@openllmsh/sdk 2.10.12-alpha.0 → 2.10.12-alpha.2
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 -7
- package/dist/client/chat.d.ts +38 -0
- package/dist/client/chat.js +77 -0
- package/dist/client/devices.d.ts +70 -0
- package/dist/client/devices.js +172 -0
- package/dist/client/http.d.ts +24 -0
- package/dist/client/http.js +34 -0
- package/dist/client/index.d.ts +43 -0
- package/dist/client/index.js +57 -0
- package/dist/client/relay-mux.d.ts +71 -0
- package/dist/client/relay-mux.js +394 -0
- package/dist/client/types.d.ts +47 -0
- package/dist/client/types.js +5 -0
- package/dist/client/vault.d.ts +68 -0
- package/dist/client/vault.js +113 -0
- package/dist/connect/chat.d.ts +24 -0
- package/dist/connect/chat.js +25 -0
- package/dist/connect/client.d.ts +50 -0
- package/dist/connect/client.js +141 -0
- package/dist/connect/config.d.ts +22 -0
- package/dist/connect/config.js +22 -0
- package/dist/connect/errors.d.ts +32 -0
- package/dist/connect/errors.js +85 -0
- package/{connect/index.ts → dist/connect/index.d.ts} +7 -7
- package/dist/connect/index.js +23 -0
- package/dist/connect/oauth.d.ts +51 -0
- package/dist/connect/oauth.js +150 -0
- package/dist/connect/routes.d.ts +38 -0
- package/dist/connect/routes.js +175 -0
- package/dist/connect/store.d.ts +26 -0
- package/dist/connect/store.js +65 -0
- package/dist/mcp/handle.d.ts +6 -0
- package/dist/mcp/handle.js +6 -0
- package/dist/mcp/index.d.ts +15 -0
- package/dist/mcp/index.js +10 -0
- package/dist/mcp/server.d.ts +84 -0
- package/dist/mcp/server.js +168 -0
- package/dist/mcp/tools/marketplace.d.ts +94 -0
- package/dist/mcp/tools/marketplace.js +303 -0
- package/dist/mcp/tools/memory-client.d.ts +109 -0
- package/dist/mcp/tools/memory-client.js +200 -0
- package/dist/mcp/tools/memory-config.d.ts +29 -0
- package/dist/mcp/tools/memory-config.js +12 -0
- package/dist/mcp/tools/memory.d.ts +92 -0
- package/dist/mcp/tools/memory.js +312 -0
- package/dist/mcp/tools/model-priority.d.ts +16 -0
- package/dist/mcp/tools/model-priority.js +87 -0
- package/dist/mcp/tools/openllm.d.ts +75 -0
- package/dist/mcp/tools/openllm.js +715 -0
- package/dist/mcp/types.d.ts +25 -0
- package/dist/mcp/types.js +7 -0
- package/dist/openapi/call.d.ts +26 -0
- package/dist/openapi/call.js +56 -0
- package/dist/openapi/index.d.ts +2 -0
- package/dist/openapi/index.js +2 -0
- package/dist/openapi/openapi.json +80213 -0
- package/dist/openapi/operations.d.ts +14 -0
- package/dist/openapi/operations.js +2060 -0
- package/dist/protocol/access.d.ts +81 -0
- package/dist/protocol/access.js +125 -0
- package/dist/protocol/agent-prompt-text.d.ts +15 -0
- package/{protocol/agent-prompt-text.ts → dist/protocol/agent-prompt-text.js} +9 -14
- package/dist/protocol/agent.d.ts +1487 -0
- package/dist/protocol/agent.js +787 -0
- package/dist/protocol/api-key-syntax.d.ts +58 -0
- package/dist/protocol/api-key-syntax.js +73 -0
- package/dist/protocol/audio.d.ts +18 -0
- package/dist/protocol/audio.js +21 -0
- package/dist/protocol/auth.d.ts +126 -0
- package/dist/protocol/auth.js +119 -0
- package/dist/protocol/changelog.d.ts +69 -0
- package/dist/protocol/changelog.js +107 -0
- package/dist/protocol/chat.d.ts +789 -0
- package/dist/protocol/chat.js +394 -0
- package/dist/protocol/client-registry.d.ts +151 -0
- package/dist/protocol/client-registry.js +238 -0
- package/dist/protocol/config.d.ts +508 -0
- package/dist/protocol/config.js +289 -0
- package/dist/protocol/control-channel.d.ts +102 -0
- package/{protocol/control-channel.ts → dist/protocol/control-channel.js} +33 -61
- package/dist/protocol/cooldown-reason.d.ts +30 -0
- package/dist/protocol/cooldown-reason.js +44 -0
- package/dist/protocol/cost.d.ts +9 -0
- package/dist/protocol/cost.js +8 -0
- package/dist/protocol/daemon-port.d.ts +2 -0
- package/dist/protocol/daemon-port.js +28 -0
- package/dist/protocol/daemon.d.ts +2034 -0
- package/dist/protocol/daemon.js +1286 -0
- package/dist/protocol/default-chains.d.ts +15 -0
- package/dist/protocol/default-chains.js +21 -0
- package/dist/protocol/device-state.d.ts +223 -0
- package/dist/protocol/device-state.js +52 -0
- package/dist/protocol/doctor-report-local.d.ts +105 -0
- package/{protocol/doctor-report-local.ts → dist/protocol/doctor-report-local.js} +92 -196
- package/dist/protocol/doctor-report-policy.d.ts +62 -0
- package/{protocol/doctor-report-policy.ts → dist/protocol/doctor-report-policy.js} +23 -76
- package/dist/protocol/doctor-report.d.ts +406 -0
- package/dist/protocol/doctor-report.js +645 -0
- package/dist/protocol/errors.d.ts +17 -0
- package/dist/protocol/errors.js +10 -0
- package/dist/protocol/execution-selection.d.ts +92 -0
- package/{protocol/execution-selection.ts → dist/protocol/execution-selection.js} +38 -63
- package/dist/protocol/hop-grant.d.ts +28 -0
- package/{protocol/hop-grant.ts → dist/protocol/hop-grant.js} +7 -13
- package/dist/protocol/image-edit-parse.d.ts +25 -0
- package/dist/protocol/image-edit-parse.js +196 -0
- package/dist/protocol/images.d.ts +288 -0
- package/dist/protocol/images.js +229 -0
- package/dist/protocol/index.d.ts +69 -0
- package/dist/protocol/index.js +74 -0
- package/dist/protocol/inference-headers.d.ts +3 -0
- package/dist/protocol/keepalive.d.ts +36 -0
- package/dist/protocol/keepalive.js +51 -0
- package/dist/protocol/local-auth.d.ts +125 -0
- package/dist/protocol/local-auth.js +227 -0
- package/dist/protocol/local-device-control.d.ts +108 -0
- package/{protocol/local-device-control.ts → dist/protocol/local-device-control.js} +40 -70
- package/dist/protocol/local-provider.d.ts +105 -0
- package/dist/protocol/local-provider.js +213 -0
- package/dist/protocol/local-session.d.ts +70 -0
- package/dist/protocol/local-session.js +87 -0
- package/dist/protocol/marketplace.d.ts +190 -0
- package/dist/protocol/marketplace.js +140 -0
- package/{protocol/media-attempt.ts → dist/protocol/media-attempt.d.ts} +11 -50
- package/dist/protocol/media-attempt.js +50 -0
- package/dist/protocol/media-bytes.d.ts +8 -0
- package/dist/protocol/media-bytes.js +21 -0
- package/dist/protocol/media-default.d.ts +56 -0
- package/dist/protocol/media-default.js +302 -0
- package/{protocol/media-dimensions.ts → dist/protocol/media-dimensions.d.ts} +11 -78
- package/dist/protocol/media-dimensions.js +90 -0
- package/dist/protocol/media-input.d.ts +20 -0
- package/dist/protocol/media-input.js +29 -0
- package/dist/protocol/media-model.d.ts +77 -0
- package/dist/protocol/media-model.js +89 -0
- package/{protocol/media-option-issue.ts → dist/protocol/media-option-issue.d.ts} +25 -53
- package/dist/protocol/media-option-issue.js +50 -0
- package/dist/protocol/media-types.d.ts +14 -0
- package/dist/protocol/media-types.js +61 -0
- package/dist/protocol/model-match.d.ts +254 -0
- package/dist/protocol/model-match.js +760 -0
- package/dist/protocol/model-settings.d.ts +153 -0
- package/dist/protocol/model-settings.js +246 -0
- package/dist/protocol/models.d.ts +1065 -0
- package/dist/protocol/models.js +638 -0
- package/dist/protocol/mux.d.ts +350 -0
- package/dist/protocol/mux.js +595 -0
- package/{protocol/plan-surface.ts → dist/protocol/plan-surface.d.ts} +12 -26
- package/dist/protocol/plan-surface.js +39 -0
- package/dist/protocol/provider-models.d.ts +8 -0
- package/{protocol/provider-models.ts → dist/protocol/provider-models.js} +7 -9
- package/dist/protocol/provider-status.d.ts +59 -0
- package/dist/protocol/provider-status.js +142 -0
- package/dist/protocol/provider-tier.d.ts +34 -0
- package/{protocol/provider-tier.ts → dist/protocol/provider-tier.js} +9 -34
- package/dist/protocol/provider-usage.d.ts +230 -0
- package/dist/protocol/provider-usage.js +214 -0
- package/dist/protocol/providers.d.ts +2496 -0
- package/dist/protocol/providers.js +723 -0
- package/dist/protocol/public-usage-index.d.ts +449 -0
- package/dist/protocol/public-usage-index.js +279 -0
- package/dist/protocol/realtime-generated.d.ts +29 -0
- package/{protocol/realtime-generated.ts → dist/protocol/realtime-generated.js} +3 -5
- package/dist/protocol/realtime.d.ts +241 -0
- package/{protocol/realtime.ts → dist/protocol/realtime.js} +88 -155
- package/dist/protocol/registry-pointer.d.ts +24 -0
- package/dist/protocol/registry-pointer.js +78 -0
- package/dist/protocol/relay.d.ts +1152 -0
- package/{protocol/relay.ts → dist/protocol/relay.js} +237 -370
- package/dist/protocol/replay-session.d.ts +4 -0
