refactor(mitm): replace inspector utilities clean-room (#11748)

Merged via /merge-batch (v3.8.51 provenance sweep). Boarded and validated together with the batch's other provenance PRs in a combined worktree — full gate suite green. Thank you.
This commit is contained in:
Diego Rodrigues de Sa e Souza
2026-08-28 05:04:49 -03:00
committed by GitHub
parent 84504e3f1f
commit 3026183120
12 changed files with 1084 additions and 770 deletions

View File

@@ -1621,7 +1621,7 @@ OmniRoute stands on the shoulders of giants. It started as a fork of **[9router]
<table>
<tr><th align="left">Project</th><th align="center">⭐</th><th align="left">How it inspired OmniRoute</th></tr>
<tr><td nowrap><b><a href="https://github.com/chouzz/llm-interceptor">llm-interceptor</a></b></td><td align="center">66</td><td>MITM interception/analysis of coding-assistant ↔ LLM traffic — our Traffic Inspector ports its SSE merge, conversation normalization, host passthrough and secret masking. The upstream's complete license text is still under provenance review.</td></tr>
<tr><td nowrap><b><a href="https://github.com/chouzz/llm-interceptor">llm-interceptor</a></b></td><td align="center">66</td><td>MITM interception/analysis of coding-assistant ↔ LLM traffic informed early Traffic Inspector requirements. Four previously derived modules — SSE merging, conversation normalization, secret masking and header sanitization — have been replaced by independent clean-room implementations based on public protocol standards. The two host-passthrough surfaces (<code>passthrough.ts</code> and <code>_internal/bypass.cjs</code>) remain OmniRoute-internal implementations classified independently; they were not rewritten as part of that replacement.</td></tr>
<tr><td nowrap><b><a href="https://github.com/InterceptSuite/ProxyBridge">ProxyBridge</a></b></td><td align="center">5,995</td><td>Transparent per-process proxy routing — inspired our crash-safe MITM teardown, socket idle-timeouts, <code>/proc</code> process attribution and TPROXY capture.</td></tr>
</table>

View File

@@ -0,0 +1 @@
- **refactor(mitm):** replace four previously derived Traffic Inspector utilities — conversation normalization, SSE merging, secret masking, and header sanitization — with independent clean-room implementations based on public provider schemas, WHATWG SSE, and RFC 6750, while keeping the two independently classified passthrough surfaces unchanged ([#11748](https://github.com/diegosouzapw/OmniRoute/pull/11748)).

View File

@@ -172,11 +172,16 @@ When set, configures `undici`'s global dispatcher with the extra CA cert, allowi
### 2.7 Secret masking (`src/mitm/maskSecrets.ts`)
Applied to all request bodies and headers **before** they enter the Traffic Inspector buffer or any log:
The independent clean-room scanner is applied to request bodies and credential headers
**before** they enter the Traffic Inspector buffer or any log. It performs a single linear pass:
- `sk-` / `ak-` / `pk-` prefixed tokens (OpenAI/Anthropic-style)
- `Authorization: Bearer <token>` headers
- Generic long tokens (≥40 chars)
- RFC 6750 `Authorization: Bearer <token>` credentials, with whole-token precedence
- Generic long opaque tokens (≥40 chars), including dotted and padded forms
`sanitizeHeaders()` lowercases retained names, joins array values deterministically, drops the
shared hop-by-hop/framing denylist (including proxy authentication), fully redacts `cookie` and
`set-cookie`, and delegates credential values to the scanner.
---

View File

@@ -174,7 +174,7 @@ This is a substantial subsystem with its own dedicated operator guide — see **
| Control | Action |
| ---------------- | --------------------------------------------------------------------- |
| ⎉ Pause | Stops rendering new requests; "X new" badge accumulates |
| 🗑 Clear | Clears the UI list (server buffer is not affected) |
| 🗑 Clear | Clears the UI list (server buffer is not affected) |
| ⬇ Export .har | Downloads current filtered list as HAR file |
| ● Record session | Starts a named recording session |
| Profile selector | LLM only / Custom hosts / All |
@@ -207,12 +207,18 @@ Custom hosts added via Mode 2 inherit their `kind` from the form input (defaults
### 4.2 SSE merger (`src/mitm/inspector/sseMerger.ts`)
**MIT port from [chouzz/llm-interceptor](https://github.com/chouzz/llm-interceptor)**
**Independent clean-room implementation.** Event parsing follows the
[WHATWG server-sent events algorithm](https://html.spec.whatwg.org/multipage/server-sent-events.html#parsing-an-event-stream),
while reconstruction follows the public [OpenAI](https://platform.openai.com/docs/api-reference/chat/create),
[Anthropic](https://platform.claude.com/docs/en/build-with-claude/streaming), and
[Gemini](https://ai.google.dev/api/generate-content#method:-models.streamgeneratecontent)
streaming schemas.
Reconstructs the final assistant message from raw SSE delta events:
- **Anthropic**: accumulates `content_block_delta` by index; handles `text_delta`, `input_json_delta` (tool calls), `thinking_delta`
- **OpenAI**: accumulates `choices[i].delta.content` and `tool_calls` by index
- **OpenAI**: accumulates Chat Completions choices/tool calls and Responses API output items
by index
- **Gemini**: accumulates `candidates[i].content.parts`
- **Unknown**: returns raw events as-is
@@ -220,7 +226,9 @@ The Response tab shows a toggle: **"Raw events ↔ Merged"**.
### 4.3 Conversation normalizer (`src/mitm/inspector/conversationNormalizer.ts`)
**MIT port from [chouzz/llm-interceptor](https://github.com/chouzz/llm-interceptor)**
**Independent clean-room implementation.** Normalization is defined by local black-box
contracts and the public OpenAI, Anthropic, and Gemini message schemas; no upstream
implementation source is used.
Converts OpenAI, Anthropic, and Gemini message formats to a single `NormalizedConversation` before rendering:
@@ -337,14 +345,14 @@ Export via `GET /api/tools/traffic-inspector/sessions/{id}/export.har` or the
Traffic Inspector shows **all intercepted HTTPS traffic**, including authorization headers and request bodies. The following controls are in place:
| Control | Details |
| ----------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
| **LOCAL_ONLY** | All routes and the WebSocket endpoint are loopback-only (enforced in `routeGuard.ts` before auth) |
| **Secret masking** | `maskSecrets()` applied to all headers and bodies before `TrafficBuffer.push()` — enabled by default (`INSPECTOR_MASK_SECRETS=true`) |
| **Body size cap** | Bodies > `INSPECTOR_MAX_BODY_KB` (default 1024 KB) are truncated with `"(truncated for performance)"` notice |
| **Sensitive header masking** | `authorization`, `cookie`, `api-key`, `x-api-key`, `proxy-authorization``Bearer ***` in Headers tab; "Show secrets" toggle |
| **CSP** | Strict Content Security Policy on Traffic Inspector pages to prevent XSS via injected response bodies |
| **No persistence by default** | The `TrafficBuffer` is in-memory and lost on server restart. Sessions are persisted only when explicitly recorded |
| Control | Details |
| ----------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
| **LOCAL_ONLY** | All routes and the WebSocket endpoint are loopback-only (enforced in `routeGuard.ts` before auth) |
| **Secret masking** | Linear `maskSecret()` scanner redacts RFC 6750 Bearer credentials, provider-prefixed keys and long opaque tokens before `TrafficBuffer.push()` |
| **Body size cap** | Bodies > `INSPECTOR_MAX_BODY_KB` (default 1024 KB) are truncated with `"(truncated for performance)"` notice |
| **Header sanitization** | Names are lowercased; framing/hop-by-hop and proxy-auth headers are dropped; cookies are fully redacted; credential values delegate to `maskSecret()` |
| **CSP** | Strict Content Security Policy on Traffic Inspector pages to prevent XSS via injected response bodies |
| **No persistence by default** | The `TrafficBuffer` is in-memory and lost on server restart. Sessions are persisted only when explicitly recorded |
### Hard Rules applied

