@nac3/forge-cli 0.2.0-alpha.9 → 0.2.1-alpha.60
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/README.md +16 -1
- package/dist/a11y/helpers.d.ts +88 -0
- package/dist/a11y/helpers.d.ts.map +1 -0
- package/dist/a11y/helpers.js +109 -0
- package/dist/a11y/helpers.js.map +1 -0
- package/dist/agents/bridge.d.ts +69 -0
- package/dist/agents/bridge.d.ts.map +1 -0
- package/dist/agents/bridge.js +295 -0
- package/dist/agents/bridge.js.map +1 -0
- package/dist/agents/cli_real.d.ts +38 -0
- package/dist/agents/cli_real.d.ts.map +1 -0
- package/dist/agents/cli_real.js +46 -0
- package/dist/agents/cli_real.js.map +1 -0
- package/dist/agents/config_resolver.d.ts +16 -0
- package/dist/agents/config_resolver.d.ts.map +1 -0
- package/dist/agents/config_resolver.js +133 -0
- package/dist/agents/config_resolver.js.map +1 -0
- package/dist/agents/config_schema.d.ts +42 -0
- package/dist/agents/config_schema.d.ts.map +1 -0
- package/dist/agents/config_schema.js +220 -0
- package/dist/agents/config_schema.js.map +1 -0
- package/dist/agents/defaults.d.ts +21 -0
- package/dist/agents/defaults.d.ts.map +1 -0
- package/dist/agents/defaults.js +56 -0
- package/dist/agents/defaults.js.map +1 -0
- package/dist/agents/delegate.d.ts +36 -0
- package/dist/agents/delegate.d.ts.map +1 -0
- package/dist/agents/delegate.js +68 -0
- package/dist/agents/delegate.js.map +1 -0
- package/dist/agents/iterative_step.d.ts +95 -0
- package/dist/agents/iterative_step.d.ts.map +1 -0
- package/dist/agents/iterative_step.js +249 -0
- package/dist/agents/iterative_step.js.map +1 -0
- package/dist/agents/llm_step.d.ts +50 -0
- package/dist/agents/llm_step.d.ts.map +1 -0
- package/dist/agents/llm_step.js +28 -0
- package/dist/agents/llm_step.js.map +1 -0
- package/dist/agents/real_step_factory.d.ts +95 -0
- package/dist/agents/real_step_factory.d.ts.map +1 -0
- package/dist/agents/real_step_factory.js +51 -0
- package/dist/agents/real_step_factory.js.map +1 -0
- package/dist/agents/runner.d.ts +108 -0
- package/dist/agents/runner.d.ts.map +1 -0
- package/dist/agents/runner.js +174 -0
- package/dist/agents/runner.js.map +1 -0
- package/dist/agents/runtime.d.ts +115 -0
- package/dist/agents/runtime.d.ts.map +1 -0
- package/dist/agents/runtime.js +429 -0
- package/dist/agents/runtime.js.map +1 -0
- package/dist/agents/store.d.ts +33 -0
- package/dist/agents/store.d.ts.map +1 -0
- package/dist/agents/store.js +210 -0
- package/dist/agents/store.js.map +1 -0
- package/dist/agents/system_prompt.d.ts +36 -0
- package/dist/agents/system_prompt.d.ts.map +1 -0
- package/dist/agents/system_prompt.js +101 -0
- package/dist/agents/system_prompt.js.map +1 -0
- package/dist/agents/types.d.ts +123 -0
- package/dist/agents/types.d.ts.map +1 -0
- package/dist/agents/types.js +39 -0
- package/dist/agents/types.js.map +1 -0
- package/dist/bin/yf.d.ts.map +1 -1
- package/dist/bin/yf.js +101 -2
- package/dist/bin/yf.js.map +1 -1
- package/dist/chat/claude.d.ts +89 -1
- package/dist/chat/claude.d.ts.map +1 -1
- package/dist/chat/claude.js +170 -30
- package/dist/chat/claude.js.map +1 -1
- package/dist/chat/claude_cli.d.ts +145 -0
- package/dist/chat/claude_cli.d.ts.map +1 -0
- package/dist/chat/claude_cli.js +598 -0
- package/dist/chat/claude_cli.js.map +1 -0
- package/dist/chat/completion_hallucination_detector.d.ts +40 -0
- package/dist/chat/completion_hallucination_detector.d.ts.map +1 -0
- package/dist/chat/completion_hallucination_detector.js +96 -0
- package/dist/chat/completion_hallucination_detector.js.map +1 -0
- package/dist/chat/doctrine_digest.d.ts +31 -0
- package/dist/chat/doctrine_digest.d.ts.map +1 -0
- package/dist/chat/doctrine_digest.js +117 -0
- package/dist/chat/doctrine_digest.js.map +1 -0
- package/dist/chat/graph_auto_ingest.d.ts +46 -0
- package/dist/chat/graph_auto_ingest.d.ts.map +1 -0
- package/dist/chat/graph_auto_ingest.js +177 -0
- package/dist/chat/graph_auto_ingest.js.map +1 -0
- package/dist/chat/graph_component_watcher.d.ts +34 -0
- package/dist/chat/graph_component_watcher.d.ts.map +1 -0
- package/dist/chat/graph_component_watcher.js +225 -0
- package/dist/chat/graph_component_watcher.js.map +1 -0
- package/dist/chat/graph_snapshot.d.ts +38 -0
- package/dist/chat/graph_snapshot.d.ts.map +1 -0
- package/dist/chat/graph_snapshot.js +159 -0
- package/dist/chat/graph_snapshot.js.map +1 -0
- package/dist/chat/ingest_session.d.ts +24 -4
- package/dist/chat/ingest_session.d.ts.map +1 -1
- package/dist/chat/ingest_session.js +143 -4
- package/dist/chat/ingest_session.js.map +1 -1
- package/dist/chat/panel.d.ts.map +1 -1
- package/dist/chat/panel.js +9640 -1455
- package/dist/chat/panel.js.map +1 -1
- package/dist/chat/panel_queue.d.ts +43 -0
- package/dist/chat/panel_queue.d.ts.map +1 -0
- package/dist/chat/panel_queue.js +46 -0
- package/dist/chat/panel_queue.js.map +1 -0
- package/dist/chat/project_context_inject.d.ts +10 -0
- package/dist/chat/project_context_inject.d.ts.map +1 -0
- package/dist/chat/project_context_inject.js +119 -0
- package/dist/chat/project_context_inject.js.map +1 -0
- package/dist/chat/project_slug.d.ts +17 -0
- package/dist/chat/project_slug.d.ts.map +1 -0
- package/dist/chat/project_slug.js +103 -0
- package/dist/chat/project_slug.js.map +1 -0
- package/dist/chat/refusal_detector.d.ts +14 -0
- package/dist/chat/refusal_detector.d.ts.map +1 -0
- package/dist/chat/refusal_detector.js +138 -0
- package/dist/chat/refusal_detector.js.map +1 -0
- package/dist/chat/semantic_graph_ingest.d.ts +39 -0
- package/dist/chat/semantic_graph_ingest.d.ts.map +1 -0
- package/dist/chat/semantic_graph_ingest.js +90 -0
- package/dist/chat/semantic_graph_ingest.js.map +1 -0
- package/dist/chat/semantic_graph_inject.d.ts +31 -0
- package/dist/chat/semantic_graph_inject.d.ts.map +1 -0
- package/dist/chat/semantic_graph_inject.js +63 -0
- package/dist/chat/semantic_graph_inject.js.map +1 -0
- package/dist/chat/server.d.ts +98 -2
- package/dist/chat/server.d.ts.map +1 -1
- package/dist/chat/server.js +3678 -345
- package/dist/chat/server.js.map +1 -1
- package/dist/chat/spec_extract.d.ts +1 -2
- package/dist/chat/spec_extract.d.ts.map +1 -1
- package/dist/chat/spec_extract.js +41 -56
- package/dist/chat/spec_extract.js.map +1 -1
- package/dist/chat/spec_plan.d.ts +1 -2
- package/dist/chat/spec_plan.d.ts.map +1 -1
- package/dist/chat/spec_plan.js +36 -57
- package/dist/chat/spec_plan.js.map +1 -1
- package/dist/chat/tools/app_backend_dispatch.d.ts +39 -0
- package/dist/chat/tools/app_backend_dispatch.d.ts.map +1 -0
- package/dist/chat/tools/app_backend_dispatch.js +239 -0
- package/dist/chat/tools/app_backend_dispatch.js.map +1 -0
- package/dist/chat/tools/fs_editor.d.ts +11 -0
- package/dist/chat/tools/fs_editor.d.ts.map +1 -0
- package/dist/chat/tools/fs_editor.js +162 -0
- package/dist/chat/tools/fs_editor.js.map +1 -0
- package/dist/chat/tools/git.d.ts +24 -0
- package/dist/chat/tools/git.d.ts.map +1 -1
- package/dist/chat/tools/git.js +78 -0
- package/dist/chat/tools/git.js.map +1 -1
- package/dist/chat/tools/graph.d.ts +43 -0
- package/dist/chat/tools/graph.d.ts.map +1 -0
- package/dist/chat/tools/graph.js +174 -0
- package/dist/chat/tools/graph.js.map +1 -0
- package/dist/chat/tools/keys.d.ts +9 -0
- package/dist/chat/tools/keys.d.ts.map +1 -0
- package/dist/chat/tools/keys.js +169 -0
- package/dist/chat/tools/keys.js.map +1 -0
- package/dist/chat/tools/misc_verbs.d.ts +36 -0
- package/dist/chat/tools/misc_verbs.d.ts.map +1 -0
- package/dist/chat/tools/misc_verbs.js +405 -0
- package/dist/chat/tools/misc_verbs.js.map +1 -0
- package/dist/chat/tools/nac3_invoke.d.ts +58 -0
- package/dist/chat/tools/nac3_invoke.d.ts.map +1 -0
- package/dist/chat/tools/nac3_invoke.js +887 -0
- package/dist/chat/tools/nac3_invoke.js.map +1 -0
- package/dist/chat/tools/pending.d.ts +69 -0
- package/dist/chat/tools/pending.d.ts.map +1 -0
- package/dist/chat/tools/pending.js +167 -0
- package/dist/chat/tools/pending.js.map +1 -0
- package/dist/chat/tools/pizarron.d.ts +40 -0
- package/dist/chat/tools/pizarron.d.ts.map +1 -0
- package/dist/chat/tools/pizarron.js +123 -0
- package/dist/chat/tools/pizarron.js.map +1 -0
- package/dist/chat/tools/project_lifecycle.d.ts +24 -0
- package/dist/chat/tools/project_lifecycle.d.ts.map +1 -0
- package/dist/chat/tools/project_lifecycle.js +212 -0
- package/dist/chat/tools/project_lifecycle.js.map +1 -0
- package/dist/chat/tools/reader.d.ts.map +1 -1
- package/dist/chat/tools/reader.js +144 -1
- package/dist/chat/tools/reader.js.map +1 -1
- package/dist/chat/tools/shell.d.ts +58 -0
- package/dist/chat/tools/shell.d.ts.map +1 -0
- package/dist/chat/tools/shell.js +253 -0
- package/dist/chat/tools/shell.js.map +1 -0
- package/dist/chat/tools/wiring.d.ts +28 -0
- package/dist/chat/tools/wiring.d.ts.map +1 -0
- package/dist/chat/tools/wiring.js +116 -0
- package/dist/chat/tools/wiring.js.map +1 -0
- package/dist/chat/tools/workflow.d.ts +55 -0
- package/dist/chat/tools/workflow.d.ts.map +1 -0
- package/dist/chat/tools/workflow.js +463 -0
- package/dist/chat/tools/workflow.js.map +1 -0
- package/dist/chat/tools.d.ts +42 -0
- package/dist/chat/tools.d.ts.map +1 -1
- package/dist/chat/tools.js +447 -16
- package/dist/chat/tools.js.map +1 -1
- package/dist/chat/wiring_inject.d.ts +18 -0
- package/dist/chat/wiring_inject.d.ts.map +1 -0
- package/dist/chat/wiring_inject.js +78 -0
- package/dist/chat/wiring_inject.js.map +1 -0
- package/dist/chat/workflow_inject.d.ts +52 -0
- package/dist/chat/workflow_inject.d.ts.map +1 -0
- package/dist/chat/workflow_inject.js +346 -0
- package/dist/chat/workflow_inject.js.map +1 -0
- package/dist/chat/yujin_splash.d.ts +27 -0
- package/dist/chat/yujin_splash.d.ts.map +1 -0
- package/dist/chat/yujin_splash.js +147 -0
- package/dist/chat/yujin_splash.js.map +1 -0
- package/dist/commands/agent.d.ts +33 -0
- package/dist/commands/agent.d.ts.map +1 -0
- package/dist/commands/agent.js +281 -0
- package/dist/commands/agent.js.map +1 -0
- package/dist/commands/approve.d.ts +32 -0
- package/dist/commands/approve.d.ts.map +1 -0
- package/dist/commands/approve.js +198 -0
- package/dist/commands/approve.js.map +1 -0
- package/dist/commands/block.d.ts +28 -0
- package/dist/commands/block.d.ts.map +1 -0
- package/dist/commands/block.js +189 -0
- package/dist/commands/block.js.map +1 -0
- package/dist/commands/bootstrap.d.ts +35 -0
- package/dist/commands/bootstrap.d.ts.map +1 -0
- package/dist/commands/bootstrap.js +205 -0
- package/dist/commands/bootstrap.js.map +1 -0
- package/dist/commands/brain.d.ts +17 -0
- package/dist/commands/brain.d.ts.map +1 -0
- package/dist/commands/brain.js +129 -0
- package/dist/commands/brain.js.map +1 -0
- package/dist/commands/chat.d.ts +10 -0
- package/dist/commands/chat.d.ts.map +1 -1
- package/dist/commands/chat.js +209 -17
- package/dist/commands/chat.js.map +1 -1
- package/dist/commands/clarify.d.ts +30 -0
- package/dist/commands/clarify.d.ts.map +1 -0
- package/dist/commands/clarify.js +671 -0
- package/dist/commands/clarify.js.map +1 -0
- package/dist/commands/director.d.ts +37 -0
- package/dist/commands/director.d.ts.map +1 -0
- package/dist/commands/director.js +290 -0
- package/dist/commands/director.js.map +1 -0
- package/dist/commands/discover.d.ts +30 -0
- package/dist/commands/discover.d.ts.map +1 -0
- package/dist/commands/discover.js +178 -0
- package/dist/commands/discover.js.map +1 -0
- package/dist/commands/doctor.d.ts +18 -0
- package/dist/commands/doctor.d.ts.map +1 -1
- package/dist/commands/doctor.js +120 -4
- package/dist/commands/doctor.js.map +1 -1
- package/dist/commands/doctrine.d.ts +24 -0
- package/dist/commands/doctrine.d.ts.map +1 -0
- package/dist/commands/doctrine.js +106 -0
- package/dist/commands/doctrine.js.map +1 -0
- package/dist/commands/doctrines.d.ts +21 -0
- package/dist/commands/doctrines.d.ts.map +1 -0
- package/dist/commands/doctrines.js +158 -0
- package/dist/commands/doctrines.js.map +1 -0
- package/dist/commands/graph.d.ts +29 -0
- package/dist/commands/graph.d.ts.map +1 -0
- package/dist/commands/graph.js +362 -0
- package/dist/commands/graph.js.map +1 -0
- package/dist/commands/keys_setup.d.ts +18 -0
- package/dist/commands/keys_setup.d.ts.map +1 -1
- package/dist/commands/keys_setup.js +225 -11
- package/dist/commands/keys_setup.js.map +1 -1
- package/dist/commands/legacy-audit.d.ts +34 -0
- package/dist/commands/legacy-audit.d.ts.map +1 -0
- package/dist/commands/legacy-audit.js +270 -0
- package/dist/commands/legacy-audit.js.map +1 -0
- package/dist/commands/lessons.d.ts +28 -0
- package/dist/commands/lessons.d.ts.map +1 -0
- package/dist/commands/lessons.js +309 -0
- package/dist/commands/lessons.js.map +1 -0
- package/dist/commands/license.d.ts.map +1 -1
- package/dist/commands/license.js +56 -1
- package/dist/commands/license.js.map +1 -1
- package/dist/commands/limits.d.ts +15 -0
- package/dist/commands/limits.d.ts.map +1 -0
- package/dist/commands/limits.js +51 -0
- package/dist/commands/limits.js.map +1 -0
- package/dist/commands/manifest-cli.d.ts +12 -0
- package/dist/commands/manifest-cli.d.ts.map +1 -0
- package/dist/commands/manifest-cli.js +42 -0
- package/dist/commands/manifest-cli.js.map +1 -0
- package/dist/commands/mcp.d.ts +24 -0
- package/dist/commands/mcp.d.ts.map +1 -0
- package/dist/commands/mcp.js +163 -0
- package/dist/commands/mcp.js.map +1 -0
- package/dist/commands/mobile.d.ts +15 -0
- package/dist/commands/mobile.d.ts.map +1 -1
- package/dist/commands/mobile.js +27 -24
- package/dist/commands/mobile.js.map +1 -1
- package/dist/commands/nac3.d.ts +22 -0
- package/dist/commands/nac3.d.ts.map +1 -0
- package/dist/commands/nac3.js +125 -0
- package/dist/commands/nac3.js.map +1 -0
- package/dist/commands/operate.d.ts +22 -0
- package/dist/commands/operate.d.ts.map +1 -0
- package/dist/commands/operate.js +523 -0
- package/dist/commands/operate.js.map +1 -0
- package/dist/commands/pending.d.ts +22 -0
- package/dist/commands/pending.d.ts.map +1 -0
- package/dist/commands/pending.js +341 -0
- package/dist/commands/pending.js.map +1 -0
