| 1 | // Package provider defines the model-backend abstraction and a registry mapping |
| 2 | // a provider "kind" to a factory. Concrete implementations live in subpackages |
| 3 | // (e.g. provider/openai) and self-register via init(). The core resolves |
| 4 | // providers by kind from config and never hardcodes 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 | "sort" |
| 17 | "strings" |
| 18 | "syscall" |
| 19 | "unicode" |
| 20 | |
| 21 | "reasonix/internal/nilutil" |
| 22 | ) |
| 23 | |
| 24 | // Role is the role of a message. |
| 25 | type Role string |
| 26 | |
| 27 | const ( |
| 28 | RoleSystem Role = "system" |
| 29 | RoleUser Role = "user" |
| 30 | RoleAssistant Role = "assistant" |
| 31 | RoleTool Role = "tool" |
| 32 | ) |
| 33 | |
| 34 | // LocalOnlyToolName/ID make display-only records safe when a newer transcript |
| 35 | // is opened by an older Reasonix binary that does not know Message.LocalOnly. |
| 36 | // Old wire normalization treats this unmatched tool result as an orphan and |
| 37 | // drops it instead of replaying partial content to the model. |
| 38 | const ( |
| 39 | LocalOnlyToolName = "__reasonix_local_only__" |
| 40 | LocalOnlyToolID = "__reasonix_local_only__" |
| 41 | ) |
| 42 | |
| 43 | // Message is a single conversation message. |
| 44 | type Message struct { |
| 45 | Role Role `json:"role"` |
| 46 | // Content is the provider-visible conversation content. Keeping this legacy |
| 47 | // field provider-visible preserves replay for older CLI/Desktop releases. |
| 48 | Content string `json:"content,omitempty"` |
| 49 | // RawContent is the user-authored form of a user turn, when it differs from |
| 50 | // Content because the host added transient context. Older releases ignore |
| 51 | // this field and still replay the provider-visible Content safely. |
| 52 | RawContent string `json:"raw_content,omitempty"` |
| 53 | // ProviderContent is a transitional field written by early Context Engine v2 |
| 54 | // builds. Loaders migrate it into Content/RawContent before normal use. |
| 55 | ProviderContent string `json:"provider_content,omitempty"` |
| 56 | Images []string `json:"images,omitempty"` // data URLs (data:<mime>;base64,…) on user (attachments) and tool (MCP image results) messages; embedded only for vision-capable models |
| 57 | ReasoningContent string `json:"reasoning_content,omitempty"` // assistant: thinking-mode chain-of-thought, round-tripped on multi-turn |
| 58 | // ReasoningID is the provider-issued identifier of the reasoning item |
| 59 | // (OpenAI Responses schema: Reasoning.id is required on input items). |
| 60 | // Captured from the streamed output item and round-tripped back into |
| 61 | // the input on subsequent turns, matching the wire schema. |
| 62 | ReasoningID string `json:"reasoning_id,omitempty"` |
| 63 | // ReasoningStatus is the final status of the reasoning item |
| 64 | // ("in_progress" | "completed") as issued by the server's done event, |
| 65 | // round-tripped back into the input alongside ReasoningID. |
| 66 | ReasoningStatus string `json:"reasoning_status,omitempty"` |
| 67 | // ReasoningSignature is an opaque, provider-issued proof that ReasoningContent |
| 68 | // is genuine model output. Anthropic requires the signed thinking block be |
| 69 | // replayed on the next turn when a tool call followed thinking; providers |
| 70 | // without signed reasoning (e.g. the openai-compatible ones) leave it empty. |
| 71 | // Round-tripped alongside ReasoningContent. |
| 72 | ReasoningSignature string `json:"reasoning_signature,omitempty"` |
| 73 | ToolCalls []ToolCall `json:"tool_calls,omitempty"` // set by assistant |
| 74 | // ResponsesItems preserves provider-issued Responses API output items that |
| 75 | // must be replayed on a stateless follow-up. Today only DeepSeek |
| 76 | // web_search_call items use this path; other providers ignore the field. |
| 77 | // Keeping the opaque JSON on the assistant turn makes resume/restart safe, |
| 78 | // while omitempty keeps old session files byte-compatible when unused. |
| 79 | ResponsesItems []json.RawMessage `json:"responses_items,omitempty"` |
| 80 | ToolCallID string `json:"tool_call_id,omitempty"` // links a tool result to its call |
| 81 | Name string `json:"name,omitempty"` // tool message: tool name |
| 82 | MemoryCitations []MemoryCitation `json:"memoryCitations,omitempty"` // local UI metadata; provider requests ignore it |
| 83 | WorkDurationMs int64 `json:"workDurationMs,omitempty"` // local UI metadata; provider requests ignore it |
| 84 | CreatedAt int64 `json:"createdAt,omitempty"` // local UI metadata; unix milliseconds; stripped before provider requests |
| 85 | Edited bool `json:"edited,omitempty"` // local UI metadata; provider requests ignore it |
| 86 | Original string `json:"original,omitempty"` // user prompt before inline edit |
| 87 | // LocalOnly marks durable transcript content that must never be sent to a |
| 88 | // model provider. Interrupted streaming output uses it so every frontend can |
| 89 | // replay what the user saw without feeding partial reasoning or tool-call |
| 90 | // arguments back into the next request. |
| 91 | LocalOnly bool `json:"local_only,omitempty"` |
| 92 | DecisionReceipt *DecisionReceipt `json:"decision_receipt,omitempty"` |
| 93 | // DecisionReceipts are local-only metadata attached to a provider-visible |
| 94 | // message. Keeping them on the existing assistant record preserves the |
| 95 | // assistant/tool-result adjacency required by current and older readers. |
| 96 | // ModelMessages strips the field before handing requests to providers. |
| 97 | DecisionReceipts []*DecisionReceipt `json:"decision_receipts,omitempty"` |
| 98 | InterruptedTurn *InterruptedTurnRecovery `json:"interrupted_turn,omitempty"` |
| 99 | // ToolExecution is local shell UI metadata on tool-result messages. It is |
| 100 | // persisted for Desktop/CLI/Serve cards and stripped by ModelMessages before |
| 101 | // any provider request so tool schemas and prompt-cache prefixes stay stable. |
| 102 | ToolExecution *ToolExecution `json:"tool_execution,omitempty"` |
| 103 | } |
| 104 | |
| 105 | // ToolExecution is host-local shell metadata mirrored from tool.ShellExecution. |
| 106 | // Provider serializers must never emit this object on the wire. |
| 107 | type ToolExecution struct { |
| 108 | Kind string `json:"kind,omitempty"` |
| 109 | Shell string `json:"shell,omitempty"` |
| 110 | ShellVersion string `json:"shellVersion,omitempty"` |
| 111 | Platform string `json:"platform,omitempty"` |
