fix(session-title): enforce framed input and deadline

2026-07-21 15:00:15 +08:00
parent d9557fa5e0
commit 9b048354b8
7 changed files with 61 additions and 16 deletions
@@ -2,7 +2,7 @@

 Optional `ctx.sessionTitle` provider that summarizes every eligible human message through `ctx.llm`. It registers the `all-user-messages` cadence and starts a new revision after each new human prompt, using seeded history as well as child-session prompts. A newer revision aborts and supersedes older work; even a provider that ignores cancellation cannot commit stale output.

-The plugin uses the complete required [shared LLM configuration](../session-title-llm/README.md#configuration). Omit both `provider` and `model` to inherit the exact route from each current logged main request, or set both to route title generation independently. If aggregate input exceeds `maxInputBytes`, the request fails instead of truncating history; automatic use warns and keeps the prior title.
+The plugin uses the complete required [shared LLM configuration](../session-title-llm/README.md#configuration). Omit both `provider` and `model` to inherit the exact route from each current logged main request, or set both to route title generation independently. If the final framed aggregate prompt exceeds `maxInputBytes`, the request fails instead of truncating history; automatic use warns and keeps the prior title.

 ## Model Experience

@@ -6,7 +6,7 @@ This package is a library, not a Cordis plugin. The provider plugins call `regis

 ## Route and failure contract

-`provider` and `model` overrides are optional but must be supplied together as non-empty strings. Without that pair, the helper uses the exact provider/model route captured from the current session's logged `request/header`; an explicit refresh before any route exists therefore needs overrides. Input exceeding `maxInputBytes` rejects instead of being truncated. Timeout, cancellation, malformed or empty output, tool calls, and non-stop finish reasons also reject; the session-title service decides whether that rejection is an automatic warning or an explicit caller failure.
+`provider` and `model` overrides are optional but must be supplied together as non-empty strings. Without that pair, the helper uses the exact provider/model route captured from the current session's logged `request/header`; an explicit refresh before any route exists therefore needs overrides. The helper measures the final JSON-framed user prompt, including seq fields, wrappers, and JSON escaping, against `maxInputBytes` before logging or dispatch instead of truncating it. Timeout and caller cancellation are rechecked while consuming the stream and after it completes, so a late successful result cannot be accepted even if an interceptor or adapter ignores abort. Malformed or empty output, tool calls, and non-stop finish reasons also reject; the session-title service decides whether that rejection is an automatic warning or an explicit caller failure.

 After route and input validation, the helper appends a log-only `session/title-llm-request` event before model dispatch. It contains the title-provider id, exact source seqs, route, system prompt, message list, and output-token cap used by the call. The append shares the title capability's per-session settlement queue, so a superseding request cannot collide with an earlier fallback, request record, or accepted-title flush. The dispatched envelope is deep-frozen to keep interceptors aligned with that record but deliberately lacks dsh-agent-loop's process-local request identity, so loop-only reconstruction observers do not compare it with the conversation header. A later model failure leaves that request record intact; validation failures that never become dispatchable requests do not create one. The event stays outside derived model history.

@@ -18,7 +18,7 @@ Every field is required except the paired route override; there are no library d
 |---|---|
 | `targetWords` | Positive target word count for non-CJK titles. |
 | `targetCjkCharacters` | Positive target character count for Chinese, Japanese, or Korean titles. |
-| `maxInputBytes` | Positive aggregate UTF-8 byte ceiling across selected messages. |
+| `maxInputBytes` | Positive UTF-8 byte ceiling for the final JSON-framed user prompt. |
 | `maxOutputTokens` | Positive auxiliary generation token cap. |
 | `timeoutMs` | Positive end-to-end deadline within the runtime timer limit. |
 | `provider`, `model` | Optional explicit route; both or neither. |
@@ -42,4 +42,4 @@ No main-request invalidation. Auxiliary cache reuse is provider-specific; the fi
 ## Known Limitations and Deferred Work