- package/{protocol/replay-session.ts → dist/protocol/replay-session.js} +1 -8
- package/dist/protocol/responses.d.ts +303 -0
- package/dist/protocol/responses.js +240 -0
- package/dist/protocol/semver.d.ts +59 -0
- package/dist/protocol/semver.js +171 -0
- package/{protocol/setup-mode.ts → dist/protocol/setup-mode.d.ts} +4 -12
- package/dist/protocol/setup-mode.js +18 -0
- package/dist/protocol/sha256-hex.d.ts +2 -0
- package/dist/protocol/sha256-hex.js +93 -0
- package/dist/protocol/stats.d.ts +711 -0
- package/dist/protocol/stats.js +353 -0
- package/dist/protocol/streaming.d.ts +16 -0
- package/dist/protocol/streaming.js +15 -0
- package/{protocol/subscription-daemon-floors.ts → dist/protocol/subscription-daemon-floors.d.ts} +8 -14
- package/dist/protocol/subscription-daemon-floors.js +4 -0
- package/dist/protocol/subscription-provider-compat.d.ts +44 -0
- package/dist/protocol/subscription-provider-compat.js +67 -0
- package/dist/protocol/subscription-provider.d.ts +15 -0
- package/dist/protocol/subscription-provider.js +17 -0
- package/dist/protocol/team-grant.d.ts +14 -0
- package/{protocol/team-grant.ts → dist/protocol/team-grant.js} +0 -3
- package/dist/protocol/team-shareable.d.ts +6 -0
- package/{protocol/team-shareable.ts → dist/protocol/team-shareable.js} +29 -33
- package/dist/protocol/transcription-encoding.d.ts +25 -0
- package/dist/protocol/transcription-encoding.js +47 -0
- package/{protocol/url-locality-dns.ts → dist/protocol/url-locality-dns.d.ts} +1 -23
- package/dist/protocol/url-locality-dns.js +33 -0
- package/dist/protocol/url-locality.d.ts +45 -0
- package/dist/protocol/url-locality.js +288 -0
- package/dist/protocol/usage-snapshots.d.ts +8 -0
- package/{protocol/usage-snapshots.ts → dist/protocol/usage-snapshots.js} +1 -2
- package/dist/protocol/vendor-cli-install.d.ts +20 -0
- package/{protocol/vendor-cli-install.ts → dist/protocol/vendor-cli-install.js} +10 -15
- package/dist/protocol/video-job-plan.d.ts +18 -0
- package/dist/protocol/video-job-plan.js +44 -0
- package/dist/protocol/videos.d.ts +107 -0
- package/dist/protocol/videos.js +154 -0
- package/dist/tunnel/channel-envelope.d.ts +53 -0
- package/dist/tunnel/channel-envelope.js +100 -0
- package/dist/tunnel/codec.d.ts +71 -0
- package/dist/tunnel/codec.js +151 -0
- package/dist/tunnel/device-anchor.d.ts +62 -0
- package/dist/tunnel/device-anchor.js +168 -0
- package/dist/tunnel/device-control.d.ts +50 -0
- package/dist/tunnel/device-control.js +88 -0
- package/dist/tunnel/device-grant.d.ts +91 -0
- package/dist/tunnel/device-grant.js +105 -0
- package/dist/tunnel/fleet.d.ts +28 -0
- package/dist/tunnel/fleet.js +41 -0
- package/dist/tunnel/index.d.ts +22 -0
- package/dist/tunnel/index.js +22 -0
- package/dist/tunnel/mux.d.ts +64 -0
- package/dist/tunnel/mux.js +449 -0
- package/dist/tunnel/peer-link.d.ts +88 -0
- package/dist/tunnel/peer-link.js +271 -0
- package/dist/tunnel/rtc-auth.d.ts +159 -0
- package/dist/tunnel/rtc-auth.js +340 -0
- package/dist/tunnel/rtc-duplex.d.ts +52 -0
- package/dist/tunnel/rtc-duplex.js +148 -0
- package/dist/tunnel/rtc-ice.d.ts +55 -0
- package/dist/tunnel/rtc-ice.js +100 -0
- package/dist/tunnel/secure.d.ts +179 -0
- package/dist/tunnel/secure.js +936 -0
- package/dist/tunnel/stream-reset-error.d.ts +6 -0
- package/dist/tunnel/stream-reset-error.js +9 -0
- package/dist/tunnel/streams.d.ts +244 -0
- package/dist/tunnel/streams.js +1071 -0
- package/dist/tunnel/team-grant.d.ts +90 -0
- package/dist/tunnel/team-grant.js +226 -0
- package/dist/tunnel/wire-bytes.d.ts +32 -0
- package/dist/tunnel/wire-bytes.js +90 -0
- package/dist/vault/api-key.d.ts +71 -0
- package/dist/vault/api-key.js +130 -0
- package/{vault/argon2.ts → dist/vault/argon2.d.ts} +18 -53
- package/dist/vault/argon2.js +82 -0
- package/dist/vault/cli-auth-seal.d.ts +35 -0
- package/dist/vault/cli-auth-seal.js +78 -0
- package/dist/vault/cred-seal.d.ts +18 -0
- package/dist/vault/cred-seal.js +86 -0
- package/dist/vault/credential.d.ts +21 -0
- package/dist/vault/credential.js +26 -0
- package/dist/vault/crypto.d.ts +51 -0
- package/dist/vault/crypto.js +156 -0
- package/dist/vault/daemon-attest.d.ts +68 -0
- package/dist/vault/daemon-attest.js +147 -0
- package/dist/vault/derive.d.ts +34 -0
- package/dist/vault/derive.js +84 -0
- package/dist/vault/device-access.d.ts +17 -0
- package/dist/vault/device-access.js +77 -0
- package/dist/vault/device-anchor.d.ts +42 -0
- package/dist/vault/device-anchor.js +90 -0
- package/dist/vault/device-control.d.ts +80 -0
- package/{vault/device-control.ts → dist/vault/device-control.js} +46 -104
- package/dist/vault/handoff.d.ts +102 -0
- package/dist/vault/handoff.js +236 -0
- package/dist/vault/index.d.ts +21 -0
- package/dist/vault/index.js +21 -0
- package/dist/vault/passkey.d.ts +82 -0
- package/{vault/passkey.ts → dist/vault/passkey.js} +63 -153
- package/dist/vault/phrase.d.ts +40 -0
- package/{vault/phrase.ts → dist/vault/phrase.js} +17 -23
- package/dist/vault/pin.d.ts +46 -0
- package/dist/vault/pin.js +92 -0
- package/dist/vault/public-key-seal.d.ts +17 -0
- package/dist/vault/public-key-seal.js +59 -0
- package/dist/vault/seed.d.ts +96 -0
- package/dist/vault/seed.js +179 -0
- package/dist/vault/team.d.ts +134 -0
- package/dist/vault/team.js +217 -0
- package/dist/vault/vault-item.d.ts +45 -0
- package/dist/vault/vault-item.js +62 -0
- package/dist/vault/vault-migrate-digest.d.ts +10 -0
- package/dist/vault/vault-migrate-digest.js +22 -0
- package/dist/vault/wrapped.d.ts +19 -0
- package/dist/vault/wrapped.js +52 -0
- package/dist/wire/adapters/messages/anthropic-map.d.ts +8 -0
- package/dist/wire/adapters/messages/anthropic-map.js +32 -0
- package/dist/wire/adapters/messages/index.d.ts +7 -0
- package/dist/wire/adapters/messages/index.js +7 -0
- package/dist/wire/adapters/messages/reasoning-from-items.d.ts +14 -0
- package/dist/wire/adapters/messages/reasoning-from-items.js +39 -0
- package/dist/wire/adapters/messages/reasoning-signature.d.ts +81 -0
- package/dist/wire/adapters/messages/reasoning-signature.js +180 -0
- package/dist/wire/adapters/messages/request.d.ts +28 -0
- package/dist/wire/adapters/messages/request.js +490 -0
- package/dist/wire/adapters/messages/response.d.ts +10 -0
- package/dist/wire/adapters/messages/response.js +177 -0
- package/dist/wire/adapters/messages/streaming.d.ts +134 -0
- package/dist/wire/adapters/messages/streaming.js +791 -0
- package/dist/wire/adapters/messages/usage.d.ts +20 -0
- package/dist/wire/adapters/messages/usage.js +10 -0
- package/dist/wire/adapters/responses/index.d.ts +3 -0
- package/dist/wire/adapters/responses/index.js +3 -0
- package/dist/wire/adapters/responses/request.d.ts +2 -0
- package/dist/wire/adapters/responses/request.js +287 -0
- package/dist/wire/adapters/responses/response.d.ts +11 -0
- package/dist/wire/adapters/responses/response.js +69 -0
- package/dist/wire/adapters/responses/streaming.d.ts +2 -0
- package/dist/wire/adapters/responses/streaming.js +267 -0
- package/dist/wire/features/account-rank.d.ts +61 -0
- package/dist/wire/features/account-rank.js +109 -0
- package/dist/wire/features/compaction/compact-request.d.ts +69 -0
- package/dist/wire/features/compaction/compact-request.js +667 -0
- package/dist/wire/features/compaction/compaction-text.d.ts +17 -0
- package/dist/wire/features/compaction/compaction-text.js +29 -0
- package/dist/wire/features/context-demote.d.ts +15 -0
- package/dist/wire/features/context-demote.js +45 -0
- package/dist/wire/features/context-overflow-strategy.d.ts +12 -0
- package/dist/wire/features/context-overflow-strategy.js +8 -0
- package/dist/wire/features/context-skip.d.ts +120 -0
- package/{wire/features/context-skip.ts → dist/wire/features/context-skip.js} +23 -57
- package/dist/wire/features/conversation-affinity.d.ts +49 -0
- package/dist/wire/features/conversation-affinity.js +121 -0
- package/dist/wire/features/max-tokens-backfill.d.ts +50 -0
- package/dist/wire/features/max-tokens-backfill.js +118 -0
- package/dist/wire/features/model-caps.d.ts +7 -0
- package/dist/wire/features/model-caps.js +40 -0
- package/dist/wire/features/model-settings.d.ts +3 -0
- package/dist/wire/features/model-settings.js +13 -0
- package/dist/wire/features/output-limit-clamp.d.ts +14 -0
- package/dist/wire/features/output-limit-clamp.js +36 -0
- package/dist/wire/features/quota-gate.d.ts +16 -0
- package/dist/wire/features/quota-gate.js +48 -0
