fix(agent): bound agent working context with autoCompactWindow

src/lib/agent/__tests__/auto-compact-window.test.ts

@@ -0,0 +1,28 @@
+import { describe, it, expect } from 'vitest';
+import { AGENT_AUTO_COMPACT_WINDOW } from '@lib/constants';
+
+/**
+ * Guards the fix for the 2026-07-09 cost regression: the agent enables the
+ * 1M-context beta, so auto-compaction only bounds per-generation input cost
+ * when `settings.autoCompactWindow` pins a smaller working window.
+ *
+ * The SDK validates this setting as an int in [100_000, 1_000_000] and
+ * *silently drops* out-of-range values (`.catch(void 0)`), which would revert
+ * to deferring compaction to the 1M ceiling — reintroducing the regression
+ * with no error. Keep the constant inside the accepted band, and well under
+ * the 1M physical window so compaction has room to run.
+ */
+describe('AGENT_AUTO_COMPACT_WINDOW', () => {
+  const SDK_MIN = 100_000;
+  const SDK_MAX = 1_000_000;
+
+  it('is an integer the SDK will accept (not silently dropped)', () => {
+    expect(Number.isInteger(AGENT_AUTO_COMPACT_WINDOW)).toBe(true);
+    expect(AGENT_AUTO_COMPACT_WINDOW).toBeGreaterThanOrEqual(SDK_MIN);
+    expect(AGENT_AUTO_COMPACT_WINDOW).toBeLessThanOrEqual(SDK_MAX);
+  });
+
+  it('stays well below the 1M physical window so compaction has headroom', () => {
+    expect(AGENT_AUTO_COMPACT_WINDOW).toBeLessThan(SDK_MAX);
+  });
+});

src/lib/agent/agent-interface.ts

@@ -18,6 +18,8 @@ import {
   WIZARD_ORCHESTRATOR_FLAG_KEY,
   WIZARD_USER_AGENT,
   DEFAULT_AGENT_MODEL,
+  CONTEXT_1M_BETA,
+  AGENT_AUTO_COMPACT_WINDOW,
 } from '@lib/constants';
 import {
   type AdditionalFeature,
@@ -964,7 +966,13 @@ export async function runAgent(
         model: agentConfig.model,
         cwd: agentConfig.workingDirectory,
         permissionMode: 'acceptEdits',
-        betas: ['context-1m-2025-08-07'],
+        betas: [CONTEXT_1M_BETA],
+        // The 1M beta gives compaction room to run without overflowing on
+        // large projects, but on its own it lets the working context (and the
+        // per-turn input we're billed for) grow toward 1M once the gateway
+        // honors the beta. Pin the auto-compact window so compaction fires at a
+        // bounded size, keeping per-generation input near the pre-1M band.
+        settings: { autoCompactWindow: AGENT_AUTO_COMPACT_WINDOW },
         mcpServers: agentConfig.mcpServers,
         agents: {
           'general-purpose': {

src/lib/agent/mcp-prompt-streaming.ts

@@ -14,7 +14,11 @@
 
 import type { AgentChunk } from '@ui/tui/services/mcp-suggested-prompts-services';
 import type { Credentials } from '@lib/wizard-session';
-import { DEFAULT_AGENT_MODEL, WIZARD_USER_AGENT } from '@lib/constants';
+import {
+  DEFAULT_AGENT_MODEL,
+  WIZARD_USER_AGENT,
+  CONTEXT_1M_BETA,
+} from '@lib/constants';
 import { HostResolution, mcpUrlFor } from '@lib/host-resolution';
 import { logToFile } from '@utils/debug';
 import { buildAgentEnv } from '@lib/agent/agent-interface';
@@ -265,8 +269,9 @@ export async function* runMcpPromptViaSdk(args: {
         permissionMode: 'acceptEdits',
         maxTurns: MAX_TURNS,
         // Match agent-interface.ts — the 1M context beta is what keeps
-        // resumed follow-up sessions from truncating after a few turns.
-        betas: ['context-1m-2025-08-07'],
+        // resumed follow-up sessions from truncating after a few turns. This
+        // flow is bounded by maxTurns above, so it needs no autoCompactWindow.
+        betas: [CONTEXT_1M_BETA],
         // Only load project-level skills/settings. Without this the SDK
         // defaults to ['user', 'project'] and a user's
         // `~/.claude/settings.json` (apiKeyHelper / env block) can

