| 1 | // Package provider defines the model-backend abstraction and a registry mappinga provider "kind" toafactory. |
| 2 | // Concrete implementations live in subpackages |
| 3 | // (e.g. provider/openai) and self-register via init(). |
| 4 | // Thecoreresolvesprovidersbykindfromconfigandneverhardcodes a specific model. |
| 5 | package provider |
| 6 | |
| 7 | import ( |
| 8 | "context" |
| 9 | "crypto/sha256" |
| 10 | "encoding/hex" |
| 11 | "encoding/json" |
| 12 | "errors" |
| 13 | "fmt" |
| 14 | "io" |
| 15 | "net" |
| 16 | "net/http" |
| 17 | "slices" |
| 18 | "sort" |
| 19 | "strings" |
| 20 | "syscall" |
| 21 | "unicode" |
| 22 | |
| 23 | "reasonix/internal/attachment" |
| 24 | "reasonix/internal/nilutil" |
| 25 | ) |
| 26 | |
| 27 | // Role is the role of a message. |
| 28 | type Role string |
| 29 | |
| 30 | const ( |
| 31 | RoleSystem Role = "system" |
| 32 | RoleUser Role = "user" |
| 33 | RoleAssistant Role = "assistant" |
| 34 | RoleTool Role = "tool" |
| 35 | ) |
| 36 | |
| 37 | // LocalOnlyToolName/ID make display-only records safe when a newer transcriptis opened by an older |
| 38 | // Reasonixbinary that does not know Message.LocalOnly. |
| 39 | // Oldwirenormalizationtreatsthisunmatchedtoolresultasanorphananddropsitinsteadofreplayingpartialcontenttothemodel. |
| 40 | const ( |
| 41 | LocalOnlyToolName = "__reasonix_local_only__" |
| 42 | LocalOnlyToolID = "__reasonix_local_only__" |
| 43 | ) |
| 44 | |
| 45 | // Message is a single conversation message. |
| 46 | type Message struct { |
| 47 | Role Role `json:"role"` |
| 48 | // ID is local transcript identity (stable across saves, reloads, and log |
| 49 | // branches). Adapters never copy it to the wire; older readers ignore it. |
| 50 | ID string `json:"id,omitempty"` |
| 51 | // Origin distinguishes real user input from host-generated user-role protocol |
| 52 | // messages. omitempty keeps legacy sessions readable by previous releases. |
| 53 | Origin MessageOrigin `json:"origin,omitempty"` |
| 54 | // Content is the provider-visible conversation content. |
| 55 | // Keepingthislegacyfieldprovider-visiblepreservesreplay for older CLI/Desktop releases. |
| 56 | Content string `json:"content,omitempty"` |
| 57 | // RawContent holds the full local original when it differs from Content. |
| 58 | // Provider projections always strip it; bounded |
| 59 | // Contentisthestablewirerepresentationandkeepssessionfilessafefor older readers. |
| 60 | RawContent string `json:"raw_content,omitempty"` |
| 61 | // ProviderContent is a transitional field written by early Context Engine v2 |
| 62 | // builds. Loaders migrate it into Content/RawContent before normal use. |
| 63 | ProviderContent string `json:"provider_content,omitempty"` |
| 64 | Images []string `json:"images,omitempty"` // vision refs: data URLs, http(s) image URLs, or Files API file-api- ids; embedded only for vision-capable models |
| 65 | // ImageInputs is the durable ordered image payload for new messages. |
| 66 | // Images remains the old-data read path. A message must not carry both. |
| 67 | ImageInputs []attachment.ImageInput `json:"image_inputs,omitempty"` |
| 68 | ReasoningContent string `json:"reasoning_content,omitempty"` // assistant: thinking-mode chain-of-thought, round-tripped on multi-turn |
| 69 | // ReasoningID is the provider-issued reasoning-item id (OpenAI Responses: |
| 70 | // Reasoning.id is required on input items), |
| 71 | // capturedfromthestreamedoutputitemandround-trippedbackintolaterinputs. |
| 72 | ReasoningID string `json:"reasoning_id,omitempty"` |
| 73 | // ReasoningStatus is the final status of the reasoning item |
| 74 | // ("in_progress" | "completed") as issued by the server's done event, |
| 75 | // round-tripped back into the input alongside ReasoningID. |
| 76 | ReasoningStatus string `json:"reasoning_status,omitempty"` |
| 77 | // ReasoningSignature is an opaque, provider-issued proof that ReasoningContentis genuine model output. |
| 78 | // Anthropic requires the signed thinking block be replayed on the next turn when a toolcallfollowedthinking; |
| 79 | // providers without signed reasoning (e.g. the openai-compatible ones) leave it empty. |
| 80 | ReasoningSignature string `json:"reasoning_signature,omitempty"` |
| 81 | ReasoningState ReasoningState `json:"reasoning_state,omitempty"` |
| 82 | ThinkingBlocks []ThinkingBlock `json:"thinking_blocks,omitempty"` |
| 83 | ToolCalls []ToolCall `json:"tool_calls,omitempty"` // set by assistant |
| 84 | // ResponsesItems preserves provider-issued Responses API output items forstateless replay. |
| 85 | // omitemptykeepsoldsession files byte-compatible. |
| 86 | ResponsesItems []json.RawMessage `json:"responses_items,omitempty"` |
| 87 | ServerSearch []ServerSearchCall `json:"server_search,omitempty"` // cards + Anthropic replay; omitempty |
| 88 | ToolCallID string `json:"tool_call_id,omitempty"` // links a tool result to its call |
| 89 | Name string `json:"name,omitempty"` // tool message: tool name |
| 90 | MemoryCitations []MemoryCitation `json:"memoryCitations,omitempty"` // local UI metadata; provider requests ignore it |
| 91 | WorkDurationMs int64 `json:"workDurationMs,omitempty"` // local UI metadata; provider requests ignore it |
| 92 | CreatedAt int64 `json:"createdAt,omitempty"` // local UI metadata; unix milliseconds; stripped before provider requests |
| 93 | Edited bool `json:"edited,omitempty"` // local UI metadata; provider requests ignore it |
| 94 | Original string `json:"original,omitempty"` // user prompt before inline edit |
| 95 | // LocalOnly marks durable transcript content that must never be sent to amodel provider. |
| 96 | // Interruptedstreamingoutputusesitsoeveryfrontendcanreplaywhattheusersawwithoutfeedingpartialreasoningortool-callargumentsbackintothenextrequest. |
| 97 | LocalOnly bool `json:"local_only,omitempty"` |
| 98 | DecisionReceipt *DecisionReceipt `json:"decision_receipt,omitempty"` |
| 99 | // DecisionReceipts are local-only metadata attached to a provider-visiblemessage. |
| 100 | // Keepingthemontheexistingassistantrecordpreservestheassistant/tool-resultadjacencyrequiredbycurrentandolderreaders. |
| 101 | // ModelMessages strips the field before handing requests to providers. |
| 102 | DecisionReceipts []*DecisionReceipt `json:"decision_receipts,omitempty"` |
| 103 | InterruptedTurn *InterruptedTurnRecovery `json:"interrupted_turn,omitempty"` |
| 104 | // FinalReadinessRecovery is durable host state on a LocalOnly sentinel. |
