create-walle 0.9.21 → 0.9.23
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 +27 -5
- package/package.json +2 -2
- package/template/CLAUDE.md +2 -2
- package/template/LICENSE +1 -1
- package/template/bin/ctm-dev-cleanup.js +24 -3
- package/template/bin/ctm-launch.sh +13 -0
- package/template/bin/dev.sh +156 -18
- package/template/bin/node-bin.sh +84 -0
- package/template/bin/pin-node.sh +51 -0
- package/template/claude-task-manager/api-prompts.js +1203 -182
- package/template/claude-task-manager/api-reviews.js +109 -15
- package/template/claude-task-manager/approval-agent.js +1360 -280
- package/template/claude-task-manager/bin/restart-ctm.sh +64 -23
- package/template/claude-task-manager/bin/storage-migration-supervisor.js +338 -0
- package/template/claude-task-manager/db.js +4417 -295
- package/template/claude-task-manager/docs/app-update-refresh-protocol.md +69 -0
- package/template/claude-task-manager/docs/approval-ai-refinement.md +138 -0
- package/template/claude-task-manager/docs/approval-rescue-loop.md +74 -0
- package/template/claude-task-manager/docs/codex-operational-warning-health.md +107 -0
- package/template/claude-task-manager/docs/codex-resume-state-guard-design.md +17 -12
- package/template/claude-task-manager/docs/codex-terminal-render-controller-handoff.md +311 -0
- package/template/claude-task-manager/docs/coding-agent-hooks-architecture.md +418 -0
- package/template/claude-task-manager/docs/conversation-import-freshness.md +20 -0
- package/template/claude-task-manager/docs/google-workspace-auth-health.md +77 -0
- package/template/claude-task-manager/docs/image-paste-ux.md +13 -0
- package/template/claude-task-manager/docs/ipad-web-preview.md +88 -0
- package/template/claude-task-manager/docs/main-loop-offload-architecture.md +66 -0
- package/template/claude-task-manager/docs/microsoft-dev-tunnel-phone-access-design.md +274 -519
- package/template/claude-task-manager/docs/mobile-live-streaming.md +27 -5
- package/template/claude-task-manager/docs/mobile-remote-submission-lifecycle.md +69 -0
- package/template/claude-task-manager/docs/phone-access-design.md +53 -15
- package/template/claude-task-manager/docs/phone-passkey-identity.md +122 -0
- package/template/claude-task-manager/docs/phone-setup.md +3 -0
- package/template/claude-task-manager/docs/prompt-editing-tree-design.md +25 -1
- package/template/claude-task-manager/docs/remote-desktop-access-design.md +268 -0
- package/template/claude-task-manager/docs/restart-lifecycle-architecture.md +95 -0
- package/template/claude-task-manager/docs/runtime-work-control-plane.md +53 -0
- package/template/claude-task-manager/docs/session-interactive-wait-surfaces.md +38 -0
- package/template/claude-task-manager/docs/session-needs-you-dismissal.md +84 -0
- package/template/claude-task-manager/docs/session-render-state-management-design.md +91 -3
- package/template/claude-task-manager/docs/session-standup-command-center-design.md +25 -1
- package/template/claude-task-manager/docs/session-title-authority.md +32 -0
- package/template/claude-task-manager/docs/session-workspace-binding.md +33 -0
- package/template/claude-task-manager/docs/skill-intent-resolution-design.md +72 -0
- package/template/claude-task-manager/docs/walle-mcp-supervisor-health.md +86 -0
- package/template/claude-task-manager/docs/walle-relay-phone-access-design.md +24 -15
- package/template/claude-task-manager/docs/walle-session-history-hydration.md +114 -0
- package/template/claude-task-manager/docs/walle-session-input-queue.md +104 -0
- package/template/claude-task-manager/docs/walle-session-model-catalog.md +90 -0
- package/template/claude-task-manager/docs/walle-session-model-preferences.md +15 -6
- package/template/claude-task-manager/git-utils.js +897 -27
- package/template/claude-task-manager/lib/agent-capabilities.js +33 -0
- package/template/claude-task-manager/lib/agent-cli-cache.js +37 -7
- package/template/claude-task-manager/lib/agent-hooks-installer.js +26 -2
- package/template/claude-task-manager/lib/agent-presets.js +17 -1
- package/template/claude-task-manager/lib/all-sessions-query.js +108 -0
- package/template/claude-task-manager/lib/approval-ai-refinement.js +488 -0
- package/template/claude-task-manager/lib/approval-self-adapt.js +168 -0
- package/template/claude-task-manager/lib/async-semaphore.js +44 -0
- package/template/claude-task-manager/lib/auth-context.js +5 -0
- package/template/claude-task-manager/lib/auth-rate-limit.js +47 -4
- package/template/claude-task-manager/lib/auth-rules.js +29 -2
- package/template/claude-task-manager/lib/auto-approval-verifier.js +129 -16
- package/template/claude-task-manager/lib/background-llm.js +144 -17
- package/template/claude-task-manager/lib/branch-inventory.js +212 -0
- package/template/claude-task-manager/lib/claude-desktop-sessions.js +15 -3
- package/template/claude-task-manager/lib/coalesce-sync-frames.js +151 -0
- package/template/claude-task-manager/lib/codex-launch-health.js +762 -0
- package/template/claude-task-manager/lib/codex-transcript-pager.js +51 -0
- package/template/claude-task-manager/lib/codex-zst.js +124 -0
- package/template/claude-task-manager/lib/coding-agent-models.js +233 -30
- package/template/claude-task-manager/lib/connection-health.js +232 -0
- package/template/claude-task-manager/lib/conversation-blob-parser.js +42 -0
- package/template/claude-task-manager/lib/conversation-tail-merge.js +89 -26
- package/template/claude-task-manager/lib/ctm-session-context-api.js +39 -10
- package/template/claude-task-manager/lib/cursor-conversation-store.js +354 -0
- package/template/claude-task-manager/lib/db-owner-worker-client.js +315 -0
- package/template/claude-task-manager/lib/document-review.js +141 -6
- package/template/claude-task-manager/lib/escalation-review.js +152 -0
- package/template/claude-task-manager/lib/graceful-shutdown.js +159 -0
- package/template/claude-task-manager/lib/headless-term-service.js +678 -0
- package/template/claude-task-manager/lib/heavy-worker-fallback.js +38 -0
- package/template/claude-task-manager/lib/jsonl-conversation-parser.js +542 -0
- package/template/claude-task-manager/lib/jsonl-range-reader.js +112 -0
- package/template/claude-task-manager/lib/main-db-census.js +216 -0
- package/template/claude-task-manager/lib/message-pagination.js +106 -4
- package/template/claude-task-manager/lib/microsoft-dev-tunnel-setup.js +750 -26
- package/template/claude-task-manager/lib/mobile-auth-api.js +274 -7
- package/template/claude-task-manager/lib/mobile-auth-store.js +592 -10
- package/template/claude-task-manager/lib/mobile-notification-dispatcher.js +15 -0
- package/template/claude-task-manager/lib/model-overview-brain-fallback.js +311 -0
- package/template/claude-task-manager/lib/model-overview-cache.js +141 -0
- package/template/claude-task-manager/lib/models-health-routing-notice.js +126 -0
- package/template/claude-task-manager/lib/node-pin-guard.js +93 -0
- package/template/claude-task-manager/lib/perf-tracker.js +242 -6
- package/template/claude-task-manager/lib/permission-match.js +76 -0
- package/template/claude-task-manager/lib/permission-sync.js +133 -20
- package/template/claude-task-manager/lib/process-title.js +35 -0
- package/template/claude-task-manager/lib/prompt-executions-query.js +25 -0
- package/template/claude-task-manager/lib/prompt-index-disk-cache.js +44 -0
- package/template/claude-task-manager/lib/prompt-intent.js +132 -0
- package/template/claude-task-manager/lib/provider-user-context.js +34 -0
- package/template/claude-task-manager/lib/read-pool-client.js +313 -0
- package/template/claude-task-manager/lib/readpool-breaker.js +31 -0
- package/template/claude-task-manager/lib/recent-sessions-breaker.js +12 -0
- package/template/claude-task-manager/lib/remote-feedback-client.js +72 -0
- package/template/claude-task-manager/lib/remote-relay-protocol.js +37 -4
- package/template/claude-task-manager/lib/remote-relay-store.js +159 -0
- package/template/claude-task-manager/lib/remote-submission-observer.js +278 -0
- package/template/claude-task-manager/lib/restart-guard.js +109 -0
- package/template/claude-task-manager/lib/restore-interruption-detector.js +439 -0
- package/template/claude-task-manager/lib/restore-policy.js +13 -0
- package/template/claude-task-manager/lib/restore-resume-batch.js +74 -0
- package/template/claude-task-manager/lib/restore-runtime.js +68 -0
- package/template/claude-task-manager/lib/restore-storm.js +34 -0
- package/template/claude-task-manager/lib/resume-cwd.js +36 -0
- package/template/claude-task-manager/lib/resume-preflight.js +313 -0