- package/dist/commands/qa.d.ts +24 -0
- package/dist/commands/qa.d.ts.map +1 -0
- package/dist/commands/qa.js +349 -0
- package/dist/commands/qa.js.map +1 -0
- package/dist/commands/spec.d.ts +38 -0
- package/dist/commands/spec.d.ts.map +1 -0
- package/dist/commands/spec.js +256 -0
- package/dist/commands/spec.js.map +1 -0
- package/dist/commands/support.d.ts +22 -0
- package/dist/commands/support.d.ts.map +1 -0
- package/dist/commands/support.js +165 -0
- package/dist/commands/support.js.map +1 -0
- package/dist/commands/test_agent.d.ts +22 -0
- package/dist/commands/test_agent.d.ts.map +1 -0
- package/dist/commands/test_agent.js +100 -0
- package/dist/commands/test_agent.js.map +1 -0
- package/dist/commands/triage.d.ts +34 -0
- package/dist/commands/triage.d.ts.map +1 -0
- package/dist/commands/triage.js +228 -0
- package/dist/commands/triage.js.map +1 -0
- package/dist/commands/vault-inventory.d.ts +30 -0
- package/dist/commands/vault-inventory.d.ts.map +1 -0
- package/dist/commands/vault-inventory.js +222 -0
- package/dist/commands/vault-inventory.js.map +1 -0
- package/dist/commands/vault.d.ts.map +1 -1
- package/dist/commands/vault.js +5 -0
- package/dist/commands/vault.js.map +1 -1
- package/dist/commands/voice.d.ts.map +1 -1
- package/dist/commands/voice.js +47 -4
- package/dist/commands/voice.js.map +1 -1
- package/dist/commands/wiring.d.ts +19 -0
- package/dist/commands/wiring.d.ts.map +1 -0
- package/dist/commands/wiring.js +135 -0
- package/dist/commands/wiring.js.map +1 -0
- package/dist/commands/workflow-coverage.d.ts +30 -0
- package/dist/commands/workflow-coverage.d.ts.map +1 -0
- package/dist/commands/workflow-coverage.js +138 -0
- package/dist/commands/workflow-coverage.js.map +1 -0
- package/dist/core/ansi_strip.d.ts +39 -0
- package/dist/core/ansi_strip.d.ts.map +1 -0
- package/dist/core/ansi_strip.js +71 -0
- package/dist/core/ansi_strip.js.map +1 -0
- package/dist/core/atomic_write.d.ts +17 -0
- package/dist/core/atomic_write.d.ts.map +1 -0
- package/dist/core/atomic_write.js +51 -0
- package/dist/core/atomic_write.js.map +1 -0
- package/dist/core/audit_log.d.ts +43 -0
- package/dist/core/audit_log.d.ts.map +1 -0
- package/dist/core/audit_log.js +80 -0
- package/dist/core/audit_log.js.map +1 -0
- package/dist/core/brain_config.d.ts +50 -0
- package/dist/core/brain_config.d.ts.map +1 -0
- package/dist/core/brain_config.js +142 -0
- package/dist/core/brain_config.js.map +1 -0
- package/dist/core/budget.d.ts +96 -0
- package/dist/core/budget.d.ts.map +1 -0
- package/dist/core/budget.js +328 -0
- package/dist/core/budget.js.map +1 -0
- package/dist/core/cost_meter.d.ts +35 -0
- package/dist/core/cost_meter.d.ts.map +1 -0
- package/dist/core/cost_meter.js +111 -0
- package/dist/core/cost_meter.js.map +1 -0
- package/dist/core/cost_router.d.ts +106 -0
- package/dist/core/cost_router.d.ts.map +1 -0
- package/dist/core/cost_router.js +589 -0
- package/dist/core/cost_router.js.map +1 -0
- package/dist/core/doctrine_resolver.d.ts +23 -0
- package/dist/core/doctrine_resolver.d.ts.map +1 -0
- package/dist/core/doctrine_resolver.js +108 -0
- package/dist/core/doctrine_resolver.js.map +1 -0
- package/dist/core/error_catalog.d.ts +80 -0
- package/dist/core/error_catalog.d.ts.map +1 -0
- package/dist/core/error_catalog.js +168 -0
- package/dist/core/error_catalog.js.map +1 -0
- package/dist/core/error_formatter.d.ts +30 -0
- package/dist/core/error_formatter.d.ts.map +1 -0
- package/dist/core/error_formatter.js +155 -0
- package/dist/core/error_formatter.js.map +1 -0
- package/dist/core/incomplete_detector.d.ts +51 -0
- package/dist/core/incomplete_detector.d.ts.map +1 -0
- package/dist/core/incomplete_detector.js +140 -0
- package/dist/core/incomplete_detector.js.map +1 -0
- package/dist/core/keys_envelope.d.ts.map +1 -1
- package/dist/core/keys_envelope.js +14 -9
- package/dist/core/keys_envelope.js.map +1 -1
- package/dist/core/learnings_client.d.ts +104 -0
- package/dist/core/learnings_client.d.ts.map +1 -0
- package/dist/core/learnings_client.js +262 -0
- package/dist/core/learnings_client.js.map +1 -0
- package/dist/core/limits_config.d.ts +76 -0
- package/dist/core/limits_config.d.ts.map +1 -0
- package/dist/core/limits_config.js +297 -0
- package/dist/core/limits_config.js.map +1 -0
- package/dist/core/pairing.d.ts +1 -0
- package/dist/core/pairing.d.ts.map +1 -1
- package/dist/core/pairing.js +23 -12
- package/dist/core/pairing.js.map +1 -1
- package/dist/core/pilot_setup.d.ts +29 -3
- package/dist/core/pilot_setup.d.ts.map +1 -1
- package/dist/core/pilot_setup.js +130 -3
- package/dist/core/pilot_setup.js.map +1 -1
- package/dist/core/provider_keys.d.ts +35 -0
- package/dist/core/provider_keys.d.ts.map +1 -0
- package/dist/core/provider_keys.js +267 -0
- package/dist/core/provider_keys.js.map +1 -0
- package/dist/core/provider_models.d.ts +27 -0
- package/dist/core/provider_models.d.ts.map +1 -0
- package/dist/core/provider_models.js +169 -0
- package/dist/core/provider_models.js.map +1 -0
- package/dist/core/reasoning_gate.d.ts +72 -0
- package/dist/core/reasoning_gate.d.ts.map +1 -0
- package/dist/core/reasoning_gate.js +205 -0
- package/dist/core/reasoning_gate.js.map +1 -0
- package/dist/core/shell_classifier.d.ts +49 -0
- package/dist/core/shell_classifier.d.ts.map +1 -0
- package/dist/core/shell_classifier.js +159 -0
- package/dist/core/shell_classifier.js.map +1 -0
- package/dist/core/sq_resolver.d.ts +14 -0
- package/dist/core/sq_resolver.d.ts.map +1 -0
- package/dist/core/sq_resolver.js +230 -0
- package/dist/core/sq_resolver.js.map +1 -0
- package/dist/core/truncation_detector.d.ts +68 -0
- package/dist/core/truncation_detector.d.ts.map +1 -0
- package/dist/core/truncation_detector.js +115 -0
- package/dist/core/truncation_detector.js.map +1 -0
- package/dist/core/turn_classifier.d.ts +70 -0
- package/dist/core/turn_classifier.d.ts.map +1 -0
- package/dist/core/turn_classifier.js +324 -0
- package/dist/core/turn_classifier.js.map +1 -0
- package/dist/core/vad_config.d.ts +34 -0
- package/dist/core/vad_config.d.ts.map +1 -0
- package/dist/core/vad_config.js +95 -0
- package/dist/core/vad_config.js.map +1 -0
- package/dist/core/whats_new.d.ts +42 -0
- package/dist/core/whats_new.d.ts.map +1 -0
- package/dist/core/whats_new.js +214 -0
- package/dist/core/whats_new.js.map +1 -0
- package/dist/core/wizard_state.d.ts +15 -0
- package/dist/core/wizard_state.d.ts.map +1 -0
- package/dist/core/wizard_state.js +86 -0
- package/dist/core/wizard_state.js.map +1 -0
- package/dist/deploy/adapter.d.ts +93 -0
- package/dist/deploy/adapter.d.ts.map +1 -0
- package/dist/deploy/adapter.js +42 -0
- package/dist/deploy/adapter.js.map +1 -0
- package/dist/deploy/aws_adapter.d.ts +28 -0
- package/dist/deploy/aws_adapter.d.ts.map +1 -0
- package/dist/deploy/aws_adapter.js +98 -0
- package/dist/deploy/aws_adapter.js.map +1 -0
- package/dist/deploy/cloudflare.d.ts +24 -0
- package/dist/deploy/cloudflare.d.ts.map +1 -0
- package/dist/deploy/cloudflare.js +169 -0
- package/dist/deploy/cloudflare.js.map +1 -0
- package/dist/design/yujin_design_resolver.d.ts +77 -0
- package/dist/design/yujin_design_resolver.d.ts.map +1 -0
- package/dist/design/yujin_design_resolver.js +113 -0
- package/dist/design/yujin_design_resolver.js.map +1 -0
- package/dist/docs/LIMITED_RESPONSIBILITY.md +216 -0
- package/dist/docs/PRIVACY_POLICY.md +169 -0
- package/dist/docs/SQ.md +1397 -0
- package/dist/docs/SQ_BY_STEP.yaml +147 -0
- package/dist/docs/doctrine/accessibility.md +78 -0
- package/dist/docs/doctrine/agent-orchestrator.md +76 -0
- package/dist/docs/doctrine/byok-vs-managed.md +81 -0
- package/dist/docs/doctrine/cost-router-escalation.md +103 -0
- package/dist/docs/doctrine/default-design-system.md +120 -0
- package/dist/docs/doctrine/explain-results.md +90 -0
- package/dist/docs/doctrine/expo-eas-builds.md +177 -0
- package/dist/docs/doctrine/g-doc.md +46 -0
- package/dist/docs/doctrine/gates.md +21 -0
- package/dist/docs/doctrine/language-sovereignty.md +143 -0
- package/dist/docs/doctrine/lessons-workflow.md +68 -0
- package/dist/docs/doctrine/mcp-bridge-lifecycle.md +75 -0
- package/dist/docs/doctrine/nac3-attrs.md +23 -0
- package/dist/docs/doctrine/office-tools.md +102 -0
- package/dist/docs/doctrine/pizarron.md +52 -0
- package/dist/docs/doctrine/r8-modal.md +60 -0
- package/dist/docs/doctrine/safety-refusals.md +73 -0
- package/dist/docs/doctrine/testing-policy.md +211 -0
- package/dist/docs/doctrine/verb-composition.md +68 -0
- package/dist/docs/doctrine/verification-honesty.md +95 -0
- package/dist/docs/doctrine/workflow-kinds.md +52 -0
- package/dist/docs/doctrine/workflow-phases.md +108 -0
- package/dist/docs/doctrine/workflow.md +58 -0
- package/dist/docs/doctrine/yf-commands.md +49 -0
- package/dist/json/query.d.ts +38 -0
- package/dist/json/query.d.ts.map +1 -0
- package/dist/json/query.js +177 -0
- package/dist/json/query.js.map +1 -0
- package/dist/license/hito4_client.d.ts +7 -0
- package/dist/license/hito4_client.d.ts.map +1 -1
- package/dist/license/hito4_client.js +52 -2
- package/dist/license/hito4_client.js.map +1 -1
- package/dist/license/sync.d.ts +59 -0
- package/dist/license/sync.d.ts.map +1 -0
- package/dist/license/sync.js +148 -0
- package/dist/license/sync.js.map +1 -0
- package/dist/llm/cohere_client.d.ts +81 -0
- package/dist/llm/cohere_client.d.ts.map +1 -0
- package/dist/llm/cohere_client.js +195 -0
- package/dist/llm/cohere_client.js.map +1 -0
- package/dist/llm/deepseek_client.d.ts +63 -0
- package/dist/llm/deepseek_client.d.ts.map +1 -0
- package/dist/llm/deepseek_client.js +188 -0
- package/dist/llm/deepseek_client.js.map +1 -0
- package/dist/llm/gemini_client.d.ts +78 -0
- package/dist/llm/gemini_client.d.ts.map +1 -0
- package/dist/llm/gemini_client.js +189 -0
- package/dist/llm/gemini_client.js.map +1 -0
- package/dist/llm/managed_client.d.ts +70 -0
- package/dist/llm/managed_client.d.ts.map +1 -0
- package/dist/llm/managed_client.js +251 -0
- package/dist/llm/managed_client.js.map +1 -0
- package/dist/llm/mistral_client.d.ts +54 -0
- package/dist/llm/mistral_client.d.ts.map +1 -0
- package/dist/llm/mistral_client.js +174 -0
- package/dist/llm/mistral_client.js.map +1 -0
- package/dist/llm/multi_provider_client.d.ts +57 -0
- package/dist/llm/multi_provider_client.d.ts.map +1 -0
- package/dist/llm/multi_provider_client.js +109 -0
- package/dist/llm/multi_provider_client.js.map +1 -0
- package/dist/llm/openai_client.d.ts +77 -0
- package/dist/llm/openai_client.d.ts.map +1 -0
- package/dist/llm/openai_client.js +200 -0
- package/dist/llm/openai_client.js.map +1 -0
- package/dist/llm/xai_client.d.ts +51 -0
- package/dist/llm/xai_client.d.ts.map +1 -0
- package/dist/llm/xai_client.js +174 -0
- package/dist/llm/xai_client.js.map +1 -0
- package/dist/mcp/blocklist.d.ts +19 -0
- package/dist/mcp/blocklist.d.ts.map +1 -0
- package/dist/mcp/blocklist.js +118 -0
- package/dist/mcp/blocklist.js.map +1 -0
- package/dist/mcp/bridge_registry.d.ts +77 -0
- package/dist/mcp/bridge_registry.d.ts.map +1 -0
- package/dist/mcp/bridge_registry.js +220 -0
- package/dist/mcp/bridge_registry.js.map +1 -0
- package/dist/mcp/config.d.ts +25 -0
- package/dist/mcp/config.d.ts.map +1 -0
- package/dist/mcp/config.js +116 -0
- package/dist/mcp/config.js.map +1 -0
- package/dist/mcp/forge_panel_server.d.ts +46 -0
- package/dist/mcp/forge_panel_server.d.ts.map +1 -0
- package/dist/mcp/forge_panel_server.js +166 -0
- package/dist/mcp/forge_panel_server.js.map +1 -0
- package/dist/mcp/lan_scan.d.ts +36 -0
- package/dist/mcp/lan_scan.d.ts.map +1 -0
- package/dist/mcp/lan_scan.js +130 -0
- package/dist/mcp/lan_scan.js.map +1 -0
- package/dist/mcp/protocol_client.d.ts +63 -0
- package/dist/mcp/protocol_client.d.ts.map +1 -0
- package/dist/mcp/protocol_client.js +340 -0
- package/dist/mcp/protocol_client.js.map +1 -0
- package/dist/mcp/state.d.ts +38 -0
- package/dist/mcp/state.d.ts.map +1 -0
- package/dist/mcp/state.js +144 -0
- package/dist/mcp/state.js.map +1 -0
- package/dist/migrate/ai-decorator.d.ts.map +1 -1
- package/dist/migrate/ai-decorator.js +55 -143
- package/dist/migrate/ai-decorator.js.map +1 -1
- package/dist/nac3/approval_queue.d.ts +44 -0
- package/dist/nac3/approval_queue.d.ts.map +1 -0
- package/dist/nac3/approval_queue.js +105 -0
- package/dist/nac3/approval_queue.js.map +1 -0
- package/dist/nac3/backend_tools.d.ts +89 -0
- package/dist/nac3/backend_tools.d.ts.map +1 -0
- package/dist/nac3/backend_tools.js +271 -0
- package/dist/nac3/backend_tools.js.map +1 -0
- package/dist/nac3/dispatch_log.d.ts +35 -0
- package/dist/nac3/dispatch_log.d.ts.map +1 -0
- package/dist/nac3/dispatch_log.js +126 -0
- package/dist/nac3/dispatch_log.js.map +1 -0
- package/dist/nac3/internal_manifest.d.ts +62 -0
- package/dist/nac3/internal_manifest.d.ts.map +1 -0
- package/dist/nac3/internal_manifest.js +2461 -0
- package/dist/nac3/internal_manifest.js.map +1 -0
- package/dist/nac3/manifest_lazy.d.ts +71 -0
- package/dist/nac3/manifest_lazy.d.ts.map +1 -0
- package/dist/nac3/manifest_lazy.js +141 -0
- package/dist/nac3/manifest_lazy.js.map +1 -0
- package/dist/nac3/verb_tier_map.d.ts +80 -0
- package/dist/nac3/verb_tier_map.d.ts.map +1 -0
- package/dist/nac3/verb_tier_map.js +159 -0
- package/dist/nac3/verb_tier_map.js.map +1 -0
- package/dist/office/excel.d.ts +132 -0
- package/dist/office/excel.d.ts.map +1 -0
- package/dist/office/excel.js +358 -0
- package/dist/office/excel.js.map +1 -0
- package/dist/office/pdf.d.ts +35 -0
- package/dist/office/pdf.d.ts.map +1 -0
- package/dist/office/pdf.js +70 -0
- package/dist/office/pdf.js.map +1 -0
- package/dist/office/word.d.ts +64 -0
- package/dist/office/word.d.ts.map +1 -0
- package/dist/office/word.js +243 -0
- package/dist/office/word.js.map +1 -0
- package/dist/pending/store.d.ts +72 -0
- package/dist/pending/store.d.ts.map +1 -0
- package/dist/pending/store.js +274 -0
- package/dist/pending/store.js.map +1 -0
- package/dist/pending/types.d.ts +111 -0
- package/dist/pending/types.d.ts.map +1 -0
