v2: scrum-master execution for Claude Code — parallel dev agents, review gates, and a live board. Your AI ships multi-phase plans start to finish, no matter how many sessions it takes.
Install in 10 seconds ↓When an AI agent's context window fills up, the system compresses earlier messages and the agent keeps working on subtly wrong assumptions. On big plans the damage compounds: phases run one at a time even when they're independent, and the agent burns its own context reading source files until its early decisions turn to fog.
The agent re-asks things it already resolved an hour ago, wasting time and breaking flow.
Implementations silently deviate from decisions made earlier in the session when those decisions get compressed away.
The agent breaks things it just built because it forgot why they were built that way.
v2 turns your session into a scrum master. It dispatches parallel dev agents with disjoint file ownership, reviews every phase before merging, ships as it goes, and keeps one live board so any session can resume from two file reads.
Create or import an implementation plan
Parallel dev agents, each owning its files
Adversarial gate, then ship while the next sprint runs
Live state; any session resumes from two reads
One BOARD.md per plan holds phase states, commit SHAs, decisions, and learnings with exact file:line references. Updated at every sync point, so a crash loses one sprint at most.
scripts/board-check.js verifies every merged claim, the progress count, the ship SHA, and the door state against git itself. A board that lies blocks dispatch until it's fixed.
Lite for half-day jobs (one file, one gate), solo for small serial plans, scrum for parallel agent orchestration. Route by flag or by scope.
The strongest model orchestrates, one tier down develops, rote lanes go two tiers down, and the riskiest review gates use the orchestrator's own model. Every model earns its cost.
Every phase is adversarially reviewed before its merge. Fixes go back to the same warm agent; two failed rounds parks the phase for a human instead of looping.
Install with a single command. No daemons, no setup wizards. Plans and boards are markdown in your repo, plus one small audit script.
End-of-day sweep: parallel reviewers over the day's commits, confident fixes applied, tests, deploy through your release path, summary for the morning. Fully unattended.
Sessions run in shifts that end at clean board boundaries. The next session resumes with full acuity instead of a compacted one guessing at what it forgot.
Already discussing architecture? Use --from-context to distill the conversation into a structured plan. Add --now to skip confirmations.
Completed plans stay useful. Bug reports get structured debugging with maintenance logs, not ad-hoc guessing.
Live state, not an archaeology dig. This is everything a fresh session reads to take over, and the checker verifies it against git before any work is dispatched.
Runs sprint after sprint and keeps the board current. Invoke it bare to resume the most recent active plan; go status just reports.
/justkeepswimming:go my-feature
Pauses at sprint boundaries. Review results, adjust the plan, decide when to continue.
/justkeepswimming:go my-feature --interactive
| Just Keep Swimming | Heavy workflow systems | |
|---|---|---|
| Installation | 1 plugin command | 30+ files |
| State tracking | One live board, audited against git | STATE.md, ROADMAP.md, CONTEXT.md, etc. |
| Concepts to learn | Plan, Board, Sprint, Merge window | Projects, Milestones, Phases, Waves, Audits |
| Parallelism | File-ownership lanes, review gates, one git owner | Unmanaged, or none |
| Setup time | 10 seconds | Plugin install + configuration |
| Dependencies | None required | Plugin ecosystem |
| Commands | 2 commands + a status call | 10+ specialized agents |
Most AI agents fail the same way: they pattern-match to common fixes and try them one by one. Just Keep Swimming ships with six problem-solving principles injected into every session — changing how the agent approaches problems.
Gather facts before forming hypotheses. Every minute understanding the problem saves ten minutes of wrong-direction fixes.
List every condition required for success and verify each one. The broken condition is always the one nobody checks because it's "too basic."
Run diagnostics with an open mind. Read the actual output — don't skim for confirmation of what you already believe.
Systems fail without useful errors all the time. They fall back, return generic messages, or silently skip the broken path. Find the quiet one.
Fix the root cause and only the root cause. The best fix is the smallest change. If you changed more than needed, you didn't understand the problem well enough.
If the user has tried everything, the obvious solutions are exhausted. The answer lives in the layer nobody checked yet. Don't re-run the playbook.
Install as a plugin or copy two files. No dependencies.
/plugin marketplace add cyber-qais/justkeepswimming
/plugin install justkeepswimming
Open any Claude Code session and run:
/justkeepswimming:go my-feature
mkdir -p ~/.claude/commands/justkeepswimming
cp commands/justkeepswimming/*.md ~/.claude/commands/justkeepswimming/
npm install justkeepswimming