[BUG] Insanely high usage (56% 5h pro) after "continue" (Claude Sonnet 4.6 - auto effort) with just a few file edits
Resolved 💬 11 comments Opened Apr 6, 2026 by milankubin Closed Apr 24, 2026
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
Honestly, the usage is so random:
Often the usage speaks initially on a fresh session - no skills loaded, and just two mcp's: Serena and Context7.
Plugins: typescript-lsp, pyright-lsp.
Serena lists some files in initial prompt for reference (small .md memory files).
Initial prompt is a few paragraphs worth of small list of memory (.md) files and when to load them.
They are loaded on-demand (AGENT discretion).
It tends to settle down around 25% and then get pretty good usage out of Sonnet.
But this is what I would expect from an Opus 4.6 high effort planning prompt.
This is just insane.
I typed "continue" went AFK to have lunch, expecting a few small fixes, or it waiting for permissions.
But it stopped after some tiny edits.
Serena tool calls are usually lighter than file reads as well, and these files are not huge anyway.
Code is pretty well chunked up and organized.
Explain this...
● Bash(git status && git log --oneline -12)
⎿ On branch dev
Your branch is ahead of 'bitbucket/dev' by 7 commits.
(use "git push" to publish your local commits)
… +17 lines (ctrl+o to expand)
⎿ You're out of extra usage · resets 2pm (Europe/Amsterdam)
/upgrade or /extra-usage to finish what you’re working on.
✻ Cogitated for 4m 13s2 PM :
❯ continue
● Working tree is clean. The STATUS has P9-6.4h (TYPE_CHECKING generator fix) and P9-6.13 (dev-up check) as next tasks. Dev-up will fail
at BRK-2 regardless — let me do P9-6.4h first so the generator is correct before re-running it, then dev-up.
Loading the needed memories:
● serena - Read Memory (MCP)(memory_name: "backend/src/core/codegen/orm_dto_generator")
⎿ {
"result": "# orm_dto_generator — ORM and DTO Generator\n\n## What it is\n\n`backend/src/core/codegen/orm_dto_generator.py` —
the code generator that produces SQLAlchemy ORM models and Pydantic DTO models from feature config YAMLs.\n\nClass: `OrmDtoGe
… +14 lines (ctrl+o to expand)
● serena - Read Memory (MCP)(memory_name: "global/prescriptive/authoring/python/type_checking_imports")
⎿ {
"result": "# TYPE_CHECKING Import Pattern\n\nUse `if TYPE_CHECKING:` for imports needed only in annotations (avoids circular
\nimports, zero runtime cost):\n\n```python\nfrom typing import TYPE_CHECKING\n\nif TYPE_CHECKING:\n from src.core.features
… +5 lines (ctrl+o to expand)
● Now let me find the relevant generator code:
● serena - Get Symbols Overview (MCP)(relative_path: "backend/src/core/codegen/orm_dto_generator.py", depth: 1)
⎿ {
"result": "{\"Constant\": [\"GENERATED_FILE_MARKER\", \"GENERATED_PACKAGE_MARKER\", \"GENERATED_AT_PATTERN\", \"GM001\", \"G
M002\", \"GM003\", \"GM004\", \"GM005\", \"GM006\", \"GM007\", \"_BEGIN_MARKER_TPL\", \"_END_MARKER_TPL\", \"_LEGACY_BEGIN\",
… +20 lines (ctrl+o to expand)
● serena - Find Symbol (MCP)(name_path_pattern: "OrmDtoGenerator/_render_model_module", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", include_body: true)
⎿ {
"result": "[{\"name_path\": \"OrmDtoGenerator/_render_model_module\", \"kind\": \"Method\", \"relative_path\": \"backend/src
/core/codegen/orm_dto_generator.py\", \"body_location\": {\"start_line\": 481, \"end_line\": 548}, \"body\": \"@classmethod\\n
… +21 lines (ctrl+o to expand)
● Now I need to understand _ModelRenderContext and how relationships are resolved:
● serena - Find Symbol (MCP)(name_path_pattern: "_ModelRenderContext", relative_path: "backend/src/core/codegen/orm_dto_generator.py",
include_body: true)
⎿ {
"result": "[{\"name_path\": \"_ModelRenderContext\", \"kind\": \"Class\", \"relative_path\": \"backend/src/core/codegen/orm_