- package/dist/wire/index.d.ts +44 -0
- package/dist/wire/index.js +44 -0
- package/dist/wire/lib/attempt-outcome.d.ts +19 -0
- package/dist/wire/lib/attempt-outcome.js +18 -0
- package/{wire/lib/canonical/content-part.ts → dist/wire/lib/canonical/content-part.d.ts} +12 -18
- package/dist/wire/lib/canonical/content-part.js +11 -0
- package/dist/wire/lib/canonical/count-tokens-capability.d.ts +52 -0
- package/dist/wire/lib/canonical/count-tokens-capability.js +109 -0
- package/{wire/lib/canonical/encoding-select.ts → dist/wire/lib/canonical/encoding-select.d.ts} +7 -24
- package/dist/wire/lib/canonical/encoding-select.js +26 -0
- package/dist/wire/lib/canonical/image-signature.d.ts +20 -0
- package/dist/wire/lib/canonical/image-signature.js +184 -0
- package/dist/wire/lib/canonical/image-sse.d.ts +88 -0
- package/dist/wire/lib/canonical/image-sse.js +173 -0
- package/{wire/lib/canonical/message.ts → dist/wire/lib/canonical/message.d.ts} +2 -20
- package/dist/wire/lib/canonical/message.js +31 -0
- package/dist/wire/lib/canonical/token-estimate.d.ts +98 -0
- package/dist/wire/lib/canonical/token-estimate.js +230 -0
- package/dist/wire/lib/encrypted-content.d.ts +22 -0
- package/dist/wire/lib/encrypted-content.js +55 -0
- package/dist/wire/lib/error-class.d.ts +57 -0
- package/dist/wire/lib/error-class.js +97 -0
- package/dist/wire/lib/error-envelope.d.ts +20 -0
- package/dist/wire/lib/error-envelope.js +53 -0
- package/dist/wire/lib/forwarded-headers.d.ts +16 -0
- package/{wire/lib/forwarded-headers.ts → dist/wire/lib/forwarded-headers.js} +62 -72
- package/dist/wire/lib/provider-error-rules.d.ts +21 -0
- package/dist/wire/lib/provider-error-rules.js +20 -0
- package/{wire/lib/refusal.ts → dist/wire/lib/refusal.d.ts} +3 -13
- package/dist/wire/lib/refusal.js +27 -0
- package/dist/wire/lib/responses-apply-patch.d.ts +57 -0
- package/dist/wire/lib/responses-apply-patch.js +81 -0
- package/dist/wire/lib/responses-namespace.d.ts +27 -0
- package/{wire/lib/responses-namespace.ts → dist/wire/lib/responses-namespace.js} +16 -32
- package/dist/wire/lib/secret-scrub.d.ts +23 -0
- package/dist/wire/lib/secret-scrub.js +72 -0
- package/dist/wire/lib/streaming/accumulate.d.ts +12 -0
- package/dist/wire/lib/streaming/accumulate.js +187 -0
- package/dist/wire/lib/streaming/finish-reason.d.ts +20 -0
- package/dist/wire/lib/streaming/finish-reason.js +15 -0
- package/{wire/lib/streaming/peek.ts → dist/wire/lib/streaming/peek.d.ts} +13 -82
- package/dist/wire/lib/streaming/peek.js +100 -0
- package/dist/wire/lib/streaming/provider-decode.d.ts +57 -0
- package/dist/wire/lib/streaming/provider-decode.js +210 -0
- package/dist/wire/lib/streaming/response-stream.d.ts +16 -0
- package/dist/wire/lib/streaming/response-stream.js +64 -0
- package/dist/wire/lib/streaming/sse.d.ts +130 -0
- package/dist/wire/lib/streaming/sse.js +600 -0
- package/dist/wire/lib/streaming/strip-tool-calls.d.ts +14 -0
- package/dist/wire/lib/streaming/strip-tool-calls.js +75 -0
- package/dist/wire/lib/streaming/upstream-error.d.ts +32 -0
- package/dist/wire/lib/streaming/upstream-error.js +45 -0
- package/dist/wire/lib/strip-subagent-isolation.d.ts +16 -0
- package/dist/wire/lib/strip-subagent-isolation.js +240 -0
- package/dist/wire/lib/thinking-signature.d.ts +56 -0
- package/{wire/lib/thinking-signature.ts → dist/wire/lib/thinking-signature.js} +77 -84
- package/dist/wire/lib/token-ruler/bpe-core.d.ts +105 -0
- package/dist/wire/lib/token-ruler/bpe-core.js +398 -0
- package/dist/wire/lib/token-ruler/index.d.ts +2 -0
- package/dist/wire/lib/token-ruler/index.js +2 -0
- package/dist/wire/lib/token-ruler/ruler.d.ts +36 -0
- package/dist/wire/lib/token-ruler/ruler.js +50 -0
- package/dist/wire/lib/token-ruler/vocab/claude-ranks.d.ts +14 -0
- package/dist/wire/lib/token-ruler/vocab/claude-ranks.js +16 -0
- package/{wire/lib/token-ruler/vocab/claude.ts → dist/wire/lib/token-ruler/vocab/claude.d.ts} +2 -12
- package/dist/wire/lib/token-ruler/vocab/claude.js +8 -0
- package/dist/wire/lib/token-ruler/vocab/o200k-ranks.d.ts +3 -0
- package/dist/wire/lib/token-ruler/vocab/o200k-ranks.js +18 -0
- package/dist/wire/lib/token-ruler/vocab/o200k.d.ts +11 -0
- package/dist/wire/lib/token-ruler/vocab/o200k.js +6 -0
- package/dist/wire/lib/tool-name-map.d.ts +44 -0
- package/dist/wire/lib/tool-name-map.js +186 -0
- package/dist/wire/lib/tool-schema.d.ts +15 -0
- package/dist/wire/lib/tool-schema.js +307 -0
- package/dist/wire/lib/usage-observer.d.ts +9 -0
- package/dist/wire/lib/usage-observer.js +108 -0
- package/dist/wire/providers/anthropic/adaptive-thinking.d.ts +125 -0
- package/dist/wire/providers/anthropic/adaptive-thinking.js +221 -0
- package/dist/wire/providers/anthropic/beta-headers.d.ts +44 -0
- package/{wire/providers/anthropic/beta-headers.ts → dist/wire/providers/anthropic/beta-headers.js} +57 -79
- package/dist/wire/providers/anthropic/index.d.ts +6 -0
- package/{wire/providers/anthropic/index.ts → dist/wire/providers/anthropic/index.js} +6 -6
- package/dist/wire/providers/anthropic/request.d.ts +12 -0
- package/dist/wire/providers/anthropic/request.js +456 -0
- package/dist/wire/providers/anthropic/response.d.ts +3 -0
- package/dist/wire/providers/anthropic/response.js +126 -0
- package/dist/wire/providers/anthropic/streaming.d.ts +28 -0
- package/dist/wire/providers/anthropic/streaming.js +268 -0
- package/dist/wire/providers/anthropic/web-search-domains.d.ts +7 -0
- package/dist/wire/providers/anthropic/web-search-domains.js +36 -0
- package/dist/wire/providers/chatgpt/common.d.ts +27 -0
- package/{wire/providers/chatgpt/common.ts → dist/wire/providers/chatgpt/common.js} +69 -89
- package/dist/wire/providers/chatgpt/index.d.ts +4 -0
- package/{wire/providers/chatgpt/index.ts → dist/wire/providers/chatgpt/index.js} +4 -4
- package/dist/wire/providers/chatgpt/request.d.ts +132 -0
- package/dist/wire/providers/chatgpt/request.js +696 -0
- package/dist/wire/providers/chatgpt/streaming.d.ts +69 -0
- package/dist/wire/providers/chatgpt/streaming.js +761 -0
- package/dist/wire/providers/chatgpt/web-search.d.ts +19 -0
- package/dist/wire/providers/chatgpt/web-search.js +55 -0
- package/dist/wire/providers/grok/web-search.d.ts +23 -0
- package/dist/wire/providers/grok/web-search.js +74 -0
- package/dist/wire/providers/kimi/web-search.d.ts +48 -0
- package/dist/wire/providers/kimi/web-search.js +106 -0
- package/dist/wire/providers/upstream-deny.d.ts +61 -0
- package/dist/wire/providers/upstream-deny.js +293 -0
- package/dist/wire/providers/upstream-request.d.ts +104 -0
- package/dist/wire/providers/upstream-request.js +411 -0
- package/package.json +105 -23
- package/ARCHITECTURE.md +0 -43
- package/client/chat.ts +0 -140
- package/client/devices.ts +0 -269
- package/client/http.ts +0 -72
- package/client/index.ts +0 -129
- package/client/relay-mux.ts +0 -547
- package/client/types.ts +0 -51
- package/client/vault.ts +0 -239
- package/connect/chat.ts +0 -58
- package/connect/client.ts +0 -240
- package/connect/config.ts +0 -47
- package/connect/errors.ts +0 -121
- package/connect/oauth.ts +0 -265
- package/connect/routes.ts +0 -234
- package/connect/store.ts +0 -107
- package/mcp/handle.ts +0 -24
- package/mcp/index.ts +0 -76
- package/mcp/server.ts +0 -319
- package/mcp/tools/marketplace.ts +0 -439
- package/mcp/tools/memory-client.ts +0 -297
- package/mcp/tools/memory-config.ts +0 -32
- package/mcp/tools/memory.ts +0 -369
- package/mcp/tools/model-priority.ts +0 -88
- package/mcp/tools/openllm.ts +0 -883
- package/mcp/types.ts +0 -28
- package/openapi/call.ts +0 -90
- package/openapi/index.ts +0 -8
- package/openapi/openapi.json +0 -80213
- package/openapi/operations.ts +0 -2075
- package/protocol/CONTRIBUTING.md +0 -21
- package/protocol/README.md +0 -134
- package/protocol/access.ts +0 -186
- package/protocol/agent.ts +0 -991
- package/protocol/api-key-syntax.ts +0 -105
- package/protocol/assets/openllm-light.svg +0 -18
- package/protocol/assets/openllm.svg +0 -18
- package/protocol/audio.ts +0 -29
- package/protocol/auth.ts +0 -142
- package/protocol/changelog.ts +0 -155
- package/protocol/chat.ts +0 -458
- package/protocol/client-registry.ts +0 -356
- package/protocol/config.ts +0 -345
- package/protocol/cooldown-reason.ts +0 -78
- package/protocol/cost.ts +0 -10
- package/protocol/daemon-port.ts +0 -31
- package/protocol/daemon.ts +0 -1524
- package/protocol/default-chains.ts +0 -66
- package/protocol/device-state.ts +0 -85
