AI coders aren't held back by models. They're held back by context engineering hell: hand-maintaining context memory files, re-explaining your repo every session, and watching institutional knowledge evaporate between branches. We built the Flux framework to fix the operational layer. It turns ad-hoc AI coding into session-based, Git-native workflows that auto-prime the model with the right code, docs, and prior lessons, track decisions as you work, and synthesize durable learnings when you're done - so the next session starts smarter by default. The result is less librarian work, fewer repeat mistakes, faster onboarding, and a development loop where your team's hard-won fixes actually compound over time.
Check out this 3-minute video on Flux Framework
AI coders are powerful—but we've been propping them up with manual context engineering:
CLAUDE.md) and hope the model reads them.That treadmill taxes your best engineers, slows onboarding, and makes repeated mistakes feel inevitable. The problem isn't the models - it's the operational layer around them.
The Flux addresses the shortcomings of the operational layer.
Flux replaces ad-hoc prompt rituals with learning-focused sessions: you start with a session goal, the framework primes the model with the right code/docs/lessons, spins up an isolated Git branch, tracks progress as you work, and on completion synthesizes lessons into durable org memory.
Your very next AI-coding session starts smarter - as if by magic (spoiler, it's not magic).
Session Start
→ Context Priming
→ Development Work (human + AI)
→ Updates (progress, git, decisions)
→ Completion
→ Knowledge Synthesis (lessons, docs index)
Each session begins with the goal, which the framework parses to pinpoint the relevant domains in your codebase. It then discovers the pertinent files, pulls in the right documentation and historical lessons, and automatically spins up an isolated Git branch for the work. With that, the model is primed with exactly the context it needs—no manual memory-file wrangling, no copy-paste rituals.
As you build, the system records what actually happened: code changes, decisions taken, and the challenges you hit along the way. It captures insights and emerging lessons in-line with the work, while keeping tabs on Git status and task completion so your progress stays auditable and easy to resume.
When you wrap, the framework generates a comprehensive session summary and promotes the new insights into project/org memory. Documentation gets richer with each pass, reflecting the improved understanding from the session. The net effect is a repo that "remembers", so future sessions initialize with stronger starting context and fewer repeated mistakes.
Works today with Claude Code; porting to other agents is mostly path/wrapper changes.
git clone https://github.com/carveragents/flux.git
cp -R flux/commands ~/.claude/
/session:start fix memory leak in data processing
/session:update
/session:end
- Produces a comprehensive session summary.
- Updates (or creates) `docs/LESSONS.md` with deduped, structured learnings.
- Builds/updates `docs/README.md` as a navigational index.
- Clears the active session pointer.
session_name from your goal (e.g., feat-auth-refactor).docs/README.md and docs/LESSONS.md (if present) to prime the model..claude/.sessions/<YYYY-MM-DD-HHMM>-<session_name>.md and marks it current.(--porcelain), task counts, issues, solutions.docs/LESSONS.md → merges new learnings with existing ones (no duplicates).docs/README.md → a pointer index to everything in docs/, never duplicating content..claude/.sessions/.current-session./git:commit:Conventional commits + emoji, suggests splitting logically distinct diffs./git:merge-cleanup:Merge current branch to main, push, and clean up.Start:
/session:start fix memory leak in data processing
- Branch: `bugfix-memory-leak-data-processing`
- Context: loads relevant data-processing files, prior lessons on memory/streaming (if any exist in the project documentation/memory files).
Development phase (with updates):
- You (human + AI) do the work as usual
/session:update
Update – 2025-09-24 14:40
Summary: Replaced list accumulation with generator-based streaming in the pipeline.
Git Changes:
Todos (2 completed, 1 pending):
Issues and Solutions:
End:
/session:end
Generates a full summary and updates:
Next time anyone starts a session touching data pipelines, the relevant lessons are automatically surfaced at /session:start.
flux/
├── README.md
├── LICENSE
└── commands/
├── git/
│ ├── commit.md
│ └── merge-cleanup.md
└── sessions/
├── current.md
├── end.md
├── help.md
├── list.md
├── start.md
└── update.md
docs/README.md, docs/LESSONS.md (if present).docs/README.md is a pointer map, not a dumping ground.docs/LESSONS.md without duplication.current-session is cleared (not deleted) on session end to avoid accidental orphan states./git:commit inspects diffs and session context to ensure atomic commits.All logic lives in the ‘command contracts’ (commands/.../*.md). To adapt:
git status work as-is.docs/LESSONS.md is repo-scoped; multi-repo knowledge roll-ups are a natural next step.
If you're tired of context engineering hell and want compounding AI-assisted development, try Flux on your next task.
GitHub Repo + Quick Start: Flux Framework