| 105 | // ModelMessages removes it before provider serialization. |
| 106 | FinalReadinessRecovery *FinalReadinessRecovery `json:"final_readiness_recovery,omitempty"` |
| 107 | ProtocolRecovery json.RawMessage `json:"protocol_recovery,omitempty"` |
| 108 | ReadPause *ReadPause `json:"read_pause,omitempty"` |
| 109 | // ToolExecution is local shell UI metadata on tool-result messages. It ispersisted for |
| 110 | // Desktop/CLI/Servecards and stripped by |
| 111 | // ModelMessagesbeforeanyproviderrequestsotoolschemasandprompt-cacheprefixes stay stable. |
| 112 | ToolExecution *ToolExecution `json:"tool_execution,omitempty"` |
| 113 | ToolRunState ToolRunState `json:"tool_run_state,omitempty"` |
| 114 | // PresentedFiles is trusted, versioned host metadata produced by the |
| 115 | // built-in present tool. It is persisted for replay and stripped from |
| 116 | // provider requests. Older binaries ignore this optional object. |
| 117 | PresentedFiles *PresentedFilesMetadata `json:"presented_files,omitempty"` |
| 118 | // ReadResult is a persisted, host-only reader delivery envelope for diagnostics. |
| 119 | // ModelMessages strips it; provider serializers must never emit it on the wire. |
| 120 | ReadResult json.RawMessage `json:"read_result,omitempty"` |
| 121 | // ToolDiagnostic is persisted host recovery data, stripped by ModelMessages. |
| 122 | ToolDiagnostic json.RawMessage `json:"tool_diagnostic,omitempty"` |
| 123 | // ReadCompletion is a display-only terminal coverage receipt. It authorizes |
| 124 | // neither historical writes nor continuation in another run. |
| 125 | ReadCompletion *ReadCompletion `json:"read_completion,omitempty"` |
| 126 | // MCPApp is the local MCP Apps presentation for results from App-capableservers. Persisted for |
| 127 | // Desktopcardsand stripped by ModelMessages; |
| 128 | // provider serializers must never emit it on the wire. |
| 129 | MCPApp *MCPAppPresentation `json:"mcp_app,omitempty"` |
| 130 | // VisionSummary is durable local metadata generated by an optional imageunderstanding prepass. |
| 131 | // Itiscopiedinto provider-visible Content by theturn owner and stripped at the provider boundary. |
| 132 | VisionSummary *VisionSummary `json:"vision_summary,omitempty"` |
| 133 | } |
| 134 | |
| 135 | // VisionSummary is a bounded, provider-independent description of one userturn's image attachments. |
| 136 | // Itcontains no image bytes, paths, or reasoning. |
| 137 | type VisionSummary struct { |
| 138 | Version int `json:"version"` |
| 139 | PromptVersion string `json:"prompt_version"` |
| 140 | ModelRef string `json:"model_ref"` |
| 141 | ImageDigests []string `json:"image_digests"` |
| 142 | Summary string `json:"summary"` |
| 143 | CreatedAt int64 `json:"created_at"` |
| 144 | } |
| 145 | |
| 146 | // ToolExecution is host-local shell metadata mirrored from tool.ShellExecution. |
| 147 | // Provider serializers must never emit this object on the wire. |
| 148 | type ToolExecution struct { |
| 149 | Kind string `json:"kind,omitempty"` |
| 150 | Shell string `json:"shell,omitempty"` |
| 151 | ShellVersion string `json:"shellVersion,omitempty"` |
| 152 | Platform string `json:"platform,omitempty"` |
| 153 | SupportsAndAnd bool `json:"supportsAndAnd"` |
| 154 | State string `json:"state,omitempty"` |
| 155 | FailurePhase string `json:"failurePhase,omitempty"` |
| 156 | ExitCode *int `json:"exitCode,omitempty"` |
| 157 | OutputTail string `json:"outputTail,omitempty"` |
| 158 | MutationRisk string `json:"mutationRisk,omitempty"` |
| 159 | Verification string `json:"verification,omitempty"` |
| 160 | DurationMs int64 `json:"durationMs,omitempty"` |
| 161 | } |
| 162 | |
| 163 | // DecisionReceipt is durable, provider-excluded evidence of a user-ownedapproval decision. |
| 164 | // Itintentionallycontains only bounded labels and theoutcome, |
| 165 | // neverfree-formguidanceorprovider-visiblecontent. |
| 166 | type DecisionReceipt struct { |
| 167 | ID string `json:"id"` |
| 168 | Kind string `json:"kind"` |
| 169 | Tool string `json:"tool,omitempty"` |
| 170 | Subject string `json:"subject,omitempty"` |
| 171 | Outcome string `json:"outcome"` |
| 172 | } |
| 173 | |
| 174 | // MemoryCitation is local display metadata for memories that influenced anassistant turn. |
| 175 | // Providerimplementations must not forward it to model APIs. |
| 176 | type MemoryCitation struct { |
| 177 | ID string `json:"id,omitempty"` |
| 178 | Source string `json:"source"` |
| 179 | LineStart int `json:"lineStart,omitempty"` |
| 180 | LineEnd int `json:"lineEnd,omitempty"` |
| 181 | Note string `json:"note,omitempty"` |
| 182 | Kind string `json:"kind,omitempty"` |
| 183 | } |
| 184 | |
| 185 | // ParseImageDataURL splits a `data:<media-type>;base64,<payload>` URL into itsmedia type and base64 payload. |
| 186 | // ok is false for anything that isn't a base64 |
| 187 | // data URL — providers that need the split (Anthropic) skip those silently. |
| 188 | func ParseImageDataURL(dataURL string) (mediaType, base64Data string, ok bool) { |
| 189 | rest, found := strings.CutPrefix(dataURL, "data:") |
| 190 | if !found { |
| 191 | return "", "", false |
| 192 | } |
| 193 | meta, payload, found := strings.Cut(rest, ",") |
| 194 | if !found { |
| 195 | return "", "", false |
| 196 | } |
| 197 | mt, found := strings.CutSuffix(meta, ";base64") |
| 198 | if !found || mt == "" { |
| 199 | return "", "", false |
| 200 | } |
| 201 | return mt, payload, true |
| 202 | } |
| 203 | |
| 204 | type ToolCall struct { |
| 205 | Recovery *ToolCallRecord `json:"tool_recovery,omitempty"` // local execution evidence; stripped from model input |
| 206 | WriteIntents []json.RawMessage `json:"write_intents,omitempty"` // local versioned evidence, stripped from model input |
| 207 | ID string `json:"id"` |
| 208 | Name string `json:"name"` |
| 209 | Arguments string `json:"arguments"` |
| 210 | // ThoughtSignature is an opaque Gemini-issued proof attached to a functioncall. OpenAI-compatible |
| 211 | // Geminiendpoints require it on message replay. |
| 212 | ThoughtSignature string `json:"thought_signature,omitempty"` |
| 213 | Diff string `json:"diff,omitempty"` |
| 214 | Added int `json:"added,omitempty"` |
| 215 | Removed int `json:"removed,omitempty"` |
| 216 | // Resolved* fields are Reasonix-local display metadata for stable proxycalls such as use_capability. |
| 217 | // Provider request builders deliberatelyserialize only provider-visible fields, |
