Why Zimster exists
Agent workflows can produce good code while still spending too much time on planning loops, repeated review, and unnecessary delegation. Zimster keeps the practices that improve software quality while giving one implementation owner responsibility for the result.
The central question is practical: which steps materially improve this change, and which steps are just extra orchestration?
Design goals
- Keep one implementation owner responsible for architecture and integration.
- Use design and planning in proportion to ambiguity and risk.
- Make tests demonstrate the behavior being changed.
- Delegate only work with a useful boundary and a clear acceptance check.
- Let users choose how delegated roles map to model classes and concrete model IDs.
- Keep host-specific packaging separate from the portable workflow.
What changed in v0.7
The 0.7 line moved Zimster to portable Agent Plugin packaging and added packages for Codex, Claude Code, skills-only use, portable Agent Plugin hosts, and npm-based hosts. It also added model routing, run controls, risk-based design, and package-level verification.
Version 0.7.2 replaces the aborted 0.7.1 release. It adds hard execution stops, review-cycle accounting that cannot be reset by renaming the work, migration for 0.7.0 Git-local state, cross-platform fixes, and signed release-tag verification.
DeepSWE pilot during v0.7 development
A controlled paired DeepSWE v1.1 pilot compared Codex with Zimster's portable Agent Skills workflow against the same Codex setup without Zimster. The prespecified minimum pilot completed 24 scored runs across 12 matched pairs, with no incomplete pairs and no retries.
| Measure | Zimster | Control |
|---|---|---|
| Deterministic passes | 10 of 12 (83.33%) | 9 of 12 (75%) |
| Mean wall-clock time | 914.701 seconds | 1,031.920 seconds |
| Mean agent turns | 64.75 | 73.75 |
| Mean output tokens | 29,128.7 | 32,430.8 |
The mean runtime was 11.4% lower, agent turns 12.2% lower, and output tokens 10.2% lower in the Zimster condition. The paired pass-rate difference was +8.33 percentage points with a 95% task-cluster bootstrap interval of 0 to 25 points. The pilot was small, and the adjusted secondary comparisons were not statistically significant.
Oversized-request decomposition and optional visual treatment were added later in v0.7 development and were tested separately. The full protocol and results are in the repository's Evaluation guide.
Distribution
Version 0.7.2 ships for Codex, Claude Code, Grok, OpenCode, Pi, and Kimi Code. The workflow itself lives in 12 Agent Skills; each host adds the packaging or adapter pieces it needs.
The GitHub release files and npm package are live. Host directory listings are separate from the release itself.