| 112 | SupportsAndAnd bool `json:"supportsAndAnd"` |
| 113 | State string `json:"state,omitempty"` |
| 114 | FailurePhase string `json:"failurePhase,omitempty"` |
| 115 | ExitCode *int `json:"exitCode,omitempty"` |
| 116 | OutputTail string `json:"outputTail,omitempty"` |
| 117 | MutationRisk string `json:"mutationRisk,omitempty"` |
| 118 | Verification string `json:"verification,omitempty"` |
| 119 | DurationMs int64 `json:"durationMs,omitempty"` |
| 120 | } |
| 121 | |
| 122 | // DecisionReceipt is durable, provider-excluded evidence of a user-owned |
| 123 | // approval decision. It intentionally contains only bounded labels and the |
| 124 | // outcome, never free-form guidance or provider-visible content. |
| 125 | type DecisionReceipt struct { |
| 126 | ID string `json:"id"` |
| 127 | Kind string `json:"kind"` |
| 128 | Tool string `json:"tool,omitempty"` |
| 129 | Subject string `json:"subject,omitempty"` |
| 130 | Outcome string `json:"outcome"` |
| 131 | } |
| 132 | |
| 133 | // InterruptedTurnRecovery is the durable, provider-excluded handoff for a turn |
| 134 | // that stopped before producing a clean final answer. It contains only bounded |
| 135 | // structural facts; raw partial reasoning remains on the LocalOnly Message for |
| 136 | // display and is never copied into the recovery prompt. |
| 137 | type InterruptedTurnRecovery struct { |
| 138 | Pending bool `json:"pending,omitempty"` |
| 139 | CompletedTools []InterruptedToolSummary `json:"completed_tools,omitempty"` |
| 140 | InterruptedTools []string `json:"interrupted_tools,omitempty"` |
| 141 | DroppedPartialText bool `json:"dropped_partial_text,omitempty"` |
| 142 | DroppedPartialReasoning bool `json:"dropped_partial_reasoning,omitempty"` |
| 143 | } |
| 144 | |
| 145 | // InterruptedToolSummary records a completed, fully paired tool call without |
| 146 | // duplicating its arguments or result. The canonical assistant/tool messages |
| 147 | // immediately before the recovery record remain the source of truth. |
| 148 | type InterruptedToolSummary struct { |
| 149 | ID string `json:"id,omitempty"` |
| 150 | Name string `json:"name"` |
| 151 | Files []string `json:"files,omitempty"` |
| 152 | Added int `json:"added,omitempty"` |
| 153 | Removed int `json:"removed,omitempty"` |
| 154 | } |
| 155 | |
| 156 | // MemoryCitation is local display metadata for memories that influenced an |
| 157 | // assistant turn. Provider implementations must not forward it to model APIs. |
| 158 | type MemoryCitation struct { |
| 159 | ID string `json:"id,omitempty"` |
| 160 | Source string `json:"source"` |
| 161 | LineStart int `json:"lineStart,omitempty"` |
| 162 | LineEnd int `json:"lineEnd,omitempty"` |
| 163 | Note string `json:"note,omitempty"` |
| 164 | Kind string `json:"kind,omitempty"` |
| 165 | } |
| 166 | |
| 167 | // ParseImageDataURL splits a `data:<media-type>;base64,<payload>` URL into its |
| 168 | // media type and base64 payload. ok is false for anything that isn't a base64 |
| 169 | // data URL — providers that need the split (Anthropic) skip those silently. |
| 170 | func ParseImageDataURL(dataURL string) (mediaType, base64Data string, ok bool) { |
| 171 | rest, found := strings.CutPrefix(dataURL, "data:") |
| 172 | if !found { |
| 173 | return "", "", false |
| 174 | } |
| 175 | meta, payload, found := strings.Cut(rest, ",") |
| 176 | if !found { |
| 177 | return "", "", false |
| 178 | } |
| 179 | mt, found := strings.CutSuffix(meta, ";base64") |
| 180 | if !found || mt == "" { |
| 181 | return "", "", false |
| 182 | } |
| 183 | return mt, payload, true |
| 184 | } |
| 185 | |
| 186 | // ToolCall is a tool invocation requested by the model. Arguments is raw JSON. |
| 187 | type ToolCall struct { |
| 188 | ID string `json:"id"` |
| 189 | Name string `json:"name"` |
| 190 | Arguments string `json:"arguments"` |
| 191 | // ThoughtSignature is an opaque Gemini-issued proof attached to a function |
| 192 | // call. OpenAI-compatible Gemini endpoints require it on message replay. |
| 193 | ThoughtSignature string `json:"thought_signature,omitempty"` |
| 194 | Diff string `json:"diff,omitempty"` |
| 195 | Added int `json:"added,omitempty"` |
| 196 | Removed int `json:"removed,omitempty"` |
| 197 | // Resolved* fields are Reasonix-local display metadata for stable proxy |
| 198 | // calls such as use_capability. Provider request builders deliberately |
| 199 | // serialize only provider-visible fields, so these values never alter the |
| 200 | // provider-visible conversation or prompt-cache prefix. |
| 201 | ResolvedName string `json:"resolved_name,omitempty"` |
| 202 | CapabilityID string `json:"capability_id,omitempty"` |
| 203 | ResolvedReadOnly *bool `json:"resolved_read_only,omitempty"` |
| 204 | } |
| 205 | |
| 206 | // ToolSchema is a tool definition exposed to the model. Parameters is JSON Schema. |
| 207 | type ToolSchema struct { |
| 208 | Name string `json:"name"` |
| 209 | Description string `json:"description"` |
| 210 | Parameters json.RawMessage `json:"parameters"` |
| 211 | } |
| 212 | |
| 213 | // Request is a single completion request. |
| 214 | type Request struct { |
| 215 | Messages []Message |
| 216 | Tools []ToolSchema |
| 217 | Temperature *float64 // nil = omit; non-nil = send the value, including 0 |
| 218 | MaxTokens int |
| 219 | // ResponseFormat, when non-nil, asks the endpoint for structured JSON |
| 220 | // output (Responses: text.format.type=json_object). Nil omits the field |
| 221 | // entirely — the common path must stay byte-stable for prompt caching. |
| 222 | ResponseFormat *ResponseFormat `json:"ResponseFormat,omitempty"` |
| 223 | } |
| 224 | |
| 225 | // ResponseFormat asks a provider to constrain its output shape. |
| 226 | type ResponseFormat struct { |
| 227 | // Type is the structured format: "json_object" is the only shape the |
| 228 | // Responses endpoints currently define (MiMo/DashScope/OpenAI). |
| 229 | Type string `json:"type"` |
| 230 | } |
| 231 | |
| 232 | // DefaultReasoningOutputTokens is the conservative provider-side budget used |
| 233 | // for official reasoning APIs whose documented contract safely accepts 32K. |
| 234 | // Unknown compatible gateways must opt in through configuration instead of |
| 235 | // inheriting this value merely because they implement an OpenAI-shaped wire. |
| 236 | const DefaultReasoningOutputTokens = 32 * 1024 |
| 237 | |
| 238 | // TemperaturePtr wraps v in a pointer so callers that explicitly want a |
| 239 | // specific temperature, including 0 for deterministic output, can distinguish |
