+
catch (err) {
|
|
288
|
+
reject(new Error(`Could not parse claude CLI JSON output: ${err.message}. ` +
|
|
289
|
+
`First 300 chars: ${stdout.slice(0, 300)}`));
|
|
290
|
+
}
|
|
291
|
+
});
|
|
292
|
+
child.stdin.write(prompt);
|
|
293
|
+
child.stdin.end();
|
|
294
|
+
});
|
|
295
|
+
}
|
|
296
|
+
function isClaudePathLookupError(error) {
|
|
297
|
+
return /claude not found in PATH/i.test(error.message);
|
|
298
|
+
}
|
|
299
|
+
function alternateClaudeCommand(current) {
|
|
300
|
+
for (const candidate of CLAUDE_COMMAND_FALLBACKS) {
|
|
301
|
+
if (candidate !== current && existsSync(candidate))
|
|
302
|
+
return candidate;
|
|
303
|
+
}
|
|
304
|
+
return undefined;
|
|
305
|
+
}
|
|
306
|
+
/** Quick availability probe — true if the `claude` binary is on PATH. */
|
|
307
|
+
export async function claudeAvailable() {
|
|
308
|
+
return new Promise((resolve) => {
|
|
309
|
+
const child = spawn(resolveClaudeCommand(), ['--version'], { stdio: 'ignore' });
|
|
310
|
+
child.on('error', () => resolve(false));
|
|
311
|
+
child.on('close', (code) => resolve(code === 0));
|
|
312
|
+
});
|
|
313
|
+
}
|
|
314
|
+
//# sourceMappingURL=claude-cli.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-cli.js","sourceRoot":"","sources":["../../../src/eval/live/claude-cli.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAC3C,OAAO,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AACrC,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAC/C,OAAO,EAAE,qBAAqB,EAAE,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAoD3E,mFAAmF;AACnF,MAAM,gBAAgB,GAAG;IACvB,MAAM;IACN,MAAM;IACN,OAAO;IACP,MAAM;IACN,cAAc;IACd,MAAM;IACN,MAAM;IACN,UAAU;IACV,WAAW;IACX,MAAM;IACN,WAAW;CACZ,CAAC;AAEF,MAAM,wBAAwB,GAAG;IAC/B,uBAAuB;IACvB,0BAA0B;IAC1B,sDAAsD;CACvD,CAAC;AAEF,MAAM,UAAU,oBAAoB;IAClC,MAAM,UAAU,GAAG,OAAO,CAAC,GAAG,CAAC,eAAe,CAAC;IAC/C,IAAI,UAAU,IAAI,UAAU,CAAC,UAAU,CAAC;QAAE,OAAO,UAAU,CAAC;IAC5D,KAAK,MAAM,SAAS,IAAI,wBAAwB,EAAE,CAAC;QACjD,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,SAAS,CAAC;IAC9C,CAAC;IACD,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAO,GAAG,oBAAoB,EAAE;IAChE,MAAM,OAAO,GAAG,IAAI,GAAG,EAAU,CAAC;IAClC,IAAI,OAAO,KAAK,QAAQ;QAAE,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC;IACxD,KAAK,MAAM,SAAS,IAAI,wBAAwB,EAAE,CAAC;QACjD,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,CAAC;IAC7D,CAAC;IACD,OAAO,CAAC,GAAG,OAAO,CAAC,CAAC;AACtB,CAAC;AAED,MAAM,UAAU,mBAAmB,CACjC,IAAuB,EACvB,OAAO,GAAG,oBAAoB,EAAE;IAEhC,MAAM,WAAW,GAAG,iBAAiB,CAAC,OAAO,CAAC,CAAC;IAC/C,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAClD,OAAO;QACL,GAAG,IAAI;QACP,eAAe,EAAE,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,eAAe;QACtE,IAAI,EAAE,CAAC,GAAG,WAAW,EAAE,GAAG,YAAY,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC;KACxE,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,MAAM,cAAc,GAA2B;IACpD,KAAK,EAAE,6CAA6C;IACpD,MAAM,EAAE,gCAAgC;IACxC,IAAI,EAAE,4CAA4C;CACnD,CAAC;AAEF;;;;;GAKG;AACH,SAAS,gBAAgB,CAAC,MAAc;IACtC,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC1C,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC;IAC5C,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,UAAU,cAAc,CAAC,IAAY,EAAE,MAAe;IAC1D,MAAM,EAAE,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC;IACxC,MAAM,MAAM,GAAG,gBAAgB,CAAC,MAAM,IAAI,WAAW,CAAC,CAAC;IACvD,OAAO,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,OAAO,EAAE,GAAG,MAAM,GAAG,CAAC,CAAC;AAClE,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,MAAc,EACd,OAAyB,EAAE;IAE3B,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,CAAC,CAAC;IAClC,IAAI,OAA0B,CAAC;IAC/B,IAAI,mBAAmB,GAAG,IAAI,CAAC,eAAe,CAAC;IAC/C,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC;QACpD,IAAI,CAAC;YACH,MAAM,WAAW,GAAG,mBAAmB;gBACrC,CAAC,CAAC,EAAE,GAAG,IAAI,EAAE,eAAe,EAAE,mBAAmB,EAAE;gBACnD,CAAC,CAAC,IAAI,CAAC;YACT,OAAO,MAAM,CAAC,mBAAmB,CAAC,IAAI,CAAC,QAAQ,CAAC;gBAC9C,CAAC,CAAC,gBAAgB,CAAC,MAAM,EAAE,WAAW,CAAC;gBACvC,CAAC,CAAC,aAAa,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC;QAC1C,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,OAAO,GAAG,GAAY,CAAC;YACvB,uEAAuE;YACvE,0CAA0C;YAC1C,OAAO,CAAC,KAAK,CACX,wBAAwB,OAAO,GAAG,CAAC,IAAI,OAAO,GAAG,CAAC,YAAY,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAC9F,CAAC;YACF,IAAI,uBAAuB,CAAC,OAAO,CAAC,EAAE,CAAC;gBACrC,MAAM,QAAQ,GAAG,sBAAsB,CAAC,mBAAmB,IAAI,oBAAoB,EAAE,CAAC,CAAC;gBACvF,IAAI,QAAQ,EAAE,CAAC;oBACb,mBAAmB,GAAG,QAAQ,CAAC;oBAC/B,OAAO,CAAC,KAAK,CAAC,uDAAuD,QAAQ,EAAE,CAAC,CAAC;gBACnF,CAAC;YACH,CAAC;YACD,IAAI,OAAO,GAAG,OAAO,EAAE,CAAC;gBACtB,mEAAmE;gBACnE,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,IAAI,GAAG,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YAChE,CAAC;QACH,CAAC;IACH,CAAC;IACD,MAAM,OAAO,IAAI,IAAI,KAAK,CAAC,mBAAmB,CAAC,CAAC;AAClD,CAAC;AAUD;;;;;;;;;GASG;AACH,MAAM,UAAU,wBAAwB,CAAC,MAAyB,OAAO,CAAC,GAAG;IAC3E,MAAM,MAAM,GAAG,GAAG,CAAC,oBAAoB,IAAI,GAAG,CAAC,gBAAgB,IAAI,GAAG,CAAC,aAAa,CAAC;IACrF,MAAM,aAAa,GACjB,GAAG,CAAC,qBAAqB,IAAI,GAAG,CAAC,qBAAqB,IAAI,GAAG,CAAC,cAAc,CAAC;IAC/E,MAAM,eAAe,GAAG,GAAG,CAAC,wBAAwB,IAAI,GAAG,CAAC,iBAAiB,CAAC;IAC9E,MAAM,eAAe,GAAG,GAAG,CAAC,uBAAuB,CAAC;IAEpD,MAAM,GAAG,GAAsB,EAAE,CAAC;IAClC,IAAI,MAAM;QAAE,GAAG,CAAC,MAAM,GAAG,MAAM,CAAC;IAChC,IAAI,aAAa;QAAE,GAAG,CAAC,aAAa,GAAG,aAAa,CAAC;IACrD,IAAI,eAAe;QAAE,GAAG,CAAC,eAAe,GAAG,eAAe,CAAC;IAC3D,IAAI,eAAe;QAAE,GAAG,CAAC,eAAe,GAAG,eAAe,CAAC;IAC3D,OAAO,GAAG,CAAC;AACb,CAAC;AAED,KAAK,UAAU,gBAAgB,CAC7B,MAAc,EACd,OAAyB,EAAE;IAE3B,MAAM,KAAK,GACT,IAAI,CAAC,QAAQ,KAAK,OAAO,IAAI,IAAI,CAAC,QAAQ,KAAK,cAAc,CAAC,CAAC,CAAC,wBAAwB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IAClG,MAAM,MAAM,GAAG,MAAM,qBAAqB,CAAC;QACzC,QAAQ,EAAE,IAAI,CAAC,QAAQ;QACvB,KAAK,EAAE,IAAI,CAAC,KAAK,IAAI,SAAS;QAC9B,KAAK,EAAE,MAAM;QACb,eAAe,EAAE,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,wBAAwB,CAAC,IAAI,SAAS;QAC1E,eAAe,EAAE,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC;QAC1E,SAAS,EAAE,IAAI,CAAC,SAAS;QACzB,GAAG,KAAK;KACT,CAAC,CAAC;IACH,OAAO;QACL,IAAI,EAAE,MAAM,CAAC,IAAI;QACjB,EAAE,EAAE,IAAI;KACT,CAAC;AACJ,CAAC;AAaD,SAAS,oBAAoB,CAC3B,KAAyB;IAEzB,IAAI,KAAK,KAAK,SAAS,IAAI,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,MAAM,EAAE,CAAC;QACrF,OAAO,KAAK,CAAC;IACf,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED;;;;GAIG;AACH,SAAS,aAAa,CAAC,MAAc,EAAE,OAAyB,EAAE;IAChE,+EAA+E;IAC/E,gFAAgF;IAChF,8EAA8E;IAC9E,+EAA+E;IAC/E,mEAAmE;IACnE,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE,iBAAiB,EAAE,MAAM,EAAE,qBAAqB,EAAE,mBAAmB,EAAE,EAAE,CAAC,CAAC;IAC/F,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,IAAI,oBAAoB,EAAE,CAAC;IAC/D,MAAM,GAAG,GAAG,mBAAmB,CAAC,OAAO,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;IAEtD,+EAA+E;IAC/E,8EAA8E;IAC9E,sEAAsE;IACtE,yEAAyE;IACzE,wEAAwE;IACxE,oEAAoE;IACpE,kDAAkD;IAClD,MAAM,UAAU,GAAG,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,QAAQ,KAAK,SAAS,CAAC;IAC/D,IAAI,UAAU,EAAE,CAAC;QACf,8EAA8E;QAC9E,sEAAsE;QACtE,GAAG,CAAC,uBAAuB,GAAG,GAAG,CAAC;QAClC,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,GAAG,CAAC,UAAU,IAAI,WAAW,CAAC;QACzE,IAAI,IAAI,CAAC,KAAK;YAAE,GAAG,CAAC,eAAe,GAAG,cAAc,CAAC,IAAI,CAAC,KAAK,EAAE,GAAG,CAAC,UAAU,CAAC,CAAC;IACnF,CAAC;SAAM,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;QACtB,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;IACnC,CAAC;IACD,IAAI,IAAI,CAAC,eAAe;QAAE,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,eAAe,CAAC,CAAC;IACtE,IAAI,IAAI,CAAC,cAAc,EAAE,CAAC;QACxB,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,IAAI,CAAC,cAAc,CAAC,CAAC;IACtD,CAAC;IACD,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACtD,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,GAAG,IAAI,CAAC,YAAY,CAAC,CAAC;IACpD,CAAC;IACD,IAAI,IAAI,CAAC,QAAQ,IAAI,CAAC,UAAU,EAAE,CAAC;QACjC,wEAAwE;QACxE,wEAAwE;QACxE,wEAAwE;QACxE,6DAA6D;QAC7D,MAAM,QAAQ,GAA2B,EAAE,WAAW,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC;QACxE,IAAI,CAAC,IAAI,CAAC,YAAY,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC;IACpD,CAAC;IACD,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;QACvB,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,GAAG,gBAAgB,CAAC,CAAC;IACtD,CAAC;IAED,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,MAAM,CAAC;IAE3C,OAAO,IAAI,OAAO,CAAkB,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtD,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,EAAE,IAAI,EAAE;YACjC,KAAK,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;YAC/B,GAAG;YACH,GAAG,EAAE,IAAI,CAAC,GAAG;SACd,CAAC,CAAC;QAEH,IAAI,MAAM,GAAG,EAAE,CAAC;QAChB,IAAI,MAAM,GAAG,EAAE,CAAC;QAChB,IAAI,OAAO,GAAG,KAAK,CAAC;QAEpB,MAAM,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE;YAC5B,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,KAAK,CAAC,8BAA8B,SAAS,IAAI,CAAC,CAAC,CAAC;QACjE,CAAC,EAAE,SAAS,CAAC,CAAC;QAEd,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;QACzD,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;QAEzD,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE;YACxB,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,YAAY,CAAC,KAAK,CAAC,CAAC;YACpB,MAAM,CAAC,IAAI,KAAK,CAAC,+BAA+B,GAAG,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;QAClE,CAAC,CAAC,CAAC;QAEH,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,IAAI,EAAE,EAAE;YACzB,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,YAAY,CAAC,KAAK,CAAC,CAAC;YACpB,IAAI,IAAI,KAAK,CAAC,EAAE,CAAC;gBACf,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC;gBACpC,mEAAmE;gBACnE,oEAAoE;gBACpE,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,EAAE,CAAC;oBACnB,MAAM,CAAC,IAAI,KAAK,CAAC,qBAAqB,IAAI,oBAAoB,MAAM,EAAE,CAAC,CAAC,CAAC;oBACzE,OAAO;gBACT,CAAC;gBACD,MAAM,CAAC,IAAI,KAAK,CAAC,qBAAqB,IAAI,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC;gBAC1D,OAAO;YACT,CAAC;YACD,IAAI,CAAC;gBACH,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;gBAChC,yEAAyE;gBACzE,6DAA6D;gBAC7D,IAAI,IAAI,CAAC,gBAAgB,IAAI,IAAI,CAAC,QAAQ,KAAK,IAAI,EAAE,CAAC;oBACpD,MAAM,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;oBACnD,MAAM,OAAO,GAAG,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC;oBAC3C,OAAO,CAAC,KAAK,CACX,oCAAoC,IAAI,CAAC,QAAQ,IAAI,SAAS,KAAK;wBACjE,UAAU,MAAM,YAAY,OAAO,GAAG;wBACtC,UAAU,MAAM,CAAC,IAAI,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CACtD,CAAC;oBACF,wEAAwE;oBACxE,IACE,MAAM,KAAK,KAAK;wBAChB,MAAM,KAAK,KAAK;wBAChB,2CAA2C,CAAC,IAAI,CAAC,OAAO,CAAC,EACzD,CAAC;wBACD,MAAM,CAAC,IAAI,KAAK,CAAC,kBAAkB,MAAM,IAAI,OAAO,yBAAyB,CAAC,CAAC,CAAC;wBAChF,OAAO;oBACT,CAAC;gBACH,CAAC;gBACD,OAAO,CAAC;oBACN,IAAI,EAAE,OAAO,IAAI,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE;oBACxD,SAAS,EAAE,IAAI,CAAC,UAAU;oBAC1B,EAAE,EAAE,IAAI,CAAC,QAAQ,KAAK,IAAI,IAAI,CAAC,IAAI,CAAC,gBAAgB;oBACpD,OAAO,EAAE,OAAO,IAAI,CAAC,cAAc,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,SAAS;iBACnF,CAAC,CAAC;YACL,CAAC;YAAC,OAAO,GAAG,EAAE,CAAC;gBACb,MAAM,CACJ,IAAI,KAAK,CACP,2CAA4C,GAAa,CAAC,OAAO,IAAI;oBACnE,oBAAoB,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAC7C,CACF,CAAC;YACJ,CAAC;QACH,CAAC,CAAC,CAAC;QAEH,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;QAC1B,KAAK,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC;IACpB,CAAC,CAAC,CAAC;AACL,CAAC;AAED,SAAS,uBAAuB,CAAC,KAAY;IAC3C,OAAO,2BAA2B,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;AACzD,CAAC;AAED,SAAS,sBAAsB,CAAC,OAAe;IAC7C,KAAK,MAAM,SAAS,IAAI,wBAAwB,EAAE,CAAC;QACjD,IAAI,SAAS,KAAK,OAAO,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,SAAS,CAAC;IACvE,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,yEAAyE;AACzE,MAAM,CAAC,KAAK,UAAU,eAAe;IACnC,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE;QAC7B,MAAM,KAAK,GAAG,KAAK,CAAC,oBAAoB,EAAE,EAAE,CAAC,WAAW,CAAC,EAAE,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;QAChF,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,GAAG,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC;QACxC,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC;IACnD,CAAC,CAAC,CAAC;AACL,CAAC"}
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
import type { AgentDelegate, AgentDelegateOptions, AgentDelegateResult } from '../../runtime/types.js';
|
|
2
|
+
import type { Task } from '../../types/index.js';
|
|
3
|
+
import type { LiveBenchmark } from './live-benchmark.js';
|
|
4
|
+
/**
|
|
5
|
+
* Single-turn agent delegate backed by the headless Claude CLI.