View File

@@ -1,11 +1,9 @@
/**
* Conversation normalizer — converts OpenAI / Anthropic / Gemini request +
* response payloads into a single provider-agnostic shape.
* Provider-neutral conversation projection for Traffic Inspector views.
*
* MIT — port from https://github.com/chouzz/llm-interceptor (ui/utils.ts)
*
* Returns `null` for non-LLM requests or payloads we cannot understand —
* never throws — so the renderer can fall back to the raw view.
* This is an independent implementation of the public OpenAI, Anthropic, and
* Gemini payload shapes. It intentionally emits only the normalized block
* types consumed by the inspector UI.
*/
import { mergeStream, parseSseStream } from "./sseMerger.ts";
@@ -16,435 +14,333 @@ import type {
NormalizedTurn,
} from "./types.ts";
type JsonRecord = Record<string, unknown>;
type NormalizedRole = NormalizedTurn["role"];
function asRecord(value: unknown): Record<string, unknown> | null {
if (value && typeof value === "object" && !Array.isArray(value)) {
return value as Record<string, unknown>;
}
return null;
const WRAPPER_KEYS = ["body", "payload", "data", "request", "requestBody", "response"];
function asRecord(value: unknown): JsonRecord | null {
return value !== null && typeof value === "object" && !Array.isArray(value)
? (value as JsonRecord)
: null;
}
function tryParseJson(value: string | null | undefined): unknown {
if (!value) return null;
function tryParseJson(value: unknown): unknown {
if (typeof value !== "string") return value;
const trimmed = value.trim();
if (!trimmed) return null;
try {
return JSON.parse(value);
return JSON.parse(trimmed) as unknown;
} catch {
return null;
}
}
function normalizeRole(raw: unknown): NormalizedRole {
if (raw === "system" || raw === "user" || raw === "assistant" || raw === "tool") {
return raw;
}
if (raw === "model") return "assistant";
if (raw === "function") return "tool";
return "user";
function hasConversationShape(record: JsonRecord): boolean {
return (
"messages" in record ||
"input" in record ||
"contents" in record ||
"system" in record ||
"systemInstruction" in record ||
"choices" in record ||
"candidates" in record ||
"content" in record ||
"output" in record
);
}
/**
* OpenAI / Anthropic message content can be a string, or an array of blocks.
* Returns a list of normalized blocks.
*/
function blocksFromOpenAiContent(content: unknown): NormalizedBlock[] {
if (content == null) return [];
if (typeof content === "string") {
if (content.length === 0) return [];
return [{ type: "text", text: content }];
}
if (!Array.isArray(content)) return [];
const out: NormalizedBlock[] = [];
for (const raw of content) {
if (typeof raw === "string") {
out.push({ type: "text", text: raw });
continue;
}
const block = asRecord(raw);
if (!block) continue;
const type = block.type;
if (type === "text" || type === "output_text") {
const text = typeof block.text === "string" ? block.text : "";
out.push({ type: "text", text });
} else if (type === "input_text") {
const text = typeof block.text === "string" ? block.text : "";
out.push({ type: "text", text });
} else if (type === "tool_use") {
out.push({
type: "tool_use",
id: typeof block.id === "string" ? block.id : "",
name: typeof block.name === "string" ? block.name : "",
input: block.input ?? {},
});
} else if (type === "tool_result") {
out.push({
type: "tool_result",
tool_use_id: typeof block.tool_use_id === "string" ? block.tool_use_id : "",
content: block.content ?? null,
});
} else if (typeof block.text === "string") {
out.push({ type: "text", text: block.text });
}
}
return out;
}
/**
* OpenAI assistant messages may declare `tool_calls`. Each becomes a
* `tool_use` block alongside any text content.
*/
function appendOpenAiToolCalls(blocks: NormalizedBlock[], toolCalls: unknown): NormalizedBlock[] {
if (!Array.isArray(toolCalls)) return blocks;
for (const raw of toolCalls) {
const tc = asRecord(raw);
if (!tc) continue;
const fn = asRecord(tc.function) ?? {};
let parsedInput: unknown = {};
if (typeof fn.arguments === "string") {
try {
parsedInput = JSON.parse(fn.arguments);
} catch {
parsedInput = fn.arguments;
function unwrapPayload(value: unknown): unknown {
let current = tryParseJson(value);
for (let depth = 0; depth < 5; depth += 1) {
const record = asRecord(current);
if (!record || hasConversationShape(record)) return current;
let next: unknown = undefined;
for (const key of WRAPPER_KEYS) {
if (record[key] !== undefined) {
next = record[key];
break;
}
} else if (fn.arguments != null) {
parsedInput = fn.arguments;
}
blocks.push({
type: "tool_use",
id: typeof tc.id === "string" ? tc.id : "",
name: typeof fn.name === "string" ? fn.name : "",
input: parsedInput,
});
if (next === undefined) return current;
current = tryParseJson(next);
}
return current;
}
function normalizedRole(value: unknown): NormalizedRole | null {
if (value === "system" || value === "user" || value === "assistant" || value === "tool") {
return value;
}
if (value === "model") return "assistant";
if (value === "function") return "tool";
return null;
}
function textBlock(value: unknown): NormalizedBlock | null {
return typeof value === "string" && value.length > 0 ? { type: "text", text: value } : null;
}
function parseArguments(value: unknown): unknown {
if (typeof value !== "string") return value ?? {};
if (!value.trim()) return {};
try {
return JSON.parse(value) as unknown;
} catch {
return value;
}
}
function toolUseBlock(value: JsonRecord): NormalizedBlock | null {
const directFunction = asRecord(value.function);
const geminiFunction = asRecord(value.functionCall);
const source = directFunction ?? geminiFunction ?? value;
const name = typeof source.name === "string" ? source.name : null;
if (!name) return null;
const idValue = value.id ?? value.call_id ?? geminiFunction?.id ?? name;
const input =
source.arguments !== undefined
? parseArguments(source.arguments)
: source.args !== undefined
? source.args
: source.input !== undefined
? source.input
: {};
return {
type: "tool_use",
id: typeof idValue === "string" ? idValue : name,
name,
input,
};
}
function toolResultBlock(value: JsonRecord): NormalizedBlock | null {
const geminiResponse = asRecord(value.functionResponse);
const idValue =
value.tool_use_id ??
value.tool_call_id ??
value.call_id ??
geminiResponse?.id ??
geminiResponse?.name;
if (typeof idValue !== "string") return null;
const content =
value.output !== undefined
? value.output
: value.content !== undefined
? value.content
: geminiResponse?.response;
return { type: "tool_result", tool_use_id: idValue, content };
}
function reasoningBlocks(value: JsonRecord): NormalizedBlock[] {
const blocks: NormalizedBlock[] = [];
const summary = Array.isArray(value.summary) ? value.summary : [];
for (const entry of summary) {
const record = asRecord(entry);
const block = textBlock(record?.text);
if (block) blocks.push(block);
}
return blocks;
}
/**
* Build NormalizedTurn[] from OpenAI / Anthropic chat messages.
*/
/** Responses API reasoning items carry `summary: [{type: "summary_text", text}]`. */
function reasoningSummaryText(summary: unknown): string {
if (!Array.isArray(summary)) return "";
const parts: string[] = [];
for (const raw of summary) {
const block = asRecord(raw);
if (block && typeof block.text === "string") parts.push(block.text);
function blocksFromPart(value: unknown): NormalizedBlock[] {
if (typeof value === "string") {
const block = textBlock(value);
return block ? [block] : [];
}
return parts.join("\n\n");
}
const part = asRecord(value);
if (!part) return [];
function turnsFromOpenAiMessages(messages: unknown[]): NormalizedTurn[] {
const out: NormalizedTurn[] = [];
for (const raw of messages) {
const msg = asRecord(raw);
if (!msg) continue;
// Responses API items for tool activity/reasoning carry no `role` at
// all — they're distinguished by `type` instead. Handle these before the
// role-based branches below, which would otherwise silently drop them
// (empty `content`, no `tool_calls`, `normalizeRole(undefined)` defaults
// to "user") — the exact gap that made a real OpenClaw request's
// function_call/function_call_output items vanish from the Conversation
// Context panel entirely (2026-08-06).
if (msg.type === "function_call") {
let parsedInput: unknown = {};
if (typeof msg.arguments === "string") {
try {
parsedInput = JSON.parse(msg.arguments);
} catch {
parsedInput = msg.arguments;
}
} else if (msg.arguments != null) {
parsedInput = msg.arguments;
}
out.push({
role: "assistant",
blocks: [
{
type: "tool_use",
id: typeof msg.call_id === "string" ? msg.call_id : "",
name: typeof msg.name === "string" ? msg.name : "",
input: parsedInput,
},
],
});
continue;
}
if (msg.type === "function_call_output") {
out.push({
role: "tool",
blocks: [
{
type: "tool_result",
tool_use_id: typeof msg.call_id === "string" ? msg.call_id : "",
content: msg.output ?? null,
},
],
});
continue;
}
if (msg.type === "reasoning") {
const text = reasoningSummaryText(msg.summary);
if (!text) continue;
out.push({ role: "assistant", blocks: [{ type: "text", text }] });
continue;
}
const role = normalizeRole(msg.role);
if (msg.role === "tool" || msg.role === "function") {
const content = msg.content;
out.push({
role: "tool",
blocks: [
{
type: "tool_result",
tool_use_id:
typeof msg.tool_call_id === "string"
? msg.tool_call_id
: typeof msg.name === "string"
? msg.name
: "",
content,
},
],
});
continue;
}
const blocks = blocksFromOpenAiContent(msg.content);
if ("tool_calls" in msg) {
appendOpenAiToolCalls(blocks, msg.tool_calls);
}
if (blocks.length === 0 && msg.content == null && !("tool_calls" in msg)) {
continue;
}
out.push({ role, blocks });
const type = typeof part.type === "string" ? part.type : "";
if (type === "tool_use" || type === "function_call" || part.functionCall) {
const block = toolUseBlock(part);
return block ? [block] : [];
}
return out;
}
/**
* Gemini contents have a different shape: `[{role, parts: [{text|...}]}]`.
*/
function turnsFromGeminiContents(contents: unknown[]): NormalizedTurn[] {
const out: NormalizedTurn[] = [];
for (const raw of contents) {
const turn = asRecord(raw);
if (!turn) continue;
const role = normalizeRole(turn.role);
const blocks: NormalizedBlock[] = [];
if (Array.isArray(turn.parts)) {
for (const partRaw of turn.parts) {
const part = asRecord(partRaw);
if (!part) continue;
if (typeof part.text === "string") {
blocks.push({ type: "text", text: part.text });
} else if (part.functionCall) {
const fc = asRecord(part.functionCall) ?? {};
blocks.push({
type: "tool_use",
id: typeof fc.name === "string" ? fc.name : "",
name: typeof fc.name === "string" ? fc.name : "",
input: fc.args ?? {},
});
} else if (part.functionResponse) {
const fr = asRecord(part.functionResponse) ?? {};
blocks.push({
type: "tool_result",
tool_use_id: typeof fr.name === "string" ? fr.name : "",
content: fr.response ?? null,
});
}
}
}
if (blocks.length > 0) out.push({ role, blocks });
if (type === "tool_result" || type === "function_call_output" || part.functionResponse) {
const block = toolResultBlock(part);
return block ? [block] : [];
}
return out;
}
/**
* Anthropic Messages API requests carry a top-level `system` field (string
* or array of `{type:"text"|text}` blocks). Convert to a `system` turn.
*/
function systemTurnFromAnthropic(system: unknown): NormalizedTurn | null {
if (!system) return null;
