@vellumai/assistant 0.8.9-staging.1 → 0.8.9-staging.2

package/src/agent/loop.ts

@@ -10,13 +10,11 @@ import {
   estimateToolsTokens,
   getCalibrationProviderKey,
 } from "../context/token-estimator.js";
-import type {
-  ContextWindowManager,
-  ContextWindowResult,
-} from "../context/window-manager.js";
+import type { ContextWindowResult } from "../context/window-manager.js";
 import type { InboundActorContext } from "../daemon/conversation-runtime-assembly.js";
 import type { ToolActivityMetadata } from "../daemon/message-types/web-activity.js";
 import type { TrustContext } from "../daemon/trust-context.js";
+import { stripHistoricalWebSearchResults } from "../daemon/web-search-history.js";
 import { HOOKS } from "../plugin-api/constants.js";
 import type {
   AssistantMessageContext,
@@ -28,6 +26,7 @@ import {
   DEFAULT_COMPACTION_PLUGIN_NAME,
   defaultCompact,
 } from "../plugins/defaults/compaction/compact.js";
+import { getContextWindowManager } from "../plugins/defaults/compaction/manager-store.js";
 import postCompact from "../plugins/defaults/memory-retrieval/hooks/post-compact.js";
 import { runHook } from "../plugins/pipeline.js";
 import type { CompactionCircuitEvent } from "../plugins/types.js";
@@ -467,15 +466,6 @@ export interface AgentLoopRunOptions {
    * compactor fail-closes to excluding guardian-only attachments.
    */
   trust: TrustContext;
-  /**
-   * The conversation's {@link ContextWindowManager}, supplied by the
-   * orchestrator so mid-loop in-place compaction can run against the real
-   * per-conversation manager. Absent for callers without a compaction path
-   * (agent wakes, standalone unit tests); the loop only reads it when the
-   * mid-loop budget gate trips with {@link AgentLoopRunOptions.compactInPlace}
-   * enabled.
-   */
-  contextWindowManager?: ContextWindowManager;
   /**
    * Ad-hoc inference-profile override applied to every LLM call the loop
    * issues. When set, each `SendMessageOptions.config` carries
@@ -726,7 +716,6 @@ export class AgentLoop {
     history: Message[],
     requestId: string | undefined,
     trust: TrustContext,
-    manager: ContextWindowManager | undefined,
     signal: AbortSignal | undefined,
     onEvent: (event: AgentEvent) => void | Promise<void>,
     overrideProfile: string | null,
@@ -741,9 +730,14 @@ export class AgentLoop {
     // Record the history-stripped marker right after stripping, before the
     // pipeline runs.
     await onEvent({ type: "history_stripped" });
+    // The compaction module owns the per-conversation manager; resolve it from
+    // the store rather than holding a handle on the loop. Absent for callers
+    // without a compaction path (agent wakes, standalone unit tests), which
+    // never reach this gate.
+    const manager = getContextWindowManager(this.conversationId);
     if (manager == null) {
       throw new PluginExecutionError(
-        "default-compaction: contextWindowManager run option is missing — orchestrator must supply it before invoking compaction",
+        `default-compaction: no ContextWindowManager registered for conversation ${this.conversationId} — the compaction store must construct one before compaction runs`,
         DEFAULT_COMPACTION_PLUGIN_NAME,
       );
     }
@@ -814,7 +808,6 @@ export class AgentLoop {
       resolveOverrideProfile,
       resolveContextWindow,
       compactInPlace = false,
-      contextWindowManager,
       isNonInteractive = false,
       modelProfile = null,
       actorContext = null,
@@ -940,7 +933,6 @@ export class AgentLoop {
                   history,
                   requestId,
                   trust,
-                  contextWindowManager,
                   signal,
                   onEvent,
                   resolveEffectiveOverrideProfile() ?? null,
@@ -1115,15 +1107,10 @@ export class AgentLoop {
         );
         rlog.info({ turn: toolUseTurns }, "LLM call start");
 