- package/protocol/doctor-report.ts +0 -947
- package/protocol/errors.ts +0 -14
- package/protocol/image-edit-parse.ts +0 -263
- package/protocol/images.ts +0 -298
- package/protocol/index.ts +0 -74
- package/protocol/keepalive.ts +0 -69
- package/protocol/local-auth.ts +0 -300
- package/protocol/local-provider.ts +0 -288
- package/protocol/local-session.ts +0 -118
- package/protocol/marketplace.ts +0 -187
- package/protocol/media-bytes.ts +0 -28
- package/protocol/media-default.ts +0 -357
- package/protocol/media-input.ts +0 -61
- package/protocol/media-model.ts +0 -142
- package/protocol/media-types.ts +0 -70
- package/protocol/model-match.ts +0 -1051
- package/protocol/model-settings.ts +0 -327
- package/protocol/models.ts +0 -767
- package/protocol/mux.ts +0 -722
- package/protocol/provider-status.ts +0 -206
- package/protocol/provider-usage.ts +0 -263
- package/protocol/providers.ts +0 -901
- package/protocol/public-usage-index.ts +0 -354
- package/protocol/registry-pointer.ts +0 -118
- package/protocol/responses.ts +0 -280
- package/protocol/semver.ts +0 -182
- package/protocol/sha256-hex.ts +0 -98
- package/protocol/stats.ts +0 -406
- package/protocol/streaming.ts +0 -19
- package/protocol/subscription-provider-compat.ts +0 -104
- package/protocol/subscription-provider.ts +0 -32
- package/protocol/transcription-encoding.ts +0 -76
- package/protocol/url-locality.ts +0 -288
- package/protocol/video-job-plan.ts +0 -50
- package/protocol/videos.ts +0 -227
- package/tsconfig.json +0 -14
- package/tunnel/ARCHITECTURE.md +0 -111
- package/tunnel/assets/openllm-light.svg +0 -18
- package/tunnel/assets/openllm.svg +0 -18
- package/tunnel/channel-envelope.ts +0 -139
- package/tunnel/codec.ts +0 -191
- package/tunnel/device-anchor.ts +0 -229
- package/tunnel/device-control.ts +0 -118
- package/tunnel/device-grant.ts +0 -173
- package/tunnel/fleet.ts +0 -77
- package/tunnel/index.ts +0 -22
- package/tunnel/mux.ts +0 -580
- package/tunnel/peer-link.ts +0 -390
- package/tunnel/rtc-auth.ts +0 -415
- package/tunnel/rtc-duplex.ts +0 -187
- package/tunnel/rtc-ice.ts +0 -136
- package/tunnel/secure.ts +0 -1298
- package/tunnel/stream-reset-error.ts +0 -14
- package/tunnel/streams.ts +0 -1430
- package/tunnel/team-grant.ts +0 -302
- package/tunnel/wire-bytes.ts +0 -102
- package/vault/ARCHITECTURE.md +0 -470
- package/vault/api-key.ts +0 -190
- package/vault/cli-auth-seal.ts +0 -157
- package/vault/cred-seal.ts +0 -146
- package/vault/credential.ts +0 -55
- package/vault/crypto.ts +0 -210
- package/vault/daemon-attest.ts +0 -229
- package/vault/derive.ts +0 -113
- package/vault/device-access.ts +0 -87
- package/vault/device-anchor.ts +0 -185
- package/vault/handoff.ts +0 -396
- package/vault/index.ts +0 -21
- package/vault/pin.ts +0 -153
- package/vault/public-key-seal.ts +0 -131
- package/vault/seed.ts +0 -272
- package/vault/team.ts +0 -355
- package/vault/vault-item.ts +0 -131
- package/vault/vault-migrate-digest.ts +0 -29
- package/vault/wrapped.ts +0 -77
- package/wire/ARCHITECTURE.md +0 -90
- package/wire/CONTRIBUTING.md +0 -19
- package/wire/adapters/messages/anthropic-map.ts +0 -37
- package/wire/adapters/messages/index.ts +0 -7
- package/wire/adapters/messages/reasoning-from-items.ts +0 -38
- package/wire/adapters/messages/reasoning-signature.ts +0 -243
- package/wire/adapters/messages/request.ts +0 -574
- package/wire/adapters/messages/response.ts +0 -215
- package/wire/adapters/messages/streaming.ts +0 -1027
- package/wire/adapters/messages/usage.ts +0 -43
- package/wire/adapters/responses/index.ts +0 -3
- package/wire/adapters/responses/request.ts +0 -395
- package/wire/adapters/responses/response.ts +0 -88
- package/wire/adapters/responses/streaming.ts +0 -344
- package/wire/assets/openllm-light.svg +0 -18
- package/wire/assets/openllm.svg +0 -18
- package/wire/features/account-rank.ts +0 -166
- package/wire/features/compaction/compact-request.ts +0 -862
- package/wire/features/compaction/compaction-text.ts +0 -43
- package/wire/features/context-demote.ts +0 -47
- package/wire/features/context-overflow-strategy.ts +0 -26
- package/wire/features/conversation-affinity.ts +0 -144
- package/wire/features/max-tokens-backfill.ts +0 -141
- package/wire/features/model-caps.ts +0 -53
- package/wire/features/model-settings.ts +0 -24
- package/wire/features/output-limit-clamp.ts +0 -64
- package/wire/features/quota-gate.ts +0 -85
- package/wire/index.ts +0 -44
- package/wire/lib/attempt-outcome.ts +0 -43
- package/wire/lib/canonical/count-tokens-capability.ts +0 -138
- package/wire/lib/canonical/image-signature.ts +0 -202
- package/wire/lib/canonical/image-sse.ts +0 -216
- package/wire/lib/canonical/token-estimate.ts +0 -266
- package/wire/lib/encrypted-content.ts +0 -66
- package/wire/lib/error-class.ts +0 -154
- package/wire/lib/error-envelope.ts +0 -56
- package/wire/lib/provider-error-rules.ts +0 -48
- package/wire/lib/responses-apply-patch.ts +0 -104
- package/wire/lib/secret-scrub.ts +0 -101
- package/wire/lib/streaming/accumulate.ts +0 -221
- package/wire/lib/streaming/finish-reason.ts +0 -33
- package/wire/lib/streaming/provider-decode.ts +0 -259
- package/wire/lib/streaming/response-stream.ts +0 -71
- package/wire/lib/streaming/sse.ts +0 -699
- package/wire/lib/streaming/strip-tool-calls.ts +0 -95
- package/wire/lib/streaming/upstream-error.ts +0 -58
- package/wire/lib/strip-subagent-isolation.ts +0 -284
- package/wire/lib/token-ruler/bpe-core.ts +0 -426
- package/wire/lib/token-ruler/index.ts +0 -2
- package/wire/lib/token-ruler/ruler.ts +0 -79
- package/wire/lib/token-ruler/vocab/claude-ranks.ts +0 -16
- package/wire/lib/token-ruler/vocab/o200k-ranks.ts +0 -18
- package/wire/lib/token-ruler/vocab/o200k.ts +0 -19
- package/wire/lib/tool-name-map.ts +0 -271
- package/wire/lib/tool-schema.ts +0 -353
- package/wire/lib/usage-observer.ts +0 -125
- package/wire/providers/anthropic/adaptive-thinking.ts +0 -291
- package/wire/providers/anthropic/request.ts +0 -607
- package/wire/providers/anthropic/response.ts +0 -155
- package/wire/providers/anthropic/streaming.ts +0 -315
- package/wire/providers/anthropic/web-search-domains.ts +0 -49
- package/wire/providers/chatgpt/request.ts +0 -959
- package/wire/providers/chatgpt/streaming.ts +0 -1007
- package/wire/providers/chatgpt/web-search.ts +0 -63
- package/wire/providers/grok/web-search.ts +0 -87
- package/wire/providers/kimi/web-search.ts +0 -145
- package/wire/providers/upstream-deny.ts +0 -355
- package/wire/providers/upstream-request.ts +0 -619
- /package/{protocol/inference-headers.ts → dist/protocol/inference-headers.js} +0 -0
|
@@ -10,34 +10,16 @@
|
|
|
10
10
|
* intentional and must not be folded in here.
|
|
11
11
|
*/
|
|
12
12
|
import type { TChatMessage } from "@openllmsh/sdk/protocol";
|
|
13
|
-
|
|
14
13
|
/**
|
|
15
14
|
* Flatten a canonical message `content` (string | content-part[]) to
|
|
16
15
|
* plain text: keep only `text` parts, joined by newlines. `null` /
|
|
17
16
|
* `undefined` → `""`.
|
|
18
17
|
*/
|
|
19
|
-
export const extractMessageText
|
|
20
|
-
content: TChatMessage["content"] | null | undefined,
|
|
21
|
-
): string => {
|
|
22
|
-
if (content == null) return "";
|
|
23
|
-
if (typeof content === "string") return content;
|
|
24
|
-
return content
|
|
25
|
-
.filter((part) => part.type === "text")
|
|
26
|
-
.map((part) => (part as { type: "text"; text: string }).text)
|
|
27
|
-
.join("\n");
|
|
28
|
-
};
|
|
29
|
-
|
|
18
|
+
export declare const extractMessageText: (content: TChatMessage["content"] | null | undefined) => string;
|
|
30
19
|
/**
|
|
31
20
|
* Deserialize an OpenAI tool-call `arguments` string. Empty → `{}`;
|
|
32
21
|
* valid JSON → parsed value; unparseable → the raw string verbatim
|
|
33
22
|
* (Anthropic `tool_use.input` and the Messages adapter both accept a
|
|
34
23
|
* raw string as a last resort).
|
|
35
24
|
*/
|
|
36
|
-
export const parseToolArguments
|
|
37
|
-
if (raw.length === 0) return {};
|
|
38
|
-
try {
|
|
39
|
-
return JSON.parse(raw);
|
|
40
|
-
} catch {
|
|
41
|
-
return raw;
|
|
42
|
-
}
|
|
43
|
-
};
|
|
25
|
+
export declare const parseToolArguments: (raw: string) => unknown;
|
|
@@ -0,0 +1,31 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Flatten a canonical message `content` (string | content-part[]) to
|
|
3
|
+
* plain text: keep only `text` parts, joined by newlines. `null` /
|
|
4
|
+
* `undefined` → `""`.
