feat: preserve safe provider termination metadata

EVENTS.md

@@ -302,6 +302,67 @@ Emitted at step boundaries during multi-step tool use.
 { "type": "step_finish", "part": { "type": "step-finish" } }
 ```
 
+#### Provider termination details
+
+`step_finish.part.termination` is an optional additive diagnostic object. Its
+`normalizedReason` mirrors `part.reason`; `providerID` and `modelID` identify the
+configured provider/model. Correlate it with the enclosing `invocationID` and
+`sessionID`, and `part.messageID`. Older stored parts may omit `termination`.
+This diagnostic does not decide whether an invocation succeeded (see #108).
+
+```json
+{
+  "type": "step_finish",
+  "invocationID": "inv_fixture",
+  "sessionID": "ses_fixture",
+  "part": {
+    "type": "step-finish",
+    "messageID": "msg_fixture",
+    "reason": "error",
+    "termination": {
+      "providerID": "google-vertex",
+      "modelID": "gemini-2.5-flash",
+      "normalizedReason": "error",
+      "rawReason": { "status": "available", "value": "MALFORMED_FUNCTION_CALL", "truncated": false },
+      "requestID": { "status": "unavailable", "truncated": false },
+      "diagnostic": { "status": "redacted", "truncated": false }
+    }
+  }
+}
+```
+
+The example omits unrelated step fields. Each diagnostic field reports
+`available` (a permitted observed value), `unavailable` (missing or unsupported),
+or `redacted` (present but suppressed). Values are absent when unavailable or
+redacted. `truncated: true` means an oversized value was entirely suppressed;
+no prefix is retained. Raw reasons and request IDs are limited to 128 characters.
+Free-form provider diagnostics are always suppressed, with a 2,048-character
+threshold for the oversize flag: they can contain credentials, prompt text, or
+tool arguments that pattern-based redaction cannot reliably remove.
+
+Raw reasons are currently captured only for native `@ai-sdk/google` and
+`@ai-sdk/google-vertex` streams. Their pinned adapters support internal raw
+chunks; middleware selects the first candidate's allowlisted `finishReason`,
+records `finishMessage` presence/size, and drops every raw chunk before it reaches
+AI SDK stream consumers. Unknown raw reason strings are redacted. Other adapters
+still provide their normalized reason and explicitly report raw details as
+unavailable. No raw reason is inferred from the normalized reason.
+
+Request-ID availability uses only `x-request-id` or `x-goog-request-id` response
+headers. All nonempty values are reported as `redacted` without retaining the
+value, including ordinary IDs: arbitrary IDs cannot be reliably distinguished
+from credential material by format. Values over 128 characters also set
+`truncated: true`. The SDK's generated response ID is never presented as a
+provider request ID. No new prompts,
+reasoning, tool arguments, full responses, or response-header maps are collected
+by this diagnostic path. Existing exception diagnostics are unchanged.
+
+For #109, this delivers observed raw finish reasons and privacy-preserving
+diagnostic availability. Richer diagnostic text remains a separate policy and
+adapter-coverage decision; a redacted diagnostic cannot identify the exact
+offending tool call. Deliver after #108 so a provider error reason is paired
+with truthful process/session failure status.
+
 ## Skill Events
 
 Skills are loaded progressively. The model first sees skill names and descriptions in the tool schema. Full skill content only enters context when the model explicitly invokes the skill tool.

packages/cli/src/cli/cmd/generate.ts

@@ -1,15 +1,30 @@
 import type { CommandModule } from "yargs"
+import z from "zod"
+import { MessageV2 } from "@/session/message-v2"
+import { ProviderTermination } from "@/provider/termination"
 
 export const GenerateCommand = {
   command: "generate",
   handler: async () => {
+    // Keep OpenAPI component identities separate from legacy Zod metadata.
+    const registry = z.registry<{ id: string }>()
+    registry.add(MessageV2.StepFinishPart, { id: "StepFinishPart" })
+    registry.add(ProviderTermination.Info, { id: "ProviderTermination" })
+    const components = z.toJSONSchema(registry, {
+      metadata: z.registry(),
+      target: "draft-2020-12",
+      uri: (id) => `#/components/schemas/${id}`,
+    })
+    // Component JSON pointers are references, not standalone schema resource IDs.