-        // Strip image contentBlocks from older tool results to prevent
-        // screenshots from accumulating in the context window. The LLM
-        // already saw each image on the turn it was captured; keeping
-        // base64 blobs in history rapidly exhausts the context budget.
-        // Also strip old AX tree snapshots to keep TTFT from growing
-        // linearly with step count in computer-use sessions.
-        const providerHistory = compactAxTreeHistory(
-          stripOldMediaBlocks(history),
-        );
+        // Sanitize the outbound history right before sending: drop accumulated
+        // media, collapse old AX-tree snapshots, and convert historical
+        // web-search results to text. See {@link preModelCallSanitize}.
+        const providerHistory = preModelCallSanitize(history);
 
         // A `pre-model-call` hook (below) can defer this turn's assistant
         // output; when set, the live text stream is held so an
@@ -1936,3 +1923,32 @@ function stripOldMediaBlocks(history: Message[]): Message[] {
     };
   });
 }
+
+/**
+ * Sanitize the outbound history immediately before a provider call, bundling
+ * the pre-send transforms the loop applies to every request:
+ * - {@link stripOldMediaBlocks} drops accumulated screenshot/audio bytes from
+ *   older tool results — the model saw the media on the turn it was captured.
+ * - {@link compactAxTreeHistory} collapses all but the most recent few
+ *   `<ax-tree>` snapshots so TTFT does not grow linearly with step count.
+ * - {@link stripHistoricalWebSearchResults} converts historical
+ *   `web_search_tool_result` blocks to text summaries; Anthropic's opaque
+ *   `encrypted_content` tokens expire / are route-scoped, and replaying a stale
+ *   one is rejected with `Invalid encrypted_content in search_result block`.
+ *
+ * Transforms the outbound copy only — the durable history keeps the rich
+ * originals and each send re-derives the sanitized projection (every transform
+ * is idempotent). Because it runs unconditionally before every provider call,
+ * it is the single place where oversized media and expired web-search tokens
+ * are guaranteed to be removed from a request.
+ *
+ * This is outbound-request preparation and should eventually move to a default
+ * `pre-model-call` plugin hook ({@link HOOKS.PRE_MODEL_CALL}) once that hook's
+ * context carries the outbound message list; for now it lives inline next to
+ * the provider call it guards.
+ */
+export function preModelCallSanitize(history: Message[]): Message[] {
+  const mediaStripped = stripOldMediaBlocks(history);
+  const axCompacted = compactAxTreeHistory(mediaStripped);
+  return stripHistoricalWebSearchResults(axCompacted).messages;
+}

package/src/config/bundled-skills/acp/SKILL.md

@@ -43,20 +43,20 @@ When the user first tries to use ACP and it's not enabled, set it up automatical
 
 2. Then retry the `acp_spawn` call. Do NOT run `vellum sleep && vellum wake` - that kills the conversation.
 
-No manual binary installation is needed first: missing adapter binaries are run on demand via bun, or installed automatically (see below).
+No manual binary installation is needed first: missing adapter binaries are installed automatically (see below).
 
 ## Automatic adapter availability
 
-When `acp_spawn` finds the agent's binary missing from PATH, the assistant runs the adapter through bun instead (`bun x --bun <package>`). Bun fetches the package into its cache on first use, so there is no global install and it works on hosts without node or npm (platform-hosted assistants ship bun only). When bun itself is not on PATH, the assistant falls back to silently installing the matching npm package globally and proceeds in the same call.
+When `acp_spawn` finds the agent's binary missing from PATH, the assistant installs it once via a sandboxed bun global install and then runs the real installed binary. The install runs in a fresh empty temporary directory (never the task's project directory), with the Claude/Gemini secrets stripped from the installer environment and the registry pinned to the public npm registry, so a malicious project directory cannot hijack package resolution or capture a token. After this one-time install, the adapter is a normal trusted binary on PATH and every later spawn (and resume) uses it directly.
 
-Only the allowlisted out-of-box packages are ever run or installed this way (`@agentclientprotocol/claude-agent-acp`, `@zed-industries/codex-acp`, `@google/gemini-cli`); user-configured agents with custom commands are never fetched or installed automatically.
+Only the allowlisted out-of-box packages are ever installed this way (`@agentclientprotocol/claude-agent-acp`, `@zed-industries/codex-acp`, `@google/gemini-cli`); user-configured agents with custom commands are never installed automatically.
 
