anomalyco
Agent repeatedly re asks same question after answer is provided, fails to update plan after Glob result
Description
Environment
OS: Garuda Linux (Arch based)
Kernel: 6.18.4-zen1-1-zen
Shell: fish
OpenCode version: 1.1.25
Install path: /usr/bin/opencode
Interface: local CLI
Model: Dolphin-Mistral-24B Venice
Backend: llama.cpp
Summary The OpenCode Build agent enters a question loop and fails to update its execution plan after receiving a correct answer from the user. Even after the correct file extension is explicitly provided, the agent re asks the same question and does not proceed with the workflow.
This appears to be an agent state or planner synchronization bug, not a model misunderstanding.
Steps to reproduce
Run a Build workflow that attempts to process Substack export files.
Initial workflow attempts:
✱ Glob "substack_posts/*.md"
No markdown files are found.
User clarifies that files are HTML.
Agent performs:
✱ Glob "substack_posts/*.html"
Agent asks:
What is the correct file extension for the Substack posts?
User answers:
.html
Agent repeats the same question instead of proceeding.
Observed transcript excerpt
✱ Glob "substack_posts/.md" ✱ Glob "substack_posts/.md" No markdown files were found in the substack_posts directory.
they are html files ✱ Glob "substack_posts/*.html"
What is the correct file extension for the Substack posts? .html
✱ Glob "substack_posts/*.html" What is the correct file extension for the Substack posts? .html
The question is repeated multiple times even after the correct value is provided. The agent never advances to selecting a file, reading contents, or executing the requested transformation.
Expected behavior After the user answers .html once, the agent should:
Update its internal plan state
Treat .html as the confirmed extension
Proceed to select a file and continue the workflow
Actual behavior
Agent continues to ask the same clarification question
Planner does not converge
Workflow stalls indefinitely unless manually interrupted
Notes
The directory contains hundreds of valid .html files and they are discoverable via glob.
This appears to be a planner or agent control layer bug where updated facts from user answers are not propagated into execution state.
Similar to other agent non convergence or tool state issues, but this specific loop does not appear to be documented yet.
Also tried:
Substack post title and abstract extraction workflow
Files are in /home/innomen/AI/assistant/Substack Posts/*.html
- Select the first file matching "Substack Posts/*.html".
- Read its contents.
- Strip all HTML tags.
- Extract the post title from the
tag. If no <title> tag exists, use the first <h1> as the title.</h1> - Summarize the content to create an abstract.
- Write a new text file in the same directory named after the post title, containing the title and abstract.
- Do not ask for the file extension; assume it is .html.
✱ Glob "substack_posts/*.html"
→ Asked 1 question
▣ Build · dolphin-mistral
/init
create/update AGENTS.md /review
review changes [commit|branch|pr], defaults to uncommitted /undo
undo the last message /redo
redo the last message /compact
compact the session /rename
rename session /copy
copy session transcript to clipboard /export
export session transcript to file /timeline
jump to message /fork
fork from message /thinking
toggle thinking visibility /share
share a session /new
create a new session /models
list models /agents
list agents /session
list sessions /status
show status /mcp
toggle MCPs /theme
toggle theme /editor
open editor /connect
connect to a provider /help
show help /commands
show all commands /exit
exit the app What is the correct file extension for the Substack posts?
Build .html Dolphin-Mistral-24B Venice llama.cpp Dolphin HTML files 2. .md esc interrupt Markdown files 3. Type your own answer ↑↓ select enter submit esc dismis
Plugins
No response
OpenCode version
No response
Steps to reproduce
No response
Screenshot and/or share link
No response
Operating System
No response
Terminal
No response
