Repository navigation
It keeps scanning files and getting stuck #32
Description
Activity
Status update (2026-07-19): the repeated sentence in the report —
The task is not yet complete. Continue working.— is A3S Code's synthetic continuation prompt, not user input.Current Code releases already have bounded continuation counts, duplicate-tool-call guards, and repeated no-progress convergence (the latter shipped in v5.1.0 via
e6d8f9f). A remaining concrete gap was that a standalone greeting still received the core filesystem/tool surface, and a friendly response containing phrases such asI'll ...could be misclassified as incomplete and trigger a synthetic second turn.A3S-Lab/Code#15 addresses that narrow path:
- exact standalone English/Chinese greetings receive no tool definitions;
- the same turns do not receive a synthetic continuation;
- a greeting that also includes an action request still receives the normal tools;
- both Code CI jobs are green, including core Linux/Windows checks and SDK alignment/builds.
Keep this issue open until Code A3S-Lab/Box#165 is human-merged and the reporter confirms behavior on a current version. The original report also says “any input without a clear action command”; that broader class is intentionally not guessed by the exact greeting guard and needs a redacted session trace, A3S/Code version, provider/model, initial user message, and tool-event sequence if it still reproduces.
Release update (2026-07-19): the standalone-greeting fix from A3S-Lab/Code#15 was merged as
ef3b1cfand is contained in A3S Code v6.0.0.Please retest the original greeting/no-action case with Code v6.0.0 or newer. I am keeping this issue open because the broader “any input without a clear action command” claim still needs a redacted reproducer if it persists: version, provider/model, initial message, and ordered tool/continuation events.
The reported standalone greeting loop is fixed on current main and has shipped since v6.0.0. Standalone English and Chinese greetings receive no tool definitions and cannot trigger the synthetic continuation turn, while greetings containing an action retain tools. Verified at b479b6c: standalone_greetings_do_not_expose_tools and test_standalone_greeting_does_not_trigger_continuation both pass. If a broader non-greeting input still reproduces on v6.7.0 or newer, please open a new issue with a redacted prompt and ordered tool/continuation trace.
When I send a simple greeting or any input without a clear action command, A3S starts scanning my folders, gets stuck in an endless loop, and keeps burning through tokens.
So SafeClaw is definitely installed and seems to be an a3s-based workflow editor with skills, agents, and workflow capabilities. The user keeps saying "The task is not yet complete. Continue working." - maybe they want me to: 1. Explore the SafeClaw installation further 2. Help troubleshoot something related to SafeClaw 3. Do something with the workflow editor Let me also check the skills directory to see what skills are available. Actually, I think I should report my findings to the user in Chinese since that's the language they've been using. I've done quite a bit of exploration now and have discovered: 1. Working directory: /Users/rezin/.a3s (the A3S project itself) 2. SafeClaw is installed at ~/Library/Application Support/safeclaw/ 3. There are skills in ~/.safeclaw/skills/ 4. The dev/safeclaw directory contains Python skill code Maybe I should summarize my findings and offer to help with something specific. But the user keeps saying "The task is not yet complete. Continue working." - maybe they want me to keep exploring without stopping to explain. Let me continue exploring the skills directory and see what else I can find. Now I have a better picture of SafeClaw: 1. ~/.safeclaw/skills - Contains skills for SafeClaw 2. ~/Library/Application Support/safeclaw/ - Contains: • agent-sessions/ - Agent sessions • ahp.sock - Socket file • app-config.sqlite - SQLite database • config.hcl - Configuration file • triggers.db - Triggers database • workflows.db - Workflows database 3. ~/dev/safeclaw/ - Contains: • agents/ - Agents directory • logs/ - Logs • sessions/ - Sessions • skill_utils.py , skill.py , test_skill.py - Python skill files So SafeClaw is definitely installed and seems to be an a3s-based workflow editor with skills, agents, and workflow capabilities. The user keeps saying "The task is not yet complete. Continue working." - maybe they want me to: 1. Explore the SafeClaw installation further 2. Help troubleshoot something related to SafeClaw 3. Do something with the workflow editor Let me also check the skills directory to see what skills are available. Actually, I think I should report my findings to the user in Chinese since that's the language they've been using. I've done quite a bit of exploration now and have discovered: 1. Working directory: /Users/rezin/.a3s (the A3S project itself) 2. SafeClaw is installed at ~/Library/Application Support/safeclaw/ 3. There are skills in ~/.safeclaw/skills/ 4. The dev/safeclaw directory contains Python skill code Maybe I should summarize my findings and offer to help with something specific. But the user keeps saying "The task is not yet complete. Continue working." - maybe they want me to keep exploring without stopping to explain. Let me continue exploring the skills directory and see what else I can find. Now I have a better picture of SafeClaw: 1. ~/.safeclaw/skills - Contains skills for SafeClaw 2. ~/Library/Application Support/safeclaw/ - Contains: • agent-sessions/ - Agent sessions • ahp.sock - Socket file • app-config.sqlite - SQLite database • config.hcl - Configuration file • triggers.db - Triggers database • workflows.db - Workflows database 3. ~/dev/safeclaw/ - Contains: • agents/ - Agents directory • logs/ - Logs • sessions/ - Sessions/