| 218 | // sothesevaluesneveraltertheprovider-visible conversation orprompt-cache prefix. |
| 219 | ResolvedName string `json:"resolved_name,omitempty"` |
| 220 | CapabilityID string `json:"capability_id,omitempty"` |
| 221 | ResolvedReadOnly *bool `json:"resolved_read_only,omitempty"` |
| 222 | } |
| 223 | |
| 224 | // ToolSchema is a tool definition exposed to the model. Parameters is JSON Schema. |
| 225 | type ToolSchema struct { |
| 226 | Name string `json:"name"` |
| 227 | Description string `json:"description"` |
| 228 | Parameters json.RawMessage `json:"parameters"` |
| 229 | Deferred bool `json:"deferred,omitempty"` |
| 230 | Strict bool `json:"strict,omitempty"` |
| 231 | Namespace string `json:"namespace,omitempty"` |
| 232 | } |
| 233 | |
| 234 | // Request is a single completion request. |
| 235 | type Request struct { |
| 236 | Messages []Message |
| 237 | Tools []ToolSchema |
| 238 | Temperature *float64 // nil = omit; non-nil = send the value, including 0 |
| 239 | MaxTokens int |
| 240 | // ResponseFormat, when non-nil, asks the endpoint for structured JSONoutput (Responses: |
| 241 | // text.format.type=json_object). Nil omits the fieldentirely — |
| 242 | // thecommonpathmuststaybyte-stableforpromptcaching. |
| 243 | ResponseFormat *ResponseFormat `json:"ResponseFormat,omitempty"` |
| 244 | EffortOverride string `json:"EffortOverride,omitempty"` // per-call reasoning-depth override; adapters apply it only when the endpoint's effort vocabulary accepts it |
| 245 | ToolSearch *ToolSearch `json:"-"` |
| 246 | } |
| 247 | |
| 248 | // ResponseFormat asks a provider to constrain its output shape. |
| 249 | type ResponseFormat struct { |
| 250 | // Type is the structured format: "json_object" is the only shape the |
| 251 | // Responses endpoints currently define (MiMo/DashScope/OpenAI). |
| 252 | Type string `json:"type"` |
| 253 | } |
| 254 | |
| 255 | // Auto ladder for max_output_tokens=0. Bounds completion only; never compact_ratio. |
| 256 | // Official DeepSeek does not use this ladder: Chat/Responses omit the field |
| 257 | // (server 384K ceiling) and Anthropic sends DeepSeekMaxOutputTokens. |
| 258 | const ( |
| 259 | DefaultOrdinaryOutputTokens = 16 * 1024 // non-reasoning / non-DeepSeek |
| 260 | DefaultReasoningOutputTokens = 32 * 1024 // ordinary reasoning / MiMo |
| 261 | DefaultHighReasoningOutputTokens = 64 * 1024 // high/max effort on non-DeepSeek |
| 262 | DefaultHighOutputTokens = 128 * 1024 // explicit only; never auto |
| 263 | // DeepSeekMaxOutputTokens is the official V4 Flash/Pro completion ceiling. |
| 264 | // Pricing page: 输出长度最大 384K. K is decimal thousands, matching thedocumented 1M context = 1,000,000 tokens. |
| 265 | // Anthropic requires max_tokens. |
| 266 | DeepSeekMaxOutputTokens = 384_000 |
| 267 | ) |
| 268 | |
| 269 | // AutoOutputBudget maps max_output_tokens=0 to 16K/32K/64K for non-DeepSeekvendors. Official |
| 270 | // DeepSeekomitsthe field (Chat/Responses) or sends |
| 271 | // DeepSeekMaxOutputTokens (Anthropic). |
| 272 | func AutoOutputBudget(reasoningEnabled bool, effort string) int { |
| 273 | if !reasoningEnabled { |
| 274 | return DefaultOrdinaryOutputTokens |
| 275 | } |
| 276 | switch strings.ToLower(strings.TrimSpace(effort)) { |
| 277 | case "high", "max": |
| 278 | return DefaultHighReasoningOutputTokens |
| 279 | default: |
| 280 | return DefaultReasoningOutputTokens |
| 281 | } |
| 282 | } |
| 283 | |
| 284 | // interruptedToolResult stands in for a tool result that never landed — an |
| 285 | // assistant tool_calls turn whose execution was cut short (interrupt, crash) and |
| 286 | // later resumed. Sending such a turn unanswered trips the OpenAI/DeepSeek 400 |
| 287 | // "An assistant message with 'tool_calls' must be followed by tool messages |
| 288 | // responding to each 'tool_call_id'". |
| 289 | const interruptedToolResult = "[no result: the previous turn was interrupted before this tool call completed]" |
| 290 | |
| 291 | // SanitizeToolPairing is the provider-side alias for NormalizeMessages. |
| 292 | // Itrepairsahistorysoitsatisfiesthetool-call contract the OpenAI-compatible and |
| 293 | // Anthropic APIs enforce (every assistant tool_calls answered, no orphan toolmessages, truncated argsclosed) |
| 294 | // right before sending it to the wire — withouttouching the stored session. |
| 295 | // Keptasadistinctnamesocallsitesread as |
| 296 | // "defensive wire prep" rather than "session mutation". |
| 297 | func SanitizeToolPairing(msgs []Message) []Message { return NormalizeMessages(msgs) } |
| 298 | |
| 299 | // NormalizeMessages repairs a conversation history so it satisfies the tool-callcontract the |
| 300 | // OpenAI-compatible and Anthropic APIs enforce: |
| 301 | // everyassistanttool_callsentrymustbeansweredbyafollowingtoolmessage for its id, |
| 302 | // andatoolmessagemustfollowsuch a call. It backfills a placeholder result for anyunanswered call |
| 303 | // (sotheturnstaysintact), |
| 304 | // dropsorphan tool messages, |
| 305 | // backfills empty tool-call names from their results (#4727 — |
| 306 | // oldsessionssavedbeforeadde2d3ecancarryanemptyname), and closes truncated call-argument |
| 307 | // JSON (DeepSeek 400s on replayed half-streamed args, #3953). |
| 308 | // |
| 309 | // This is the wire-safe entry point for provider requests. Stored session loadsuse |
| 310 | // NormalizeSessionMessagessotheycansharetheassistant-turnrepairswithoutdeletingstandalonetoolmessagesthatmustround-tripthroughreasonix |
| 311 | // --resume. |
| 312 | // |
| 313 | // A well-formed history — no unanswered calls, no orphan results, no empty tool- |
| 314 | // call names, no truncated args — returns the input slice unchanged (same backingarray, zero allocation). |
| 315 | // This keeps the prefix-cache key stable for healthysessions and makes repeated normalization cheap. |
| 316 | func NormalizeMessages(msgs []Message) []Message { |
| 317 | return normalizeMessages(msgs, true) |
| 318 | } |
| 319 | |
| 320 | // NormalizeSessionMessages applies only repairs that are safe to persist in asaved session. |
| 321 | // Itsharesassistant-turn repairs with NormalizeMessages, |
| 322 | // butpreservesexistingtoolmessagesinsteadofdroppingorreordering them so |
| 323 | // Save/LoadSession remains a byte-for-byte conversation round trip for historiesthat were already on disk. |
| 324 | func NormalizeSessionMessages(msgs []Message) []Message { |
| 325 | return normalizeMessages(attachStandaloneDecisionReceipts(msgs), false) |
| 326 | } |