|
|
5
|
+
*/
|
|
6
|
+
export const extractMessageText = (content) => {
|
|
7
|
+
if (content == null)
|
|
8
|
+
return "";
|
|
9
|
+
if (typeof content === "string")
|
|
10
|
+
return content;
|
|
11
|
+
return content
|
|
12
|
+
.filter((part) => part.type === "text")
|
|
13
|
+
.map((part) => part.text)
|
|
14
|
+
.join("\n");
|
|
15
|
+
};
|
|
16
|
+
/**
|
|
17
|
+
* Deserialize an OpenAI tool-call `arguments` string. Empty → `{}`;
|
|
18
|
+
* valid JSON → parsed value; unparseable → the raw string verbatim
|
|
19
|
+
* (Anthropic `tool_use.input` and the Messages adapter both accept a
|
|
20
|
+
* raw string as a last resort).
|
|
21
|
+
*/
|
|
22
|
+
export const parseToolArguments = (raw) => {
|
|
23
|
+
if (raw.length === 0)
|
|
24
|
+
return {};
|
|
25
|
+
try {
|
|
26
|
+
return JSON.parse(raw);
|
|
27
|
+
}
|
|
28
|
+
catch {
|
|
29
|
+
return raw;
|
|
30
|
+
}
|
|
31
|
+
};
|
|
@@ -0,0 +1,98 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The TWO input-token estimators, together in one module so the choice between
|
|
3
|
+
* them is visible at the point of use. They are not interchangeable — each is
|
|
4
|
+
* wrong for the other's job:
|
|
5
|
+
*
|
|
6
|
+
* - {@link estimateBodyTokens} — the ROUTING gate. Every string in the body
|
|
7
|
+
* counts, so it OVER-estimates (discriminators, tool names, a base64 image
|
|
8
|
+
* payload). That is the point: it decides whether a hop's context window can
|
|
9
|
+
* hold the request, where over-shooting costs one extra chain step and
|
|
10
|
+
* under-shooting ships a request that 400s.
|
|
11
|
+
* - {@link estimateAnthropicInputTokens} — the CLIENT-FACING preflight
|
|
12
|
+
* (`/v1/messages/count_tokens`). Counts only what the model actually reads,
|
|
13
|
+
* so a client's context-window indicator isn't inflated by transport noise.
|
|
14
|
+
* Always uses a real BPE (default {@link DEFAULT_ENCODING} / o200k). Never
|
|
15
|
+
* use the routing estimator here: base64 image data alone would make
|
|
16
|
+
* Claude Code believe a fresh session was nearly full.
|
|
17
|
+
*/
|
|
18
|
+
/**
|
|
19
|
+
* Rough input-token estimator over a request body, format-agnostic.
|
|
20
|
+
*
|
|
21
|
+
* Walks the JSON value summing string lengths (everything else
|
|
22
|
+
* contributes zero) and divides by 4. Works for both canonical
|
|
23
|
+
* OpenAI `ChatCompletionRequest` and raw Anthropic Messages bodies
|
|
24
|
+
* because both encode all user-visible text inside string values.
|
|
25
|
+
*
|
|
26
|
+
* Used pre-fetch by the chain orchestrator to skip a hop whose
|
|
27
|
+
* provider model can't fit the request — preventing the otherwise
|
|
28
|
+
* inevitable 400 + waste of the chain entry. The estimate is
|
|
29
|
+
* intentionally conservative: in borderline cases we'd rather burn
|
|
30
|
+
* an extra chain step than ship an oversized request that 400s.
|
|
31
|
+
*/
|
|
32
|
+
import type { TRulerCache, TTokenEncoding } from "./encoding-select.js";
|
|
33
|
+
/**
|
|
34
|
+
* The ROUTING gate — ALWAYS the `chars/4` heuristic, never the BPE.
|
|
35
|
+
*
|
|
36
|
+
* This runs on every request (twice in the cloud handler, once per daemon walk),
|
|
37
|
+
* so it is deliberately arithmetic-only: a string-length walk with no
|
|
38
|
+
* tokenizer, no allocation of the joined body, and no dependence on whether
|
|
39
|
+
* some earlier request happened to warm a ruler on this isolate.
|
|
40
|
+
*
|
|
41
|
+
* It used to consult {@link peekTokenCounter} and silently upgrade itself to a
|
|
42
|
+
* real BPE pass once any ruler was warm. That made the hot path's cost — and
|
|
43
|
+
* its ANSWER — a function of process history: the first oversized request
|
|
44
|
+
* warmed the module-global counter cache and every later request on that
|
|
45
|
+
* isolate then paid full tokenization. Accuracy is now opt-in via
|
|
46
|
+
* {@link estimateBodyTokensExact}, taken only where it changes a decision
|
|
47
|
+
* (compaction) rather than ambiently everywhere.
|
|
48
|
+
*/
|
|
49
|
+
export declare const estimateBodyTokens: (body: unknown) => number;
|
|
50
|
+
/** Count the serialized wire body for cap holds. On a failed or cold ruler,
|
|
51
|
+
* preserve the historical bytes/4 estimate instead of refusing the request.
|
|
52
|
+
* The reserve/admit modes bracket both estimates so differently warmed direct
|
|
53
|
+
* leg peers cannot reject an honest caller's signed allowance.
|
|
54
|
+
*/
|
|
55
|
+
export declare const estimateCapInputTokens: (body: Uint8Array, mode?: "reserve" | "admit", cache?: TRulerCache) => number;
|
|
56
|
+
/**
|
|
57
|
+
* Warm the rulers {@link estimateCapInputTokens} uses for `mode`, each
|
|
58
|
+
* independently (a failed load leaves the other usable). Admission only reads
|
|
59
|
+
* o200k, so a supplier never pays for the Claude ruler.
|
|
60
|
+
*/
|
|
61
|
+
export declare const warmCapTokenCounter: (mode?: "reserve" | "admit") => Promise<void>;
|
|
62
|
+
/**
|
|
63
|
+
* The ACCURATE body estimate — the real BPE ruler when `encoding`'s family is
|
|
64
|
+
* warm, `chars/4` when it isn't. Opt-in and explicit: use it only where the
|
|
65
|
+
* extra precision changes an outcome and the caller has warmed the ruler
|
|
66
|
+
* deliberately (the compaction seams do, via `getTokenCounter`).
|
|
67
|
+
*
|
|
68
|
+
* Never call this on the per-request routing path — that is
|
|
69
|
+
* {@link estimateBodyTokens}, and it must stay arithmetic-only.
|
|
70
|
+
*/
|
|
71
|
+
export declare const estimateBodyTokensExact: (body: unknown, encoding?: TTokenEncoding) => number;
|
|
72
|
+
/**
|
|
73
|
+
* Estimated `input_tokens` for an Anthropic-shaped Messages body — the answer
|
|
74
|
+
* `/v1/messages/count_tokens` serves ONLY as a fallback, when the vendor's own
|
|
75
|
+
* endpoint isn't reachable (no Anthropic-wire hop at the head of the chain, no
|
|
76
|
+
* local credential, or an upstream refusal). Whenever the vendor can be asked,
|
|
77
|
+
* its exact count is relayed instead and this function is not consulted.
|
|
78
|
+
*
|
|
79
|
+
* Always acquires a BPE counter — never a silent `chars/4` heuristic. Default
|
|
80
|
+
* encoding is {@link DEFAULT_ENCODING} (o200k) regardless of inbound Messages
|
|
81
|
+
* wire shape. Pass `"claude"` only when the caller is explicitly counting for
|
|
82
|
+
* an Anthropic-family hop.
|
|
83
|
+
*
|
|
84
|
+
* Single-sourced here because BOTH preflight paths need the same number: the
|
|
85
|
+
* cloud handler (`@openllm/api/handlers/count-tokens`) and the daemon's local
|
|
86
|
+
* surface (`runCountTokens` in the walker).
|
|
87
|
+
*
|
|
88
|
+
* KNOWN LIMIT — text only. Images and documents contribute nothing, so an
|
|
89
|
+
* image-heavy (or image-only) request UNDER-counts, materially. This is
|
|
90
|
+
* deliberate: Anthropic prices an image at roughly `(w × h) / 750` tokens, and
|
|
91
|
+
* neither dimension is recoverable from a base64 blob or a URL without decoding
|
|
92
|
+
* or fetching it — so any number we put there would be fabricated, and a
|
|
93
|
+
* fabricated count is worse than a knowingly-low one for a fallback whose
|
|
94
|
+
* consumers use it as a context-window indicator. Counting the blob's own length
|
|
95
|
+
* (what {@link estimateBodyTokens} does) is not a substitute: it tracks encoding
|
|
96
|
+
* size, not token cost, and overstates by orders of magnitude.
|
|
97
|
+
*/
|
|
98
|
+
export declare const estimateAnthropicInputTokens: (body: unknown, encoding?: TTokenEncoding) => Promise<number>;
|
|
@@ -0,0 +1,230 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The TWO input-token estimators, together in one module so the choice between
|
|
3
|
+
* them is visible at the point of use. They are not interchangeable — each is
|
|
4
|
+
* wrong for the other's job:
|
|
5
|
+
*
|
|
6
|
+
* - {@link estimateBodyTokens} — the ROUTING gate. Every string in the body
|
|
7
|
+
* counts, so it OVER-estimates (discriminators, tool names, a base64 image
|
|
8
|
+
* payload). That is the point: it decides whether a hop's context window can
|
|
9
|
+
* hold the request, where over-shooting costs one extra chain step and
|
|
10
|
+
* under-shooting ships a request that 400s.
|
|
11
|
+
* - {@link estimateAnthropicInputTokens} — the CLIENT-FACING preflight
|
|
12
|
+
* (`/v1/messages/count_tokens`). Counts only what the model actually reads,
|
|
13
|
+
* so a client's context-window indicator isn't inflated by transport noise.
|
|
14
|
+
* Always uses a real BPE (default {@link DEFAULT_ENCODING} / o200k). Never
|
|
15
|
+
* use the routing estimator here: base64 image data alone would make
|
|
16
|
+
* Claude Code believe a fresh session was nearly full.
|
|
17
|
+
*/
|
|
18
|
+
import { DEFAULT_ENCODING, getTokenCounter, peekTokenCounter, } from "./encoding-select.js";
|
|
19
|
+
/**
|
|
20
|
+
* Count a single string as tokens for `encoding`, using the real BPE ruler when
|
|
21
|
+
* that family's tokenizer is already warm on this isolate, and falling back to
|
|
22
|
+
* the `chars/4` heuristic when it isn't. The synchronous fallback is what lets
|
|
23
|
+
* every caller stay synchronous; warm the ruler at request entry
|
|
24
|
+
* (`getTokenCounter`) to make the accurate path the common one. `max(1, …)` so
|
|
25
|
+
* any non-empty text counts as at least one token.
