Merge remote-tracking branch 'origin/master' into xtr/agent-loop-message-machine

packages/support/acp-snapshot/src/suite.ts

@@ -70,12 +70,13 @@ export interface Scenario {
   recorded: boolean
   /**
    * Whether replay is driven by a hand-written `replay.override.json` sidecar
-   * (a `ReplayEntry[]` that REPLACES the script derived from `session.jsonl`)
-   * — the throw/hang cases chunks cannot express. The fixture guard requires
-   * the sidecar exactly when this is set: the harness forwards the file purely
-   * on existence, so an unregistered stray sidecar would silently replace the
-   * derived script — the guard fails loud on either mismatch. Defaults to
-   * false (replay derives from the fixture's `assistant/chunk` events).
+   * (a `ReplayOverrideDoc` that replaces or patches the script derived from
+   * `session.jsonl`) — the throw/hang cases chunks cannot express. The fixture
+   * guard requires the sidecar exactly when this is set: the harness forwards
+   * the file purely on existence, so an unregistered stray sidecar would
+   * silently alter the derived script. The guard fails loud on either
+   * mismatch. Defaults to false (replay derives from the fixture's
+   * `assistant/chunk` events).
    */
   overridden?: boolean
   /**

packages/support/llm-replay/README.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write
-README.md: 901a3b7b4312fffd93e6d375c378e39064318260
-README.zh.md: b9a8068d329e28933c934e7ad352ac65641f3d23
+README.md: ce0758641f3d49a54b29415ed449e43043840f9a
+README.zh.md: 47a2b9aa211b44c4e476a1adf5a9a72d927cd0ed

packages/support/llm-replay/README.md

@@ -10,7 +10,7 @@ Its consumers are the ACP, headless `stream-json`, and TUI snapshot suites plus
 
 The fixture IS the persisted session log (`<scenario>/session.jsonl`). Its `assistant/chunk` events carry every `StreamChunk`, so grouping them by `(turn, step)` reconstructs each `stream()` call's chunk sequence (one model call per loop step). Recording is therefore "run the real agent once and harvest the `.jsonl`", done by the snapshot harness — this plugin does not record. A fixture may carry its `request/header` content tokenized to `{{system}}`/`{{tools}}` (the harness pins that content in one scenario and scrubs the rest); replay is indifferent — derivation reads only `assistant/chunk` events and the line-0 session header.
 
-Two failure modes are not reconstructable from `assistant/chunk` alone — a pure throw before any chunk (e.g. an HTTP 401, where the log holds only a `turn/end {error}` and no chunks) and a cancel/hang (timing, not chunk content). A scenario that needs those supplies an optional sidecar (`<scenario>/replay.override.json`: a `ReplayEntry[]`) that REPLACES the derived script. A `hang` entry may name `readyFile`; replay writes that empty marker after its prefix chunks reach the loop and before it waits for cancellation, so an external driver can cancel deterministically without observing a presentation update.
+Two failure modes are not reconstructable from `assistant/chunk` alone — a pure throw before any chunk (e.g. an HTTP 401, where the log holds only a `turn/end {error}` and no chunks) and a cancel/hang (timing, not chunk content). A scenario that needs those supplies an optional sidecar (`<scenario>/replay.override.json`) that either replaces the derived script (a bare `ReplayEntry[]`) or augments it (`{ patches: [{ at, entry }] }`: keep every JSONL-derived call and swap the named 0-based call indexes; `at` equal to the derived length appends the retry attempt after an injected transient throw). Patch indexes must be unique. The override document, each patch and entry, and every chunk discriminant are validated when the file loads. A `hang` entry may name `readyFile`; replay writes that empty marker after its prefix chunks reach the loop and before it waits for cancellation, so an external driver can cancel deterministically without observing a presentation update.
 
