feat(session): low-yield repeated probing detection

packages/opencode/src/session/export.ts

@@ -130,7 +130,17 @@ export namespace Export {
         omitted_attachment_count: number
       }
     }
-    diagnostics: Record<string, never>
+    diagnostics: {
+      loop?: {
+        last?: {
+          parentID: string
+          type: "same_input" | "same_target"
+          action: "block" | "stop"
+          tool: string
+          completedFailures: number
+        }
+      }
+    }
     session: Tree
   }
 
@@ -139,6 +149,64 @@ export namespace Export {
     childInfos: Session.Info[]
   }
 
+  export function deriveSnapshotDiagnostics(node: Tree): {
+    loop?: {
+      last?: {
+        parentID: string
+        type: "same_input" | "same_target"
+        action: "block" | "stop"
+        tool: string
+        completedFailures: number
+      }
+    }
+  } {
+    let lastAt = -Infinity
+    let last:
+      | {
+          parentID: string
+          type: "same_input" | "same_target"
+          action: "block" | "stop"
+          tool: string
+          completedFailures: number
+        }
+      | undefined
+    const walk = (t: Tree) => {
+      for (const message of t.messages ?? []) {
+        if (message.info.role !== "assistant") continue
+        for (const part of message.parts) {
+          if (part.type !== "tool") continue
+          const metadata = "metadata" in part.state ? part.state.metadata : undefined
+          const loop = metadata?.diagnostics?.loop as
+            | {
+                loopAction?: string
+                loopType?: string
+                loopCompletedFailures?: number
+              }
+            | undefined
+          if (!loop || (loop.loopAction !== "block" && loop.loopAction !== "stop")) continue
+          if (!loop.loopType || typeof loop.loopCompletedFailures !== "number" || !message.info.parentID) continue
+          let at = -Infinity
+          if ("time" in part.state) {
+            const t = part.state.time
+            at = "end" in t && typeof t.end === "number" ? t.end : t.start
+          }
+          if (at < lastAt) continue
+          lastAt = at
+          last = {
+            parentID: message.info.parentID,
+            type: loop.loopType === "input" ? "same_input" : "same_target",
+            action: loop.loopAction,
+            tool: part.tool,
+            completedFailures: loop.loopCompletedFailures,
+          }
+        }
+      }
+      for (const child of t.children ?? []) walk(child)
+    }
+    walk(node)
+    return last ? { loop: { last } } : {}
+  }
+
   const climbToRoot = Effect.fn("Export.climbToRoot")(function* (svc: Session.Interface, id: SessionID) {
     let current: Session.Info = yield* svc.get(id)
     while (current.parentID) {
@@ -296,7 +364,7 @@ export namespace Export {
         model_refs,
         stats: countStats(tree, ctx.count.omitted),
       },
-      diagnostics: {},
+      diagnostics: deriveSnapshotDiagnostics(tree),
       session: tree,
     } satisfies Snapshot
   })