if (typeof system === "string") {
return system.length === 0
? null
: { role: "system", blocks: [{ type: "text", text: system }] };
if (type === "reasoning") return reasoningBlocks(part);
if (
type === "text" ||
type === "input_text" ||
type === "output_text" ||
type === "summary_text" ||
type === ""
) {
const block = textBlock(part.text);
return block ? [block] : [];
}
if (!Array.isArray(system)) return null;
const blocks: NormalizedBlock[] = [];
for (const raw of system) {
const item = asRecord(raw);
if (item && typeof item.text === "string") {
blocks.push({ type: "text", text: item.text });
} else if (typeof raw === "string") {
blocks.push({ type: "text", text: raw });
}
}
if (blocks.length === 0) return null;
return { role: "system", blocks };
}
export function buildRequestTurns(body: unknown): NormalizedTurn[] | null {
const obj = asRecord(body);
if (!obj) return null;
if (Array.isArray(obj.messages)) {
const turns: NormalizedTurn[] = [];
const systemTurn = systemTurnFromAnthropic(obj.system);
if (systemTurn) turns.push(systemTurn);
turns.push(...turnsFromOpenAiMessages(obj.messages));
return turns;
}
if (Array.isArray(obj.contents)) {
const turns: NormalizedTurn[] = [];
const sysObj = asRecord(obj.systemInstruction);
if (sysObj && Array.isArray(sysObj.parts)) {
const parts: NormalizedBlock[] = [];
for (const partRaw of sysObj.parts) {
const p = asRecord(partRaw);
if (p && typeof p.text === "string") parts.push({ type: "text", text: p.text });
}
if (parts.length > 0) turns.push({ role: "system", blocks: parts });
}
turns.push(...turnsFromGeminiContents(obj.contents));
return turns;
}
if (typeof obj.prompt === "string") {
return [{ role: "user", blocks: [{ type: "text", text: obj.prompt }] }];
}
if (typeof obj.input === "string") {
return [{ role: "user", blocks: [{ type: "text", text: obj.input }] }];
}
if (Array.isArray(obj.input)) {
return turnsFromOpenAiMessages(obj.input);
}
return null;
}
function isSseResponse(req: InterceptedRequest): boolean {
const accept = req.requestHeaders["accept"] ?? req.requestHeaders["Accept"] ?? "";
const ct = req.responseHeaders["content-type"] ?? req.responseHeaders["Content-Type"] ?? "";
return (
accept.includes("event-stream") ||
ct.includes("event-stream") ||
/^\s*event:|^\s*data:/m.test(req.responseBody ?? "")
);
}
function extractAnthropicResponseTurn(message: unknown): NormalizedTurn | null {
const obj = asRecord(message);
if (!obj) return null;
const content = obj.content;
if (!Array.isArray(content)) return null;
const blocks: NormalizedBlock[] = [];
for (const raw of content) {
const block = asRecord(raw);
if (!block) continue;
if (block.type === "text" && typeof block.text === "string") {
blocks.push({ type: "text", text: block.text });
} else if (block.type === "tool_use") {
blocks.push({
type: "tool_use",
id: typeof block.id === "string" ? block.id : "",
name: typeof block.name === "string" ? block.name : "",
input: block.input ?? {},
});
} else if (block.type === "thinking" && typeof block.thinking === "string") {
blocks.push({ type: "text", text: block.thinking });
}
}
if (blocks.length === 0) return null;
return { role: "assistant", blocks };
}
function extractOpenAiResponseTurn(message: unknown): NormalizedTurn | null {
const obj = asRecord(message);
if (!obj || !Array.isArray(obj.choices)) return null;
const first = asRecord(obj.choices[0]);
if (!first) return null;
const msg = asRecord(first.message) ?? asRecord(first.delta);
if (!msg) return null;
const blocks = blocksFromOpenAiContent(msg.content);
if ("tool_calls" in msg) appendOpenAiToolCalls(blocks, msg.tool_calls);
if (blocks.length === 0) return null;
return { role: "assistant", blocks };
}
function extractGeminiResponseTurn(message: unknown): NormalizedTurn | null {
const obj = asRecord(message);
if (!obj || !Array.isArray(obj.candidates)) return null;
const first = asRecord(obj.candidates[0]);
if (!first) return null;
const content = asRecord(first.content);
if (!content || !Array.isArray(content.parts)) return null;
const blocks: NormalizedBlock[] = [];
for (const partRaw of content.parts) {
const part = asRecord(partRaw);
if (!part) continue;
if (typeof part.text === "string") {
blocks.push({ type: "text", text: part.text });
} else if (part.functionCall) {
const fc = asRecord(part.functionCall) ?? {};
blocks.push({
type: "tool_use",
id: typeof fc.name === "string" ? fc.name : "",
name: typeof fc.name === "string" ? fc.name : "",
input: fc.args ?? {},
});
}
}
if (blocks.length === 0) return null;
return { role: "assistant", blocks };
}
export function buildResponseTurns(req: InterceptedRequest): NormalizedTurn[] {
const raw = req.responseBody ?? "";
if (!raw) return [];
let payload: unknown = null;
if (isSseResponse(req)) {
const merged = mergeStream(parseSseStream(raw));
payload = merged.message ?? null;
} else {
payload = tryParseJson(raw);
}
if (!payload) return [];
const anth = extractAnthropicResponseTurn(payload);
if (anth) return [anth];
const oai = extractOpenAiResponseTurn(payload);
if (oai) return [oai];
const gem = extractGeminiResponseTurn(payload);
if (gem) return [gem];
return [];
}
/**
* Normalize an intercepted LLM request + response into a provider-agnostic
* conversation. Returns `null` for non-LLM requests or unparseable payloads.
*/
function blocksFromContent(value: unknown): NormalizedBlock[] {
if (Array.isArray(value)) return value.flatMap(blocksFromPart);
return blocksFromPart(value);
}
function turnFromMessage(value: unknown): NormalizedTurn | null {
const message = asRecord(value);
if (!message) return null;
const type = typeof message.type === "string" ? message.type : "";
if (type === "function_call") {
const block = toolUseBlock(message);
return block ? { role: "assistant", blocks: [block] } : null;
}
if (type === "function_call_output") {
const block = toolResultBlock(message);
return block ? { role: "tool", blocks: [block] } : null;
}
if (type === "reasoning") {
const blocks = reasoningBlocks(message);
return blocks.length > 0 ? { role: "assistant", blocks } : null;
}
const role = normalizedRole(message.role);
if (!role) return null;
if (role === "tool") {
const block = toolResultBlock(message);
return block ? { role, blocks: [block] } : null;
}
const blocks = blocksFromContent(message.content ?? message.parts);
for (const rawCall of Array.isArray(message.tool_calls) ? message.tool_calls : []) {
const call = asRecord(rawCall);
const block = call ? toolUseBlock(call) : null;
if (block) blocks.push(block);
}
const legacyCall = asRecord(message.function_call);
if (legacyCall) {
const block = toolUseBlock({
id: message.tool_call_id ?? legacyCall.name,
function: legacyCall,
});
if (block) blocks.push(block);
}
if (blocks.length === 0) return null;
const projectedRole =
role === "user" && blocks.every((block) => block.type === "tool_result") ? "tool" : role;
return { role: projectedRole, blocks };
}
function turnsFromItems(items: unknown[]): NormalizedTurn[] {
const turns: NormalizedTurn[] = [];
for (const item of items) {
const turn = turnFromMessage(item);
if (turn) turns.push(turn);
}
return turns;
}
function systemTurn(value: unknown): NormalizedTurn | null {
const record = asRecord(value);
const content = record?.parts ?? value;
const blocks = blocksFromContent(content);
return blocks.length > 0 ? { role: "system", blocks } : null;
}
/** Normalize a decoded request payload, or a JSON string/wrapper containing one. */
export function buildRequestTurns(body: unknown): NormalizedTurn[] | null {
const payload = asRecord(unwrapPayload(body));
if (!payload) {
const block = textBlock(typeof body === "string" ? body : null);
return block ? [{ role: "user", blocks: [block] }] : null;
}
const turns: NormalizedTurn[] = [];
const topLevelSystem = payload.systemInstruction ?? payload.system ?? payload.instructions;
if (topLevelSystem !== undefined) {
const turn = systemTurn(topLevelSystem);
if (turn) turns.push(turn);
}
if (Array.isArray(payload.messages)) turns.push(...turnsFromItems(payload.messages));
if (Array.isArray(payload.contents)) turns.push(...turnsFromItems(payload.contents));
if (Array.isArray(payload.input)) turns.push(...turnsFromItems(payload.input));
if (typeof payload.input === "string") {
const block = textBlock(payload.input);
if (block) turns.push({ role: "user", blocks: [block] });
}
return turns.length > 0 ? turns : null;
}
function turnFromAnthropicResponse(payload: JsonRecord): NormalizedTurn | null {
if (!Array.isArray(payload.content)) return null;
const blocks = blocksFromContent(payload.content);
return blocks.length > 0 ? { role: "assistant", blocks } : null;
}
function turnsFromOpenAiResponse(payload: JsonRecord): NormalizedTurn[] {
const turns: NormalizedTurn[] = [];
for (const rawChoice of Array.isArray(payload.choices) ? payload.choices : []) {
const choice = asRecord(rawChoice);
const turn = choice ? turnFromMessage(choice.message ?? choice.delta) : null;
if (turn) turns.push(turn);
}
return turns;
}
function turnsFromGeminiResponse(payload: JsonRecord): NormalizedTurn[] {
const turns: NormalizedTurn[] = [];
for (const rawCandidate of Array.isArray(payload.candidates) ? payload.candidates : []) {
const candidate = asRecord(rawCandidate);
const content = asRecord(candidate?.content);
const blocks = blocksFromContent(content?.parts);
if (blocks.length > 0) turns.push({ role: "assistant", blocks });
}
return turns;
}
function turnsFromResponsesApi(payload: JsonRecord): NormalizedTurn[] {
return turnsFromItems(Array.isArray(payload.output) ? payload.output : []);
}
function isSseResponse(req: InterceptedRequest): boolean {
const contentType =
req.responseHeaders["content-type"] ?? req.responseHeaders["Content-Type"] ?? "";
const body = req.responseBody ?? "";
return (
contentType.toLowerCase().includes("text/event-stream") ||
body.startsWith("data:") ||
body.startsWith("event:") ||
body.includes("\ndata:") ||
body.includes("\nevent:")
);
}
/** Normalize the response side of an intercepted request. */
export function buildResponseTurns(req: InterceptedRequest): NormalizedTurn[] {
if (!req.responseBody) return [];
let decoded: unknown;
if (isSseResponse(req)) {
decoded = mergeStream(parseSseStream(req.responseBody)).message;
} else {
decoded = unwrapPayload(req.responseBody);
}
const payload = asRecord(unwrapPayload(decoded));
if (!payload) return [];
const turns = turnsFromOpenAiResponse(payload);
turns.push(...turnsFromGeminiResponse(payload));
turns.push(...turnsFromResponsesApi(payload));
const anthropic = turnFromAnthropicResponse(payload);
if (anthropic) turns.push(anthropic);
if (turns.length > 0) return turns;
const nestedMessage = turnFromMessage(payload.message);
return nestedMessage ? [nestedMessage] : [];
}
export function normalizeConversation(req: InterceptedRequest): NormalizedConversation | null {
if (req.detectedKind !== "llm") return null;
const requestBody = tryParseJson(req.requestBody);
const requestTurns = buildRequestTurns(requestBody);
if (!requestTurns) return null;
const responseTurns = buildResponseTurns(req);
if (req.detectedKind !== undefined && req.detectedKind !== "llm") return null;
const request = buildRequestTurns(req.requestBody);
if (!request || request.length === 0) return null;
return {
request: requestTurns,
response: responseTurns,
request,
response: buildResponseTurns(req),
contextKey: req.contextKey ?? null,
};
}