 ## Nested agents: per-session keying
 
@@ -23,7 +23,7 @@ Replay keys every call by its calling session id (`GenerateOptions.sessionId`, s
 | Key | Type | Default | Notes |
 |---|---|---|---|
 | `file` | string | `$DSH_SNAPSHOT_FILE` | Path to the primary (parent) `session.jsonl` fixture. Required (config or env). |
-| `overrideFile` | string | `$DSH_SNAPSHOT_OVERRIDE` | Optional path to a `ReplayEntry[]` sidecar that replaces the PRIMARY session's derived script. |
+| `overrideFile` | string | `$DSH_SNAPSHOT_OVERRIDE` | Optional `ReplayOverrideDoc` sidecar for the primary session: a bare `ReplayEntry[]` replaces its derived script, while `{ patches }` augments it by call index. |
 | `childFiles` | string[] | `$DSH_SNAPSHOT_CHILD_FILES` (path-delimited) | Recorded subagent child-session logs for a nested scenario; empty for a single-session scenario. |
 | `providers` | `ReplayProviderConfig[]` | — | Optional replay-only provider and model catalog. Each model may publish `contextWindow`; configured routes dispatch through the replay adapter and never perform provider I/O. |
 | `paceMs` | number | — (burst) | Optional per-chunk delay in ms so downstream transports (e.g. the web SSE mux observed by a real browser) see genuinely incremental delivery. A realism knob only — tests must not depend on it for correctness. Non-negative integer; abort during a pace wait cancels the stream promptly. |
@@ -48,9 +48,9 @@ Replay keys every call by its calling session id (`GenerateOptions.sessionId`, s
 
 - `installLlmReplay(ctx, config)` — install the configured replay adapter or catch-all `llm/stream` listener; returns a `ReplayHandle` (`dispose()` for HMR safety plus `assertConsumed()`, the teardown check that every recorded script bound to a live session and every bound cursor drained — turning a scenario that silently drove fewer model calls than recorded into a crisp diagnostic). Use this in tests to drive replay without the Loader or env vars.
 - `loadSessionScripts(config)` — resolve the ordered `SessionScript[]` (primary + children) for a scenario, ready to bind to live sessions in first-call order.
-- `loadReplayScript(config)` — resolve the `ReplayEntry[]` for the PRIMARY session only (sidecar override if present, else derived from the JSONL; fail-loud if the fixture is missing).
+- `loadReplayScript(config)` — resolve the `ReplayEntry[]` for the primary session only (validated sidecar replacement/patches if present, else derived from the JSONL; fail-loud if the fixture is missing).
 - `deriveReplayScript(events)` / `parseSessionLog(text)` / `parseSessionHeader(text)` — the pure helpers that turn a recorded session log into a script and read its header `id`/`createdAt`. A derived group must end in a `finish` chunk; a group without one is the fingerprint of a thrown `stream()` and must instead be expressed via an override sidecar.
-- Types `ReplayEntry` / `SessionScript` / `ReplayConfig` / `ReplayProviderConfig` / `ReplayModelConfig` / `ReplayHandle` / `Config`.
+- Types `ReplayEntry` / `ReplayOverrideDoc` / `ReplayOverridePatch` / `SessionScript` / `ReplayConfig` / `ReplayProviderConfig` / `ReplayModelConfig` / `ReplayHandle` / `Config`.
 
 ## Plugin export shape
 
@@ -67,4 +67,4 @@ None; this package neither assembles nor sends a provider request.
 ## Known Limitations and Deferred Work
 
 - **First-call-order script binding assumes sequential delegation** — a cut that runs sibling subagents concurrently (or a compaction summarize call landing mid-run) would bind live sessions to recorded scripts non-deterministically; a stronger keying is deferred until such a scenario exists (`XXX(concurrent-subagents)`).
-- **Only chunk-producing calls are derivable** — a pure pre-chunk throw or a cancel/hang scenario needs the `replay.override.json` sidecar; the override replaces the PRIMARY session's script only.
+- **Only chunk-producing calls are derivable** — a pure pre-chunk throw or a cancel/hang scenario needs the `replay.override.json` sidecar. Replacement and patch forms affect only the primary session; child scripts still derive from their logs.

