Repository navigation
agent: skill tool loops on invalid id instead of falling back #61198
Description
Summary
When a user asks which skills are available, the agent attempts to load a skill with id general, which is not a valid skill id (valid ids: caveman, leanrea, opencode, report). The call fails with "Unable to load skill general", and instead of falling back to the skill list already present in context, the agent re-issues the identical failing call ~22 times in a row, never recovering. This is an agent-loop / tool-retry failure: a deterministic error result does not terminate the loop or trigger a fallback.
Environment
- opencode version: 2.2.5
- OS: Windows 11 (Windows_NT 10.0.26300, win32 x64)
- Terminal: TERM=xterm-256color (TERM_PROGRAM unset)
- Shell: Git Bash (C:\Program Files\Git\usr\bin\bash.exe); ComSpec cmd.exe
- Install/channel: latest
- Model: ollama/lfm3-omni:2b (LFM3 Omni 2B, 32k context) — local via ollama
- Active plugins: none found in config (MCP server
convexconfigured; no plugins)
Reproduction
- Configure and select the local model
ollama/lfm3-omni:2b(LFM3 Omni 2B). - Start an opencode session with that model.
- Ask: "what skills you have".
- Observe the agent call
skillwith{"id": "general"}. - The tool returns
Unable to load skill general. - The agent repeats the identical call many times (~22) instead of answering from context or stopping.
Expected Behavior
The agent should not call skill with an id not present in the available-skills list. On a deterministic skill-load failure, the loop should terminate and fall back to the in-context skill list (or a text answer), rather than retrying the same failing call.
Actual Behavior
Repeated skill tool calls with id: "general", each returning Unable to load skill general, forming a retry storm until the session ends. No useful answer is produced.
Additional Context
- The string
generalcorresponds to an agent/subagent type, not a skill id; this likely caused the confusion. - The available skills were already enumerated in the system context, so no tool call was necessary.
- Session export attached by user (session-02cfc45e-de13-4e23-a741-81dedf339c3c.md).
- Repro frequency: observed once from this session, specifically while using
ollama/lfm3-omni:2b. Likely more common with small/local models that confuse agent vs. skill namespaces; not confirmed on larger models. - Workaround: user must interrupt the session; no in-app recovery observed.
Activity
opencodejr commentedon Nov 2, 2026
Thanks for the detailed report and session export, @RoboXGamer!
I think this is a duplicate of #59339. In v2, the
doom_looppermission is still accepted in config but nothing currently triggers it. So when a model keeps re-issuing the same tool call with the same input, as happened here withskilland{"id": "general"}, nothing stops the loop.Possible duplicates and closely related issues, ordered by likelihood:
doom_loop, so repeated identical tool calls run indefinitely (same root problem)Unable to load skill <name>gives no hint when the name is a registered agent. This is thegeneralagent/skill mix-up that started your loop.There's also an open PR, #53032, which aims to stop repeated tool-call loops.
The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different.
opencodejr commentedon Nov 3, 2026
Closing this as a duplicate of #59339. We left it open for a day in case it turned out to be a different problem, and there's been no reply since.
The repeated
skillcalls with{"id": "general"}happen because v2 has no repeated-tool-call (doom_loop) detection yet, and that's what #59339 tracks. Please follow it there, along with the open PR #53032. The mix-up between thegeneralagent and a skill id is tracked in #46568.Thanks again for the report, @RoboXGamer. If you think this is a different problem, reply here and we can reopen it.