03: Chunk Normalization

このコンテンツはまだ日本語訳がありません。

Chunk normalization converts provider-native RawStreamChunk envelopes into uniform ProcessedChunk objects, extracts reasoning content, and normalizes error and tool-call payloads. It executes synchronously within BaseStreamAdapter.processChunk().

Flow and ownership

The stage 04 orchestrator loop consumes raw chunks from stage 02 and passes each to processChunk(). The method produces no side effects on orchestrator state or Discord APIs. It can update adapter-local accumulators and guards and emit diagnostic logs, while shielding the orchestrator from vendor SDK structures.

The returned ProcessedChunk carries one of four primary types:

"text": carries a visible text delta in content for buffering and Discord delivery.
"function_call": carries a normalized FunctionCall (name, args, and optional thoughtSignature).
"error": carries a normalized ProviderError with classified failure codes.
"done": carries terminal metadata (such as finish reasons) before the generator finishes.

Any chunk variant may attach thoughts (ThoughtLogEntry[]) for chain-of-thought capture and servingProvider to track upstream backends on routed endpoints.

Token usage can arrive in a trailing chunk with empty choices. OpenRouter and OpenAI-compatible adapters preserve its metadata on an empty text chunk, so the orchestrator carries the counts into StreamResult.usage for statistics and quotas. Reported usage.cost is preserved alongside the token counts.

Reasoning extraction and spill guards

Reasoning content travels through dedicated fields or in-band tags:

Structured reasoning: Adapters extract native thought blocks (such as Gemini part.thought or reasoning summaries) into ThoughtLogEntry[]. For OpenRouter, the upstream backend identifier is captured in servingProvider so unexpected reasoning leaks can be traced to specific model hosts.
Leaked reasoning tags: When models output <think> blocks into visible text deltas, ThinkBlockContentStripper redirects the enclosed text to delta.reasoning.
Tagless reasoning spills: When a provider concatenates reasoning directly onto content without tags, ReasoningContentSpillGuard detects unspaced sentence boundaries at the start of output. It strips the leaked tail while protecting backticked code identifiers.

Tool-call normalization and recovery

Tool calls stream as structured objects or raw text deltas:

Verbatim tool-call parsing: Models using custom endpoints or lacking native tool calling can output function calls directly in prose. CustomStreamAdapter runs VerbatimToolCallParser over visible content to match <toolName>(...) syntax. Text preceding the call emits as normal prose, followed by the parsed call as type: "function_call".
Truncated argument recovery: Network interruptions or socket resets can cut tool argument JSON mid-token. In adapters receiving incremental argument deltas (OpenAI-compatible, OpenRouter, Anthropic), parseAccumulatedToolArguments() falls back to tryRepairIncompleteJson(). The repair closes open containers without inventing keys and marks chunk.functionCall.argumentsTruncated = true. The tool loop inspects this flag and refuses execution of incomplete calls.

Error normalization and classification

handleProviderError() converts SDK exceptions into a unified ProviderError shape with standard categories (api_error, rate_limit, content_blocked, timeout, provider_overloaded, model_error). Specialized classifiers in src/utils/provider/providerErrorClassification.ts inspect error text to differentiate failure causes:

Model errors: Distinguishes unknown or retired model IDs from bad credentials.
Billing limits: Separates exhausted account balances from affordability ceilings that can succeed with smaller max_tokens settings.
Context overflow: Identifies token window limits that prompt history trimming.
Provider timeouts: Normalizes network read and socket timeouts into the timeout category.

Public sinks show only the provider’s localized, classified headline from createErrorDescription(). The upstream response text never reaches error embeds, interaction replies, the fallback notice, or tool results returned to the model, because an endpoint can echo the bearer value it received in an encoding that pattern redaction misses, and self-debug reads those notices back into context. The chat error embed instead offers a Show Full Error button that opens the redacted snippet in a modal, which reaches only the viewer and never enters channel history. Only someone who could replace the failing key may open it: a Manage Server member (or the DM’s user) for server credentials, and only the triggering user for a personal key. Error codes appear only when formatProviderErrorCodeForDisplay() recognizes an HTTP status or single-case enum word; anything else shows as unknown. The orchestrator reports each failed stream attempt once through reportStreamProviderFailure(), before deciding whether to show it, so normal chat, suppressed retries, and hidden image planning share one record shape and severity: a failure that needs the operator creates an error even when its notice is suppressed. The report caps the diagnostic snippet from getProviderErrorDetail() at 2,000 characters after removing the exact key that attempt sent, which may be a personal or rotated key rather than the server’s. Log messages, attached errors and metadata also remove that key before credential-pattern redaction and optional truncation. This is best-effort redaction; encoded or transformed secrets can survive. Attached errors and stack traces can be longer; log string truncation is opt-in. Expected provider failures retain warning diagnostics in JSON logs at every log level, without creating an error_logs row.

Google streaming failures attach a bounded requestDiagnostics summary: the actual request model, last five dialogue roles, final-turn part counts, prefill and tool-history presence, received chunk count, and elapsed time. The summary contains no prompt text, tool arguments, media URLs, or media bytes. Suppressed error notices carry the summary in the same report.

Empty non-private DeepSeek structured replies retain the finish reason, validated token counts, reasoning presence, output limit, and provider request ID. Missing or invalid token counts stay unknown. Expression initialization copies these diagnostics into its existing error record. Include error_metadata in error exports; expected failures remain in host JSON logs.

Logs remain private operator diagnostics. See Production tuning for log controls. /persona generate and speech synthesis also keep upstream text out of public replies: preset generators reply with a classified message and HTTP status, and speech adapters put only the status or fixed text in details, which reaches both /generate voice-message and the voice tool.

Credential-scoped recovery tips

The resolved credential source of the turn (StreamContext.textCredentialSource, "server", "personal", or "supporter") governs the troubleshooting tips displayed in error embeds:

Server scope: Tips recommend administrative actions, such as adding API keys to rotation pools or configuring server fallback chains.
Personal scope: Server key rotation tips are suppressed because standard members cannot manage server pools. Error embeds instead recommend checking personal credentials or disabling the personal text override to revert to server defaults.
Subscription scope: Subscription denials use the billing module’s classified recovery copy and suppress provider troubleshooting tips.

Source pointers

src/types/stream/interfaces.ts: ProcessedChunk, ProviderError, and BaseStreamAdapter.processChunk.
src/providers/utils/toolCallArguments.ts: parseAccumulatedToolArguments and truncation flags.
src/utils/text/jsonRepair.ts: tryRepairIncompleteJson.
src/utils/provider/providerErrorClassification.ts: error classification helpers.
src/utils/provider/providerErrorMetrics.ts: reportStreamProviderFailure and reportProviderFailure.
src/utils/discord/stream/errorUi.ts: error embed presentation, credential-scoped tips, and the footer naming whether the provider or TomoriBot failed.
src/providers/custom/customStreamAdapter.ts: verbatim tool parsing integration.