PSD Automator

Cross-platform PSD text automation for teams using both macOS and Windows.

Scope

• - Phase 1 + 2 only.
• Screenshot understanding is intentionally out of scope.
• Uses one task protocol and two execution engines:

Task Protocol

Read references/task-schema.json before running.

Minimal required fields:

• - INLINECODE1
• INLINECODE2 (layerName + newText)
• INLINECODE5 or INLINECODE6

Key optional fields:

• - workflow.sourceMode: inplace or INLINECODE9
• INLINECODE10: PNG export (mode=single or mode=layer_sets for sliced/block assets)
• INLINECODE13: zip name for sliced assets bundle
• INLINECODE14: auto / always / off (macOS Unicode path bridge)
• INLINECODE18: whether to bundle layer_sets PNGs into one zip
• INLINECODE20: screenshot path for selecting most similar exported slice

Build and Refresh PSD Index

Create or refresh local cache:

CODEBLOCK0

Incremental refresh:

CODEBLOCK1

Run a Task

Dry-run first (recommended):

CODEBLOCK2

Execute:

CODEBLOCK3

Natural-language dispatch (through OpenClaw chat command, supports .psd and .psb):

CODEBLOCK4

DingTalk image handoff (required)

When replying in DingTalk and task execution is successful with PNG output, the final response MUST include this marker with an absolute path:

CODEBLOCK5

Rules:

• - Use absolute path only (for example /Users/name/Desktop/xxx.png).
• Do not use filename-only values.
• If pngOutputPath (or first item in pngOutputPaths) is missing, clearly report failure and do not emit a fake marker.
• For mode=layer_sets, pngOutputPaths should contain all exported block images in the output folder.
• When available, use selectedPngPath as the single best-match image for screenshot workflows.
• When available, send bundleZipPath as file attachment: [DINGTALK_FILE]{"path":"<absolute_zip_path>","fileName":"<name>.zip","fileType":"zip"}[/DINGTALK_FILE].
• Keep normal human-readable summary, then append marker on a new line at the end.

OpenClaw Routing Pattern (Phase 2)

Use OpenClaw subagent routing guidance:

• - references/openclaw-subagent-routing.md

Core idea:

1. 1. Main agent parses request.
2. Resolve target machine + platform capabilities.
3. Spawn/dispatch to target subagent.
4. Subagent runs run-task.js locally.
5. Return normalized result + audit log.

Safety Baseline

• - Always support dryRun.
• Keep style lock (font and size) after text changes.
• Disable Photoshop dialogs.
• Create .bak backup before write.
• Stop on ambiguous file matches (E_FILE_AMBIGUOUS); never guess silently.
• On layer-not-found, return availableLayers + suggestedLayers.
• Emit standardized error codes from references/error-codes.md.