dto_generator.py\", \"body_location\": {\"start_line\": 153, \"end_line\": 159}, \"body\": \"@dataclass(frozen=True)\\nclass _
… +3 lines (ctrl+o to expand)
● serena - Find Symbol (MCP)(name_path_pattern: "OrmDtoGenerator/_resolve_relationship_target_class", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", include_body: true)
⎿ {
"result": "[{\"name_path\": \"OrmDtoGenerator/_resolve_relationship_target_class\", \"kind\": \"Method\", \"relative_path\":
\"backend/src/core/codegen/orm_dto_generator.py\", \"body_location\": {\"start_line\": 1211, \"end_line\": 1242}, \"body\": \
… +10 lines (ctrl+o to expand)
● serena - Find Symbol (MCP)(name_path_pattern: "OrmDtoGenerator/_render_model_relationship", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", include_body: true)
⎿ {
"result": "[{\"name_path\": \"OrmDtoGenerator/_render_model_relationship\", \"kind\": \"Method\", \"relative_path\": \"backe
nd/src/core/codegen/orm_dto_generator.py\", \"body_location\": {\"start_line\": 1316, \"end_line\": 1378}, \"body\": \"@classm
… +25 lines (ctrl+o to expand)
● serena - Find Symbol (MCP)(name_path_pattern: "OrmDtoGenerator/_collect_model_render_context", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", include_body: true)
⎿ {
"result": "[{\"name_path\": \"OrmDtoGenerator/_collect_model_render_context\", \"kind\": \"Method\", \"relative_path\": \"ba
ckend/src/core/codegen/orm_dto_generator.py\", \"body_location\": {\"start_line\": 404, \"end_line\": 438}, \"body\": \"@class
… +10 lines (ctrl+o to expand)
● serena - Find Symbol (MCP)(name_path_pattern: "OrmDtoGenerator/_render_inline_column_relationship", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", include_body: true)
⎿ {
"result": "[{\"name_path\": \"OrmDtoGenerator/_render_inline_column_relationship\", \"kind\": \"Method\", \"relative_path\":
\"backend/src/core/codegen/orm_dto_generator.py\", \"body_location\": {\"start_line\": 1262, \"end_line\": 1314}, \"body\": \
… +19 lines (ctrl+o to expand)
● Now I have full context. Let me also check the ModelEntry type to understand the relationships field:
● serena - Find Symbol (MCP)(name_path_pattern: "ModelEntry", relative_path: "backend/src/core/codegen/orm_dto_generator.py",
include_body: true)
⎿ {
"result": "[]"
}
Searched for 2 patterns (ctrl+o to expand)
● Good. Now I have everything needed. The fix requires:
1. A new helper method _collect_cross_module_type_checking_imports
2. Updated _render_model_module to use it
Let me implement:
● serena - Find Symbol (MCP)(name_path_pattern: "OrmDtoGenerator/_resolve_relationship_target_class", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", include_body: false)
⎿ {
"result": "[{\"name_path\": \"OrmDtoGenerator/_resolve_relationship_target_class\", \"kind\": \"Method\", \"relative_path\":
\"backend/src/core/codegen/orm_dto_generator.py\", \"body_location\": {\"start_line\": 1211, \"end_line\": 1242}}]"
}
● Now implementing. First insert the helper after _resolve_relationship_target_class, then update _render_model_module:
● serena - Insert After Symbol (MCP)(name_path: "OrmDtoGenerator/_resolve_relationship_target_class", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", body: "\n @classmethod\n def
_collect_cross_module_type_checking_imports(\n cls,\n model_items: list[tuple[str,
ModelEntry]],\n *,\n module_path: str,\n context: _ModelRenderContext,\n )
-> list[tuple[str, str]]:\n \"\"\"Collect cross-module relationship targets that need
TYPE_CHECKING imports.\n\n Args:\n model_items: Model names and specs for the
module being rendered.\n module_path: Dotted module path being rendered.\n
context: Global rendering context.\n\n Returns:\n Sorted unique (source_module,
class_name) pairs for TYPE_CHECKING block.\n \"\"\"\n class_to_module: dict[str, str]
= {}\n for impl_path in context.model_by_impl:\n if \".\" in impl_path:\n
mod, target_cls = impl_path.rsplit(\".\", 1)\n class_to_module[target_cls] =
