Auto-generated from codebase by
generate.py. Runpython docs/architecture/generate.pyto regenerate.
Open index.html in a browser to explore the architecture interactively with tabbed Mermaid diagrams and plugin component cards.
6 plugins. 7 agents. 7 skills. 6 engines (F1–F6). 5-axis scoring. 8 SAT assertions. 16 tests.
| Diagram | File | Description |
|---|---|---|
| High Level | highlevel.mmd | 6 plugins on top of Opus/Sonnet/Haiku agent tiers |
| Session Lifecycle | lifecycle.mmd | Craft → Refine → Converge → Test → Harden → Translate |
| Data Flow | dataflow.mmd | Prompt artifacts (prompt.*, metadata.json, learnings.md) across skills |
| Hook Bindings | hooks.mmd | Single advisory hook: PostToolUse on prompts/*/prompt.* saves |
| Plugin | Stage | Agent tier | Skill | Artifact |
|---|---|---|---|---|
| prompt-crafter | Craft | Opus + Haiku | /create |
prompt.*, metadata.json |
| prompt-refiner | Refine | Opus + Haiku | /refine |
prompt.* (v++), metadata.json |
| convergence-engine | Converge | Sonnet + Haiku | /converge |
learnings.md, updated scores |
| prompt-tester | Test | Sonnet | /test-prompt |
tests.json, pass/fail |
| prompt-harden | Harden | Sonnet red-team | /harden |
audit.json (12 attacks) |
| prompt-translate | Translate | Sonnet adapter | /translate-prompt --to <model> |
prompt.<new>, score comparison |
| Tier | Model | Used for |
|---|---|---|
| Orchestrator | Opus | Judgment, intent, technique selection (crafter, refiner) |
| Executor | Sonnet | Convergence loop, adversarial attacks, format conversion, test execution |
| Validator | Haiku | Quality gate — file completeness, metadata consistency, score freshness |
| Code | Name | Where |
|---|---|---|
| F1 | Gauss Convergence | shared/scripts/convergence.py |
| F2 | Boolean Satisfiability Overlay | run_assertions() in convergence.py |
| F3 | Cross-Domain Adaptation | prompt-translate adapter |
| F4 | Adversarial Robustness | prompt-harden red-team loop |
| F5 | Static-Dynamic Dual Verification | tester + reviewer pair |
| F6 | Gauss Accumulation (self-learning) | learnings.md aggregation |
Full derivations: docs/science/README.md.
| Verdict | Criteria |
|---|---|
| DEPLOY | σ < 0.45 and overall ≥ 9.0 and all 5 axes ≥ 7.0 and 8/8 SAT assertions pass |
| HOLD | σ ≥ 0.45 or any axis < 7.0 |
| FAIL | Reviewer flags registry mismatch / stale technique / format drift |
σ = standard deviation of 5 axis scores from 10. Axes: clarity, specificity, structure, constraints, coverage.
1. /create → prompt-crafter (Opus) drafts prompt.* + metadata.json
2. /refine → prompt-refiner (Opus) increments version, updates metadata
3. PostToolUse save → advisory hook: "convergence engine will optimize automatically"
4. /converge → convergence-engine (Sonnet) runs Gauss loop, appends learnings.md
5. /test-prompt → prompt-tester (Sonnet) runs regression cases from tests.json
6. /harden → prompt-harden (Sonnet) runs 12 adversarial attacks, emits audit.json
7. /translate-prompt → prompt-translate (Sonnet) retargets format (Claude/GPT/o-series/Gemini)
prompts/<name>/
├── prompt.<ext> production prompt, format matches target model
├── metadata.json model, tokens, cost, 5-axis scores, 8 assertions, version
├── tests.json regression test cases (≥ 3, ≥ 1 edge-case)
├── report.pdf dark-themed single-page audit (final only)
└── learnings.md F6 hypothesis/outcome log — persists across sessions
16 tests across all plugins + shared utilities:
tests/
├── convergence-engine/ Gauss loop, no-regression, accumulation
├── prompt-crafter/ technique selection, metadata schema
├── prompt-refiner/ version increment, diff generation
├── prompt-tester/ fixture execution, pass/fail routing
├── prompt-harden/ 12-attack audit, red-team coverage
├── prompt-translate/ format conversion (XML ↔ Markdown ↔ minimal ↔ few-shot)
└── run-all.sh Master runner