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 +1,170 @@
|
|
|
1
|
+
'use strict';
|
|
2
|
+
|
|
3
|
+
const VERIFICATION_TOOL_NAME_RE = /(?:test|verify|screenshot|smoke|diagnostic|lint|build|check_url|url_check|pdf_info|pdf_render_pages|make_pdf)/i;
|
|
4
|
+
const VERIFICATION_COMMAND_RE = /\b(?:test|spec|lint|build|typecheck|tsc|pytest|jest|mocha|vitest|playwright|node\s+--(?:test|check)|npm\s+(?:test|run)|pnpm\s+(?:test|run)|yarn\s+(?:test|run)|git\s+diff\s+--check|curl\s+.*(?:localhost|127\.0\.0\.1)|check_url|pdfinfo|pdftoppm|pdftotext)\b/i;
|
|
5
|
+
|
|
6
|
+
function callName(call = {}) {
|
|
7
|
+
return String(call.name || call.tool || '');
|
|
8
|
+
}
|
|
9
|
+
|
|
10
|
+
function callInput(call = {}) {
|
|
11
|
+
return call.input || call.args || {};
|
|
12
|
+
}
|
|
13
|
+
|
|
14
|
+
function inputHashFor(call = {}) {
|
|
15
|
+
if (call.inputHash) return String(call.inputHash);
|
|
16
|
+
try {
|
|
17
|
+
return JSON.stringify(callInput(call)).slice(0, 500);
|
|
18
|
+
} catch {
|
|
19
|
+
return String(callInput(call)).slice(0, 500);
|
|
20
|
+
}
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
function hasResultEvidence(call = {}) {
|
|
24
|
+
return Object.prototype.hasOwnProperty.call(call, 'result')
|
|
25
|
+
|| Object.prototype.hasOwnProperty.call(call, 'ok')
|
|
26
|
+
|| Object.prototype.hasOwnProperty.call(call, 'success')
|
|
27
|
+
|| Object.prototype.hasOwnProperty.call(call, 'exitCode')
|
|
28
|
+
|| Boolean(call.error);
|
|
29
|
+
}
|
|
30
|
+
|
|
31
|
+
function toolResultSucceeded(call = {}) {
|
|
32
|
+
const result = Object.prototype.hasOwnProperty.call(call, 'result') ? call.result : call;
|
|
33
|
+
if (call.error === true || typeof call.error === 'string') return false;
|
|
34
|
+
if (!hasResultEvidence(call)) return false;
|
|
35
|
+
if (result && typeof result === 'object') {
|
|
36
|
+
if (result.error) return false;
|
|
37
|
+
if (result.ok === false || result.success === false) return false;
|
|
38
|
+
if (typeof result.exitCode === 'number' && result.exitCode !== 0) return false;
|
|
39
|
+
}
|
|
40
|
+
if (call.ok === false || call.success === false) return false;
|
|
41
|
+
if (typeof call.exitCode === 'number' && call.exitCode !== 0) return false;
|
|
42
|
+
return true;
|
|
43
|
+
}
|
|
44
|
+
|
|
45
|
+
function isVerificationToolCall(call = {}) {
|
|
46
|
+
const name = callName(call);
|
|
47
|
+
const input = String(inputHashFor(call));
|
|
48
|
+
if (name === 'browser_screenshot' || name === 'browser_smoke_test' || name === 'check_url' || name === 'url_check'
|
|
49
|
+
|| name === 'pdf_info' || name === 'pdf_render_pages' || name === 'make_pdf') return true;
|
|
50
|
+
if (name === 'run_shell') return VERIFICATION_COMMAND_RE.test(input);
|
|
51
|
+
return VERIFICATION_TOOL_NAME_RE.test(name);
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
function isSuccessfulVerificationToolCall(call = {}) {
|
|
55
|
+
return isVerificationToolCall(call) && toolResultSucceeded(call);
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
function hasVerificationEvidence(toolCallHistory = []) {
|
|
59
|
+
return (toolCallHistory || []).some(isSuccessfulVerificationToolCall);
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
function hasFailedVerificationAttempt(toolCallHistory = []) {
|
|
63
|
+
return (toolCallHistory || [])
|
|
64
|
+
.some((call) => isVerificationToolCall(call) && hasResultEvidence(call) && !toolResultSucceeded(call));