-Manual installation is fallback guidance for unusual setups: bun and npm both unavailable, restricted global installs, or an auto-install failure (the failure reason is surfaced in the tool result).
+Manual installation is fallback guidance for unusual setups: bun unavailable, restricted global installs, or an auto-install failure (the failure reason is surfaced in the tool result).
 
 ```bash
-npm i -g @agentclientprotocol/claude-agent-acp   # claude
-npm i -g @zed-industries/codex-acp               # codex
-npm i -g @google/gemini-cli                      # gemini
+bun add -g @agentclientprotocol/claude-agent-acp   # claude
+bun add -g @zed-industries/codex-acp               # codex
+bun add -g @google/gemini-cli                       # gemini
 ```
 
 ## Claude setup
@@ -71,7 +71,7 @@ When the token is missing, do NOT ask the user to paste it into chat. Collect it
 
 ## Codex setup
 
-The `codex-acp` adapter is fetched automatically when missing, but it shells out to the underlying `codex` CLI, which must also be on PATH:
+The `codex-acp` adapter is installed automatically when missing, but it shells out to the underlying `codex` CLI, which must also be on PATH:
 
 1. **Install the Codex CLI** (version 0.111 or higher) via OpenAI's distribution channel of choice. The adapter will fail if `codex` isn't on PATH.
 
@@ -82,7 +82,7 @@ The `codex-acp` adapter is fetched automatically when missing, but it shells out
 
 ## Gemini setup
 
-Gemini CLI speaks ACP natively (`gemini --acp`) - there is no separate adapter binary. The CLI itself is fetched from `@google/gemini-cli` when missing.
+Gemini CLI speaks ACP natively (`gemini --acp`) - there is no separate adapter binary. The CLI itself is installed from `@google/gemini-cli` when missing.
 
 **Authenticate** with an API key through the credential store (the primary path):
 
@@ -120,14 +120,14 @@ A workspace `acp.agents.gemini` override is only for non-secret customization (c
 
 ## Updating an adapter
 
-This applies only to adapters installed globally via npm; adapters run via bun are fetched on first use and have no global install to update. If `acp_spawn` reports that an adapter is outdated, ask the user before updating. To update:
+Adapters are installed once via a bun global install. To update one to the latest version, ask the user first, then re-install it globally:
 
 ```bash
-npm i -g @agentclientprotocol/claude-agent-acp@latest
+bun add -g @agentclientprotocol/claude-agent-acp@latest
 # or
-npm i -g @zed-industries/codex-acp@latest
+bun add -g @zed-industries/codex-acp@latest
 # or
-npm i -g @google/gemini-cli@latest
+bun add -g @google/gemini-cli@latest
 ```
 
 Then retry the `acp_spawn` call.
@@ -140,7 +140,7 @@ Then retry the `acp_spawn` call.
 
 ## Discoverability
 
-Use `acp_list_agents` to see what's set up and what's missing. It returns each available agent profile, whether ACP is enabled, whether the agent is runnable (its binary is on PATH, or the assistant can run it via bun), and an install hint if not. This is the right tool to call when deciding between `claude`, `codex`, and `gemini`, or when the user asks "what coding agents do I have?"
+Use `acp_list_agents` to see what's set up and what's missing. It returns each available agent profile, whether ACP is enabled, whether the agent's binary is on PATH (missing binaries are installed automatically on first spawn), and an install hint if not. This is the right tool to call when deciding between `claude`, `codex`, and `gemini`, or when the user asks "what coding agents do I have?"
 