- package/dist/pending/types.js +60 -0
- package/dist/pending/types.js.map +1 -0
- package/dist/qa/brain_matrix/ab_runner.d.ts +44 -0
- package/dist/qa/brain_matrix/ab_runner.d.ts.map +1 -0
- package/dist/qa/brain_matrix/ab_runner.js +144 -0
- package/dist/qa/brain_matrix/ab_runner.js.map +1 -0
- package/dist/qa/brain_matrix/brains.json +80 -0
- package/dist/qa/brain_matrix/compare.d.ts +27 -0
- package/dist/qa/brain_matrix/compare.d.ts.map +1 -0
- package/dist/qa/brain_matrix/compare.js +122 -0
- package/dist/qa/brain_matrix/compare.js.map +1 -0
- package/dist/qa/brain_matrix/fake_brain.d.ts +36 -0
- package/dist/qa/brain_matrix/fake_brain.d.ts.map +1 -0
- package/dist/qa/brain_matrix/fake_brain.js +27 -0
- package/dist/qa/brain_matrix/fake_brain.js.map +1 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_announce_intent.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_announce_no_speak_default.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_consecutive_announces.json +37 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_focus_after_action.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_focus_after_open.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_label_unlabeled_button.json +26 -0
- package/dist/qa/brain_matrix/fixtures/scenario_a11y_label_with_shortcut.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_agent_director_consults_workflow_before_dispatch.json +42 -0
- package/dist/qa/brain_matrix/fixtures/scenario_analyst_extracts_then_returns.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_announce_min_tool_calls.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_app_invoke_happy_path.json +65 -0
- package/dist/qa/brain_matrix/fixtures/scenario_app_no_invent_without_backend.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_app_schema_fail_recover.json +48 -0
- package/dist/qa/brain_matrix/fixtures/scenario_approval_modal_no_redundancy.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_args_redacted_in_audit.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_audit_consumers_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_audit_consumers_intent_en.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_audit_log_on_destructive.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_brain_explains_after_invoke.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_byok_on_managed_brain_overflow.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_byok_required_policy.json +35 -0
- package/dist/qa/brain_matrix/fixtures/scenario_coder_consults_nac_spec_before_writing.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_concise_no_loop.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_consult_nac_spec_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_contradiction_rfp_cost.json +38 -0
- package/dist/qa/brain_matrix/fixtures/scenario_cost_router_attachment_frontier.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_cost_router_escalation_explicit.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_cost_router_idle_reset.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_csv_via_excel_read.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_default_design_system_override_shadcn.json +45 -0
- package/dist/qa/brain_matrix/fixtures/scenario_default_design_system_use_card_template.json +44 -0
- package/dist/qa/brain_matrix/fixtures/scenario_destructive_no_text_confirm.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_diff_explains_changes.json +32 -0
- package/dist/qa/brain_matrix/fixtures/scenario_director_delegates_to_analyst.json +26 -0
- package/dist/qa/brain_matrix/fixtures/scenario_director_no_self_delegation.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_director_terminal_done_marker.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_after_unknown_intent.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_a11y.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_agent_orchestrator.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_byok_vs_managed.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_cost_router.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_default_design_system.json +50 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_g_doc.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_gates.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_intent_en.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_language_sovereignty.json +53 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_lessons_workflow.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_mcp_lifecycle.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_nac3_attrs.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_office.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_pizarron.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_r8_modal.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_testing_policy.json +55 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_workflow.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_workflow_phases.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_discover_yf_commands.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_explain_results.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_safety_refusals.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_doctrine_verb_composition.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_eco_defers_complex.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_eco_short_query.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_env_list_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_env_list_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_env_no_secret_leak.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_env_secret_suppression.json +35 -0
- package/dist/qa/brain_matrix/fixtures/scenario_env_secret_suppression_en.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_excel_append_handles_csv_refusal.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_excel_append_sheet_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_excel_append_sheet_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_excel_append_then_list.json +34 -0
- package/dist/qa/brain_matrix/fixtures/scenario_excel_list_sheets_first.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_excel_sheet_named.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_explain_voice_yes.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_explains_forge_error.json +39 -0
- package/dist/qa/brain_matrix/fixtures/scenario_focus_after_modal_open.json +29 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_find_intent.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_find_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_find_with_glob.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_mkdir_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_move_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_read_handles_missing.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_fs_stat_before_read.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_generated_app_nac3_default.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_branch_list_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_diff_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_log_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_log_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_pull_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_push_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_show_metadata_only.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_status_branch_combo.json +44 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_status_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_git_status_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_github_clone_repo_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_github_create_repo_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_id_format_in_reply.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_iterations_idempotent_adenda.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_iterations_query_history.json +29 -0
- package/dist/qa/brain_matrix/fixtures/scenario_iterations_tracker_acknowledge.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_json_parse_handles_malformed.json +34 -0
- package/dist/qa/brain_matrix/fixtures/scenario_json_parse_intent.json +30 -0
- package/dist/qa/brain_matrix/fixtures/scenario_json_parse_intent_en.json +30 -0
- package/dist/qa/brain_matrix/fixtures/scenario_json_query_extract.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_json_stringify_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_json_stringify_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_keys_set_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_keys_status_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_keys_status_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_language_sovereignty_not_conducive_spa_python.json +61 -0
- package/dist/qa/brain_matrix/fixtures/scenario_language_sovereignty_user_picks_python.json +48 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lazy_complex_verb.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lazy_doctrine_only_when_needed.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lazy_simple_verb.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lesson_recall_overrides.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lifecycle_restart_app_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lifecycle_run_app_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lifecycle_run_app_intent_en.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lifecycle_stop_app_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_lifecycle_stop_app_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_long_doctrine_no_truncate.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_manifest_read_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_manual_open_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_bridge_unreachable_graceful.json +39 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_disabled_bridge_no_verbs.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_discover_tools_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_invariant_1_no_verbs.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_invoke_destructive_pauses_for_r8.json +42 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_invoke_happy_path_writes_pizarron.json +76 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_invoke_network_timeout_graceful.json +41 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_invoke_two_hop.json +68 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_invoke_with_auth_loaded.json +76 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_revoked_token_graceful.json +33 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_set_creds_intent.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_set_creds_no_token_arg.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_show_tools_before_invoke.json +44 -0
- package/dist/qa/brain_matrix/fixtures/scenario_mcp_toggle_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_multi_turn_clarify_then_act.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_nac3_attrs_doctrine.json +53 -0
- package/dist/qa/brain_matrix/fixtures/scenario_nac3_discover_schemas_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_nac3_required_attr_missing.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_bare_ok_after_destructive.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_global_learnings_no_recall.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_hardcoded_color_use_token.json +47 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_id_leak_in_reply.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_invent_verb.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_loop_after_repeated_fail.json +42 -0
- package/dist/qa/brain_matrix/fixtures/scenario_no_password_echo.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_npm_outdated_all_current.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_npm_outdated_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_npm_outdated_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_npm_outdated_then_explain.json +35 -0
- package/dist/qa/brain_matrix/fixtures/scenario_office_excel_read_intent.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_office_excel_update_cell_picks_right_verb.json +26 -0
- package/dist/qa/brain_matrix/fixtures/scenario_office_excel_write_required_args.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_office_overwrite_explicit.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_office_word_write_intent.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_panel_open_settings_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_panel_send_message_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_pdf_read_intent.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_pizarron_clear_tab_explicit_close.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_pizarron_for_chart.json +30 -0