packages/opencode/src/session/loop-renderer.ts

@@ -0,0 +1,126 @@
+import { SessionDiagnostics } from "./diagnostics"
+
+export namespace LoopRenderer {
+  export type RenderInput = {
+    tool: string
+    state: SessionDiagnostics.SignatureState
+    locale?: string
+  }
+
+  export function render(input: RenderInput): string {
+    const { tool, state, locale } = input
+    const errorLine = firstErrorLine(state.lastError)
+    const scrubbedError = errorLine ? scrubErrorText(errorLine) : undefined
+    const truncatedError = scrubbedError ? SessionDiagnostics.truncateForRenderer(scrubbedError) : undefined
+    const isZh = (locale ?? "").toLowerCase().startsWith("zh")
+
+    if (tool === "webfetch") {
+      const rawURL = extractURL(state.lastInput)
+      const cleanedURL = rawURL ? stripQueryAndFragment(rawURL) : undefined
+      const truncatedURL = cleanedURL ? SessionDiagnostics.truncateForRenderer(cleanedURL) : undefined
+      if (state.kind === "input") {
+        if (isZh) {
+          return truncatedURL
+            ? `我重复调用了相同的请求 ${state.completedFailures} 次没成功，已停止。请求：webfetch ${truncatedURL}`
+            : `我重复调用了相同的请求 ${state.completedFailures} 次没成功，已停止。`
+        }
+        return truncatedURL
+          ? `I made the same request ${state.completedFailures} times without success and stopped. Request: webfetch ${truncatedURL}`
+          : `I made the same request ${state.completedFailures} times without success and stopped.`
+      }
+      if (isZh) {
+        if (truncatedURL && truncatedError)
+          return `我重复抓取同一个目标 ${state.completedFailures} 次都失败，已停止。目标：${truncatedURL} 错误：${truncatedError}`
+        if (truncatedURL)
+          return `我重复抓取同一个目标 ${state.completedFailures} 次都失败，已停止。目标：${truncatedURL}`
+        if (truncatedError)
+          return `我重复抓取同一个目标 ${state.completedFailures} 次都失败，已停止。错误：${truncatedError}`
+        return `我重复抓取同一个目标 ${state.completedFailures} 次都失败，已停止。`
+      }
+      if (truncatedURL && truncatedError)
+        return `I failed to fetch the same target ${state.completedFailures} times and stopped. Target: ${truncatedURL} Error: ${truncatedError}`
+      if (truncatedURL)
+        return `I failed to fetch the same target ${state.completedFailures} times and stopped. Target: ${truncatedURL}`
+      if (truncatedError)
+        return `I failed to fetch the same target ${state.completedFailures} times and stopped. Error: ${truncatedError}`
+      return `I failed to fetch the same target ${state.completedFailures} times and stopped.`
+    }
+
+    if (state.kind === "target") {
+      if (isZh) {
+        return truncatedError
+          ? `我重复在同一个目标上失败了 ${state.completedFailures} 次，已停止。工具：${tool} 错误：${truncatedError}`
+          : `我重复在同一个目标上失败了 ${state.completedFailures} 次，已停止。工具：${tool}`
+      }
+      return truncatedError
+        ? `I failed against the same target ${state.completedFailures} times and stopped. Tool: ${tool} Error: ${truncatedError}`
+        : `I failed against the same target ${state.completedFailures} times and stopped. Tool: ${tool}`
+    }
+
+    if (isZh) {
+      return truncatedError
+        ? `我重复调用了 ${tool} ${state.completedFailures} 次都失败，已停止。最近一次错误：${truncatedError}`
+        : `我重复调用了 ${tool} ${state.completedFailures} 次都失败，已停止。`
+    }
+    return truncatedError
+      ? `I called ${tool} ${state.completedFailures} times without success and stopped. Last error: ${truncatedError}`
+      : `I called ${tool} ${state.completedFailures} times without success and stopped.`
+  }
+
+  function extractURL(value: unknown): string | undefined {
+    if (typeof value === "string") return value
+    if (!value || typeof value !== "object") return undefined
+    const r = value as Record<string, unknown>
+    if (typeof r.url === "string") return r.url
+    if (typeof r.href === "string") return r.href
+    return undefined
+  }
+
+  function stripQueryAndFragment(url: string): string {
+    try {
+      const u = new URL(url)
+      return `${u.protocol}//${u.host}${u.pathname}`
+    } catch {
+      const q = url.indexOf("?")
+      const f = url.indexOf("#")
+      const cuts = [q, f].filter((i) => i >= 0)
+      if (!cuts.length) return url
+      return url.slice(0, Math.min(...cuts))
+    }
+  }
+
+  // Strip query strings and fragments from any URLs embedded in free text (error messages,
+  // stack traces, etc.). Tokens often live in `?token=...` or `#access_token=...`. The `i`
+  // flag covers uppercase scheme variants like `HTTPS://`.
+  function scrubURLsInText(text: string): string {
+    return text.replace(/(https?:\/\/[^\s,;)]+)/gi, (match) => stripQueryAndFragment(match))
+  }
+
+  // Multi-pass scrub for tool error text. Tool errors echo URLs with tokens, fully-qualified
+  // file paths, quoted user input, and Bearer/Basic auth headers. Mask each class so the
+  // synthetic stop summary does not become a leak vector. Path detection uses a negative
+  // lookbehind so `Error:at /tmp/x` and `failed at '/home/alice/key'` get scrubbed too — not
+  // only paths preceded by start-of-line or whitespace. Path char class allows spaces
+  // (real paths: `/Users/alice/My Documents/...`) and stops at structural delimiters
+  // (quote, comma, semicolon, colon, paren/bracket, newline) — over-scrub of trailing
+  // descriptive text is preferred to under-scrub leaving secrets visible.
+  function scrubErrorText(text: string): string {
+    return scrubURLsInText(text)
+      .replace(/(?:Bearer|Basic|token|api[_-]?key)[\s:=]+[^\s'"`,;)]+/gi, "<token>")
+      .replace(/['"`][^'"`]*['"`]/g, "<quoted>")
+      .replace(/\b[A-Za-z]:\\[^'"`,;:)\]\n]+/g, "<path>")
+      // Forward-slash Windows paths (`C:/Users/...`) — JS path normalization on Windows can
+      // produce these, and the general path regex below excludes `:` as preceding char to
+      // avoid stripping URL paths, so drive-letter forward-slash paths need their own pass.
+      .replace(/\b[A-Za-z]:\/[^'"`,;:)\]\n]+/g, "<path>")
+      // Relative paths starting with `./` or `../`. Without this pass the absolute-path
+      // regex below only catches the `/...` suffix and leaves the leading dots visible.
+      .replace(/(?<![\w.])\.\.?\/[^'"`,;:)\]\n]+/g, "<path>")
+      .replace(/(?<![\w/:])\/[^'"`,;:)\]\n]+/g, "<path>")
+  }
+
+  function firstErrorLine(error: unknown): string | undefined {
+    const line = SessionDiagnostics.firstLine(error)
+    return line === "" ? undefined : line
+  }
+}