|
|
6
|
+
*
|
|
7
|
+
* Atlas injects its learned knowledge into `options.systemContext`; this
|
|
8
|
+
* delegate prepends it to the task prompt, runs `claude -p`, then **verifies the
|
|
9
|
+
* answer against ground truth** via the benchmark. The verified correctness
|
|
10
|
+
* becomes the returned `success`, so it flows into `trajectory.outcome.success`
|
|
11
|
+
* — meaning the learning pipeline learns from correctly-labeled outcomes, not
|
|
12
|
+
* the agent's self-report.
|
|
13
|
+
*
|
|
14
|
+
* The runner sets `currentTask` immediately before each `atlas.solve`, which is
|
|
15
|
+
* safe because the eval runs tasks strictly sequentially.
|
|
16
|
+
*/
|
|
17
|
+
export declare class ClaudeBenchmarkDelegate implements AgentDelegate {
|
|
18
|
+
private readonly benchmark;
|
|
19
|
+
private readonly opts;
|
|
20
|
+
private currentTask;
|
|
21
|
+
private lastCostUsd;
|
|
22
|
+
private totalCostUsd;
|
|
23
|
+
constructor(benchmark: LiveBenchmark, opts?: {
|
|
24
|
+
model?: string;
|
|
25
|
+
timeoutMs?: number;
|
|
26
|
+
provider?: string;
|
|
27
|
+
awsRegion?: string;
|
|
28
|
+
/** Per-call retry override (failFast → 0 = single attempt). */
|
|
29
|
+
cliRetries?: number;
|
|
30
|
+
});
|
|
31
|
+
/** Set the task being solved next (its gold answer is read during verify). */
|
|
32
|
+
setCurrentTask(task: Task): void;
|
|
33
|
+
/** Cumulative USD cost across all calls (best-effort, from the CLI). */
|
|
34
|
+
getTotalCostUsd(): number;
|
|
35
|
+
execute(prompt: string, options: AgentDelegateOptions): Promise<AgentDelegateResult>;
|
|
36
|
+
}
|
|
37
|
+
export declare function createClaudeBenchmarkDelegate(benchmark: LiveBenchmark, opts?: {
|
|
38
|
+
model?: string;
|
|
39
|
+
timeoutMs?: number;
|
|
40
|
+
}): ClaudeBenchmarkDelegate;
|
|
41
|
+
//# sourceMappingURL=claude-delegate.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-delegate.d.ts","sourceRoot":"","sources":["../../../src/eval/live/claude-delegate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,aAAa,EACb,oBAAoB,EACpB,mBAAmB,EACpB,MAAM,wBAAwB,CAAC;AAChC,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,sBAAsB,CAAC;AACjD,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAGzD;;;;;;;;;;;;GAYG;AACH,qBAAa,uBAAwB,YAAW,aAAa;IAMzD,OAAO,CAAC,QAAQ,CAAC,SAAS;IAC1B,OAAO,CAAC,QAAQ,CAAC,IAAI;IANvB,OAAO,CAAC,WAAW,CAAqB;IACxC,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,YAAY,CAAK;gBAGN,SAAS,EAAE,aAAa,EACxB,IAAI,GAAE;QACrB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,+DAA+D;QAC/D,UAAU,CAAC,EAAE,MAAM,CAAC;KAChB;IAGR,8EAA8E;IAC9E,cAAc,CAAC,IAAI,EAAE,IAAI,GAAG,IAAI;IAIhC,wEAAwE;IACxE,eAAe,IAAI,MAAM;IAInB,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,oBAAoB,GAAG,OAAO,CAAC,mBAAmB,CAAC;CA0B3F;AAED,wBAAgB,6BAA6B,CAC3C,SAAS,EAAE,aAAa,EACxB,IAAI,CAAC,EAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,SAAS,CAAC,EAAE,MAAM,CAAA;CAAE,GAC5C,uBAAuB,CAEzB"}
|
|
@@ -0,0 +1,59 @@
|
|
|
1
|
+
import { runClaude } from './claude-cli.js';
|
|
2
|
+
/**
|
|
3
|
+
* Single-turn agent delegate backed by the headless Claude CLI.
