refly（refly 搜索）

Refly

Rules

1. CLI only - Use refly <command>, never call API directly.
2. Trust JSON - Only trust CLI JSON (ok, payload, error, hint).
3. No fabricated IDs - Never invent workflow/run/node IDs.
4. No tokens - Never print or request auth tokens.
5. Stop on error - If ok=false, stop and show hint.

Available Commands

Tip: Get installationId (skpi-xxx) from refly skill installations.

Command Options

Recommended: Use --files --latest to get files from the final output without processing all toolcalls.

Skill Categories & Execution Patterns

Choose the execution pattern based on the skill's output type:

Pattern A: File Generation Skills

Use for: nano-banana-pro, fal-image, fal-video, fal-audio, seedream-image, kling-video, wan-video

# Step 1: Run and capture RUN_ID
RESULT=$(refly skill run --id <installationId> --input '<json>')
RUN_ID=$(echo "$RESULT" | jq -r '.payload.workflowExecutions[0].id')

# Step 2: Wait for completion
refly workflow status "$RUN_ID" --watch --interval 30000

# Step 3: Get files and download to Desktop
FILES=$(refly workflow toolcalls "$RUN_ID" --files --latest | jq -r '.payload.files[]')
echo "$FILES" | jq -c '.' | while read -r file; do
  FILE_ID=$(echo "$file" | jq -r '.fileId')
  FILE_NAME=$(echo "$file" | jq -r '.name')
  if [ -n "$FILE_ID" ] && [ "$FILE_ID" != "null" ]; then
    refly file download "$FILE_ID" -o "$HOME/Desktop/${FILE_NAME}"
    open "$HOME/Desktop/${FILE_NAME}"
  fi
done

Pattern B: Text/Data Skills

Use for: jina, perplexity, weather-report, exa, research skills

# Step 1: Run and capture RUN_ID
RESULT=$(refly skill run --id <installationId> --input '<json>')
RUN_ID=$(echo "$RESULT" | jq -r '.payload.workflowExecutions[0].id')

# Step 2: Wait for completion
refly workflow status "$RUN_ID" --watch --interval 30000

# Step 3: Extract text content from toolcalls
CONTENT=$(refly workflow toolcalls "$RUN_ID" --files --latest | jq -r '.payload.nodes[].content')
echo "$CONTENT"

Pattern C: Action Skills

Use for: send-email, slack, microsoft-teams, zoom, calendar, CRM integrations

# Step 1: Run and capture RUN_ID
RESULT=$(refly skill run --id <installationId> --input '<json>')
RUN_ID=$(echo "$RESULT" | jq -r '.payload.workflowExecutions[0].id')

# Step 2: Wait for completion
refly workflow status "$RUN_ID" --watch --interval 30000

# Step 3: Confirm action status
STATUS=$(refly workflow detail "$RUN_ID" | jq -r '.payload.status')
echo "Action completed with status: $STATUS"

ID Types:

Required behavior:

• skill run returns RUN_ID (we-xxx) in .payload.workflowExecutions[0].id
• Use we-xxx for all workflow commands (status, detail, toolcalls)
• Choose execution pattern (A/B/C) based on skill category
• File skills: Download to ~/Desktop/ and open to show user
• Text skills: Extract .payload.nodes[].content for text output
• Action skills: Confirm .payload.status for completion

Directory Structure

~/.refly/skills/
├── base/                       # Base skill files (this symlink target)
│   ├── SKILL.md
│   └── rules/
│       ├── execution.md
│       ├── workflow.md
│       ├── node.md
│       ├── file.md
│       └── skill.md
└── <skill-name>/               # Domain skill directories
    └── SKILL.md

~/.claude/skills/
├── refly → ~/.refly/skills/base/           # Base skill symlink
└── <skill-name> → ~/.refly/skills/<name>/  # Domain skill symlinks

Routing

User intent -> match domain skill (name/trigger) in ~/.claude/skills/ -> read domain skill SKILL.md -> execute via refly skill run -> return CLI-verified result.

References

• rules/execution.md - Skill execution patterns and error handling
• rules/workflow.md - Workflow command reference
• rules/node.md - Node command reference
• rules/file.md - File command reference
• rules/skill.md - Customized Skill command reference