pi-coralbricks-provider 1.0.6 → 1.0.7

package/README.md

@@ -23,7 +23,6 @@ _A [pi](https://github.com/earendil-works/pi-coding-agent) provider extension fo
 - **Per-family thinking levels** — zai-style `thinking` control for GLM (including a *real* off switch), `reasoning_effort` for Kimi K3 and GPT-OSS
 - **Accurate cost tracking** — pricing mirrors Coral's published rates, and cached reads are **$0 on every model**, so pi's computed cost matches Coral's own `usage.cost` to the token
 - **Self-healing model sync** — stale-while-revalidate from the authenticated `/v1/models` (or the unauthenticated [public catalog](https://www.coralbricks.ai/api/public/models) before auth), hot-swapped at session start
-- **Streaming repair** — transparently fixes Coral's gpt-oss tool-call delta index fragmentation so streamed tool calls always accumulate correctly
 - **synbad-validated** — [synbad](https://github.com/synthetic-lab/synbad) tool-calling and reasoning-parsing evals pass 15/15 on GLM 5.2, GLM 5.3, and Kimi K3 in both unary and streaming modes
 
 ## Installation
@@ -152,16 +151,6 @@ Coral's gateway follows the OpenAI Chat Completions API:
 - **`supportsStrictMode: false`** — Kimi K3 (no strict JSON-schema tool definitions).
 - **`requiresReasoningContentOnAssistantMessages: true`** — Kimi K3.
 
-### Streaming Tool-Call Repair
-
-On the raw wire, Coral occasionally emits a streamed tool call's final arguments fragment on a **new delta index** instead of continuing the existing one (`"index": 1` mid-call), which fragments the call under any spec-compliant accumulator. Observed on gpt-oss-120b; GLM and Kimi streams are correct.
-
-This extension's `streamSimple` pipes SSE responses through a repair stream that rewrites id-less, name-less tool-call deltas claiming a fresh index onto the last real call's index. Deltas that do carry an id/name (new calls, parallel calls) pass through untouched, and streams that are already correct are byte-identical. Verify against a live model:
-
-```bash
-CORALBRICKS_API_KEY=cb_... bun run scripts/probe-stream-fix.ts [model-id]
-```
-
 ### Patch Overrides & Custom Models
 
 - **`patch.json`** — per-model overrides applied on top of `models.json` (reasoning flags, pricing corrections, compat settings, thinking level maps). Currently empty — the curated defaults match the live API.
@@ -178,7 +167,7 @@ Validated with [synbad](https://github.com/synthetic-lab/synbad) — Synthetic's
 | GLM 5.2 FP4 | 15/15 ✅ | 15/15 ✅ | |
 | GLM 5.3 FP4 | 15/15 ✅ | 15/15 ✅ | |
 | Kimi K3 | 15/15 ✅ | 15/15 ✅ | |
-| GPT-OSS 120B | 14/15 ⚠️ | 10–11/15 ❌ raw | repaired in-extension (see above) |
+| GPT-OSS 120B | 14/15 ⚠️ | pending re-eval | gateway tool-call index bug is fixed upstream |
 
 The remaining gpt-oss quirk is model-side: it answers "Paris and London" with a single batched tool call even when `parallel_tool_calls: true` — not a gateway bug.
 

package/index.ts

@@ -424,84 +424,6 @@ async function revalidateModels(apiKey: string | undefined, embeddedModels: Json
   return merged;
 }
 