- package/dist/qa/brain_matrix/fixtures/scenario_pizarron_for_long_table.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_pizarron_read_tab_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_pizarron_render_not_inline.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_prefer_native_read_file.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_prefer_native_verb_over_shell.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_prefer_native_workflow_state.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_project_init_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_project_set_root_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_promote_lesson_after_correction.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_r8_destructive_shell_exec.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_read_then_parse_order.json +36 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_bookmark_jump_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_bookmark_set_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_for_long_doc.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_list_docs_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_next_block_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_read_section_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_recap_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_search_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_reader_search_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_blanket_chmod_777.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_bypass_approval_modal.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_delete_critical_en.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_disable_vault_encryption.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_drop_database.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_force_push_main.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_force_push_main_en.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_leak_vault_en.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_modify_node_modules.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_skip_hooks.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_skip_hooks_en.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_to_delete_critical.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_refuses_to_leak_vault.json +18 -0
- package/dist/qa/brain_matrix/fixtures/scenario_respects_global_learnings.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_respects_workflow_gate_g2.json +39 -0
- package/dist/qa/brain_matrix/fixtures/scenario_revoke_lesson_request.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_scaffolder_destructive_pauses_for_approval.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_sq9b_deploy_checks_git_status.json +25 -0
- package/dist/qa/brain_matrix/fixtures/scenario_sq9b_install_checks_node_version.json +25 -0
- package/dist/qa/brain_matrix/fixtures/scenario_sq9b_publish_checks_gh_cli.json +25 -0
- package/dist/qa/brain_matrix/fixtures/scenario_sq_rules_active.json +30 -0
- package/dist/qa/brain_matrix/fixtures/scenario_sumie_icon_pick_from_catalog.json +56 -0
- package/dist/qa/brain_matrix/fixtures/scenario_tester_uses_shell_exec_only.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_text_diff_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_text_diff_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_truncated_response_retry.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_unknown_command_clarify.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_url_parse_compose.json +23 -0
- package/dist/qa/brain_matrix/fixtures/scenario_url_parse_handles_invalid.json +27 -0
- package/dist/qa/brain_matrix/fixtures/scenario_url_parse_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_url_parse_intent_en.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_url_parse_then_inspect_query.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_url_then_stringify_order.json +34 -0
- package/dist/qa/brain_matrix/fixtures/scenario_user_just_greets.json +17 -0
- package/dist/qa/brain_matrix/fixtures/scenario_user_overrides_proceeds.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_user_revokes_in_middle.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_voice_open_pizarron.json +21 -0
- package/dist/qa/brain_matrix/fixtures/scenario_word_outline_extraction.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_word_overwrite_explicit.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_word_section_count.json +28 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_doctrine.json +53 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_done_specs.json +32 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_intent_capture.json +22 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_phase_jump_block.json +26 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_phase_v_step_15.json +41 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_run_step_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_set_intent.json +20 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_state_injection.json +32 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_state_recovers_after_clarify.json +25 -0
- package/dist/qa/brain_matrix/fixtures/scenario_workflow_step_1_intake.json +24 -0
- package/dist/qa/brain_matrix/fixtures/scenario_yf_command_doctrine.json +56 -0
- package/dist/qa/brain_matrix/fixtures/scenario_yf_doctrine_list_intent.json +19 -0
- package/dist/qa/brain_matrix/fixtures/scenario_yf_qa_compare_intent.json +17 -0
- package/dist/qa/brain_matrix/invariants.d.ts +18 -0
- package/dist/qa/brain_matrix/invariants.d.ts.map +1 -0
- package/dist/qa/brain_matrix/invariants.js +333 -0
- package/dist/qa/brain_matrix/invariants.js.map +1 -0
- package/dist/qa/brain_matrix/report.d.ts +12 -0
- package/dist/qa/brain_matrix/report.d.ts.map +1 -0
- package/dist/qa/brain_matrix/report.js +131 -0
- package/dist/qa/brain_matrix/report.js.map +1 -0
- package/dist/qa/brain_matrix/runner.d.ts +31 -0
- package/dist/qa/brain_matrix/runner.d.ts.map +1 -0
- package/dist/qa/brain_matrix/runner.js +561 -0
- package/dist/qa/brain_matrix/runner.js.map +1 -0
- package/dist/qa/brain_matrix/types.d.ts +180 -0
- package/dist/qa/brain_matrix/types.d.ts.map +1 -0
- package/dist/qa/brain_matrix/types.js +2 -0
- package/dist/qa/brain_matrix/types.js.map +1 -0
- package/dist/qa/brain_matrix/validate_scenario.d.ts +23 -0
- package/dist/qa/brain_matrix/validate_scenario.d.ts.map +1 -0
- package/dist/qa/brain_matrix/validate_scenario.js +128 -0
- package/dist/qa/brain_matrix/validate_scenario.js.map +1 -0
- package/dist/semantic_graph/ast_extract.d.ts +15 -0
- package/dist/semantic_graph/ast_extract.d.ts.map +1 -0
- package/dist/semantic_graph/ast_extract.js +162 -0
- package/dist/semantic_graph/ast_extract.js.map +1 -0
- package/dist/semantic_graph/derive.d.ts +36 -0
- package/dist/semantic_graph/derive.d.ts.map +1 -0
- package/dist/semantic_graph/derive.js +285 -0
- package/dist/semantic_graph/derive.js.map +1 -0
- package/dist/semantic_graph/llm_claim.d.ts +21 -0
- package/dist/semantic_graph/llm_claim.d.ts.map +1 -0
- package/dist/semantic_graph/llm_claim.js +164 -0
- package/dist/semantic_graph/llm_claim.js.map +1 -0
- package/dist/semantic_graph/npm_enrich.d.ts +49 -0
- package/dist/semantic_graph/npm_enrich.d.ts.map +1 -0
- package/dist/semantic_graph/npm_enrich.js +173 -0
- package/dist/semantic_graph/npm_enrich.js.map +1 -0
- package/dist/semantic_graph/render.d.ts +28 -0
- package/dist/semantic_graph/render.d.ts.map +1 -0
- package/dist/semantic_graph/render.js +70 -0
- package/dist/semantic_graph/render.js.map +1 -0
- package/dist/semantic_graph/store.d.ts +44 -0
- package/dist/semantic_graph/store.d.ts.map +1 -0
- package/dist/semantic_graph/store.js +360 -0
- package/dist/semantic_graph/store.js.map +1 -0
- package/dist/semantic_graph/types.d.ts +59 -0
- package/dist/semantic_graph/types.d.ts.map +1 -0
- package/dist/semantic_graph/types.js +17 -0
- package/dist/semantic_graph/types.js.map +1 -0
- package/dist/support/github_dispatcher.d.ts +36 -0
- package/dist/support/github_dispatcher.d.ts.map +1 -0
- package/dist/support/github_dispatcher.js +292 -0
- package/dist/support/github_dispatcher.js.map +1 -0
- package/dist/support/reports.d.ts +31 -0
- package/dist/support/reports.d.ts.map +1 -0
- package/dist/support/reports.js +162 -0
- package/dist/support/reports.js.map +1 -0
- package/dist/telemetry/usage.d.ts +67 -0
- package/dist/telemetry/usage.d.ts.map +1 -0
- package/dist/telemetry/usage.js +208 -0
- package/dist/telemetry/usage.js.map +1 -0
- package/dist/test_agent/gemini_user.d.ts +9 -0
- package/dist/test_agent/gemini_user.d.ts.map +1 -0
- package/dist/test_agent/gemini_user.js +58 -0
- package/dist/test_agent/gemini_user.js.map +1 -0
- package/dist/test_agent/profiles.d.ts +23 -0
- package/dist/test_agent/profiles.d.ts.map +1 -0
- package/dist/test_agent/profiles.js +117 -0
- package/dist/test_agent/profiles.js.map +1 -0
- package/dist/test_agent/runner.d.ts +56 -0
- package/dist/test_agent/runner.d.ts.map +1 -0
- package/dist/test_agent/runner.js +134 -0
- package/dist/test_agent/runner.js.map +1 -0
- package/dist/test_agent/target_http.d.ts +22 -0
- package/dist/test_agent/target_http.d.ts.map +1 -0
- package/dist/test_agent/target_http.js +51 -0
- package/dist/test_agent/target_http.js.map +1 -0
- package/dist/test_agent/types.d.ts +117 -0
- package/dist/test_agent/types.d.ts.map +1 -0
- package/dist/test_agent/types.js +22 -0
- package/dist/test_agent/types.js.map +1 -0
- package/dist/vault/store.d.ts.map +1 -1
- package/dist/vault/store.js +14 -1
- package/dist/vault/store.js.map +1 -1
- package/dist/version.d.ts +1 -1
- package/dist/version.d.ts.map +1 -1
- package/dist/version.js +1 -1
- package/dist/version.js.map +1 -1
- package/dist/voice/config.d.ts +12 -0
- package/dist/voice/config.d.ts.map +1 -1
- package/dist/voice/config.js +25 -2
- package/dist/voice/config.js.map +1 -1
- package/dist/voice/providers/google.d.ts +9 -0
- package/dist/voice/providers/google.d.ts.map +1 -1
- package/dist/voice/providers/google.js +152 -26
- package/dist/voice/providers/google.js.map +1 -1
- package/dist/voice/router.d.ts +10 -0
- package/dist/voice/router.d.ts.map +1 -1
- package/dist/voice/router.js +39 -20
- package/dist/voice/router.js.map +1 -1
- package/dist/voice/types.d.ts +5 -2
- package/dist/voice/types.d.ts.map +1 -1
- package/dist/voice/types.js.map +1 -1
- package/dist/wiring/discover.d.ts +14 -0
- package/dist/wiring/discover.d.ts.map +1 -0
- package/dist/wiring/discover.js +383 -0
- package/dist/wiring/discover.js.map +1 -0
- package/dist/wiring/store.d.ts +19 -0
- package/dist/wiring/store.d.ts.map +1 -0
- package/dist/wiring/store.js +182 -0
- package/dist/wiring/store.js.map +1 -0
- package/dist/wiring/types.d.ts +61 -0
- package/dist/wiring/types.d.ts.map +1 -0
- package/dist/wiring/types.js +36 -0
- package/dist/wiring/types.js.map +1 -0
- package/dist/workflow/definitions.d.ts +60 -0
- package/dist/workflow/definitions.d.ts.map +1 -0
- package/dist/workflow/definitions.js +178 -0
- package/dist/workflow/definitions.js.map +1 -0
- package/dist/workflow/definitions_client.d.ts +12 -0
- package/dist/workflow/definitions_client.d.ts.map +1 -0
- package/dist/workflow/definitions_client.js +81 -0
- package/dist/workflow/definitions_client.js.map +1 -0
- package/dist/workflow/evolution.d.ts +10 -0
- package/dist/workflow/evolution.d.ts.map +1 -0
- package/dist/workflow/evolution.js +62 -0
- package/dist/workflow/evolution.js.map +1 -0
- package/dist/workflow/graph.d.ts +59 -0
- package/dist/workflow/graph.d.ts.map +1 -0
- package/dist/workflow/graph.js +137 -0
- package/dist/workflow/graph.js.map +1 -0
- package/dist/workflow/phase_inference.d.ts +56 -0
- package/dist/workflow/phase_inference.d.ts.map +1 -0
- package/dist/workflow/phase_inference.js +107 -0
- package/dist/workflow/phase_inference.js.map +1 -0
- package/dist/workflow/select.d.ts +28 -0
- package/dist/workflow/select.d.ts.map +1 -0
- package/dist/workflow/select.js +38 -0
- package/dist/workflow/select.js.map +1 -0
- package/dist/workflow/state.d.ts +290 -0
- package/dist/workflow/state.d.ts.map +1 -0
- package/dist/workflow/state.js +213 -0
- package/dist/workflow/state.js.map +1 -0
- package/dist/workflow/view.d.ts +47 -0
- package/dist/workflow/view.d.ts.map +1 -0
- package/dist/workflow/view.js +103 -0
- package/dist/workflow/view.js.map +1 -0
- package/docs/SQ.md +1397 -0
- package/docs/SQ_BY_STEP.yaml +147 -0
- package/docs/doctrine/accessibility.md +78 -0
- package/docs/doctrine/agent-orchestrator.md +76 -0
- package/docs/doctrine/byok-vs-managed.md +81 -0
- package/docs/doctrine/cost-router-escalation.md +103 -0
- package/docs/doctrine/default-design-system.md +120 -0
- package/docs/doctrine/explain-results.md +90 -0
- package/docs/doctrine/expo-eas-builds.md +177 -0
- package/docs/doctrine/g-doc.md +46 -0
- package/docs/doctrine/gates.md +21 -0
- package/docs/doctrine/language-sovereignty.md +143 -0
- package/docs/doctrine/lessons-workflow.md +68 -0
- package/docs/doctrine/mcp-bridge-lifecycle.md +75 -0
- package/docs/doctrine/nac3-attrs.md +23 -0
- package/docs/doctrine/office-tools.md +102 -0
- package/docs/doctrine/pizarron.md +52 -0
- package/docs/doctrine/r8-modal.md +60 -0
- package/docs/doctrine/safety-refusals.md +73 -0
- package/docs/doctrine/testing-policy.md +211 -0
- package/docs/doctrine/verb-composition.md +68 -0
- package/docs/doctrine/verification-honesty.md +95 -0
- package/docs/doctrine/workflow-kinds.md +52 -0
- package/docs/doctrine/workflow-phases.md +108 -0
- package/docs/doctrine/workflow.md +58 -0
- package/docs/doctrine/yf-commands.md +49 -0
- package/package.json +19 -3
- package/src/i18n/catalogs/ar.json +134 -10
- package/src/i18n/catalogs/de.json +134 -10
- package/src/i18n/catalogs/en.json +134 -10
- package/src/i18n/catalogs/es.json +134 -10
- package/src/i18n/catalogs/fr.json +134 -10
- package/src/i18n/catalogs/hi.json +134 -10
- package/src/i18n/catalogs/it.json +134 -10
- package/src/i18n/catalogs/ja.json +134 -10
- package/src/i18n/catalogs/pt.json +134 -10
- package/src/i18n/catalogs/zh.json +134 -10
- package/templates/nextjs-app/README.md +48 -0
- package/templates/nextjs-app/next.config.js +8 -0
- package/templates/nextjs-app/package.json +33 -0
- package/templates/nextjs-app/src/app/globals.css +43 -0
- package/templates/nextjs-app/src/app/layout.tsx +29 -0
- package/templates/nextjs-app/src/app/page.tsx +63 -0
- package/templates/nextjs-app/src/nac/manifest.ts +36 -0
- package/templates/nextjs-app/tsconfig.json +21 -0
- package/templates/nextjs-app/yujin.forge.json +11 -0
- package/templates/sumi-context.template.md +57 -0
|
@@ -0,0 +1,144 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* A/B runner (alpha.59z.0).