 ## Working directory
 

package/src/config/bundled-skills/acp/TOOLS.json

@@ -3,7 +3,7 @@
   "tools": [
     {
       "name": "acp_spawn",
-      "description": "Spawn an external coding agent (e.g. Claude Code, Codex, Gemini) via ACP to work on a task. Default profiles ship for `claude` (`claude-agent-acp`), `codex` (`codex-acp`), and `gemini` (`gemini --acp`); the assistant resolves the agent id to the right binary. The agent runs as a subprocess and streams results back. Use this when you want to delegate a coding task to an external agent that has its own tools, file editing, and terminal access. If a default agent's binary is missing, the assistant runs it via bun (fetching the package on first use, no global install) or falls back to an npm auto-install, proceeding in the same call; if neither works, an actionable install hint is returned - do NOT alter `agents.<id>.command` to swap binaries. If ACP is disabled, follow the setup instructions in SKILL.md.",
+      "description": "Spawn an external coding agent (e.g. Claude Code, Codex, Gemini) via ACP to work on a task. Default profiles ship for `claude` (`claude-agent-acp`), `codex` (`codex-acp`), and `gemini` (`gemini --acp`); the assistant resolves the agent id to the right binary. The agent runs as a subprocess and streams results back. Use this when you want to delegate a coding task to an external agent that has its own tools, file editing, and terminal access. If a default agent's binary is missing, the assistant installs it once via a sandboxed bun global install and proceeds in the same call; if that fails (e.g. bun unavailable), an actionable install hint is returned - do NOT alter `agents.<id>.command` to swap binaries. If ACP is disabled, follow the setup instructions in SKILL.md.",
       "category": "orchestration",
       "risk": "high",
       "input_schema": {
@@ -87,7 +87,7 @@
     },
     {
       "name": "acp_list_agents",
-      "description": "Lists ACP coding agents available to spawn. Each entry includes whether ACP is enabled, whether the agent is runnable (its binary is on PATH, or the assistant can run it via bun), and an install command if not. Use this to decide between 'claude', 'codex', and 'gemini' or to surface setup steps to the user.",
+      "description": "Lists ACP coding agents available to spawn. Each entry includes whether ACP is enabled, whether the agent's binary is on PATH (missing binaries are installed automatically on first spawn), and an install command if not. Use this to decide between 'claude', 'codex', and 'gemini' or to surface setup steps to the user.",
       "category": "orchestration",
       "risk": "low",
       "input_schema": {

package/src/config/bundled-skills/app-builder/SKILL.md

@@ -203,7 +203,7 @@ render(<App />, document.getElementById("app")!);
 
 #### `app_create` accepts EXACTLY these 7 keys — nothing else
 
-`name` (required), `description`, `schema_json`, `source_files`, `preview`, `auto_open`, `change_summary`.
+`name` (optional — defaults to the `preview` title, else "New App"), `description`, `schema_json`, `source_files`, `preview`, `auto_open`, `change_summary`.
 
 Anything else fails with `Invalid input for tool "app_create": Unknown parameter "X"`. The retired keys models still reach for:
 
@@ -251,10 +251,10 @@ Editing an existing app means reusing its `app_id` — never `app_create`. Resol
 
 - **`file_edit`** — targeted changes (styles, fixes, small features), full path `/workspace/data/apps/<slug>/src/...`
 - **`file_write`** — new files or full rewrites
-- **Rename / metadata** — edit `/workspace/data/apps/<slug>.json` directly. Not a new app.
+- **`app_update`** — rename or change description/schema, and/or rewrite files via `source_files`, in one call. It recompiles for you, so no separate `app_refresh` is needed. Use this instead of editing `<slug>.json` by hand.
 - **Full rebrand** — still iteration, edit the existing files.
 
-Then `app_refresh(app_id)` ONCE. If the change is substantial, `app_open(app_id, open_mode: "preview")` for a fresh card; for small tweaks the existing card stays valid.
+Then `app_refresh(app_id)` ONCE (unless you used `app_update`, which already recompiled). If the change is substantial, `app_open(app_id, open_mode: "preview")` for a fresh card; for small tweaks the existing card stays valid.
 