| 327 | |
| 328 | // attachStandaloneDecisionReceipts migrates the short-lived receipt encodingthat stored a |
| 329 | // LocalOnlyassistantmessage between an assistant tool call andits result. |
| 330 | // Foldingthatmetadataintothelatestassistantmessagerepairsalready-writtensessionsbeforetool-pairnormalizationcanfabricateaplaceholder. |
| 331 | // Healthyhistoriesreturntheoriginalsliceunchanged. |
| 332 | func attachStandaloneDecisionReceipts(msgs []Message) []Message { |
| 333 | target := -1 |
| 334 | needsMigration := false |
| 335 | for i, m := range msgs { |
| 336 | switch { |
| 337 | case m.Role == RoleUser && !m.LocalOnly: |
| 338 | target = -1 |
| 339 | case m.Role == RoleAssistant && !m.LocalOnly: |
| 340 | target = i |
| 341 | case target >= 0 && m.LocalOnly && m.DecisionReceipt != nil: |
| 342 | needsMigration = true |
| 343 | } |
| 344 | if needsMigration { |
| 345 | break |
| 346 | } |
| 347 | } |
| 348 | if !needsMigration { |
| 349 | return msgs |
| 350 | } |
| 351 | |
| 352 | out := make([]Message, 0, len(msgs)) |
| 353 | target = -1 |
| 354 | for _, m := range msgs { |
| 355 | switch { |
| 356 | case m.Role == RoleUser && !m.LocalOnly: |
| 357 | target = -1 |
| 358 | case m.Role == RoleAssistant && !m.LocalOnly: |
| 359 | out = append(out, m) |
| 360 | target = len(out) - 1 |
| 361 | continue |
| 362 | case target >= 0 && m.LocalOnly && m.DecisionReceipt != nil: |
| 363 | receipts := append([]*DecisionReceipt(nil), out[target].DecisionReceipts...) |
| 364 | out[target].DecisionReceipts = append(receipts, m.DecisionReceipt) |
| 365 | continue |
| 366 | } |
| 367 | out = append(out, m) |
| 368 | } |
| 369 | return out |
| 370 | } |
| 371 | |
| 372 | func normalizeMessages(msgs []Message, dropOrphanTools bool) []Message { |
| 373 | if normalized, ok := tryNormalizeFastPath(msgs, dropOrphanTools); ok { |
| 374 | return normalized // well-formed: pass through without allocating |
| 375 | } |
| 376 | out := make([]Message, 0, len(msgs)) |
| 377 | for i := 0; i < len(msgs); { |
| 378 | m := msgs[i] |
| 379 | if m.LocalOnly { |
| 380 | if !dropOrphanTools { |
| 381 | out = append(out, m) |
| 382 | } |
| 383 | i++ |
| 384 | continue |
| 385 | } |
| 386 | if m.Role == RoleAssistant && len(m.ToolCalls) > 0 { |
| 387 | j := i + 1 |
| 388 | for j < len(msgs) && msgs[j].Role == RoleTool && !msgs[j].LocalOnly { |
| 389 | j++ |
| 390 | } |
| 391 | // Backfill empty tool-call names from the corresponding toolresults so the model sees which tool was invoked |
| 392 | // (#4727). |
| 393 | // The wire-format fix (openai.go) ensures empty fields arenever omitted, so this backfill is a |
| 394 | // UXimprovement, not acorrectness requirement. |
| 395 | calls := backfillToolCallNames(m.ToolCalls, msgs[i+1:j]) |
| 396 | m.ToolCalls = calls |
| 397 | out = append(out, repairToolCallArgs(m)) |
| 398 | if dropOrphanTools { |
| 399 | out = append(out, pairToolResults(calls, msgs[i+1:j])...) |
| 400 | } else { |
| 401 | out = append(out, sessionToolResults(calls, msgs[i+1:j])...) |
| 402 | } |
| 403 | i = j |
| 404 | continue |
| 405 | } |
| 406 | if m.Role == RoleTool { |
| 407 | if !dropOrphanTools { |
| 408 | out = append(out, m) |
| 409 | } |
| 410 | // Orphan tool message: provider sends drop it; session loads preserve it. |
| 411 | i++ |
| 412 | continue |
| 413 | } |
| 414 | out = append(out, m) |
| 415 | i++ |
| 416 | } |
| 417 | return out |
| 418 | } |
| 419 | |
| 420 | // tryNormalizeFastPath reports whether msgs needs no repair and, if so, |
| 421 | // returnsitas-issothecallercanskipallocating. Healthy tool-call/tool-resultturns pass through unchanged; |
| 422 | // malformed turns take the slow path. |
| 423 | func tryNormalizeFastPath(msgs []Message, dropOrphanTools bool) ([]Message, bool) { |
| 424 | for i := 0; i < len(msgs); { |
| 425 | m := msgs[i] |
| 426 | if m.LocalOnly { |
| 427 | if dropOrphanTools { |
| 428 | return nil, false |
| 429 | } |
| 430 | i++ |
| 431 | continue |
| 432 | } |
| 433 | if m.Role == RoleAssistant && len(m.ToolCalls) > 0 { |
| 434 | j := i + 1 |
| 435 | for j < len(msgs) && msgs[j].Role == RoleTool && !msgs[j].LocalOnly { |
| 436 | j++ |
| 437 | } |
| 438 | if !toolTurnWellFormed(m.ToolCalls, msgs[i+1:j]) || needsToolCallArgRepair(m.ToolCalls) { |
| 439 | return nil, false |
| 440 | } |
| 441 | i = j |
| 442 | continue |
| 443 | } |
| 444 | if m.Role == RoleTool && dropOrphanTools { |
| 445 | return nil, false |
| 446 | } |
| 447 | i++ |
| 448 | } |
| 449 | return msgs, true |
| 450 | } |
| 451 | |
| 452 | func toolTurnWellFormed(calls []ToolCall, results []Message) bool { |
| 453 | if len(calls) != len(results) { |
| 454 | return false |
| 455 | } |
| 456 | for _, tc := range calls { |
| 457 | if tc.Name == "" { |
| 458 | return false |
| 459 | } |
| 460 | } |
| 461 | for k, tc := range calls { |
| 462 | if results[k].ToolCallID != tc.ID { |
| 463 | return false |
| 464 | } |
| 465 | if results[k].Name != tc.Name { |
| 466 | return false |
| 467 | } |
| 468 | } |
| 469 | return true |
| 470 | } |
| 471 | |
| 472 | func needsToolCallArgRepair(calls []ToolCall) bool { |
| 473 | for _, tc := range calls { |
| 474 | if tc.Arguments != "" && !json.Valid([]byte(tc.Arguments)) { |
| 475 | return true |
| 476 | } |
| 477 | } |
| 478 | return false |
| 479 | } |
| 480 | |
| 481 | // repairToolCallArgs returns m with any undecodable tool-call Arguments closedinto valid JSON |
| 482 | // (copy-on-write; the caller's history is never mutated). Emptyarguments pass through — some gateways send |
| 483 | // "" for no-arg tools. |
| 484 | func repairToolCallArgs(m Message) Message { |
| 485 | broken := false |
| 486 | for _, tc := range m.ToolCalls { |
| 487 | if tc.Arguments != "" && !json.Valid([]byte(tc.Arguments)) { |
| 488 | broken = true |
| 489 | break |
| 490 | } |
| 491 | } |
| 492 | if !broken { |
| 493 | return m |
| 494 | } |
| 495 | calls := make([]ToolCall, len(m.ToolCalls)) |
| 496 | copy(calls, m.ToolCalls) |
| 497 | for i := range calls { |
| 498 | if calls[i].Arguments == "" || json.Valid([]byte(calls[i].Arguments)) { |
| 499 | continue |
| 500 | } |
| 501 | calls[i].Arguments = closeTruncatedJSON(calls[i].Arguments) |
| 502 | } |
| 503 | m.ToolCalls = calls |
| 504 | return m |
| 505 | } |
| 506 | |
| 507 | // closeTruncatedJSON best-effort completes a JSON document cut off mid-stream |
| 508 | // (unterminated string, open braces, dangling comma/colon); anything stillinvalid after closing degrades to |
| 509 | // "{}". |