|
|
4
|
+
*
|
|
5
|
+
* Atlas injects its learned knowledge into `options.systemContext`; this
|
|
6
|
+
* delegate prepends it to the task prompt, runs `claude -p`, then **verifies the
|
|
7
|
+
* answer against ground truth** via the benchmark. The verified correctness
|
|
8
|
+
* becomes the returned `success`, so it flows into `trajectory.outcome.success`
|
|
9
|
+
* — meaning the learning pipeline learns from correctly-labeled outcomes, not
|
|
10
|
+
* the agent's self-report.
|
|
11
|
+
*
|
|
12
|
+
* The runner sets `currentTask` immediately before each `atlas.solve`, which is
|
|
13
|
+
* safe because the eval runs tasks strictly sequentially.
|
|
14
|
+
*/
|
|
15
|
+
export class ClaudeBenchmarkDelegate {
|
|
16
|
+
benchmark;
|
|
17
|
+
opts;
|
|
18
|
+
currentTask = null;
|
|
19
|
+
lastCostUsd = 0;
|
|
20
|
+
totalCostUsd = 0;
|
|
21
|
+
constructor(benchmark, opts = {}) {
|
|
22
|
+
this.benchmark = benchmark;
|
|
23
|
+
this.opts = opts;
|
|
24
|
+
}
|
|
25
|
+
/** Set the task being solved next (its gold answer is read during verify). */
|
|
26
|
+
setCurrentTask(task) {
|
|
27
|
+
this.currentTask = task;
|
|
28
|
+
}
|
|
29
|
+
/** Cumulative USD cost across all calls (best-effort, from the CLI). */
|
|
30
|
+
getTotalCostUsd() {
|
|
31
|
+
return this.totalCostUsd;
|
|
32
|
+
}
|
|
33
|
+
async execute(prompt, options) {
|
|
34
|
+
// systemContext carries Atlas's injected playbooks/experiences — keep it
|
|
35
|
+
// first so accumulated learning shapes the agent's approach.
|
|
36
|
+
const fullPrompt = [options.systemContext, prompt].filter(Boolean).join('\n\n');
|
|
37
|
+
const res = await runClaude(fullPrompt, {
|
|
38
|
+
model: this.opts.model,
|
|
39
|
+
timeoutMs: this.opts.timeoutMs ?? options.timeoutMs,
|
|
40
|
+
provider: this.opts.provider,
|
|
41
|
+
awsRegion: this.opts.awsRegion,
|
|
42
|
+
retries: this.opts.cliRetries,
|
|
43
|
+
});
|
|
44
|
+
this.lastCostUsd = res.costUsd ?? 0;
|
|
45
|
+
this.totalCostUsd += this.lastCostUsd;
|
|
46
|
+
const verdict = this.currentTask
|
|
47
|
+
? this.benchmark.verify(this.currentTask, res.text)
|
|
48
|
+
: { correct: false, extracted: undefined };
|
|
49
|
+
return {
|
|
50
|
+
success: verdict.correct,
|
|
51
|
+
output: res.text,
|
|
52
|
+
structured: { extracted: verdict.extracted, costUsd: this.lastCostUsd },
|
|
53
|
+
};
|
|
54
|
+
}
|
|
55
|
+
}
|
|
56
|
+
export function createClaudeBenchmarkDelegate(benchmark, opts) {
|
|
57
|
+
return new ClaudeBenchmarkDelegate(benchmark, opts);
|
|
58
|
+
}
|
|
59
|
+
//# sourceMappingURL=claude-delegate.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-delegate.js","sourceRoot":"","sources":["../../../src/eval/live/claude-delegate.ts"],"names":[],"mappings":"AAOA,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAE5C;;;;;;;;;;;;GAYG;AACH,MAAM,OAAO,uBAAuB;IAMf;IACA;IANX,WAAW,GAAgB,IAAI,CAAC;IAChC,WAAW,GAAG,CAAC,CAAC;IAChB,YAAY,GAAG,CAAC,CAAC;IAEzB,YACmB,SAAwB,EACxB,OAOb,EAAE;QARW,cAAS,GAAT,SAAS,CAAe;QACxB,SAAI,GAAJ,IAAI,CAOf;IACL,CAAC;IAEJ,8EAA8E;IAC9E,cAAc,CAAC,IAAU;QACvB,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC;IAC1B,CAAC;IAED,wEAAwE;IACxE,eAAe;QACb,OAAO,IAAI,CAAC,YAAY,CAAC;IAC3B,CAAC;IAED,KAAK,CAAC,OAAO,CAAC,MAAc,EAAE,OAA6B;QACzD,yEAAyE;QACzE,6DAA6D;QAC7D,MAAM,UAAU,GAAG,CAAC,OAAO,CAAC,aAAa,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAEhF,MAAM,GAAG,GAAG,MAAM,SAAS,CAAC,UAAU,EAAE;YACtC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,KAAK;YACtB,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,SAAS;YACnD,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;YAC5B,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS;YAC9B,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,UAAU;SAC9B,CAAC,CAAC;QAEH,IAAI,CAAC,WAAW,GAAG,GAAG,CAAC,OAAO,IAAI,CAAC,CAAC;QACpC,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,WAAW,CAAC;QAEtC,MAAM,OAAO,GAAG,IAAI,CAAC,WAAW;YAC9B,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,EAAE,GAAG,CAAC,IAAI,CAAC;YACnD,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,SAAS,EAAE,CAAC;QAE7C,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,MAAM,EAAE,GAAG,CAAC,IAAI;YAChB,UAAU,EAAE,EAAE,SAAS,EAAE,OAAO,CAAC,SAAS,EAAE,OAAO,EAAE,IAAI,CAAC,WAAW,EAAE;SACxE,CAAC;IACJ,CAAC;CACF;AAED,MAAM,UAAU,6BAA6B,CAC3C,SAAwB,EACxB,IAA6C;IAE7C,OAAO,IAAI,uBAAuB,CAAC,SAAS,EAAE,IAAI,CAAC,CAAC;AACtD,CAAC"}
|
|
@@ -0,0 +1,24 @@
|
|
|
1
|
+
import type { AgentBackend, AgentSession, AgentSpawnConfig } from '../../runtime/types.js';
|
|
2
|
+
/**
|
|
3
|
+
* General Claude CLI backend for offline workspace-analysis agents.
