Skip to content

fix(sparc-service): strip agent-injected keys before SPARC evaluates tool calls #738

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
vz-ibm wants to merge 2 commits into
base: main
Choose a base branch
from
+45 −0
Open

fix(sparc-service): strip agent-injected keys before SPARC evaluates tool calls #738

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
691da50
feat(sparc-service): add SPARC_LOG_REQUESTS and SPARC_STRIP_TOOL_ARG_…
vz-ibm Jul 19, 2026
91262d0
fix(sparc-service): configure root logger so application log.info() o…
vz-ibm Jul 19, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions authbridge/sparc-service/sparc_service/__main__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,8 @@


def main() -> None:
import logging
logging.basicConfig(level=logging.INFO)
Comment on lines +11 to +12

@coderabbitai coderabbitai Bot Aug 5, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

rg -n -C 5 \
  'basicConfig|sys\.stdout|sys\.stderr|stream=|log_config' \
  authbridge/sparc-service

Repository: rossoctl/cortex

Length of output: 1023

🌐 Web query:

Python logging.basicConfig default stream stderr documentation

💡 Result:

Yes, the Python logging.basicConfig function defaults to using sys.stderr [1]. When basicConfig is called without a filename argument, it configures the logging system by creating a StreamHandler [2][1]. If the optional stream keyword argument is not provided to basicConfig, this StreamHandler defaults to sys.stderr [1]. This behavior is inherited from the StreamHandler class itself, which defaults to sys.stderr when no stream is explicitly specified [3][1]. If you wish to send logs to standard output instead, you can explicitly pass sys.stdout to the stream argument [4]: import logging import sys logging.basicConfig(stream=sys.stdout)

Citations:

Route application logs to stdout.

logging.basicConfig() defaults to sys.stderr, so the sparc_service.logger call in this module does not meet the stdout contract. Add stream=sys.stdout, or route this logger through Uvicorn’s logging configuration. Verify the deployed entry point uses loguru or sparc_service.logger directly.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@authbridge/sparc-service/sparc_service/__main__.py` around lines 11 - 12,
Update the logging setup in the module-level initialization around
logging.basicConfig so application logs are explicitly routed to sys.stdout by
configuring its stream, while preserving the existing INFO level. Ensure the
deployed entry point’s direct sparc_service.logger usage follows this stdout
configuration rather than defaulting to stderr.

settings = Settings.from_env()
uvicorn.run("sparc_service.api:app", host=settings.host, port=settings.port, log_level="info")

Expand Down
43 changes: 43 additions & 0 deletions authbridge/sparc-service/sparc_service/api.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,9 @@

from __future__ import annotations

import json
import logging
import os

from fastapi import FastAPI, HTTPException
from fastapi.concurrency import run_in_threadpool
Expand All @@ -19,6 +21,37 @@

log = logging.getLogger(__name__)

# When SPARC_LOG_REQUESTS=true, log the full incoming ReflectRequest JSON so
# you can inspect exactly what the caller sends (useful for diagnosing
# unexpected tool argument keys). Disabled by default — payloads can be large.
_LOG_REQUESTS = os.getenv("SPARC_LOG_REQUESTS", "").strip().lower() in {"1", "true", "yes"}

# When SPARC_STRIP_TOOL_ARG_KEYS is set (comma-separated key names), those keys
# are removed from every tool_calls[].function.arguments JSON object before the
# request reaches SPARC. Use to drop agent-injected keys that are not in the
# tool spec and would cause SPARC to reject the call.
# Example: SPARC_STRIP_TOOL_ARG_KEYS=session_id,request_id
_STRIP_KEYS: frozenset[str] = frozenset(
k.strip() for k in os.getenv("SPARC_STRIP_TOOL_ARG_KEYS", "").split(",") if k.strip()
)


def _strip_tool_arg_keys(tool_calls: list[dict], keys: frozenset[str]) -> list[dict]:
"""Return a copy of tool_calls with the named argument keys removed."""
result = []
for tc in tool_calls:
fn = tc.get("function", {})
raw_args = fn.get("arguments", "")
Comment on lines +43 to +44