| 240 | // that intent from "not set, use the provider default". |
| 241 | func TemperaturePtr(v float64) *float64 { return &v } |
| 242 | |
| 243 | // OptionalTemperature returns nil when v is zero, matching the historical |
| 244 | // config behavior where 0 meant "not configured", and a pointer otherwise. |
| 245 | func OptionalTemperature(v float64) *float64 { |
| 246 | if v == 0 { |
| 247 | return nil |
| 248 | } |
| 249 | return &v |
| 250 | } |
| 251 | |
| 252 | // interruptedToolResult stands in for a tool result that never landed — an |
| 253 | // assistant tool_calls turn whose execution was cut short (interrupt, crash) and |
| 254 | // later resumed. Sending such a turn unanswered trips the OpenAI/DeepSeek 400 |
| 255 | // "An assistant message with 'tool_calls' must be followed by tool messages |
| 256 | // responding to each 'tool_call_id'". |
| 257 | const interruptedToolResult = "[no result: the previous turn was interrupted before this tool call completed]" |
| 258 | |
| 259 | // SanitizeToolPairing is the provider-side alias for NormalizeMessages. It repairs |
| 260 | // a history so it satisfies the tool-call contract the OpenAI-compatible and |
| 261 | // Anthropic APIs enforce (every assistant tool_calls answered, no orphan tool |
| 262 | // messages, truncated args closed) right before sending it to the wire — without |
| 263 | // touching the stored session. Kept as a distinct name so call sites read as |
| 264 | // "defensive wire prep" rather than "session mutation". |
| 265 | func SanitizeToolPairing(msgs []Message) []Message { return NormalizeMessages(msgs) } |
| 266 | |
| 267 | // ModelMessages removes durable display-only records before a request is |
| 268 | // handed to any provider. Healthy sessions without such records keep their |
| 269 | // original backing slice, preserving the allocation and prompt-cache fast path. |
| 270 | func ModelMessages(msgs []Message) []Message { |
| 271 | needsCopy := false |
| 272 | for _, m := range msgs { |
| 273 | if m.LocalOnly || m.RawContent != "" || m.ProviderContent != "" || m.DecisionReceipt != nil || len(m.DecisionReceipts) > 0 || m.ToolExecution != nil { |
| 274 | needsCopy = true |
| 275 | break |
| 276 | } |
| 277 | } |
| 278 | if !needsCopy { |
| 279 | return msgs |
| 280 | } |
| 281 | out := make([]Message, 0, len(msgs)) |
| 282 | for _, candidate := range msgs { |
| 283 | if candidate.LocalOnly { |
| 284 | continue |
| 285 | } |
| 286 | if candidate.ProviderContent != "" { |
| 287 | candidate.Content = candidate.ProviderContent |
| 288 | candidate.ProviderContent = "" |
| 289 | } |
| 290 | candidate.RawContent = "" |
| 291 | candidate.DecisionReceipt = nil |
| 292 | candidate.DecisionReceipts = nil |
| 293 | // Local shell metadata must never enter provider request bytes. |
| 294 | candidate.ToolExecution = nil |
| 295 | out = append(out, candidate) |
| 296 | } |
| 297 | return out |
| 298 | } |
| 299 | |
| 300 | // NormalizeMessages repairs a conversation history so it satisfies the tool-call |
| 301 | // contract the OpenAI-compatible and Anthropic APIs enforce: every assistant |
| 302 | // tool_calls entry must be answered by a following tool message for its id, and a |
| 303 | // tool message must follow such a call. It backfills a placeholder result for any |
| 304 | // unanswered call (so the turn stays intact), drops orphan tool messages, |
| 305 | // backfills empty tool-call names from their results (#4727 — old sessions saved |
| 306 | // before adde2d3e can carry an empty name), 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 loads |
| 310 | // use NormalizeSessionMessages so they can share the assistant-turn repairs |
| 311 | // without deleting standalone tool messages that must round-trip through |
| 312 | // reasonix --resume. |
| 313 | // |
| 314 | // A well-formed history — no unanswered calls, no orphan results, no empty tool- |
| 315 | // call names, no truncated args — returns the input slice unchanged (same backing |
| 316 | // array, zero allocation). This keeps the prefix-cache key stable for healthy |
| 317 | // sessions and makes repeated normalization cheap. |
| 318 | func NormalizeMessages(msgs []Message) []Message { |
| 319 | return normalizeMessages(msgs, true) |
| 320 | } |
| 321 | |
| 322 | // NormalizeSessionMessages applies only repairs that are safe to persist in a |
| 323 | // saved session. It shares assistant-turn repairs with NormalizeMessages, but |
| 324 | // preserves existing tool messages instead of dropping or reordering them so |
| 325 | // Save/LoadSession remains a byte-for-byte conversation round trip for histories |
| 326 | // that were already on disk. |
| 327 | func NormalizeSessionMessages(msgs []Message) []Message { |
| 328 | return normalizeMessages(attachStandaloneDecisionReceipts(msgs), false) |
| 329 | } |
| 330 | |
| 331 | // attachStandaloneDecisionReceipts migrates the short-lived receipt encoding |
| 332 | // that stored a LocalOnly assistant message between an assistant tool call and |
| 333 | // its result. Folding that metadata into the latest assistant message repairs |
| 334 | // already-written sessions before tool-pair normalization can fabricate a |
| 335 | // placeholder. Healthy histories return the original slice unchanged. |
| 336 | func attachStandaloneDecisionReceipts(msgs []Message) []Message { |
| 337 | target := -1 |
| 338 | needsMigration := false |
| 339 | for i, m := range msgs { |
| 340 | switch { |
| 341 | case m.Role == RoleUser && !m.LocalOnly: |
| 342 | target = -1 |
| 343 | case m.Role == RoleAssistant && !m.LocalOnly: |
| 344 | target = i |
| 345 | case target >= 0 && m.LocalOnly && m.DecisionReceipt != nil: |
| 346 | needsMigration = true |
| 347 | } |
| 348 | if needsMigration { |
| 349 | break |
| 350 | } |
| 351 | } |
| 352 | if !needsMigration { |
| 353 | return msgs |
| 354 | } |
| 355 | |
| 356 | out := make([]Message, 0, len(msgs)) |
| 357 | target = -1 |
| 358 | for _, m := range msgs { |
| 359 | switch { |
| 360 | case m.Role == RoleUser && !m.LocalOnly: |
| 361 | target = -1 |
| 362 | case m.Role == RoleAssistant && !m.LocalOnly: |
| 363 | out = append(out, m) |
| 364 | target = len(out) - 1 |
| 365 | continue |
| 366 | case target >= 0 && m.LocalOnly && m.DecisionReceipt != nil: |
| 367 | receipts := append([]*DecisionReceipt(nil), out[target].DecisionReceipts...) |
| 368 | out[target].DecisionReceipts = append(receipts, m.DecisionReceipt) |
| 369 | continue |
| 370 | } |
| 371 | out = append(out, m) |
| 372 | } |
| 373 | return out |
| 374 | } |
| 375 | |