+    for (const schema of Object.values(components.schemas)) delete schema.$id
     const specs = {
       openapi: "3.1.1",
       info: {
         title: "aictrl",
         version: "1.0.0",
       },
       paths: {},
+      components,
     }
     const json = JSON.stringify(specs, null, 2)
 

packages/cli/src/provider/termination.ts

@@ -0,0 +1,123 @@
+import type { LanguageModelV2Middleware } from "@ai-sdk/provider"
+import z from "zod"
+
+export namespace ProviderTermination {
+  const VALUE_LIMIT = 128
+  const DIAGNOSTIC_LIMIT = 2048
+  function Field(limit: number) {
+    return z.object({
+      status: z.enum(["available", "unavailable", "redacted"]),
+      value: z.string().max(limit).optional(),
+      truncated: z.boolean(),
+    })
+  }
+  type Field = z.infer<ReturnType<typeof Field>>
+
+  export const Info = z
+    .object({
+      providerID: z.string(),
+      modelID: z.string(),
+      normalizedReason: z.string(),
+      rawReason: Field(VALUE_LIMIT),
+      requestID: Field(VALUE_LIMIT),
+      diagnostic: Field(DIAGNOSTIC_LIMIT),
+    })
+    .meta({ ref: "ProviderTermination" })
+  export type Info = z.infer<typeof Info>
+
+  const reasons = new Set([
+    "FINISH_REASON_UNSPECIFIED",
+    "STOP",
+    "MAX_TOKENS",
+    "SAFETY",
+    "RECITATION",
+    "LANGUAGE",
+    "OTHER",
+    "BLOCKLIST",
+    "PROHIBITED_CONTENT",
+    "SPII",
+    "MALFORMED_FUNCTION_CALL",
+    "IMAGE_SAFETY",
+    "UNEXPECTED_TOOL_CALL",
+    "NO_IMAGE",
+    "IMAGE_PROHIBITED_CONTENT",
+    "IMAGE_OTHER",
+    "IMAGE_RECITATION",
+  ])
+
+  function field(value: unknown, allowed: (value: string) => boolean, limit = VALUE_LIMIT): Field {
+    if (typeof value !== "string" || !value) return { status: "unavailable", truncated: false }
+    if (value.length > limit || !allowed(value)) return { status: "redacted", truncated: value.length > limit }
+    return { status: "available", value, truncated: false }
+  }
+
+  function record(value: unknown): Record<string, unknown> | undefined {
+    if (typeof value !== "object" || value === null || Array.isArray(value)) return
+    return value as Record<string, unknown>
+  }
+
+  export function middleware(input: { providerID: string; modelID: string; npm: string }): LanguageModelV2Middleware {
+    const google = input.npm === "@ai-sdk/google" || input.npm === "@ai-sdk/google-vertex"
+    return {
+      async transformParams({ type, params }) {
+        return type === "stream" && google ? { ...params, includeRawChunks: true } : params
+      },
+      async wrapStream({ doStream }) {
+        const result = await doStream()
+        let rawReason = field(undefined, () => false)
+        let diagnostic = field(undefined, () => false)
+        const headers = result.response?.headers
+        // A provider/proxy can echo arbitrary credentials into an ID header.
+        // Format allowlists cannot distinguish an opaque ID from an opaque key.
+        const requestID = field(headers?.["x-request-id"] || headers?.["x-goog-request-id"], () => false)
+        return {
+          ...result,
+          stream: result.stream.pipeThrough(
+            new TransformStream({
+              transform(chunk, controller) {
+                if (chunk.type === "raw") {
+                  if (google) {
+                    const candidates = record(chunk.rawValue)?.candidates
+                    const candidate = Array.isArray(candidates) ? record(candidates[0]) : undefined
+                    if (candidate?.finishReason != null) {
+                      rawReason = field(candidate.finishReason, (value) => reasons.has(value))
+                      // Provider messages can contain prompts or tool arguments. Retain only
+                      // presence/size information; do not collect their free-form contents.
+                      diagnostic = field(candidate.finishMessage, () => false, DIAGNOSTIC_LIMIT)
+                    }
+                  }
+                  // Raw payloads must never escape this boundary to stream consumers.
