Loki Mode - OpenClaw Skill
When to use
- User asks to "build", "implement", or "develop" a feature from a PRD
- User provides a requirements document and wants autonomous execution
- User says "loki mode" or references autonomous development
- User wants to run a full SDLC cycle on a codebase
Prerequisites
loki CLI installed on the host (via npm install -g loki-mode or Homebrew)
- One of: Claude Code or opencode installed
- Corresponding API key set (ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY)
How to invoke
Start a session
Use the bash tool with background mode:
bash(command: "loki start <prd-path> --bg --yes --no-dashboard", pty: true, background: true, workdir: "<project-dir>")
Key flags:
--bg: Background mode (session outlives the tool call)
--yes: Skip confirmation prompts
--no-dashboard: Avoid port conflicts in sandboxed environments
--provider <claude|opencode>: Select AI provider (default: claude)
--budget <amount>: Set cost limit in USD (auto-pause when exceeded)
Monitor progress
Poll status every 30 seconds:
bash(command: "loki status --json", workdir: "<project-dir>")
The JSON output contains:
version: Loki Mode version string
status: inactive, running, paused, stopped, completed, unknown
phase: Current SDLC phase (e.g., BOOTSTRAP, DISCOVERY, ARCHITECTURE, DEVELOPMENT, QA, DEPLOYMENT)
iteration: Current iteration number
provider: Which AI provider is active (claude, opencode)
pid: Process ID of the running session (null if not running)
elapsed_time: Seconds since session start
dashboard_url: URL of the web dashboard (null if disabled)
task_counts: Object with total, completed, failed, pending counts
For budget tracking (not in JSON output), read the budget file directly:
bash(command: "cat .loki/metrics/budget.json 2>/dev/null || echo '{}'", workdir: "<project-dir>")
Budget JSON fields: budget_limit, budget_used
Report progress to channel
After each poll, summarize changes:
- Phase transitions ("Moved from ARCHITECTURE to DEVELOPMENT")
- Task completion counts ("12/20 tasks complete, 0 failed")
- Elapsed time ("Running for 45 minutes")
- Error states that need attention (failed tasks > 0, status is unknown)
If budget tracking is active, include cost in updates:
- "Estimated cost: $4.50 / $50.00 budget"
Control commands
- Pause:
bash(command: "loki pause", workdir: "<project-dir>")
- Resume:
bash(command: "loki resume", workdir: "<project-dir>")
- Stop:
bash(command: "loki stop", workdir: "<project-dir>")
- Status:
bash(command: "loki status", workdir: "<project-dir>")
- Logs:
bash(command: "loki logs --tail 50", workdir: "<project-dir>")
Session complete
When status becomes "stopped" or "completed":
- Run
loki status --json for final summary
- Run
git log --oneline -20 to show commits made
- Report final task counts, elapsed time, and duration
- If council verdict exists, include it:
cat .loki/council/report.md
Critical rules
- ALWAYS use --bg flag (session must outlive the tool call)
- ALWAYS use --yes flag (no confirmation prompts in non-interactive channels)
- NEVER run loki in the OpenClaw workspace directory itself
- Poll status rather than watching stdout (background mode detaches)
- If session crashes, check
loki logs before restarting
- Respect budget limits -- include cost in every progress update when tracking is active
- The --no-dashboard flag is recommended to avoid port conflicts in sandboxed environments
1---2name: loki-mode-openclaw3description: "Launch Loki Mode autonomous SDLC agent. Handles PRD-to-deployment with minimal human intervention. Invoke for multi-phase development tasks, bug fixing campaigns, or full product builds."4metadata:5 source: https://skillmd.com/skills/asklokesh/loki-mode-openclaw6---78# Loki Mode - OpenClaw Skill910## When to use11- User asks to "build", "implement", or "develop" a feature from a PRD12- User provides a requirements document and wants autonomous execution13- User says "loki mode" or references autonomous development14- User wants to run a full SDLC cycle on a codebase1516## Prerequisites17- `loki` CLI installed on the host (via `npm install -g loki-mode` or Homebrew)18- One of: Claude Code or opencode installed19- Corresponding API key set (ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY)2021## How to invoke2223### Start a session24Use the bash tool with background mode:25```26bash(command: "loki start <prd-path> --bg --yes --no-dashboard", pty: true, background: true, workdir: "<project-dir>")27```2829Key flags:30- `--bg`: Background mode (session outlives the tool call)31- `--yes`: Skip confirmation prompts32- `--no-dashboard`: Avoid port conflicts in sandboxed environments33- `--provider <claude|opencode>`: Select AI provider (default: claude)34- `--budget <amount>`: Set cost limit in USD (auto-pause when exceeded)3536### Monitor progress37Poll status every 30 seconds:38```39bash(command: "loki status --json", workdir: "<project-dir>")40```4142The JSON output contains:43- `version`: Loki Mode version string44- `status`: inactive, running, paused, stopped, completed, unknown45- `phase`: Current SDLC phase (e.g., BOOTSTRAP, DISCOVERY, ARCHITECTURE, DEVELOPMENT, QA, DEPLOYMENT)46- `iteration`: Current iteration number47- `provider`: Which AI provider is active (claude, opencode)48- `pid`: Process ID of the running session (null if not running)49- `elapsed_time`: Seconds since session start50- `dashboard_url`: URL of the web dashboard (null if disabled)51- `task_counts`: Object with `total`, `completed`, `failed`, `pending` counts5253For budget tracking (not in JSON output), read the budget file directly:54```55bash(command: "cat .loki/metrics/budget.json 2>/dev/null || echo '{}'", workdir: "<project-dir>")56```57Budget JSON fields: `budget_limit`, `budget_used`5859### Report progress to channel60After each poll, summarize changes:61- Phase transitions ("Moved from ARCHITECTURE to DEVELOPMENT")62- Task completion counts ("12/20 tasks complete, 0 failed")63- Elapsed time ("Running for 45 minutes")64- Error states that need attention (failed tasks > 0, status is unknown)6566If budget tracking is active, include cost in updates:67- "Estimated cost: $4.50 / $50.00 budget"6869### Control commands70- Pause: `bash(command: "loki pause", workdir: "<project-dir>")`71- Resume: `bash(command: "loki resume", workdir: "<project-dir>")`72- Stop: `bash(command: "loki stop", workdir: "<project-dir>")`73- Status: `bash(command: "loki status", workdir: "<project-dir>")`74- Logs: `bash(command: "loki logs --tail 50", workdir: "<project-dir>")`7576### Session complete77When status becomes "stopped" or "completed":781. Run `loki status --json` for final summary792. Run `git log --oneline -20` to show commits made803. Report final task counts, elapsed time, and duration814. If council verdict exists, include it: `cat .loki/council/report.md`8283## Critical rules84- ALWAYS use --bg flag (session must outlive the tool call)85- ALWAYS use --yes flag (no confirmation prompts in non-interactive channels)86- NEVER run loki in the OpenClaw workspace directory itself87- Poll status rather than watching stdout (background mode detaches)88- If session crashes, check `loki logs` before restarting89- Respect budget limits -- include cost in every progress update when tracking is active90- The --no-dashboard flag is recommended to avoid port conflicts in sandboxed environments