| 376 | func normalizeMessages(msgs []Message, dropOrphanTools bool) []Message { |
| 377 | if normalized, ok := tryNormalizeFastPath(msgs, dropOrphanTools); ok { |
| 378 | return normalized // well-formed: pass through without allocating |
| 379 | } |
| 380 | out := make([]Message, 0, len(msgs)) |
| 381 | for i := 0; i < len(msgs); { |
| 382 | m := msgs[i] |
| 383 | if m.LocalOnly { |
| 384 | if !dropOrphanTools { |
| 385 | out = append(out, m) |
| 386 | } |
| 387 | i++ |
| 388 | continue |
| 389 | } |
| 390 | if m.Role == RoleAssistant && len(m.ToolCalls) > 0 { |
| 391 | j := i + 1 |
| 392 | for j < len(msgs) && msgs[j].Role == RoleTool && !msgs[j].LocalOnly { |
| 393 | j++ |
| 394 | } |
| 395 | // Backfill empty tool-call names from the corresponding tool |
| 396 | // results so the model sees which tool was invoked (#4727). |
| 397 | // The wire-format fix (openai.go) ensures empty fields are |
| 398 | // never omitted, so this backfill is a UX improvement, not a |
| 399 | // correctness requirement. |
| 400 | calls := backfillToolCallNames(m.ToolCalls, msgs[i+1:j]) |
| 401 | m.ToolCalls = calls |
| 402 | out = append(out, repairToolCallArgs(m)) |
| 403 | if dropOrphanTools { |
| 404 | out = append(out, pairToolResults(calls, msgs[i+1:j])...) |
| 405 | } else { |
| 406 | out = append(out, sessionToolResults(calls, msgs[i+1:j])...) |
| 407 | } |
| 408 | i = j |
| 409 | continue |
| 410 | } |
| 411 | if m.Role == RoleTool { |
| 412 | if !dropOrphanTools { |
| 413 | out = append(out, m) |
| 414 | } |
| 415 | // Orphan tool message: provider sends drop it; session loads preserve it. |
| 416 | i++ |
| 417 | continue |
| 418 | } |
| 419 | out = append(out, m) |
| 420 | i++ |
| 421 | } |
| 422 | return out |
| 423 | } |
| 424 | |
| 425 | // tryNormalizeFastPath reports whether msgs needs no repair and, if so, returns |
| 426 | // it as-is so the caller can skip allocating. Healthy tool-call/tool-result |
| 427 | // turns pass through unchanged; malformed turns take the slow path. |
| 428 | func tryNormalizeFastPath(msgs []Message, dropOrphanTools bool) ([]Message, bool) { |
| 429 | for i := 0; i < len(msgs); { |
| 430 | m := msgs[i] |
| 431 | if m.LocalOnly { |
| 432 | if dropOrphanTools { |
| 433 | return nil, false |
| 434 | } |
| 435 | i++ |
| 436 | continue |
| 437 | } |
| 438 | if m.Role == RoleAssistant && len(m.ToolCalls) > 0 { |
| 439 | j := i + 1 |
| 440 | for j < len(msgs) && msgs[j].Role == RoleTool && !msgs[j].LocalOnly { |
| 441 | j++ |
| 442 | } |
| 443 | if !toolTurnWellFormed(m.ToolCalls, msgs[i+1:j]) || needsToolCallArgRepair(m.ToolCalls) { |
| 444 | return nil, false |
| 445 | } |
| 446 | i = j |
| 447 | continue |
| 448 | } |
| 449 | if m.Role == RoleTool && dropOrphanTools { |
| 450 | return nil, false |
| 451 | } |
| 452 | i++ |
| 453 | } |
| 454 | return msgs, true |
| 455 | } |
| 456 | |
| 457 | func toolTurnWellFormed(calls []ToolCall, results []Message) bool { |
| 458 | if len(calls) != len(results) { |
| 459 | return false |
| 460 | } |
| 461 | for _, tc := range calls { |
| 462 | if tc.Name == "" { |
| 463 | return false |
| 464 | } |
| 465 | } |
| 466 | for k, tc := range calls { |
| 467 | if results[k].ToolCallID != tc.ID { |
| 468 | return false |
| 469 | } |
| 470 | if results[k].Name != tc.Name { |
| 471 | return false |
| 472 | } |
| 473 | } |
| 474 | return true |
| 475 | } |
| 476 | |
| 477 | func needsToolCallArgRepair(calls []ToolCall) bool { |
| 478 | for _, tc := range calls { |
| 479 | if tc.Arguments != "" && !json.Valid([]byte(tc.Arguments)) { |
| 480 | return true |
| 481 | } |
| 482 | } |
| 483 | return false |
| 484 | } |
| 485 | |
| 486 | // repairToolCallArgs returns m with any undecodable tool-call Arguments closed |
| 487 | // into valid JSON (copy-on-write; the caller's history is never mutated). Empty |
| 488 | // arguments pass through — some gateways send "" for no-arg tools. |
| 489 | func repairToolCallArgs(m Message) Message { |
| 490 | broken := false |
| 491 | for _, tc := range m.ToolCalls { |
| 492 | if tc.Arguments != "" && !json.Valid([]byte(tc.Arguments)) { |
| 493 | broken = true |
| 494 | break |
| 495 | } |
| 496 | } |
| 497 | if !broken { |
| 498 | return m |
| 499 | } |
| 500 | calls := make([]ToolCall, len(m.ToolCalls)) |
| 501 | copy(calls, m.ToolCalls) |
| 502 | for i := range calls { |
| 503 | if calls[i].Arguments == "" || json.Valid([]byte(calls[i].Arguments)) { |
| 504 | continue |
| 505 | } |
| 506 | calls[i].Arguments = closeTruncatedJSON(calls[i].Arguments) |
| 507 | } |
| 508 | m.ToolCalls = calls |
| 509 | return m |
| 510 | } |
| 511 | |
| 512 | // closeTruncatedJSON best-effort completes a JSON document cut off mid-stream |
| 513 | // (unterminated string, open braces, dangling comma/colon); anything still |
| 514 | // invalid after closing degrades to "{}". |
| 515 | func closeTruncatedJSON(s string) string { |
| 516 | var stack []byte |
| 517 | inStr, esc := false, false |
| 518 | for i := 0; i < len(s); i++ { |
| 519 | c := s[i] |
| 520 | if inStr { |
| 521 | switch { |
| 522 | case esc: |
| 523 | esc = false |
| 524 | case c == '\\': |
| 525 | esc = true |
| 526 | case c == '"': |
| 527 | inStr = false |
| 528 | } |
| 529 | continue |
| 530 | } |
| 531 | switch c { |
| 532 | case '"': |
| 533 | inStr = true |
| 534 | case '{': |
| 535 | stack = append(stack, '}') |
| 536 | case '[': |
| 537 | stack = append(stack, ']') |
| 538 | case '}', ']': |
| 539 | if len(stack) > 0 { |
| 540 | stack = stack[:len(stack)-1] |
| 541 | } |
| 542 | } |
| 543 | } |
| 544 | out := s |
| 545 | if esc { |
| 546 | out = out[:len(out)-1] |
| 547 | } |
| 548 | if inStr { |
| 549 | out += `"` |
| 550 | } |
| 551 | trimmed := strings.TrimRight(out, " \t\r\n") |
| 552 | switch { |
| 553 | case strings.HasSuffix(trimmed, ","): |
| 554 | out = trimmed[:len(trimmed)-1] |
| 555 | case strings.HasSuffix(trimmed, ":"): |
| 556 | out = trimmed + "null" |
| 557 | } |
| 558 | for i := len(stack) - 1; i >= 0; i-- { |
| 559 | out += string(stack[i]) |
| 560 | } |
| 561 | if !json.Valid([]byte(out)) { |
| 562 | return "{}" |
| 563 | } |
| 564 | return out |
| 565 | } |
| 566 | |
| 567 | // pairToolResults answers each tool_call with its result, backfilling a |
| 568 | // placeholder for any unanswered one. Distinct non-empty ids pair by id (so |
| 569 | // reordered results re-sort to call order); empty or duplicate ids pair by |
| 570 | // position instead — some gateways stream tool calls by index with no id, and a |
| 571 | // map keyed on id would collapse those results into one (call order is preserved |