packages/support/llm-replay/README.zh.md

@@ -10,7 +10,7 @@
 
 Fixture 就是持久化会话日志（`<scenario>/session.jsonl`）。其 `assistant/chunk` 事件携带每个 `StreamChunk`，因此按 `(turn, step)` 对其分组可重建每次 `stream()` 调用的分片序列（每个 loop 步骤一次模型调用）。因此，录制操作是「运行一次真实 agent 并收集 `.jsonl`」，由快照 harness 完成；该插件不执行录制。Fixture 的 `request/header` 内容可能被 token 化为 `{{system}}`/`{{tools}}`（harness 在一个场景中固定该内容，并擦除其余场景）；回放对此并不关心，因为派生只读取 `assistant/chunk` 事件和第 0 行会话 header。
 
-有两种失败 mode 无法仅从 `assistant/chunk` 重建：在任何分片前纯抛出（例如 HTTP 401，日志只包含 `turn/end {error}` 而没有分片），以及 cancel/hang（是时序，而非分片内容）。需要这些的场景提供可选 sidecar（`<scenario>/replay.override.json`：一个 `ReplayEntry[]`），以替换派生脚本。`hang` 条目可以指定 `readyFile`；在其前缀分片到达 loop 后、等待取消前，回放会写入该空标记，使外部驱动器可以在不观察展示更新的情况下确定性取消。
+有两种失败 mode 无法仅从 `assistant/chunk` 重建：在任何分片前纯抛出（例如 HTTP 401，日志只包含 `turn/end {error}` 而没有分片），以及 cancel/hang（是时序，而非分片内容）。需要这些的场景提供可选 sidecar（`<scenario>/replay.override.json`），它要么替换派生脚本（裸 `ReplayEntry[]`），要么增补派生脚本（`{ patches: [{ at, entry }] }`：保留全部由 JSONL 派生的调用，仅在点名的调用索引处换入，索引从 0 计；`at` 等于派生长度时为追加，正是注入的瞬态抛出之后那次重试尝试所占的槽位）。Patch 索引必须互不重复。覆写文档、每个 patch 与每个条目，以及每个分片的判别字段都会在文件加载时接受校验。`hang` 条目可以指定 `readyFile`；在其前缀分片到达 loop 后、等待取消前，回放会写入该空标记，使外部驱动器可以在不观察展示更新的情况下确定性取消。
 
 ## 嵌套 agent：每会话键控
 
@@ -23,7 +23,7 @@ Fixture 就是持久化会话日志（`<scenario>/session.jsonl`）。其 `assis
 | 键 | 类型 | 默认值 | 说明 |
 |---|---|---|---|
 | `file` | string | `$DSH_SNAPSHOT_FILE` | 主（父）`session.jsonl` fixture 的路径。必需（配置或 env）。 |
-| `overrideFile` | string | `$DSH_SNAPSHOT_OVERRIDE` | 替换主会话派生脚本的 `ReplayEntry[]` sidecar 可选路径。 |
+| `overrideFile` | string | `$DSH_SNAPSHOT_OVERRIDE` | 主会话的可选 `ReplayOverrideDoc` sidecar：裸 `ReplayEntry[]` 替换其派生脚本，`{ patches }` 则按调用索引增补该脚本。 |
 | `childFiles` | string[] | `$DSH_SNAPSHOT_CHILD_FILES` (path-delimited) | 嵌套场景中已记录的 subagent 子会话日志；单会话场景为空。 |
 | `providers` | `ReplayProviderConfig[]` | 无 | 可选的仅回放提供方和模型目录。每个模型可以发布 `contextWindow`；已配置路由通过回放适配器分派，绝不执行提供方 I/O。 |
 | `paceMs` | number | 无（突发） | 可选的每分片毫秒延迟，使下游传输（例如真实浏览器观察的 web SSE mux）看到真正的增量传递。它只是仿真开关，测试不得依赖它保证正确性。值必须是非负整数；pace 等待期间中止会迅速取消流。 |
@@ -48,9 +48,9 @@ Fixture 就是持久化会话日志（`<scenario>/session.jsonl`）。其 `assis
 