|
|
26
|
+
*/
|
|
27
|
+
const rulerTokens = (text, encoding) => {
|
|
28
|
+
if (text.length === 0)
|
|
29
|
+
return 0;
|
|
30
|
+
const counter = peekTokenCounter(encoding);
|
|
31
|
+
if (counter !== null)
|
|
32
|
+
return Math.max(1, counter.count(text));
|
|
33
|
+
return Math.max(1, Math.ceil(text.length / 4));
|
|
34
|
+
};
|
|
35
|
+
const stringCharsFromAny = (v) => {
|
|
36
|
+
if (typeof v === "string")
|
|
37
|
+
return v.length;
|
|
38
|
+
if (Array.isArray(v)) {
|
|
39
|
+
let n = 0;
|
|
40
|
+
for (const x of v)
|
|
41
|
+
n += stringCharsFromAny(x);
|
|
42
|
+
return n;
|
|
43
|
+
}
|
|
44
|
+
if (v !== null && typeof v === "object") {
|
|
45
|
+
let n = 0;
|
|
46
|
+
for (const k of Object.keys(v)) {
|
|
47
|
+
n += stringCharsFromAny(v[k]);
|
|
48
|
+
}
|
|
49
|
+
return n;
|
|
50
|
+
}
|
|
51
|
+
return 0;
|
|
52
|
+
};
|
|
53
|
+
/**
|
|
54
|
+
* Collect every string in a JSON value, joined by "\n" (a token boundary the
|
|
55
|
+
* BPE won't merge across), so the ruler counts them as one pass instead of
|
|
56
|
+
* summing chars. Mirrors {@link stringCharsFromAny}'s walk exactly.
|
|
57
|
+
*/
|
|
58
|
+
const collectStrings = (v, out) => {
|
|
59
|
+
if (typeof v === "string") {
|
|
60
|
+
out.push(v);
|
|
61
|
+
return;
|
|
62
|
+
}
|
|
63
|
+
if (Array.isArray(v)) {
|
|
64
|
+
for (const x of v)
|
|
65
|
+
collectStrings(x, out);
|
|
66
|
+
return;
|
|
67
|
+
}
|
|
68
|
+
if (v !== null && typeof v === "object") {
|
|
69
|
+
for (const k of Object.keys(v)) {
|
|
70
|
+
collectStrings(v[k], out);
|
|
71
|
+
}
|
|
72
|
+
}
|
|
73
|
+
};
|
|
74
|
+
/**
|
|
75
|
+
* The ROUTING gate — ALWAYS the `chars/4` heuristic, never the BPE.
|
|
76
|
+
*
|
|
77
|
+
* This runs on every request (twice in the cloud handler, once per daemon walk),
|
|
78
|
+
* so it is deliberately arithmetic-only: a string-length walk with no
|
|
79
|
+
* tokenizer, no allocation of the joined body, and no dependence on whether
|
|
80
|
+
* some earlier request happened to warm a ruler on this isolate.
|
|
81
|
+
*
|
|
82
|
+
* It used to consult {@link peekTokenCounter} and silently upgrade itself to a
|
|
83
|
+
* real BPE pass once any ruler was warm. That made the hot path's cost — and
|
|
84
|
+
* its ANSWER — a function of process history: the first oversized request
|
|
85
|
+
* warmed the module-global counter cache and every later request on that
|
|
86
|
+
* isolate then paid full tokenization. Accuracy is now opt-in via
|
|
87
|
+
* {@link estimateBodyTokensExact}, taken only where it changes a decision
|
|
88
|
+
* (compaction) rather than ambiently everywhere.
|
|
89
|
+
*/
|
|
90
|
+
export const estimateBodyTokens = (body) => Math.ceil(stringCharsFromAny(body) / 4);
|
|
91
|
+
/** Count the serialized wire body for cap holds. On a failed or cold ruler,
|
|
92
|
+
* preserve the historical bytes/4 estimate instead of refusing the request.
|
|
93
|
+
* The reserve/admit modes bracket both estimates so differently warmed direct
|
|
94
|
+
* leg peers cannot reject an honest caller's signed allowance.
|
|
95
|
+
*/
|
|
96
|
+
export const estimateCapInputTokens = (body, mode = "reserve", cache) => {
|
|
97
|
+
const fallback = Math.ceil(body.byteLength / 4);
|
|
98
|
+
const counter = peekTokenCounter(DEFAULT_ENCODING, cache);
|
|
99
|
+
const claude = mode === "reserve" ? peekTokenCounter("claude", cache) : null;
|
|
100
|
+
if (counter === null && claude === null)
|
|
101
|
+
return fallback;
|
|
102
|
+
const text = new TextDecoder().decode(body);
|
|
103
|
+
if (mode === "admit")
|
|
104
|
+
return Math.min(counter?.count(text) ?? fallback, fallback);
|
|
105
|
+
return Math.max(fallback, counter?.count(text) ?? 0, claude?.count(text) ?? 0);
|
|
106
|
+
};
|
|
107
|
+
/**
|
|
108
|
+
* Warm the rulers {@link estimateCapInputTokens} uses for `mode`, each
|
|
109
|
+
* independently (a failed load leaves the other usable). Admission only reads
|
|
110
|
+
* o200k, so a supplier never pays for the Claude ruler.
|
|
111
|
+
*/
|
|
112
|
+
export const warmCapTokenCounter = async (mode = "reserve") => {
|
|
113
|
+
await Promise.allSettled([
|
|
114
|
+
getTokenCounter(DEFAULT_ENCODING),
|
|
115
|
+
...(mode === "reserve" ? [getTokenCounter("claude")] : []),
|
|
116
|
+
]);
|
|
117
|
+
};
|
|
118
|
+
/**
|
|
119
|
+
* The ACCURATE body estimate — the real BPE ruler when `encoding`'s family is
|
|
120
|
+
* warm, `chars/4` when it isn't. Opt-in and explicit: use it only where the
|
|
121
|
+
* extra precision changes an outcome and the caller has warmed the ruler
|
|
122
|
+
* deliberately (the compaction seams do, via `getTokenCounter`).
|
|
123
|
+
*
|
|
124
|
+
* Never call this on the per-request routing path — that is
|
|
125
|
+
* {@link estimateBodyTokens}, and it must stay arithmetic-only.
|
|
126
|
+
*/
|
|
127
|
+
export const estimateBodyTokensExact = (body, encoding = DEFAULT_ENCODING) => {
|
|
128
|
+
if (peekTokenCounter(encoding) === null) {
|
|
129
|
+
// Cold — the ruler was never warmed, so fall back to the cheap heuristic.
|
|
130
|
+
return Math.ceil(stringCharsFromAny(body) / 4);
|
|
131
|
+
}
|
|
132
|
+
const parts = [];
|
|
133
|
+
collectStrings(body, parts);
|
|
134
|
+
return rulerTokens(parts.join("\n"), encoding);
|
|
135
|
+
};
|
|
136
|
+
/**
|
|
137
|
+
* The text an Anthropic content value actually puts in front of the model:
|
|
138
|
+
* `text` blocks, a `tool_use`'s serialized input, a `tool_result`'s body
|
|
139
|
+
* (recursively — it may itself be a block array), and prior `thinking`. Block
|
|
140
|
+
* types the model does not read as text (image/document sources) contribute
|
|
141
|
+
* nothing, which is exactly why this is not `estimateBodyTokens`.
|
|
142
|
+
*/
|
|
143
|
+
const anthropicContentText = (content) => {
|
|
144
|
+
if (typeof content === "string")
|
|
145
|
+
return content;
|
|
146
|
+
if (!Array.isArray(content))
|
|
147
|
+
return "";
|
|
148
|
+
const parts = [];
|
|
149
|
+
for (const block of content) {
|
|
150
|
+
if (block === null || typeof block !== "object")
|
|
151
|
+
continue;
|
|
152
|
+
const b = block;
|
|
153
|
+
if (b.type === "text" && typeof b.text === "string")
|
|
154
|
+
parts.push(b.text);
|
|
155
|
+
else if (b.type === "tool_use") {
|
|
156
|
+
// The name is billed alongside the arguments — both are in the transcript
|
|
157
|
+
// the model reads back.
|
|
158
|
+
if (typeof b.name === "string")
|
|
159
|
+
parts.push(b.name);
|
|
160
|
+
parts.push(JSON.stringify(b.input ?? {}));
|
|
161
|
+
}
|
|
162
|
+
else if (b.type === "tool_result") {
|
|
163
|
+
const inner = b.content;
|
|
164
|
+
if (typeof inner === "string")
|
|
165
|
+
parts.push(inner);
|
|
166
|
+
else if (Array.isArray(inner))
|
|
167
|
+
parts.push(anthropicContentText(inner));
|
|
168
|
+
}
|
|
169
|
+
else if (b.type === "thinking" && typeof b.thinking === "string") {
|
|
170
|
+
parts.push(b.thinking);
|
|
171
|
+
}
|
|
172
|
+
}
|
|
173
|
+
return parts.join("\n");
|
|
174
|
+
};
|
|
175
|
+
const visibleTranscriptOf = (body) => {
|
|
176
|
+
if (body === null || typeof body !== "object")
|
|
177
|
+
return "";
|
|
178
|
+
const b = body;
|
|
179
|
+
const sections = [];
|
|
180
|
+
const system = b.system;
|
|
181
|
+
if (typeof system === "string")
|
|
182
|
+
sections.push(system);
|
|
183
|
+
else if (Array.isArray(system))
|
|
184
|
+
sections.push(anthropicContentText(system));
|
|
185
|
+
if (Array.isArray(b.messages)) {
|
|
186
|
+
for (const m of b.messages) {
|
|
187
|
+
if (m === null || typeof m !== "object")
|
|
188
|
+
continue;
|
|
189
|
+
sections.push(anthropicContentText(m.content));
|
|
190
|
+
}
|
|
191
|
+
}
|
|
192
|
+
if (Array.isArray(b.tools) && b.tools.length > 0) {
|
|
193
|
+
sections.push(JSON.stringify(b.tools));
|
|
194
|
+
}
|
|
195
|
+
return sections.filter((s) => s.length > 0).join("\n");
|
|
196
|
+
};
|
|
197
|
+
const countTranscript = (joined, counter) => joined.length === 0 ? 0 : Math.max(1, counter.count(joined));
|
|
198
|
+
/**
|
|
199
|
+
* Estimated `input_tokens` for an Anthropic-shaped Messages body — the answer
|
|
200
|
+
* `/v1/messages/count_tokens` serves ONLY as a fallback, when the vendor's own
|
|
201
|
+
* endpoint isn't reachable (no Anthropic-wire hop at the head of the chain, no
|
|
202
|
+
* local credential, or an upstream refusal). Whenever the vendor can be asked,
|
|
203
|
+
* its exact count is relayed instead and this function is not consulted.