| 572 | // because the loop appends results in call order). |
| 573 | func pairToolResults(calls []ToolCall, avail []Message) []Message { |
| 574 | out := make([]Message, 0, len(calls)) |
| 575 | if idDistinct(calls) { |
| 576 | byID := make(map[string]Message, len(avail)) |
| 577 | for _, r := range avail { |
| 578 | byID[r.ToolCallID] = r |
| 579 | } |
| 580 | for _, tc := range calls { |
| 581 | if r, ok := byID[tc.ID]; ok { |
| 582 | r.Name = tc.Name |
| 583 | out = append(out, r) |
| 584 | } else { |
| 585 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 586 | } |
| 587 | } |
| 588 | return out |
| 589 | } |
| 590 | for k, tc := range calls { |
| 591 | if k < len(avail) { |
| 592 | r := avail[k] |
| 593 | r.ToolCallID = tc.ID |
| 594 | r.Name = tc.Name |
| 595 | out = append(out, r) |
| 596 | } else { |
| 597 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 598 | } |
| 599 | } |
| 600 | return out |
| 601 | } |
| 602 | |
| 603 | // sessionToolResults preserves every stored tool result and appends placeholders |
| 604 | // only for calls that have no recorded answer. Load-time normalization must not |
| 605 | // drop or reorder user history; provider sends can still use pairToolResults for |
| 606 | // strict wire formatting. |
| 607 | func sessionToolResults(calls []ToolCall, avail []Message) []Message { |
| 608 | out := append([]Message(nil), avail...) |
| 609 | if idDistinct(calls) { |
| 610 | answered := make(map[string]struct{}, len(avail)) |
| 611 | for _, r := range avail { |
| 612 | answered[r.ToolCallID] = struct{}{} |
| 613 | } |
| 614 | for _, tc := range calls { |
| 615 | if _, ok := answered[tc.ID]; !ok { |
| 616 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 617 | } |
| 618 | } |
| 619 | return out |
| 620 | } |
| 621 | for k := len(avail); k < len(calls); k++ { |
| 622 | tc := calls[k] |
| 623 | out = append(out, Message{Role: RoleTool, ToolCallID: tc.ID, Name: tc.Name, Content: interruptedToolResult}) |
| 624 | } |
| 625 | return out |
| 626 | } |
| 627 | |
| 628 | // backfillToolCallNames returns calls with any empty Name filled in from the |
| 629 | // matching tool result (by id, then by position). Old sessions (#4727) may have |
| 630 | // saved assistant tool-calls with an empty name; backfilling gives the model |
| 631 | // useful context during replay. The common case (no empty names) returns the |
| 632 | // input unchanged without allocating. Unpaired calls keep their empty name, |
| 633 | // which the wire-format fix (openai.go) handles gracefully. |
| 634 | func backfillToolCallNames(calls []ToolCall, results []Message) []ToolCall { |
| 635 | missing := false |
| 636 | for _, c := range calls { |
| 637 | if c.Name == "" { |
| 638 | missing = true |
| 639 | break |
| 640 | } |
| 641 | } |
| 642 | if !missing { |
| 643 | return calls |
| 644 | } |
| 645 | out := make([]ToolCall, len(calls)) |
| 646 | copy(out, calls) |
| 647 | if idDistinct(calls) { |
| 648 | byID := make(map[string]string, len(results)) |
| 649 | for _, r := range results { |
| 650 | if r.Name != "" { |
| 651 | byID[r.ToolCallID] = r.Name |
| 652 | } |
| 653 | } |
| 654 | for k := range out { |
| 655 | if out[k].Name == "" { |
| 656 | if n, ok := byID[out[k].ID]; ok { |
| 657 | out[k].Name = n |
| 658 | } |
| 659 | } |
| 660 | } |
| 661 | return out |
| 662 | } |
| 663 | // Fallback: positional pairing (same order as pairToolResults). |
| 664 | for k := range out { |
| 665 | if out[k].Name == "" && k < len(results) { |
| 666 | out[k].Name = results[k].Name |
| 667 | } |
| 668 | } |
| 669 | return out |
| 670 | } |
| 671 | |
| 672 | // idDistinct reports whether every call carries a non-empty id unique within the |
| 673 | // batch — the condition under which id-keyed pairing is safe. |
| 674 | func idDistinct(calls []ToolCall) bool { |
| 675 | seen := make(map[string]struct{}, len(calls)) |
| 676 | for _, tc := range calls { |
| 677 | if tc.ID == "" { |
| 678 | return false |
| 679 | } |
| 680 | if _, dup := seen[tc.ID]; dup { |
| 681 | return false |
| 682 | } |
| 683 | seen[tc.ID] = struct{}{} |
| 684 | } |
| 685 | return true |
| 686 | } |
| 687 | |
| 688 | // ChunkType identifies the kind of a streamed increment. |
| 689 | type ChunkType int |
| 690 | |
| 691 | const ( |
| 692 | ChunkText ChunkType = iota // text delta |
| 693 | ChunkReasoning // thinking-mode reasoning delta (before the visible answer) |
| 694 | ChunkToolCallStart // a tool call has begun (ToolCall: ID+Name; args still streaming) |
| 695 | ChunkToolCallArgsDelta // progress while a call's arguments stream (ToolCall: ID+Name; ArgChars: cumulative) |
| 696 | ChunkToolCall // one complete tool call |
| 697 | ChunkUsage // token usage for the completion |
| 698 | ChunkDone // completion finished normally |
| 699 | ChunkError // an error occurred |
| 700 | ChunkResponsesItem // a complete provider-issued Responses API output item for stateless replay |
| 701 | ) |
| 702 | |
| 703 | // Usage reports token accounting for a completion. Cache hit/miss come from |
| 704 | // either DeepSeek's top-level prompt_cache_{hit,miss}_tokens or the OpenAI/MiMo |
| 705 | // standard prompt_tokens_details.cached_tokens — the openai provider normalises |
| 706 | // both shapes into these fields. ReasoningTokens is the thinking-mode subset of |
| 707 | // CompletionTokens reported by thinking-capable models. FinishReason carries |
| 708 | // the model's last reported choices[0].finish_reason so the agent can surface |
| 709 | // abnormal terminations ("length", "content_filter", "repetition_truncation"). |
| 710 | // Estimated marks counts reconstructed locally because the provider's terminal |
| 711 | // usage record did not arrive; exact provider usage leaves it false. |
| 712 | type Usage struct { |
| 713 | PromptTokens int |
| 714 | CompletionTokens int |
| 715 | TotalTokens int |
| 716 | CacheHitTokens int // prompt tokens served from cache |
| 717 | CacheMissTokens int // prompt tokens not cached, including CacheWriteTokens |
| 718 | CacheWriteTokens int // subset of CacheMissTokens used to create provider cache entries |
| 719 | CacheWriteBilledTokens float64 // cache-write charge expressed in ordinary input-token equivalents |
| 720 | ReasoningTokens int // subset of CompletionTokens spent on chain-of-thought |
| 721 | FinishReason string // "stop", "tool_calls", "length", "content_filter", "repetition_truncation", … |
| 722 | Estimated bool |
| 723 | // RequestCount is the number of provider requests represented by this |
| 724 | // aggregate. Zero means one request for backward compatibility. Recovery |