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
function normalizeToolCallEvidence(call = {}, result) {
|
|
68
|
+
const normalized = {
|
|
69
|
+
name: callName(call),
|
|
70
|
+
inputHash: inputHashFor(call),
|
|
71
|
+
};
|
|
72
|
+
if (call.id || call.toolCallId) normalized.id = call.id || call.toolCallId;
|
|
73
|
+
if (Object.prototype.hasOwnProperty.call(call, 'input')) normalized.input = call.input;
|
|
74
|
+
if (Object.prototype.hasOwnProperty.call(call, 'args')) normalized.args = call.args;
|
|
75
|
+
if (Object.prototype.hasOwnProperty.call(call, 'error')) normalized.error = call.error;
|
|
76
|
+
if (arguments.length > 1) normalized.result = result;
|
|
77
|
+
else if (Object.prototype.hasOwnProperty.call(call, 'result')) normalized.result = call.result;
|
|
78
|
+
if (hasResultEvidence(normalized)) normalized.ok = toolResultSucceeded(normalized);
|
|
79
|
+
return normalized;
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
function hasSuccessfulToolEvidence(toolCallHistory = [], toolName = '') {
|
|
83
|
+
const expected = String(toolName || '');
|
|
84
|
+
if (!expected) return false;
|
|
85
|
+
return (toolCallHistory || []).some((call) => callName(call) === expected && toolResultSucceeded(call));
|
|
86
|
+
}
|
|
87
|
+
|
|
88
|
+
function collectToolEvidence(toolCallHistory = [], toolName = '') {
|
|
89
|
+
const expected = String(toolName || '');
|
|
90
|
+
return (toolCallHistory || [])
|
|
91
|
+
.filter((call) => callName(call) === expected)
|
|
92
|
+
.map((call) => ({
|
|
93
|
+
name: callName(call),
|
|
94
|
+
ok: toolResultSucceeded(call),
|
|
95
|
+
result: call.result && typeof call.result === 'object' ? call.result : call,
|
|
96
|
+
}));
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
function buildAcceptanceContract({
|
|
100
|
+
request = '',
|
|
101
|
+
subtask = {},
|
|
102
|
+
changedFiles = [],
|
|
103
|
+
frontend = false,
|
|
104
|
+
requiresFileChanges = false,
|
|
105
|
+
} = {}) {
|
|
106
|
+
const explicit = subtask.acceptance && typeof subtask.acceptance === 'object' ? subtask.acceptance : {};
|
|
107
|
+
const text = `${request}\n${subtask.title || ''}\n${subtask.prompt || ''}`;
|
|
108
|
+
const frontendTask = frontend || /\b(?:website|web\s?page|frontend|ui|ux|html|css|javascript|responsive|layout|browser)\b/i.test(text);
|
|
109
|
+
const validators = new Set(Array.isArray(explicit.validators) ? explicit.validators : []);
|
|
110
|
+
if (frontendTask) {
|
|
111
|
+
validators.add('frontend.static_contract');
|
|
112
|
+
validators.add('frontend.screenshot_evidence');
|
|
113
|
+
validators.add('frontend.browser_runtime');
|
|
114
|
+
}
|
|
115
|
+
const writePolicy = explicit.write_policy || explicit.writePolicy || (
|
|
116
|
+
requiresFileChanges ? 'must-write' : 'may-write'
|
|
117
|
+
);
|
|
118
|
+
return {
|
|
119
|
+
taskKind: explicit.task_kind || explicit.taskKind || (frontendTask ? 'frontend-ui' : 'code'),
|
|
120
|
+
writePolicy,
|
|
121
|
+
validators: [...validators],
|
|
122
|
+
requiresFrontendValidation: frontendTask,
|
|
123
|
+
changedFiles: [...new Set(changedFiles || [])],
|
|
124
|
+
recoveryPolicy: explicit.recovery_policy || explicit.recoveryPolicy || 'retry-with-validator-failure',
|
|
125
|
+
};
|
|
126
|
+
}
|
|
127
|
+
|
|
128
|
+
function validatorFailure(name, message, detail = {}) {
|
|
129
|
+
return {
|
|
130
|
+
name,
|
|
131
|
+
ok: false,
|
|
132
|
+
message,
|
|
133
|
+
detail,
|
|
134
|
+
};
|
|
135
|
+
}
|
|
136
|
+
|
|
137