-// Streaming: Coral gateway tool-call index repair
-//
-// Some Coral models (observed on gpt-oss-120b) emit a streamed tool call's
-// final arguments fragment on a NEW delta index instead of continuing the
-// existing one:
-//
-//   {"tool_calls":[{"id":"call_1","index":0,"function":{"name":"get_weather","arguments":"{\"location\": \"Par"}}]}
-//   {"tool_calls":[{"index":1,"function":{"arguments":"is\"}"}}]}   ← wrong index
-//
-// OpenAI semantics say a delta can only START a tool call when it carries an
-// id (and a function name); continuations omit them. pi-ai's accumulator
-// merges by index, so a mis-indexed fragment materializes a second, truncated
-// tool call. Per choice, we track the index of the last fragment that carried
-// id/name and rewrite id-less, name-less deltas that claim a different index
-// onto it. Deltas that DO carry id/name (new calls, parallel calls) pass
-// through untouched, as do Coral streams that are already correct (GLM/Kimi).
-
-function repairToolCallIndices(chunk: any, lastToolCallIndex: Map<number, number>): boolean {
-  const choices = chunk?.choices;
-  if (!Array.isArray(choices)) return false;
-  let changed = false;
-  for (const choice of choices) {
-    const toolCalls = choice?.delta?.tool_calls;
-    if (!Array.isArray(toolCalls) || toolCalls.length === 0) continue;
-    const choiceIndex = typeof choice.index === "number" ? choice.index : 0;
-    for (const tc of toolCalls) {
-      const hasId = typeof tc?.id === "string" && tc.id.length > 0;
-      const hasName = typeof tc?.function?.name === "string" && tc.function.name.length > 0;
-      if (hasId || hasName) {
-        lastToolCallIndex.set(choiceIndex, tc.index);
-        continue;
-      }
-      const previous = lastToolCallIndex.get(choiceIndex);
-      if (previous !== undefined && typeof tc?.index === "number" && tc.index !== previous) {
-        tc.index = previous;
-        changed = true;
-      }
-    }
-  }
-  return changed;
-}
-
-function createToolCallIndexFixer(): TransformStream<Uint8Array, Uint8Array> {
-  const decoder = new TextDecoder();
-  const encoder = new TextEncoder();
-  const lastToolCallIndex = new Map<number, number>();
-  let buffer = "";
-
-  const fixLine = (line: string): string => {
-    if (!line.startsWith("data:") || !line.includes("\"tool_calls\"")) return line;
-    const payload = line.slice(5).trim();
-    if (!payload || payload === "[DONE]") return line;
-    let chunk: any;
-    try {
-      chunk = JSON.parse(payload);
-    } catch {
-      return line;
-    }
-    const changed = repairToolCallIndices(chunk, lastToolCallIndex);
-    return changed ? "data: " + JSON.stringify(chunk) : line;
-  };
-
-  return new TransformStream<Uint8Array, Uint8Array>({
-    transform(chunk, controller) {
-      buffer += decoder.decode(chunk, { stream: true });
-      let newline: number;
-      while ((newline = buffer.indexOf("\n")) >= 0) {
-        const line = buffer.slice(0, newline);
-        buffer = buffer.slice(newline + 1);
-        controller.enqueue(encoder.encode(fixLine(line) + "\n"));
-      }
-    },
-    flush(controller) {
-      if (buffer.length > 0) controller.enqueue(encoder.encode(fixLine(buffer)));
-    },
-  });
-}
-
 export function streamCoral(
   model: any,
   context: any,
@@ -522,24 +444,8 @@ export function streamCoral(
   const reasoningEffort = clampedReasoning === "off" ? undefined : clampedReasoning;
   const { reasoning: _reasoning, ...streamOptions } = options ?? {};
 
-  // Per-request fetch interceptor (never a global patch — concurrent agent and
-  // helper-model streams must not see each other's wrappers).
-  const upstreamFetch = streamOptions.fetch ?? globalThis.fetch;
-  const coralFetch: typeof globalThis.fetch = async (input, init) => {
-    const response = await upstreamFetch(input as RequestInfo | URL, init);
-    const url = typeof input === "string" ? input : input instanceof URL ? input.toString() : (input as Request).url;
-    const isEventStream = (response.headers.get("content-type") ?? "").includes("text/event-stream");
-    if (!response.body || !isEventStream || !url.includes("/chat/completions")) return response;
-    return new Response(response.body.pipeThrough(createToolCallIndexFixer()), {
-      headers: response.headers,
-      status: response.status,
-      statusText: response.statusText,
-    });
-  };
-
   return streamOpenAICompletions(model, context, {
     ...streamOptions,
-    fetch: coralFetch,
     reasoningEffort,
     apiKey,
   } as any);
@@ -596,4 +502,4 @@ export default function (pi: ExtensionAPI) {
   });
 }
 