@coderabbitai coderabbitai Bot Aug 5, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

rg -n -C 8 \
  'class ReflectRequest|tool_calls|_strip_tool_arg_keys|model_copy' \
  authbridge/sparc-service --glob '*.py'

Repository: rossoctl/cortex

Length of output: 18484

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '--- api.py relevant section ---\n'
sed -n '1,85p' authbridge/sparc-service/sparc_service/api.py | cat -n

printf '\n--- api model/type definitions ---\n'
sed -n '1,80p' authbridge/sparc-service/sparc_service/models.py | cat -n

printf '\n--- tests around request bodies ---\n'
sed -n '36,75p' authbridge/sparc-service/tests/test_api.py | cat -n

python3 - <<'PY'
import json
from pathlib import Path

models = Path('authbridge/sparc-service/sparc_service/models.py').read_text()
for line_no, line in enumerate(models.splitlines(), 1):
    if 'tool_calls:' in line or 'tool_specs:' in line or 'tool call' in line:
        print(f'{models}:{line_no}:{line}')
PY

Repository: rossoctl/cortex

Length of output: 244

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- api.py relevant section ---'
sed -n '1,85p' authbridge/sparc-service/sparc_service/api.py | cat -n

printf '%s\n' ''
printf '%s\n' '--- api model/type definitions ---'
sed -n '1,80p' authbridge/sparc-service/sparc_service/models.py | cat -n

printf '%s\n' ''
printf '%s\n' '--- tests around request bodies ---'
sed -n '1,75p' authbridge/sparc-service/tests/test_api.py | cat -n

python3 - <<'PY'
from pathlib import Path

models = Path('authbridge/sparc-service/sparc_service/models.py').read_text()
for line_no, line in enumerate(models.splitlines(), 1):
    if 'tool_calls:' in line or 'tool_specs:' in line or 'tool call' in line:
        print(f'{models}:{line_no}:{line}')
PY

Repository: rossoctl/cortex

Length of output: 20440

Validate tool-call objects before accessing nested fields.

ReflectRequest.tool_calls accepts a list of dict[str, Any], but _strip_tool_arg_keys still reads tc.get("function") and then fn.get("arguments") outside the try. A malformed tool call such as {"function": null} or {"function": 1} can raise inside _strip_tool_arg_keys while stripping is enabled, before the endpoint’s error handling. Check that tc and tc["function"] are mappings before calling .get("arguments", ...) and treat the rest consistently.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@authbridge/sparc-service/sparc_service/api.py` around lines 43 - 44, Update
_strip_tool_arg_keys to validate that each tool call and its function value are
mappings before invoking .get on them. Handle malformed entries such as null or
scalar function values consistently with the existing stripping behavior, and
keep argument processing inside the endpoint’s established error-handling path.

try:
args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
if isinstance(args, dict):
args = {k: v for k, v in args.items() if k not in keys}
new_args = json.dumps(args) if isinstance(args, dict) else raw_args
except (json.JSONDecodeError, TypeError):
new_args = raw_args
result.append({**tc, "function": {**fn, "arguments": new_args}})
return result


def create_app(engine: ReflectionEngine | None = None) -> FastAPI:
"""Build the FastAPI app. Inject ``engine`` in tests; defaults to env config."""
Expand Down Expand Up @@ -51,6 +84,16 @@ def readyz() -> dict[str, object]:

@app.post("/reflect", response_model=ReflectResponse)
async def reflect(request: ReflectRequest) -> ReflectResponse:
if _LOG_REQUESTS:
log.info("incoming reflect request: %s", request.model_dump_json())

if _STRIP_KEYS and request.tool_calls:
request = request.model_copy(
update={"tool_calls": _strip_tool_arg_keys(request.tool_calls, _STRIP_KEYS)}
)
if _LOG_REQUESTS:
log.info("after strip (%s): tool_calls=%s", sorted(_STRIP_KEYS), request.tool_calls)

# SPARCReflectionComponent.process is synchronous (and CPU/IO bound on the
# LLM call); run it off the event loop so the service stays responsive.
try:
Expand Down
Loading