+                  return
+                }
+                if (chunk.type !== "finish") {
+                  controller.enqueue(chunk)
+                  return
+                }
+                const termination: Info = {
+                  providerID: input.providerID,
+                  modelID: input.modelID,
+                  normalizedReason: chunk.finishReason,
+                  rawReason,
+                  requestID,
+                  diagnostic,
+                }
+                controller.enqueue({
+                  ...chunk,
+                  providerMetadata: {
+                    ...chunk.providerMetadata,
+                    aictrl: { ...chunk.providerMetadata?.aictrl, termination },
+                  },
+                })
+              },
+            }),
+          ),
+        }
+      },
+    }
+  }
+
+  export function from(metadata: unknown): Info | undefined {
+    const result = Info.safeParse(record(record(metadata)?.aictrl)?.termination)
+    return result.success ? result.data : undefined
+  }
+}

packages/cli/src/session/llm.ts

@@ -22,6 +22,7 @@ import { SystemPrompt } from "./system"
 import { Flag } from "@/flag/flag"
 import { PermissionNext } from "@/permission/next"
 import { Auth } from "@/auth"
+import { ProviderTermination } from "@/provider/termination"
 
 export namespace LLM {
   const log = Log.create({ service: "llm" })
@@ -234,6 +235,11 @@ export namespace LLM {
       model: wrapLanguageModel({
         model: language,
         middleware: [
+          ProviderTermination.middleware({
+            providerID: input.model.providerID,
+            modelID: input.model.id,
+            npm: input.model.api.npm,
+          }),
           {
             async transformParams(args) {
               if (args.type === "stream") {

packages/cli/src/session/message-v2.ts

@@ -9,6 +9,7 @@ import { fn } from "@/util/fn"
 import { Database, eq, desc, inArray } from "@/storage/db"
 import { MessageTable, PartTable } from "./session.sql"
 import { ProviderTransform } from "@/provider/transform"
+import { ProviderTermination } from "@/provider/termination"
 import { STATUS_CODES } from "http"
 import { Storage } from "@/storage/storage"
 import { ProviderError } from "@/provider/error"
@@ -241,6 +242,7 @@ export namespace MessageV2 {
   export const StepFinishPart = PartBase.extend({
     type: z.literal("step-finish"),
     reason: z.string(),
+    termination: ProviderTermination.Info.optional(),
     snapshot: z.string().optional(),
     cost: z.number(),
     tokens: z.object({

packages/cli/src/session/processor.ts

@@ -16,6 +16,7 @@ import { SessionCompaction } from "./compaction"
 import { PermissionNext } from "@/permission/next"
 import { Question } from "@/question"
 import { NamedError } from "@aictrl/util/error"
+import { ProviderTermination } from "@/provider/termination"
 
 export namespace SessionProcessor {
   const DOOM_LOOP_THRESHOLD = 3
@@ -255,6 +256,7 @@ export namespace SessionProcessor {
                   await Session.updatePart({
                     id: Identifier.ascending("part"),
                     reason: value.finishReason,
+                    termination: ProviderTermination.from(value.providerMetadata),
                     snapshot: await Snapshot.track(),
                     messageID: input.assistantMessage.id,
                     sessionID: input.assistantMessage.sessionID,

packages/cli/test/cli/generate.test.ts

@@ -0,0 +1,37 @@
+import { expect, test } from "bun:test"
+import path from "path"
+
+test("generate publishes the termination schema used by SDK codegen", async () => {
+  const proc = Bun.spawn(
+    ["bun", "run", "--conditions=browser", path.resolve(import.meta.dir, "../../src/index.ts"), "generate"],
+    {
+      stdout: "pipe",
+      stderr: "pipe",
+    },
+  )
+  const [stdout, stderr, exit] = await Promise.all([
+    new Response(proc.stdout).text(),
+    new Response(proc.stderr).text(),
+    proc.exited,
+  ])
+  expect(exit, stderr).toBe(0)
+  const spec = JSON.parse(stdout)
+  const part = spec.components.schemas.StepFinishPart
+  const diagnostic = spec.components.schemas.ProviderTermination
+  expect(spec.openapi).toBe("3.1.1")
+  expect(part.properties.termination).toEqual({ $ref: "#/components/schemas/ProviderTermination" })
+  expect(part.$schema).toBe("https://json-schema.org/draft/2020-12/schema")
+  expect(diagnostic.$schema).toBe("https://json-schema.org/draft/2020-12/schema")
+  expect(part.id).toBeUndefined()
+  expect(diagnostic.id).toBeUndefined()
+  expect(part.$id).toBeUndefined()
+  expect(diagnostic.$id).toBeUndefined()
+  expect(part.properties.type.const).toBe("step-finish")
+  expect(stdout).not.toMatch(/"ref"\s*:/)
+  expect(stdout).not.toMatch(/"nullable"\s*:/)
+  expect(part.required).not.toContain("termination")
+  expect(diagnostic.required).toContain("normalizedReason")
+  expect(diagnostic.properties.rawReason.properties.value.maxLength).toBe(128)
+  expect(diagnostic.properties.diagnostic.properties.value.maxLength).toBe(2048)
+  expect(diagnostic.properties.requestID.properties.status.enum).toEqual(["available", "unavailable", "redacted"])
+})