|
|
4
|
+
*
|
|
5
|
+
* Unlike benchmark delegates, this backend does not verify task success against
|
|
6
|
+
* a benchmark. It runs Claude in the supplied workspace cwd and lets templates
|
|
7
|
+
* validate the files they asked the agent to write.
|
|
8
|
+
*/
|
|
9
|
+
export declare class ClaudeWorkspaceBackend implements AgentBackend {
|
|
10
|
+
private readonly opts;
|
|
11
|
+
readonly name = "claude-workspace";
|
|
12
|
+
readonly supportedTypes: string[];
|
|
13
|
+
private readonly sessions;
|
|
14
|
+
constructor(opts?: {
|
|
15
|
+
model?: string;
|
|
16
|
+
timeoutMs?: number;
|
|
17
|
+
provider?: string;
|
|
18
|
+
awsRegion?: string;
|
|
19
|
+
});
|
|
20
|
+
isAvailable(): Promise<boolean>;
|
|
21
|
+
spawn(config: AgentSpawnConfig): Promise<AgentSession>;
|
|
22
|
+
getSession(sessionId: string): Promise<AgentSession | undefined>;
|
|
23
|
+
}
|
|
24
|
+
//# sourceMappingURL=claude-workspace-backend.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-workspace-backend.d.ts","sourceRoot":"","sources":["../../../src/eval/live/claude-workspace-backend.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACV,YAAY,EAEZ,YAAY,EACZ,gBAAgB,EACjB,MAAM,wBAAwB,CAAC;AAGhC;;;;;;GAMG;AACH,qBAAa,sBAAuB,YAAW,YAAY;IAOvD,OAAO,CAAC,QAAQ,CAAC,IAAI;IANvB,QAAQ,CAAC,IAAI,sBAAsB;IACnC,QAAQ,CAAC,cAAc,WAAmB;IAE1C,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAmC;gBAGzC,IAAI,GAAE;QACrB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,SAAS,CAAC,EAAE,MAAM,CAAC;KACf;IAGF,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC;IAI/B,KAAK,CAAC,MAAM,EAAE,gBAAgB,GAAG,OAAO,CAAC,YAAY,CAAC;IAuDtD,UAAU,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,GAAG,SAAS,CAAC;CAGvE"}
|
|
@@ -0,0 +1,76 @@
|
|
|
1
|
+
import { randomUUID } from 'node:crypto';
|
|
2
|
+
import { runClaude } from './claude-cli.js';
|
|
3
|
+
/**
|
|
4
|
+
* General Claude CLI backend for offline workspace-analysis agents.
|
|
5
|
+
*
|
|
6
|
+
* Unlike benchmark delegates, this backend does not verify task success against
|
|
7
|
+
* a benchmark. It runs Claude in the supplied workspace cwd and lets templates
|
|
8
|
+
* validate the files they asked the agent to write.
|
|
9
|
+
*/
|
|
10
|
+
export class ClaudeWorkspaceBackend {
|
|
11
|
+
opts;
|
|
12
|
+
name = 'claude-workspace';
|
|
13
|
+
supportedTypes = ['claude-code'];
|
|
14
|
+
sessions = new Map();
|
|
15
|
+
constructor(opts = {}) {
|
|
16
|
+
this.opts = opts;
|
|
17
|
+
}
|
|
18
|
+
async isAvailable() {
|
|
19
|
+
return true;
|
|
20
|
+
}
|
|
21
|
+
async spawn(config) {
|
|
22
|
+
const startTime = new Date();
|
|
23
|
+
const session = {
|
|
24
|
+
id: randomUUID(),
|
|
25
|
+
agentType: config.agentType,
|
|
26
|
+
task: config.task,
|
|
27
|
+
state: 'running',
|
|
28
|
+
messages: [],
|
|
29
|
+
toolCalls: [],
|
|
30
|
+
startTime,
|
|
31
|
+
metadata: { backend: this.name },
|
|
32
|
+
};
|
|
33
|
+
this.sessions.set(session.id, session);
|
|
34
|
+
try {
|
|
35
|
+
const prompt = [
|
|
36
|
+
config.systemPromptAdditions,
|
|
37
|
+
'You are running as a workspace automation agent. Required output files must be created on disk in this working directory. Use file tools to write the requested output files; do not only describe the result in your final answer. The caller validates files on disk and ignores prose unless the files exist.',
|
|
38
|
+
config.task.description,
|
|
39
|
+
]
|
|
40
|
+
.filter(Boolean)
|
|
41
|
+
.join('\n\n');
|
|
42
|
+
const result = await runClaude(prompt, {
|
|
43
|
+
model: this.opts.model,
|
|
44
|
+
timeoutMs: config.timeout ?? this.opts.timeoutMs,
|
|
45
|
+
provider: this.opts.provider,
|
|
46
|
+
awsRegion: this.opts.awsRegion,
|
|
47
|
+
cwd: config.cwd,
|
|
48
|
+
permissionMode: 'acceptEdits',
|
|
49
|
+
allowedTools: ['Read', 'Write', 'Edit', 'Glob', 'Grep'],
|
|
50
|
+
});
|
|
51
|
+
const message = {
|
|
52
|
+
role: 'assistant',
|
|
53
|
+
content: result.text,
|
|
54
|
+
timestamp: new Date(),
|
|
55
|
+
};
|
|
56
|
+
session.messages.push(message);
|
|
57
|
+
session.result = result.text;
|
|
58
|
+
session.state = result.ok ? 'completed' : 'failed';
|
|
59
|
+
session.endTime = new Date();
|
|
60
|
+
if (!result.ok) {
|
|