| 725 | // paths that merge multiple attempts set the exact count. |
| 726 | RequestCount int |
| 727 | // Context* fields describe the latest single-request shape for context |
| 728 | // gauges and rebind telemetry. When zero, consumers fall back to the |
| 729 | // billable Prompt/Completion/… fields. Multi-attempt sampling recovery |
| 730 | // sets PromptTokens (etc.) to the billable aggregate and fills Context* |
| 731 | // from the final attempt only. |
| 732 | ContextPromptTokens int |
| 733 | ContextCompletionTokens int |
| 734 | ContextReasoningTokens int |
| 735 | ContextCacheHitTokens int |
| 736 | ContextCacheMissTokens int |
| 737 | } |
| 738 | |
| 739 | // ContextFillTokens returns the latest-attempt context fill (prompt+completion) |
| 740 | // used by status bars and context panels. Falls back to billable totals when |
| 741 | // no Context* fields were set (single-attempt / legacy usage events). |
| 742 | func (u *Usage) ContextFillTokens() int { |
| 743 | if u == nil { |
| 744 | return 0 |
| 745 | } |
| 746 | if u.ContextPromptTokens > 0 || u.ContextCompletionTokens > 0 { |
| 747 | return u.ContextPromptTokens + u.ContextCompletionTokens |
| 748 | } |
| 749 | return u.PromptTokens + u.CompletionTokens |
| 750 | } |
| 751 | |
| 752 | // ContextPromptForGauge returns the latest-attempt prompt size for context |
| 753 | // displays. Falls back to PromptTokens when ContextPromptTokens is unset. |
| 754 | func (u *Usage) ContextPromptForGauge() int { |
| 755 | if u == nil { |
| 756 | return 0 |
| 757 | } |
| 758 | if u.ContextPromptTokens > 0 { |
| 759 | return u.ContextPromptTokens |
| 760 | } |
| 761 | return u.PromptTokens |
| 762 | } |
| 763 | |
| 764 | // Pricing is a provider's per-1M-token rates, used to estimate spend. Currency |
| 765 | // is a display symbol or ISO-like code (default "¥"). toml tags let config decode it. |
| 766 | type Pricing struct { |
| 767 | CacheHit float64 `toml:"cache_hit"` // per 1M cached prompt tokens |
| 768 | Input float64 `toml:"input"` // per 1M uncached prompt tokens |
| 769 | Output float64 `toml:"output"` // per 1M completion tokens |
| 770 | Currency string `toml:"currency"` |
| 771 | } |
| 772 | |
| 773 | // Cost estimates the spend for a usage record. |
| 774 | func (p *Pricing) Cost(u *Usage) float64 { |
| 775 | if p == nil || u == nil { |
| 776 | return 0 |
| 777 | } |
| 778 | hit := u.CacheHitTokens |
| 779 | miss := u.CacheMissTokens |
| 780 | if hit+miss == 0 && u.PromptTokens > 0 { |
| 781 | miss = u.PromptTokens |
| 782 | } else if miss == 0 && hit > 0 && u.PromptTokens > hit { |
| 783 | miss = u.PromptTokens - hit |
| 784 | } |
| 785 | // CacheMissTokens intentionally remains the raw prompt-token denominator |
| 786 | // used by cache hit-rate displays, so cache writes are included there. For |
| 787 | // cost, split those writes back out and replace them with their provider- |
| 788 | // supplied input-token equivalent (for example Anthropic's 1.25x 5-minute |
| 789 | // writes or 2x 1-hour writes). Older providers leave both fields at zero and |
| 790 | // keep the legacy one-input-rate behavior. A write count without billed |
| 791 | // units also falls back to 1x for backward compatibility. |
| 792 | write := u.CacheWriteTokens |
| 793 | if write < 0 { |
| 794 | write = 0 |
| 795 | } |
| 796 | if write > miss { |
| 797 | write = miss |
| 798 | } |
| 799 | billedWrite := 0.0 |
| 800 | if write > 0 { |
| 801 | billedWrite = u.CacheWriteBilledTokens |
| 802 | if billedWrite <= 0 { |
| 803 | billedWrite = float64(write) |
| 804 | } |
| 805 | } |
| 806 | inputTokenUnits := float64(miss-write) + billedWrite |
| 807 | return (float64(hit)*p.CacheHit + |
| 808 | inputTokenUnits*p.Input + |
| 809 | float64(u.CompletionTokens)*p.Output) / 1e6 |
| 810 | } |
| 811 | |
| 812 | // Symbol returns the currency display symbol, defaulting to "¥". |
| 813 | func (p *Pricing) Symbol() string { |
| 814 | if p == nil || p.Currency == "" { |
| 815 | return "¥" |
| 816 | } |
| 817 | return currencySymbol(p.Currency) |
| 818 | } |
| 819 | |
| 820 | func currencySymbol(currency string) string { |
| 821 | value := strings.TrimSpace(currency) |
| 822 | if value == "" { |
| 823 | return "¥" |
| 824 | } |
| 825 | switch strings.ToLower(value) { |
| 826 | case "cny", "rmb", "yuan", "renminbi", "cnh": |
| 827 | return "¥" |
| 828 | case "usd", "dollar", "dollars", "us dollar", "us dollars", "us$": |
| 829 | return "$" |
| 830 | case "eur", "euro", "euros": |
| 831 | return "€" |
| 832 | case "gbp", "pound", "pounds", "sterling": |
| 833 | return "£" |
| 834 | case "jpy", "yen": |
| 835 | return "¥" |
| 836 | } |
| 837 | switch value { |
| 838 | case "￥", "¥": |
| 839 | return "¥" |
| 840 | case "$", "€", "£": |
| 841 | return value |
| 842 | } |
| 843 | // any embedded currency sign → keep as-is (compact symbols like A$, HK$). |
| 844 | for _, r := range value { |
| 845 | if unicode.Is(unicode.Sc, r) { |
| 846 | return value |
| 847 | } |
| 848 | } |
| 849 | if isThreeLetterCurrencyCode(value) { |
| 850 | return strings.ToUpper(value) + " " |
| 851 | } |
| 852 | return "¥" |
| 853 | } |
| 854 | |
| 855 | func isThreeLetterCurrencyCode(value string) bool { |
| 856 | if len(value) != 3 { |
| 857 | return false |
| 858 | } |
| 859 | for _, r := range value { |
| 860 | if (r < 'a' || r > 'z') && (r < 'A' || r > 'Z') { |
| 861 | return false |
| 862 | } |
| 863 | } |
| 864 | return true |
| 865 | } |
| 866 | |
| 867 | // Chunk is a single streamed event. Read the field matching Type. |
| 868 | type Chunk struct { |
| 869 | Type ChunkType |
| 870 | Text string // ChunkText, ChunkReasoning |
| 871 | Signature string // ChunkReasoning: opaque proof for the reasoning (Anthropic thinking signature), when issued |
| 872 | // ReasoningID/ReasoningStatus ride the final ChunkReasoning of a turn |
| 873 | // (empty Text): the provider-issued reasoning item id/status captured |
| 874 | // from the SSE stream, so the Agent can persist them into the session |
| 875 | // and the next turn's input reasoning item round-trips them (review |
| 876 | // #7234 — OpenAI Responses schema marks Reasoning.id required). |
| 877 | ReasoningID string // ChunkReasoning: provider-issued reasoning item id |
| 878 | ReasoningStatus string // ChunkReasoning: final reasoning item status ("completed") |
| 879 | ToolCall *ToolCall // ChunkToolCallStart (ID+Name only), ChunkToolCallArgsDelta (ID+Name), ChunkToolCall (complete) |
| 880 | ArgChars int // ChunkToolCallArgsDelta: cumulative argument characters received for this call |