| 510 | func closeTruncatedJSON(s string) string { |
| 511 | var stack []byte |
| 512 | inStr, esc := false, false |
| 513 | for i := range len(s) { |
| 514 | c := s[i] |
| 515 | if inStr { |
| 516 | switch { |
| 517 | case esc: |
| 518 | esc = false |
| 519 | case c == '\\': |
| 520 | esc = true |
| 521 | case c == '"': |
| 522 | inStr = false |
| 523 | } |
| 524 | continue |
| 525 | } |
| 526 | switch c { |
| 527 | case '"': |
| 528 | inStr = true |
| 529 | case '{': |
| 530 | stack = append(stack, '}') |
| 531 | case '[': |
| 532 | stack = append(stack, ']') |
| 533 | case '}', ']': |
| 534 | if len(stack) > 0 { |
| 535 | stack = stack[:len(stack)-1] |
| 536 | } |
| 537 | } |
| 538 | } |
| 539 | out := s |
| 540 | if esc { |
| 541 | out = out[:len(out)-1] |
| 542 | } |
| 543 | if inStr { |
| 544 | out += `"` |
| 545 | } |
| 546 | trimmed := strings.TrimRight(out, " \t\r\n") |
| 547 | switch { |
| 548 | case strings.HasSuffix(trimmed, ","): |
| 549 | out = trimmed[:len(trimmed)-1] |
| 550 | case strings.HasSuffix(trimmed, ":"): |
| 551 | out = trimmed + "null" |
| 552 | } |
| 553 | for _, v := range slices.Backward(stack) { |
| 554 | out += string(v) |
| 555 | } |
| 556 | if !json.Valid([]byte(out)) { |
| 557 | return "{}" |
| 558 | } |
| 559 | return out |
| 560 | } |
| 561 | |
| 562 | // pairToolResults answers each tool_call with its result, backfilling aplaceholder for any unanswered one. |
| 563 | // Distinct non-empty ids pair by id (soreordered results re-sort to call order); |
| 564 | // emptyorduplicateidspairbyposition instead — some gatewaysstream tool calls by index with no id, |
| 565 | // andamapkeyed on id would collapse those results into one |
| 566 | // (callorder is preservedbecause the loop appendsresults in call order). |
| 567 | func pairToolResults(calls []ToolCall, avail []Message) []Message { |
| 568 | out := make([]Message, 0, len(calls)) |
| 569 | if idDistinct(calls) { |
| 570 | byID := make(map[string]Message, len(avail)) |
| 571 | for _, r := range avail { |
| 572 | byID[r.ToolCallID] = r |
| 573 | } |
| 574 | for _, tc := range calls { |
| 575 | if r, ok := byID[tc.ID]; ok { |
| 576 | r.Name = tc.Name |
| 577 | out = append(out, r) |
| 578 | } else { |
| 579 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 580 | } |
| 581 | } |
| 582 | return out |
| 583 | } |
| 584 | for k, tc := range calls { |
| 585 | if k < len(avail) { |
| 586 | r := avail[k] |
| 587 | r.ToolCallID = tc.ID |
| 588 | r.Name = tc.Name |
| 589 | out = append(out, r) |
| 590 | } else { |
| 591 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 592 | } |
| 593 | } |
| 594 | return out |
| 595 | } |
| 596 | |
| 597 | // sessionToolResultspreserveseverystoredtoolresultandappendsplaceholdersonlyforcallsthathavenorecordedanswer. |
| 598 | // Load-timenormalizationmustnotdrop or reorder user history; |
| 599 | // providersendscanstillusepairToolResultsforstrictwireformatting. |
| 600 | func sessionToolResults(calls []ToolCall, avail []Message) []Message { |
| 601 | out := append([]Message(nil), avail...) |
| 602 | if idDistinct(calls) { |
| 603 | answered := make(map[string]struct{}, len(avail)) |
| 604 | for _, r := range avail { |
| 605 | answered[r.ToolCallID] = struct{}{} |
| 606 | } |
| 607 | for _, tc := range calls { |
| 608 | if _, ok := answered[tc.ID]; !ok { |
| 609 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 610 | } |
| 611 | } |
| 612 | return out |
| 613 | } |
| 614 | for k := len(avail); k < len(calls); k++ { |
| 615 | tc := calls[k] |
| 616 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 617 | } |
| 618 | return out |
| 619 | } |
| 620 | |
| 621 | // backfillToolCallNames returns calls with any empty Name filled in from thematching tool result (by id, |
| 622 | // then by position). Old sessions (#4727) may havesaved assistant tool-calls with an empty name; |
| 623 | // backfillinggives the modeluseful context during replay. |
| 624 | // The common case (no empty names) returns theinput unchanged without allocating. |
| 625 | // Unpairedcallskeeptheirempty name, |
| 626 | // which the wire-format fix (openai.go) handles gracefully. |
| 627 | func backfillToolCallNames(calls []ToolCall, results []Message) []ToolCall { |
| 628 | missing := false |
| 629 | for _, c := range calls { |
| 630 | if c.Name == "" { |
| 631 | missing = true |
| 632 | break |
| 633 | } |
| 634 | } |
| 635 | if !missing { |
| 636 | return calls |
| 637 | } |
| 638 | out := make([]ToolCall, len(calls)) |
| 639 | copy(out, calls) |
| 640 | if idDistinct(calls) { |
| 641 | byID := make(map[string]string, len(results)) |
| 642 | for _, r := range results { |
| 643 | if r.Name != "" { |
| 644 | byID[r.ToolCallID] = r.Name |
| 645 | } |
| 646 | } |
| 647 | for k := range out { |
| 648 | if out[k].Name == "" { |
| 649 | if n, ok := byID[out[k].ID]; ok { |
| 650 | out[k].Name = n |
| 651 | } |
| 652 | } |
| 653 | } |
| 654 | return out |
| 655 | } |
| 656 | // Fallback: positional pairing (same order as pairToolResults). |
| 657 | for k := range out { |
| 658 | if out[k].Name == "" && k < len(results) { |
| 659 | out[k].Name = results[k].Name |
| 660 | } |
| 661 | } |
| 662 | return out |
| 663 | } |
| 664 | |
| 665 | // idDistinct reports whether every call carries a non-empty id unique within thebatch — |
| 666 | // theconditionunderwhich id-keyed pairing is safe. |
| 667 | func idDistinct(calls []ToolCall) bool { |
| 668 | seen := make(map[string]struct{}, len(calls)) |
| 669 | for _, tc := range calls { |
| 670 | if tc.ID == "" { |
| 671 | return false |
| 672 | } |
| 673 | if _, dup := seen[tc.ID]; dup { |
| 674 | return false |
| 675 | } |
| 676 | seen[tc.ID] = struct{}{} |
| 677 | } |
| 678 | return true |
| 679 | } |
| 680 | |
| 681 | // ChunkType identifies the kind of a streamed increment. |
| 682 | type ChunkType int |
| 683 | |
| 684 | const ( |
| 685 | ChunkText ChunkType = iota // text delta |
| 686 | ChunkReasoning // thinking-mode reasoning delta (before the visible answer) |
| 687 | ChunkToolCallStart // a tool call has begun (ToolCall: ID+Name; args still streaming) |
| 688 | ChunkToolCallArgsDelta // progress while a call's arguments stream (ToolCall: ID+Name; ArgChars: cumulative) |
| 689 | ChunkToolCall // one complete tool call |
| 690 | ChunkUsage // token usage for the completion |
| 691 | ChunkDone // completion finished normally |
| 692 | ChunkError // an error occurred |
| 693 | ChunkResponsesItem // a complete provider-issued Responses API output item for stateless replay |