- package/template/claude-task-manager/lib/runtime-work-registry.js +444 -0
- package/template/claude-task-manager/lib/sanitize-openai-auth.js +31 -0
- package/template/claude-task-manager/lib/scheduler.js +21 -1
- package/template/claude-task-manager/lib/scrollback-snapshot-store.js +159 -0
- package/template/claude-task-manager/lib/serial-task-queue.js +64 -0
- package/template/claude-task-manager/lib/server-listeners.js +239 -0
- package/template/claude-task-manager/lib/session-capture.js +42 -7
- package/template/claude-task-manager/lib/session-content-backfill.js +131 -0
- package/template/claude-task-manager/lib/session-history.js +388 -43
- package/template/claude-task-manager/lib/session-host-manager.js +287 -0
- package/template/claude-task-manager/lib/session-image-refs.js +209 -0
- package/template/claude-task-manager/lib/session-jobs.js +399 -59
- package/template/claude-task-manager/lib/session-prompt-index.js +137 -0
- package/template/claude-task-manager/lib/session-restore.js +53 -0
- package/template/claude-task-manager/lib/session-standup.js +123 -23
- package/template/claude-task-manager/lib/session-state-bus.js +14 -0
- package/template/claude-task-manager/lib/session-stream.js +64 -16
- package/template/claude-task-manager/lib/session-timeline-summary.js +260 -0
- package/template/claude-task-manager/lib/session-token-usage.js +494 -0
- package/template/claude-task-manager/lib/session-workspace-binding.js +356 -0
- package/template/claude-task-manager/lib/setup-network-config.js +9 -0
- package/template/claude-task-manager/lib/size-cap.js +45 -0
- package/template/claude-task-manager/lib/size-cap.test.js +62 -0
- package/template/claude-task-manager/lib/skill-autocomplete.js +180 -1
- package/template/claude-task-manager/lib/skill-intent-resolver.js +304 -0
- package/template/claude-task-manager/lib/sqlite-driver.js +19 -3
- package/template/claude-task-manager/lib/standup-attention.js +7 -3
- package/template/claude-task-manager/lib/status-authority.js +39 -0
- package/template/claude-task-manager/lib/status-hooks.js +4 -0
- package/template/claude-task-manager/lib/storage-migration.js +235 -0
- package/template/claude-task-manager/lib/structured-capture.js +298 -0
- package/template/claude-task-manager/lib/sync-io-census.js +163 -0
- package/template/claude-task-manager/lib/tailscale-setup.js +6 -0
- package/template/claude-task-manager/lib/terminal-activity-evidence.js +33 -0
- package/template/claude-task-manager/lib/terminal-choice.js +364 -0
- package/template/claude-task-manager/lib/terminal-control-sanitize.js +17 -0
- package/template/claude-task-manager/lib/terminal-fingerprint.js +48 -0
- package/template/claude-task-manager/lib/terminal-output-flush.js +84 -0
- package/template/claude-task-manager/lib/timeline-order.js +122 -0
- package/template/claude-task-manager/lib/transcript-store.js +348 -43
- package/template/claude-task-manager/lib/transport-security.js +84 -1
- package/template/claude-task-manager/lib/wait-state.js +184 -0
- package/template/claude-task-manager/lib/walle-client.js +47 -5
- package/template/claude-task-manager/lib/walle-ctm-history.js +564 -4
- package/template/claude-task-manager/lib/walle-external-actions.js +135 -16
- package/template/claude-task-manager/lib/walle-history-hydration.js +46 -0
- package/template/claude-task-manager/lib/walle-native-health.js +403 -0
- package/template/claude-task-manager/lib/walle-repair.js +701 -0
- package/template/claude-task-manager/lib/walle-session-cache.js +109 -0
- package/template/claude-task-manager/lib/walle-session-context.js +57 -21
- package/template/claude-task-manager/lib/walle-session-model-catalog.js +34 -0
- package/template/claude-task-manager/lib/walle-supervisor.js +539 -63
- package/template/claude-task-manager/lib/walle-transcript.js +52 -0
- package/template/claude-task-manager/lib/worktree-active-sync.js +11 -7
- package/template/claude-task-manager/lib/worktree-cwd.js +32 -1
- package/template/claude-task-manager/package.json +1 -1
- package/template/claude-task-manager/prompt-harvest.js +89 -66
- package/template/claude-task-manager/providers/claude-code.js +51 -3
- package/template/claude-task-manager/providers/cursor.js +140 -45
- package/template/claude-task-manager/public/css/reviews.css +551 -61
- package/template/claude-task-manager/public/css/setup.css +191 -0
- package/template/claude-task-manager/public/css/walle-session.css +865 -10
- package/template/claude-task-manager/public/css/walle.css +154 -0
- package/template/claude-task-manager/public/designs/ai-providers-consolidation-v2.html +830 -0
- package/template/claude-task-manager/public/index.html +18516 -2058
- package/template/claude-task-manager/public/ipad.html +363 -0
- package/template/claude-task-manager/public/js/document-review-links.js +301 -0
- package/template/claude-task-manager/public/js/image-normalize.js +69 -36
- package/template/claude-task-manager/public/js/message-renderer.js +1265 -77
- package/template/claude-task-manager/public/js/prompts.js +66 -29
- package/template/claude-task-manager/public/js/reviews.js +901 -133
- package/template/claude-task-manager/public/js/session-activity-utils.js +11 -1
- package/template/claude-task-manager/public/js/session-search-utils.js +94 -10
- package/template/claude-task-manager/public/js/session-status-precedence.js +23 -5
- package/template/claude-task-manager/public/js/setup.js +1273 -176
- package/template/claude-task-manager/public/js/stream-view.js +691 -73
- package/template/claude-task-manager/public/js/terminal-reconciler.js +210 -0
- package/template/claude-task-manager/public/js/walle-session.js +2455 -158
- package/template/claude-task-manager/public/js/walle.js +455 -28
- package/template/claude-task-manager/public/m/app.css +2909 -262
- package/template/claude-task-manager/public/m/app.js +6601 -398
- package/template/claude-task-manager/public/m/claim.html +224 -17
- package/template/claude-task-manager/public/m/index.html +117 -21
- package/template/claude-task-manager/public/m/sw.js +3 -1
- package/template/claude-task-manager/public/manifest.json +2 -2
- package/template/claude-task-manager/public/prompts.html +30 -14
- package/template/claude-task-manager/queue-engine.js +507 -28
- package/template/claude-task-manager/scripts/repair-claude-session-images.js +27 -8
- package/template/claude-task-manager/server.js +14341 -2197
- package/template/claude-task-manager/session-integrity.js +160 -18
- package/template/claude-task-manager/session-search-ranking.js +1 -0
- package/template/claude-task-manager/session-utils.js +25 -5
- package/template/claude-task-manager/workers/approval-blocklist.js +96 -6
- package/template/claude-task-manager/workers/approval-widget-validator.js +14 -8
- package/template/claude-task-manager/workers/conversation-import-worker.js +11 -50
- package/template/claude-task-manager/workers/db-owner-worker.js +386 -0
- package/template/claude-task-manager/workers/harvest-worker.js +9 -55
- package/template/claude-task-manager/workers/headless-term-worker.js +9 -530
- package/template/claude-task-manager/workers/read-pool-worker.js +387 -0
- package/template/claude-task-manager/workers/scrollback-worker.js +11 -72
- package/template/claude-task-manager/workers/session-host-process.js +146 -0
- package/template/claude-task-manager/workers/session-integrity-worker.js +10 -54
- package/template/claude-task-manager/workers/state-detectors/base.js +18 -1
- package/template/claude-task-manager/workers/state-detectors/claude-code.js +182 -9
- package/template/claude-task-manager/workers/state-detectors/codex.js +150 -2
- package/template/claude-task-manager/workers/state-detectors/cursor.js +127 -0
- package/template/claude-task-manager/workers/state-detectors/gemini.js +21 -0
- package/template/claude-task-manager/workers/state-detectors/index.js +29 -0
- package/template/claude-task-manager/workers/state-detectors/opencode.js +103 -0
- package/template/docs/design/markdown-review-pane.md +206 -0
- package/template/docs/designs/2026-05-17-portkey-gateway-provider-ux.md +129 -38
- package/template/docs/designs/2026-05-20-mobile-worktree-finish-command.md +27 -0
- package/template/docs/designs/2026-05-22-ai-configuration-consolidation.md +248 -0
- package/template/docs/designs/ai-configuration-consolidation-mock.html +812 -0
- package/template/docs/private-memory-and-pii-policy.md +69 -0
- package/template/package.json +2 -1
- package/template/scripts/check-private-data.js +201 -0
- package/template/shared/sqlite-owner-guard.js +30 -0
- package/template/shared/sqlite-owner-write-queue.js +225 -0
