Composable recipes for your code factory.
Impstack defines how agents receive work, isolate changes, report evidence, and review each other. Choose agent profiles independently of the workflow. Use native Codex app tools, terminal agents through Herdr, or a hybrid assignment that mixes both. The hybrid example uses an app coordinator with terminal implementation lanes.
The harness owns execution, permissions, authentication, and context. Impstack supplies configuration validation and task handoffs. It does not start a replacement agent runtime.
The name comes from "imperfect operator." Human review and release decisions remain part of the method. Backpass is an optional source of corrections, alongside other retrospective tools.
- Configure a factory and validate its handoffs.
- Understand the architecture and its boundaries.
- Install the existing personal preset.
The factory CLI runs from this checkout with Python and its standard library. The existing bootstrap installs the personal skill and harness preset. Factory module declarations do not change that bootstrap.
git clone https://github.com/ivankqw/impstack.git ~/impstack
cd ~/impstack
bin/factory --helpflowchart TD
I[Issue and acceptance criteria] --> R[Workflow recipe]
P[Agent profiles and role assignments] --> H[Task handoffs]
R --> H
H --> N[Native Codex app tools]
H --> T[Herdr terminal agents]
T --> C[Claude Code / Codex CLI / OpenCode]
N --> E[Result and verification evidence]
C --> E
E --> V[Independent defect and standards review]
V --> D[Draft pull request]
D --> O[Authorized merge and operator release]
| Shared requirement | Replaceable choice |
|---|---|
| Explicit task and acceptance criteria | GitHub, Linear, or a local work record |
| Isolated implementation and review workspaces | Harness worktrees or a worktree helper |
| Recorded profile for each role | Harness, execution method, provider, model, and native options |
| Independent review with executed evidence | Reviewer profile and review tools |
| Durable task result with unresolved work | Native task tools or Herdr with report files |
| Explicit merge authority and operator releases | Repository policy and deployment tooling |
| Corrections remain reviewable | Backpass, reflect, or manual retrospective |
Model names are local choices. Validate them in the selected harness before dispatch. A valid factory config does not prove model availability or runtime readiness.
The installer configures Claude Code, Codex, and OpenCode instructions, skills, and MCP declarations. It also links Claude Code and Codex hook files for operator registration. A factory profile declares a handoff. Commission the selected harness before you use it.
Herdr enables the terminal execution path. Native app handoffs do not require Herdr.
Use --with-herdr when you install to restore the Herdr skill.
It does not install the Herdr runtime.
See factory configuration for the executable contract and legacy model presets for the existing prose configurations.
I write a skill when no upstream skill covers the job.
Skills I own
| skill | job |
|---|---|
factory-workflow |
Execute selected profiles through native tools or Herdr with verified handoffs. |
cleanup-crew |
Keep the issue tracker aligned with current work. |
commission |
Probe a machine, check its fixed contract, and write its local record. |
dogfood-local |
Run a local app and verify the real user path. |
lane-orchestration |
Run implementation lanes from briefs through review and cleanup. |
browser-tooling |
Give agent panes browser inspection tools. |
recommission |
Re-probe a machine record and ship at most one proven correction. |
Upstream skills by source
| source and credit | skills |
|---|---|
| pstack by Lauren "poteto" Tan, contributors, and Michael Denyer's Claude port | Routing, review, verification, agent workflows, and engineering principles. |
| Herdr | herdr. |
| Matt Pocock | ask-matt, code-review, codebase-design, diagnosing-bugs, domain-modeling, grill-me, grill-with-docs, grilling, handoff, implement, improve-codebase-architecture, prototype, research, resolving-merge-conflicts, setup-matt-pocock-skills, tdd, teach, to-questionnaire, to-spec, to-tickets, triage, wait-what, wayfinder, wizard, writing-for-agents. |
| Vercel Labs | find-skills. |
| Cursor | deslop. |
| Hardik Pandya | stop-slop. |
| Microsoft | microsoft-foundry. |
| shadcn | improve. |
| Leon | brandkit, design-taste-frontend, high-end-visual-design, imagegen-frontend-mobile, imagegen-frontend-web, image-to-code, redesign-existing-projects. |
| Peter Bakaus | impeccable. |
| Aiden Bai | improve-react. |
| Saurabh Kumar | cyclomatic-complexity. |
| Dietrich Gebert | ponytail-audit, ponytail-debt, ponytail-gain, ponytail-help, ponytail-review. |
- Thesis explains the code-factory direction and its boundaries.
- How it works explains each layer and its source file.
- Install installs, verifies, updates, and removes the setup.
- Credits names upstream authors, sources, and licenses.
- Maintaining gives repository editing and verification rules.
- Experiments records bounded workflow comparisons and their limits.
Fork this setup and overfit it to yourself. Keep what fits your work and replace my assumptions with yours.
The repository uses the MIT License. See LICENSE. NOTICE records adapted material.