| 694 | ChunkServerSearch // provider-executed web_search; not a client tool call |
| 695 | ) |
| 696 | |
| 697 | // Usage reports token accounting for a completion. Cache hit/miss come fromeither |
| 698 | // DeepSeek'stop-levelprompt_cache_{hit,miss}_tokens or the |
| 699 | // OpenAI/MiMostandardprompt_tokens_details.cached_tokens — |
| 700 | // theopenaiprovidernormalisesbothshapesintothesefields. |
| 701 | // ReasoningTokens is the thinking-mode subset of |
| 702 | // CompletionTokens reported by thinking-capable models. |
| 703 | // FinishReasoncarriesthemodel'slastreportedchoices[0].finish_reason sotheagentcansurfaceabnormalterminations |
| 704 | // ("length", "content_filter", "repetition_truncation"). |
| 705 | // Estimated marks counts reconstructed locally because the provider's terminalusage record did not arrive; |
| 706 | // exact provider usage leaves it false. |
| 707 | type Usage struct { |
| 708 | Unknown bool `json:"unknown,omitempty"` // at least one request had no provider usage |
| 709 | PromptTokens int |
| 710 | CompletionTokens int |
| 711 | TotalTokens int |
| 712 | CacheHitTokens int // prompt tokens served from cache |
| 713 | CacheMissTokens int // prompt tokens not cached, including CacheWriteTokens |
| 714 | CacheWriteTokens int // subset of CacheMissTokens used to create provider cache entries |
| 715 | CacheWriteBilledTokens float64 // cache-write charge expressed in ordinary input-token equivalents |
| 716 | ReasoningTokens int // subset of CompletionTokens spent on chain-of-thought |
| 717 | FinishReason string // "stop", "tool_calls", "length", "content_filter", "repetition_truncation", … |
| 718 | Estimated bool |
| 719 | // RequestCount is the number of provider requests represented by thisaggregate. |
| 720 | // Zeromeansonerequestforbackward compatibility. Recoverypaths that merge multiple attempts settheexactcount. |
| 721 | RequestCount int |
| 722 | // Context* fields describe the latest single-request shape for contextgauges and rebind telemetry. Whenzero, |
| 723 | // consumers fall back to thebillable Prompt/Completion/… fields. Multi-attempt sampling recoverysets |
| 724 | // PromptTokens (etc.) tothebillable aggregate and fills Context* |
| 725 | // from the final attempt only. |
| 726 | ContextPromptTokens int |
| 727 | ContextCompletionTokens int |
| 728 | ContextReasoningTokens int |
| 729 | ContextCacheHitTokens int |
| 730 | ContextCacheMissTokens int |
| 731 | } |
| 732 | |
| 733 | // ContextFillTokens returns the latest prompt occupancy used by context gauges. |
| 734 | func (u *Usage) ContextFillTokens() int { |
| 735 | return u.LatestPromptTokens() |
| 736 | } |
| 737 | |
| 738 | // LatestPromptTokens returns the latest-attempt prompt size for context-awareruntime decisions. Falls backto |
| 739 | // PromptTokens for single-attempt legacy usage. |
| 740 | func (u *Usage) LatestPromptTokens() int { |
| 741 | if u == nil { |
| 742 | return 0 |
| 743 | } |
| 744 | if u.ContextPromptTokens > 0 { |
| 745 | return u.ContextPromptTokens |
| 746 | } |
| 747 | return u.PromptTokens |
| 748 | } |
| 749 | |
| 750 | // Pricing is a provider's per-1M-token rates, used to estimate spend. Currencyis a display symbol or |
| 751 | // ISO-like code (default "¥"). toml tags let config decode it. |
| 752 | type Pricing struct { |
| 753 | CacheHit float64 `toml:"cache_hit"` // per 1M cached prompt tokens |
| 754 | Input float64 `toml:"input"` // per 1M uncached prompt tokens |
| 755 | Output float64 `toml:"output"` // per 1M completion tokens |
| 756 | Currency string `toml:"currency"` |
| 757 | } |
| 758 | |
| 759 | // Cost estimates the spend for a usage record. Compatibility adapter only — |
| 760 | // new host code must consume billing.CostQuote instead of aggregating floats. |
| 761 | func (p *Pricing) Cost(u *Usage) float64 { |
| 762 | if p == nil || u == nil { |
| 763 | return 0 |
| 764 | } |
| 765 | // Keepthehistoricalfloatpathbyte-stableforteststhatassertexactfloatresultswithoutgoingthroughthefixed-pointquotelayer. |
| 766 | hit := u.CacheHitTokens |
| 767 | miss := u.CacheMissTokens |
| 768 | if hit+miss == 0 && u.PromptTokens > 0 { |
| 769 | miss = u.PromptTokens |
| 770 | } else if miss == 0 && hit > 0 && u.PromptTokens > hit { |
| 771 | miss = u.PromptTokens - hit |
| 772 | } |
| 773 | // CacheMissTokens intentionally remains the raw prompt-token denominatorused by cache hit-rate displays, |
| 774 | // socache writes are included there. Forcost, split those writes back out and replace themwiththeirprovider- |
| 775 | // supplied input-token equivalent (for example Anthropic's 1.25x 5-minutewrites or 2x 1-hour writes). |
| 776 | // Olderproviders leave both fields at zero andkeep the legacy one-input-rate behavior. |
| 777 | // Awritecountwithoutbilledunits also falls back to 1xforbackward compatibility. |
| 778 | write := min(max(u.CacheWriteTokens, 0), miss) |
| 779 | billedWrite := 0.0 |
| 780 | if write > 0 { |
| 781 | billedWrite = u.CacheWriteBilledTokens |
| 782 | if billedWrite <= 0 { |
| 783 | billedWrite = float64(write) |
| 784 | } |
| 785 | } |
| 786 | inputTokenUnits := float64(miss-write) + billedWrite |
| 787 | return (float64(hit)*p.CacheHit + |
| 788 | inputTokenUnits*p.Input + |
| 789 | float64(u.CompletionTokens)*p.Output) / 1e6 |
| 790 | } |
| 791 | |
| 792 | // Symbol returns the currency display symbol, defaulting to "¥". |
| 793 | func (p *Pricing) Symbol() string { |
| 794 | if p == nil || p.Currency == "" { |
| 795 | return "¥" |
| 796 | } |
| 797 | return currencySymbol(p.Currency) |
| 798 | } |
| 799 | |
| 800 | func currencySymbol(currency string) string { |
| 801 | value := strings.TrimSpace(currency) |
| 802 | if value == "" { |
| 803 | return "¥" |
| 804 | } |
| 805 | switch strings.ToLower(value) { |
| 806 | case "cny", "rmb", "yuan", "renminbi", "cnh": |
| 807 | return "¥" |
| 808 | case "usd", "dollar", "dollars", "us dollar", "us dollars", "us$": |
| 809 | return "$" |
| 810 | case "eur", "euro", "euros": |
| 811 | return "€" |
| 812 | case "gbp", "pound", "pounds", "sterling": |
| 813 | return "£" |
| 814 | case "jpy", "yen": |
| 815 | return "¥" |
| 816 | } |
| 817 | switch value { |
| 818 | case "￥", "¥": |
| 819 | return "¥" |
| 820 | case "$", "€", "£": |
| 821 | return value |
| 822 | } |
| 823 | // any embedded currency sign → keep as-is (compact symbols like A$, HK$). |
| 824 | for _, r := range value { |
| 825 | if unicode.Is(unicode.Sc, r) { |
| 826 | return value |
| 827 | } |
| 828 | } |
