Skill Pipelines

Route: /skills

Skill Pipelines integrate the cursor-kenji / skills.sh agent-skill ecosystem directly into Mushi triage. When a bug report is classified, Mushi recommends the most relevant Cursor agent skills — and you can attach a skill chain to any report, visualize it as a live React Flow pipeline, and execute it either as a handoff to your local Cursor agent or automatically via Cursor Cloud.

How it works


git repo skills/SKILL.md (cursor-kenji or any skills.sh-compatible repo)
  → skill-sync syncs SKILL.md files into the agent_skills catalog (daily cron)
  → classify-report Stage 2 recommends top 3 skills per report (pgvector similarity)
  → You start a pipeline run from the Report Detail page (or via CLI/MCP)
  → Mushi composes a "run packet" — skill instructions + full report context
  → Console renders the chain as a live React Flow pipeline
  → Steps update in real-time as your Cursor agent checks them in

The three tabs

Catalog

Browse the full skill catalog grouped by category (workflow, debug, test, audit, enhance, …). Search by name or description. Start a pipeline from any skill.

Pipelines

All pipeline runs for your project — past and active. Click a run to see the live React Flow visualization with step-level status (pending → running → passed / failed). Subscribe to real-time updates without refreshing.

Sources

Manage the GitHub repos whose SKILL.md files are synced into your project’s catalog. The default seed is kensaurus/cursor-kenji. Any skills.sh -compatible repo works.

Execution modes

Mode	Description	When to use
Handoff (default)	Mushi composes a run packet you paste into Cursor	Local development, free, full Cursor IDE experience
Cloud (opt-in)	Each step dispatches a Cursor Cloud agent run automatically	CI pipelines, async execution, PM-driven automation

Triage links (`?skill=`)

Share a pre-loaded triage link with any team member:


https://console.mushimushi.dev/reports/<report-id>?skill=workflow-fix-and-ship

When opened, the report detail page pre-selects the skill and prompts to start the pipeline. Useful for PMs handing off to devs, or for embedding in Slack/Linear notifications.

CLI quick reference


# Browse the skill catalog
mushi skills list
mushi skills list --category workflow
mushi skills list --search "fix bug"
 
# View full skill instructions
mushi skills show workflow-fix-and-ship
 
# Trigger skill sync (after adding a new source)
mushi skills sync
 
# Start a pipeline for a report
mushi pipeline start <reportId> --skill workflow-fix-and-ship
mushi pipeline start <reportId> --skill workflow-fix-and-ship --mode cloud
 
# Watch a pipeline run and print the context packet
mushi pipeline watch <runId>
 
# Check in a step (report your progress back to Mushi)
mushi pipeline checkin <runId> --step 0 --status passed
mushi pipeline checkin <runId> --step 1 --status failed --notes "Build failed: missing dep"

MCP tools

Use these from your Cursor agent to close the loop without leaving the IDE:

Tool	Scope	Description
`list_skills`	read	Browse the skill catalog
`get_skill`	read	Full skill body + chain
`start_skill_pipeline`	write	Attach a skill to a report and get the run packet
`get_pipeline_run`	read	Fetch the context packet and step statuses
`checkin_pipeline_step`	write	Report step completion — PM sees it live in the console

PDCA-style improvement loop

If a pipeline step fails, the console offers Improve and retry — this creates a child run (with parent_run_id and incremented iteration) so the PM can loop a feature through a skill like workflow-build-feature repeatedly, each iteration accumulating context from the previous run.

Security

Skill catalog content comes exclusively from allowlisted GitHub repos configured in the Sources tab. Arbitrary URLs are never fetched. Every SKILL.md body is scanned for leaked API keys before storage.

Allowlisted sources only — no arbitrary URL ingestion
Secret-pattern scan before embedding or storing any SKILL.md body
Skill content flows into LLM prompts only — never executed against raw user strings (air-gap preserved)
Description length enforced at ≤ 1024 chars per the Agent Skills spec
Run packet capped at 40,000 chars to prevent context overflows

Adding a new skill source

Go to Skill Pipelines → Sources

Click Add source and enter the GitHub owner/repo slug (e.g. myorg/my-skills-repo).

Trigger a sync

Click Sync now or run mushi skills sync. The skill-sync edge function fetches all skills/*/SKILL.md files, parses frontmatter, embeds descriptions, and upserts the catalog.

Browse the new skills

Switch to the Catalog tab. New skills appear grouped by category prefix.