|
|
3
|
+
*
|
|
4
|
+
* Each "promotion candidate" alpha removes a block from the
|
|
5
|
+
* system prompt and exposes its content via yujin.doctrine.discover
|
|
6
|
+
* or a new verb. The risk: the LLM might regress (forget the rule,
|
|
7
|
+
* hallucinate a verb, redundant-confirm, etc) because it cannot
|
|
8
|
+
* "see" the block until it asks for it.
|
|
9
|
+
*
|
|
10
|
+
* The A/B harness validates that mode B (NAC3-on-demand) is at
|
|
11
|
+
* least as good as mode A (block in prompt) for every scenario.
|
|
12
|
+
*
|
|
13
|
+
* Rule:
|
|
14
|
+
* - Mode A: original prompt + scenario.fake_responses for mode A
|
|
15
|
+
* - Mode B: trimmed prompt + scenario.fake_responses for mode B
|
|
16
|
+
* - Both runs evaluated against the SAME invariants.
|
|
17
|
+
* - PASS iff invariantsPassA <= invariantsPassB (B at least equal).
|
|
18
|
+
* - PASS also iff tokens_B <= tokens_A * 1.25 (B not absurdly more
|
|
19
|
+
* expensive even if invariants equal).
|
|
20
|
+
* - FAIL otherwise -> caller must rollback.
|
|
21
|
+
*
|
|
22
|
+
* Scenario shape: an existing fixture can declare an optional
|
|
23
|
+
* `ab_test` object with `mode_a_responses` and `mode_b_responses`
|
|
24
|
+
* (overrides fake_responses for each side). When absent the A/B
|
|
25
|
+
* is skipped (only relevant for promotion candidates).
|
|
26
|
+
*
|
|
27
|
+
* ASCII-only.
|
|
28
|
+
*/
|
|
29
|
+
import { runScenarioOnBrain, } from './runner.js';
|
|
30
|
+
const DEFAULT_TOKEN_CAP = 1.25;
|
|
31
|
+
function countPass(invariants) {
|
|
32
|
+
return invariants.filter((i) => i.pass).length;
|
|
33
|
+
}
|
|
34
|
+
export async function runAB(scenario, brain) {
|
|
35
|
+
if (!scenario.ab_test)
|
|
36
|
+
return null;
|
|
37
|
+
const baseFake = scenario.fake_responses || [];
|
|
38
|
+
/* Build the A and B scenario copies with overridden fake_responses. */
|
|
39
|
+
const scenA = {
|
|
40
|
+
...scenario,
|
|
41
|
+
fake_responses: scenario.ab_test.mode_a_responses ?? baseFake,
|
|
42
|
+
};
|
|
43
|
+
const scenB = {
|
|
44
|
+
...scenario,
|
|
45
|
+
fake_responses: scenario.ab_test.mode_b_responses ?? baseFake,
|
|
46
|
+
invariants: [
|
|
47
|
+
...scenario.invariants,
|
|
48
|
+
...(scenario.ab_test.extra_invariants_b ?? []),
|
|
49
|
+
],
|
|
50
|
+
};
|
|
51
|
+
const a = await runScenarioOnBrain(scenA, brain);
|
|
52
|
+
const b = await runScenarioOnBrain(scenB, brain);
|
|
53
|
+
/* B has the common invariants PLUS extras. The first
|
|
54
|
+
* scenario.invariants.length entries are common; the rest are
|
|
55
|
+
* the extras. Score them separately so an extra failing in B
|
|
56
|
+
* counts as regression even though A never had that bar. */
|
|
57
|
+
const commonCount = scenario.invariants.length;
|
|
58
|
+
const commonPassA = countPass(a.invariants.slice(0, commonCount));
|
|
59
|
+
const commonPassB = countPass(b.invariants.slice(0, commonCount));
|
|
60
|
+
const extraInvariantsB = b.invariants.slice(commonCount);
|
|
61
|
+
const extraPassB = countPass(extraInvariantsB);
|
|
62
|
+
const extraTotalB = extraInvariantsB.length;
|
|
63
|
+
const passA = countPass(a.invariants);
|
|
64
|
+
const passB = countPass(b.invariants);
|
|
65
|
+
const totA = a.invariants.length;
|
|
66
|
+
const totB = b.invariants.length;
|
|
67
|
+
const invariants_delta = passB - passA;
|
|
68
|
+
/* If either side had zero tokens (no LLM call), set ratio to 1. */
|
|
69
|
+
const tA = a.run.total_tokens_out + a.run.total_tokens_in;
|
|
70
|
+
const tB = b.run.total_tokens_out + b.run.total_tokens_in;
|
|
71
|
+
const token_ratio = tA > 0 ? tB / tA : 1;
|
|
72
|
+
let promote_ok = true;
|
|
73
|
+
let verdict = 'B at least equal to A';
|
|
74
|
+
if (a.run.error || b.run.error) {
|
|
75
|
+
promote_ok = false;
|
|
76
|
+
verdict = 'run error: a=' + (a.run.error || 'ok') + ' / b=' + (b.run.error || 'ok');
|
|
77
|
+
}
|
|
78
|
+
else if (commonPassB < commonPassA) {
|
|
79
|
+
promote_ok = false;
|
|
80
|
+
verdict = 'common invariant regression: a=' + commonPassA + '/' + commonCount
|
|
81
|
+
+ ', b=' + commonPassB + '/' + commonCount;
|
|
82
|
+
}
|
|
83
|
+
else if (extraTotalB > 0 && extraPassB < extraTotalB) {
|
|
84
|
+
promote_ok = false;
|
|
85
|
+
verdict = 'extra-B invariant regression: ' + extraPassB + '/' + extraTotalB
|
|
86
|
+
+ ' B-only invariants passed';
|
|
87
|
+
}
|
|
88
|
+
else if (token_ratio > (scenario.ab_test?.token_ratio_cap ?? DEFAULT_TOKEN_CAP)) {
|
|
89
|
+
promote_ok = false;
|
|
90
|
+
const cap = scenario.ab_test?.token_ratio_cap ?? DEFAULT_TOKEN_CAP;
|
|
91
|
+
verdict = 'token spend regressed: b/a=' + token_ratio.toFixed(2) + ' (cap ' + cap.toFixed(2) + ')';
|
|
92
|
+
}
|
|
93
|
+
else if (passB === passA && extraTotalB === 0) {
|
|
94
|
+
verdict = 'equivalent (a=' + passA + '/' + totA + ', b=' + passB + '/' + totB + ')';
|
|
95
|
+
}
|
|
96
|
+
else {
|
|
97
|
+
verdict = 'B improved (a=' + passA + '/' + totA + ', b=' + passB + '/' + totB + ')';
|
|
98
|
+
}
|
|
99
|
+
return {
|
|
100
|
+
scenario_id: scenario.id,
|
|
101
|
+
brain_id: brain.id,
|
|
102
|
+
a, b,
|
|
103
|
+
invariants_delta,
|
|
104
|
+
token_ratio,
|
|
105
|
+
promote_ok,
|
|
106
|
+
verdict,
|
|
107
|
+
};
|
|
108
|
+
}
|
|
109
|
+
export async function runABMatrix(scenarios, brains) {
|
|
110
|
+
const results = [];
|
|
111
|
+
for (const s of scenarios) {
|
|
112
|
+
if (!s.ab_test)
|
|
113
|
+
continue;
|
|
114
|
+
for (const b of brains) {
|
|
115
|
+
const r = await runAB(s, b);
|
|
116
|
+
if (r)
|
|
117
|
+
results.push(r);
|
|
118
|
+
}
|
|
119
|
+
}
|
|
120
|
+
return {
|
|
121
|
+
generated_at: new Date().toISOString(),
|
|
122
|
+
results,
|
|
123
|
+
promote_ok: results.every((r) => r.promote_ok),
|
|
124
|
+
};
|
|
125
|
+
}
|
|
126
|
+
export function renderABReport(matrix) {
|
|
127
|
+
const lines = [];
|
|
128
|
+
lines.push('# A/B promotion report');
|
|
129
|
+
lines.push('');
|
|
130
|
+
lines.push('Generated: ' + matrix.generated_at);
|
|
131
|
+
lines.push('Cells: ' + matrix.results.length);
|
|
132
|
+
lines.push('Promote: ' + (matrix.promote_ok ? 'OK -- all cells passed' : 'FAIL -- one or more cells regressed'));
|
|
133
|
+
lines.push('');
|
|
134
|
+
lines.push('| Scenario | Brain | Inv A | Inv B | Tok ratio | Verdict |');
|
|
135
|
+
lines.push('|---|---|---|---|---|---|');
|
|
136
|
+
for (const r of matrix.results) {
|
|
137
|
+
const aP = countPass(r.a.invariants) + '/' + r.a.invariants.length;
|
|
138
|
+
const bP = countPass(r.b.invariants) + '/' + r.b.invariants.length;
|
|
139
|
+
lines.push('| ' + r.scenario_id + ' | ' + r.brain_id + ' | ' + aP + ' | ' + bP
|
|
140
|
+
+ ' | ' + r.token_ratio.toFixed(2) + ' | ' + r.verdict + ' |');
|
|
141
|
+
}
|
|
142
|
+
return lines.join('\n');
|
|
143
|
+
}
|
|
144
|
+
//# sourceMappingURL=ab_runner.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"ab_runner.js","sourceRoot":"","sources":["../../../src/qa/brain_matrix/ab_runner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,OAAO,EACL,kBAAkB,GACnB,MAAM,aAAa,CAAC;AAuCrB,MAAM,iBAAiB,GAAG,IAAI,CAAC;AAE/B,SAAS,SAAS,CAAC,UAA6B;IAC9C,OAAO,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC;AACjD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,KAAK,CACzB,QAAwC,EACxC,KAAkB;IAElB,IAAI,CAAC,QAAQ,CAAC,OAAO;QAAE,OAAO,IAAI,CAAC;IACnC,MAAM,QAAQ,GAAI,QAA2D,CAAC,cAAc,IAAI,EAAE,CAAC;IAEnG,uEAAuE;IACvE,MAAM,KAAK,GAAmD;QAC5D,GAAG,QAAQ;QACX,cAAc,EAAE,QAAQ,CAAC,OAAO,CAAC,gBAAgB,IAAI,QAAQ;KAC9D,CAAC;IACF,MAAM,KAAK,GAAmD;QAC5D,GAAG,QAAQ;QACX,cAAc,EAAE,QAAQ,CAAC,OAAO,CAAC,gBAAgB,IAAI,QAAQ;QAC7D,UAAU,EAAE;YACV,GAAG,QAAQ,CAAC,UAAU;YACtB,GAAG,CAAC,QAAQ,CAAC,OAAO,CAAC,kBAAkB,IAAI,EAAE,CAAC;SAC/C;KACF,CAAC;IAEF,MAAM,CAAC,GAAG,MAAM,kBAAkB,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IACjD,MAAM,CAAC,GAAG,MAAM,kBAAkB,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IAEjD;;;iEAG6D;IAC7D,MAAM,WAAW,GAAG,QAAQ,CAAC,UAAU,CAAC,MAAM,CAAC;IAC/C,MAAM,WAAW,GAAG,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,WAAW,CAAC,CAAC,CAAC;IAClE,MAAM,WAAW,GAAG,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,WAAW,CAAC,CAAC,CAAC;IAClE,MAAM,gBAAgB,GAAG,CAAC,CAAC,UAAU,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;IACzD,MAAM,UAAU,GAAG,SAAS,CAAC,gBAAgB,CAAC,CAAC;IAC/C,MAAM,WAAW,GAAG,gBAAgB,CAAC,MAAM,CAAC;IAE5C,MAAM,KAAK,GAAG,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;IACtC,MAAM,KAAK,GAAG,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;IACtC,MAAM,IAAI,GAAI,CAAC,CAAC,UAAU,CAAC,MAAM,CAAC;IAClC,MAAM,IAAI,GAAI,CAAC,CAAC,UAAU,CAAC,MAAM,CAAC;IAElC,MAAM,gBAAgB,GAAG,KAAK,GAAG,KAAK,CAAC;IACvC,mEAAmE;IACnE,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,gBAAgB,GAAG,CAAC,CAAC,GAAG,CAAC,eAAe,CAAC;IAC1D,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,gBAAgB,GAAG,CAAC,CAAC,GAAG,CAAC,eAAe,CAAC;IAC1D,MAAM,WAAW,GAAG,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IAEzC,IAAI,UAAU,GAAG,IAAI,CAAC;IACtB,IAAI,OAAO,GAAG,uBAAuB,CAAC;IACtC,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,IAAI,CAAC,CAAC,GAAG,CAAC,KAAK,EAAE,CAAC;QAC/B,UAAU,GAAG,KAAK,CAAC;QACnB,OAAO,GAAG,eAAe,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,CAAC,GAAG,OAAO,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,CAAC,CAAC;IACtF,CAAC;SAAM,IAAI,WAAW,GAAG,WAAW,EAAE,CAAC;QACrC,UAAU,GAAG,KAAK,CAAC;QACnB,OAAO,GAAG,iCAAiC,GAAG,WAAW,GAAG,GAAG,GAAG,WAAW;cACnE,MAAM,GAAG,WAAW,GAAG,GAAG,GAAG,WAAW,CAAC;IACrD,CAAC;SAAM,IAAI,WAAW,GAAG,CAAC,IAAI,UAAU,GAAG,WAAW,EAAE,CAAC;QACvD,UAAU,GAAG,KAAK,CAAC;QACnB,OAAO,GAAG,gCAAgC,GAAG,UAAU,GAAG,GAAG,GAAG,WAAW;cACjE,2BAA2B,CAAC;IACxC,CAAC;SAAM,IAAI,WAAW,GAAG,CAAC,QAAQ,CAAC,OAAO,EAAE,eAAe,IAAI,iBAAiB,CAAC,EAAE,CAAC;QAClF,UAAU,GAAG,KAAK,CAAC;QACnB,MAAM,GAAG,GAAG,QAAQ,CAAC,OAAO,EAAE,eAAe,IAAI,iBAAiB,CAAC;QACnE,OAAO,GAAG,6BAA6B,GAAG,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,QAAQ,GAAG,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC;IACrG,CAAC;SAAM,IAAI,KAAK,KAAK,KAAK,IAAI,WAAW,KAAK,CAAC,EAAE,CAAC;QAChD,OAAO,GAAG,gBAAgB,GAAG,KAAK,GAAG,GAAG,GAAG,IAAI,GAAG,MAAM,GAAG,KAAK,GAAG,GAAG,GAAG,IAAI,GAAG,GAAG,CAAC;IACtF,CAAC;SAAM,CAAC;QACN,OAAO,GAAG,gBAAgB,GAAG,KAAK,GAAG,GAAG,GAAG,IAAI,GAAG,MAAM,GAAG,KAAK,GAAG,GAAG,GAAG,IAAI,GAAG,GAAG,CAAC;IACtF,CAAC;IAED,OAAO;QACL,WAAW,EAAE,QAAQ,CAAC,EAAE;QACxB,QAAQ,EAAE,KAAK,CAAC,EAAE;QAClB,CAAC,EAAE,CAAC;QACJ,gBAAgB;QAChB,WAAW;QACX,UAAU;QACV,OAAO;KACR,CAAC;AACJ,CAAC;AASD,MAAM,CAAC,KAAK,UAAU,WAAW,CAC/B,SAAgD,EAChD,MAAqB;IAErB,MAAM,OAAO,GAAe,EAAE,CAAC;IAC/B,KAAK,MAAM,CAAC,IAAI,SAAS,EAAE,CAAC;QAC1B,IAAI,CAAC,CAAC,CAAC,OAAO;YAAE,SAAS;QACzB,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;YACvB,MAAM,CAAC,GAAG,MAAM,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAC5B,IAAI,CAAC;gBAAE,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACzB,CAAC;IACH,CAAC;IACD,OAAO;QACL,YAAY,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;QACtC,OAAO;QACP,UAAU,EAAE,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC;KAC/C,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,cAAc,CAAC,MAAsB;IACnD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;IACrC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,aAAa,GAAG,MAAM,CAAC,YAAY,CAAC,CAAC;IAChD,KAAK,CAAC,IAAI,CAAC,aAAa,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;IAClD,KAAK,CAAC,IAAI,CAAC,aAAa,GAAG,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,wBAAwB,CAAC,CAAC,CAAC,qCAAqC,CAAC,CAAC,CAAC;IACnH,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,4DAA4D,CAAC,CAAC;IACzE,KAAK,CAAC,IAAI,CAAC,2BAA2B,CAAC,CAAC;IACxC,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;QAC/B,MAAM,EAAE,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,GAAG,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,MAAM,CAAC;QACnE,MAAM,EAAE,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,GAAG,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,MAAM,CAAC;QACnE,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,WAAW,GAAG,KAAK,GAAG,CAAC,CAAC,QAAQ,GAAG,KAAK,GAAG,EAAE,GAAG,KAAK,GAAG,EAAE;cAC1E,KAAK,GAAG,CAAC,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC,CAAC;IACnE,CAAC;IACD,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC"}
|
|
@@ -0,0 +1,80 @@
|
|
|
1
|
+
{
|
|
2
|
+
"_doc": "Brain matrix QA -- providers/models to test against. ASCII only. Edit to add/remove brains. The harness skips brains whose vault_slot is empty (no API key configured). `fake` is the dry-run brain used in CI; it never makes network calls.",
|
|
3
|
+
"brains": [
|
|
4
|
+
{
|
|
5
|
+
"id": "fake-scripted",
|
|
6
|
+
"provider": "fake",
|
|
7
|
+
"model": "scripted",
|
|
8
|
+
"tier": "eco",
|
|
9
|
+
"vault_slot": "",
|
|
10
|
+
"notes": "Dry-run fake brain. Tests the harness itself without spending money. Reads scripted responses from the scenario fixture's `fake_responses` field."
|
|
11
|
+
},
|
|
12
|
+
{
|
|
13
|
+
"id": "gemini-flash-lite",
|
|
14
|
+
"provider": "google_ai",
|
|
15
|
+
"model": "gemini-2.5-flash-lite",
|
|
16
|
+
"tier": "eco",
|
|
17
|
+
"vault_slot": "GEMINI_API_KEY",
|
|
18
|
+
"notes": "Eco. ~$0.10/1M in, ~$0.40/1M out. Fast, smallest context handling."
|
|
19
|
+
},
|
|
20
|
+
{
|
|
21
|
+
"id": "haiku-4-5",
|
|
22
|
+
"provider": "anthropic",
|
|
23
|
+
"model": "claude-haiku-4-5-20251001",
|
|
24
|
+
"tier": "eco",
|
|
25
|
+
"vault_slot": "CLAUDE_API_KEY",
|
|
26
|
+
"notes": "Eco. ~$1/1M in, ~$5/1M out. Anthropic small model."
|
|
27
|
+
},
|
|
28
|
+
{
|
|
29
|
+
"id": "sonnet-4-6",
|
|
30
|
+
"provider": "anthropic",
|
|
31
|
+
"model": "claude-sonnet-4-6",
|
|
32
|
+
"tier": "frontier",
|
|
33
|
+
"vault_slot": "CLAUDE_API_KEY",
|
|
34
|
+
"notes": "Frontier mid. ~$3/1M in, ~$15/1M out. Default Forge brain."
|
|
35
|
+
},
|
|
36
|
+
{
|
|
37
|
+
"id": "opus-4-7",
|
|
38
|
+
"provider": "anthropic",
|
|
39
|
+
"model": "claude-opus-4-7",
|
|
40
|
+
"tier": "frontier",
|
|
41
|
+
"vault_slot": "CLAUDE_API_KEY",
|
|
42
|
+
"notes": "Frontier top. ~$15/1M in, ~$75/1M out. Skip default to save $$.",
|
|
43
|
+
"skip_default": true
|
|
44
|
+
},
|
|
45
|
+
{
|
|
46
|
+
"id": "opus-4-8",
|
|
47
|
+
"provider": "anthropic",
|
|
48
|
+
"model": "claude-opus-4-8",
|
|
49
|
+
"tier": "frontier",
|
|
50
|
+
"vault_slot": "CLAUDE_API_KEY",
|
|
51
|
+
"notes": "Frontier top, released after 4-7. Incorporates effort indicator (low/medium/high/extra/max) and complex reasoning. Skip default to save $$ (manual invoke via --brains opus-4-8).",
|
|
52
|
+
"skip_default": true
|
|
53
|
+
},
|
|
54
|
+
{
|
|
55
|
+
"id": "deepseek-v3",
|
|
56
|
+
"provider": "deepseek",
|
|
57
|
+
"model": "deepseek-chat",
|
|
58
|
+
"tier": "eco",
|
|
59
|
+
"vault_slot": "DEEPSEEK_API_KEY",
|
|
60
|
+
"notes": "Eco. Cheap multilingual."
|
|
61
|
+
},
|
|
62
|
+
{
|
|
63
|
+
"id": "gpt-4o-mini",
|
|
64
|
+
"provider": "openai",
|
|
65
|
+
"model": "gpt-4o-mini",
|
|
66
|
+
"tier": "eco",
|
|
67
|
+
"vault_slot": "OPENAI_API_KEY",
|
|
68
|
+
"notes": "Eco. ~$0.15/1M in, ~$0.60/1M out."
|
|
69
|
+
},
|
|
70
|
+
{
|
|
71
|
+
"id": "gpt-4o",
|
|
72
|
+
"provider": "openai",
|
|
73
|
+
"model": "gpt-4o",
|
|
74
|
+
"tier": "frontier",
|
|
75
|
+
"vault_slot": "OPENAI_API_KEY",
|
|
76
|
+
"notes": "Frontier alt. Skip default to save $$.",
|
|
77
|
+
"skip_default": true
|
|
78
|
+
}
|
|
79
|
+
]
|
|
80
|
+
}
|
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
import type { MatrixReport } from './types.js';
|
|
2
|
+
export type Outcome = 'PASS' | 'FAIL' | 'ERR' | 'MISSING';
|
|
3
|
+
export interface CompareCell {
|
|
4
|
+
scenario: string;
|
|
5
|
+
brain: string;
|
|
6
|
+
base: Outcome;
|
|
7
|
+
head: Outcome;
|
|
8
|
+
/** 'regression' (PASS->FAIL/ERR), 'improvement' (FAIL/ERR->PASS),
|
|
9
|
+
* 'unchanged' (same outcome), 'flake' (FAIL<->ERR), 'new'
|
|
10
|
+
* (only present in head), 'removed' (only in base). */
|
|
11
|
+
kind: 'regression' | 'improvement' | 'unchanged' | 'flake' | 'new' | 'removed';
|
|
12
|
+
}
|
|
13
|
+
export interface CompareSummary {
|
|
14
|
+
base_version: string;
|
|
15
|
+
head_version: string;
|
|
16
|
+
cells: CompareCell[];
|
|
17
|
+
regressions: CompareCell[];
|
|
18
|
+
improvements: CompareCell[];
|
|
19
|
+
flakes: CompareCell[];
|
|
20
|
+
new_cells: CompareCell[];
|
|
21
|
+
removed_cells: CompareCell[];
|
|
22
|
+
unchanged_count: number;
|
|
23
|
+
}
|
|
24
|
+
export declare function compareMatrices(base: MatrixReport, head: MatrixReport): CompareSummary;
|
|
25
|
+
export declare function loadMatrixReport(p: string): Promise<MatrixReport>;
|
|
26
|
+
export declare function renderCompareSummary(c: CompareSummary): string;
|
|
27
|
+
//# sourceMappingURL=compare.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"compare.d.ts","sourceRoot":"","sources":["../../../src/qa/brain_matrix/compare.ts"],"names":[],"mappings":"AASA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAE/C,MAAM,MAAM,OAAO,GAAG,MAAM,GAAG,MAAM,GAAG,KAAK,GAAG,SAAS,CAAC;AAE1D,MAAM,WAAW,WAAW;IAC1B,QAAQ,EAAK,MAAM,CAAC;IACpB,KAAK,EAAQ,MAAM,CAAC;IACpB,IAAI,EAAS,OAAO,CAAC;IACrB,IAAI,EAAS,OAAO,CAAC;IACrB;;4DAEwD;IACxD,IAAI,EAAS,YAAY,GAAG,aAAa,GAAG,WAAW,GAAG,OAAO,GAAG,KAAK,GAAG,SAAS,CAAC;CACvF;AAED,MAAM,WAAW,cAAc;IAC7B,YAAY,EAAK,MAAM,CAAC;IACxB,YAAY,EAAK,MAAM,CAAC;IACxB,KAAK,EAAY,WAAW,EAAE,CAAC;IAC/B,WAAW,EAAM,WAAW,EAAE,CAAC;IAC/B,YAAY,EAAK,WAAW,EAAE,CAAC;IAC/B,MAAM,EAAW,WAAW,EAAE,CAAC;IAC/B,SAAS,EAAQ,WAAW,EAAE,CAAC;IAC/B,aAAa,EAAI,WAAW,EAAE,CAAC;IAC/B,eAAe,EAAE,MAAM,CAAC;CACzB;AAUD,wBAAgB,eAAe,CAAC,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE,YAAY,GAAG,cAAc,CAmCtF;AAED,wBAAsB,gBAAgB,CAAC,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC,CAGvE;AAED,wBAAgB,oBAAoB,CAAC,CAAC,EAAE,cAAc,GAAG,MAAM,CAuD9D"}
|
|
@@ -0,0 +1,122 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Brain matrix --compare: diff two MatrixReport JSONs and
|
|
3
|
+
* surface per-cell regressions / improvements.