 - `installLlmReplay(ctx, config)`：安装已配置回放适配器或 catch-all `llm/stream` 监听器；返回 `ReplayHandle`（包含用于 HMR 安全的 `dispose()`，以及 `assertConsumed()` 拆卸检查；后者确保每个已记录脚本都绑定到实时会话，且每个已绑定游标都已耗尽，从而将场景静默驱动的模型调用少于记录数转换为明确诊断）。在测试中使用它，可以不通过 Loader 或 env var 驱动回放。
 - `loadSessionScripts(config)`：解析场景的有序 `SessionScript[]` （主级 + 子级），准备按首次调用顺序绑定到实时会话。
-- `loadReplayScript(config)`：只解析主会话的 `ReplayEntry[]` （如果存在则使用 sidecar override，否则从 JSONL 派生；fixture 缺失时快速失败）。
+- `loadReplayScript(config)`：只解析主会话的 `ReplayEntry[]` （如果存在则使用经校验的 sidecar 替换或 patch，否则从 JSONL 派生；fixture 缺失时快速失败）。
 - `deriveReplayScript(events)` / `parseSessionLog(text)` / `parseSessionHeader(text)`：将已记录会话日志转换为脚本并读取其 header `id`/`createdAt` 的纯辅助工具。派生分组必须以 `finish` 分片结束；没有该分片的分组是已抛出 `stream()` 的指纹，必须改用 override sidecar 表达。
-- 类型 `ReplayEntry` / `SessionScript` / `ReplayConfig` / `ReplayProviderConfig` / `ReplayModelConfig` / `ReplayHandle` / `Config`。
+- 类型 `ReplayEntry` / `ReplayOverrideDoc` / `ReplayOverridePatch` / `SessionScript` / `ReplayConfig` / `ReplayProviderConfig` / `ReplayModelConfig` / `ReplayHandle` / `Config`。
 
 ## 插件导出形态
 
@@ -67,4 +67,4 @@ Fixture 就是持久化会话日志（`<scenario>/session.jsonl`）。其 `assis
 ## 已知限制与待完成工作
 
 - **首次调用顺序脚本绑定假设串行委托**：并发运行同级 subagent 的 cut（或运行中落地的压缩摘要调用）会非确定性地将实时会话绑定到已记录脚本；在这种场景出现前暂不实现更强的键控（`XXX(concurrent-subagents)`）。
-- **只有生产分片的调用可派生**：纯分片前抛出或 cancel/hang 场景需要 `replay.override.json` sidecar；override 只替换主会话的脚本。
+- **只有生产分片的调用可派生**：纯分片前抛出或 cancel/hang 场景需要 `replay.override.json` sidecar。替换和 patch 两种形式都只影响主会话；子会话脚本仍从各自日志派生。

packages/support/llm-replay/src/index.ts

@@ -59,10 +59,11 @@ export interface ReplayConfig {
    */
   file: string
   /**
-   * Optional `ReplayEntry[]` sidecar that REPLACES the derived script for the
-   * PRIMARY session. Used by the two single-session scenarios not expressible as
-   * `assistant/chunk` (pure throw-before-chunk, cancel/hang). Absent for normal
-   * and nested scenarios.
+   * Optional sidecar for the PRIMARY session: a bare `ReplayEntry[]` replaces
+   * the derived script; `{ patches }` keeps it and swaps the named call
+   * indexes ({@link ReplayOverrideDoc}). Used by single-session scenarios not
+   * expressible as `assistant/chunk` (throw-before-chunk, cancel/hang,
+   * injected transient failures). Absent for normal and nested scenarios.
    */
   overrideFile?: string
   /**
@@ -200,26 +201,157 @@ export function deriveReplayScript(events: SessionEvent[]): ReplayEntry[] {
 }
 