- package/template/shared/sqlite-storage-policy.js +111 -0
- package/template/shared/sqlite-write-lock.js +428 -0
- package/template/wall-e/agent-runners/claude-code.js +5 -0
- package/template/wall-e/agent.js +166 -22
- package/template/wall-e/api-walle.js +524 -70
- package/template/wall-e/auth/provider-flows.js +11 -1
- package/template/wall-e/bin/walle-mcp-stdio.js +341 -17
- package/template/wall-e/brain.js +1614 -141
- package/template/wall-e/chat/attachment-blocks.js +96 -0
- package/template/wall-e/chat/attachments.js +2 -1
- package/template/wall-e/chat/capability-resolver.js +7 -7
- package/template/wall-e/chat/context-messages.js +28 -0
- package/template/wall-e/chat/conversation-frame.js +630 -0
- package/template/wall-e/chat/provider-messages.js +125 -0
- package/template/wall-e/chat.js +1002 -233
- package/template/wall-e/coding/acceptance-contract.js +170 -0
- package/template/wall-e/coding/acp-adapter.js +1 -1
- package/template/wall-e/coding/agent-catalog.js +3 -0
- package/template/wall-e/coding/artifact-store.js +93 -0
- package/template/wall-e/coding/capability-router.js +120 -0
- package/template/wall-e/coding/coding-run-controller.js +423 -0
- package/template/wall-e/coding/compaction-service.js +157 -12
- package/template/wall-e/coding/frontend-verification.js +258 -0
- package/template/wall-e/coding/lifecycle-hooks.js +75 -0
- package/template/wall-e/coding/local-preview-contract.js +157 -0
- package/template/wall-e/coding/permission-service.js +57 -13
- package/template/wall-e/coding/prompt-bundle.js +19 -1
- package/template/wall-e/coding/prompt-section-registry.js +227 -0
- package/template/wall-e/coding/provider-compat.js +15 -0
- package/template/wall-e/coding/runtime-events.js +224 -0
- package/template/wall-e/coding/runtime-mode.js +3 -0
- package/template/wall-e/coding/side-git-snapshot.js +160 -4
- package/template/wall-e/coding/snapshot-service.js +143 -1
- package/template/wall-e/coding/stream-processor.js +388 -34
- package/template/wall-e/coding/task-tool.js +141 -4
- package/template/wall-e/coding/tool-execution-controller.js +365 -0
- package/template/wall-e/coding/tool-registry.js +43 -5
- package/template/wall-e/coding/user-hooks.js +217 -0
- package/template/wall-e/coding-orchestrator.js +1330 -221
- package/template/wall-e/coding-prompts.js +20 -4
- package/template/wall-e/context/context-builder.js +15 -2
- package/template/wall-e/decision/confidence.js +1 -1
- package/template/wall-e/docs/coding-acceptance-contract.md +41 -0
- package/template/wall-e/docs/external-action-controller.md +26 -6
- package/template/wall-e/docs/telemetry-lifecycle.md +8 -2
- package/template/wall-e/embeddings.js +591 -53
- package/template/wall-e/external-action-controller.js +12 -0
- package/template/wall-e/http/auth.js +1 -0
- package/template/wall-e/http/chat-api.js +46 -11
- package/template/wall-e/http/model-admin.js +836 -34
- package/template/wall-e/lib/boot-profile.js +88 -0
- package/template/wall-e/lib/event-loop-monitor.js +93 -0
- package/template/wall-e/lib/service-health.js +194 -0
- package/template/wall-e/llm/anthropic.js +130 -5
- package/template/wall-e/llm/client.js +266 -63
- package/template/wall-e/llm/default-fallback.js +382 -0
- package/template/wall-e/llm/health.js +19 -0
- package/template/wall-e/llm/message-guard.js +78 -0
- package/template/wall-e/llm/model-catalog.js +252 -1
- package/template/wall-e/llm/openai.js +26 -4
- package/template/wall-e/llm/portkey-sync.js +654 -0
- package/template/wall-e/llm/provider-error.js +30 -2
- package/template/wall-e/llm/registry.js +5 -1
- package/template/wall-e/llm/request-compat.js +67 -0
- package/template/wall-e/loops/backfill.js +79 -23
- package/template/wall-e/loops/brain-optimize.js +67 -0
- package/template/wall-e/loops/ingest.js +25 -10
- package/template/wall-e/loops/question-digest.js +160 -0
- package/template/wall-e/loops/reflect.js +6 -4
- package/template/wall-e/loops/think.js +39 -12
- package/template/wall-e/mcp-server.js +318 -36
- package/template/wall-e/memory/ctm-context-client.js +52 -14
- package/template/wall-e/memory/ctm-operational-context.js +237 -0
- package/template/wall-e/memory/ctm-prompt-executions-client.js +128 -0
- package/template/wall-e/memory/ctm-session-context.js +111 -63
- package/template/wall-e/prompts/coding/deepseek.txt +3 -0
- package/template/wall-e/prompts/coding/gemini.txt +6 -0
- package/template/wall-e/prompts/coding/gpt.txt +6 -0
- package/template/wall-e/prompts/coding/local.txt +7 -0
- package/template/wall-e/runtime/decision-hooks.js +115 -0
- package/template/wall-e/runtime/devbox-gateway.js +82 -8
- package/template/wall-e/runtime/prompt-manifest.js +86 -0
- package/template/wall-e/runtime/tool-executor.js +269 -0
- package/template/wall-e/runtime/tool-result-envelope.js +138 -0
- package/template/wall-e/runtime/transcript-projection.js +60 -0
- package/template/wall-e/runtime/walle-runtime.js +224 -0
- package/template/wall-e/scripts/db-optimize/migrate.js +162 -0
- package/template/wall-e/scripts/db-optimize/recall-eval.js +117 -0
- package/template/wall-e/server.js +15 -0
- package/template/wall-e/session-files.js +9 -0
- package/template/wall-e/skills/_bundled/google-calendar/run.js +1 -1
- package/template/wall-e/skills/_bundled/gws-workspace/run.js +1 -1
- package/template/wall-e/skills/_bundled/slack-mentions/run.js +76 -6
- package/template/wall-e/skills/claude-code-reader.js +7 -3
- package/template/wall-e/skills/script-skill-runner.js +10 -0
- package/template/wall-e/skills/skill-planner.js +38 -0
- package/template/wall-e/tools/builtin-middleware.js +19 -9
- package/template/wall-e/tools/local-tools.js +1428 -16
- package/template/wall-e/tools/permission-checker.js +73 -5
- package/template/wall-e/tools/question-manager.js +117 -7
- package/template/wall-e/training/harvester.js +12 -28
- package/template/wall-e/training/replay.js +25 -80
- package/template/website/index.html +10 -10
- package/template/wall-e/eval/ab-test.js +0 -203
- package/template/wall-e/eval/agent-runner.js +0 -772
- package/template/wall-e/eval/agent-scorer.js +0 -461
- package/template/wall-e/eval/aggregator.js +0 -414
- package/template/wall-e/eval/allowed-test-commands.js +0 -34
- package/template/wall-e/eval/benchmark-generator.js +0 -113
- package/template/wall-e/eval/benchmarks/chat-eval.json +0 -1662
- package/template/wall-e/eval/benchmarks/chat.json +0 -82
- package/template/wall-e/eval/benchmarks/coding-agent-real.json +0 -1
- package/template/wall-e/eval/benchmarks/coding-agent.json +0 -1581
- package/template/wall-e/eval/benchmarks/coding.json +0 -122
- package/template/wall-e/eval/benchmarks/memory-retrieval.json +0 -234
- package/template/wall-e/eval/benchmarks/reasoning.json +0 -82
- package/template/wall-e/eval/benchmarks/swebench-lite-30.json +0 -212
- package/template/wall-e/eval/benchmarks.js +0 -669
- package/template/wall-e/eval/cc-replay.js +0 -719
- package/template/wall-e/eval/chat-eval.js +0 -525
- package/template/wall-e/eval/check-keys.js +0 -15
- package/template/wall-e/eval/check-providers.js +0 -42
- package/template/wall-e/eval/codex-cli-baseline.js +0 -669
- package/template/wall-e/eval/coding-agent-real.js +0 -570
- package/template/wall-e/eval/context-compactor.js +0 -251
- package/template/wall-e/eval/debug-agent003.js +0 -68
- package/template/wall-e/eval/diagnostics.js +0 -216
- package/template/wall-e/eval/eval-orchestrator.js +0 -642
- package/template/wall-e/eval/evaluate.js +0 -202
- package/template/wall-e/eval/evaluator.js +0 -373
- package/template/wall-e/eval/exporter.js +0 -212
- package/template/wall-e/eval/fixtures/express-basic/package.json +0 -9
- package/template/wall-e/eval/fixtures/express-basic/server.js +0 -115
- package/template/wall-e/eval/fixtures/express-basic/test.js +0 -83
- package/template/wall-e/eval/fixtures/express-buggy/package.json +0 -9
- package/template/wall-e/eval/fixtures/express-buggy/server.js +0 -113
- package/template/wall-e/eval/fixtures/express-buggy/test.js +0 -83
- package/template/wall-e/eval/fixtures/express-buggy-items/package.json +0 -9
- package/template/wall-e/eval/fixtures/express-buggy-items/server.js +0 -112
- package/template/wall-e/eval/fixtures/express-buggy-items/test.js +0 -83
- package/template/wall-e/eval/fixtures/express-buggy-search/package.json +0 -9
- package/template/wall-e/eval/fixtures/express-buggy-search/server.js +0 -121
- package/template/wall-e/eval/fixtures/express-buggy-search/test.js +0 -83
- package/template/wall-e/eval/fixtures/express-rename-data/data.js +0 -34
- package/template/wall-e/eval/fixtures/express-rename-data/package.json +0 -9
