Independent developer Ekko237 has released four consecutive updates to imagegen-swarm, a CLI skill for Claude Code and Codex that distributes image generation tasks across six different engines. The updates, spanning versions 1.3.0 through 1.3.3, were not part of a pre-planned roadmap but were driven entirely by community feedback, personal usage patterns, and automated security scans. The tool allows users to batch process images using Codex, Jimeng, Grok, Gemini, Muse, and Lovart, leveraging existing accounts to avoid separate API costs.
Series Mode and Style Consistency
Version 1.3.0 introduced 'series mode' after a user on X pointed out that multi-engine batches often lacked a cohesive look. This feature applies a unified style guide to all engines and uses a local Pillow-based scoring system to reject images that deviate too far from the reference, with a default threshold of 0.72. However, Ekko237 initially hesitated to enable this by default, realizing that his own daily batch of 50 font images relied on style variance. Consequently, version 1.3.1 refined the logic to ask for user confirmation before enforcing consistency, ensuring mixed-style jobs aren't inadvertently homogenized.
Critical Budget and Security Fixes
The subsequent updates addressed more severe operational risks. Version 1.3.2 fixed a budget leak identified by ClawScan, where discarded paid images (due to style mismatch or errors) were incorrectly releasing their budget reservation, allowing actual spend to exceed configured caps. The fix ensures that any paid image returned by an API keeps its budget slot reserved, even if the file is later discarded, preventing overspend during heavy retry cycles. Meanwhile, version 1.3.3 tightened security for the Muse engine, which uses a headless browser. The update restricts approval permissions to exact image saves with the label 'Allow once', blocking any compound requests or ambiguous free-text matches to prevent unintended file operations.
Builder Insights and Limitations
Ekko237 notes that the current style scoring is rudimentary, tested only on solid-color images, and hasn't been validated on complex character sets. Similarly, the Muse security patch currently fails closed, refusing all live approval cards because the developer lacks real-world samples to calibrate the extractor, forcing jobs to retry on other engines. A benchmark on an 8GB M1 MacBook Pro showed a 50-image batch completing generation in 2.3 minutes, though this excludes screening and retries.
Key Takeaways
- Version 1.3.2 fixes a critical bug where discarded paid images could cause budget overruns.
- Series mode is opt-in to preserve style variance in mixed batches, triggered by user confirmation.
- Muse security is now 'fail-closed', rejecting all approval cards until real-world data is available.
- The tool relies on existing engine accounts, eliminating the need for a dedicated image API key.
The Bottom Line
This is a masterclass in shipping public tools: treat user comments and security scanners as your best QA team, but always default to 'fail-closed' when you lack real-world data to validate your assumptions.