 /**
- * Build the replay script for the PRIMARY session: the sidecar override if
- * present, otherwise the script derived from the recorded session JSONL.
- * Fail-loud if the JSONL fixture is missing (the scenario was never recorded) —
- * never silently returns an empty script, so a coverage hole can't masquerade
- * as a passing replay.
+ * One positional patch in an augmentation sidecar: replaces the derived
+ * entry at call index `at` (0-based) with `entry`, or appends when `at`
+ * equals the derived length (an extra recorded-after-the-fact call, e.g. the
+ * retry attempt following an injected transient throw).
+ */
+export interface ReplayOverridePatch {
+  /** 0-based call index into the derived script; == length appends. */
+  at: number
+  /** The replacement (or appended) entry at that call position. */
+  entry: ReplayEntry
+}
+
+/**
+ * Override sidecar document: either a whole-script replacement (a
+ * bare `ReplayEntry[]`) or the augmentation form `{ patches }`, which keeps
+ * the JSONL-derived script and swaps only the named call indexes — the shape
+ * for "turn N errors, everything else replays as recorded".
+ */
+export type ReplayOverrideDoc = ReplayEntry[] | { patches: ReplayOverridePatch[] }
+
+const REPLAY_CHUNK_TYPES = new Set<StreamChunk['type']>([
+  'block-start',
+  'text-delta',
+  'reasoning-delta',
+  'tool-call-delta',
+  'block-end',
+  'usage',
+  'finish',
+])
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value)
+}
+
+function hasExactKeys(value: Record<string, unknown>, keys: readonly string[]): boolean {
+  return Object.keys(value).length === keys.length && keys.every(key => Object.hasOwn(value, key))
+}
+
+function invalidOverride(file: string, location: string, detail: string): never {
+  throw new Error(`llm-replay: invalid override ${file}: ${location} ${detail}`)
+}
+
+function readChunks(value: unknown, file: string, location: string): StreamChunk[] {
+  if (!Array.isArray(value)) invalidOverride(file, location, 'chunks must be an array')
+  for (const [index, chunk] of value.entries()) {
+    if (!isRecord(chunk)
+      || typeof chunk['type'] !== 'string'
+      || !REPLAY_CHUNK_TYPES.has(chunk['type'] as StreamChunk['type'])) {
+      invalidOverride(file, `${location}.chunks[${index}]`, 'must have a known StreamChunk type')
+    }
+  }
+  return value as StreamChunk[]
+}
+
+function readReplayEntry(value: unknown, file: string, location: string): ReplayEntry {
+  if (!isRecord(value)) invalidOverride(file, location, 'must be an object')
+  switch (value['kind']) {
+    case 'chunks': {
+      if (!hasExactKeys(value, ['kind', 'chunks'])) invalidOverride(file, location, 'has invalid chunks-entry fields')
+      return { kind: 'chunks', chunks: readChunks(value['chunks'], file, location) }
+    }
+    case 'throw': {
+      if (!hasExactKeys(value, ['kind', 'chunks', 'message', 'code'])) {
+        invalidOverride(file, location, 'has invalid throw-entry fields')
+      }
+      if (typeof value['message'] !== 'string' || value['message'].length === 0) {
+        invalidOverride(file, location, 'message must be a non-empty string')
+      }
+      if (typeof value['code'] !== 'string' || value['code'].length === 0) {
+        invalidOverride(file, location, 'code must be a non-empty string')
+      }
+      return {
+        kind: 'throw',
+        chunks: readChunks(value['chunks'], file, location),
+        message: value['message'],
+        code: value['code'],
+      }
+    }
+    case 'hang': {
+      const readyFile = value['readyFile']
+      const keys = readyFile === undefined ? ['kind'] : ['kind', 'readyFile']
+      if (!hasExactKeys(value, keys)) invalidOverride(file, location, 'has invalid hang-entry fields')
