diff --git a/Agent.md b/Agent.md
index ea6ef940..a394fc86 100644
--- a/Agent.md
+++ b/Agent.md
@@ -119,7 +119,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.token; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (1133) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (1136) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (95: 45 daemon_client + 20 conn-manager + 8 integration + 6 build-config + 7 gui-state + 3 preload-api + 4 boot-contract + 2 theme-guard) — syntax: `node --check main.js preload.js daemon_client.js`
 Renderer: `cd emrg/gui/renderer && npm run typecheck && npm test` (448: 5 snapshot-store + 9 utils + 3 ErrorBoundary + 2 App smoke + 11 commands + 4 copywriting + 11 i18n + 11 markdown + 15 transcript + 7 TranscriptView + 15 history + 22 composer + 14 Composer + 12 sidebar + 17 Sidebar + 9 fileTree + 9 FileTree + 16 resultPanel + 8 ResultPanel + 27 workspaceView + 10 WorkspaceView + 10 dialog + 6 Dialog + 9 ConfirmDialog + 9 RenameDialog + 10 dialogLists + 3 HelpDialog + 9 MemoryDialog + 6 SkillsDialog + 9 openSession + 6 WelcomeDialog + 8 OpenSessionDialog + 7 NewSessionDialog + 7 rewind + 8 RewindDialog + 7 GithubDeviceDialog + 15 daemonBridge + 7 DaemonBridgeProvider + 25 Shell + 15 DialogHost + 20 SettingsPanel + 6 TaskFormDialog + 5 RantDialog + 4 vendorMarkdown) + `npm run build` → `renderer/dist/`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
diff --git a/emrg/server/daemon.py b/emrg/server/daemon.py
index 24319c08..e034804a 100644
--- a/emrg/server/daemon.py
+++ b/emrg/server/daemon.py
@@ -3143,11 +3143,27 @@ def _detect_silent_anchor_drift(self, session, real_pt: int, estimate: int) -> N
             session.session_id, shift, real_pt, old_real, estimate, old_est,
         )
 
+    # Conservative per-image token allowance for vision content (Codex
+    # #41003 class bug): a pasted image in an image_url block costs the API
+    # hundreds-to-thousands of tokens, but the char-based estimator counted
+    # it as ~0 (content was a list, not a str). Auto-compact then never fired
+    # on image-heavy sessions and the context could overflow. The exact bill
+    # depends on resolution/detail (OpenAI high-detail ≈ 85 + 170/tile, max
+    # ~765); 1000 covers typical pasted screenshots. The usage anchor
+    # (self._usage_anchors) self-corrects the residual on the next real
+    # API usage, so this only needs to be in the right ballpark.
+    _TOKENS_PER_IMAGE = 1000
+
     def _estimate_tokens(self, messages: list[dict]) -> int:
         """Rough token estimation from OpenAI-format messages.
 
         Character-aware: CJK ≈ 2 chars/token, ASCII ≈ 4 chars/token.
         Adds +3 tokens per message for role/content metadata overhead.
+
+        Vision-aware: content may be a list of parts (OpenAI format:
+        {"type": "text" | "image_url", ...}) — text parts are char-counted,
+        each image_url block gets a fixed _TOKENS_PER_IMAGE allowance, and
+        unknown dict parts fall back to JSON char-counting.
         """
         total = 0
         for m in messages:
@@ -3155,11 +3171,32 @@ def _estimate_tokens(self, messages: list[dict]) -> int:
             content = m.get("content") or ""
             if isinstance(content, str):
                 total += self._count_chars_for_tokens(content)
+            elif isinstance(content, list):
+                total += self._estimate_content_parts(content)
             for tc in (m.get("tool_calls") or []):
                 tc_str = json.dumps(tc, ensure_ascii=False)
                 total += self._count_chars_for_tokens(tc_str)
         return total
 
+    @staticmethod
+    def _estimate_content_parts(parts: list) -> int:
+        """Estimate tokens for OpenAI vision-format content parts."""
+        total = 0
+        for part in parts:
+            if isinstance(part, str):
+                total += EmrgServer._count_chars_for_tokens(part)
+            elif isinstance(part, dict):
+                ptype = part.get("type")
+                if ptype == "text" and isinstance(part.get("text"), str):
+                    total += EmrgServer._count_chars_for_tokens(part["text"])
+                elif ptype == "image_url":
+                    total += EmrgServer._TOKENS_PER_IMAGE
+                else:
+                    total += EmrgServer._count_chars_for_tokens(
+                        json.dumps(part, ensure_ascii=False)
+                    )
+        return total
+
     @staticmethod
     def _estimate_text(text: str) -> int:
         """Rough token count for a plain text string."""
diff --git a/tests/test_daemon.py b/tests/test_daemon.py
index da6c2ee2..372dc1a5 100644
--- a/tests/test_daemon.py
+++ b/tests/test_daemon.py
@@ -973,6 +973,49 @@ def test_estimate_tokens_multiple_messages():
     assert result == 7
 
 
+def test_estimate_tokens_vision_image_block():
+    """Vision content: text parts char-counted, image_url blocks get allowance.
+
+    Codex #41003-class fix: before, list-form content was skipped entirely
+    (only +3 overhead), so pasted images estimated at ~0 tokens and
+    auto-compact never fired on image-heavy sessions.
+    """
+    server = _make_server()
+    msgs = [{
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "what is in this screenshot"},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+        ],
+    }]
+    result = server._estimate_tokens(msgs)
+    text_tokens = server._count_chars_for_tokens("what is in this screenshot")
+    # 3 overhead + text tokens + 1 image allowance
+    assert result == 3 + text_tokens + server._TOKENS_PER_IMAGE
+
+
+def test_estimate_tokens_vision_multiple_images():
+    """Three image blocks → three allowances (plus overhead)."""
+    server = _make_server()
+    msgs = [{
+        "role": "user",
+        "content": [
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,BBBB"}},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,CCCC"}},
+        ],
+    }]
+    result = server._estimate_tokens(msgs)
+    assert result == 3 + 3 * server._TOKENS_PER_IMAGE
+
+
+def test_estimate_tokens_vision_empty_list():
+    """Empty content list → just overhead (no crash)."""
+    server = _make_server()
+    msgs = [{"role": "user", "content": []}]
+    assert server._estimate_tokens(msgs) == 3
+
+
 # ── _estimate_single ──────────────────────────────────────────────