- package/template/wall-e/eval/fixtures/express-rename-data/server.js +0 -97
- package/template/wall-e/eval/fixtures/express-rename-data/test.js +0 -88
- package/template/wall-e/eval/fixtures/express-xss/package.json +0 -12
- package/template/wall-e/eval/fixtures/express-xss/server.js +0 -90
- package/template/wall-e/eval/fixtures/express-xss/test.js +0 -67
- package/template/wall-e/eval/fixtures/express-xss/views/profile.ejs +0 -9
- package/template/wall-e/eval/fixtures/fullstack-app/config/default.js +0 -9
- package/template/wall-e/eval/fixtures/fullstack-app/config/test.js +0 -13
- package/template/wall-e/eval/fixtures/fullstack-app/package.json +0 -11
- package/template/wall-e/eval/fixtures/fullstack-app/public/css/style.css +0 -137
- package/template/wall-e/eval/fixtures/fullstack-app/public/index.html +0 -46
- package/template/wall-e/eval/fixtures/fullstack-app/public/js/app.js +0 -121
- package/template/wall-e/eval/fixtures/fullstack-app/public/js/auth.js +0 -71
- package/template/wall-e/eval/fixtures/fullstack-app/public/js/items.js +0 -80
- package/template/wall-e/eval/fixtures/fullstack-app/public/js/users.js +0 -46
- package/template/wall-e/eval/fixtures/fullstack-app/public/login.html +0 -45
- package/template/wall-e/eval/fixtures/fullstack-app/public/register.html +0 -38
- package/template/wall-e/eval/fixtures/fullstack-app/scripts/migrate.js +0 -23
- package/template/wall-e/eval/fixtures/fullstack-app/scripts/seed.js +0 -46
- package/template/wall-e/eval/fixtures/fullstack-app/server/db.js +0 -99
- package/template/wall-e/eval/fixtures/fullstack-app/server/index.js +0 -94
- package/template/wall-e/eval/fixtures/fullstack-app/server/middleware/auth.js +0 -19
- package/template/wall-e/eval/fixtures/fullstack-app/server/middleware/logger.js +0 -19
- package/template/wall-e/eval/fixtures/fullstack-app/server/router.js +0 -50
- package/template/wall-e/eval/fixtures/fullstack-app/server/routes/auth.js +0 -69
- package/template/wall-e/eval/fixtures/fullstack-app/server/routes/health.js +0 -23
- package/template/wall-e/eval/fixtures/fullstack-app/server/routes/items.js +0 -88
- package/template/wall-e/eval/fixtures/fullstack-app/server/routes/users.js +0 -75
- package/template/wall-e/eval/fixtures/fullstack-app/server/test.js +0 -198
- package/template/wall-e/eval/fixtures/fullstack-app/server/utils/response.js +0 -34
- package/template/wall-e/eval/fixtures/fullstack-app/server/utils/validate.js +0 -26
- package/template/wall-e/eval/fixtures/fullstack-app/server.js +0 -8
- package/template/wall-e/eval/fixtures/fullstack-app/test.js +0 -12
- package/template/wall-e/eval/fixtures/monorepo-basic/package.json +0 -8
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/api/data.js +0 -58
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/api/middleware.js +0 -46
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/api/package.json +0 -8
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/api/routes.js +0 -64
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/api/server.js +0 -56
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/api/test.js +0 -116
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/cli/commands.js +0 -61
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/cli/index.js +0 -62
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/cli/output.js +0 -43
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/cli/package.json +0 -11
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/cli/test.js +0 -44
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/shared/formatters.js +0 -43
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/shared/index.js +0 -12
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/shared/package.json +0 -5
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/shared/test.js +0 -55
- package/template/wall-e/eval/fixtures/monorepo-basic/packages/shared/validators.js +0 -29
- package/template/wall-e/eval/fixtures/monorepo-basic/test.js +0 -46
- package/template/wall-e/eval/fixtures/node-cli/index.js +0 -78
- package/template/wall-e/eval/fixtures/node-cli/package.json +0 -10
- package/template/wall-e/eval/fixtures/node-cli/test.js +0 -57
- package/template/wall-e/eval/fixtures/node-typed/package.json +0 -8
- package/template/wall-e/eval/fixtures/node-typed/src/handlers.js +0 -31
- package/template/wall-e/eval/fixtures/node-typed/src/utils.js +0 -33
- package/template/wall-e/eval/fixtures/node-typed/test.js +0 -36
- package/template/wall-e/eval/fixtures/python-flask/app.py +0 -14
- package/template/wall-e/eval/fixtures/python-flask/requirements.txt +0 -2
- package/template/wall-e/eval/fixtures/python-flask/test_app.py +0 -25
- package/template/wall-e/eval/fixtures/wall-e-subset/brain.js +0 -105
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/aggregator.js +0 -101
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/benchmarks/chat.json +0 -20
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/benchmarks/coding.json +0 -32
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/benchmarks.js +0 -64
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/fixtures/simple-project/package.json +0 -6
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/fixtures/simple-project/server.js +0 -31
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/fixtures/simple-project/test.js +0 -18
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/fixtures/simple-project/utils.js +0 -34
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/runner.js +0 -104
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/scorer.js +0 -73
- package/template/wall-e/eval/fixtures/wall-e-subset/eval/test.js +0 -134
- package/template/wall-e/eval/fixtures/wall-e-subset/llm/client.js +0 -99
- package/template/wall-e/eval/fixtures/wall-e-subset/llm/providers.js +0 -63
- package/template/wall-e/eval/fixtures/wall-e-subset/llm/test.js +0 -70
- package/template/wall-e/eval/fixtures/wall-e-subset/package.json +0 -10
- package/template/wall-e/eval/fixtures/wall-e-subset/test.js +0 -86
- package/template/wall-e/eval/harvester.js +0 -685
- package/template/wall-e/eval/head-to-head.js +0 -388
- package/template/wall-e/eval/humaneval-adapter.js +0 -321
- package/template/wall-e/eval/list-models.js +0 -31
- package/template/wall-e/eval/livecodebench-adapter.js +0 -291
- package/template/wall-e/eval/mail-integration.js +0 -443
- package/template/wall-e/eval/manifest.js +0 -186
- package/template/wall-e/eval/meta-harness/adapters/coding-agent.js +0 -57
- package/template/wall-e/eval/meta-harness/bootstrap-snapshot.js +0 -149
- package/template/wall-e/eval/meta-harness/candidate-store.js +0 -117
- package/template/wall-e/eval/meta-harness/cli.js +0 -86
- package/template/wall-e/eval/meta-harness/domain-spec.js +0 -154
- package/template/wall-e/eval/meta-harness/domains/coding-agent.domain.json +0 -84
- package/template/wall-e/eval/meta-harness/examples/env-bootstrap-candidate.js +0 -29
- package/template/wall-e/eval/meta-harness/experience-store.js +0 -174
- package/template/wall-e/eval/meta-harness/frontier.js +0 -96
- package/template/wall-e/eval/meta-harness/harness-interface.js +0 -90
- package/template/wall-e/eval/meta-harness/leakage-guard.js +0 -80
- package/template/wall-e/eval/meta-harness/optimizer.js +0 -207
- package/template/wall-e/eval/meta-harness/proposer-runner.js +0 -110
- package/template/wall-e/eval/meta-harness/reporting.js +0 -58
- package/template/wall-e/eval/meta-harness/telemetry.js +0 -27
- package/template/wall-e/eval/meta-harness/validation.js +0 -81
- package/template/wall-e/eval/promoter.js +0 -228
- package/template/wall-e/eval/provider-normalizer.js +0 -33
- package/template/wall-e/eval/replay.js +0 -395
- package/template/wall-e/eval/run-agent-benchmarks.js +0 -386
- package/template/wall-e/eval/run-codex-cli-baseline.js +0 -177
- package/template/wall-e/eval/run-coding-agent-real.js +0 -187
- package/template/wall-e/eval/run-eval.js +0 -435
- package/template/wall-e/eval/run-model-comparison.js +0 -142
- package/template/wall-e/eval/session-evaluator.js +0 -187
- package/template/wall-e/eval/session-miner.js +0 -207
- package/template/wall-e/eval/session-retrieval-benchmark.js +0 -150
- package/template/wall-e/eval/session-transcripts.js +0 -509
- package/template/wall-e/eval/shadow.js +0 -161
- package/template/wall-e/eval/swebench-adapter.js +0 -345
- package/template/wall-e/eval/swebench-docker.js +0 -192
- package/template/wall-e/eval/train.py +0 -320
- package/template/wall-e/eval/trainer.js +0 -232
- package/template/wall-e/eval/weekly-eval-loop.js +0 -241
|
@@ -0,0 +1,630 @@
|
|
|
1
|
+
'use strict';