+      if (readyFile !== undefined && (typeof readyFile !== 'string' || readyFile.length === 0)) {
+        invalidOverride(file, location, 'readyFile must be a non-empty string')
+      }
+      return { kind: 'hang', ...(readyFile === undefined ? {} : { readyFile }) }
+    }
+    default:
+      return invalidOverride(file, location, `has unknown kind ${JSON.stringify(value['kind'])}`)
+  }
+}
+
+function readOverrideDoc(value: unknown, file: string): ReplayOverrideDoc {
+  if (Array.isArray(value)) return value.map((entry, index) => readReplayEntry(entry, file, `entry ${index}`))
+  if (!isRecord(value) || !hasExactKeys(value, ['patches']) || !Array.isArray(value['patches'])) {
+    return invalidOverride(file, 'document', 'must be a ReplayEntry[] or { patches: [...] }')
+  }
+  return {
+    patches: value['patches'].map((value, index): ReplayOverridePatch => {
+      const location = `patch ${index}`
+      if (!isRecord(value) || !hasExactKeys(value, ['at', 'entry'])) {
+        return invalidOverride(file, location, 'must contain exactly at and entry')
+      }
+      const at = value['at']
+      if (typeof at !== 'number' || !Number.isSafeInteger(at) || at < 0) {
+        return invalidOverride(file, location, 'at must be a non-negative safe integer')
+      }
+      return { at, entry: readReplayEntry(value['entry'], file, `${location}.entry`) }
+    }),
+  }
+}
+
+/**
+ * Load the PRIMARY session's replay script: the sidecar override when present
+ * (whole-script replacement or `{ patches }` augmentation over the derived
+ * script), else the script derived from the session JSONL (fail-loud when the
+ * fixture is missing).
  * @param config - the fixture paths; only `file` and `overrideFile` are consulted.
- * @returns the primary session's replay entries.
+ * @returns the resolved primary-session script.
  */
 export function loadReplayScript(config: ReplayConfig): ReplayEntry[] {
   if (config.overrideFile !== undefined && existsSync(config.overrideFile)) {
-    const parsed: unknown = JSON.parse(readFileSync(config.overrideFile, 'utf8'))
-    if (!Array.isArray(parsed)) {
-      throw new Error(`llm-replay: override is not a JSON array: ${config.overrideFile}`)
+    const doc = readOverrideDoc(JSON.parse(readFileSync(config.overrideFile, 'utf8')) as unknown, config.overrideFile)
+    if (Array.isArray(doc)) return doc
+    const script = deriveScriptFromFile(config.file)
+    const derivedLength = script.length
+    const seenIndexes = new Set<number>()
+    for (const patch of doc.patches) {
+      if (patch.at > derivedLength) {
+        throw new Error(
+          `llm-replay: override patch index ${String(patch.at)} out of range `
+          + `(derived script has ${derivedLength} call(s); == length appends): ${config.overrideFile}`,
+        )
+      }
+      if (seenIndexes.has(patch.at)) {
+        throw new Error(`llm-replay: duplicate override patch index ${patch.at}: ${config.overrideFile}`)
+      }
+      seenIndexes.add(patch.at)
+      script[patch.at] = patch.entry
     }
-    return parsed as ReplayEntry[]
+    return script
   }
-  if (!existsSync(config.file)) {
-    throw new Error(`llm-replay: fixture not found: ${config.file} — run \`pnpm run test:snapshot:record\` first`)
+  return deriveScriptFromFile(config.file)
+}
+
+/** Derive the primary script from the session JSONL, failing loud on a missing fixture. */
+function deriveScriptFromFile(file: string): ReplayEntry[] {
+  if (!existsSync(file)) {
+    throw new Error(`llm-replay: fixture not found: ${file} — run \`pnpm run test:snapshot:record\` first`)
   }
-  return deriveReplayScript(parseSessionLog(readFileSync(config.file, 'utf8')))
+  return deriveReplayScript(parseSessionLog(readFileSync(file, 'utf8')))
 }
 