|
|
204
|
+
*
|
|
205
|
+
* Always acquires a BPE counter — never a silent `chars/4` heuristic. Default
|
|
206
|
+
* encoding is {@link DEFAULT_ENCODING} (o200k) regardless of inbound Messages
|
|
207
|
+
* wire shape. Pass `"claude"` only when the caller is explicitly counting for
|
|
208
|
+
* an Anthropic-family hop.
|
|
209
|
+
*
|
|
210
|
+
* Single-sourced here because BOTH preflight paths need the same number: the
|
|
211
|
+
* cloud handler (`@openllm/api/handlers/count-tokens`) and the daemon's local
|
|
212
|
+
* surface (`runCountTokens` in the walker).
|
|
213
|
+
*
|
|
214
|
+
* KNOWN LIMIT — text only. Images and documents contribute nothing, so an
|
|
215
|
+
* image-heavy (or image-only) request UNDER-counts, materially. This is
|
|
216
|
+
* deliberate: Anthropic prices an image at roughly `(w × h) / 750` tokens, and
|
|
217
|
+
* neither dimension is recoverable from a base64 blob or a URL without decoding
|
|
218
|
+
* or fetching it — so any number we put there would be fabricated, and a
|
|
219
|
+
* fabricated count is worse than a knowingly-low one for a fallback whose
|
|
220
|
+
* consumers use it as a context-window indicator. Counting the blob's own length
|
|
221
|
+
* (what {@link estimateBodyTokens} does) is not a substitute: it tracks encoding
|
|
222
|
+
* size, not token cost, and overstates by orders of magnitude.
|
|
223
|
+
*/
|
|
224
|
+
export const estimateAnthropicInputTokens = async (body, encoding = DEFAULT_ENCODING) => {
|
|
225
|
+
const joined = visibleTranscriptOf(body);
|
|
226
|
+
if (joined.length === 0)
|
|
227
|
+
return 0;
|
|
228
|
+
const counter = await getTokenCounter(encoding);
|
|
229
|
+
return countTranscript(joined, counter);
|
|
230
|
+
};
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Recovery for a replayed `reasoning.encrypted_content` the upstream can't
|
|
3
|
+
* decrypt. The OpenAI Responses backend (ChatGPT/Codex) and the Grok chat
|
|
4
|
+
* proxy only decrypt encrypted reasoning THEY minted for that
|
|
5
|
+
* account+model+session. On a fallback that switches account/model, the prior
|
|
6
|
+
* hop's encrypted reasoning replays into a hop that can't decrypt it and the
|
|
7
|
+
* backend 400s. Stripping the resumable state and retrying makes the request
|
|
8
|
+
* succeed (the model just re-reasons) instead of surfacing a hard error — the
|
|
9
|
+
* exact recovery the official partner client performs
|
|
10
|
+
* (audit 2026-07-14-codex-upstream-wire §F2).
|
|
11
|
+
*/
|
|
12
|
+
export declare const isEncryptedContentError: (rawBody: string) => boolean;
|
|
13
|
+
/** True when a built Responses body carries a `reasoning` input item with
|
|
14
|
+
* non-empty `encrypted_content` — i.e. stripping it would change the request
|
|
15
|
+
* (so a retry is worth attempting). */
|
|
16
|
+
export declare const responsesBodyHasEncryptedContent: (body: unknown) => boolean;
|
|
17
|
+
/** Return a copy of a built Responses body with `encrypted_content` removed
|
|
18
|
+
* from every `reasoning` input item (the item itself stays — its `summary`
|
|
19
|
+
* keeps the visible chain). The orphan `id` is dropped too: with
|
|
20
|
+
* `store: false` Codex treats a named reasoning item as a persisted lookup
|
|
21
|
+
* and 400s ("Item not found") when the ciphertext is gone. */
|
|
22
|
+
export declare const stripResponsesEncryptedContent: (body: unknown) => unknown;
|
|
@@ -0,0 +1,55 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Recovery for a replayed `reasoning.encrypted_content` the upstream can't
|
|
3
|
+
* decrypt. The OpenAI Responses backend (ChatGPT/Codex) and the Grok chat
|
|
4
|
+
* proxy only decrypt encrypted reasoning THEY minted for that
|
|
5
|
+
* account+model+session. On a fallback that switches account/model, the prior
|
|
6
|
+
* hop's encrypted reasoning replays into a hop that can't decrypt it and the
|
|
7
|
+
* backend 400s. Stripping the resumable state and retrying makes the request
|
|
8
|
+
* succeed (the model just re-reasons) instead of surfacing a hard error — the
|
|
9
|
+
* exact recovery the official partner client performs
|
|
10
|
+
* (audit 2026-07-14-codex-upstream-wire §F2).
|
|
11
|
+
*/
|
|
12
|
+
/** Does a 400 body signal an undecryptable replayed reasoning item? Matched
|
|
13
|
+
* loosely across the OpenAI + xAI phrasings (`invalid_encrypted_content`,
|
|
14
|
+
* `thinking_signature_invalid`, "could not decrypt", "failed to decrypt"). */
|
|
15
|
+
const DECRYPT_FAILURE_RE = /encrypted[_\s-]?content|thinking[_\s-]?signature|(?:could|failed|unable)[^.]*?decrypt|decrypt[^.]*?fail/i;
|
|
16
|
+
export const isEncryptedContentError = (rawBody) => rawBody.length > 0 && DECRYPT_FAILURE_RE.test(rawBody);
|
|
17
|
+
/** True when a built Responses body carries a `reasoning` input item with
|
|
18
|
+
* non-empty `encrypted_content` — i.e. stripping it would change the request
|
|
19
|
+
* (so a retry is worth attempting). */
|
|
20
|
+
export const responsesBodyHasEncryptedContent = (body) => {
|
|
21
|
+
if (body === null || typeof body !== "object")
|
|
22
|
+
return false;
|
|
23
|
+
const input = body.input;
|
|
24
|
+
if (!Array.isArray(input))
|
|
25
|
+
return false;
|
|
26
|
+
return input.some((item) => item !== null &&
|
|
27
|
+
typeof item === "object" &&
|
|
28
|
+
item.type === "reasoning" &&
|
|
29
|
+
typeof item.encrypted_content ===
|
|
30
|
+
"string" &&
|
|
31
|
+
(item.encrypted_content.length ?? 0) >
|
|
32
|
+
0);
|
|
33
|
+
};
|
|
34
|
+
/** Return a copy of a built Responses body with `encrypted_content` removed
|
|
35
|
+
* from every `reasoning` input item (the item itself stays — its `summary`
|
|
36
|
+
* keeps the visible chain). The orphan `id` is dropped too: with
|
|
37
|
+
* `store: false` Codex treats a named reasoning item as a persisted lookup
|
|
38
|
+
* and 400s ("Item not found") when the ciphertext is gone. */
|
|
39
|
+
export const stripResponsesEncryptedContent = (body) => {
|
|
40
|
+
if (body === null || typeof body !== "object")
|
|
41
|
+
return body;
|
|
42
|
+
const b = body;
|
|
43
|
+
if (!Array.isArray(b.input))
|
|
44
|
+
return body;
|
|
45
|
+
const input = b.input.map((item) => {
|
|
46
|
+
if (item === null ||
|
|
47
|
+
typeof item !== "object" ||
|
|
48
|
+
item.type !== "reasoning") {
|
|
49
|
+
return item;
|
|
50
|
+
}
|
|
51
|
+
const { encrypted_content: _dropped, id: _removedId, ...rest } = item;
|
|
52
|
+
return rest;
|
|
53
|
+
});
|
|
54
|
+
return { ...body, input };
|
|
55
|
+
};
|
|
@@ -0,0 +1,57 @@
|
|
|
1
|
+
import type { TCooldownReason, TErrorEnvelope } from "@openllmsh/sdk/protocol";
|
|
2
|
+
import { type TProviderErrorRule } from "./provider-error-rules.js";
|
|
3
|
+
/**
|
|
4
|
+
* Classifies an uncommitted upstream hop failure. Lives in
|
|
5
|
+
* `@openllmsh/sdk/wire` so the cloud orchestrator and coreless daemon walker
|
|
6
|
+
* share one policy.
|
|
7
|
+
*
|
|
8
|
+
* Once the gateway has validated the inbound request, an upstream rejection
|
|
9
|
+
* cannot prove that every other provider/model will reject the canonical
|
|
10
|
+
* request. Therefore every non-abort failure is classified as `transient`
|
|
11
|
+
* and the cooldown-policy action (`walk_no_cool` / `cool_and_advance` /
|
|
12
|
+
* `retry_in_place` / `surface`) decides whether the walk continues. Only
|
|
13
|
+
* request-deterministic reasons (`request_invalid`, `unprocessable`) surface
|
|
14
|
+
* immediately; capability / unfamiliar 4xx (`upstream_rejection`) and
|
|
15
|
+
* oversized payloads (`payload_too_large`) walk without cooling so a sibling
|
|
16
|
+
* model can still serve. This decision is deliberately independent of
|
|
17
|
+
* provider names for the defaults; optional `provider` rules may refine the
|
|
18
|
+
* reason tag before those defaults run.
|
|
19
|
+
*
|
|
20
|
+
* A committed response never reaches this classifier. After output has been
|
|
21
|
+
* committed, the serving transport owns any failure and cannot restart on a
|
|
22
|
+
* different model safely.
|
|
23
|
+
*
|
|
24
|
+
* Some Envoy edge-connect failures wear a 5xx status despite describing a
|
|
25
|
+
* network failure between the edge and its upstream. The narrow conjunction
|
|
26
|
+
* below keeps those failures from cooling a model, while ordinary vendor 5xx
|
|
27
|
+
* responses remain `server_error` and retain their cooldown.