|
|
2
|
+
|
|
3
|
+
const FRAME_SCHEMA_VERSION = 1;
|
|
4
|
+
const MAX_CONTEXT_MESSAGES = 60;
|
|
5
|
+
const MAX_RECOMMENDATION_NAMES = 12;
|
|
6
|
+
|
|
7
|
+
const DOMAIN_PATTERNS = [
|
|
8
|
+
['email', /\b(e-?mail|inbox|reply|respond|thread|recipient|cc|bcc|gmail|mail)\b/i],
|
|
9
|
+
['calendar', /\b(calendar|meeting|schedule|appointment|event|availability|free|busy)\b/i],
|
|
10
|
+
['coding', /\b(code|repo|repository|branch|commit|merge|test|build|bug|fix|implement|refactor|deploy|website|app|ui|ux)\b/i],
|
|
11
|
+
['travel', /\b(trip|travel|flight|hotel|lodging|airbnb|vrbo|stay|rental car|car rental|airport|cruise|booking|reservation|itinerary|destination)\b/i],
|
|
12
|
+
['recommendation', /\b(recommend|recommendation|top\s+\d+|best|rank|choose|pick|options?)\b/i],
|
|
13
|
+
];
|
|
14
|
+
|
|
15
|
+
const TRAVEL_RE = /\b(trip|travel|flight|hotel|lodging|airbnb|vrbo|stay|rental car|car rental|airport|cruise|booking|reservation|itinerary|destination)\b/i;
|
|
16
|
+
const RECOMMENDATION_RE = /\b(recommend|recommendation|top\s+\d+|best|rank|choose|pick|options?)\b/i;
|
|
17
|
+
const LODGING_RE = /\b(hotel|lodging|airbnb|vrbo|vacation rental|short[- ]term rental|stay|staying|room|resort|inn|booking|book directly|direct booking)\b/i;
|
|
18
|
+
const TRANSPORT_RE = /\b(flight|airport|rental car|car rental|drop ?off|pick ?up|drive|driving|terminal)\b/i;
|
|
19
|
+
const CORRECTION_RE = /\b(doesn'?t change|does not change|same|unchanged|wrong|not what|lost|different|instead|rather than|not just|not limited|doesn'?t have to be|can be|go back|stick with|you missed|you ignored)\b/i;
|
|
20
|
+
|
|
21
|
+
const STOP_LOCATION_WORDS = new Set([
|
|
22
|
+
'Given', 'Give', 'Can', 'Should', 'What', 'When', 'Where', 'Why', 'How', 'The',
|
|
23
|
+
'This', 'That', 'There', 'Hotel', 'Hotels', 'Airbnb', 'Vrbo', 'Dec', 'Jan',
|
|
24
|
+
'Feb', 'Mar', 'Apr', 'May', 'Jun', 'Jul', 'Aug', 'Sep', 'Sept', 'Oct', 'Nov',
|
|
25
|
+
'Monday', 'Tuesday', 'Wednesday', 'Thursday', 'Friday', 'Saturday', 'Sunday',
|
|
26
|
+
'Plan', 'Help', 'Recommend', 'Recommendation', 'Recommendations', 'Top', 'Book',
|
|
27
|
+
'Booking', 'Direct', 'New',
|
|
28
|
+
]);
|
|
29
|
+
|
|
30
|
+
function parseSessionMetadata(raw) {
|
|
31
|
+
if (!raw) return {};
|
|
32
|
+
if (typeof raw === 'object') return raw;
|
|
33
|
+
try {
|
|
34
|
+
const parsed = JSON.parse(raw);
|
|
35
|
+
return parsed && typeof parsed === 'object' ? parsed : {};
|
|
36
|
+
} catch {
|
|
37
|
+
return {};
|
|
38
|
+
}
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
function normalizeSemanticMessages(messages, currentMessage = '') {
|
|
42
|
+
const raw = Array.isArray(messages) ? messages : [];
|
|
43
|
+
const normalized = [];
|
|
44
|
+
for (const item of raw) {
|
|
45
|
+
if (!item || typeof item !== 'object') continue;
|
|
46
|
+
const role = String(item.role || '').trim();
|
|
47
|
+
if (role !== 'user' && role !== 'assistant') continue;
|
|
48
|
+
const content = contentToText(item.content != null ? item.content : item.text);
|
|
49
|
+
if (!content.trim()) continue;
|
|
50
|
+
normalized.push({ role, content: content.trim() });
|
|
51
|
+
}
|
|
52
|
+
const current = String(currentMessage || '').trim();
|
|
53
|
+
if (current) {
|
|
54
|
+
const last = normalized[normalized.length - 1];
|
|
55
|
+
if (!last || last.role !== 'user' || last.content !== current) {
|
|
56
|
+
normalized.push({ role: 'user', content: current });
|
|
57
|
+
} else {
|
|
58
|
+
last.content = current;
|
|
59
|
+
}
|
|
60
|
+
}
|
|
61
|
+
return normalized.slice(-MAX_CONTEXT_MESSAGES);
|
|
62
|
+
}
|
|
63
|
+
|
|
64
|
+
function contentToText(content) {
|
|
65
|
+
if (content == null) return '';
|
|
66
|
+
if (typeof content === 'string') return content;
|
|
67
|
+
if (Array.isArray(content)) {
|
|
68
|
+
return content.map((part) => {
|
|
69
|
+
if (!part) return '';
|
|
70
|
+
if (typeof part === 'string') return part;
|
|
71
|
+
if (part.type === 'text') return part.text || '';
|
|
72
|
+
if (part.type === 'tool_result') return part.content || '';
|
|
73
|
+
if (part.type === 'tool_use') return `[tool:${part.name || 'unknown'}]`;
|
|
74
|
+
return '';
|
|
75
|
+
}).filter(Boolean).join('\n');
|
|
76
|
+
}
|
|
77
|
+
try { return JSON.stringify(content); } catch { return String(content); }
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
function buildConversationRuntime({ messages, currentMessage, previousFrame, channel, cwd, taskType } = {}) {
|
|
81
|
+
const semanticMessages = normalizeSemanticMessages(messages, currentMessage);
|
|
82
|
+
const frame = deriveConversationFrame(semanticMessages, {
|
|
83
|
+
previousFrame,
|
|
84
|
+
channel,
|
|
85
|
+
cwd,
|
|
86
|
+
taskType,
|
|
87
|
+
});
|
|
88
|
+
return {
|
|
89
|
+
schemaVersion: FRAME_SCHEMA_VERSION,
|
|
90
|
+
messages: semanticMessages,
|
|
91
|
+
frame,
|
|
92
|
+
validators: validatorsForFrame(frame),
|
|
93
|
+
};
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
function deriveConversationFrame(messages, opts = {}) {
|
|
97
|
+
const previous = normalizeFrame(opts.previousFrame);
|
|
98
|
+
const currentUser = [...messages].reverse().find((msg) => msg.role === 'user')?.content || '';
|
|
99
|
+
const allText = messages.map((msg) => msg.content).join('\n');
|
|
100
|
+
const currentDomain = inferDomain(currentUser) || previous.domain || inferDomain(allText) || 'general';
|
|
101
|
+
const isCorrection = CORRECTION_RE.test(currentUser);
|
|
102
|
+
const now = new Date().toISOString();
|
|
103
|
+
|
|
104
|
+
const frame = {
|
|
105
|
+
schemaVersion: FRAME_SCHEMA_VERSION,
|
|
106
|
+
updatedAt: now,
|
|
107
|
+
domain: currentDomain,
|
|
108
|
+
activeGoal: inferActiveGoal(currentUser, previous.activeGoal),
|
|
109
|
+
scope: {
|
|
110
|
+
...(previous.scope || {}),
|
|
111
|
+
...(opts.cwd ? { projectPath: opts.cwd } : {}),
|
|
112
|
+
},
|
|
113
|
+
constraints: [...(previous.constraints || [])],
|
|
114
|
+
recentCorrections: [...(previous.recentCorrections || [])],
|
|
115
|
+
inactiveRelatedFrames: [...(previous.inactiveRelatedFrames || [])],
|
|
116
|
+
evidenceRefs: [],
|
|
117
|
+
previousRecommendations: Array.isArray(previous.previousRecommendations)
|
|
118
|
+
? previous.previousRecommendations.slice(0, MAX_RECOMMENDATION_NAMES)
|
|
119
|
+
: [],
|
|
120
|
+
confidence: previous.confidence || 0.45,
|
|
121
|
+
};
|
|
122
|
+
|
|
123
|
+
const travelLike = currentDomain === 'travel' || TRAVEL_RE.test(allText) || RECOMMENDATION_RE.test(currentUser);
|
|
124
|
+
if (travelLike) applyTravelFrame(frame, messages, currentUser, isCorrection);
|
|
125
|
+
applyActionFrame(frame, messages, currentUser);
|
|
126
|
+
|
|
127
|
+
const constraints = inferConstraints(currentUser);
|
|
128
|
+
for (const constraint of constraints) addUnique(frame.constraints, constraint);
|
|
129
|
+
if (isCorrection) addRecentCorrection(frame, currentUser);
|
|
130
|
+
|
|
131
|
+
const explicitDates = extractDates(allText);
|
|
132
|
+
if (explicitDates.length) frame.scope.dates = explicitDates.slice(-4);
|
|
133
|
+
const travelerSummary = extractTravelerSummary(allText);
|
|
134
|
+
if (travelerSummary) frame.scope.travelers = travelerSummary;
|
|
135
|
+
|
|
136
|
+
frame.previousRecommendations = collectPreviousRecommendations(messages)
|
|
137
|
+
.concat(frame.previousRecommendations || [])
|
|
138
|
+
.filter(uniqueByLower)
|
|
139
|
+
.slice(0, MAX_RECOMMENDATION_NAMES);
|
|
140
|
+
|
|
141
|
+
frame.confidence = scoreFrameConfidence(frame);
|
|
142
|
+
return pruneFrame(frame);
|
|
143
|
+
}
|
|
144
|
+
|
|
145
|
+
function normalizeFrame(frame) {
|
|
146
|
+
if (!frame || typeof frame !== 'object') return {};
|
|
147
|
+
return {
|
|
148
|
+
schemaVersion: frame.schemaVersion || FRAME_SCHEMA_VERSION,
|
|
149
|
+
domain: typeof frame.domain === 'string' ? frame.domain : undefined,
|
|
150
|
+
activeGoal: typeof frame.activeGoal === 'string' ? frame.activeGoal : undefined,
|
|
151
|
+
scope: frame.scope && typeof frame.scope === 'object' ? { ...frame.scope } : {},
|
|
152
|
+
constraints: Array.isArray(frame.constraints) ? frame.constraints.filter(Boolean).map(String) : [],
|
|
153
|
+
recentCorrections: Array.isArray(frame.recentCorrections) ? frame.recentCorrections.filter(Boolean).map(String).slice(-5) : [],
|
|
154
|
+
inactiveRelatedFrames: Array.isArray(frame.inactiveRelatedFrames) ? frame.inactiveRelatedFrames.filter(Boolean).map(String).slice(-8) : [],
|
|
155
|
+
evidenceRefs: Array.isArray(frame.evidenceRefs) ? frame.evidenceRefs : [],
|
|
156
|
+
previousRecommendations: Array.isArray(frame.previousRecommendations) ? frame.previousRecommendations.filter(Boolean).map(String) : [],
