https://github.com/duck-lint/coding-agent-harness
This harness prevents common failure modes with iterative AI-assisted coding. Especially helpful for users with no formal coding language syntax experience.
https://github.com/duck-lint/coding-agent-harness
ai-assisted-development coordination harness-engineering harness-framework sub-agents
Last synced: about 14 hours ago
JSON representation
This harness prevents common failure modes with iterative AI-assisted coding. Especially helpful for users with no formal coding language syntax experience.
- Host: GitHub
- URL: https://github.com/duck-lint/coding-agent-harness
- Owner: duck-lint
- Created: 2026-05-10T15:07:16.000Z (24 days ago)
- Default Branch: master
- Last Pushed: 2026-05-30T18:26:40.000Z (4 days ago)
- Last Synced: 2026-05-30T19:13:44.344Z (4 days ago)
- Topics: ai-assisted-development, coordination, harness-engineering, harness-framework, sub-agents
- Homepage: https://youtu.be/XiF06_yM488?si=Pxh8pqdHvqghgJdO
- Size: 262 KB
- Stars: 2
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Coding Agent Harness Template
This repo packages a lightweight external cognition harness for Codex-driven coding work.
It contains:
- `AGENTS.md`: always-on orchestrator guidance for the user-facing chat
- `harness-skills/project-manager/`: advisory skill for strict admissibility and project trajectory reports
- `harness-skills/decision-matrix/`: first-class decision aid for structured option comparison
- `harness-skills/seed-repo/`: skill and script for seeding a target repo with `harness/`
- `subagents/`: distributable TOML templates for the project-manager advisory role plus planner, implementer, reviewer, adversary, and archivist roles
## Architecture
The root orchestrator owns the user conversation and role routing. The project-manager skill is advisory: it produces strict reports about admissibility, thesis/tension, approval boundaries, affected surfaces, and next admissible transitions. The subagents execute bounded jobs inside their own authority.
This keeps the UX in one chat while preserving separate context and purpose:
- root `AGENTS.md` handles conversation and orchestration
- `$project-manager` handles project direction and admissibility review
- `decision-matrix` handles structured comparison when the user needs to choose among several options
- installed subagents handle project-direction review, planning, implementation, review, adversarial checks, and archival memory
- `$seed-repo` installs repo-local harness memory into target projects
## Seeding A Repo
Use the seed skill script from this repo:
```powershell
python .\harness-skills\seed-repo\scripts\seed-repo.py --target C:\path\to\target-repo
```
The script copies:
```text
harness-skills/seed-repo/assets/repo-harness-template/repo-harness/
```
into:
```text
/harness/
```
It aborts if `/harness/` already exists unless `--force` is explicitly supplied.
## Subagent Installation
The TOML files under `subagents/` are install templates. To let the root orchestrator spawn these roles, install them into:
```text
~/.codex/agents/
```
The seed script can do this with:
```powershell
python .\harness-skills\seed-repo\scripts\seed-repo.py --target C:\path\to\target-repo --install-subagents
```
Installing subagents changes user-global Codex config. The orchestrator or seed skill must ask for explicit approval before running that option.
## Seeded Harness Layout
```text
harness/
README.md
harness-runtime.md
sub-agents.md
archive-policy.md
known-failures.md
open-decisions.md
canon/
type-system-operational.md
bridge-schema.md
implementation-projects/
active/
archive/
templates/
implementation-plan-template.md
implementation-tracker-template.md
project-spec/
template-governance-primitives.md
template-project-spec.md
```
`harness/` is the canonical repo-local continuity store. Do not duplicate project state in repo-root `memories/`, host memory files, or chat-only summaries.