| 881 | ResponsesItem json.RawMessage // ChunkResponsesItem: opaque validated Responses API output item |
| 882 | Usage *Usage // ChunkUsage |
| 883 | Err error // ChunkError |
| 884 | } |
| 885 | |
| 886 | // Fixed stream-interrupt reasons for observability. Values are a closed enum |
| 887 | // and must never carry URLs, tool arguments, file paths, or raw error text. |
| 888 | const ( |
| 889 | StreamInterruptConnectionReset = "connection_reset" |
| 890 | StreamInterruptPrematureEOF = "premature_eof" |
| 891 | StreamInterruptIdleTimeout = "idle_timeout" |
| 892 | ) |
| 893 | |
| 894 | // StreamInterruptedError marks that the current sampling attempt never reached |
| 895 | // a clean provider terminal event and is therefore uncommitted. The Agent may |
| 896 | // replay the exact same provider request. Providers must not perform body-phase |
| 897 | // request replay themselves — that lives at the Agent layer so retry budgets, |
| 898 | // UI rollback, and tool execution stay single-owner. context.Canceled, auth, |
| 899 | // 4xx/schema errors, and unparseable complete protocol payloads must not use |
| 900 | // this type. |
| 901 | type StreamInterruptedError struct { |
| 902 | Err error |
| 903 | Reason string // one of the StreamInterrupt* constants; may be empty for older callers |
| 904 | } |
| 905 | |
| 906 | func (e *StreamInterruptedError) Error() string { |
| 907 | if e == nil || e.Err == nil { |
| 908 | return "stream interrupted" |
| 909 | } |
| 910 | return e.Err.Error() |
| 911 | } |
| 912 | |
| 913 | func (e *StreamInterruptedError) Unwrap() error { |
| 914 | if e == nil { |
| 915 | return nil |
| 916 | } |
| 917 | return e.Err |
| 918 | } |
| 919 | |
| 920 | // StreamInterrupt wraps err as a StreamInterruptedError with a fixed reason. |
| 921 | func StreamInterrupt(err error, reason string) error { |
| 922 | if err == nil { |
| 923 | return nil |
| 924 | } |
| 925 | return &StreamInterruptedError{Err: err, Reason: reason} |
| 926 | } |
| 927 | |
| 928 | // StreamInterruptReason returns the fixed reason when err is a stream |
| 929 | // interruption, or empty otherwise. |
| 930 | func StreamInterruptReason(err error) string { |
| 931 | var interrupted *StreamInterruptedError |
| 932 | if !errors.As(err, &interrupted) || interrupted == nil { |
| 933 | return "" |
| 934 | } |
| 935 | if interrupted.Reason != "" { |
| 936 | return interrupted.Reason |
| 937 | } |
| 938 | return ClassifyStreamInterrupt(interrupted.Err) |
| 939 | } |
| 940 | |
| 941 | // ClassifyStreamInterrupt maps a transport error onto a fixed reason enum. |
| 942 | // Prefer attaching Reason at the emit site; this is a best-effort fallback. |
| 943 | func ClassifyStreamInterrupt(err error) string { |
| 944 | if err == nil { |
| 945 | return StreamInterruptPrematureEOF |
| 946 | } |
| 947 | msg := strings.ToLower(err.Error()) |
| 948 | switch { |
| 949 | case strings.Contains(msg, "stalled") || strings.Contains(msg, "idle timeout") || strings.Contains(msg, "no data for"): |
| 950 | return StreamInterruptIdleTimeout |
| 951 | case errors.Is(err, io.ErrUnexpectedEOF) || errors.Is(err, io.EOF) || strings.Contains(msg, "before completion") || strings.Contains(msg, "unexpected eof"): |
| 952 | return StreamInterruptPrematureEOF |
| 953 | case errors.Is(err, net.ErrClosed) || errors.Is(err, syscall.ECONNRESET) || errors.Is(err, syscall.ECONNABORTED) || |
| 954 | strings.Contains(msg, "connection reset") || strings.Contains(msg, "forcibly closed") || strings.Contains(msg, "broken pipe"): |
| 955 | return StreamInterruptConnectionReset |
| 956 | default: |
| 957 | if IsConnReset(err) { |
| 958 | return StreamInterruptConnectionReset |
| 959 | } |
| 960 | return StreamInterruptPrematureEOF |
| 961 | } |
| 962 | } |
| 963 | |
| 964 | func IsStreamInterrupted(err error) bool { |
| 965 | var interrupted *StreamInterruptedError |
| 966 | return errors.As(err, &interrupted) |
| 967 | } |
| 968 | |
| 969 | // Provider is a chat-capable model backend. |
| 970 | type Provider interface { |
| 971 | // Name returns the provider instance name, e.g. "deepseek" / "mimo". |
| 972 | Name() string |
| 973 | // Stream starts a streaming completion, pushing increments on the channel. |
| 974 | // Cancelling ctx must abort the underlying request; a closed channel marks |
| 975 | // the end of the completion. |
| 976 | Stream(ctx context.Context, req Request) (<-chan Chunk, error) |
| 977 | } |
| 978 | |
| 979 | // ToolCallReasoningPolicy is optionally implemented by providers whose protocol |
| 980 | // replays the provider-issued reasoning block on assistant tool_calls turns |
| 981 | // (DeepSeek thinking mode). The agent uses it to archive the original reasoning |
| 982 | // text on those turns (a display-translated copy must not round-trip to the |
| 983 | // API) and to warn when a turn arrives with none — the request still succeeds |
| 984 | // because the wire layer always emits the reasoning_content key for such turns, |
| 985 | // but the model loses its chain-of-thought context. Most providers leave this |
| 986 | // unset; callers must treat it as false. |
| 987 | type ToolCallReasoningPolicy interface { |
| 988 | RequiresToolCallReasoning() bool |
| 989 | } |
| 990 | |
| 991 | // RequiresToolCallReasoning reports whether p replays reasoning_content on |
| 992 | // assistant tool_calls turns sent back in history. |
| 993 | func RequiresToolCallReasoning(p Provider) bool { |
| 994 | if nilutil.IsNil(p) { |
| 995 | return false |
| 996 | } |
| 997 | policy, ok := p.(ToolCallReasoningPolicy) |
| 998 | return ok && policy.RequiresToolCallReasoning() |
| 999 | } |
| 1000 | |
| 1001 | // ReasoningRoundTripPolicy is optionally implemented by providers that require |
| 1002 | // every assistant message to preserve provider-issued reasoning in later |
| 1003 | // requests. This is broader than ToolCallReasoningPolicy, which covers only |
| 1004 | // assistant tool_calls turns. |
| 1005 | type ReasoningRoundTripPolicy interface { |
| 1006 | RequiresReasoningRoundTrip() bool |
| 1007 | } |
| 1008 | |
| 1009 | // RequiresReasoningRoundTrip reports whether raw provider reasoning must be |
| 1010 | // retained and replayed on all assistant messages. |
| 1011 | func RequiresReasoningRoundTrip(p Provider) bool { |
| 1012 | if nilutil.IsNil(p) { |
| 1013 | return false |
| 1014 | } |
| 1015 | policy, ok := p.(ReasoningRoundTripPolicy) |
| 1016 | return ok && policy.RequiresReasoningRoundTrip() |
| 1017 | } |
| 1018 | |
| 1019 | // MissingToolCallReasoningWarningPolicy is optionally implemented by providers |