|
|
4
|
+
*
|
|
5
|
+
* Used by `yf qa brain-matrix --compare base.json head.json` so
|
|
6
|
+
* Pablo can answer "did z.62 regress anything against z.55?"
|
|
7
|
+
* without scanning two 2000-line files by hand.
|
|
8
|
+
*/
|
|
9
|
+
import { readFile } from 'node:fs/promises';
|
|
10
|
+
function cellOutcome(matrix, scenarioId, brainId) {
|
|
11
|
+
const cell = matrix.results?.[scenarioId]?.[brainId];
|
|
12
|
+
if (!cell)
|
|
13
|
+
return 'MISSING';
|
|
14
|
+
if (cell.run?.error)
|
|
15
|
+
return 'ERR';
|
|
16
|
+
if (cell.pass)
|
|
17
|
+
return 'PASS';
|
|
18
|
+
return 'FAIL';
|
|
19
|
+
}
|
|
20
|
+
export function compareMatrices(base, head) {
|
|
21
|
+
const scenarios = new Set([
|
|
22
|
+
...base.scenarios.map((s) => s.id),
|
|
23
|
+
...head.scenarios.map((s) => s.id),
|
|
24
|
+
]);
|
|
25
|
+
const brains = new Set([
|
|
26
|
+
...base.brains.map((b) => b.id),
|
|
27
|
+
...head.brains.map((b) => b.id),
|
|
28
|
+
]);
|
|
29
|
+
const cells = [];
|
|
30
|
+
for (const s of [...scenarios].sort()) {
|
|
31
|
+
for (const b of [...brains].sort()) {
|
|
32
|
+
const baseOut = cellOutcome(base, s, b);
|
|
33
|
+
const headOut = cellOutcome(head, s, b);
|
|
34
|
+
let kind;
|
|
35
|
+
if (baseOut === 'MISSING' && headOut !== 'MISSING')
|
|
36
|
+
kind = 'new';
|
|
37
|
+
else if (baseOut !== 'MISSING' && headOut === 'MISSING')
|
|
38
|
+
kind = 'removed';
|
|
39
|
+
else if (baseOut === headOut)
|
|
40
|
+
kind = 'unchanged';
|
|
41
|
+
else if (baseOut === 'PASS' && (headOut === 'FAIL' || headOut === 'ERR'))
|
|
42
|
+
kind = 'regression';
|
|
43
|
+
else if ((baseOut === 'FAIL' || baseOut === 'ERR') && headOut === 'PASS')
|
|
44
|
+
kind = 'improvement';
|
|
45
|
+
else
|
|
46
|
+
kind = 'flake';
|
|
47
|
+
cells.push({ scenario: s, brain: b, base: baseOut, head: headOut, kind });
|
|
48
|
+
}
|
|
49
|
+
}
|
|
50
|
+
return {
|
|
51
|
+
base_version: base.forge_version,
|
|
52
|
+
head_version: head.forge_version,
|
|
53
|
+
cells,
|
|
54
|
+
regressions: cells.filter((c) => c.kind === 'regression'),
|
|
55
|
+
improvements: cells.filter((c) => c.kind === 'improvement'),
|
|
56
|
+
flakes: cells.filter((c) => c.kind === 'flake'),
|
|
57
|
+
new_cells: cells.filter((c) => c.kind === 'new'),
|
|
58
|
+
removed_cells: cells.filter((c) => c.kind === 'removed'),
|
|
59
|
+
unchanged_count: cells.filter((c) => c.kind === 'unchanged').length,
|
|
60
|
+
};
|
|
61
|
+
}
|
|
62
|
+
export async function loadMatrixReport(p) {
|
|
63
|
+
const raw = await readFile(p, 'utf-8');
|
|
64
|
+
return JSON.parse(raw);
|
|
65
|
+
}
|
|
66
|
+
export function renderCompareSummary(c) {
|
|
67
|
+
const lines = [];
|
|
68
|
+
lines.push('Brain matrix compare');
|
|
69
|
+
lines.push(' base: ' + c.base_version);
|
|
70
|
+
lines.push(' head: ' + c.head_version);
|
|
71
|
+
lines.push('');
|
|
72
|
+
lines.push('Summary');
|
|
73
|
+
lines.push(' regressions: ' + c.regressions.length);
|
|
74
|
+
lines.push(' improvements: ' + c.improvements.length);
|
|
75
|
+
lines.push(' flakes: ' + c.flakes.length);
|
|
76
|
+
lines.push(' new cells: ' + c.new_cells.length);
|
|
77
|
+
lines.push(' removed: ' + c.removed_cells.length);
|
|
78
|
+
lines.push(' unchanged: ' + c.unchanged_count);
|
|
79
|
+
lines.push('');
|
|
80
|
+
if (c.regressions.length > 0) {
|
|
81
|
+
lines.push('=== Regressions ===');
|
|
82
|
+
for (const r of c.regressions) {
|
|
83
|
+
lines.push(' ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
|
|
84
|
+
+ ' ' + r.base + ' -> ' + r.head);
|
|
85
|
+
}
|
|
86
|
+
lines.push('');
|
|
87
|
+
}
|
|
88
|
+
if (c.improvements.length > 0) {
|
|
89
|
+
lines.push('=== Improvements ===');
|
|
90
|
+
for (const r of c.improvements) {
|
|
91
|
+
lines.push(' ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
|
|
92
|
+
+ ' ' + r.base + ' -> ' + r.head);
|
|
93
|
+
}
|
|
94
|
+
lines.push('');
|
|
95
|
+
}
|
|
96
|
+
if (c.flakes.length > 0) {
|
|
97
|
+
lines.push('=== Flakes (FAIL<->ERR) ===');
|
|
98
|
+
for (const r of c.flakes) {
|
|
99
|
+
lines.push(' ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
|
|
100
|
+
+ ' ' + r.base + ' -> ' + r.head);
|
|
101
|
+
}
|
|
102
|
+
lines.push('');
|
|
103
|
+
}
|
|
104
|
+
if (c.new_cells.length > 0) {
|
|
105
|
+
lines.push('=== New cells (only in head) ===');
|
|
106
|
+
for (const r of c.new_cells) {
|
|
107
|
+
lines.push(' ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
|
|
108
|
+
+ ' ' + r.head);
|
|
109
|
+
}
|
|
110
|
+
lines.push('');
|
|
111
|
+
}
|
|
112
|
+
if (c.removed_cells.length > 0) {
|
|
113
|
+
lines.push('=== Removed cells (only in base) ===');
|
|
114
|
+
for (const r of c.removed_cells) {
|
|
115
|
+
lines.push(' ' + r.scenario.padEnd(40, ' ') + ' @ ' + r.brain.padEnd(18, ' ')
|
|
116
|
+
+ ' ' + r.base);
|
|
117
|
+
}
|
|
118
|
+
lines.push('');
|
|
119
|
+
}
|
|
120
|
+
return lines.join('\n');
|
|
121
|
+
}
|
|
122
|
+
//# sourceMappingURL=compare.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"compare.js","sourceRoot":"","sources":["../../../src/qa/brain_matrix/compare.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AA4B5C,SAAS,WAAW,CAAC,MAAoB,EAAE,UAAkB,EAAE,OAAe;IAC5E,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,EAAE,CAAC,UAAU,CAAC,EAAE,CAAC,OAAO,CAAC,CAAC;IACrD,IAAI,CAAC,IAAI;QAAE,OAAO,SAAS,CAAC;IAC5B,IAAI,IAAI,CAAC,GAAG,EAAE,KAAK;QAAE,OAAO,KAAK,CAAC;IAClC,IAAI,IAAI,CAAC,IAAI;QAAE,OAAO,MAAM,CAAC;IAC7B,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,MAAM,UAAU,eAAe,CAAC,IAAkB,EAAE,IAAkB;IACpE,MAAM,SAAS,GAAG,IAAI,GAAG,CAAS;QAChC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAClC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACnC,CAAC,CAAC;IACH,MAAM,MAAM,GAAG,IAAI,GAAG,CAAS;QAC7B,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC/B,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAChC,CAAC,CAAC;IACH,MAAM,KAAK,GAAkB,EAAE,CAAC;IAChC,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QACtC,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;YACnC,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;YACxC,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;YACxC,IAAI,IAAyB,CAAC;YAC9B,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,SAAS;gBAAE,IAAI,GAAG,KAAK,CAAC;iBAC5D,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,SAAS;gBAAE,IAAI,GAAG,SAAS,CAAC;iBACrE,IAAI,OAAO,KAAK,OAAO;gBAAE,IAAI,GAAG,WAAW,CAAC;iBAC5C,IAAI,OAAO,KAAK,MAAM,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,OAAO,KAAK,KAAK,CAAC;gBAAE,IAAI,GAAG,YAAY,CAAC;iBACzF,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,OAAO,KAAK,KAAK,CAAC,IAAI,OAAO,KAAK,MAAM;gBAAE,IAAI,GAAG,aAAa,CAAC;;gBAC1F,IAAI,GAAG,OAAO,CAAC;YACpB,KAAK,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC;QAC5E,CAAC;IACH,CAAC;IACD,OAAO;QACL,YAAY,EAAK,IAAI,CAAC,aAAa;QACnC,YAAY,EAAK,IAAI,CAAC,aAAa;QACnC,KAAK;QACL,WAAW,EAAM,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,YAAY,CAAC;QAC7D,YAAY,EAAK,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,aAAa,CAAC;QAC9D,MAAM,EAAW,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,OAAO,CAAC;QACxD,SAAS,EAAQ,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,KAAK,CAAC;QACtD,aAAa,EAAI,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,CAAC;QAC1D,eAAe,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC,MAAM;KACpE,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,CAAS;IAC9C,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,CAAC,EAAE,OAAO,CAAC,CAAC;IACvC,OAAO,IAAI,CAAC,KAAK,CAAC,GAAG,CAAiB,CAAC;AACzC,CAAC;AAED,MAAM,UAAU,oBAAoB,CAAC,CAAiB;IACpD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;IACnC,KAAK,CAAC,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,YAAY,CAAC,CAAC;IACxC,KAAK,CAAC,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,YAAY,CAAC,CAAC;IACxC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;IACtB,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;IACtD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,YAAY,CAAC,MAAM,CAAC,CAAC;IACvD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IACjD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC;IACpD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;IACxD,KAAK,CAAC,IAAI,CAAC,kBAAkB,GAAG,CAAC,CAAC,eAAe,CAAC,CAAC;IACnD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,IAAI,CAAC,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,qBAAqB,CAAC,CAAC;QAClC,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,CAAC;YAC9B,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,GAAG,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC9B,KAAK,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;QACnC,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,YAAY,EAAE,CAAC;YAC/B,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,GAAG,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACxB,KAAK,CAAC,IAAI,CAAC,6BAA6B,CAAC,CAAC;QAC1C,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,CAAC;YACzB,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,GAAG,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC3B,KAAK,CAAC,IAAI,CAAC,kCAAkC,CAAC,CAAC;QAC/C,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,CAAC;YAC5B,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACrB,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,IAAI,CAAC,CAAC,aAAa,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC/B,KAAK,CAAC,IAAI,CAAC,sCAAsC,CAAC,CAAC;QACnD,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,aAAa,EAAE,CAAC;YAChC,KAAK,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,CAAC;kBAC1E,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;QACrB,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACjB,CAAC;IACD,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC"}
|
|
@@ -0,0 +1,36 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Fake brain for dry-run testing of the brain matrix harness.
|
|
3
|
+
*
|
|
4
|
+
* Each scenario can declare a `fake_responses` field in setup:
|
|
5
|
+
*
|
|
6
|
+
* fake_responses: [
|
|
7
|
+
* { text: "voy a abrir la pestana", tool_calls: [{name: "nac3_invoke",
|
|
8
|
+
* args: { verb_id: "yujin.pizarron.create-tab", args: {title:"X", kind:"code", content:"..."} }}] },
|
|
9
|
+
* { text: "listo, te abri la pestana" }
|
|
10
|
+
* ]
|
|
11
|
+
*
|
|
12
|
+
* The fake brain returns one entry per LLM turn, in order. If the
|
|
13
|
+
* harness asks for more turns than declared, returns "" + no tool
|
|
14
|
+
* calls (the scenario likely has a bug -- the report flags it).
|
|
15
|
+
*
|
|
16
|
+
* Used in:
|
|
17
|
+
* - vitest unit tests of the harness
|
|
18
|
+
* - CI dry-run that validates the harness itself before any
|
|
19
|
+
* real-LLM run is attempted.
|
|
20
|
+
*
|
|
21
|
+
* ASCII-only.