|
|
157
|
+
confidence: Number(frame.confidence || 0) || undefined,
|
|
158
|
+
};
|
|
159
|
+
}
|
|
160
|
+
|
|
161
|
+
function inferDomain(text) {
|
|
162
|
+
for (const [domain, re] of DOMAIN_PATTERNS) {
|
|
163
|
+
if (re.test(text || '')) return domain;
|
|
164
|
+
}
|
|
165
|
+
return null;
|
|
166
|
+
}
|
|
167
|
+
|
|
168
|
+
function inferActiveGoal(text, fallback = '') {
|
|
169
|
+
const value = String(text || '');
|
|
170
|
+
if (/\b(reply|respond)\b/i.test(value) && /\b(email|mail|thread)\b/i.test(value)) return 'reply to email thread';
|
|
171
|
+
if (/\b(send|draft|write)\b/i.test(value) && /\b(email|message)\b/i.test(value)) return 'draft or send outbound message';
|
|
172
|
+
if (/\b(calendar|schedule|create.*event|add.*event)\b/i.test(value)) return 'schedule calendar event';
|
|
173
|
+
if (/\b(hotel|lodging|airbnb|vrbo|stay)\b/i.test(value) && RECOMMENDATION_RE.test(value)) return 'recommend lodging options';
|
|
174
|
+
if (TRAVEL_RE.test(value)) return 'plan travel';
|
|
175
|
+
if (/\b(fix|implement|build|debug|refactor|test)\b/i.test(value)) return 'complete coding task';
|
|
176
|
+
if (RECOMMENDATION_RE.test(value)) return 'recommend options';
|
|
177
|
+
return fallback || 'answer current user request';
|
|
178
|
+
}
|
|
179
|
+
|
|
180
|
+
function applyTravelFrame(frame, messages, currentUser, isCorrection) {
|
|
181
|
+
const latestLodgingLocation = findLatestLocationForContext(messages, LODGING_RE);
|
|
182
|
+
if (latestLodgingLocation) frame.scope.lodgingLocation = latestLodgingLocation;
|
|
183
|
+
|
|
184
|
+
const latestTravelDestination = findLatestTravelDestination(messages);
|
|
185
|
+
if (latestTravelDestination && !frame.scope.destination) frame.scope.destination = latestTravelDestination;
|
|
186
|
+
|
|
187
|
+
const inactiveLocations = findInactiveTravelLocations(messages, frame.scope.lodgingLocation || frame.scope.destination);
|
|
188
|
+
for (const loc of inactiveLocations) addUnique(frame.inactiveRelatedFrames, loc);
|
|
189
|
+
|
|
190
|
+
if (/\bdirect(?:ly)?\b/i.test(currentUser) && /\b(book|booking|hotel|lodging|reservation)\b/i.test(currentUser)) {
|
|
191
|
+
addUnique(frame.constraints, 'booking_channel:direct');
|
|
192
|
+
}
|
|
193
|
+
if (/\bairbnb|vrbo|vacation rental|short[- ]term rental|home rental|condo\b/i.test(currentUser)) {
|
|
194
|
+
addUnique(frame.constraints, 'include_supply_type:alternative_rentals');
|
|
195
|
+
}
|
|
196
|
+
if (/\bnot\s+(?:just|only|limited to)\s+hilton|doesn'?t have to be hilton|not hilton-only\b/i.test(currentUser)) {
|
|
197
|
+
addUnique(frame.constraints, 'not_brand_limited:hilton');
|
|
198
|
+
}
|
|
199
|
+
if (/\b(new|different|changed|doesn'?t change|same)\b/i.test(currentUser) && RECOMMENDATION_RE.test(currentUser + ' recommendation')) {
|
|
200
|
+
addUnique(frame.constraints, 'requires_novel_recommendations');
|
|
201
|
+
}
|
|
202
|
+
if (isCorrection) addUnique(frame.constraints, 'latest_user_correction_wins');
|
|
203
|
+
}
|
|
204
|
+
|
|
205
|
+
function applyActionFrame(frame, messages, currentUser) {
|
|
206
|
+
if (frame.domain === 'email') {
|
|
207
|
+
if (/\b(reply|respond|original email|same thread)\b/i.test(currentUser)) {
|
|
208
|
+
addUnique(frame.constraints, 'email_action:reply_to_original_thread');
|
|
209
|
+
}
|
|
210
|
+
const recipients = extractEmails(messages.map(m => m.content).join('\n'));
|
|
211
|
+
if (recipients.length) frame.scope.recipients = recipients.slice(-8);
|
|
212
|
+
}
|
|
213
|
+
if (frame.domain === 'calendar') {
|
|
214
|
+
const calendars = [];
|
|
215
|
+
if (/\bwork\b|examplecorp/i.test(currentUser)) calendars.push('work');
|
|
216
|
+
if (/\bpersonal\b|gmail/i.test(currentUser)) calendars.push('personal');
|
|
217
|
+
if (calendars.length) frame.scope.calendars = calendars;
|
|
218
|
+
}
|
|
219
|
+
if (frame.domain === 'coding' && !frame.scope.projectPath) {
|
|
220
|
+
const pathMatch = messages.map(m => m.content).join('\n').match(/(?:cwd|worktree|repo|project)\s*[:=]\s*(`?)(\/[^\s`]+)\1/i);
|
|
221
|
+
if (pathMatch) frame.scope.projectPath = pathMatch[2];
|
|
222
|
+
}
|
|
223
|
+
}
|
|
224
|
+
|
|
225
|
+
function inferConstraints(text) {
|
|
226
|
+
const value = String(text || '');
|
|
227
|
+
const constraints = [];
|
|
228
|
+
if (/\bno approval needed|just do it|go ahead\b/i.test(value)) constraints.push('user_authorized_execution');
|
|
229
|
+
if (/\bdon'?t send\b|\bdo not send\b/i.test(value)) constraints.push('external_action:draft_only');
|
|
230
|
+
if (/\btest (it|your changes)|validate|verify\b/i.test(value)) constraints.push('verification_required');
|
|
231
|
+
return constraints;
|
|
232
|
+
}
|
|
233
|
+
|
|
234
|
+
function findLatestLocationForContext(messages, contextRe) {
|
|
235
|
+
const ordered = [
|
|
236
|
+
...[...messages].filter(msg => msg.role === 'user').reverse(),
|
|
237
|
+
...[...messages].filter(msg => msg.role !== 'user').reverse(),
|
|
238
|
+
];
|
|
239
|
+
for (const msg of ordered) {
|
|
240
|
+
const text = msg.content || '';
|
|
241
|
+
if (!contextRe.test(text) && !TRAVEL_RE.test(text)) continue;
|
|
242
|
+
const explicit = findPrepositionalLocation(text);
|
|
243
|
+
if (explicit) return explicit;
|
|
244
|
+
const locations = extractLocationCandidates(text);
|
|
245
|
+
const contextual = locations.find((loc) => appearsNearContext(text, loc, contextRe));
|
|
246
|
+
if (contextual) return contextual;
|
|
247
|
+
}
|
|
248
|
+
return '';
|
|
249
|
+
}
|
|
250
|
+
|
|
251
|
+
function findLatestTravelDestination(messages) {
|
|
252
|
+
const ordered = [
|
|
253
|
+
...[...messages].filter(msg => msg.role === 'user').reverse(),
|
|
254
|
+
...[...messages].filter(msg => msg.role !== 'user').reverse(),
|
|
255
|
+
];
|
|
256
|
+
for (const msg of ordered) {
|
|
257
|
+
const text = msg.content || '';
|
|
258
|
+
if (!TRAVEL_RE.test(text) && !RECOMMENDATION_RE.test(text)) continue;
|
|
259
|
+
const explicit = findPrepositionalLocation(text);
|
|
260
|
+
if (explicit) return explicit;
|
|
261
|
+
const locations = extractLocationCandidates(text);
|
|
262
|
+
if (locations.length) return locations[0];
|
|
263
|
+
}
|
|
264
|
+
return '';
|
|
265
|
+
}
|
|
266
|
+
|
|
267
|
+
function findInactiveTravelLocations(messages, activeLocation) {
|
|
268
|
+
const active = normalizeComparable(activeLocation);
|
|
269
|
+
const out = [];
|
|
270
|
+
for (const msg of messages) {
|
|
271
|
+
const text = msg.content || '';
|
|
272
|
+
if (!TRANSPORT_RE.test(text)) continue;
|
|
273
|
+
for (const loc of extractLocationCandidates(text)) {
|
|
274
|
+
if (normalizeComparable(loc) && normalizeComparable(loc) !== active) addUnique(out, loc);
|
|
275
|
+
}
|
|
276
|
+
}
|
|
277
|
+
return out.slice(-6);
|
|
278
|
+
}
|
|
279
|
+
|
|
280
|
+
function extractLocationCandidates(text) {
|
|
281
|
+
const value = stripUrls(String(text || ''));
|
|
282
|
+
const out = [];
|
|
283
|
+
const airportRe = /\b[A-Z]{3}\b/g;
|
|
284
|
+
for (const match of value.matchAll(airportRe)) addUnique(out, match[0]);
|
|
285
|
+
|
|
286
|
+
const properRe = /\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3}\b/g;
|
|
287
|
+
for (const match of value.matchAll(properRe)) {
|
|
288
|
+
const candidate = match[0].trim();
|
|
289
|
+
const first = candidate.split(/\s+/)[0];
|
|
290
|
+
if (STOP_LOCATION_WORDS.has(first)) continue;
|
|
291
|
+
if (candidate.length < 4 || candidate.length > 60) continue;
|
|
292
|
+
if (/^(Wall|You|Today|Thanks|Please|Hi|Hello|Best|Review|Summary|Final|Action|Status)$/i.test(candidate)) continue;
|
|
293
|
+
addUnique(out, candidate);
|
|
294
|
+
}
|
|
295
|
+
return out;
|
|
296
|
+
}
|
|
297
|
+
|
|
298
|
+
function findPrepositionalLocation(text) {
|
|
299
|
+
const value = stripUrls(String(text || ''));
|
|
300
|
+
const patterns = [
|
|
301
|
+
/\b(?:in|near|around|for|to)\s+([A-Z][A-Za-z]+(?:\s+[A-Z][A-Za-z]+){0,3}|[A-Z]{3})\b(?=[^\n]{0,80}\b(?:hotel|lodging|airbnb|vrbo|stay|recommend|trip|travel|booking|reservation)\b)/i,
|
|
302
|
+
/\b(?:hotel|lodging|airbnb|vrbo|stay|recommend|trip|travel|booking|reservation)\b[^\n]{0,80}\b(?:in|near|around|for|to)\s+([A-Z][A-Za-z]+(?:\s+[A-Z][A-Za-z]+){0,3}|[A-Z]{3})\b/i,
|
|
303
|
+
/\b([A-Z][A-Za-z]+(?:\s+[A-Z][A-Za-z]+){0,3}|[A-Z]{3})\s+(?:hotel|hotels|lodging|airbnb|vrbo|stay|trip|travel|booking|reservation|recommendations?)\b/i,
|
|
304
|
+
];
|
|
305
|
+
for (const pattern of patterns) {
|
|
306
|
+
const match = value.match(pattern);
|
|
307
|
+
if (match?.[1]) {
|
|
308
|
+
const location = cleanLocation(match[1]);
|
|
309
|
+
if (isLikelyLocationCandidate(location)) return location;