View File

@@ -1,11 +1,8 @@
/**
* SSE merger — reconstructs complete LLM response from streaming SSE chunks.
* Server-sent event parsing and provider response reconstruction.
*
* MIT — port from https://github.com/chouzz/llm-interceptor (merger.py)
*
* Detects API format by chunk shape (not URL — robust to URL rewrite) and
* rebuilds Anthropic / OpenAI / Gemini responses. Falls back to a raw event
* list when the format is unrecognised so the caller never crashes.
* This module is an independent implementation based on the WHATWG event-stream
* algorithm and the public OpenAI, Anthropic, and Gemini streaming schemas.
*/
export type ApiFormat = "anthropic" | "openai" | "gemini" | "unknown";
@@ -23,294 +20,475 @@ export interface MergedResponse {
raw?: SseEvent[];
}
function asRecord(value: unknown): Record<string, unknown> | null {
if (value && typeof value === "object" && !Array.isArray(value)) {
return value as Record<string, unknown>;
}
return null;
type JsonRecord = Record<string, unknown>;
interface AnthropicBlockState {
block: JsonRecord;
partialInput: string;
}
/**
* Inspect chunk shapes to determine the upstream API. Matches on the first
* recognisable hint; returns `"unknown"` if none match.
*/
export function detectApiFormat(chunks: SseEvent[]): ApiFormat {
for (const c of chunks) {
const j = asRecord(c.json);
if (!j) continue;
if (j.type === "message_start" || j.type === "content_block_delta") return "anthropic";
if (Array.isArray(j.choices)) {
const first = j.choices[0];
if (first && typeof first === "object" && "delta" in first) return "openai";
interface OpenAiToolState {
value: JsonRecord;
functionValue: JsonRecord;
}
interface OpenAiChoiceState {
value: JsonRecord;
message: JsonRecord;
tools: Map<number, OpenAiToolState>;
}
function asRecord(value: unknown): JsonRecord | null {
return value !== null && typeof value === "object" && !Array.isArray(value)
? (value as JsonRecord)
: null;
}
function asArray(value: unknown): unknown[] {
return Array.isArray(value) ? value : [];
}
function asIndex(value: unknown, fallback: number): number {
return typeof value === "number" && Number.isInteger(value) && value >= 0 ? value : fallback;
}
function appendString(record: JsonRecord, key: string, value: unknown): void {
if (typeof value !== "string") return;
const current = typeof record[key] === "string" ? record[key] : "";
record[key] = current + value;
}
function mergeRecord(previous: unknown, next: unknown): JsonRecord {
return { ...(asRecord(previous) ?? {}), ...(asRecord(next) ?? {}) };
}
function dispatchSseEvent(
events: SseEvent[],
dataLines: string[],
sawDataField: boolean,
eventName: string
): void {
if (!sawDataField) return;
const data = dataLines.join("\n");
const event: SseEvent = { data };
if (eventName) event.event = eventName;
if (data !== "[DONE]") {
try {
event.json = JSON.parse(data) as unknown;
} catch {
// Raw data is still useful to callers when a provider emits a sentinel or malformed JSON.
}
if (Array.isArray(j.candidates)) return "gemini";
}
events.push(event);
}
/** Parse a complete `text/event-stream` payload using WHATWG field semantics. */
export function parseSseStream(raw: string): SseEvent[] {
const input = raw.charCodeAt(0) === 0xfeff ? raw.slice(1) : raw;
const events: SseEvent[] = [];
let dataLines: string[] = [];
let eventName = "";
let sawDataField = false;
let lineStart = 0;
const processLine = (line: string): void => {
if (line.length === 0) {
dispatchSseEvent(events, dataLines, sawDataField, eventName);
dataLines = [];
eventName = "";
sawDataField = false;
return;
}
if (line.startsWith(":")) return;
const colon = line.indexOf(":");
const field = colon === -1 ? line : line.slice(0, colon);
let value = colon === -1 ? "" : line.slice(colon + 1);
if (value.startsWith(" ")) value = value.slice(1);
if (field === "data") {
dataLines.push(value);
sawDataField = true;
} else if (field === "event") {
eventName = value;
}
// `id`, `retry`, comments, and extension fields do not alter the public event shape.
};
for (let index = 0; index < input.length; index += 1) {
const code = input.charCodeAt(index);
if (code !== 0x0a && code !== 0x0d) continue;
processLine(input.slice(lineStart, index));
if (code === 0x0d && input.charCodeAt(index + 1) === 0x0a) index += 1;
lineStart = index + 1;
}
// WHATWG does not dispatch an event that lacks its terminating blank line.
return events;
}
export function detectApiFormat(chunks: SseEvent[]): ApiFormat {
for (const chunk of chunks) {
const namedEvent = chunk.event ?? "";
if (
namedEvent === "message_start" ||
namedEvent === "message_delta" ||
namedEvent === "message_stop" ||
namedEvent.startsWith("content_block_")
) {
return "anthropic";
}
const payload = asRecord(chunk.json);
if (!payload) continue;
const type = typeof payload.type === "string" ? payload.type : "";
if (
type === "message_start" ||
type === "message_delta" ||
type === "message_stop" ||
type.startsWith("content_block_")
) {
return "anthropic";
}
if (Array.isArray(payload.choices) || type.startsWith("response.")) return "openai";
if (Array.isArray(payload.candidates) || asRecord(payload.usageMetadata)) return "gemini";
}
return "unknown";
}
/**
* Parse a raw SSE stream (the response body string captured by the proxy)
* into discrete events. Empty blocks and `[DONE]` terminators are skipped
* silently; malformed JSON payloads are kept as raw `data` (no `json`).
*/
export function parseSseStream(raw: string): SseEvent[] {
const events: SseEvent[] = [];
if (!raw) return events;
// SSE blocks separated by blank lines — accept both LF and CRLF.
for (const block of raw.split(/\r?\n\r?\n/)) {
if (!block.trim()) continue;
const ev: SseEvent = {};
for (const line of block.split(/\r?\n/)) {
if (line.startsWith("event:")) {
ev.event = line.slice(6).trim();
} else if (line.startsWith("data:")) {
ev.data = (ev.data ?? "") + line.slice(5).trim();
export function rebuildAnthropic(chunks: SseEvent[]): MergedResponse {
let message: JsonRecord = { type: "message", role: "assistant", content: [] };
const blocks = new Map<number, AnthropicBlockState>();
for (const chunk of chunks) {
const payload = asRecord(chunk.json);
if (!payload) continue;
const type = typeof payload.type === "string" ? payload.type : "";
if (type === "message_start") {
const startedMessage = asRecord(payload.message);
if (startedMessage) {
message = { ...startedMessage };
for (const [position, value] of asArray(startedMessage.content).entries()) {
const block = asRecord(value);
if (block) blocks.set(position, { block: { ...block }, partialInput: "" });
}
}
}
if (ev.data === undefined) continue;
if (ev.data === "[DONE]") {
events.push(ev);
continue;
}
try {
ev.json = JSON.parse(ev.data);
} catch {
// keep raw data only
const index = asIndex(payload.index, blocks.size);
if (type === "content_block_start") {
const contentBlock = asRecord(payload.content_block);
if (contentBlock) {
blocks.set(index, { block: { ...contentBlock }, partialInput: "" });
}
continue;
}
events.push(ev);
}
return events;
}
interface AnthropicBlock {
type: string;
text?: string;
thinking?: string;
id?: string;
name?: string;
input?: unknown;
}
/**
* Rebuild an Anthropic Messages API response from streaming events.
* Handles `text_delta`, `thinking_delta`, and `input_json_delta` deltas;
* applies `JSON.parse` (best-effort) on accumulated tool-use input.
*/
export function rebuildAnthropic(chunks: SseEvent[]): MergedResponse {
const blocks: AnthropicBlock[] = [];
let message: Record<string, unknown> | null = null;
const inputJsonByIndex: Record<number, string> = {};
for (const c of chunks) {
const j = asRecord(c.json);
if (!j) continue;
const t = j.type;
if (t === "message_start") {
const m = asRecord(j.message);