| 829 | if isThreeLetterCurrencyCode(value) { |
| 830 | return strings.ToUpper(value) + " " |
| 831 | } |
| 832 | return "¥" |
| 833 | } |
| 834 | |
| 835 | func isThreeLetterCurrencyCode(value string) bool { |
| 836 | if len(value) != 3 { |
| 837 | return false |
| 838 | } |
| 839 | for _, r := range value { |
| 840 | if (r < 'a' || r > 'z') && (r < 'A' || r > 'Z') { |
| 841 | return false |
| 842 | } |
| 843 | } |
| 844 | return true |
| 845 | } |
| 846 | |
| 847 | // Chunk is a single streamed event. Read the field matching Type. |
| 848 | type Chunk struct { |
| 849 | ThinkingBlock *ThinkingBlock |
| 850 | ReasoningState ReasoningState |
| 851 | Type ChunkType |
| 852 | Text string // ChunkText, ChunkReasoning |
| 853 | Signature string // ChunkReasoning: opaque proof for the reasoning (Anthropic thinking signature), when issued |
| 854 | // ReasoningID/ReasoningStatus ride the final ChunkReasoning of a turn |
| 855 | // (empty Text): the provider-issued reasoning item id/status capturedfrom the SSE stream, so the |
| 856 | // Agentcanpersist them into the sessionand the next turn's input reasoning item round-trips them (review |
| 857 | // #7234 — OpenAI Responses schema marks Reasoning.id required). |
| 858 | ReasoningID string // ChunkReasoning: provider-issued reasoning item id |
| 859 | ReasoningStatus string // ChunkReasoning: final reasoning item status ("completed") |
| 860 | ToolCall *ToolCall // ChunkToolCallStart (ID+Name only), ChunkToolCallArgsDelta (ID+Name), ChunkToolCall (complete) |
| 861 | ArgChars int // ChunkToolCallArgsDelta: cumulative argument characters received for this call |
| 862 | ResponsesItem json.RawMessage // ChunkResponsesItem: opaque validated Responses API output item |
| 863 | ServerSearch *ServerSearchCall // ChunkServerSearch: display card + replay payload |
| 864 | Usage *Usage // ChunkUsage |
| 865 | Err error // ChunkError |
| 866 | } |
| 867 | |
| 868 | // Fixed stream-interrupt reasons for observability. Values are a closed enumand must never carry URLs, |
| 869 | // toolarguments, file paths, or raw error text. |
| 870 | const ( |
| 871 | StreamInterruptConnectionReset = "connection_reset" |
| 872 | StreamInterruptPrematureEOF = "premature_eof" |
| 873 | StreamInterruptIdleTimeout = "idle_timeout" |
| 874 | ) |
| 875 | |
| 876 | // StreamInterruptedErrormarksthatthecurrentsamplingattemptneverreachedacleanproviderterminaleventandisthereforeuncommitted. |
| 877 | // The |
| 878 | // Agentmayreplay the exact same provider request. Providers must notperformbody-phaserequestreplaythemselves |
| 879 | // — |
| 880 | // that lives at the Agent layer so retry budgets, |
| 881 | // UI rollback, and tool execution stay single-owner. context.Canceled, auth, |
| 882 | // 4xx/schema errors, and unparseable complete protocol payloads must not usethis type. |
| 883 | type StreamInterruptedError struct { |
| 884 | Err error |
| 885 | Reason string // one of the StreamInterrupt* constants; may be empty for older callers |
| 886 | } |
| 887 | |
| 888 | func (e *StreamInterruptedError) Error() string { |
| 889 | if e == nil || e.Err == nil { |
| 890 | return "stream interrupted" |
| 891 | } |
| 892 | return e.Err.Error() |
| 893 | } |
| 894 | |
| 895 | func (e *StreamInterruptedError) Unwrap() error { |
| 896 | if e == nil { |
| 897 | return nil |
| 898 | } |
| 899 | return e.Err |
| 900 | } |
| 901 | |
| 902 | // StreamInterrupt wraps err as a StreamInterruptedError with a fixed reason. |
| 903 | func StreamInterrupt(err error, reason string) error { |
| 904 | if err == nil { |
| 905 | return nil |
| 906 | } |
| 907 | return &StreamInterruptedError{Err: err, Reason: reason} |
| 908 | } |
| 909 | |
| 910 | // StreamInterruptReason returns the fixed reason when err is a streaminterruption, or empty otherwise. |
| 911 | func StreamInterruptReason(err error) string { |
| 912 | var interrupted *StreamInterruptedError |
| 913 | if !errors.As(err, &interrupted) || interrupted == nil { |
| 914 | return "" |
| 915 | } |
| 916 | if interrupted.Reason != "" { |
| 917 | return interrupted.Reason |
| 918 | } |
| 919 | return ClassifyStreamInterrupt(interrupted.Err) |
| 920 | } |
| 921 | |
| 922 | // ClassifyStreamInterrupt maps a transport error onto a fixed reason enum. |
| 923 | // Prefer attaching Reason at the emit site; this is a best-effort fallback. |
| 924 | func ClassifyStreamInterrupt(err error) string { |
| 925 | if err == nil { |
| 926 | return StreamInterruptPrematureEOF |
| 927 | } |
| 928 | msg := strings.ToLower(err.Error()) |
| 929 | switch { |
| 930 | case strings.Contains(msg, "stalled") || strings.Contains(msg, "idle timeout") || strings.Contains(msg, "no data for"): |
| 931 | return StreamInterruptIdleTimeout |
| 932 | case errors.Is(err, io.ErrUnexpectedEOF) || errors.Is(err, io.EOF) || strings.Contains(msg, "before completion") || strings.Contains(msg, "unexpected eof"): |
| 933 | return StreamInterruptPrematureEOF |
| 934 | case errors.Is(err, net.ErrClosed) || errors.Is(err, syscall.ECONNRESET) || errors.Is(err, syscall.ECONNABORTED) || |
| 935 | strings.Contains(msg, "connection reset") || strings.Contains(msg, "forcibly closed") || strings.Contains(msg, "broken pipe"): |
| 936 | return StreamInterruptConnectionReset |
| 937 | default: |
| 938 | if IsConnReset(err) { |
| 939 | return StreamInterruptConnectionReset |
| 940 | } |
| 941 | return StreamInterruptPrematureEOF |
| 942 | } |
| 943 | } |
| 944 | |
| 945 | func IsStreamInterrupted(err error) bool { |
| 946 | var interrupted *StreamInterruptedError |
| 947 | return errors.As(err, &interrupted) |
| 948 | } |
| 949 | |
| 950 | // Provider is a chat-capable model backend. |
| 951 | type Provider interface { |
| 952 | // Name returns the provider instance name, e.g. "deepseek" / "mimo". |
| 953 | Name() string |
| 954 | // Stream starts a streaming completion, pushing increments on the channel. |
| 955 | // Cancelling ctx must abort the underlying request; a closed channel marksthe end of the completion. |
| 956 | Stream(ctx context.Context, req Request) (<-chan Chunk, error) |
| 957 | } |
| 958 | |
| 959 | // ToolCallReasoningPolicyisoptionallyimplementedbyproviderswhoseprotocolreplaystheprovider-issuedreasoningblockonassistanttool_callsturns |
| 960 | // (DeepSeek thinking mode). The agent uses it to archive the original reasoningtext on those turns |
| 961 | // (adisplay-translated copy must not round-trip to the |
| 962 | // API) and to detect turns that arrive with none. |