 - The helper accepts text output only and rejects tool calls; structured-output adapters and provider-specific prompt variants are not exposed.
- It enforces a byte ceiling for the whole selected input rather than clipping individual messages or applying a retention policy.
+- It enforces a byte ceiling for the whole framed user prompt rather than clipping individual messages or applying a retention policy.
@@ -58,7 +58,7 @@ export interface SessionTitleLlmConfig {
  readonly targetWords: number
  /** Target character count for Chinese, Japanese, or Korean titles. */
  readonly targetCjkCharacters: number
-  /** Maximum total UTF-8 bytes across selected source-message text. */
+  /** Maximum UTF-8 bytes in the final JSON-framed user prompt. */
  readonly maxInputBytes: number
  /** Auxiliary generation output-token cap. */
  readonly maxOutputTokens: number
@@ -242,14 +242,15 @@ export async function generateSessionTitleWithLlm(
  if (selectedMessages.length === 0) {
    throw new Error('session-title-llm: at least one source message is required')
  }
-  const inputBytes = selectedMessages.reduce((total, message) => total + Buffer.byteLength(message.text, 'utf8'), 0)
+  const framedInput = frameMessages(selectedMessages)
+  const inputBytes = Buffer.byteLength(framedInput, 'utf8')
  if (inputBytes > config.maxInputBytes) {
    throw new Error(`session-title-llm: input is ${inputBytes} bytes, exceeding maxInputBytes ${config.maxInputBytes}`)
  }
  const route = resolveRoute(config, request)
  const messages: Message[] = [{
    role: 'user',
-    content: [{ type: 'text', text: frameMessages(selectedMessages) }],
+    content: [{ type: 'text', text: framedInput }],
  }]
  const system = systemPrompt(config)
  using callDeadline = deadline(request.signal, config.timeoutMs, SESSION_TITLE_TIMEOUT_CODE)
@@ -272,7 +273,11 @@ export async function generateSessionTitleWithLlm(
  }, callDeadline.signal)
  callDeadline.signal.throwIfAborted()
  const assembler = new BlockAssembler()
-  for await (const chunk of ctx.llm.stream(options)) assembler.push(chunk)
+  for await (const chunk of ctx.llm.stream(options)) {
+    callDeadline.signal.throwIfAborted()
+    assembler.push(chunk)
+  }
+  callDeadline.signal.throwIfAborted()
  const terminalError = finishError(assembler.finish)
  if (terminalError !== undefined) throw terminalError
  const blocks = assembler.message().content
@@ -48,6 +48,17 @@ class CooperativeAdapter extends LlmAdapter {
  }
 }

+class DelayedSuccessAdapter extends LlmAdapter {
+  constructor(private readonly delayMs: number) {
+    super()
+  }
+
+  override async * stream(): AsyncIterable<StreamChunk> {
+    await new Promise<void>(resolve => setTimeout(resolve, this.delayMs))
+    yield * SCRIPT
+  }
+}
+
 const SCRIPT: StreamChunk[] = [
  { type: 'block-start', index: 0, blockType: 'text' },
  { type: 'text-delta', index: 0, text: '  五个字标题  ' },
@@ -162,21 +173,24 @@ describe('generateSessionTitleWithLlm', () => {
      })
  })

-  it('uses paired explicit overrides and rejects an oversized input without calling the model', async () => {
+  it('uses paired explicit overrides and bounds the final framed input before model dispatch', async () => {
    const ctx = new Context()
    await ctx.plugin(SessionStore)
    await ctx.plugin(LlmService)
    const adapter = new RecordingAdapter(SCRIPT)
    ctx.llm.registerAdapter(['explicit-route'], adapter)
+    const oversized = request(ctx)
+    const [selected] = oversized.messages
+    if (selected === undefined) throw new Error('expected one selected message')
+    const rawInputBytes = Buffer.byteLength(selected.text, 'utf8')
    const config = resolveSessionTitleLlmConfig({
      ...CONFIG,
      provider: 'explicit-route',
      model: 'explicit-model',
-      maxInputBytes: 4,
+      maxInputBytes: rawInputBytes,
    })

-    const oversized = request(ctx)
-    await expect(generateSessionTitleWithLlm(ctx, config, oversized, oversized.messages, TITLE_PROVIDER))
+    await expect(generateSessionTitleWithLlm(ctx, config, oversized, [selected], TITLE_PROVIDER))
      .rejects.toThrow(/input.*bytes.*maxInputBytes/i)
    expect(adapter.requests).toEqual([])
    expect(oversized.session.events.some(event => event.type === 'session/title-llm-request')).toBe(false)
@@ -322,4 +336,30 @@ describe('generateSessionTitleWithLlm', () => {
      vi.useRealTimers()
    }
  })
+
+  it('rejects a successful stream that completes after the configured deadline', async () => {
+    vi.useFakeTimers()
+    try {
+      const ctx = new Context()
+      await ctx.plugin(SessionStore)
+      await ctx.plugin(LlmService)
+      ctx.llm.registerAdapter(['current-route'], new DelayedSuccessAdapter(20))
+      const providerRequest = request(ctx)
+      const pending = generateSessionTitleWithLlm(
+        ctx,
+        resolveSessionTitleLlmConfig({ ...CONFIG, timeoutMs: 10 }),
+        providerRequest,
+        providerRequest.messages,
+        TITLE_PROVIDER,
+      )
+      const rejected = expect(pending).rejects.toMatchObject({
+        code: SESSION_TITLE_TIMEOUT_CODE,
+        timeoutMs: 10,
+      })
+      await vi.advanceTimersByTimeAsync(20)
+      await rejected
+    } finally {
+      vi.useRealTimers()
+    }
+  })
 })