 /**
@@ -359,9 +491,8 @@ async function* replayEntry(entry: ReplayEntry, signal: AbortSignal | undefined,
       })
       /* v8 ignore next -- unreachable: the hang promise only ever rejects (on abort), never resolves; control never reaches here */
       return
+    /* v8 ignore next -- sidecar entries are validated before they reach the closed local union. */
     default:
-      // Closed local union: an unknown kind means malformed (hand-edited or
-      // drifted) sidecar data — fail loud with a runtime diagnostic.
       return assertNever(entry, 'llm-replay replay entry')
   }
 }

packages/support/llm-replay/tests/llm-replay.spec.ts

@@ -203,11 +203,91 @@ describe('loadReplayScript', () => {
     expect(() => loadReplayScript({ file: join(dir, 'absent.jsonl') })).toThrow(/fixture not found/)
   })
 
-  it('throws when the override is not a JSON array', () => {
+  it('rejects an override document that is neither supported form', () => {
     writeFileSync(file, sessionJsonl([]), 'utf8')
     const overrideFile = join(dir, 'replay.override.json')
     writeFileSync(overrideFile, '{"not":"array"}', 'utf8')
-    expect(() => loadReplayScript({ file, overrideFile })).toThrow(/not a JSON array/)
+    expect(() => loadReplayScript({ file, overrideFile })).toThrow(/document must be a ReplayEntry\[\] or \{ patches/)
+  })
+
+  it('patches form: swaps the named call index and keeps derived siblings', () => {
+    const callB: StreamChunk[] = [
+      { type: 'block-start', index: 0, blockType: 'text' },
+      { type: 'text-delta', index: 0, text: 'two' },
+      { type: 'finish', reason: { kind: 'stop' } },
+    ]
+    let seq = 1
+    writeFileSync(file, sessionJsonl([
+      ...TEXT_CHUNKS.map(c => chunkEvent(seq++, 1, 1, c)),
+      ...callB.map(c => chunkEvent(seq++, 1, 2, c)),
+    ]), 'utf8')
+    const overrideFile = join(dir, 'replay.override.json')
+    writeFileSync(overrideFile, JSON.stringify({
+      patches: [{ at: 0, entry: { kind: 'throw', chunks: [], message: 'transient', code: 'SERVER' } }],
+    }), 'utf8')
+    expect(loadReplayScript({ file, overrideFile })).toEqual([
+      { kind: 'throw', chunks: [], message: 'transient', code: 'SERVER' },
+      { kind: 'chunks', chunks: callB },
+    ])
+  })
+
+  it('patches form: at == derived length appends (the retry-attempt slot)', () => {
+    writeFileSync(file, sessionJsonl(TEXT_CHUNKS.map((c, i) => chunkEvent(i + 1, 1, 1, c))), 'utf8')
+    const overrideFile = join(dir, 'replay.override.json')
+    writeFileSync(overrideFile, JSON.stringify({
+      patches: [
+        { at: 0, entry: { kind: 'throw', chunks: [], message: '429', code: 'RATE_LIMIT' } },
+        { at: 1, entry: { kind: 'chunks', chunks: TEXT_CHUNKS } },
+      ],
+    }), 'utf8')
+    expect(loadReplayScript({ file, overrideFile })).toEqual([
+      { kind: 'throw', chunks: [], message: '429', code: 'RATE_LIMIT' },
+      { kind: 'chunks', chunks: TEXT_CHUNKS },
+    ])
+  })
+
+  it('patches form: an out-of-range index fails loud with the derived length', () => {
+    writeFileSync(file, sessionJsonl(TEXT_CHUNKS.map((c, i) => chunkEvent(i + 1, 1, 1, c))), 'utf8')
+    const overrideFile = join(dir, 'replay.override.json')
+    writeFileSync(overrideFile, JSON.stringify({ patches: [{ at: 2, entry: { kind: 'hang' } }] }), 'utf8')
+    expect(() => loadReplayScript({ file, overrideFile })).toThrow(/patch index 2 out of range.*1 call/s)
+  })
+