message = m ? { ...m } : {};
} else if (t === "content_block_start") {
const idx = typeof j.index === "number" ? j.index : blocks.length;
const cb = asRecord(j.content_block);
const block: AnthropicBlock = { type: "text" };
if (cb) {
for (const [k, v] of Object.entries(cb)) (block as Record<string, unknown>)[k] = v;
if (type === "content_block_delta") {
const delta = asRecord(payload.delta);
if (!delta) continue;
const state = blocks.get(index) ?? { block: {}, partialInput: "" };
const deltaType = typeof delta.type === "string" ? delta.type : "";
if (deltaType === "text_delta") appendString(state.block, "text", delta.text);
if (deltaType === "thinking_delta") appendString(state.block, "thinking", delta.thinking);
if (deltaType === "signature_delta") {
appendString(state.block, "signature", delta.signature);
}
if (block.type === "text" && block.text === undefined) block.text = "";
if (block.type === "thinking" && block.thinking === undefined) block.thinking = "";
if (block.type === "tool_use" && block.input === undefined) block.input = {};
blocks[idx] = block;
} else if (t === "content_block_delta") {
const idx = typeof j.index === "number" ? j.index : 0;
const d = asRecord(j.delta);
if (!d) continue;
// Ensure a block slot exists (some streams skip content_block_start).
const slot = blocks[idx] ?? (blocks[idx] = { type: "text", text: "" });
const dType = d.type;
if (dType === "text_delta" && typeof d.text === "string") {
slot.text = (slot.text ?? "") + d.text;
} else if (dType === "thinking_delta" && typeof d.thinking === "string") {
slot.thinking = (slot.thinking ?? "") + d.thinking;
} else if (dType === "input_json_delta" && typeof d.partial_json === "string") {
inputJsonByIndex[idx] = (inputJsonByIndex[idx] ?? "") + d.partial_json;
if (deltaType === "input_json_delta" && typeof delta.partial_json === "string") {
state.partialInput += delta.partial_json;
}
} else if (t === "content_block_stop") {
const idx = typeof j.index === "number" ? j.index : 0;
const slot = blocks[idx];
if (slot && slot.type === "tool_use" && inputJsonByIndex[idx]) {
blocks.set(index, state);
continue;
}
if (type === "content_block_stop") {
const state = blocks.get(index);
if (state?.partialInput) {
try {
slot.input = JSON.parse(inputJsonByIndex[idx]);
state.block.input = JSON.parse(state.partialInput) as unknown;
} catch {
// keep accumulated string for forensic visibility
slot.input = inputJsonByIndex[idx];
state.block.input = state.partialInput;
}
}
} else if (t === "message_delta") {
if (!message) message = {};
const d = asRecord(j.delta);
if (d && typeof d.stop_reason === "string") {
message.stop_reason = d.stop_reason;
}
const usage = asRecord(j.usage);
if (usage) {
const prev = asRecord(message.usage) ?? {};
message.usage = { ...prev, ...usage };
continue;
}
if (type === "message_delta") {
Object.assign(message, asRecord(payload.delta) ?? {});
if (payload.usage !== undefined) {
message.usage = mergeRecord(message.usage, payload.usage);
}
}
}
const filledBlocks = blocks.filter((b) => b !== undefined);
return {
format: "anthropic",
message: { ...(message ?? {}), content: filledBlocks },
message.content = [...blocks.entries()]
.sort(([left], [right]) => left - right)
.map(([, state]) => state.block);
return { format: "anthropic", message };
}
function getOpenAiChoice(
choices: Map<number, OpenAiChoiceState>,
index: number
): OpenAiChoiceState {
const existing = choices.get(index);
if (existing) return existing;
const created: OpenAiChoiceState = {
value: { index },
message: { role: "assistant", content: "" },
tools: new Map(),
};
choices.set(index, created);
return created;
}
interface OpenAiToolCall {
index: number;
id?: string;
type?: string;
function: { name: string; arguments: string };
function mergeOpenAiTools(state: OpenAiChoiceState, toolDeltas: unknown[]): void {
for (const [position, rawTool] of toolDeltas.entries()) {
const tool = asRecord(rawTool);
if (!tool) continue;
const index = asIndex(tool.index, position);
const current = state.tools.get(index) ?? { value: { index }, functionValue: {} };
for (const [key, value] of Object.entries(tool)) {
if (key !== "function" && key !== "index" && value !== undefined) {
current.value[key] = value;
}
}
const functionDelta = asRecord(tool.function);
if (functionDelta) {
if (typeof functionDelta.name === "string") {
appendString(current.functionValue, "name", functionDelta.name);
}
if (typeof functionDelta.arguments === "string") {
appendString(current.functionValue, "arguments", functionDelta.arguments);
}
current.value.function = current.functionValue;
}
state.tools.set(index, current);
}
}
interface OpenAiChoice {
index: number;
message: {
role: string;
content: string;
tool_calls?: OpenAiToolCall[];
refusal?: string;
function rebuildOpenAiResponses(chunks: SseEvent[]): JsonRecord {
let response: JsonRecord = { object: "response", output: [] };
const items = new Map<number, JsonRecord>();
const contentByItem = new Map<number, Map<number, JsonRecord>>();
const getItem = (outputIndex: number): JsonRecord => {
const existing = items.get(outputIndex);
if (existing) return existing;
const created: JsonRecord = { type: "message", role: "assistant", content: [] };
items.set(outputIndex, created);
return created;
};
finish_reason: string | null;
const getPart = (outputIndex: number, contentIndex: number): JsonRecord => {
let content = contentByItem.get(outputIndex);
if (!content) {
content = new Map();
contentByItem.set(outputIndex, content);
}
const existing = content.get(contentIndex);
if (existing) return existing;
const created: JsonRecord = { type: "output_text", text: "" };
content.set(contentIndex, created);
return created;
};
const mergeItem = (outputIndex: number, rawItem: unknown): void => {
const item = asRecord(rawItem);
if (!item) return;
const current = getItem(outputIndex);
for (const [key, value] of Object.entries(item)) {
if (key !== "content" && value !== undefined) current[key] = value;
}
for (const [contentIndex, rawPart] of asArray(item.content).entries()) {
const part = asRecord(rawPart);
if (part) Object.assign(getPart(outputIndex, contentIndex), part);
}
};
for (const chunk of chunks) {
const payload = asRecord(chunk.json);
if (!payload) continue;
const embeddedResponse = asRecord(payload.response);
if (embeddedResponse) {
for (const [key, value] of Object.entries(embeddedResponse)) {
if (key !== "output" && value !== undefined) response[key] = value;
}
for (const [outputIndex, item] of asArray(embeddedResponse.output).entries()) {
mergeItem(outputIndex, item);
}
}
const outputIndex = asIndex(payload.output_index, 0);
const contentIndex = asIndex(payload.content_index, 0);
if (
payload.type === "response.output_item.added" ||
payload.type === "response.output_item.done"
) {
mergeItem(outputIndex, payload.item);
}
if (
payload.type === "response.content_part.added" ||
payload.type === "response.content_part.done"
) {
const part = asRecord(payload.part);
if (part) Object.assign(getPart(outputIndex, contentIndex), part);
}
if (payload.type === "response.output_text.delta") {
appendString(getPart(outputIndex, contentIndex), "text", payload.delta);
}
if (payload.type === "response.output_text.done" && typeof payload.text === "string") {
getPart(outputIndex, contentIndex).text = payload.text;
}
if (payload.type === "response.refusal.delta") {
const part = getPart(outputIndex, contentIndex);
part.type = "refusal";
appendString(part, "refusal", payload.delta);
}
if (payload.type === "response.refusal.done" && typeof payload.refusal === "string") {
const part = getPart(outputIndex, contentIndex);
part.type = "refusal";
part.refusal = payload.refusal;
}
if (payload.type === "response.function_call_arguments.delta") {
