IssueLoop

Finds failing tests (batch or live). Splits failures into independent tickets via an LLM, with automatic priority fallback across providers. Stores them. Hands them out one at a time. Optionally proposes and applies fixes under an explicit, auditable permission allowlist.

---

Why this exists

Bugs sit unfixed because someone has to notice them first. IssueLoop reads the test suite directly — a failing test is the bug report. It can also watch a live-running process or log file, for bugs that only show up outside a test run. No human has to spot it, and no human has to manually triage which failures are actually distinct problems.

Install

Not yet on PyPI — install from git:

python3 -m venv .venv
source .venv/bin/activate
pip install "git+https://github.com/onenot8/issueLoop@phase-2"

Or for local development:

git clone https://github.com/onenot8/issueLoop
cd issueLoop
pip install -e ".[dev]"

Local SQLite backend, zero setup. No database account, no API key required to try it.

Quickstart

import issueloop

issueloop.use(
    database="local",                     # or "supabase" for a hosted backup
    llm={"providers": [                   # priority list — falls back automatically
        {"provider": "anthropic", "model": "claude-sonnet-4-6", "api_key": "sk-..."},
        {"provider": "ollama", "model": "qwen2.5-coder:7b"},   # free, local, default if omitted
    ]},
    notify={"webhook": "https://your-endpoint"},   # fires if IssueLoop itself breaks
)

issueloop.scan_repo("repos/myrepo")
issueloop.run_tests("myrepo")
issueloop.create_tickets("myrepo")

ticket = issueloop.get_top_error("myrepo")   # claims one, marks in_progress
issueloop.resolve(ticket["id"])              # your side fixed it

Live monitoring

Watch a process IssueLoop spawns itself, or tail a log file an already-running process writes to:

from issueloop import watch_process, watch_log_file, stop_watch

handle = watch_process("myrepo", "python3 main.py", cwd="/path/to/repo", debounce_seconds=3.0)
# ...
stop_watch(handle)

Or from the CLI, in the foreground:

issueloop watch myrepo --command "python3 main.py" --cwd /path/to/repo --debounce 3.0
issueloop watch myrepo --log-file /path/to/app.log --debounce 3.0

Errors are debounced (default 3s of quiet) so a single multi-line traceback becomes one ticket, not dozens. Detected errors land in the same log format the batch test runner uses — issueloop tickets picks them up with no extra steps.

Fix-apply layer

Off by default — every command needs an explicit allowlist entry:

# config/permission.yaml
per_repo:
  myrepo:
    allowed_patterns:
      - 'sed -i .* somefile\.py'

issueloop.propose_fix(ticket_id, "sed -i 's/old/new/' somefile.py")
result = issueloop.apply_fix(ticket_id)
# {"status": "resolved" | "retry" | "escalated" | "denied", ...}

apply_fix re-runs the ticket's associated test after applying the fix and only resolves it if the test actually passes. Failures retry up to a configurable limit, then escalate to needs_human. Every permission decision is audited — issueloop.get_permission_audit_log().

Use it from the CLI

issueloop check-env
issueloop scan repos/myrepo
issueloop test myrepo
issueloop tickets myrepo
issueloop next myrepo
issueloop watch myrepo --command "..." | --log-file ...
issueloop cleanup --days 30

Use it from Node, Go, or anything else

issueloop serve --port 8787

const r = await fetch("http://localhost:8787/errors/top?repo=myrepo");
const ticket = await r.json();

serve is localhost-only, no auth — put it behind your own gateway if you ever expose it further.

Full API

Bug query — get_all_bugs, get_unresolved_bugs, get_resolved_bugs, get_failed_bugs, get_bugs_needing_human, get_bugs_by_status, get_bugs_by_priority, get_bug, get_bug_count, get_bug_count_by_status, search_bugs, get_oldest_bug, get_newest_bug

Lifecycle — get_top_error, get_all_errors, resolve, fail, escalate, reassign, retry_bug, get_bug_attempts, bulk_resolve

Fix-apply — propose_fix, apply_fix, check_permission, get_permission_audit_log

Scan / test — scan_repo, get_file_inventory, run_tests, run_single_test, create_tickets

Live monitoring — watch_process, watch_log_file, stop_watch, list_active_watchers

LLM / tokens — get_token_consumption, get_token_consumption_by_provider, get_llm_call_history, get_llm_provider_status

Database — cleanup, purge_repo, get_database_stats, export_bugs, reap_stale_bugs, rotate_logs

Notifications — get_crash_log, get_notification_config

Config / utility — use, get_config, list_repos, health_check

Config — every option

Field	Values	Default
database	"local", "supabase"	"local"
database_path	any path	data/issueloop.db
retention_days	int	30
llm.providers	list of provider dicts, priority order	single ollama default
llm.provider	"ollama", "anthropic", "openai"	"ollama"
llm.apiKey	string	none (required for anthropic/openai)
llm.tokenSize	int	1024
notify.webhook	URL	none

Both apiKey/api_key and tokenSize/token_size work — camelCase or snake_case, your call. A single llm={...} dict still works exactly as before; wrap multiple in llm={"providers": [...]} for automatic fallback.

Config files (config/permission.yaml, config/provider_config.yaml) resolve in this order: an explicit env var (ISSUELOOP_PERMISSION_PATH / ISSUELOOP_PROVIDER_CONFIG_PATH) → ./config/ relative to your current directory → config/ in the IssueLoop checkout → bundled package defaults. This means a real pip install (not just -e .) still works correctly even without a project-local config/ directory.

Why trust this

• Every module has a test, and the tests run against real behavior — a real SQLite file, a real HTTP server on a real socket — not just mocks pretending things work.
• Real bugs were caught and fixed by actually running this code, not by reading it — including a prompt that only ever sent stderr to the LLM triage step while pytest (and most test runners) report failures on stdout, and a config-bundling gap that made the permission system silently deny everything under a real (non-editable) install. Both have regression coverage now.
• Local-first by default. Nothing phones home, nothing requires an account to try.
• Read the code — it's small enough to actually read. That's on purpose. File-by-file build order: BUILD_ORDER.md.

Testing

pytest tests/ -v

Full walkthrough, including live monitoring and the fix-apply layer: TESTING.md.

License

MIT — see LICENSE.