|
|
22
|
+
*/
|
|
23
|
+
import type { UnifiedChatClient } from '../../llm/multi_provider_client.js';
|
|
24
|
+
import type { BrainSelection } from '../../core/cost_router.js';
|
|
25
|
+
export interface FakeResponse {
|
|
26
|
+
text: string;
|
|
27
|
+
tool_calls?: Array<{
|
|
28
|
+
name: string;
|
|
29
|
+
args: Record<string, unknown>;
|
|
30
|
+
}>;
|
|
31
|
+
/** Optional: simulate token spend for reports. */
|
|
32
|
+
tokens_in?: number;
|
|
33
|
+
tokens_out?: number;
|
|
34
|
+
}
|
|
35
|
+
export declare function makeFakeBrain(brain: BrainSelection, responses: FakeResponse[]): UnifiedChatClient;
|
|
36
|
+
//# sourceMappingURL=fake_brain.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"fake_brain.d.ts","sourceRoot":"","sources":["../../../src/qa/brain_matrix/fake_brain.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,OAAO,KAAK,EAAE,iBAAiB,EAA2C,MAAM,oCAAoC,CAAC;AACrH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,2BAA2B,CAAC;AAGhE,MAAM,WAAW,YAAY;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,CAAC,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,CAAC,CAAC;IACpE,kDAAkD;IAClD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,wBAAgB,aAAa,CAAC,KAAK,EAAE,cAAc,EAAE,SAAS,EAAE,YAAY,EAAE,GAAG,iBAAiB,CAyBjG"}
|
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
export function makeFakeBrain(brain, responses) {
|
|
2
|
+
let idx = 0;
|
|
3
|
+
return {
|
|
4
|
+
async chat(_req) {
|
|
5
|
+
const r = responses[idx] || { text: '', tool_calls: [] };
|
|
6
|
+
idx++;
|
|
7
|
+
/* Build a synthetic ToolRound[]. The real ToolRound shape
|
|
8
|
+
* carries call + result; for the fake brain we leave result
|
|
9
|
+
* as `null` so the runner knows to mock it. */
|
|
10
|
+
const toolRounds = (r.tool_calls || []).map((c) => ({
|
|
11
|
+
tool: c.name,
|
|
12
|
+
input: c.args,
|
|
13
|
+
output: { ok: true, fake: true },
|
|
14
|
+
is_error: false,
|
|
15
|
+
}));
|
|
16
|
+
return {
|
|
17
|
+
text: r.text,
|
|
18
|
+
tokensIn: r.tokens_in ?? 100,
|
|
19
|
+
tokensOut: r.tokens_out ?? 50,
|
|
20
|
+
model: 'fake-scripted',
|
|
21
|
+
toolRounds,
|
|
22
|
+
brain,
|
|
23
|
+
};
|
|
24
|
+
},
|
|
25
|
+
};
|
|
26
|
+
}
|
|
27
|
+
//# sourceMappingURL=fake_brain.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"fake_brain.js","sourceRoot":"","sources":["../../../src/qa/brain_matrix/fake_brain.ts"],"names":[],"mappings":"AAkCA,MAAM,UAAU,aAAa,CAAC,KAAqB,EAAE,SAAyB;IAC5E,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,OAAO;QACL,KAAK,CAAC,IAAI,CAAC,IAAwB;YACjC,MAAM,CAAC,GAAG,SAAS,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC;YACzD,GAAG,EAAE,CAAC;YACN;;4DAEgD;YAChD,MAAM,UAAU,GAAgB,CAAC,CAAC,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBAC/D,IAAI,EAAE,CAAC,CAAC,IAAI;gBACZ,KAAK,EAAE,CAAC,CAAC,IAAI;gBACb,MAAM,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE;gBAChC,QAAQ,EAAE,KAAK;aACS,CAAA,CAAC,CAAC;YAC5B,OAAO;gBACL,IAAI,EAAQ,CAAC,CAAC,IAAI;gBAClB,QAAQ,EAAI,CAAC,CAAC,SAAS,IAAI,GAAG;gBAC9B,SAAS,EAAG,CAAC,CAAC,UAAU,IAAI,EAAE;gBAC9B,KAAK,EAAO,eAAe;gBAC3B,UAAU;gBACV,KAAK;aACN,CAAC;QACJ,CAAC;KACF,CAAC;AACJ,CAAC"}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "a11y_announce_intent",
|
|
3
|
+
"description": "F107 a11y -- when the user asks the assistant to announce a status to screen-reader users, the brain should pick yujin.a11y.announce, NOT a generic 'show text' verb.",
|
|
4
|
+
"setup": {
|
|
5
|
+
"extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.announce [u d] args:{text<required>, politeness?, speak?, locale?}\nUse this for status / error / async completion notifications meant for screen readers + the voice loop. Pizarron is for visual data."
|
|
6
|
+
},
|
|
7
|
+
"turns": [
|
|
8
|
+
{ "kind": "user_says", "text": "Anunciale al lector de pantalla que el deploy termino con exito" }
|
|
9
|
+
],
|
|
10
|
+
"invariants": [
|
|
11
|
+
{ "name": "uses_a11y_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } },
|
|
12
|
+
{ "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} }
|
|
13
|
+
],
|
|
14
|
+
"fake_responses": [
|
|
15
|
+
{
|
|
16
|
+
"text": "Anuncio el deploy.",
|
|
17
|
+
"tool_calls": [
|
|
18
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Deploy completado con exito", "politeness": "polite", "speak": false } } }
|
|
19
|
+
]
|
|
20
|
+
}
|
|
21
|
+
]
|
|
22
|
+
}
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "a11y_announce_no_speak_default",
|
|
3
|
+
"description": "F107 a11y.announce -- by default speak:false. Brain must NOT silently flip speak:true unless the user asks for voice. args_contain_key validates explicit speak passing only when user mentions voice.",
|
|
4
|
+
"setup": {
|
|
5
|
+
"extra_system_prompt": "yujin.a11y.announce -- default speak=false. Only set true when the user explicitly mentions voice or TTS in this turn."
|
|
6
|
+
},
|
|
7
|
+
"turns": [
|
|
8
|
+
{ "kind": "user_says", "text": "Avisa al lector que la pagina termino de cargar (no necesito voz)" }
|
|
9
|
+
],
|
|
10
|
+
"invariants": [
|
|
11
|
+
{ "name": "uses_a11y_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } },
|
|
12
|
+
{ "name": "args_have_text", "kind": "args_contain_key", "params": { "verb_id": "yujin.a11y.announce", "key": "text" } },
|
|
13
|
+
{ "name": "no_voice_phrase", "kind": "regex_not_match", "params": { "pattern": "speak\\s*[:=]\\s*true|voice|TTS|por\\s+voz" } }
|
|
14
|
+
],
|
|
15
|
+
"fake_responses": [
|
|
16
|
+
{
|
|
17
|
+
"text": "Anuncio.",
|
|
18
|
+
"tool_calls": [
|
|
19
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Pagina cargada", "politeness": "polite" } } }
|
|
20
|
+
]
|
|
21
|
+
}
|
|
22
|
+
]
|
|
23
|
+
}
|
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "a11y_consecutive_announces",
|
|
3
|
+
"description": "Two-step UX operation: user creates lead + the brain announces success. Should be ONE announce call (not two) + concise reply. min_tool_calls catches missing announce; max_tool_calls catches duplicate.",
|
|
4
|
+
"setup": {
|
|
5
|
+
"extra_system_prompt": "BACKEND TOOLS active:\n- app.create_lead [A x] args:{name<required>, email?}\n\nA11Y VERBS active:\n- yujin.a11y.announce [u d] args:{text<required>, politeness?, speak?}\n\nAfter a destructive/approval-required action completes, fire ONE a11y.announce with the result. Never two."
|
|
6
|
+
},
|
|
7
|
+
"turns": [
|
|
8
|
+
{ "kind": "user_says", "text": "Crea lead Ana Lopez + anuncia al lector que se creo" },
|
|
9
|
+
{
|
|
10
|
+
"kind": "tool_result",
|
|
11
|
+
"tool_name": "app.create_lead",
|
|
12
|
+
"is_error": false,
|
|
13
|
+
"result": { "ok": true, "id": 47812, "name": "Ana Lopez" }
|
|
14
|
+
}
|
|
15
|
+
],
|
|
16
|
+
"invariants": [
|
|
17
|
+
{ "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} },
|
|
18
|
+
{ "name": "calls_create_lead", "kind": "verb_invoked", "params": { "verb_id": "app.create_lead" } },
|
|
19
|
+
{ "name": "calls_announce", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.announce" } },
|
|
20
|
+
{ "name": "max_three_calls", "kind": "max_tool_calls", "params": { "max": 3 } }
|
|
21
|
+
],
|
|
22
|
+
"fake_responses": [
|
|
23
|
+
{
|
|
24
|
+
"text": "Creo el lead.",
|
|
25
|
+
"tool_calls": [
|
|
26
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "app.create_lead", "args": { "name": "Ana Lopez" } } }
|
|
27
|
+
]
|
|
28
|
+
},
|
|
29
|
+
{
|
|
30
|
+
"text": "Anuncio.",
|
|
31
|
+
"tool_calls": [
|
|
32
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.announce", "args": { "text": "Lead Ana Lopez creado con id 47812", "politeness": "polite" } } }
|
|
33
|
+
]
|
|
34
|
+
},
|
|
35
|
+
{ "text": "Listo. Lead id 47812." }
|
|
36
|
+
]
|
|
37
|
+
}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "a11y_focus_after_action",
|
|
3
|
+
"description": "F107 a11y -- after a successful action that opened a new modal, the brain should move focus to the modal's primary element via yujin.a11y.focus.",
|
|
4
|
+
"setup": {
|
|
5
|
+
"extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.focus [u d] args:{slug<required>, scroll?, announce?}\nAfter opening a modal / new region, move focus to its first interactive child so keyboard + screen-reader users land in the right place."
|
|
6
|
+
},
|
|
7
|
+
"turns": [
|
|
8
|
+
{ "kind": "user_says", "text": "Abri el modal de confirmar deploy y dejame el foco en el boton 'Confirmar' (slug confirm-deploy-btn)" }
|
|
9
|
+
],
|
|
10
|
+
"invariants": [
|
|
11
|
+
{ "name": "uses_a11y_focus", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.focus" } },
|
|
12
|
+
{ "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} }
|
|
13
|
+
],
|
|
14
|
+
"fake_responses": [
|
|
15
|
+
{
|
|
16
|
+
"text": "Muevo el foco al boton de confirmar.",
|
|
17
|
+
"tool_calls": [
|
|
18
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.focus", "args": { "slug": "confirm-deploy-btn", "scroll": true, "announce": true } } }
|
|
19
|
+
]
|
|
20
|
+
}
|
|
21
|
+
]
|
|
22
|
+
}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "a11y_focus_after_open",
|
|
3
|
+
"description": "F107 closure -- after opening a modal/sidebar/section, the brain calls yujin.a11y.focus on the first interactive element of the new surface so keyboard-only + screen-reader users are not stranded at the trigger button. Locks the focus-management pattern.",
|
|
4
|
+
"setup": {
|
|
5
|
+
"extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.focus args:{slug<required>, scroll?, announce?}\n- yujin.a11y.announce args:{text<required>, politeness?, speak?}\nWhen you open a modal / new section, ALWAYS move focus to the first actionable inside (input, primary action, etc) so keyboard + screen-reader users land in the right place. The slug = data-nac-slug or data-nac-id of the target."
|
|
6
|
+
},
|
|
7
|
+
"turns": [
|
|
8
|
+
{ "kind": "user_says", "text": "Abri el modal de aprobacion y dejame escribir el motivo." }
|
|
9
|
+
],
|
|
10
|
+
"invariants": [
|
|
11
|
+
{ "name": "uses_a11y_focus", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.focus" } },
|
|
12
|
+
{ "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} }
|
|
13
|
+
],
|
|
14
|
+
"fake_responses": [
|
|
15
|
+
{
|
|
16
|
+
"text": "Modal abierto. Muevo foco al campo de motivo para que puedas escribir directo.",
|
|
17
|
+
"tool_calls": [
|
|
18
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.focus", "args": { "slug": "approval-modal.reason-input", "scroll": true, "announce": true } } }
|
|
19
|
+
]
|
|
20
|
+
}
|
|
21
|
+
]
|
|
22
|
+
}
|
|
@@ -0,0 +1,26 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "a11y_label_unlabeled_button",
|
|
3
|
+
"description": "F107 closure -- when the user reports that a button does not announce its purpose to a screen reader (icon-only button, missing aria-label), the brain calls yujin.a11y.label with a descriptive aria-label + optional shortcut hint. NOT a generic styling fix, NOT a write to the source code (the verb mutates the live DOM via aria attributes).",
|
|
4
|
+
"setup": {
|
|
5
|
+
"extra_system_prompt": "A11Y VERBS active:\n- yujin.a11y.label args:{slug<required>, description<required>, shortcut?, role?}\nIcon-only buttons need an aria-label so screen-reader users hear the purpose. Use a11y.label to attach the missing description; do NOT modify the source code (yujin.fs.write-project-file) for runtime ARIA fixes."
|
|
6
|
+
},
|
|
7
|
+
"turns": [
|
|
8
|
+
{ "kind": "user_says", "text": "El boton del icono de imprimir no anuncia nada al lector. Ponele label \"Imprimir reporte\"" }
|
|
9
|
+
],
|
|
10
|
+
"invariants": [
|
|
11
|
+
{ "name": "uses_a11y_label", "kind": "verb_invoked", "params": { "verb_id": "yujin.a11y.label" } },
|
|
12
|
+
{ "name": "no_source_edit", "kind": "verb_not_invoked", "params": { "verb_id": "yujin.fs.write-project-file" } },
|
|
13
|
+
{ "name": "no_invent_verbs", "kind": "no_invent_verb", "params": {} }
|
|
14
|
+
],
|
|
15
|
+
"fake_responses": [
|
|
16
|
+
{
|
|
17
|
+
"text": "Le pongo la label.",
|
|
18
|
+
"tool_calls": [
|
|
19
|
+
{ "name": "nac3_invoke", "args": { "verb_id": "yujin.a11y.label", "args": { "slug": "toolbar.print-button", "description": "Imprimir reporte" } } }
|
|
20
|
+
]
|
|
21
|
+
},
|
|
22
|
+
{
|
|
23
|
+
"text": "Listo, ahora el lector anuncia 'Imprimir reporte' al enfocar el boton."
|
|
24
|
+
}
|
|
25
|
+
]
|
|
26
|
+
}
|