-export { applyPatch, buildModels, mergeWithEmbedded, transformApiModel, transformCatalogModel, activeDeprecatedModels, withDeprecated, parseContextWindow, repairToolCallIndices, PROVIDER_ID, BASE_URL };
+export { applyPatch, buildModels, mergeWithEmbedded, transformApiModel, transformCatalogModel, activeDeprecatedModels, withDeprecated, parseContextWindow, PROVIDER_ID, BASE_URL };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-coralbricks-provider",
-  "version": "1.0.6",
+  "version": "1.0.7",
   "description": "CoralBricks provider extension for pi - Access GLM 5.2/5.3, Kimi K3, and GPT-OSS 120B through the Coral Inference API",
   "author": "monotykamary",
   "homepage": "https://github.com/monotykamary/pi-coralbricks-provider#readme",
@@ -40,7 +40,6 @@
     "custom-models.json",
     "patch.json",
     "scripts/update-models.js",
-    "scripts/probe-stream-fix.ts",
     "README.md",
     "LICENSE"
   ],

package/scripts/probe-stream-fix.ts

@@ -1,62 +0,0 @@
-/**
- * Live probe: stream gpt-oss-120b through streamCoral (the extension's
- * streamSimple path) and verify the tool-call index bug is repaired — the
- * accumulated message must contain exactly one complete tool call.
- */
-import { streamCoral } from "../index.ts";
-import modelsData from "../models.json" with { type: "json" };
-
-const apiKey = process.env.CORALBRICKS_API_KEY;
-if (!apiKey) throw new Error("CORALBRICKS_API_KEY not set");
-
-const modelId = process.argv[2] ?? "gpt-oss-120b";
-const model = {
-  ...(modelsData as any[]).find((m) => m.id === modelId)!,
-  provider: "coralbricks",
-  baseUrl: "https://inference.coralbricks.ai/v1",
-  api: "openai-completions",
-};
-
-const context = {
-  systemPrompt: "You are a helpful assistant.",
-  messages: [
-    { role: "user", content: [{ type: "text", text: "What is the weather in Paris? Use the tool." }] },
-  ],
-  tools: [
-    {
-      name: "get_weather",
-      description: "Get current weather for a location",
-      parameters: {
-        type: "object",
-        properties: { location: { type: "string", description: "City name" } },
-        required: ["location"],
-      } as any,
-    },
-  ],
-};
-
-const events = streamCoral(model, context, { apiKey, reasoning: "high" } as any);
-let finalMessage: any = null;
-for await (const event of events) {
-  if (event.type === "done") finalMessage = event.message;
-  if (event.type === "error") finalMessage = event.error;
-}
-
-if (!finalMessage) throw new Error("no final message");
-if (finalMessage.stopReason === "error") {
-  throw new Error("stream error: " + finalMessage.errorMessage);
-}
-
-const calls = (finalMessage.content ?? []).filter((c: any) => c.type === "toolCall");
-console.log("stopReason:", finalMessage.stopReason);
-console.log("toolCalls:", calls.length);
-for (const call of calls) {
-  console.log("  -", call.name, JSON.stringify(call.arguments));
-}
-
-if (calls.length !== 1) throw new Error(`expected exactly 1 tool call, got ${calls.length}`);
-const args = typeof calls[0].arguments === "string" ? JSON.parse(calls[0].arguments) : calls[0].arguments;
-if (!args.location || !/paris/i.test(String(args.location))) {
-  throw new Error("arguments malformed: " + JSON.stringify(calls[0].arguments));
-}
-console.log("PASS: streamCoral produced exactly one complete tool call");