[codex] clarify web search action labels

packages/agent-config/src/config.ts

@@ -136,6 +136,13 @@ export interface PersistedSession {
     cacheReadTokens: number
     cacheWriteTokens: number
   }
+  /**
+   * Calibration factor for the in-composer Context gauge — actual
+   * `input_tokens` / pre-turn estimate from the most recent completed
+   * turn. Persisted so a resumed session keeps its calibration instead
+   * of starting from 1.0 again.
+   */
+  contextEstimateScale?: number
 }
 
 // ── Main config ─────────────────────────────────────────────────────

packages/agent-core/src/agent.ts

@@ -228,6 +228,100 @@ export interface ToolCallbacks {
   getParentForkContext?: () => ParentForkContext | undefined
 }
 
+/**
+ * Names of every tool registered by `buildTools` itself (i.e. not from
+ * MCP servers or OAuth connector managers). Used by `categorizeTools`
+ * to split a flattened tool list into "system" vs. "MCP/connector" for
+ * the Context gauge popover. Keep this in sync when adding/removing
+ * built-in tools above. New tools added to `buildAntonCoreTools` should
+ * also land here.
+ */
+export const BUILT_IN_TOOL_NAMES = new Set<string>([
+  // Filesystem + shell + git + http
+  'shell',
+  'read',
+  'write',
+  'edit',
+  'glob',
+  'list',
+  'grep',
+  'git',
+  'http_api',
+  // Browser + media
+  'browser',
+  'image',
+  'clipboard',
+  // Web (Anton-native)
+  'web_search',
+  'web_research',
+  // Task / planning
+  'todo',
+  'task_tracker',
+  'plan',
+  'ask_user',
+  'artifact',
+  // Sub-agent / routine
+  'sub_agent',
+  'spawn_sub_agent',
+  'shared_state',
+  'routine',
+  'deliver_result',
+  // Anton platform
+  'memory',
+  'notification',
+  'database',
+  'publish',
+  'skill',
+  'set_session_title',
+  'update_project_context',
+  'activate_workflow',
+  // Pi-SDK low-level (rarely surfaced but possible)
+  'filesystem',
+  'process',
+  'network',
+])
+
+export interface CategorizedToolSizes {
+  /** Built-in tool count. */
+  systemToolCount: number
+  /** MCP-server + direct-OAuth-connector tool count. */
+  mcpToolCount: number
+  /** Sum of `name + description + JSON.stringify(parameters)` lengths for built-ins. */
+  systemToolChars: number
+  /** Same, for MCP/connector tools. */
+  mcpToolChars: number
+}
+
+/**
+ * Split a flattened `AgentTool[]` (as returned by `buildTools`) into
+ * built-in vs. MCP/connector buckets and report char-budget per bucket.
+ * The Context gauge popover divides the chars by 4 to estimate tokens.
+ */
+export function categorizeTools(tools: AgentTool[]): CategorizedToolSizes {
+  let systemToolCount = 0
+  let mcpToolCount = 0
+  let systemToolChars = 0
+  let mcpToolChars = 0
+  for (const tool of tools) {
+    const isBuiltIn = BUILT_IN_TOOL_NAMES.has(tool.name)
+    let chars = tool.name.length + (tool.description?.length ?? 0)
+    try {
+      chars += JSON.stringify(tool.parameters ?? {}).length
+    } catch {
+      // Parameter schemas should always serialize, but be defensive — a
+      // bad schema must not break the gauge.
+    }
+    if (isBuiltIn) {
+      systemToolCount += 1
+      systemToolChars += chars
+    } else {
+      mcpToolCount += 1
+      mcpToolChars += chars
+    }
+  }
+  return { systemToolCount, mcpToolCount, systemToolChars, mcpToolChars }
+}
+
 export function buildTools(
   config: AgentConfig,
   callbacks?: ToolCallbacks,

packages/agent-core/src/harness/codex-harness-session.ts

@@ -978,6 +978,7 @@ export class CodexHarnessSession {
               totalTokens?: number
             }
             last?: { inputTokens?: number; outputTokens?: number; cachedInputTokens?: number }
+            modelContextWindow?: number
           }
         }
       | undefined
@@ -1000,6 +1001,40 @@ export class CodexHarnessSession {
         cacheWriteTokens: 0,
       },
     })
+
+    // Drive the in-composer Context gauge for harness sessions. The
+    // harness CLI doesn't expose how its own prompt is split, so the
+    // popover renders a 2-row breakdown (Messages + Free space).
+    //
+    // Prefer `last.inputTokens` (per-turn input that just hit the model)
+    // over `total.inputTokens` for the gauge, because `total` is the
+    // SUM of input across every turn this thread (including cached
+    // tokens) — on long conversations it routinely exceeds the actual
+    // prompt size and would drive the gauge past 100%. Fall back to
+    // `total` only when `last` isn't reported.
+    const contextWindow = p?.tokenUsage?.modelContextWindow ?? 0
+    if (contextWindow > 0) {
+      const lastInput = p?.tokenUsage?.last?.inputTokens
+      const totalInput = p?.tokenUsage?.total?.inputTokens ?? 0
+      const messages = Math.min(
+        contextWindow,
+        typeof lastInput === 'number' && lastInput > 0 ? lastInput : totalInput,
+      )
+      this.emit({
+        type: 'context_update',
+        breakdown: {
+          contextWindow,
+          systemPrompt: 0,
+          systemTools: 0,
+          mcpTools: 0,
+          skills: 0,
+          memoryFiles: 0,
+          messages,
+          autocompactBuffer: 0,
+          source: 'harness',
+        },
+      })
+    }
   }
 
   private onCompacted() {

packages/agent-core/src/prompt-layers.ts

@@ -713,6 +713,49 @@ Your native tools (filesystem, shell, code editing, git, web search, etc.) remai
   )
 }
 
+// ── Per-layer size accounting (for the Context gauge popover) ──────
+
+/**
+ * Char-budget breakdown of every layer that contributes to a Pi-SDK
+ * session's prompt. Returned alongside the assembled string so the
+ * server can populate `ContextBreakdown` without re-running the layer
+ * builders. Char counts get divided by 4 at the call site to estimate
+ * tokens — same heuristic as `estimateMessageTokens` in compaction.ts.
+ */
+export interface SessionPromptLayerSizes {
+  identity: number
+  workspaceRules: number
+  userRules: number
+  currentContext: number
+  surface: number
+  memory: number
+  projectMemoryInstructions: number
+  agentContext: number
+  connectors: number
+  projectTypeGuidelines: number
+  referenceKnowledge: number
+  skills: number
+  workflows: number
+}
+
+export function emptyPromptLayerSizes(): SessionPromptLayerSizes {
+  return {
+    identity: 0,
+    workspaceRules: 0,
+    userRules: 0,
+    currentContext: 0,
+    surface: 0,
+    memory: 0,
+    projectMemoryInstructions: 0,
+    agentContext: 0,
+    connectors: 0,
+    projectTypeGuidelines: 0,
+    referenceKnowledge: 0,
+    skills: 0,
+    workflows: 0,
+  }
+}
+
 // ── High-level entry point ──────────────────────────────────────────
 
 export interface HarnessContextPromptOpts {