Skip to content

About

Lean project decomposition, compact planning, and test-first implementation for AI coding agents.

Topics

Resources

Stars

7 stars

Watchers

1 watching

Forks

Repository files navigation

Zagrosi Forge

Forge turns briefs into reviewed plans and tested code, in Codex and Claude Code.

Forge turns requests into compact plans, clear code, and verified changes. For features, fixes and cleanup in Codex and Claude Code. Lean by default: focused context, targeted checks, one integration result.

Install · Use · Multiple models · Teams · Workflows · Engineering · Releases · Contribute

Install

Requires Python 3.11+ and plugins. On Windows, replace python3 with python or py -3. Restart after installing or updating.

Claude Code

claude plugin marketplace add zagrosi-code/zagrosi-forge
claude plugin install zagrosi-forge@zagrosi

Update with claude plugin marketplace update zagrosi, then claude plugin update zagrosi-forge@zagrosi.

Codex

codex plugin marketplace add zagrosi-code/zagrosi-forge
codex plugin add zagrosi-forge@zagrosi
Install or develop from a local checkout
git clone https://github.com/zagrosi-code/zagrosi-forge.git
cd zagrosi-forge

Load this checkout in Claude Code for one session:

claude --plugin-dir "/absolute/path/to/zagrosi-forge"

From the checkout in Codex:

python3 scripts/zagrosi_skills.py install --pretty
# Preview without writing:
python3 scripts/zagrosi_skills.py install --dry-run --pretty
# Refresh from the updated checkout:
git pull --ff-only
python3 scripts/zagrosi_skills.py update-check --pretty
python3 scripts/zagrosi_skills.py self-update --pretty

Use

Describe your task; request standard or deep when needed. Try the first task.

Claude Code

/zagrosi-forge:zagrosi-forge Add this feature and finish verification.
/zagrosi-forge:zagrosi-cleanup Simplify this subsystem while preserving behavior.

Codex

Use $zagrosi-forge:zagrosi-forge to add this feature and finish verification.
Use $zagrosi-forge:zagrosi-cleanup to simplify this subsystem while preserving behavior.

Both hosts share skills, runtime and standards. Use zagrosi-project, zagrosi-plan, and zagrosi-implement for individual phases. Authorized work continues through delivery; planning-only requests stop at the plan.

Multiple models

Ask Forge for independent Codex, Claude or Gemini reviews. Reviewers receive the same bounded packet:

Implement this feature with independent Codex and Claude reviews. Keep one writer; report missing access.

Forge reuses native CLI logins. Choose model IDs or native defaults; missing access and unreported identity stay visible. No silent fallback. Add --check-auth for login status. See provider setup.

python3 scripts/zagrosi_skills.py provider-status --check-cli --pretty

Teams

See teammates' tasks and reserve work across clones, worktrees and both hosts. Ask Forge to enable collaboration, commit .forge/team.json, and have teammates join. Share reviewed plans through Git; prepare private workspaces for independent section work. No extra service.

Every depth and cleanup shares reservations. Stale tasks retain ownership; offline status grants no clearance. Notes remain in Git history. See setup, handoffs and recovery.

Workflows

Project splits a brief into specs; Plan defines a compact contract; Implement tests, builds, reviews, and records each section.

Workflow Output
Project project-manifest.md and child spec.md files
Single-section plan sections/index.md and one canonical section with tests, evidence, decisions, risks, and review
Multi-section plan Shared plan and ordered sections
Implement Code, tests, and machine-readable section records
Cleanup Simpler responsibilities and preserved behavior, with regression evidence
Depth Use
lean — default Minimum sufficient artifacts and focused checks
standard Wider research, traceability, or coordination
deep High-risk or architecture-heavy work

Depth changes investigation and review. Every mode retains requirements, ownership, acceptance, and verification without duplicate ledgers.

Complete contracts before coding. See compact-plan format, depth standards, and example briefs.

Resume and compatibility

Continue from the planning directory in either host. Checkpoints retain progress; ownership conflicts serialize work. Changed code/contracts require fresh verification. Legacy evidence cannot establish new verified completion.

Run helpers from the Forge checkout; use absolute plan paths.

python3 scripts/zagrosi_skills.py status --path "/absolute/path/to/planning/01-auth" --pretty
python3 scripts/zagrosi_skills.py next-section --planning-dir "/absolute/path/to/planning/01-auth" --pretty
python3 scripts/zagrosi_skills.py commands --phase implement --verbose --pretty

fast aliases lean; existing zagrosi-*, deep-*, DEEP_META, and migrated claude-* artifacts remain accepted. codex_review aliases agent_review.

Detached execution retains its pinned evidence protocol and physical contracts; privileged handoff requires Darwin/arm64 and APFS.

Engineering

The engineering standard follows Ponytail: trace callers, reuse code, prefer direct flow. Simplify duplication and responsibilities with regression coverage. Preservation checks run before edits, then replay unchanged checks and commands; required pairs gate completion. Capture cannot prove coverage or authorship. Run feature checks and integration. Cleanup can cover repositories; PR review covers the diff and callers.

context-brief and implementation-packet keep complete sections, mapped source requirements and linked contracts. Budget: 2,000 words (--max-words); broken links fail and oversized sections return an adjusted retry. Failures retain blockers, recovery commands and complete reports (--full-output).

Runtime loading verifies SHA-256 source bindings; writes are locked and atomic.

Helper benchmarks measure fixed fixtures. Coding trials separate code quality, workflow completion, time, and usage. Latest observations retain failures; adapter support alone does not establish faster or better model output.

Contribute

Shared workflows: skills/; runtime: scripts/forge/; checks: tests/ and examples/; host metadata: .codex-plugin/ and .claude-plugin/.

After runtime/test edits, run python3 tools/update_runtime_manifest.py. After adding/removing package members, stage intended files and run python3 tools/update_package_manifest.py. CI covers Linux/macOS/Windows and native host packaging. Native validators require their CLIs.

python3 tools/update_runtime_manifest.py --check
python3 tools/update_package_manifest.py --check
npm ci --prefix tools
uv run --with pytest python -m pytest
python3 scripts/zagrosi_skills.py doctor --plugin-root . --strict --pretty
python3 scripts/zagrosi_skills.py release-check --plugin-root . --pretty
plugin-scanner verify .
claude plugin validate . --strict

MIT License · Attribution

About

Lean project decomposition, compact planning, and test-first implementation for AI coding agents.

Topics

Resources

Stars

7 stars

Watchers

1 watching

Forks

Releases

Sponsor this project

Packages

Used by

Contributors

Languages