const item = getItem(outputIndex);
item.type = "function_call";
appendString(item, "arguments", payload.delta);
}
if (
payload.type === "response.function_call_arguments.done" &&
typeof payload.arguments === "string"
) {
const item = getItem(outputIndex);
item.type = "function_call";
item.arguments = payload.arguments;
}
}
response.output = [...items.entries()]
.sort(([left], [right]) => left - right)
.map(([outputIndex, item]) => {
const content = contentByItem.get(outputIndex);
if (content && content.size > 0) {
item.content = [...content.entries()]
.sort(([left], [right]) => left - right)
.map(([, part]) => part);
}
return item;
});
return response;
}
/**
* Rebuild an OpenAI Chat Completions response from streaming events.
* Accumulates content text and tool-call fragments per choice/index.
*/
export function rebuildOpenAI(chunks: SseEvent[]): MergedResponse {
const choicesByIdx: Record<number, OpenAiChoice> = {};
let model: string | null = null;
let usage: unknown = null;
let id: string | null = null;
const hasChatChunks = chunks.some((chunk) => Array.isArray(asRecord(chunk.json)?.choices));
if (!hasChatChunks) {
return { format: "openai", message: rebuildOpenAiResponses(chunks) };
}
for (const c of chunks) {
const j = asRecord(c.json);
if (!j) continue;
if (typeof j.model === "string") model = j.model;
if (typeof j.id === "string") id = j.id;
if (j.usage != null) usage = j.usage;
if (!Array.isArray(j.choices)) continue;
const result: JsonRecord = {};
const choices = new Map<number, OpenAiChoiceState>();
for (const raw of j.choices) {
const ch = asRecord(raw);
if (!ch) continue;
const idx = typeof ch.index === "number" ? ch.index : 0;
const slot = (choicesByIdx[idx] ??= {
index: idx,
message: { role: "assistant", content: "" },
finish_reason: null,
});
const delta = asRecord(ch.delta) ?? {};
if (typeof delta.role === "string") slot.message.role = delta.role;
if (typeof delta.content === "string") slot.message.content += delta.content;
if (typeof delta.refusal === "string") {
slot.message.refusal = (slot.message.refusal ?? "") + delta.refusal;
}
if (Array.isArray(delta.tool_calls)) {
slot.message.tool_calls ??= [];
for (const tcRaw of delta.tool_calls) {
const tc = asRecord(tcRaw);
if (!tc) continue;
const ti = typeof tc.index === "number" ? tc.index : 0;
const tcSlot =
slot.message.tool_calls[ti] ??
(slot.message.tool_calls[ti] = {
index: ti,
function: { name: "", arguments: "" },
});
if (typeof tc.id === "string") tcSlot.id = tc.id;
if (typeof tc.type === "string") tcSlot.type = tc.type;
const fn = asRecord(tc.function);
if (fn) {
if (typeof fn.name === "string") tcSlot.function.name += fn.name;
if (typeof fn.arguments === "string") tcSlot.function.arguments += fn.arguments;
}
for (const chunk of chunks) {
const payload = asRecord(chunk.json);
if (!payload) continue;
for (const [key, value] of Object.entries(payload)) {
if (key !== "choices" && key !== "usage" && value !== undefined) result[key] = value;
}
if (payload.usage !== undefined) result.usage = payload.usage;
for (const [position, rawChoice] of asArray(payload.choices).entries()) {
const choice = asRecord(rawChoice);
if (!choice) continue;
const index = asIndex(choice.index, position);
const state = getOpenAiChoice(choices, index);
const delta = asRecord(choice.delta);
if (delta) {
if (typeof delta.role === "string") state.message.role = delta.role;
appendString(state.message, "content", delta.content);
appendString(state.message, "refusal", delta.refusal);
mergeOpenAiTools(state, asArray(delta.tool_calls));
const functionCall = asRecord(delta.function_call);
if (functionCall) {
const current = asRecord(state.message.function_call) ?? {};
appendString(current, "name", functionCall.name);
appendString(current, "arguments", functionCall.arguments);
state.message.function_call = current;
}
}
if (typeof ch.finish_reason === "string") slot.finish_reason = ch.finish_reason;
}
}
return {
format: "openai",
message: {
id,
model,
choices: Object.values(choicesByIdx).sort((a, b) => a.index - b.index),
usage,
},
};
}
/**
* Rebuild a Gemini `generateContent`-style response from streaming events.
* Concatenates all parts across emitted candidates into a single candidate.
*/
export function rebuildGemini(chunks: SseEvent[]): MergedResponse {
const parts: unknown[] = [];
let usageMetadata: unknown = null;
let finishReason: unknown = null;
let modelVersion: string | null = null;
for (const c of chunks) {
const j = asRecord(c.json);
if (!j) continue;
if (j.usageMetadata != null) usageMetadata = j.usageMetadata;
if (typeof j.modelVersion === "string") modelVersion = j.modelVersion;
if (!Array.isArray(j.candidates)) continue;
for (const candRaw of j.candidates) {
const cand = asRecord(candRaw);
if (!cand) continue;
if (cand.finishReason != null) finishReason = cand.finishReason;
const content = asRecord(cand.content);
if (!content) continue;
const ps = content.parts;
if (Array.isArray(ps)) {
for (const p of ps) parts.push(p);
for (const [key, value] of Object.entries(choice)) {
if (key !== "delta" && value !== null && value !== undefined) state.value[key] = value;
}
}
}
return {
format: "gemini",
message: {
candidates: [
{
content: { parts, role: "model" },
...(finishReason != null ? { finishReason } : {}),
},
],
...(modelVersion ? { modelVersion } : {}),
...(usageMetadata != null ? { usageMetadata } : {}),
},
};
result.choices = [...choices.entries()]
.sort(([left], [right]) => left - right)
.map(([, state]) => {
if (state.tools.size > 0) {
state.message.tool_calls = [...state.tools.entries()]
.sort(([left], [right]) => left - right)
.map(([, tool]) => tool.value);
}
return { ...state.value, message: state.message };
});
return { format: "openai", message: result };
}
export function rebuildGemini(chunks: SseEvent[]): MergedResponse {
const result: JsonRecord = {};
const candidates = new Map<number, JsonRecord>();
const parts = new Map<number, unknown[]>();
for (const chunk of chunks) {
const payload = asRecord(chunk.json);
if (!payload) continue;
for (const [key, value] of Object.entries(payload)) {
if (key !== "candidates" && value !== undefined) result[key] = value;
}
for (const [position, rawCandidate] of asArray(payload.candidates).entries()) {
const candidate = asRecord(rawCandidate);
if (!candidate) continue;
const index = asIndex(candidate.index, position);
const current = candidates.get(index) ?? { index };
const content = asRecord(candidate.content);
const currentParts = parts.get(index) ?? [];
if (content) currentParts.push(...asArray(content.parts));
parts.set(index, currentParts);
for (const [key, value] of Object.entries(candidate)) {
if (key !== "content" && value !== undefined) current[key] = value;
}
if (content) {
current.content = {
...asRecord(current.content),
...content,
parts: currentParts,
};
}
candidates.set(index, current);
}
}
result.candidates = [...candidates.entries()]
.sort(([left], [right]) => left - right)
.map(([, candidate]) => candidate);
return { format: "gemini", message: result };
}
/**
* Merge an array of SSE events into a single rebuilt response. Returns
* `{ format: "unknown", raw }` (no throw) for unrecognised shapes.
*/
export function mergeStream(chunks: SseEvent[]): MergedResponse {
const format = detectApiFormat(chunks);
switch (format) {
case "anthropic":
return rebuildAnthropic(chunks);
case "openai":
return rebuildOpenAI(chunks);
case "gemini":
return rebuildGemini(chunks);
default:
return { format: "unknown", raw: chunks };
}
if (format === "anthropic") return rebuildAnthropic(chunks);
if (format === "openai") return rebuildOpenAI(chunks);
if (format === "gemini") return rebuildGemini(chunks);
return { format: "unknown", raw: chunks };
}

