One PRD in. Architecture, code, tests, and acceptance evidence out — then a self-converging loop hardens it until it clears a measurable quality bar.
CodeTrace is an AI-first delivery pipeline: a reusable, version-controlled system
of agents, skills, commands, rules, and memory under .claude/. To prove the pipeline
works, it generates its own payload — a lightweight, local-first, zero-dependency
Python function tracer under src/.
The pipeline is the deliverable. The library is the proof.
PRD → Intent → Architecture → Build Spec → Task Slices → Coding → Verify → Accept → ⟳ Converge
Every stage is three files: a command (entry), a skill (how), an agent (who), wrapped in non-negotiable rules and resumable memory.
/seechen --run # run the full pipeline from docs/PRD.md
/converge --run # harden a green milestone until it convergesThe pipeline generated the library end to end, then the convergence loop hardened it across three rounds:
77/77 tests · 99% line coverage · 77.3% mutation · ruff clean · mypy 0 errors · complexity ≤ B → CONVERGED
The loop caught what a green suite hid — a type-safety hole, a complexity hotspot, and a coverage illusion: 99% of lines executed, but only 64.5% of the logic was actually asserted. Hardening lifted mutation to 77.3%, past the gate, and the loop stopped.
→ Case Study — the full narrative · Audit evidence — per-round reports
| Document | What it covers |
|---|---|
| PRD | Product source of truth |
| Workflow | Stage-by-stage definitions |
| Pipeline Components | Every agent and skill in detail — inputs, outputs, boundaries |
| Case Study | The end-to-end run, round by round |
| Convergence Loop | Rubric, gates, stop conditions, driver |
| Diagrams | Flow + sequence diagrams, rendered inline (PlantUML source in docs/diagrams/) |
CodeTrace/
├── .claude/ # the pipeline: agents, skills, commands, rules, memory, docs
├── docs/ # PRD, workflow, component reference, case study, diagrams
├── specs/ # generated outputs: intent, architecture, build, acceptance, audit
├── src/ tests/ # the generated library and its test suite
└── CLAUDE.md # repository entry point
Follow the rules in .claude/rules/: task-appropriate branches,
Conventional Commits, English for reusable constraint files, and update
.claude/memory/pipeline-state.md after each completed stage.
MIT — see LICENSE.