src/lib/agent/runner/harness/anthropic/README.md

@@ -24,7 +24,10 @@ Both entry points are implemented:
   user's OAuth token (`CLAUDE_CODE_OAUTH_TOKEN` + `ANTHROPIC_BASE_URL`).
   Bedrock fallback via `x-posthog-use-bedrock-fallback: true`.
 - **Context window:** 1M-context beta (`context-1m-2025-08-07`) so large
-  projects don't overflow during compaction.
+  projects don't overflow during compaction. The billed working context is
+  bounded separately by `settings.autoCompactWindow` (`AGENT_AUTO_COMPACT_WINDOW`,
+  200K) — auto-compaction fires at that size rather than deferring to the 1M
+  ceiling, keeping per-generation input cost near the pre-1M band.
 - **Custom headers:** wizard flags (`X-POSTHOG-FLAG-*`) and metadata
   (`X-POSTHOG-PROPERTY-*`) piggyback on every gateway request for tracing.
 - **Model routing:** `AgentConfig.modelOverride` accepts any gateway model id

src/lib/agent/runner/harness/pi/index.ts

@@ -23,6 +23,7 @@ import {
   POSTHOG_PROPERTY_HEADER_PREFIX,
   WIZARD_REMARK_EVENT_NAME,
   WIZARD_USER_AGENT,
+  CONTEXT_1M_BETA,
 } from '@lib/constants';
 import { analytics } from '@utils/analytics';
 import { AgentErrorType } from '@lib/agent/agent-interface';
@@ -153,7 +154,9 @@ function buildGatewayHeaders(
     'x-posthog-use-bedrock-fallback': 'true',
     // 1M context window, same as the anthropic edition — pi otherwise runs at
     // 200k and overflows on larger projects (the post-run compaction failures).
-    'anthropic-beta': 'context-1m-2025-08-07',
+    // Note: pi drives its own compaction (pi-coding-agent), so the anthropic
+    // edition's `autoCompactWindow` cost guard does not apply here.
+    'anthropic-beta': CONTEXT_1M_BETA,
   };
   for (const [key, value] of Object.entries(wizardMetadata)) {
     const name = key.startsWith(POSTHOG_PROPERTY_HEADER_PREFIX)

src/lib/constants.ts

@@ -44,6 +44,27 @@ export const GPT5_4_MODEL = 'openai/gpt-5.4';
  */
 export const GPT5_MINI_MODEL = 'openai/gpt-5-mini';
 
+/**
+ * Anthropic 1M-context beta. Enabled on the agent so a run has physical
+ * headroom for compaction to succeed on large projects — but it does NOT set
+ * the working-context size the model actually pays for on every turn; that is
+ * `AGENT_AUTO_COMPACT_WINDOW` below.
+ */
+export const CONTEXT_1M_BETA = 'context-1m-2025-08-07';
+
+/**
+ * Working-context ceiling (tokens) at which the SDK auto-compacts, passed via
+ * `Options.settings.autoCompactWindow`. Decoupled from the physical window: the
+ * 1M beta ({@link CONTEXT_1M_BETA}) gives compaction room to run, while this
+ * keeps the per-turn input the run is billed for near the ~110-200K band.
+ *
+ * Without it, once the gateway honors the 1M beta the SDK defers compaction to
+ * ~1M, so a long linear run's context climbs past 600K and per-generation cost
+ * balloons (the 2026-07-09 spend regression). Must stay within the SDK's
+ * accepted range (100K-1M).
+ */
+export const AGENT_AUTO_COMPACT_WINDOW = 200_000;
+
 // ── Agent runner routing axes ────────────────────────────────────────
 
 /**