 > ⚠️ **`skill_load("app-builder")` is required before every `app_*` call** (including the first `app_create`). The skill can auto-unload between turns; without the reload, `app_refresh` / `app_open` error with "not currently active." It's idempotent — call it every time.
 

package/src/config/bundled-skills/app-builder/TOOLS.json

@@ -11,7 +11,7 @@
         "properties": {
           "name": {
             "type": "string",
-            "description": "Name of the app"
+            "description": "Optional name of the app. When omitted, defaults to the preview card title (or \"New App\" if there is none); rename later via app_update."
           },
           "description": {
             "type": "string",
@@ -87,11 +87,52 @@
             "description": "Retired. Multi-page apps route inside the Preact app under src/components/. Passing this returns a guidance error."
           }
         },
-        "required": ["name"]
+        "required": []
       },
       "executor": "tools/app-create.ts",
       "execution_target": "host"
     },
+    {
+      "name": "app_update",
+      "description": "Update an existing app: change its name, description, or schema, and/or rewrite source files, then recompile. Use this to iterate on an app you already created instead of calling app_create again (which would make a duplicate). Resolve the app_id with app_list first if you don't have it.",
+      "category": "apps",
+      "risk": "low",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "app_id": {
+            "type": "string",
+            "description": "The ID of the app to update"
+          },
+          "name": {
+            "type": "string",
+            "description": "New name for the app. Omit to leave the name unchanged."
+          },
+          "description": {
+            "type": "string",
+            "description": "New description for the app. Omit to leave the description unchanged."
+          },
+          "schema_json": {
+            "type": "string",
+            "description": "New JSON schema for the app data structure. Omit to leave the schema unchanged."
+          },
+          "source_files": {
+            "type": "object",
+            "description": "Map of relative file paths to file contents to write under the app directory (e.g. 'src/main.tsx', 'src/components/App.tsx', 'src/styles.css'). Supply changed source here instead of separate file_write calls. The app is recompiled after writing.",
+            "additionalProperties": {
+              "type": "string"
+            }
+          },
+          "change_summary": {
+            "type": "string",
+            "description": "Short summary of what changed, using git conventional commit format (e.g. 'fix: correct totals', 'feat: add dark mode'). Used as the version history entry."
+          }
+        },
+        "required": ["app_id"]
+      },
+      "executor": "tools/app-update.ts",
+      "execution_target": "host"
+    },
     {
       "name": "app_delete",
       "description": "Delete a persistent app and all its records.",

package/src/config/bundled-skills/app-builder/tools/app-update.ts

@@ -0,0 +1,18 @@
+import { setAppCommitMessage } from "../../../../memory/app-git-service.js";
+import * as appStore from "../../../../memory/app-store.js";
+import type { AppUpdateInput } from "../../../../tools/apps/executors.js";
+import { executeAppUpdate } from "../../../../tools/apps/executors.js";
+import type {
+  ToolContext,
+  ToolExecutionResult,
+} from "../../../../tools/types.js";
+
+export async function run(
+  input: Record<string, unknown>,
+  context: ToolContext,
+): Promise<ToolExecutionResult> {
+  if (typeof input.change_summary === "string" && input.change_summary.trim()) {
+    setAppCommitMessage(context.conversationId, input.change_summary.trim());
+  }
+  return executeAppUpdate(input as unknown as AppUpdateInput, appStore);
+}

package/src/config/bundled-tool-registry.ts

@@ -25,6 +25,7 @@ import * as appDelete from "./bundled-skills/app-builder/tools/app-delete.js";
 import * as appGenerateIcon from "./bundled-skills/app-builder/tools/app-generate-icon.js";
 import * as appList from "./bundled-skills/app-builder/tools/app-list.js";
 import * as appRefresh from "./bundled-skills/app-builder/tools/app-refresh.js";
+import * as appUpdate from "./bundled-skills/app-builder/tools/app-update.js";
 // ── app-control ────────────────────────────────────────────────────────────────
 import * as appControlClick from "./bundled-skills/app-control/tools/app-control-click.js";
 import * as appControlCombo from "./bundled-skills/app-control/tools/app-control-combo.js";
@@ -140,6 +141,7 @@ export const bundledToolRegistry = new Map<string, SkillToolScript>([
 
