Open-source spec-driven skills plugin. Works on both Codex and Claude Code.
| Field | Value | Notes |
|---|---|---|
| Repository | github.com/devnomad-byte/techneering |
Open-source repo |
| Plugin name | tn |
Both Codex and Claude Code identifier |
| Skill call prefix | tn: |
Prefix for all skills (e.g. tn:compass) |
Codex reads this AGENTS.md before doing any work, and loads Techneering's skills from ./skills/. A SessionStart hook (in hooks/hooks.json, which Codex auto-discovers by default) also injects the tn:compass gateway at session start, mirroring the Claude Code experience. If the hook doesn't fire for any reason, the rules below still apply.
Why there's both AGENTS.md and CLAUDE.md (no conflict): Each host loads only its own file. Codex auto-loads AGENTS.md (this file); it does not auto-load CLAUDE.md. Claude Code auto-loads CLAUDE.md; it does not auto-load AGENTS.md. They serve the same purpose for two different hosts, so both ship in the repo. The hooks/hooks.json is shared — Codex injects CLAUDE_PLUGIN_ROOT for compatibility, so the one hook file works on both.
The single most important rule: Before responding to any user request, check whether a Techneering skill applies — and if it does, follow that skill. Start every session by consulting tn:compass to pick the right development path.
| Skill | Call | Type | Description |
|---|---|---|---|
| Compass | tn:compass |
Gateway | Auto-injected on session start, selects development path |
| Scout | tn:scout |
Flow | Standalone exploration mode, pure thinking, writes no files |
| Draft | tn:draft |
Flow | Generate proposal + specs + design + tasks |
| Forge | tn:forge |
Flow | Entry point: chains blueprint → assemble |
| Audit | tn:audit |
Flow | Dual-layer verification: spec compliance + requirement tracing |
| Vault | tn:vault |
Flow | delta merge → move into archive/ |
| Spark | tn:spark |
Process | Socratic questioning + project style detection + framework recommendation |
| Blueprint | tn:blueprint |
Process | Break into bite-sized tasks |
| Diagnose | tn:diagnose |
Rigid | 4 phases: root cause → pattern → hypothesis → fix |
| Gate | tn:gate |
Rigid | Iron law: run command → read output → claim → show menu |
| Redgreen | tn:redgreen |
Rigid | Red-green-refactor cycle |
| Assemble | tn:assemble |
Execution | Subagent dispatch + two-stage review + parallel scheduling |
| Inspect | tn:inspect |
Execution | Global cross-task review: consistency, architecture, security |
| Isolate | tn:isolate |
Execution | Create isolated workspace + verify test baseline |
| Ship | tn:ship |
Wrap-up | Branch integration: merge / PR / keep / discard |
| Craft | tn:craft |
Domain | Auto-stacked when a frontend task is detected |
| Path | Use case | Flow |
|---|---|---|
| Complete | Large changes | spark → draft → forge → gate → user chooses next flow |
| Lightweight | Medium changes | draft → forge → gate → user chooses next flow |
| Fix | Urgent bugs | diagnose → redgreen → gate → user chooses next flow |
| Direct | Trivial changes | direct execution → gate |
After gate passes, present a menu for the user to choose:
- Full verification (audit → inspect → ship → vault) — production code
- Standard flow (audit → ship → vault) — internal projects
- Quick commit (ship → vault) — small changes/prototypes
- Verify only (audit) — just check, don't commit
After each skill completes, its Next Step points to the next skill. The full transition rules live in the tn:compass gateway skill (./skills/compass/SKILL.md).
All skills follow a unified structure: frontmatter → core principle → steps → red flags → quick reference → Next Step → guardrails.
Rigid skills (redgreen/diagnose/gate) must have a red-flags list. All skills must have Next Step + guardrails.
On the first draft, techneering/config.yaml is generated with two core sections:
| Field | Purpose | Source |
|---|---|---|
coding_conventions |
Code style constraints (naming, structure, patterns, formatting) | spark detection / simplified detection / framework defaults |
quality_standards |
External quality standards (Alibaba Taishan, PEP 8, etc.) | Recommended after spark detects the tech stack |
Standard source options: Java → ali-taishan-java, Python → ali-taishan-python + pep8, Go → gofmt, frontend → user's choice.
Full-flow injection: draft writes → assemble implementer/reviewer read and enforce → inspect cross-checks globally.