| 963 | // Whetheranexplicitemptyvalueisavalidfinalfallbackisaseparate protocol capability. |
| 964 | // Mostprovidersleavethisunset; callers must treat it as false. |
| 965 | type ToolCallReasoningPolicy interface { |
| 966 | RequiresToolCallReasoning() bool |
| 967 | } |
| 968 | |
| 969 | // RequiresToolCallReasoningreportswhetherpreplaysreasoning_contentonassistanttool_callsturnssentbackinhistory. |
| 970 | func RequiresToolCallReasoning(p Provider) bool { |
| 971 | if nilutil.IsNil(p) { |
| 972 | return false |
| 973 | } |
| 974 | policy, ok := p.(ToolCallReasoningPolicy) |
| 975 | return ok && policy.RequiresToolCallReasoning() |
| 976 | } |
| 977 | |
| 978 | // ReasoningRoundTripPolicyisoptionallyimplementedbyprovidersthatrequireeveryassistantmessagetopreserveprovider-issuedreasoninginlaterrequests. |
| 979 | // This is broader than ToolCallReasoningPolicy, which covers onlyassistant tool_calls turns. |
| 980 | type ReasoningRoundTripPolicy interface { |
| 981 | RequiresReasoningRoundTrip() bool |
| 982 | } |
| 983 | |
| 984 | // RequiresReasoningRoundTripreportswhetherrawproviderreasoningmustberetainedandreplayedonallassistantmessages. |
| 985 | func RequiresReasoningRoundTrip(p Provider) bool { |
| 986 | if nilutil.IsNil(p) { |
| 987 | return false |
| 988 | } |
| 989 | policy, ok := p.(ReasoningRoundTripPolicy) |
| 990 | return ok && policy.RequiresReasoningRoundTrip() |
| 991 | } |
| 992 | |
| 993 | // MissingToolCallReasoningWarningPolicyisoptionallyimplementedbyproviderswhosereplayprotocolrequiresreasoning_content, |
| 994 | // but whose active model maynot reliably emit it. The legacy Warning name is retainedforsourcecompatibility; |
| 995 | // theagentnowusesthispolicy for silent bounded recovery andemits no user-visible protocol notice. |
| 996 | type MissingToolCallReasoningWarningPolicy interface { |
| 997 | WarnOnMissingToolCallReasoning() bool |
| 998 | } |
| 999 | |
| 1000 | // MissingToolCallReasoningWarningIdentityPolicy optionally supplies the stable, |
| 1001 | // non-credential configuration identity used to rate-limit missing-reasoningrecovery attempts. |
| 1002 | // Thelegacynamepreserves adapters and persisted state. |
| 1003 | // Implementations may include adapter kind, endpoint, model, and thinkingcontrols; |
| 1004 | // therawidentityneverleavesmemory and is hashed beforepersistence. |
| 1005 | type MissingToolCallReasoningWarningIdentityPolicy interface { |
| 1006 | MissingToolCallReasoningWarningIdentity() string |
| 1007 | } |
| 1008 | |
| 1009 | // WarnOnMissingToolCallReasoningreportswhetheratool_callsturnwithemptyreasoning_contentshouldentersilentrecovery. |
| 1010 | // Its legacy name ispreservedfor provider implementations compiled against the original diagnostic API. |
| 1011 | func WarnOnMissingToolCallReasoning(p Provider) bool { |
| 1012 | if nilutil.IsNil(p) { |
| 1013 | return false |
| 1014 | } |
| 1015 | policy, ok := p.(MissingToolCallReasoningWarningPolicy) |
| 1016 | if ok { |
| 1017 | return policy.WarnOnMissingToolCallReasoning() |
| 1018 | } |
| 1019 | return RequiresToolCallReasoning(p) |
| 1020 | } |
| 1021 | |
| 1022 | // MissingToolCallReasoningWarningFingerprintreturnsanopaquestablekeyforoneproviderconfiguration'srecoverycooldown. |
| 1023 | // Concrete adapters distinguishendpoint/model/protocol changes; |
| 1024 | // providerswithouttheoptionalpolicyretainasafetype-and-namefallback. |
| 1025 | // The legacy name preserves the on-disk statecontract. |
| 1026 | // Thedigestpreventslocalstatefromexposingrawendpointsormodel identifiers. |
| 1027 | func MissingToolCallReasoningWarningFingerprint(p Provider) string { |
| 1028 | if nilutil.IsNil(p) { |
| 1029 | return "" |
| 1030 | } |
| 1031 | identity := fmt.Sprintf("%T\x00%s", p, strings.TrimSpace(p.Name())) |
| 1032 | if policy, ok := p.(MissingToolCallReasoningWarningIdentityPolicy); ok { |
| 1033 | if configured := strings.TrimSpace(policy.MissingToolCallReasoningWarningIdentity()); configured != "" { |
| 1034 | identity = configured |
| 1035 | } |
| 1036 | } |
| 1037 | digest := sha256.Sum256([]byte(identity)) |
| 1038 | return hex.EncodeToString(digest[:]) |
| 1039 | } |
| 1040 | |
| 1041 | // Config is a resolved provider instance configuration. |
| 1042 | type Config struct { |
| 1043 | // HTTPClient supplies immutable credential-proxy transport without changing serialization or vendor identity. |
| 1044 | HTTPClient *http.Client |
| 1045 | Name string // stable instance id, e.g. "deepseek-anthropic" |
| 1046 | DisplayName string // user-editable label; empty falls back to Name |
| 1047 | Protocol string // configured wire adapter id |
| 1048 | BaseURL string // OpenAI-compatible endpoint |
| 1049 | Model string // model id |
| 1050 | APIKey string // resolved from api_key_env |
| 1051 | Extra map[string]any // kind-specific options |
| 1052 | // ModelInfo is adapter-owned metadata for the exact model instance. It is |
| 1053 | // optional so existing third-party factories remain source-compatible. |
| 1054 | ModelInfo *ModelInfo |
| 1055 | } |
| 1056 | |
| 1057 | // Factory builds a Provider from a resolved Config. |
| 1058 | type Factory func(cfg Config) (Provider, error) |
| 1059 | |
| 1060 | var registry = map[string]Factory{} |
| 1061 | |
| 1062 | // Register adds a factory under a kind (e.g. "openai"). Intended for init(). |
| 1063 | // It panics on a duplicate kind, since that is a compile-time wiring mistake. |
| 1064 | func Register(kind string, f Factory) { |
| 1065 | if _, dup := registry[kind]; dup { |
| 1066 | panic("provider: duplicate kind " + kind) |
| 1067 | } |
| 1068 | registry[kind] = f |
| 1069 | } |
| 1070 | |
| 1071 | // New instantiates the provider of the given kind. |
| 1072 | func New(kind string, cfg Config) (Provider, error) { |
| 1073 | f, ok := registry[kind] |
| 1074 | if !ok { |
| 1075 | return nil, fmt.Errorf("provider: unknown kind %q (registered: %v)", kind, Kinds()) |
| 1076 | } |
| 1077 | p, err := f(cfg) |
| 1078 | if err != nil { |
| 1079 | return nil, err |
| 1080 | } |
| 1081 | if nilutil.IsNil(p) { |
| 1082 | return nil, fmt.Errorf("provider: factory %q returned nil provider", kind) |
| 1083 | } |
| 1084 | return p, nil |
| 1085 | } |
| 1086 | |
| 1087 | // Kinds returns the registered kinds, sorted. |
| 1088 | func Kinds() []string { |
| 1089 | out := make([]string, 0, len(registry)) |
| 1090 | for k := range registry { |
| 1091 | out = append(out, k) |
| 1092 | } |
| 1093 | sort.Strings(out) |
| 1094 | return out |
| 1095 | } |
| 1096 |