   // app-builder
   ["app-builder:tools/app-create.ts", appCreate],
+  ["app-builder:tools/app-update.ts", appUpdate],
   ["app-builder:tools/app-delete.ts", appDelete],
   ["app-builder:tools/app-refresh.ts", appRefresh],
   ["app-builder:tools/app-generate-icon.ts", appGenerateIcon],

package/src/config/call-site-defaults.ts

@@ -17,7 +17,6 @@ export const CALL_SITE_DEFAULTS: Record<LLMCallSite, CallSiteDefaultConfig> = {
   patternScan: { profile: "balanced" },
   narrativeRefinement: { profile: "balanced" },
   callAgent: { profile: "balanced" },
-  proactiveArtifactBuild: { profile: "balanced" },
   memoryConsolidation: { profile: "balanced" },
   identityIntro: { profile: "balanced" },
   emptyStateGreeting: { profile: "balanced" },
@@ -42,7 +41,6 @@ export const CALL_SITE_DEFAULTS: Record<LLMCallSite, CallSiteDefaultConfig> = {
   },
 
   filingAgent: { profile: "cost-optimized" },
-  proactiveArtifactDecision: { profile: "cost-optimized" },
   memoryExtraction: { profile: "cost-optimized" },
   memoryRetrieval: { profile: "cost-optimized" },
   memoryRetrospective: { profile: "cost-optimized" },

package/src/config/feature-flag-registry.json

@@ -140,7 +140,7 @@
     },
     {
       "id": "multi-platform-assistant",
-      "scope": "assistant",
+      "scope": "client",
       "key": "multi-platform-assistant",
       "label": "Multi-Platform Assistant Switcher",
       "description": "Enable the menu-bar assistant switcher for managing multiple platform-hosted assistants (new/switch/retire)",

package/src/config/schemas/call-site-catalog.ts

@@ -300,20 +300,6 @@ const CATALOG_RECORD: CatalogRecord = {
     description: "General-purpose LLM inference call site for skill use.",
     domain: "skills",
   },
-  proactiveArtifactDecision: {
-    id: "proactiveArtifactDecision",
-    displayName: "Proactive Artifact Decision",
-    description:
-      "Decides what personalized artifact to build for new users based on conversation context.",
-    domain: "agentLoop",
-  },
-  proactiveArtifactBuild: {
-    id: "proactiveArtifactBuild",
-    displayName: "Proactive Artifact Build",
-    description:
-      "Builds the personalized artifact in a background conversation with tool access.",
-    domain: "agentLoop",
-  },
   homeGreeting: {
     id: "homeGreeting",
     displayName: "Home Greeting",

package/src/config/schemas/llm.ts

@@ -77,8 +77,6 @@ export const LLMCallSiteEnum = z.enum([
   "meetChatOpportunity",
   "inference",
   "trustRuleSuggestion",
-  "proactiveArtifactDecision",
-  "proactiveArtifactBuild",
   "homeGreeting",
   "homeSuggestedPrompts",
 ]);

package/src/context/compactor.ts

@@ -22,6 +22,7 @@ import { optimizeImageForTransport } from "../agent/image-optimize.js";
 import type { CompactionConfig } from "../config/schemas/compaction.js";
 import type { LLMCallSite } from "../config/schemas/llm.js";
 import { filterMessagesForUntrustedActor } from "../daemon/message-provenance.js";
+import { stripHistoricalWebSearchResults } from "../daemon/web-search-history.js";
 import {
   getAttachmentContent,
   getAttachmentMetadataForMessage,
@@ -685,6 +686,25 @@ function extractTextFromResponse(content: ContentBlock[]): string {
     .join("\n");
 }
 
+// Build the outbound message list for a compaction provider call: convert
+// historical web_search_tool_result blocks to text, then append the
+// summarization instruction at the tail.
+//
+// Anthropic's opaque `encrypted_content` tokens are route-scoped and expire, so
+// replaying a stale one is rejected with `Invalid encrypted_content in
+// search_result block`. Sanitizing here — the single seam every compaction
+// provider call funnels through — keeps that error away from both the
+// assistant-driven and emergency summarization calls. Only this outbound copy
+// is sanitized; tail resolution and the persisted compaction result read the
+// caller's original messages, so durable history keeps the rich blocks. The
+// instruction stays at the tail so the prefix cache stays warm.
+function buildCompactionRequest(
+  history: Message[],
+  instruction: Message,
+): Message[] {
+  return [...stripHistoricalWebSearchResults(history).messages, instruction];
+}
+
 export async function runAssistantDrivenCompaction(
   args: CompactionRunArgs,
 ): Promise<CompactionRunResult> {
@@ -718,9 +738,7 @@ export async function runAssistantDrivenCompaction(
     manifestText,
   );
 
