Autonomous deep-work sessions for OpenClaw. Set a topic, a duration, and walk away. Sprint spawns workers iteratively, detects stalls, tracks goals, and delivers a final synthesis report — all hands-off.

🟡 Status: Beta. Proven across 5+ production sprints including overnight autonomous runs. Ready for unsupervised use with the right goal structure — see Overnight Use below.

• Iterative worker spawning — Director runs every 20–30 min, spawns a worker subagent, waits for output
• Goal tracking — Define goals upfront via goals.json; Sprint tracks completion across iterations
• Stall detection — If a worker doesn't produce output in 610s, stall counter increments; 3 consecutive stalls triggers escalation
• Mutation support — Director can propose goal changes mid-sprint; mutations queue for operator ACK before acting
• Synthesis — When the sprint window closes, synthesis script reads all iter-N.md files and produces a SPRINT-REPORT.md
• Discord thread — Sprint auto-creates a thread for live progress updates

# OpenClaw
git clone https://github.com/jeremyknows/sprint-openclaw ~/.openclaw/skills/sprint-openclaw

Requirements:

• OpenClaw runtime with sessions_spawn support
• flock — director lock protocol (brew install util-linux on macOS)
• jq — state JSON parsing (brew install jq)
• Discord plugin configured (for thread creation + iteration updates)

Verify:

which flock && echo "✅ flock" || echo "❌ missing — brew install util-linux"
which jq && echo "✅ jq" || echo "❌ missing — brew install jq"

Run a 2-hour sprint on: research OAuth 2.1 hybrid auth patterns for VeeFriends
Sprint on implementing Redis rate limiting — 4 hours, high autonomy
Start a sprint: analyze our top 10 Discord messages from last week

# Basic sprint
bash ${CLAUDE_SKILL_DIR}/scripts/sprint-init.sh \
  --topic "Research React 19 performance patterns" \
  --duration "2h"

# With goals file and autonomy level
bash ${CLAUDE_SKILL_DIR}/scripts/sprint-init.sh \
  --topic "Implement Redis-backed API rate limiting" \
  --duration "6h" \
  --autonomy-level "high" \
  --goals /path/to/goals.json

Start at low. Promote after 1 successful sprint at each level.

# Quick status
cat ~/.openclaw/data/sprints/<sprint-id>/state.json | jq '{status, iterationCount, consecutiveStalls}'

# Read latest iteration
cat ~/.openclaw/data/sprints/<sprint-id>/iter-$(ls ~/.openclaw/data/sprints/<sprint-id>/iter-*.md | wc -l | tr -d ' ').md

# Watch director log
tail -f ~/.openclaw/data/sprints/<sprint-id>/director.log

~/.openclaw/
  data/
    sprints/
      <sprint-id>/
        state.json          # Live state (status, goals, stalls, mutations)
        goals.json          # Initial goals
        iter-1.md, ...      # Worker outputs per iteration
        director.log        # Director run history
        SPRINT-REPORT.md    # Final synthesis (on completion)

skills/sprint/
  SKILL.md
  scripts/
    sprint-init.sh          # Initialize sprint
    sprint-director.sh      # Cron-driven iteration controller
    sprint-synthesize.sh    # Final report generation
    sprint-worker-prompt.md # Worker subagent prompt template
    sprint-runbook.md       # Detailed operational procedures
  docs/
    GUIDE.md                # Goals schema, autonomy details, monitoring
    TROUBLESHOOTING.md      # Failure modes and recovery procedures

Sprint has been validated for unsupervised overnight runs. Two conditions matter:

Goals must be specific, not thematic. Each goal/iteration should map to one concrete deliverable — a file, a fix, a research finding. Open-ended goals like "improve the analytics dashboard" cause Iter-1 to stall on scoping rather than producing output. If you wouldn't know whether a goal was "done" at a glance, rewrite it.

Set MAX_ITERATIONS. Without a cap, a sprint on a wide topic can run indefinitely. Set it in goals.json or pass --max-iterations to sprint-init.sh. 4–6 iterations is a good overnight window.

{
  "maxIterations": 5,
  "goals": [
    "Audit /src/lib/metric-definitions.ts — list all missing Analytics/Retention entries",
    "Add missing MetricHelpButton entries for the top 5 Analytics metrics",
    "Fix HolderOverlapMatrix to respect the collections prop"
  ]
}

• flock required — Lock protocol silently fails without it; concurrent directors corrupt state.json
• No cost ceiling — Long high-autonomy sprints can spawn many subagents. Monitor costs actively.
• 610s worker timeout — Hard-coded. Long research tasks (many URLs, large codebases) will stall.
• Manual archiving — Sprint data accumulates in data/sprints/ indefinitely. Archive manually.
• Synthesis needs ≥1 iter — If all workers time out, synthesis produces an empty report.
• Thematic goals cause stalls — Workers spend Iter-1 scoping instead of producing. Write goals that are already scoped.

Full procedures: docs/TROUBLESHOOTING.md

v0.3.0 · MIT · jeremyknows