|
+
function validatorPass(name, message, detail = {}) {
|
|
138
|
+
return {
|
|
139
|
+
name,
|
|
140
|
+
ok: true,
|
|
141
|
+
message,
|
|
142
|
+
detail,
|
|
143
|
+
};
|
|
144
|
+
}
|
|
145
|
+
|
|
146
|
+
function summarizeValidatorFailures(report = {}) {
|
|
147
|
+
return (report.validators || [])
|
|
148
|
+
.filter((item) => item && item.ok === false)
|
|
149
|
+
.map((item) => item.message || item.name)
|
|
150
|
+
.filter(Boolean);
|
|
151
|
+
}
|
|
152
|
+
|
|
153
|
+
module.exports = {
|
|
154
|
+
callName,
|
|
155
|
+
callInput,
|
|
156
|
+
inputHashFor,
|
|
157
|
+
hasResultEvidence,
|
|
158
|
+
toolResultSucceeded,
|
|
159
|
+
isVerificationToolCall,
|
|
160
|
+
isSuccessfulVerificationToolCall,
|
|
161
|
+
hasVerificationEvidence,
|
|
162
|
+
hasFailedVerificationAttempt,
|
|
163
|
+
normalizeToolCallEvidence,
|
|
164
|
+
hasSuccessfulToolEvidence,
|
|
165
|
+
collectToolEvidence,
|
|
166
|
+
buildAcceptanceContract,
|
|
167
|
+
validatorFailure,
|
|
168
|
+
validatorPass,
|
|
169
|
+
summarizeValidatorFailures,
|
|
170
|
+
};
|
|
@@ -64,6 +64,72 @@ class ArtifactStore {
|
|
|
64
64
|
metadata: safeJsonValue(metadata),
|
|
65
65
|
};
|
|
66
66
|
}
|
|
67
|
+
|
|
68
|
+
storeArtifact({
|
|
69
|
+
sessionId = '',
|
|
70
|
+
toolCallId = '',
|
|
71
|
+
toolName = '',
|
|
72
|
+
kind = 'artifact',
|
|
73
|
+
sourcePath = '',
|
|
74
|
+
path: artifactPath = '',
|
|
75
|
+
content = null,
|
|
76
|
+
mimeType = 'application/octet-stream',
|
|
77
|
+
bytes = null,
|
|
78
|
+
sha256 = '',
|
|
79
|
+
metadata = {},
|
|
80
|
+
} = {}) {
|
|
81
|
+
const inputPath = sourcePath || artifactPath || '';
|
|
82
|
+
let buffer = null;
|
|
83
|
+
let ext = extensionForArtifact(inputPath, mimeType);
|
|
84
|
+
if (content !== null && content !== undefined) {
|
|
85
|
+
buffer = Buffer.isBuffer(content) ? content : Buffer.from(String(content));
|
|
86
|
+
} else if (inputPath && fs.existsSync(inputPath) && fs.statSync(inputPath).isFile()) {
|
|
87
|
+
buffer = fs.readFileSync(inputPath);
|
|
88
|
+
if (!mimeType || mimeType === 'application/octet-stream') mimeType = mimeTypeForPath(inputPath, mimeType);
|
|
89
|
+
} else if (Number.isFinite(Number(bytes)) && sha256) {
|
|
90
|
+
return {
|
|
91
|
+
artifactId: `art_${String(sha256).slice(0, 16)}`,
|
|
92
|
+
kind: String(kind || 'artifact'),
|
|
93
|
+
sessionId: String(sessionId || ''),
|
|
94
|
+
toolCallId: String(toolCallId || ''),
|
|
95
|
+
toolName: String(toolName || ''),
|
|
96
|
+
path: String(inputPath || ''),
|
|
97
|
+
bytes: Number(bytes),
|
|
98
|
+
sha256: String(sha256),
|
|
99
|
+
mimeType,
|
|
100
|
+
createdAt: this.now(),
|
|
101
|
+
metadata: safeJsonValue(metadata),
|
|
102
|
+
};
|
|
103
|
+
} else {
|
|
104
|
+
throw new Error('storeArtifact requires content, an existing sourcePath/path, or bytes+sha256 metadata');
|
|
105
|
+
}
|
|
106
|
+
|
|
107
|
+
const realBytes = buffer.length;
|
|
108
|
+
const realSha = sha256 || crypto.createHash('sha256').update(buffer).digest('hex');
|
|
109
|
+
const artifactId = `art_${realSha.slice(0, 16)}`;
|
|
110
|
+
const dir = path.join(this.artifactsDir, safeName(sessionId || 'session'));
|
|
111
|
+
fs.mkdirSync(dir, { recursive: true });
|
|
112
|
+
const filePath = path.join(dir, `${safeName(toolCallId || toolName || kind || artifactId)}-${realSha.slice(0, 12)}.${ext}`);
|
|
113
|
+
if (inputPath && path.resolve(inputPath) === path.resolve(filePath)) {
|
|
114
|
+
// Already in the artifact store.