-  // Append instruction at the tail — prefix unchanged, so prefix cache
-  // stays warm.
-  const requestMessages = [...args.messages, instruction];
+  const requestMessages = buildCompactionRequest(args.messages, instruction);
 
   let response: ProviderResponse;
   try {
@@ -1090,7 +1108,7 @@ export async function runEmergencyCompaction(
     role: "user",
     content: [{ type: "text", text: EMERGENCY_COMPACTION_PROMPT }],
   };
-  const requestMessages = [...prefix, instruction];
+  const requestMessages = buildCompactionRequest(prefix, instruction);
 
   let response: ProviderResponse;
   try {

package/src/daemon/conversation-agent-loop.ts

@@ -76,6 +76,11 @@ import { enqueueMemoryRetrospectiveOnCompaction } from "../memory/memory-retrosp
 import { HOOKS } from "../plugin-api/constants.js";
 import type { UserPromptSubmitContext } from "../plugin-api/types.js";
 import { defaultCompact } from "../plugins/defaults/compaction/compact.js";
+import {
+  createInitialReducerState,
+  reduceContextOverflow,
+  type ReducerState,
+} from "../plugins/defaults/compaction/context-overflow-reducer.js";
 import { deepRepairHistory } from "../plugins/defaults/history-repair/terminal.js";
 import userPromptSubmitMemoryRetrieval, {
   type MemoryRetrievalHookContext,
@@ -98,11 +103,6 @@ import { getWorkspaceGitService } from "../workspace/git-service.js";
 import { commitTurnChanges } from "../workspace/turn-commit.js";
 import { cleanAssistantContent } from "./assistant-attachments.js";
 import { resolveOverflowAction } from "./context-overflow-policy.js";
-import {
-  createInitialReducerState,
-  reduceContextOverflow,
-  type ReducerState,
-} from "./context-overflow-reducer.js";
 import type { Conversation } from "./conversation.js";
 import {
   createEventHandlerState,
@@ -152,7 +152,6 @@ import {
   UNSENDABLE_IMAGE_NOTE,
 } from "./persist-unsendable-image.js";
 import { resolveTrustClass, type TrustContext } from "./trust-context.js";
-import { stripHistoricalWebSearchResults } from "./web-search-history.js";
 
 const log = getLogger("conversation-agent-loop");
 
@@ -881,16 +880,10 @@ export async function runAgentLoopImpl(
     const isTrustedActor = resolveTrustClass(ctx.trustContext) === "guardian";
     let currentInjectionMode: InjectionMode = "full";
     const memoryCtx: MemoryRetrievalHookContext = {
-      graphMemory: ctx.graphMemory,
-      config: getConfig(),
       onEvent,
-      isTrustedActor,
       conversationId: ctx.conversationId,
       userMessageId,
       logger: rlog,
-      // An external cancel aborts `prepareMemory` instead of letting it run
-      // to completion after the turn has already been torn down.
-      signal: abortController.signal,
       latestMessages: ctx.messages,
       requestId: reqId,
       mode: currentInjectionMode,
@@ -907,32 +900,6 @@ export async function runAgentLoopImpl(
     // message list downstream.
     let runMessages = memoryCtx.latestMessages;
 
