一键在 Manus 中运行任何 Skill

image-job-dispatcher

星标1

分支0

更新时间2026年4月9日 06:10

Dispatch one-image-per-worker jobs for the presentation image pipeline. Use when you need to register image jobs, fan out isolated Codex subagents, and keep MCP state as the source of truth.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

goodand

goodand/my-image-parser

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Image Job Dispatcher

Overview

Use this skill when a batch of presentation images must be processed through isolated workers. This skill owns dispatch only. It does not decide final approvals or commit rename and metadata changes.

Use This Skill When

A presentation export has already produced image files.
Each image should run in its own worker context.
Job state must be tracked through MCPs instead of chat text.
The parent agent needs to resume safely after interruption.

Inputs

Image list or extracted image directory
Stable job_id
Stable image_id per file
Registry location from registry/session_paths.json
Optional shard manifest when a multi-worker batch has already been materialized

Required MCPs

agent-task-manager
conport

Optional MCPs

filesystem for path inspection
cv-mcp only when the dispatcher also performs a smoke check

Workflow

Read the registry and pipeline spec.
If needed, materialize shard JSONL files and issued task packets before worker fanout.
Register the parent job and per-image records through MCP-backed state.
Dispatch one worker per image or per shard-owned row using stable identifiers only.
Treat MCP state and worker-owned artifacts as the completion signal.
Hand off only finished items to the auditor or commit manager.

Dispatch Rules

Use job_id and image_id as worker inputs.
Keep image content and captions in the registry, not in worker prompts.
Do not trust free-form worker text as the completion signal.
A worker is complete only when MCP state and evidence say it is complete.
Retry failed rows individually. Do not re-run the full batch without reason.
For parallel runs, one shard JSONL, one worker ledger, and one _responses/ directory belong to exactly one worker.
Only the main agent may aggregate shared summaries after parallel workers finish.

Handoff Targets

Use image-worker for single-image processing.
Use image-result-auditor after worker completion.
Use image-commit-manager only after human approval is recorded.

References

references/runtime.md
references/troubleshooting.md
references/shard-strategy.md
references/task-packet-contract.md
references/cohort-expansion-truth-sync.md
checklists/parallel-preflight.md
knowledge_bases/image-job-dispatcher-knowledge-base-at2026-03-28-10-55.md
evals/evals.json

Extension Boundary

This skill may also guide the coordinator when a newly closed image must be folded back into an already-canonical experiment cohort. That extension is limited to truth synchronization:

rebuild aggregate bundle truth from finished per-image artifacts
rebuild downstream consumer truth from the same canonical inputs
patch closure and registry surfaces so shared state matches the widened cohort

It still does not own semantic judging, file commit, or baseline promotion.

同仓库更多 Skills

同仓库

component-split-ocr-review

goodand/my-image-parser

Export deterministic component-level OCR evidence for one transparent or semi-transparent image by splitting alpha-connected components, writing a component table, and running OCR on each component. Use when bounded component evidence must be exported before any downstream refinement or caption rerun.

2026-04-131

macos-ocr-evidence

goodand/my-image-parser

Extract bounded OCR evidence from local image files with macos-ocr-mcp, using macOCR only as a fallback reference when MCP coverage is insufficient. Use when downstream refinement or review needs visible-text evidence, not final semantic decisions.

2026-04-131

object-isolation-correction

goodand/my-image-parser

Wrap object-isolation correction as a repo-specific workflow. Use when a local image already exists, the current isolation result is imperfect, and the next step is to choose between ImageSorcery-first, imagegen-first, or a hybrid correction path with a bounded correction packet and worker run.

2026-04-131

parser-sidecar-to-canonical-schema-promotion

goodand/my-image-parser

Promote a bounded parser raw sidecar artifact into the canonical Table -> Row -> Cell schema. Use when a parser MCP returned markdown plus nested JSON sidecars and the project must normalize that output before any downstream wrapper, worksheet, or RAG consumer reads it.

2026-04-131

transparent-component-triage

goodand/my-image-parser

Triage a bounded batch of PPT-extracted images to find the small subset that is mechanically sufficient for alpha-only component splitting. Use when conservative preprocessing is required and only review-gated deterministic candidates should advance beyond the full-image baseline.

2026-04-131

vscode-fabriqa-foam-workflow

goodand/my-image-parser

Operate an existing markdown review surface in VS Code with fabriqa for same-page image editing and Foam for backlinks and graph navigation, without redefining machine-truth semantics.

2026-04-131

name	image-job-dispatcher
description	Dispatch one-image-per-worker jobs for the presentation image pipeline. Use when you need to register image jobs, fan out isolated Codex subagents, and keep MCP state as the source of truth.

Image Job Dispatcher

Overview

Use this skill when a batch of presentation images must be processed through isolated workers. This skill owns dispatch only. It does not decide final approvals or commit rename and metadata changes.

Use This Skill When

A presentation export has already produced image files.
Each image should run in its own worker context.
Job state must be tracked through MCPs instead of chat text.
The parent agent needs to resume safely after interruption.

Inputs

Image list or extracted image directory
Stable job_id
Stable image_id per file
Registry location from registry/session_paths.json
Optional shard manifest when a multi-worker batch has already been materialized

Required MCPs

agent-task-manager
conport

Optional MCPs

filesystem for path inspection
cv-mcp only when the dispatcher also performs a smoke check

Workflow

Read the registry and pipeline spec.
If needed, materialize shard JSONL files and issued task packets before worker fanout.
Register the parent job and per-image records through MCP-backed state.
Dispatch one worker per image or per shard-owned row using stable identifiers only.
Treat MCP state and worker-owned artifacts as the completion signal.
Hand off only finished items to the auditor or commit manager.

Dispatch Rules

Use job_id and image_id as worker inputs.
Keep image content and captions in the registry, not in worker prompts.
Do not trust free-form worker text as the completion signal.
A worker is complete only when MCP state and evidence say it is complete.
Retry failed rows individually. Do not re-run the full batch without reason.
For parallel runs, one shard JSONL, one worker ledger, and one _responses/ directory belong to exactly one worker.
Only the main agent may aggregate shared summaries after parallel workers finish.

Handoff Targets

Use image-worker for single-image processing.
Use image-result-auditor after worker completion.
Use image-commit-manager only after human approval is recorded.

References

references/runtime.md
references/troubleshooting.md
references/shard-strategy.md
references/task-packet-contract.md
references/cohort-expansion-truth-sync.md
checklists/parallel-preflight.md
knowledge_bases/image-job-dispatcher-knowledge-base-at2026-03-28-10-55.md
evals/evals.json

Extension Boundary

This skill may also guide the coordinator when a newly closed image must be folded back into an already-canonical experiment cohort. That extension is limited to truth synchronization:

rebuild aggregate bundle truth from finished per-image artifacts
rebuild downstream consumer truth from the same canonical inputs
patch closure and registry surfaces so shared state matches the widened cohort

It still does not own semantic judging, file commit, or baseline promotion.