|
|
115
|
+
} else {
|
|
116
|
+
fs.writeFileSync(filePath, buffer);
|
|
117
|
+
}
|
|
118
|
+
return {
|
|
119
|
+
artifactId,
|
|
120
|
+
kind: String(kind || 'artifact'),
|
|
121
|
+
sessionId: String(sessionId || ''),
|
|
122
|
+
toolCallId: String(toolCallId || ''),
|
|
123
|
+
toolName: String(toolName || ''),
|
|
124
|
+
path: filePath,
|
|
125
|
+
originalPath: inputPath ? String(inputPath) : '',
|
|
126
|
+
bytes: realBytes,
|
|
127
|
+
sha256: realSha,
|
|
128
|
+
mimeType,
|
|
129
|
+
createdAt: this.now(),
|
|
130
|
+
metadata: safeJsonValue(metadata),
|
|
131
|
+
};
|
|
132
|
+
}
|
|
67
133
|
}
|
|
68
134
|
|
|
69
135
|
function defaultArtifactsDir(env = process.env) {
|
|
@@ -109,12 +175,16 @@ function publicArtifactRef(artifact) {
|
|
|
109
175
|
return {
|
|
110
176
|
artifactId: artifact.artifactId,
|
|
111
177
|
kind: artifact.kind,
|
|
178
|
+
sessionId: artifact.sessionId,
|
|
112
179
|
path: artifact.path,
|
|
180
|
+
originalPath: artifact.originalPath,
|
|
113
181
|
bytes: artifact.bytes,
|
|
114
182
|
sha256: artifact.sha256,
|
|
115
183
|
mimeType: artifact.mimeType,
|
|
116
184
|
toolCallId: artifact.toolCallId,
|
|
117
185
|
toolName: artifact.toolName,
|
|
186
|
+
createdAt: artifact.createdAt,
|
|
187
|
+
metadata: safeJsonValue(artifact.metadata || {}),
|
|
118
188
|
};
|
|
119
189
|
}
|
|
120
190
|
|
|
@@ -139,6 +209,29 @@ function safeName(value) {
|
|
|
139
209
|
.slice(0, 80) || 'artifact';
|
|
140
210
|
}
|
|
141
211
|
|
|
212
|
+
function extensionForArtifact(filePath, mimeType) {
|
|
213
|
+
const ext = path.extname(String(filePath || '')).replace(/^\./, '').toLowerCase();
|
|
214
|
+
if (ext) return safeName(ext).slice(0, 12) || 'bin';
|
|
215
|
+
if (/pdf/i.test(mimeType)) return 'pdf';
|
|
216
|
+
if (/png/i.test(mimeType)) return 'png';
|
|
217
|
+
if (/jpe?g/i.test(mimeType)) return 'jpg';
|
|
218
|
+
if (/html/i.test(mimeType)) return 'html';
|
|
219
|
+
if (/json/i.test(mimeType)) return 'json';
|
|
220
|
+
if (/text/i.test(mimeType)) return 'txt';
|
|
221
|
+
return 'bin';
|
|
222
|
+
}
|
|
223
|
+
|
|
224
|
+
function mimeTypeForPath(filePath, fallback) {
|
|
225
|
+
const ext = path.extname(String(filePath || '')).toLowerCase();
|
|
226
|
+
if (ext === '.pdf') return 'application/pdf';
|
|
227
|
+
if (ext === '.png') return 'image/png';
|
|
228
|
+
if (ext === '.jpg' || ext === '.jpeg') return 'image/jpeg';
|
|
229
|
+
if (ext === '.html' || ext === '.htm') return 'text/html';
|
|
230
|
+
if (ext === '.json') return 'application/json';
|
|
231
|
+
if (ext === '.txt' || ext === '.md') return 'text/plain';
|
|
232
|
+
return fallback || 'application/octet-stream';
|
|
233
|
+
}
|
|
234
|
+
|
|
142
235
|
function safeJsonValue(value) {
|
|
143
236
|
try {
|
|
144
237
|
return JSON.parse(JSON.stringify(value));
|
|
@@ -0,0 +1,120 @@
|
|
|
1
|
+
'use strict';
|
|
2
|
+
|
|
3
|
+
const FRONTEND_EXT_RE = /\.(html?|css|scss|sass|less|jsx|tsx|vue|svelte|astro)$/i;
|
|
4
|
+
const PDF_EXT_RE = /\.pdf$/i;
|
|
5
|
+
|
|
6
|
+
const FRONTEND_KEYWORDS_RE = new RegExp([
|
|
7
|
+
'\\b(',
|
|
8
|
+