|
|
28
|
+
*/
|
|
29
|
+
export type TFallbackClass = {
|
|
30
|
+
readonly kind: "transient";
|
|
31
|
+
readonly reason: TCooldownReason;
|
|
32
|
+
} | {
|
|
33
|
+
readonly kind: "abort";
|
|
34
|
+
};
|
|
35
|
+
export type TClassifierInput = {
|
|
36
|
+
readonly status: number;
|
|
37
|
+
readonly envelope?: TErrorEnvelope | undefined;
|
|
38
|
+
readonly providerFormat: "openai" | "anthropic";
|
|
39
|
+
readonly aborted: boolean;
|
|
40
|
+
readonly provider?: string | undefined;
|
|
41
|
+
};
|
|
42
|
+
/** True when an error body says the account is out of quota (the same
|
|
43
|
+
* phrasing that upgrades a 401/402/403/429 to `quota_exhausted`). */
|
|
44
|
+
export declare const isQuotaExhaustedText: (text: string) => boolean;
|
|
45
|
+
/**
|
|
46
|
+
* The vendor error phrasings that mean "request too large for this model's
|
|
47
|
+
* context window". Exported so callers that classify a raw upstream body — e.g.
|
|
48
|
+
* the daemon walker's last-resort compaction trigger — reuse the SAME matcher
|
|
49
|
+
* instead of maintaining a drifting copy.
|
|
50
|
+
*/
|
|
51
|
+
export declare const CONTEXT_OVERFLOW_BODY: RegExp;
|
|
52
|
+
/**
|
|
53
|
+
* Pure reason tagging. Accepts an optional `rules` override so tests can
|
|
54
|
+
* exercise provider rules without mutating the empty shipped table.
|
|
55
|
+
*/
|
|
56
|
+
export declare const reasonFor: (i: TClassifierInput, rules?: readonly TProviderErrorRule[]) => TCooldownReason;
|
|
57
|
+
export declare const classifyHopError: (i: TClassifierInput, rules?: readonly TProviderErrorRule[]) => TFallbackClass;
|
|
@@ -0,0 +1,97 @@
|
|
|
1
|
+
import { errorRulesFor, providerRuleMatches, } from "./provider-error-rules.js";
|
|
2
|
+
const QUOTA_BODY = /usage limit|usage balance|balance exhausted|quota|billing cycle|upgrade your plan|out of credits|purchase extra usage|insufficient_quota|insufficient (?:credits?|balance|funds)|credit balance|payment required|top[\s-]?up (?:your )?(?:credits?|balance|account)/i;
|
|
3
|
+
/** Request-shape failures that every model would reject identically.
|
|
4
|
+
* Capability / model-local refusals ("Unsupported parameter: …",
|
|
5
|
+
* "Unknown parameter: …") stay out — those walk as `upstream_rejection`. */
|
|
6
|
+
const REQUEST_INVALID_BODY = /missing required parameter|could not parse the json|invalid json body|extra inputs are not permitted/i;
|
|
7
|
+
/** True when an error body says the account is out of quota (the same
|
|
8
|
+
* phrasing that upgrades a 401/402/403/429 to `quota_exhausted`). */
|
|
9
|
+
export const isQuotaExhaustedText = (text) => QUOTA_BODY.test(text);
|
|
10
|
+
/**
|
|
11
|
+
* The vendor error phrasings that mean "request too large for this model's
|
|
12
|
+
* context window". Exported so callers that classify a raw upstream body — e.g.
|
|
13
|
+
* the daemon walker's last-resort compaction trigger — reuse the SAME matcher
|
|
14
|
+
* instead of maintaining a drifting copy.
|
|
15
|
+
*/
|
|
16
|
+
export const CONTEXT_OVERFLOW_BODY = /maximum (prompt|context) length|exceeds the context window|too many tokens|reduce the length of the messages|prompt is too long|exceeds? (?:the )?model token limit|token limit: [\d,]+/i;
|
|
17
|
+
// Envoy's first sentence is version-invariant; the trailing reset-reason
|
|
18
|
+
// clause drifts between Envoy versions ("reset reason: …" vs "retried and the
|
|
19
|
+
// latest reset reason: …"), so the matcher pins only the first sentence.
|
|
20
|
+
const ENVOY_CONNECT_ERROR_BODY = [
|
|
21
|
+
/upstream connect error/i,
|
|
22
|
+
/disconnect\/reset before headers/i,
|
|
23
|
+
];
|
|
24
|
+
const isEnvoyConnectError = (message) => ENVOY_CONNECT_ERROR_BODY.every((pattern) => pattern.test(message));
|
|
25
|
+
/**
|
|
26
|
+
* Pure reason tagging. Accepts an optional `rules` override so tests can
|
|
27
|
+
* exercise provider rules without mutating the empty shipped table.
|
|
28
|
+
*/
|
|
29
|
+
export const reasonFor = (i, rules = errorRulesFor(i.provider)) => {
|
|
30
|
+
if (i.envelope?.error?.code === "content_filter")
|
|
31
|
+
return "content_filter";
|
|
32
|
+
// A 2xx the runner still couldn't use (an undecodable 200): the provider
|
|
33
|
+
// ACCEPTED the request and may have billed it, so walking on would send —
|
|
34
|
+
// and bill — it again. Only an explicit in-band refusal (content_filter,
|
|
35
|
+
// above) walks from a 2xx; no body matcher or provider rule may turn an
|
|
36
|
+
// accepted request into a walk. 3xx falls through to the normal rules
|
|
37
|
+
// (typically `upstream_rejection` → walk).
|
|
38
|
+
if (i.status >= 200 && i.status < 300)
|
|
39
|
+
return "accepted_unreadable";
|
|
40
|
+
const message = i.envelope?.error?.message ?? "";
|
|
41
|
+
if (CONTEXT_OVERFLOW_BODY.test(message))
|
|
42
|
+
return "context_overflow";
|
|
43
|
+
for (const rule of rules) {
|
|
44
|
+
if (providerRuleMatches(rule, {
|
|
45
|
+
status: i.status,
|
|
46
|
+
code: i.envelope?.error?.code,
|
|
47
|
+
message,
|
|
48
|
+
})) {
|
|
49
|
+
return rule.reason;
|
|
50
|
+
}
|
|
51
|
+
}
|
|
52
|
+
if (i.status === 0)
|
|
53
|
+
return "network";
|
|
54
|
+
if (i.status === 408)
|
|
55
|
+
return "timeout";
|
|
56
|
+
if (i.status === 429) {
|
|
57
|
+
return QUOTA_BODY.test(message) ? "quota_exhausted" : "rate_limit";
|
|
58
|
+
}
|
|
59
|
+
if (i.status >= 500) {
|
|
60
|
+
return isEnvoyConnectError(message) ? "network" : "server_error";
|
|
61
|
+
}
|
|
62
|
+
// 426 Upgrade Required: the upstream refuses THIS client (xAI's CLI proxy
|
|
63
|
+
// gates on the device's installed grok version). Like an auth failure it is
|
|
64
|
+
// hop-local — another provider/device can serve — so it cools and advances
|
|
65
|
+
// instead of falling into a catch-all that used to surface and halt the chain.
|
|
66
|
+
if (i.status === 401 ||
|
|
67
|
+
i.status === 403 ||
|
|
68
|
+
i.status === 407 ||
|
|
69
|
+
i.status === 426) {
|
|
70
|
+
return QUOTA_BODY.test(message) ? "quota_exhausted" : "auth";
|
|
71
|
+
}
|
|
72
|
+
if (i.status === 402) {
|
|
73
|
+
return QUOTA_BODY.test(message) ? "quota_exhausted" : "payment";
|
|
74
|
+
}
|
|
75
|
+
if (i.status === 404)
|
|
76
|
+
return "not_found";
|
|
77
|
+
if (i.status === 413)
|
|
78
|
+
return "payload_too_large";
|
|
79
|
+
if (i.status === 422)
|
|
80
|
+
return "unprocessable";
|
|
81
|
+
// Any remaining 4xx: quota / request-invalid body matchers before the
|
|
82
|
+
// catch-all. A credits-exhausted OpenRouter 400 must cool-and-advance, not
|
|
83
|
+
// stop the walk; a missing-parameter 400 still surfaces.
|
|
84
|
+
if (i.status >= 400 && i.status < 500) {
|
|
85
|
+
if (QUOTA_BODY.test(message))
|
|
86
|
+
return "quota_exhausted";
|
|
87
|
+
if (REQUEST_INVALID_BODY.test(message))
|
|
88
|
+
return "request_invalid";
|
|
89
|
+
}
|
|
90
|
+
return "upstream_rejection";
|
|
91
|
+
};
|
|
92
|
+
export const classifyHopError = (i, rules) => i.aborted
|
|
93
|
+
? { kind: "abort" }
|
|
94
|
+
: {
|
|
95
|
+
kind: "transient",
|
|
96
|
+
reason: reasonFor(i, rules ?? errorRulesFor(i.provider)),
|
|
97
|
+
};
|
|
@@ -0,0 +1,20 @@
|
|
|
1
|
+
import type { TErrorEnvelope } from "@openllmsh/sdk/protocol";
|
|
2
|
+
/**
|
|
3
|
+
* Short single-line snippet from an upstream body (max chars, with ellipsis).
|
|
4
|
+
* Shared by the envelope parser (classification) and the daemon walker
|
|
5
|
+
* (trail/UI headers) so the trim/ellipsis rules cannot drift.
|
|
6
|
+
*/
|
|
7
|
+
export declare const bodySnippet: (raw: string, max?: number) => string;
|
|
8
|
+
/**
|
|
9
|
+
* Best-effort error-envelope extraction for reason tagging. Parsing is never
|
|
10
|
+
* a routing gate: an unknown provider body still represents an uncommitted
|
|
11
|
+
* candidate failure and therefore walks.
|
|
12
|
+
*
|
|
13
|
+
* A STRING-valued `error` is lifted into the canonical envelope: xAI's CLI
|
|
14
|
+
* proxy answers an exhausted balance with `{"error":"Grok Build usage balance
|
|
15
|
+
* exhausted"}`, and dropping that prose left the classifier with an empty
|
|
16
|
+
* message — tagging the hop `payment` (60s) instead of `quota_exhausted`.
|
|
17
|
+
* Non-JSON bodies become a truncated snippet with type `upstream_error`
|
|
18
|
+
* (500 chars so a late quota phrase still classifies).
|
|
19
|
+
*/
|
|
20
|
+
export declare const errorEnvelopeFrom: (raw: string) => TErrorEnvelope | undefined;
|