| 1020 | // whose replay protocol requires reasoning_content, but whose active model may |
| 1021 | // not reliably emit it. The legacy Warning name is retained for source |
| 1022 | // compatibility; the agent now uses this policy for silent bounded recovery and |
| 1023 | // emits no user-visible protocol notice. |
| 1024 | type MissingToolCallReasoningWarningPolicy interface { |
| 1025 | WarnOnMissingToolCallReasoning() bool |
| 1026 | } |
| 1027 | |
| 1028 | // MissingToolCallReasoningWarningIdentityPolicy optionally supplies the stable, |
| 1029 | // non-credential configuration identity used to rate-limit missing-reasoning |
| 1030 | // recovery attempts. The legacy name preserves adapters and persisted state. |
| 1031 | // Implementations may include adapter kind, endpoint, model, and thinking |
| 1032 | // controls; the raw identity never leaves memory and is hashed before |
| 1033 | // persistence. |
| 1034 | type MissingToolCallReasoningWarningIdentityPolicy interface { |
| 1035 | MissingToolCallReasoningWarningIdentity() string |
| 1036 | } |
| 1037 | |
| 1038 | // WarnOnMissingToolCallReasoning reports whether a tool_calls turn with empty |
| 1039 | // reasoning_content should enter silent recovery. Its legacy name is preserved |
| 1040 | // for provider implementations compiled against the original diagnostic API. |
| 1041 | func WarnOnMissingToolCallReasoning(p Provider) bool { |
| 1042 | if nilutil.IsNil(p) { |
| 1043 | return false |
| 1044 | } |
| 1045 | policy, ok := p.(MissingToolCallReasoningWarningPolicy) |
| 1046 | if ok { |
| 1047 | return policy.WarnOnMissingToolCallReasoning() |
| 1048 | } |
| 1049 | return RequiresToolCallReasoning(p) |
| 1050 | } |
| 1051 | |
| 1052 | // MissingToolCallReasoningWarningFingerprint returns an opaque stable key for |
| 1053 | // one provider configuration's recovery cooldown. Concrete adapters distinguish |
| 1054 | // endpoint/model/protocol changes; providers without the optional policy retain |
| 1055 | // a safe type-and-name fallback. The legacy name preserves the on-disk state |
| 1056 | // contract. The digest prevents local state from exposing raw endpoints or |
| 1057 | // model identifiers. |
| 1058 | func MissingToolCallReasoningWarningFingerprint(p Provider) string { |
| 1059 | if nilutil.IsNil(p) { |
| 1060 | return "" |
| 1061 | } |
| 1062 | identity := fmt.Sprintf("%T\x00%s", p, strings.TrimSpace(p.Name())) |
| 1063 | if policy, ok := p.(MissingToolCallReasoningWarningIdentityPolicy); ok { |
| 1064 | if configured := strings.TrimSpace(policy.MissingToolCallReasoningWarningIdentity()); configured != "" { |
| 1065 | identity = configured |
| 1066 | } |
| 1067 | } |
| 1068 | digest := sha256.Sum256([]byte(identity)) |
| 1069 | return hex.EncodeToString(digest[:]) |
| 1070 | } |
| 1071 | |
| 1072 | // Config is a resolved provider instance configuration. |
| 1073 | type Config struct { |
| 1074 | Name string // instance name, e.g. "deepseek" |
| 1075 | BaseURL string // OpenAI-compatible endpoint |
| 1076 | Model string // model id |
| 1077 | APIKey string // resolved from api_key_env |
| 1078 | Extra map[string]any // kind-specific options |
| 1079 | } |
| 1080 | |
| 1081 | // AuthError reports that a provider rejected the API key (HTTP 401/403). Its |
| 1082 | // message is already user-facing and actionable — it names the provider and, |
| 1083 | // when known, the environment variable the key comes from — and it carries the |
| 1084 | // server's own reason as Body, because relay gateways explain *why* the key was |
| 1085 | // rejected ("token expired", key not entitled to the model) in the response |
| 1086 | // body. Body is deliberately NOT part of Error(): servers echo masked key |
| 1087 | // fragments in auth bodies, and the ambient error string flows into logs, |
| 1088 | // status lines, and traces where key material must never propagate. Display |
| 1089 | // layers that want the reason read Body and extract it themselves. Providers |
| 1090 | // should return this (rather than a generic status error) for auth failures. |
| 1091 | type AuthError struct { |
| 1092 | Provider string // the provider instance name, e.g. "deepseek" |
| 1093 | KeyEnv string // the api_key_env the key is read from, when known |
| 1094 | KeySource string // human-readable source of KeyEnv, when known |
| 1095 | Status int // the HTTP status (401 or 403) |
| 1096 | HasKey bool // a non-empty key was sent — the server rejected it, vs. no key configured at all |
| 1097 | Body string // trimmed response-body snippet, the server's verbatim reason when it gave one |
| 1098 | } |
| 1099 | |
| 1100 | func (e *AuthError) Error() string { |
| 1101 | key := "the API key" |
| 1102 | if e.KeyEnv != "" { |
| 1103 | key = e.KeyEnv |
| 1104 | } |
| 1105 | if e.KeySource != "" { |
| 1106 | key += " from " + e.KeySource |
| 1107 | } |
| 1108 | return fmt.Sprintf("authentication failed for provider %q (HTTP %d): %s is invalid or expired — update it (in .env or your environment) and retry, or run `reasonix setup`", |
| 1109 | e.Provider, e.Status, key) |
| 1110 | } |
| 1111 | |
| 1112 | // Factory builds a Provider from a resolved Config. |
| 1113 | type Factory func(cfg Config) (Provider, error) |
| 1114 | |
| 1115 | var registry = map[string]Factory{} |
| 1116 | |
| 1117 | // Register adds a factory under a kind (e.g. "openai"). Intended for init(). |
| 1118 | // It panics on a duplicate kind, since that is a compile-time wiring mistake. |
| 1119 | func Register(kind string, f Factory) { |
| 1120 | if _, dup := registry[kind]; dup { |
| 1121 | panic("provider: duplicate kind " + kind) |
| 1122 | } |
| 1123 | registry[kind] = f |
| 1124 | } |
| 1125 | |
| 1126 | // New instantiates the provider of the given kind. |
| 1127 | func New(kind string, cfg Config) (Provider, error) { |
| 1128 | f, ok := registry[kind] |
| 1129 | if !ok { |
| 1130 | return nil, fmt.Errorf("provider: unknown kind %q (registered: %v)", kind, Kinds()) |
| 1131 | } |
| 1132 | p, err := f(cfg) |
| 1133 | if err != nil { |
| 1134 | return nil, err |
| 1135 | } |
| 1136 | if nilutil.IsNil(p) { |
| 1137 | return nil, fmt.Errorf("provider: factory %q returned nil provider", kind) |
| 1138 | } |
| 1139 | return p, nil |
| 1140 | } |
| 1141 | |
| 1142 | // Kinds returns the registered kinds, sorted. |
| 1143 | func Kinds() []string { |
| 1144 | out := make([]string, 0, len(registry)) |
| 1145 | for k := range registry { |
| 1146 | out = append(out, k) |
| 1147 | } |
| 1148 | sort.Strings(out) |
| 1149 | return out |
| 1150 | } |
| 1151 |