61
|
+
session.error = 'Claude CLI returned an error result';
|
|
62
|
+
}
|
|
63
|
+
}
|
|
64
|
+
catch (error) {
|
|
65
|
+
session.state = 'failed';
|
|
66
|
+
session.error = error instanceof Error ? error.message : String(error);
|
|
67
|
+
session.endTime = new Date();
|
|
68
|
+
}
|
|
69
|
+
this.sessions.set(session.id, session);
|
|
70
|
+
return session;
|
|
71
|
+
}
|
|
72
|
+
async getSession(sessionId) {
|
|
73
|
+
return this.sessions.get(sessionId);
|
|
74
|
+
}
|
|
75
|
+
}
|
|
76
|
+
//# sourceMappingURL=claude-workspace-backend.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-workspace-backend.js","sourceRoot":"","sources":["../../../src/eval/live/claude-workspace-backend.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAOzC,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAE5C;;;;;;GAMG;AACH,MAAM,OAAO,sBAAsB;IAOd;IANV,IAAI,GAAG,kBAAkB,CAAC;IAC1B,cAAc,GAAG,CAAC,aAAa,CAAC,CAAC;IAEzB,QAAQ,GAAG,IAAI,GAAG,EAAwB,CAAC;IAE5D,YACmB,OAKb,EAAE;QALW,SAAI,GAAJ,IAAI,CAKf;IACL,CAAC;IAEJ,KAAK,CAAC,WAAW;QACf,OAAO,IAAI,CAAC;IACd,CAAC;IAED,KAAK,CAAC,KAAK,CAAC,MAAwB;QAClC,MAAM,SAAS,GAAG,IAAI,IAAI,EAAE,CAAC;QAC7B,MAAM,OAAO,GAAiB;YAC5B,EAAE,EAAE,UAAU,EAAE;YAChB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,KAAK,EAAE,SAAS;YAChB,QAAQ,EAAE,EAAE;YACZ,SAAS,EAAE,EAAE;YACb,SAAS;YACT,QAAQ,EAAE,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE;SACjC,CAAC;QACF,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC;QAEvC,IAAI,CAAC;YACH,MAAM,MAAM,GAAG;gBACb,MAAM,CAAC,qBAAqB;gBAC5B,kTAAkT;gBAClT,MAAM,CAAC,IAAI,CAAC,WAAW;aACxB;iBACE,MAAM,CAAC,OAAO,CAAC;iBACf,IAAI,CAAC,MAAM,CAAC,CAAC;YAEhB,MAAM,MAAM,GAAG,MAAM,SAAS,CAAC,MAAM,EAAE;gBACrC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,KAAK;gBACtB,SAAS,EAAE,MAAM,CAAC,OAAO,IAAI,IAAI,CAAC,IAAI,CAAC,SAAS;gBAChD,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;gBAC5B,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS;gBAC9B,GAAG,EAAE,MAAM,CAAC,GAAG;gBACf,cAAc,EAAE,aAAa;gBAC7B,YAAY,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;aACxD,CAAC,CAAC;YAEH,MAAM,OAAO,GAAiB;gBAC5B,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,MAAM,CAAC,IAAI;gBACpB,SAAS,EAAE,IAAI,IAAI,EAAE;aACtB,CAAC;YACF,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;YAC/B,OAAO,CAAC,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC;YAC7B,OAAO,CAAC,KAAK,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC;YACnD,OAAO,CAAC,OAAO,GAAG,IAAI,IAAI,EAAE,CAAC;YAC7B,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,CAAC;gBACf,OAAO,CAAC,KAAK,GAAG,qCAAqC,CAAC;YACxD,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,CAAC,KAAK,GAAG,QAAQ,CAAC;YACzB,OAAO,CAAC,KAAK,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;YACvE,OAAO,CAAC,OAAO,GAAG,IAAI,IAAI,EAAE,CAAC;QAC/B,CAAC;QAED,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC;QACvC,OAAO,OAAO,CAAC;IACjB,CAAC;IAED,KAAK,CAAC,UAAU,CAAC,SAAiB;QAChC,OAAO,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;IACtC,CAAC;CACF"}
|
|
@@ -0,0 +1,64 @@
|
|
|
1
|
+
import { type Task } from '../../types/index.js';
|
|
2
|
+
import type { EvalTask } from '../types.js';
|
|
3
|
+
import type { ExtendedLiveBenchmark, BenchmarkVerdict } from './live-benchmark.js';
|
|
4
|
+
import type { BenchmarkDelegate } from './live-runner.js';
|
|
5
|
+
import { EnvBridge } from './env-bridge.js';
|
|
6
|
+
export declare class ClBenchBenchmark implements ExtendedLiveBenchmark {
|
|
7
|
+
readonly name = "clbench";
|
|
8
|
+
readonly interactive = true;
|
|
9
|
+
readonly preserveOrder = true;
|
|
10
|
+
readonly playbookMinTrajectories = 3;
|
|
11
|
+
private bridge;
|
|
12
|
+
private readonly maxSteps;
|
|
13
|
+
private readonly reinject;
|
|
14
|
+
private readonly historyWindow;
|
|
15
|
+
constructor(opts?: {
|
|
16
|
+
maxSteps?: number;
|
|
17
|
+
reinject?: boolean;
|
|
18
|
+
historyWindow?: number;
|
|
19
|
+
});
|
|
20
|
+
setup(): Promise<void>;
|
|
21
|
+
teardown(): Promise<void>;
|
|
22
|
+
private getBridge;
|
|
23
|
+
freshBridge(): Promise<EnvBridge>;
|
|
24
|
+
/**
|
|
25
|
+
* Load conversations as tasks. Applies a seeded Fisher-Yates shuffle so
|
|
26
|
+
* different seeds produce different orderings (reproducible per seed).