mod\n\n cross_module: set[tuple[str, str]] = set()\n\n for model_name, model_spec in
model_items:\n for rel_spec in (model_spec.relationships or {}).values():\n
target_class = cls._resolve_relationship_target_class(\n
current_class_name=model_name,\n current_model=model_spec,\n
target=rel_spec.target,\n context=context,\n )\n if
target_class == model_name:\n continue\n target_module =
class_to_module.get(target_class)\n if target_module and target_module !=
module_path:\n cross_module.add((target_module, target_class))\n\n for
col_spec in model_spec.columns.values():\n if col_spec.fk is None or
col_spec.fk.relationship is None:\n continue\n target_class =
cls._resolve_relationship_target_class(\n current_class_name=model_name,\n
current_model=model_spec,\n target=col_spec.fk.target,\n
context=context,\n )\n if target_class == model_name:\n
continue\n target_module = class_to_module.get(target_class)\n
if target_module and target_module != module_path:\n
cross_module.add((target_module, target_class))\n\n return sorted(cross_module)\n")
⎿ {
"result": "OK"
}
● serena - Replace Symbol Body (MCP)(name_path: "OrmDtoGenerator/_render_model_module", relative_path:
"backend/src/core/codegen/orm_dto_generator.py", body: "@classmethod\n def
_render_model_module(\n cls,\n model_items: list[tuple[str, ModelEntry]],\n
*,\n source_hints: list[str],\n module_path: str,\n context:
_ModelRenderContext,\n ) -> str:\n \"\"\"Render one generated model module.\n\n
Args:\n model_items: Model names and validated specs for one module.\n
source_hints: YAML source hints included in the header.\n module_path: Dotted module path
being rendered.\n context: Global rendering context for relationship and inheritance\n
resolution.\n\n Returns:\n Generated model module source code.\n
\"\"\"\n\n used_mixins = {\n mixin_name\n for _, model_spec in
model_items\n for mixin_name in model_spec.mixins\n }\n mixin_import_lines
= [\n f\"from {cls.MIXIN_IMPORT_PATHS[mixin_name]} import {mixin_name}\"\n for
mixin_name in sorted(used_mixins)\n if mixin_name in cls.MIXIN_IMPORT_PATHS\n ]\n
docstring_lines = [\n \"Generated at: <generated>\",\n
GENERATED_FILE_MARKER,\n *[f\"Source YAML: `{hint}`\" for hint in source_hints],\n
\"Edit model definitions in YAML and regenerate via OrmDtoGenerator.\",\n ]\n
header_lines = ['\"\"\"', *docstring_lines, '\"\"\"', \"\"]\n\n cross_module_imports =
cls._collect_cross_module_type_checking_imports(\n model_items, module_path=module_path,
context=context\n )\n typing_import = (\n \"from typing import
TYPE_CHECKING, ClassVar\"\n if cross_module_imports\n else \"from typing
import ClassVar\"\n )\n import_lines = [\n \"from datetime import date,
datetime, time\",\n typing_import,\n \"import uuid\",\n \"\",\n
\"import sqlalchemy as sa\",\n \"from pydantic import JsonValue\",\n
\"from sqlalchemy.dialects import postgresql\",\n \"from sqlalchemy.orm import Mapped,
mapped_column, relationship\",\n \"from pgvector.sqlalchemy import Vector\",\n
\"\",\n *mixin_import_lines,\n \"from src.core.tsm_db.database import
Base\",\n ]\n if mixin_import_lines:\n import_lines.append(\"\")\n
if cross_module_imports:\n import_lines.extend([\n \"if
TYPE_CHECKING:\",\n *[f\" from {mod} import {cls_name}\" for mod, cls_name in
cross_module_imports],\n \"\",\n ])\n lines = [line for line in
header_lines + import_lines if line is not None]\n\n for model_name, model_spec in
model_items:\n lines.extend(\n cls._render_model_class(\n
model_name,\n model_spec,\n context=context,\n
)\n )\n lines.extend([\"\", \"\"])\n\n return
\"\\n\".join(lines).rstrip() + \"\\n\"")
⎿ {
"result": "OK"
}
● Now verify it runs correctly:
● Bash(cd **********************backend && source .venv/bin/activate.fish 2>/dev/null;
.venv/bin/python -m src.core.codegen.orm_dt…)
⎿ Wrote 50 generated model files.
