Save Your Progress
Your AI coding partner forgets everything. Bonfire remembers.
The Problem
AI agents are stateless. Every conversation starts from scratch. The agent doesn't remember what you decided yesterday, why you chose that architecture, what blockers you hit, or where you left off.
You end up re-explaining context, re-making decisions, and watching your AI partner repeat the same mistakes. I wrote about this in November--a pattern for maintaining context across sessions using Markdown files and slash commands[1].
The Solution
Bonfire is a plugin for Claude Code and OpenCode that maintains a living context document--read at session start, updated at session end. Your AI partner picks up exactly where you left off. It's like a saved game for your work.
/bonfire:startreads context/bonfire:endsaves contextThat's the core. No complex setup. No external services. Just Markdown files in your repo. It complements your issue tracker--use that for what needs to be done, use Bonfire for where you are and what you decided.
Beyond Start and End
The original pattern was two commands and a Markdown file[2]. The plugin does more--but every addition addresses the same insight: the work has shifted from writing to directing.
Claude Code and OpenCode can write code for twenty minutes and often get it right. What's expensive is everything around it--building the wrong thing, missing edge cases, losing context, discovering misalignment after the code is written. The cost moved from typing to thinking.
These features are responses to that shift:
- Session state--preserve context so you don't re-explain
- Spec interviews--research the codebase, then surface ambiguity before it derails[3]
- Work review--catch blindspots before they ship
- Subagents--isolate heavy operations so sessions last longer[4]
- Proactive skills--read context automatically, suggest archiving when done
All feedback loops. All addressing where the bottleneck actually is now.
/bonfire:start | Read context, scaffold on first run |
/bonfire:end | Update context, commit changes |
/bonfire:spec | Create implementation spec |
/bonfire:review | Find blindspots and quick wins |
/bonfire:document | Document a codebase topic |
/bonfire:archive | Archive completed work |
/bonfire:configure | Change project settings |
Install Bonfire
Claude Code
# Install the plugin claude plugin marketplace add vieko/bonfire claude plugin install bonfire@vieko # Start tracking your project /bonfire:start # Save your progress /bonfire:end
OpenCode
# Install the plugin (project or global) bunx opencode-bonfire install # Project install bunx opencode-bonfire install --global # Global install # Start tracking your project /bonfire-start # Save your progress /bonfire-end
Both platforms use the same .bonfire/ directory and CLAUDE.md file format. You can switch between Claude Code and OpenCode freely--your context travels with you.
Bonfire animation by Jon Romero Ruiz.
Commands shown use : in Claude Code and - in OpenCode.
[1] My colleague Aman suggested turning the pattern into a tool. I built create-sessions-dir first, which evolved into the Bonfire plugin.
[2] Adam Wolff from the Claude Code team put this well: “Most of the fancy stuff you can do with bespoke tools, custom agents, and elaborate skills can also be achieved with a prompt.”
[3] Inspired by Thariq's spec-based workflow--start minimal, have Claude interview you to resolve ambiguity, then execute in a fresh session. I added a research phase so questions are informed by codebase context.
[4] Boris Cherny, who created Claude Code, uses subagents for code review--parallel instances checking different aspects, then more to challenge the findings. The plugin applies this pattern for research and review, isolating heavy operations from the main conversation.