+  it('validates patch and entry shapes at the file boundary', () => {
+    writeFileSync(file, sessionJsonl([]), 'utf8')
+    const overrideFile = join(dir, 'replay.override.json')
+    const invalid: Array<{ doc: unknown; message: RegExp }> = [
+      { doc: null, message: /document must be/ },
+      { doc: { patches: [null] }, message: /patch 0 must contain exactly at and entry/ },
+      { doc: { patches: [{ at: -1, entry: { kind: 'hang' } }] }, message: /at must be a non-negative safe integer/ },
+      { doc: { patches: [{ at: 1.5, entry: { kind: 'hang' } }] }, message: /at must be a non-negative safe integer/ },
+      { doc: [42], message: /entry 0 must be an object/ },
+      { doc: [{ kind: 'chunks', chunks: 'nope' }], message: /chunks must be an array/ },
+      { doc: [{ kind: 'chunks', chunks: [], extra: true }], message: /invalid chunks-entry fields/ },
+      { doc: [{ kind: 'chunks', chunks: [{ type: 'bogus' }] }], message: /known StreamChunk type/ },
+      { doc: [{ kind: 'throw', chunks: [], message: 'nope', code: 'AUTH', extra: true }], message: /invalid throw-entry fields/ },
+      { doc: [{ kind: 'throw', chunks: [], message: '', code: 'AUTH' }], message: /message must be a non-empty string/ },
+      { doc: [{ kind: 'throw', chunks: [], message: 'nope', code: '' }], message: /code must be a non-empty string/ },
+      { doc: [{ kind: 'hang', extra: true }], message: /invalid hang-entry fields/ },
+      { doc: [{ kind: 'hang', readyFile: 1 }], message: /readyFile must be a non-empty string/ },
+      { doc: [{ kind: 'bogus' }], message: /unknown kind/ },
+    ]
+    for (const { doc, message } of invalid) {
+      writeFileSync(overrideFile, JSON.stringify(doc), 'utf8')
+      expect(() => loadReplayScript({ file, overrideFile })).toThrow(message)
+    }
+  })
+
+  it('rejects duplicate patch indexes instead of silently taking the last one', () => {
+    writeFileSync(file, sessionJsonl(TEXT_CHUNKS.map((c, i) => chunkEvent(i + 1, 1, 1, c))), 'utf8')
+    const overrideFile = join(dir, 'replay.override.json')
+    writeFileSync(overrideFile, JSON.stringify({
+      patches: [
+        { at: 0, entry: { kind: 'hang' } },
+        { at: 0, entry: { kind: 'throw', chunks: [], message: 'busy', code: 'SERVER' } },
+      ],
+    }), 'utf8')
+    expect(() => loadReplayScript({ file, overrideFile })).toThrow(/duplicate override patch index 0/)
   })
 })
 
@@ -364,16 +444,14 @@ describe('installLlmReplay (through the real LlmService)', () => {
       .toEqual([{ type: 'finish', reason: { kind: 'stop' } }])
   })
 
-  it('throws on a malformed sidecar entry kind (the assertNever guard)', async () => {
+  it('rejects a malformed sidecar entry kind before installing replay', async () => {
     writeFileSync(file, sessionJsonl([]), 'utf8')
     const overrideFile = join(dir, 'replay.override.json')
     // A kind the union does not know — hand-edited/drifted sidecar data.
     writeFileSync(overrideFile, JSON.stringify([{ kind: 'bogus' }]), 'utf8')
     const ctx = new Context()
     await ctx.plugin(LlmService)
-    installLlmReplay(ctx, { file, overrideFile })
-    await expect(drain(ctx.llm.stream({ provider: 'm', model: 'm', messages: [] })))
-      .rejects.toThrow(/llm-replay replay entry/)
+    expect(() => installLlmReplay(ctx, { file, overrideFile })).toThrow(/unknown kind/)
   })
 
   it('rejects a hang entry when the signal fires DURING the wait (abort listener path)', async () => {