image-job-dispatcher

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

1. Read the registry and pipeline spec.
2. If needed, materialize shard JSONL files and issued task packets before worker fanout.
3. Register the parent job and per-image records through MCP-backed state.
4. Dispatch one worker per image or per shard-owned row using stable identifiers only.
5. Treat MCP state and worker-owned artifacts as the completion signal.
6. 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.