● Generator runs clean with 50 files. Let me verify the TYPE_CHECKING block appears correctly in a generated file with cross-module
relationships:
Searched for 1 pattern, read 1 file (ctrl+o to expand)
● That's a legacy file from the old generator. Let me check one from our new generator:
Read 1 file (ctrl+o to expand)
● TYPE_CHECKING, ClassVar in the typing import, and the if TYPE_CHECKING: block correctly importing User from the user module. Now
update STATUS and commit:
● Update(*************************/STATUS_plan-9.md)
⎿ Added 1 line, removed 1 line
141
142 ### Component: Generator enhancement — `TYPE_CHECKING` imports
143
144 -- [ ] P9-6.4h Add `TYPE_CHECKING` imports to `_render_model_module` in `orm_dto_generator.py` — generated ORM files em
-it `Mapped["User"]` etc. with no import; fix: build `class_name→module_path` lookup from `context.model_by_impl`, coll
-ect cross-module relationship targets per module, emit `from typing import TYPE_CHECKING, ClassVar` + `if TYPE_CHECKIN
-G:` import block; affects all generated ORM files (Python 3.14)
144 +- [x] P9-6.4h Add `TYPE_CHECKING` imports to `_render_model_module` in `orm_dto_generator.py` — added `_collect_cross_
+module_type_checking_imports` helper; generator now emits `TYPE_CHECKING, ClassVar` + `if TYPE_CHECKING:` block for cr
+oss-module relationship targets; generator ran clean (50 files); verified in `workspace_orm.py` (`User` import)
145
146 ### Component: Generated ORM models and Pydantic schemas (`orm/`, `dto/`)
147
● Bash(cd **********************************&& git add backend/src/core/codegen/orm_dto_generator.py
backend/src/**********************************o…)
⎿ M backend/src/core/codegen/orm_dto_generator.py
M backend/src**********************************STATUS_plan-9.md
M backend/src/core/**********************************_orm.py
… +19 lines (ctrl+o to expand)
● Good — the regenerated ORM files show the TYPE_CHECKING additions. Stage and commit everything together:
● Bash(git add backend/src/ && git commit -m "$(cat <<'EOF'
plan-9(P9-6.4h): generator TYPE_CHECKING imports for cross-module relationships…)
⎿ [dev aef0f8c0] plan-9(P9-6.4h): generator TYPE_CHECKING imports for cross-module relationships
22 files changed, 239 insertions(+), 375 deletions(-)
● aef0f8c0. P9-6.4h done. Next is P9-6.13 — starting make dev-up to verify what currently breaks (BRK-2 import failures expected). Shall
I proceed?
✻ Cogitated for 3m 42s
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
❯
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
🔋:2.2k/90.3k | 💰:2.47/50.00 | 📨:22/746 | ⌛️:4h 16m | 🤖:sonnet(Sonnet 4.6)
⏵⏵ accept edits on (shift+tab to cycle)
14:49 <img width="1242" height="286" alt="Image" src="https://github.com/user-attachments/assets/5e172291-9514-422a-a1be-2b2e7af8f673" />
### What Should Happen?
FIx the random high usage problems
### Error Messages/Logs
```shell
Steps to Reproduce
- Run Claude Code with --model sonnet.
- Set /effort auto
- Hit the 5h limit within a session.
- Press continue
- Cry
Claude Model
Sonnet (default)
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
v2.1.92
Platform
Anthropic API
Operating System
Ubuntu/Debian Linux
Terminal/Shell
Other
Additional Information
_No response_
This issue has 11 comments on GitHub. Read the full discussion on GitHub ↗