-    // The `remember` tool handles scratchpad-style memory writes directly to the graph.
-
-    // Reducer state, tool-token budget, and calibration provider key consumed
-    // by the post-rejection convergence loop further down. The tool-token
-    // budget is resolved once per turn (the resolved tool set is stable across
-    // the turn); the calibration key matches the key recorded by `handleUsage`
-    // for wrapper providers (OpenRouter routing to Anthropic → key is
-    // `"anthropic"`).
-    let reducerState: ReducerState | undefined;
-    const toolTokenBudget = ctx.agentLoop.getToolTokenBudget(runMessages);
-    const estimationProviderName = getCalibrationProviderKey(ctx.provider);
-
-    // Replace historical web_search_tool_result blocks with text summaries.
-    // The opaque `encrypted_content` tokens Anthropic attaches to each result
-    // expire / are route-scoped; replaying a stale token is rejected with
-    // `Invalid encrypted_content in search_result block`. Titles + URLs
-    // preserve enough context for the model on follow-up turns.
-    const webSearchStrip = stripHistoricalWebSearchResults(runMessages);
-    if (webSearchStrip.stats.blocksStripped > 0) {
-      rlog.info(
-        { phase: "pre_run", ...webSearchStrip.stats },
-        "Converted historical web_search_tool_result blocks to text summaries",
-      );
-      runMessages = webSearchStrip.messages;
-    }
-
     // user-prompt-submit hook: plugins may transform `runMessages` right
     // before the agent loop receives them. Fires once per user turn at the
     // primary `agentLoop.run` only — the re-entry / retry calls further down
@@ -957,6 +924,16 @@ export async function runAgentLoopImpl(
     );
     runMessages = finalUserPromptCtx.latestMessages;
 
+    // Reducer state, tool-token budget, and calibration provider key consumed
+    // by the post-rejection convergence loop further down. The tool-token
+    // budget is resolved once per turn (the resolved tool set is stable across
+    // the turn); the calibration key matches the key recorded by `handleUsage`
+    // for wrapper providers (OpenRouter routing to Anthropic → key is
+    // `"anthropic"`).
+    let reducerState: ReducerState | undefined;
+    const toolTokenBudget = ctx.agentLoop.getToolTokenBudget(runMessages);
+    const estimationProviderName = getCalibrationProviderKey(ctx.provider);
+
     const shouldGenerateTitle = isReplaceableTitle(
       getConversation(ctx.conversationId)?.title ?? null,
     );
@@ -1021,7 +998,6 @@ export async function runAgentLoopImpl(
           onCheckpoint,
           callSite: turnCallSite,
           trust: loopTrust,
-          contextWindowManager: ctx.contextWindowManager,
           overrideProfile: turnOverrideProfile,
           resolveOverrideProfile: resolveCurrentOverrideProfile,
           resolveContextWindow,
@@ -1085,8 +1061,6 @@ export async function runAgentLoopImpl(
       // intentionally deferred until there's a concrete plugin-level use case.
       const retryRepair = deepRepairHistory(updatedHistory);
       runMessages = retryRepair.messages;
-      const retryStrip = stripHistoricalWebSearchResults(runMessages);
-      runMessages = retryStrip.messages;
       state.orderingErrorDetected = false;
       state.deferredOrderingError = null;
 
@@ -1380,14 +1354,6 @@ export async function runAgentLoopImpl(
         if (isTrustedActor && currentInjectionMode !== "minimal") {
           ctx.graphMemory.retrackCachedNodes();
         }
-        const convergenceStrip = stripHistoricalWebSearchResults(runMessages);
-        if (convergenceStrip.stats.blocksStripped > 0) {
-          rlog.info(
-            { phase: "convergence", ...convergenceStrip.stats },
-            "Converted historical web_search_tool_result blocks to text summaries",
-          );
-          runMessages = convergenceStrip.messages;
-        }
         state.contextTooLargeDetected = false;
         yieldedForBudget = false;
 
@@ -1469,14 +1435,6 @@ export async function runAgentLoopImpl(
           if (isTrustedActor && currentInjectionMode !== "minimal") {
             ctx.graphMemory.retrackCachedNodes();
           }
-          const fallbackStrip = stripHistoricalWebSearchResults(runMessages);
-          if (fallbackStrip.stats.blocksStripped > 0) {
-            rlog.info(
-              { phase: "fail_gracefully_compact", ...fallbackStrip.stats },
-              "Converted historical web_search_tool_result blocks to text summaries",
-            );
-            runMessages = fallbackStrip.messages;
-          }
           state.contextTooLargeDetected = false;
 
           updatedHistory = await runAgentLoop(runMessages);