View File

@@ -1,31 +1,122 @@
/**
* Secret masking utilities for MITM traffic inspection.
* Applied to all headers/bodies before any log or broadcast.
* Regex patterns are pre-compiled (order matters: BEARER first).
* Linear secret scanner for captured Traffic Inspector text.
*
* Pattern sources: plano 11 §4.8 (origin: llm-interceptor proxy.py:310)
* Bearer credentials follow RFC 6750's token alphabet and take precedence over
* heuristic key masking so no suffix of a credential survives a partial match.
*/
// Pre-compiled regex patterns — ORDER IS SIGNIFICANT (BEARER must run first).
// BEARER matches the token after a standalone "Bearer " — NOT only after a
// literal "authorization:" prefix. sanitizeHeaders() masks header *values*
// ("Bearer <token>") with the key already stripped, so a prefix-anchored regex
// never fired there and short/opaque-but-<40 tokens leaked into the inspector
// (found by the AgentBridge live capture). The char class is bounded + linear
// (no nested quantifiers) to stay ReDoS-safe.
const BEARER = /(\bBearer\s+)[A-Za-z0-9._~+/-]+=*/gi;
const SK_KEY = /\b(sk|ak|pk)-[A-Za-z0-9_-]{16,}\b/g;
const LONG_TOKEN = /\b[A-Za-z0-9_-]{40,}\b/g;
const PREFIXED_KEY_BODY_MIN = 16;
const OPAQUE_TOKEN_MIN = 40;
/**
* Mask secrets in a string value.
* - Bearer tokens: replaces the token after any "Bearer " with "***"
* - sk-/ak-/pk- keys: keeps first 6 chars + last 2 chars
* - Long opaque tokens (≥40 chars): keeps first 4 chars + last 2 chars
*/
interface Match {
start: number;
end: number;
replacement: string;
}
function isAsciiLetter(code: number): boolean {
return (code >= 0x41 && code <= 0x5a) || (code >= 0x61 && code <= 0x7a);
}
function isAsciiDigit(code: number): boolean {
return code >= 0x30 && code <= 0x39;
}
function isKeyCharCode(code: number): boolean {
return isAsciiLetter(code) || isAsciiDigit(code) || code === 0x2d || code === 0x5f;
}
function isBearerCoreCharCode(code: number): boolean {
return isKeyCharCode(code) || code === 0x2e || code === 0x7e || code === 0x2b || code === 0x2f;
}
function isOpaqueCharCode(code: number): boolean {
return isBearerCoreCharCode(code);
}
function equalsAsciiIgnoreCase(value: string, index: number, expected: string): boolean {
if (index + expected.length > value.length) return false;
for (let offset = 0; offset < expected.length; offset += 1) {
const actualCode = value.charCodeAt(index + offset);
const expectedCode = expected.charCodeAt(offset);
const folded = actualCode >= 0x41 && actualCode <= 0x5a ? actualCode + 0x20 : actualCode;
if (folded !== expectedCode) return false;
}
return true;
}
function bearerMatch(value: string, index: number): Match | null {
if (!equalsAsciiIgnoreCase(value, index, "bearer")) return null;
if (index > 0 && isKeyCharCode(value.charCodeAt(index - 1))) return null;
const schemeEnd = index + 6;
if (value.charCodeAt(schemeEnd) !== 0x20) return null;
let tokenStart = schemeEnd;
while (value.charCodeAt(tokenStart) === 0x20) tokenStart += 1;
let end = tokenStart;
while (end < value.length && isBearerCoreCharCode(value.charCodeAt(end))) end += 1;
if (end === tokenStart) return null;
while (value.charCodeAt(end) === 0x3d) end += 1;
return { start: tokenStart, end, replacement: "***" };
}
function prefixedKeyMatch(value: string, index: number): Match | null {
if (index > 0 && isOpaqueCharCode(value.charCodeAt(index - 1))) return null;
const prefix = value.slice(index, index + 3);
if (prefix !== "sk-" && prefix !== "ak-" && prefix !== "pk-") return null;
let end = index + 3;
while (end < value.length && isOpaqueCharCode(value.charCodeAt(end))) end += 1;
if (end - (index + 3) < PREFIXED_KEY_BODY_MIN) return null;
while (value.charCodeAt(end) === 0x3d) end += 1;
const token = value.slice(index, end);
return {
start: index,
end,
replacement: `${token.slice(0, 6)}${token.slice(-2)}`,
};
}
function opaqueTokenMatch(value: string, index: number): Match | null {
if (!isOpaqueCharCode(value.charCodeAt(index))) return null;
if (index > 0 && isOpaqueCharCode(value.charCodeAt(index - 1))) return null;
let end = index + 1;
while (end < value.length && isOpaqueCharCode(value.charCodeAt(end))) end += 1;
if (end - index < OPAQUE_TOKEN_MIN) return null;
while (value.charCodeAt(end) === 0x3d) end += 1;
const token = value.slice(index, end);
return {
start: index,
end,
replacement: `${token.slice(0, 4)}${token.slice(-2)}`,
};
}
/** Mask Bearer credentials, provider-style keys, and long opaque tokens. */
export function maskSecret(value: string): string {
return value
.replace(BEARER, "$1***")
.replace(SK_KEY, (m) => `${m.slice(0, 6)}${m.slice(-2)}`)
.replace(LONG_TOKEN, (m) => `${m.slice(0, 4)}${m.slice(-2)}`);
let output = "";
let literalStart = 0;
let index = 0;
while (index < value.length) {
const match =
bearerMatch(value, index) ?? prefixedKeyMatch(value, index) ?? opaqueTokenMatch(value, index);
if (!match) {
index += 1;
continue;
}
output += value.slice(literalStart, match.start);
output += match.replacement;
index = match.end;
literalStart = match.end;
}
if (literalStart === 0) return value;
return output + value.slice(literalStart);
}

View File

@@ -1,61 +1,37 @@
import type { IncomingHttpHeaders } from "node:http";
import { isForbiddenUpstreamHeaderName } from "@/shared/constants/upstreamHeaders";
import { maskSecret } from "./maskSecrets";
import { isForbiddenUpstreamHeaderName } from "../shared/constants/upstreamHeaders.ts";
import { maskSecret } from "./maskSecrets.ts";
/**
* Header names whose values must be masked (case-insensitive).
* These carry credentials/tokens that must not appear in logs or broadcasts.
*/
const SECRET_HEADER_NAMES = new Set([
const MASKED_CREDENTIAL_HEADERS = new Set([
"authorization",
"cookie",
// `set-cookie` is the RESPONSE-side credential header — upstream session/CSRF
// cookies must be masked too, else they leak verbatim into inspector JSON.
"set-cookie",
"x-api-key",
"api-key",
"bearer",
"proxy-authorization",
"x-api-key",
"x-auth-token",
"x-goog-api-key",
]);
const FULLY_REDACTED_HEADERS = new Set(["cookie", "set-cookie"]);
function isSecretHeader(name: string): boolean {
return SECRET_HEADER_NAMES.has(name.toLowerCase());
function sanitizeCredentialValue(value: string): string {
const masked = maskSecret(value);
return masked === value ? "[REDACTED]" : masked;
}
/**
* Sanitize HTTP headers for safe logging/broadcasting.
*
* - Removes headers in the upstream denylist (hop-by-hop, Host, etc.)
* - Applies maskSecret() to values of authorization/cookie/key headers
* - Coerces array values to comma-joined strings
* - Returns a plain Record<string, string> (never undefined values)
*/
export function sanitizeHeaders(
headers: IncomingHttpHeaders | Record<string, string | string[] | undefined>,
headers: IncomingHttpHeaders | Record<string, string | string[] | undefined>
): Record<string, string> {
const result: Record<string, string> = {};
for (const [key, value] of Object.entries(headers)) {
if (value === undefined || value === null) continue;
const lowerKey = key.toLowerCase();
// Remove denylist headers (hop-by-hop, framing)
if (isForbiddenUpstreamHeaderName(lowerKey)) continue;
// Normalize array values
const strValue = Array.isArray(value) ? value.join(", ") : String(value);
// Mask secret header values
if (isSecretHeader(lowerKey)) {
// `set-cookie` carries an entire session/CSRF cookie value that maskSecret's
// format heuristics (Bearer / sk- / ≥40-char) do NOT catch, so it would leak
// verbatim into inspector JSON — fully redact it (SECURITY_AUDIT M6).
result[lowerKey] = lowerKey === "set-cookie" ? "[REDACTED]" : maskSecret(strValue);
} else {
result[lowerKey] = strValue;
const sanitized: Record<string, string> = {};
for (const [rawName, rawValue] of Object.entries(headers)) {
const name = rawName.trim().toLowerCase();
if (!name) continue;
if (isForbiddenUpstreamHeaderName(name)) continue;
if (FULLY_REDACTED_HEADERS.has(name)) {
sanitized[name] = "[REDACTED]";
continue;
}
}
if (rawValue === undefined) continue;
return result;
const value = Array.isArray(rawValue) ? rawValue.join(", ") : rawValue;
sanitized[name] = MASKED_CREDENTIAL_HEADERS.has(name) ? sanitizeCredentialValue(value) : value;
}
return sanitized;
}

View File

@@ -165,6 +165,26 @@ test("normalizes Responses API reasoning items (no `role` field) into an assista
assert.equal((conv.request[0].blocks[0] as { text: string }).text, "Thinking about the request.");
});
test("normalizes Responses API instructions as a system turn before input", () => {
const req = makeReq({
path: "/v1/responses",
requestBody: JSON.stringify({
instructions: "Answer tersely.",
input: [{ role: "user", content: [{ type: "input_text", text: "Hello" }] }],
}),
});
const conv = normalizeConversation(req);
assert.ok(conv);
assert.deepEqual(
conv.request.map((turn) => turn.role),
["system", "user"]
);
assert.equal(
(conv.request[0].blocks[0] as { type: "text"; text: string }).text,
"Answer tersely."
);
});
test("normalizes Anthropic request with top-level system + tool_use response", () => {
const req = makeReq({
host: "api.anthropic.com",
@@ -190,6 +210,29 @@ test("normalizes Anthropic request with top-level system + tool_use response", (
assert.equal(conv.response[0].blocks[1].type, "tool_use");
});
test("normalizes Anthropic user-carried tool_result content as a tool turn", () => {
const req = makeReq({
host: "api.anthropic.com",
path: "/v1/messages",
requestBody: JSON.stringify({
messages: [
{
role: "user",
content: [{ type: "tool_result", tool_use_id: "toolu_1", content: "sunny" }],
},
],
}),
});
const conv = normalizeConversation(req);
assert.ok(conv);
assert.equal(conv.request[0].role, "tool");
assert.deepEqual(conv.request[0].blocks[0], {
type: "tool_result",
tool_use_id: "toolu_1",
content: "sunny",
});
});
test("normalizes Gemini request contents + functionCall response", () => {
const req = makeReq({
host: "generativelanguage.googleapis.com",