|
|
27
|
+
*/
|
|
28
|
+
loadTasks(opts: {
|
|
29
|
+
count: number;
|
|
30
|
+
seed: number;
|
|
31
|
+
}): Promise<EvalTask[]>;
|
|
32
|
+
verify(_task: Task, _output: string): BenchmarkVerdict;
|
|
33
|
+
getBridgeInstance(): EnvBridge | null;
|
|
34
|
+
delegateFactory: (_benchmark: ExtendedLiveBenchmark, opts?: {
|
|
35
|
+
model?: string;
|
|
36
|
+
timeoutMs?: number;
|
|
37
|
+
provider?: string;
|
|
38
|
+
awsRegion?: string;
|
|
39
|
+
cliRetries?: number;
|
|
40
|
+
}) => BenchmarkDelegate;
|
|
41
|
+
}
|
|
42
|
+
export declare function createClBenchBenchmark(opts?: {
|
|
43
|
+
maxSteps?: number;
|
|
44
|
+
reinject?: boolean;
|
|
45
|
+
historyWindow?: number;
|
|
46
|
+
}): ClBenchBenchmark;
|
|
47
|
+
/**
|
|
48
|
+
* Structured per-step prompt for CL-Bench CRM conversations.
|
|
49
|
+
* Surfaces tool catalog, conversation context, and injected knowledge.
|
|
50
|
+
*/
|
|
51
|
+
export declare function buildClBenchPrompt(p: {
|
|
52
|
+
knowledge?: string;
|
|
53
|
+
instruction: string;
|
|
54
|
+
category: string;
|
|
55
|
+
complexity: string;
|
|
56
|
+
history: Array<{
|
|
57
|
+
action: string;
|
|
58
|
+
obs: string;
|
|
59
|
+
}>;
|
|
60
|
+
obs: string;
|
|
61
|
+
availableTools?: string[];
|
|
62
|
+
context?: Record<string, unknown>;
|
|
63
|
+
}): string;
|
|
64
|
+
//# sourceMappingURL=clbench.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"clbench.d.ts","sourceRoot":"","sources":["../../../src/eval/live/clbench.ts"],"names":[],"mappings":"AAEA,OAAO,EAML,KAAK,IAAI,EACV,MAAM,sBAAsB,CAAC;AAE9B,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAC5C,OAAO,KAAK,EAAE,qBAAqB,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACnF,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAC;AAC1D,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAuE5C,qBAAa,gBAAiB,YAAW,qBAAqB;IAC5D,QAAQ,CAAC,IAAI,aAAa;IAC1B,QAAQ,CAAC,WAAW,QAAQ;IAC5B,QAAQ,CAAC,aAAa,QAAQ;IAC9B,QAAQ,CAAC,uBAAuB,KAAK;IAErC,OAAO,CAAC,MAAM,CAA0B;IACxC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAClC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAU;IACnC,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;gBAGrC,IAAI,GAAE;QACJ,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,QAAQ,CAAC,EAAE,OAAO,CAAC;QACnB,aAAa,CAAC,EAAE,MAAM,CAAC;KACnB;IAOF,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;IAItB,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC;YAKjB,SAAS;IAiBjB,WAAW,IAAI,OAAO,CAAC,SAAS,CAAC;IAMvC;;;OAGG;IACG,SAAS,CAAC,IAAI,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,GAAG,OAAO,CAAC,QAAQ,EAAE,CAAC;IA4C3E,MAAM,CAAC,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,GAAG,gBAAgB;IAItD,iBAAiB,IAAI,SAAS,GAAG,IAAI;IAIrC,eAAe,GACb,YAAY,qBAAqB,EACjC,OAAO;QACL,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,UAAU,CAAC,EAAE,MAAM,CAAC;KACrB,KACA,iBAAiB,CAWlB;CACH;AAED,wBAAgB,sBAAsB,CAAC,IAAI,CAAC,EAAE;IAC5C,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,aAAa,CAAC,EAAE,MAAM,CAAC;CACxB,GAAG,gBAAgB,CAEnB;AAwID;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,CAAC,EAAE;IACpC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,KAAK,CAAC;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IAChD,GAAG,EAAE,MAAM,CAAC;IACZ,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACnC,GAAG,MAAM,CAkCT"}
|