[Bug]: forge doesn't get context from the command I run

[syabro]

Bug Description

~/Downloads
❯ rm test                                                                                                                                󱙺 FORGE  gpt-5.4-mini
rm: cannot remove 'test': Is a directory

~/Downloads
❯ : fix it                                                                                                                               󱙺 FORGE  gpt-5.4-mini
⏺ [11:11:07] Initialize 729a620c-de9b-4d67-a287-b8af9efab6db
Investigating the task

I'm trying to figure out what "fix it" means without any context. I need to explore the codebase first, and maybe I'll need to ask for clarification later. There might be an issue in the current folder, or perhaps something in Downloads that's related. Since the user said to "fix it," I need to identify what "it" refers to, like a failing test or build. I'll look into the directories using the tools available to see what I can find.Deciding on command usage

Steps to Reproduce

log above:

• new session
• rm <dirnane>
• ask forge to gix it

Expected Behavior

context passed, forge can do something about it

Actual Behavior

context not passed, forge doesn't know about my commands

Error Logs

Forge Version

2.1.0

Operating System

Linux (Arch)

OS Version

No response

AI Provider

OpenAI

Model

gpt-5 mini

Installation Method

Other

Configuration

[tusharmath]

Thanks @syabro this would be a great feature to have. Im thinking of approaches:

Information to pass on

1. Pass on the user initiated command (Can be done by accessing history)
2. Pass on the user initiated command + the result of the command (Will require a PTY implementation)

Ways to pass on

1. Auto insert with user prompt
2. Provide a tool for the agent to read

Considerations

• It should be configurable, not everyone would like to give access to every command that is being executed.

Am I missing something?

[tusharmath]

Support PTY would not be easy and im not sure if it would be possible to display the output and simultaneously also store it for later use.

I think a simple solution here would be to do it via our ZSH plugin. We already have a hook to intercept the commands. We can store if the output of the last command was successfull or not in a variable. this information can then be passed on to the forge-cli as a parameter and rendered via its user prompt.

[syabro]

@tusharmath sorry, but I thought it works like Warp

from the docs https://forgecode.dev/docs/zsh-support/

...
# Run your tools as part of the investigation
ps aux | grep server
htop -p $(pgrep server)

# Continue with full context
: now optimize the memory allocations you identified in the server struct

[tusharmath]

The example was meant to express that between the two commands that were prefixed with the : char, there is continuity. That being said I think we can build a WARP like experience for sure where we can provide additional context to the agent about the last command.

[zhaog100]

/attempt

[soongyintong]

✅ Context Fix Guide Submitted!

PR: #2664

Problem Analysis

Issue: When user runs commands outside forge, then says "fix it", forge has no context.

Root Cause: Forge only tracks:

• Messages within forge conversation
• Tool calls made through forge
• Explicitly attached files

Forge does NOT track:

• External shell history
• Commands run outside forge
• Error messages from terminal

Solution Options

1. Add Shell History to Context (Recommended)

• Read recent commands from .bash_history, .zsh_history, etc.
• Add as droppable context message
• Security: filter sensitive patterns

2. Improve Error Messages

• Guide forge to ask for clarification
• Example: "Could you share what command you ran and what error you saw?"

3. Add :context Command

• Explicit command to capture shell context
• User workflow: forge :context fix it

Files

• CONTEXT_FIX_GUIDE.md (192 lines)
• Includes code samples, test cases, security considerations

Ready for review! 🎉