|
|
310
|
+
}
|
|
311
|
+
}
|
|
312
|
+
return '';
|
|
313
|
+
}
|
|
314
|
+
|
|
315
|
+
function appearsNearContext(text, location, contextRe) {
|
|
316
|
+
if (!location) return false;
|
|
317
|
+
const value = String(text || '');
|
|
318
|
+
const idx = value.toLowerCase().indexOf(location.toLowerCase());
|
|
319
|
+
if (idx < 0) return false;
|
|
320
|
+
const window = value.slice(Math.max(0, idx - 100), Math.min(value.length, idx + location.length + 100));
|
|
321
|
+
return contextRe.test(window);
|
|
322
|
+
}
|
|
323
|
+
|
|
324
|
+
function cleanLocation(value) {
|
|
325
|
+
const clean = String(value || '')
|
|
326
|
+
.replace(/[.,;:!?)}\]]+$/g, '')
|
|
327
|
+
.trim();
|
|
328
|
+
const parts = clean.split(/\s+/).filter(Boolean);
|
|
329
|
+
while (parts.length > 1 && STOP_LOCATION_WORDS.has(parts[0])) parts.shift();
|
|
330
|
+
if (parts.length === 1 && STOP_LOCATION_WORDS.has(parts[0])) return '';
|
|
331
|
+
return parts.join(' ');
|
|
332
|
+
}
|
|
333
|
+
|
|
334
|
+
function isLikelyLocationCandidate(value) {
|
|
335
|
+
const clean = String(value || '').trim();
|
|
336
|
+
if (!clean) return false;
|
|
337
|
+
if (/^[A-Z]{3}$/.test(clean)) return true;
|
|
338
|
+
return /^[A-Z]/.test(clean) && !STOP_LOCATION_WORDS.has(clean.split(/\s+/)[0]);
|
|
339
|
+
}
|
|
340
|
+
|
|
341
|
+
function stripUrls(value) {
|
|
342
|
+
return String(value || '').replace(/https?:\/\/\S+/g, ' ');
|
|
343
|
+
}
|
|
344
|
+
|
|
345
|
+
function extractDates(text) {
|
|
346
|
+
const value = String(text || '');
|
|
347
|
+
const out = [];
|
|
348
|
+
for (const match of value.matchAll(/\b20\d{2}-\d{2}-\d{2}\b/g)) addUnique(out, match[0]);
|
|
349
|
+
for (const match of value.matchAll(/\b(?:Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Sept|Oct|Nov|Dec)[a-z]*\.?\s+\d{1,2}(?:\s*(?:-|to|through|–)\s*(?:(?:Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Sept|Oct|Nov|Dec)[a-z]*\.?\s*)?\d{1,2})?(?:,\s*20\d{2})?\b/gi)) {
|
|
350
|
+
addUnique(out, match[0]);
|
|
351
|
+
}
|
|
352
|
+
return out.slice(-8);
|
|
353
|
+
}
|
|
354
|
+
|
|
355
|
+
function extractTravelerSummary(text) {
|
|
356
|
+
const value = String(text || '');
|
|
357
|
+
const parts = [];
|
|
358
|
+
const adults = value.match(/\b(\d+)\s+adults?\b/i);
|
|
359
|
+
if (adults) parts.push(`${adults[1]} adult${adults[1] === '1' ? '' : 's'}`);
|
|
360
|
+
const children = value.match(/\b(\d+)\s+(?:kids?|children|child)\b/i);
|
|
361
|
+
if (children) parts.push(`${children[1]} child${children[1] === '1' ? '' : 'ren'}`);
|
|
362
|
+
const age = value.match(/\b(\d{1,2})\s*years?\s*old\b/i);
|
|
363
|
+
if (age && !children) parts.push(`${age[1]}-year-old traveler`);
|
|
364
|
+
return parts.join(' + ');
|
|
365
|
+
}
|
|
366
|
+
|
|
367
|
+
function extractEmails(text) {
|
|
368
|
+
return [...new Set(String(text || '').match(/[A-Z0-9._%+-]+@[A-Z0-9.-]+\.[A-Z]{2,}/gi) || [])];
|
|
369
|
+
}
|
|
370
|
+
|
|
371
|
+
function collectPreviousRecommendations(messages) {
|
|
372
|
+
const names = [];
|
|
373
|
+
for (const msg of messages) {
|
|
374
|
+
if (msg.role !== 'assistant') continue;
|
|
375
|
+
for (const name of extractRecommendationNames(msg.content)) addUnique(names, name);
|
|
376
|
+
}
|
|
377
|
+
return names.slice(-MAX_RECOMMENDATION_NAMES);
|
|
378
|
+
}
|
|
379
|
+
|
|
380
|
+
function extractRecommendationNames(text) {
|
|
381
|
+
const out = [];
|
|
382
|
+
const value = String(text || '');
|
|
383
|
+
for (const match of value.matchAll(/\*\*([^*\n]{3,80})\*\*/g)) {
|
|
384
|
+
const name = cleanRecommendationName(match[1]);
|
|
385
|
+
if (name) addUnique(out, name);
|
|
386
|
+
}
|
|
387
|
+
for (const line of value.split('\n')) {
|
|
388
|
+
const match = line.match(/^\s*(?:[-*]|\d+[.)])\s+(?:\*\*)?([^:*—-][^:—\n]{2,80}?)(?:\*\*)?(?:\s*[:—-]|\s+\(|$)/);
|
|
389
|
+
if (!match) continue;
|
|
390
|
+
const name = cleanRecommendationName(match[1]);
|
|
391
|
+
if (name) addUnique(out, name);
|
|
392
|
+
}
|
|
393
|
+
return out.slice(0, MAX_RECOMMENDATION_NAMES);
|
|
394
|
+
}
|
|
395
|
+
|
|
396
|
+
function cleanRecommendationName(value) {
|
|
397
|
+
const name = String(value || '')
|
|
398
|
+
.replace(/^[^\wA-Z]+/, '')
|
|
399
|
+
.replace(/\s+/g, ' ')
|
|
400
|
+
.trim();
|
|
401
|
+
if (!name || name.length < 3 || name.length > 80) return '';
|
|
402
|
+
if (/^(why|pros|cons|price|location|recommendation|option|summary|final|notes?|action|status|detail|field)$/i.test(name)) return '';
|
|
403
|
+
if (!/[A-Z]/.test(name)) return '';
|
|
404
|
+
return name;
|
|
405
|
+
}
|
|
406
|
+
|
|
407
|
+
function addRecentCorrection(frame, text) {
|
|
408
|
+
const sentence = String(text || '').replace(/\s+/g, ' ').trim().slice(0, 240);
|
|
409
|
+
if (!sentence) return;
|
|
410
|
+
addUnique(frame.recentCorrections, sentence);
|
|
411
|
+
if (frame.recentCorrections.length > 5) frame.recentCorrections = frame.recentCorrections.slice(-5);
|
|
412
|
+
}
|
|
413
|
+
|
|
414
|
+
function scoreFrameConfidence(frame) {
|
|
415
|
+
let score = 0.35;
|
|
416
|
+
const semanticScopeKeys = Object.keys(frame.scope || {}).filter((key) => key !== 'projectPath');
|
|
417
|
+
if (frame.domain && frame.domain !== 'general') score += 0.15;
|
|
418
|
+
if (frame.activeGoal && frame.activeGoal !== 'answer current user request') score += 0.1;
|
|
419
|
+
if (semanticScopeKeys.length > 0) score += 0.15;
|
|
420
|
+
if ((frame.constraints || []).length > 0) score += 0.1;
|
|
421
|
+
if (frame.domain === 'coding' && frame.scope?.projectPath) score += 0.1;
|
|
422
|
+
if (frame.scope?.lodgingLocation || frame.scope?.destination) score += 0.1;
|
|
423
|
+
return Math.max(0.1, Math.min(0.95, Number(score.toFixed(2))));
|
|
424
|
+
}
|
|
425
|
+
|
|
426
|
+
function pruneFrame(frame) {
|
|
427
|
+
const out = JSON.parse(JSON.stringify(frame || {}));
|
|
428
|
+
out.constraints = (out.constraints || []).filter(Boolean).slice(-16);
|
|
429
|
+
out.recentCorrections = (out.recentCorrections || []).filter(Boolean).slice(-5);
|
|
430
|
+
out.inactiveRelatedFrames = (out.inactiveRelatedFrames || []).filter(Boolean).slice(-8);
|
|
431
|
+
out.previousRecommendations = (out.previousRecommendations || []).filter(Boolean).slice(0, MAX_RECOMMENDATION_NAMES);
|
|
432
|
+
out.evidenceRefs = (out.evidenceRefs || []).slice(-8);
|
|
433
|
+
return out;
|
|
434
|
+
}
|
|
435
|
+
|
|
436
|
+
function validatorsForFrame(frame) {
|
|
437
|
+
const validators = [];
|
|
438
|
+
if (frame?.domain === 'travel' || frame?.activeGoal?.includes('recommend') || frame?.scope?.lodgingLocation) {
|
|
439
|
+
validators.push('travel_recommendation_scope');
|
|
440
|
+
}
|
|
441
|
+
if (frame?.domain === 'email') validators.push('email_action_scope');
|
|
442
|
+
if (frame?.domain === 'calendar') validators.push('calendar_action_scope');
|
|
443
|
+
if (frame?.domain === 'coding') validators.push('coding_completion_scope');
|
|
444
|
+
return validators;
|
|
445
|
+
}
|
|
446
|
+
|
|
447
|
+
function buildConversationFramePrompt(frame) {
|
|
448
|
+
if (!frame || frame.confidence < 0.45) return '';
|
|
449
|
+
const lines = [];
|
|
450
|
+
lines.push('## Active Conversation Frame');
|
|
451
|
+
lines.push('This frame is the source of truth for the current task. The latest user correction wins over older assistant suggestions, memories, artifacts, and neighboring CTM tabs.');
|
|
452
|
+
lines.push(`Domain: ${frame.domain || 'general'}`);
|
|
453
|
+
lines.push(`Active goal: ${frame.activeGoal || 'answer current user request'}`);
|
|
454
|
+
const scopeLines = Object.entries(frame.scope || {})
|
|
455
|
+
.filter(([, value]) => value != null && value !== '' && (!Array.isArray(value) || value.length))
|
|
456
|
+
.map(([key, value]) => `- ${key}: ${Array.isArray(value) ? value.join(', ') : value}`);
|
|
457
|
+
if (scopeLines.length) lines.push(`Scope:\n${scopeLines.join('\n')}`);
|
|
458
|
+
if (frame.constraints?.length) lines.push(`Constraints:\n${frame.constraints.map(c => `- ${c}`).join('\n')}`);
|
|
459
|
+
if (frame.recentCorrections?.length) lines.push(`Recent user corrections:\n${frame.recentCorrections.map(c => `- ${c}`).join('\n')}`);
|
|
460
|
+
if (frame.inactiveRelatedFrames?.length) lines.push(`Related but inactive context:\n${frame.inactiveRelatedFrames.map(c => `- ${c}`).join('\n')}`);
|
|
461
|
+
if (frame.previousRecommendations?.length) lines.push(`Previously recommended candidates to avoid repeating unless justified:\n${frame.previousRecommendations.map(c => `- ${c}`).join('\n')}`);
|
|
462
|
+
lines.push('Before finalizing, verify the answer matches this frame. If it conflicts, repair the answer rather than claiming success.');