'website|webpage|web\\s?page|landing\\s?page|homepage|portfolio\\s?(site|website|page)?',
|
|
9
|
+
'|hero\\s?(section|banner|image)|nav(igation)?\\s?bar|gallery|lightbox|carousel',
|
|
10
|
+
'|UI|UX|user\\s?interface|user\\s?experience|front[\\s-]?end|frontend',
|
|
11
|
+
'|stylesheet|stylesheets|responsive\\s?(design|layout)|design\\s?system',
|
|
12
|
+
')\\b',
|
|
13
|
+
].join(''), 'i');
|
|
14
|
+
const FRONTEND_NEGATIVE_RE = /\b(schema|algorithm|protocol|api|grpc|sql|migration|pipeline|backend\s?service|microservice)\b/i;
|
|
15
|
+
|
|
16
|
+
const PDF_KEYWORDS_RE = /\b(?:pdf|portable document|document export|print(?:able)? document|export as pdf|make (?:a )?pdf|generate (?:a )?pdf|render (?:a )?pdf)\b/i;
|
|
17
|
+
const PDF_GENERATION_RE = /\b(?:make|generate|create|export|render|design|format|style|polish|typeset|print)\b[\s\S]{0,80}\bpdf\b|\bpdf\b[\s\S]{0,80}\b(?:make|generate|create|export|render|design|format|style|polish|typeset|print)\b/i;
|
|
18
|
+
const PDF_INGESTION_RE = /\b(?:read|inspect|summari[sz]e|analy[sz]e|extract|review)\b[\s\S]{0,80}\bpdf\b|\bpdf\b[\s\S]{0,80}\b(?:read|inspect|summari[sz]e|analy[sz]e|extract|review)\b/i;
|
|
19
|
+
|
|
20
|
+
const MANIFESTS = Object.freeze({
|
|
21
|
+
frontend_design: Object.freeze({
|
|
22
|
+
id: 'frontend_design',
|
|
23
|
+
label: 'Web/UI visual artifact work',
|
|
24
|
+
promptSections: ['frontend-design', 'browser-verification'],
|
|
25
|
+
tools: ['start_static_server', 'check_url', 'browser_screenshot'],
|
|
26
|
+
requiredArtifacts: ['screenshot.desktop', 'screenshot.mobile'],
|
|
27
|
+
completionGate: 'visual_artifact_verified',
|
|
28
|
+
}),
|
|
29
|
+
pdf_document: Object.freeze({
|
|
30
|
+
id: 'pdf_document',
|
|
31
|
+
label: 'PDF document generation and design',
|
|
32
|
+
promptSections: ['document-design', 'pdf-generation', 'pdf-verification'],
|
|
33
|
+
tools: ['make_pdf', 'pdf_info', 'pdf_render_pages', 'pdf_read_pages'],
|
|
34
|
+
requiredArtifacts: ['pdf.file', 'pdf.page_preview'],
|
|
35
|
+
completionGate: 'document_artifact_verified',
|
|
36
|
+
}),
|
|
37
|
+
pdf_ingestion: Object.freeze({
|
|
38
|
+
id: 'pdf_ingestion',
|
|
39
|
+
label: 'PDF ingestion and page inspection',
|
|
40
|
+
promptSections: ['pdf-ingestion'],
|
|
41
|
+
tools: ['pdf_info', 'pdf_read_pages', 'pdf_render_pages'],
|
|
42
|
+
requiredArtifacts: ['pdf.file'],
|
|
43
|
+
completionGate: 'source_document_verified',
|
|
44
|
+
}),
|
|
45
|
+
});
|
|
46
|
+
|
|
47
|
+
function routeArtifactCapabilities({ prompt = '', taskFileHints = [], projectInfo = null } = {}) {
|
|
48
|
+
const hints = Array.isArray(taskFileHints) ? taskFileHints.map(String) : [];
|
|
49
|
+
const text = String(prompt || '');
|
|
50
|
+
const capabilities = [];
|
|
51
|
+
|
|
52
|
+
if (detectFrontend(text, hints, projectInfo)) capabilities.push(MANIFESTS.frontend_design);
|
|
53
|
+
|
|
54
|
+
if (detectPdf(text, hints)) {
|
|
55
|
+
const mode = detectPdfMode(text, hints);
|
|
56
|
+
capabilities.push(mode === 'ingestion' ? MANIFESTS.pdf_ingestion : MANIFESTS.pdf_document);