View File

@@ -15,9 +15,7 @@ function jsonChunks(payloads: unknown[]): SseEvent[] {
}
test("detectApiFormat — message_start → anthropic", () => {
const chunks = jsonChunks([
{ type: "message_start", message: { id: "msg_1" } },
]);
const chunks = jsonChunks([{ type: "message_start", message: { id: "msg_1" } }]);
assert.equal(detectApiFormat(chunks), "anthropic");
});
@@ -28,16 +26,24 @@ test("detectApiFormat — content_block_delta → anthropic", () => {
assert.equal(detectApiFormat(chunks), "anthropic");
});
test("detectApiFormat — Anthropic SSE event name is authoritative when JSON type is absent", () => {
const chunks: SseEvent[] = [{ event: "message_start", data: "{}", json: {} }];
assert.equal(detectApiFormat(chunks), "anthropic");
});
test("detectApiFormat — choices[].delta → openai", () => {
const chunks = jsonChunks([
{ choices: [{ index: 0, delta: { content: "Hi" } }] },
]);
const chunks = jsonChunks([{ choices: [{ index: 0, delta: { content: "Hi" } }] }]);
assert.equal(detectApiFormat(chunks), "openai");
});
test("detectApiFormat — candidates → gemini", () => {
const chunks = jsonChunks([{ candidates: [{ content: { parts: [{ text: "Hello" }] } }] }]);
assert.equal(detectApiFormat(chunks), "gemini");
});
test("detectApiFormat — Gemini usage-only terminal chunk remains identifiable", () => {
const chunks = jsonChunks([
{ candidates: [{ content: { parts: [{ text: "Hello" }] } }] },
{ usageMetadata: { promptTokenCount: 3, candidatesTokenCount: 2 }, modelVersion: "gemini" },
]);
assert.equal(detectApiFormat(chunks), "gemini");
});
@@ -57,7 +63,11 @@ test("rebuildAnthropic — concat text_delta by index", () => {
]);
const merged = rebuildAnthropic(chunks);
assert.equal(merged.format, "anthropic");
const msg = merged.message as { content: Array<{ text: string }>; stop_reason: string; usage: { output_tokens: number } };
const msg = merged.message as {
content: Array<{ text: string }>;
stop_reason: string;
usage: { output_tokens: number };
};
assert.equal(msg.content[0].text, "Hello world");
assert.equal(msg.stop_reason, "end_turn");
assert.equal(msg.usage.output_tokens, 7);
@@ -71,12 +81,22 @@ test("rebuildAnthropic — input_json_delta merges and JSON.parses", () => {
index: 0,
content_block: { type: "tool_use", id: "tu_1", name: "search" },
},
{ type: "content_block_delta", index: 0, delta: { type: "input_json_delta", partial_json: '{"q":' } },
{ type: "content_block_delta", index: 0, delta: { type: "input_json_delta", partial_json: '"hi"}' } },
{
type: "content_block_delta",
index: 0,
delta: { type: "input_json_delta", partial_json: '{"q":' },
},
{
type: "content_block_delta",
index: 0,
delta: { type: "input_json_delta", partial_json: '"hi"}' },
},
{ type: "content_block_stop", index: 0 },
]);
const merged = rebuildAnthropic(chunks);
const msg = merged.message as { content: Array<{ type: string; input: { q: string }; name: string }> };
const msg = merged.message as {
content: Array<{ type: string; input: { q: string }; name: string }>;
};
assert.equal(msg.content[0].type, "tool_use");
assert.equal(msg.content[0].name, "search");
assert.deepEqual(msg.content[0].input, { q: "hi" });
@@ -86,7 +106,11 @@ test("rebuildAnthropic — thinking_delta accumulates", () => {
const chunks = jsonChunks([
{ type: "content_block_start", index: 0, content_block: { type: "thinking" } },
{ type: "content_block_delta", index: 0, delta: { type: "thinking_delta", thinking: "I am " } },
{ type: "content_block_delta", index: 0, delta: { type: "thinking_delta", thinking: "thinking..." } },
{
type: "content_block_delta",
index: 0,
delta: { type: "thinking_delta", thinking: "thinking..." },
},
]);
const merged = rebuildAnthropic(chunks);
const msg = merged.message as { content: Array<{ thinking: string }> };
@@ -95,7 +119,11 @@ test("rebuildAnthropic — thinking_delta accumulates", () => {
test("rebuildOpenAI — concat delta.content per choice index", () => {
const chunks = jsonChunks([
{ id: "c1", model: "gpt-4", choices: [{ index: 0, delta: { role: "assistant", content: "Hel" } }] },
{
id: "c1",
model: "gpt-4",
choices: [{ index: 0, delta: { role: "assistant", content: "Hel" } }],
},
{ choices: [{ index: 0, delta: { content: "lo" }, finish_reason: null }] },
{ choices: [{ index: 0, delta: {}, finish_reason: "stop" }] },
{ usage: { prompt_tokens: 3, completion_tokens: 1 } },
@@ -123,7 +151,14 @@ test("rebuildOpenAI — merges tool_calls per (choice, tool) index", () => {
{
index: 0,
delta: {
tool_calls: [{ index: 0, id: "tc_1", type: "function", function: { name: "search", arguments: '{"q":' } }],
tool_calls: [
{
index: 0,
id: "tc_1",
type: "function",
function: { name: "search", arguments: '{"q":' },
},
],
},
},
],
@@ -152,6 +187,46 @@ test("rebuildOpenAI — merges tool_calls per (choice, tool) index", () => {
assert.equal(msg.choices[0].message.tool_calls[0].function.arguments, '{"q":"hi"}');
});
test("rebuildOpenAI — reconstructs Responses API text items and completion metadata", () => {
const chunks = jsonChunks([
{ type: "response.created", response: { id: "resp_1", object: "response", output: [] } },
{
type: "response.output_item.added",
output_index: 0,
item: { id: "msg_1", type: "message", role: "assistant", content: [] },
},
{
type: "response.content_part.added",
output_index: 0,
content_index: 0,
part: { type: "output_text", text: "" },
},
{ type: "response.output_text.delta", output_index: 0, content_index: 0, delta: "Hel" },
{ type: "response.output_text.delta", output_index: 0, content_index: 0, delta: "lo" },
{
type: "response.completed",
response: {
id: "resp_1",
object: "response",
status: "completed",
output: [],
usage: { input_tokens: 2, output_tokens: 1, total_tokens: 3 },
},
},
]);
const merged = rebuildOpenAI(chunks);
const response = merged.message as {
status: string;
usage: { total_tokens: number };
output: Array<{ id: string; content: Array<{ text: string }> }>;
};
assert.equal(response.status, "completed");
assert.equal(response.usage.total_tokens, 3);
assert.equal(response.output[0].id, "msg_1");
assert.equal(response.output[0].content[0].text, "Hello");
});
test("rebuildGemini — merges parts across candidates", () => {
const chunks = jsonChunks([
{ candidates: [{ content: { parts: [{ text: "Hello " }] } }] },
@@ -171,10 +246,7 @@ test("rebuildGemini — merges parts across candidates", () => {
});
test("mergeStream — unknown format returns raw fallback (no crash)", () => {
const chunks: SseEvent[] = [
{ data: "garbage" },
{ data: "{}", json: { foo: 1 } },
];
const chunks: SseEvent[] = [{ data: "garbage" }, { data: "{}", json: { foo: 1 } }];
const merged = mergeStream(chunks);
assert.equal(merged.format, "unknown");
assert.ok(Array.isArray(merged.raw));
@@ -182,10 +254,7 @@ test("mergeStream — unknown format returns raw fallback (no crash)", () => {
});
test("parseSseStream — parses event/data blocks separated by blank lines", () => {
const raw =
"event: foo\ndata: 1\n\n" +
'data: {"x":1}\n\n' +
"data: [DONE]\n\n";
const raw = "event: foo\ndata: 1\n\n" + 'data: {"x":1}\n\n' + "data: [DONE]\n\n";
const events = parseSseStream(raw);
assert.equal(events.length, 3);
assert.equal(events[0].event, "foo");

View File

@@ -39,6 +39,23 @@ test("maskSecret — long opaque token (≥40 chars) is masked", () => {
assert.ok(result.length < longToken.length);
});
test("maskSecret — dotted opaque tokens are redacted as one integral credential", () => {
const token =
"eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM0NTY3ODkwIiwic2NvcGUiOiJhZG1pbiJ9.signature1234567890";
const result = maskSecret(token);
assert.ok(result.startsWith("eyJh"));
assert.ok(result.endsWith("…90"));
assert.ok(!result.includes("eyJzdWIi"), "the middle JWT-like segment must not survive");
assert.equal((result.match(/…/g) ?? []).length, 1, "the credential must be masked once");
});
test("maskSecret — padded opaque tokens include trailing padding in one redaction", () => {
const token = `${"A".repeat(42)}==`;
const result = maskSecret(token);
assert.equal(result, "AAAA…==");
assert.equal((result.match(/…/g) ?? []).length, 1);
});
test("maskSecret — string without secrets is unchanged", () => {
const safe = "Content-Type: application/json";
assert.equal(maskSecret(safe), safe);

View File

@@ -26,9 +26,31 @@ test("sanitizeHeaders — Set-Cookie header name is matched case-insensitively",
assert.equal(out["set-cookie"], "[REDACTED]");
});
test("sanitizeHeaders — request cookies are fully redacted regardless of token shape", () => {
const out = sanitizeHeaders({ Cookie: "session=short; csrf=also-short" });
assert.deepEqual(out, { cookie: "[REDACTED]" });
});
test("sanitizeHeaders — authorization bearer token is still masked, not leaked", () => {
const out = sanitizeHeaders({ authorization: "Bearer sk-proj-abcdefghijklmnop" });
assert.ok(!out["authorization"].includes("sk-proj-abcdefghijklmnop"), "bearer token must be masked");
assert.ok(
!out["authorization"].includes("sk-proj-abcdefghijklmnop"),
"bearer token must be masked"
);
});
test("sanitizeHeaders — drops framing, hop-by-hop, and proxy credential headers", () => {
const out = sanitizeHeaders({
Host: "provider.example",
Connection: "keep-alive",
"Content-Length": "42",
"Proxy-Authorization": "Basic c2VjcmV0",
"Proxy-Authenticate": "Basic realm=proxy",
TE: "trailers",
Upgrade: "websocket",
"X-Keep": "visible",
});
assert.deepEqual(out, { "x-keep": "visible" });
});
test("sanitizeHeaders — non-secret headers pass through unchanged", () => {
@@ -36,3 +58,11 @@ test("sanitizeHeaders — non-secret headers pass through unchanged", () => {
assert.equal(out["content-type"], "application/json");
assert.equal(out["x-request-id"], "req-42");
});
test("sanitizeHeaders — lowercases names, joins arrays predictably, and omits undefined", () => {
const out = sanitizeHeaders({
"X-Trace": ["edge-a", "edge-b"],
"X-Missing": undefined,
});
assert.deepEqual(out, { "x-trace": "edge-a, edge-b" });
});