|
|
463
|
+
return `\n\n${lines.join('\n')}`;
|
|
464
|
+
}
|
|
465
|
+
|
|
466
|
+
function validateFrameAnswer(text, frame) {
|
|
467
|
+
const violations = [];
|
|
468
|
+
const value = String(text || '');
|
|
469
|
+
if (!value.trim() || !frame) return { ok: true, violations };
|
|
470
|
+
|
|
471
|
+
if (validatorsForFrame(frame).includes('travel_recommendation_scope')) {
|
|
472
|
+
validateTravelRecommendationAnswer(value, frame, violations);
|
|
473
|
+
}
|
|
474
|
+
if (validatorsForFrame(frame).includes('email_action_scope')) {
|
|
475
|
+
validateEmailActionAnswer(value, frame, violations);
|
|
476
|
+
}
|
|
477
|
+
if (validatorsForFrame(frame).includes('calendar_action_scope')) {
|
|
478
|
+
validateCalendarActionAnswer(value, frame, violations);
|
|
479
|
+
}
|
|
480
|
+
if (validatorsForFrame(frame).includes('coding_completion_scope')) {
|
|
481
|
+
validateCodingAnswer(value, frame, violations);
|
|
482
|
+
}
|
|
483
|
+
|
|
484
|
+
return { ok: violations.length === 0, violations };
|
|
485
|
+
}
|
|
486
|
+
|
|
487
|
+
function validateTravelRecommendationAnswer(text, frame, violations) {
|
|
488
|
+
const activeLocation = frame.scope?.lodgingLocation || frame.scope?.destination || '';
|
|
489
|
+
const recommendationLike = RECOMMENDATION_RE.test(text) || LODGING_RE.test(text);
|
|
490
|
+
if (activeLocation && recommendationLike) {
|
|
491
|
+
const hasActive = containsComparable(text, activeLocation);
|
|
492
|
+
const inactiveHit = (frame.inactiveRelatedFrames || []).find((loc) => loc && containsComparable(text, loc));
|
|
493
|
+
if (inactiveHit && !hasActive) {
|
|
494
|
+
violations.push({
|
|
495
|
+
code: 'scope_drift',
|
|
496
|
+
message: `Answer appears to use inactive location "${inactiveHit}" instead of active location "${activeLocation}".`,
|
|
497
|
+
});
|
|
498
|
+
}
|
|
499
|
+
}
|
|
500
|
+
if ((frame.constraints || []).includes('include_supply_type:alternative_rentals')) {
|
|
501
|
+
if (!/\b(airbnb|vrbo|vacation rental|short[- ]term rental|home rental|condo|rental home|rental house|apartment rental)\b/i.test(text)) {
|
|
502
|
+
violations.push({
|
|
503
|
+
code: 'missing_allowed_supply_type',
|
|
504
|
+
message: 'User explicitly allowed Airbnb/alternative rentals, but the answer does not include that supply type.',
|
|
505
|
+
});
|
|
506
|
+
}
|
|
507
|
+
}
|
|
508
|
+
if ((frame.constraints || []).includes('requires_novel_recommendations')) {
|
|
509
|
+
const previous = (frame.previousRecommendations || []).filter(Boolean);
|
|
510
|
+
if (previous.length > 0) {
|
|
511
|
+
const repeated = previous.filter((name) => containsComparable(text, name));
|
|
512
|
+
const currentNames = extractRecommendationNames(text);
|
|
513
|
+
const newNames = currentNames.filter((name) => !previous.some((prev) => normalizeComparable(prev) === normalizeComparable(name)));
|
|
514
|
+
if (repeated.length >= Math.min(2, previous.length) && newNames.length < 2) {
|
|
515
|
+
violations.push({
|
|
516
|
+
code: 'not_novel_recommendations',
|
|
517
|
+
message: `User asked for changed recommendations, but the answer repeats prior candidates: ${repeated.slice(0, 4).join(', ')}.`,
|
|
518
|
+
});
|
|
519
|
+
}
|
|
520
|
+
}
|
|
521
|
+
}
|
|
522
|
+
if ((frame.constraints || []).includes('not_brand_limited:hilton')) {
|
|
523
|
+
const mentionsHilton = /\bhilton\b/i.test(text);
|
|
524
|
+
const hasNonHiltonAlternative = /\b(airbnb|vrbo|marriott|hyatt|ihg|independent|boutique|vacation rental|condo|rental)\b/i.test(text);
|
|
525
|
+
if (mentionsHilton && !hasNonHiltonAlternative) {
|
|
526
|
+
violations.push({
|
|
527
|
+
code: 'brand_limited',
|
|
528
|
+
message: 'User said the answer does not have to be Hilton-only, but the response remains brand-limited.',
|
|
529
|
+
});
|
|
530
|
+
}
|
|
531
|
+
}
|
|
532
|
+
}
|
|
533
|
+
|
|
534
|
+
function validateEmailActionAnswer(text, frame, violations) {
|
|
535
|
+
if (!(frame.constraints || []).includes('email_action:reply_to_original_thread')) return;
|
|
536
|
+
if (/\b(new email|separate email|fresh email|new thread)\b/i.test(text) && !/\boriginal thread|same thread|reply\b/i.test(text)) {
|
|
537
|
+
violations.push({
|
|
538
|
+
code: 'email_thread_drift',
|
|
539
|
+
message: 'User asked to reply/respond to the original email thread, but the answer describes a new email/thread.',
|
|
540
|
+
});
|
|
541
|
+
}
|
|
542
|
+
}
|
|
543
|
+
|
|
544
|
+
function validateCalendarActionAnswer(text, frame, violations) {
|
|
545
|
+
const calendars = frame.scope?.calendars || [];
|
|
546
|
+
if (calendars.length < 2) return;
|
|
547
|
+
for (const cal of calendars) {
|
|
548
|
+
if (!new RegExp(`\\b${escapeRegExp(cal)}\\b`, 'i').test(text)) {
|
|
549
|
+
violations.push({
|
|
550
|
+
code: 'calendar_target_missing',
|
|
551
|
+
message: `User requested calendar target "${cal}", but the final answer does not mention it.`,
|
|
552
|
+
});
|
|
553
|
+
}
|
|
554
|
+
}
|
|
555
|
+
}
|
|
556
|
+
|
|
557
|
+
function validateCodingAnswer(text, frame, violations) {
|
|
558
|
+
if (!(frame.constraints || []).includes('verification_required')) return;
|
|
559
|
+
const claimsCompletion = /\b(done|complete|completed|fixed|implemented)\b/i.test(text);
|
|
560
|
+
const mentionsVerificationEvidence = /\b(test(?:ed|s)?|verified|validated|checked|ran|screenshot|browser_screenshot|check_url|build|lint|typecheck|evidence)\b/i.test(text);
|
|
561
|
+
if (claimsCompletion && !mentionsVerificationEvidence) {
|
|
562
|
+
violations.push({
|
|
563
|
+
code: 'missing_verification_evidence',
|
|
564
|
+
message: 'User asked for verification, but the final completion claim does not mention validation evidence.',
|
|
565
|
+
});
|
|
566
|
+
}
|
|
567
|
+
}
|
|
568
|
+
|
|
569
|
+
function buildFrameRepairPrompt({ frame, violations, originalUserMessage } = {}) {
|
|
570
|
+
const issueLines = (violations || []).map((v) => `- ${v.code}: ${v.message}`).join('\n') || '- frame_mismatch';
|
|
571
|
+
const frameJson = JSON.stringify(frame || {}, null, 2);
|
|
572
|
+
return [
|
|
573
|
+
'Your previous final answer failed Wall-E active-frame validation.',
|
|
574
|
+
'Rewrite the final answer only. Do not mention this validator or internal frame.',
|
|
575
|
+
'The latest user message and Active Conversation Frame are authoritative over older assistant replies, memories, artifacts, and neighboring sessions.',
|
|
576
|
+
'',
|
|
577
|
+
`Latest user message:\n${String(originalUserMessage || '').slice(0, 1200)}`,
|
|
578
|
+
'',
|
|
579
|
+
`Active Conversation Frame:\n${frameJson.slice(0, 4000)}`,
|
|
580
|
+
'',
|
|
581
|
+
`Violations:\n${issueLines}`,
|
|
582
|
+
'',
|
|
583
|
+
'Return a corrected final answer that satisfies every violation. Keep the same response language as the latest user message.',
|
|
584
|
+
].join('\n');
|
|
585
|
+
}
|
|
586
|
+
|
|
587
|
+
function containsComparable(text, needle) {
|
|
588
|
+
const source = normalizeComparable(text);
|
|
589
|
+
const target = normalizeComparable(needle);
|
|
590
|
+
if (!source || !target) return false;
|
|
591
|
+
if (source.includes(target)) return true;
|
|
592
|
+
if (/^[a-z]{3}$/.test(target)) return new RegExp(`\\b${escapeRegExp(target)}\\b`, 'i').test(String(text || ''));
|
|
593
|
+
return false;
|
|
594
|
+
}
|
|
595
|
+
|
|
596
|
+
function normalizeComparable(value) {
|
|
597
|
+
return String(value || '')
|
|
598
|
+
.toLowerCase()
|
|
599
|
+
.replace(/[^a-z0-9]+/g, ' ')
|
|
600
|
+
.trim();
|
|
601
|
+
}
|
|
602
|
+
|
|
603
|
+
function addUnique(list, value) {
|
|
604
|
+
const clean = typeof value === 'string' ? value.trim() : value;
|
|
605
|
+
if (!clean) return;
|
|
606
|
+
const key = normalizeComparable(clean);
|
|
607
|
+
if (!key) return;
|
|
608
|
+
if (!list.some((item) => normalizeComparable(item) === key)) list.push(clean);
|
|
609
|
+
}
|
|
610
|
+
|
|
611
|
+
function uniqueByLower(value, index, array) {
|
|
612
|
+
const key = normalizeComparable(value);
|
|
613
|
+
return key && array.findIndex((item) => normalizeComparable(item) === key) === index;
|
|
614
|
+
}
|
|
615
|
+
|
|
616
|
+
function escapeRegExp(value) {
|
|
617
|
+
return String(value || '').replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
|
|
618
|
+
}
|
|
619
|
+
|
|
620
|
+
module.exports = {
|
|
621
|
+
FRAME_SCHEMA_VERSION,
|
|
622
|
+
buildConversationRuntime,
|
|
623
|
+
buildConversationFramePrompt,
|
|
624
|
+
buildFrameRepairPrompt,
|
|
625
|
+
deriveConversationFrame,
|
|
626
|
+
extractRecommendationNames,
|
|
627
|
+
normalizeSemanticMessages,
|
|
628
|
+
parseSessionMetadata,
|
|
629
|
+
validateFrameAnswer,
|
|
630
|
+
};
|