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
return dedupeCapabilities(capabilities);
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
function detectFrontend(text, hints, projectInfo) {
|
|
63
|
+
if (hints.some((hint) => FRONTEND_EXT_RE.test(hint))) return true;
|
|
64
|
+
if (projectInfo?.framework && /\b(?:react|vue|svelte|astro|next|vite)\b/i.test(projectInfo.framework)) {
|
|
65
|
+
if (FRONTEND_KEYWORDS_RE.test(text)) return true;
|
|
66
|
+
}
|
|
67
|
+
if (!FRONTEND_KEYWORDS_RE.test(text)) return false;
|
|
68
|
+
if (FRONTEND_NEGATIVE_RE.test(text) && !hints.some((hint) => FRONTEND_EXT_RE.test(hint))) return false;
|
|
69
|
+
return true;
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
function detectPdf(text, hints) {
|
|
73
|
+
return hints.some((hint) => PDF_EXT_RE.test(hint)) || PDF_KEYWORDS_RE.test(text);
|
|
74
|
+
}
|
|
75
|
+
|
|
76
|
+
function detectPdfMode(text, hints) {
|
|
77
|
+
if (PDF_GENERATION_RE.test(text)) return 'generation';
|
|
78
|
+
if (hints.some((hint) => PDF_EXT_RE.test(hint)) || PDF_INGESTION_RE.test(text)) return 'ingestion';
|
|
79
|
+
return 'generation';
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
function buildArtifactCapabilityContext(capabilities = []) {
|
|
83
|
+
const routed = dedupeCapabilities(capabilities);
|
|
84
|
+
if (!routed.length) return '';
|
|
85
|
+
const lines = ['# Artifact Capability Routes'];
|
|
86
|
+
lines.push('This task was routed to typed artifact capabilities. Treat their tools and evidence gates as part of the completion contract, not optional suggestions.');
|
|
87
|
+
for (const capability of routed) {
|
|
88
|
+
lines.push('');
|
|
89
|
+
lines.push(`## ${capability.label}`);
|
|
90
|
+
lines.push(`- id: \`${capability.id}\``);
|
|
91
|
+
lines.push(`- use tools: ${capability.tools.map((tool) => `\`${tool}\``).join(', ')}`);
|
|
92
|
+
lines.push(`- required evidence before claiming success: ${capability.requiredArtifacts.map((item) => `\`${item}\``).join(', ')}`);
|
|
93
|
+
lines.push(`- completion gate: \`${capability.completionGate}\``);
|
|
94
|
+
}
|
|
95
|
+
lines.push('');
|
|
96
|
+
lines.push('For generated visual/document artifacts, create or verify the artifact with tools, then inspect the rendered output or report the exact tool-backed blocker.');
|
|
97
|
+
return lines.join('\n');
|
|
98
|
+
}
|
|
99
|
+
|
|
100
|
+
function dedupeCapabilities(capabilities = []) {
|
|
101
|
+
const seen = new Set();
|
|
102
|
+
const out = [];
|
|
103
|
+
for (const capability of capabilities || []) {
|
|
104
|
+
if (!capability?.id || seen.has(capability.id)) continue;
|
|
105
|
+
seen.add(capability.id);
|
|
106
|
+
out.push({ ...capability });
|
|
107
|
+
}
|
|
108
|
+
return out;
|
|
109
|
+
}
|
|
110
|
+
|
|
111
|
+
function hasCapability(capabilities, id) {
|
|
112
|
+
return (capabilities || []).some((capability) => capability?.id === id);
|
|
113
|
+
}
|
|
114
|
+
|
|
115
|
+
module.exports = {
|
|
116
|
+
MANIFESTS,
|
|
117
|
+
routeArtifactCapabilities,
|
|
118
|
+
buildArtifactCapabilityContext,
|
|
119
|
+
hasCapability,
|
|
120
|
+
};
|