UNPKG

paqad-ai

Version:

Spec-driven development framework — AI agents that think before they type

github.com/Eliyce/paqad-ai

Eliyce/paqad-ai

41 lines (32 loc) • 3.15 kB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
# Quality Ratchet — measure → tool mapping The quality ratchet records four measures at the project's real level today and only ever allows equal-or-better. It reuses existing per-language tools — it does **not** reinvent metric tooling. Each measure is normalised to a **deficiency count where lower is better** (violations / orphan files / disabled strict flags), so "worse" is always "the number went up". ## The four measures and where each comes from | Measure | What it means | How it is measured (per pack) | Confidence | | ---------------- | -------------------------------- | ------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------- | | `tangledness` | how tangled the code is | the language's complexity tool, run via rule-scripts (TS/JS: ESLint `complexity`; Python: pylint/radon; Rust: clippy; …) | `mature` where a real tool exists, else `lower` | | `dead_code` | dead / unused code | **consumed** from the traceability reachability/orphan output — never re-scanned | `mature` | | `risky_patterns` | risky patterns | existing security / risky-pattern lint + pentest (OWASP) findings | `mature` where wired, else `lower` | | `strictness` | how strictly the code is written | the strict-config surface, read directly (TS: `tsconfig` strict flags; others: equivalent) | `mature` for TS, else `lower` | A measure with no tool for the stack is recorded **lower-confidence / blocked** — never a fabricated number — and a blocked measure never blocks the gate. ## Per-pack wiring - **This repo (TypeScript)** — strictness from `tsconfig` strict flags; dead code from traceability; tangledness from ESLint `complexity`; risky patterns from the security lint + pentest findings. - **Other languages** — wire the complexity and risky-pattern runners through the pack's rule-scripts; mark measures with weak tooling `lower` confidence (mirrors mutation testing). Strictness and dead code carry over as above. ## The rule (for agents) - A change is allowed only if every measure is **equal or better** than the recorded baseline. New / changed code is held to at least the existing level (clean-as-you-code), so the recorded level can only tighten. - The `fast` lane is not blocked by complexity / dead-code / risky noise, but it **cannot loosen the baseline** (a strictness-loosening fast change still trips). - A genuine, legitimate need to worsen a measure opens a `quality.ratchet_exception` Decision Pause — approved once, reused for the same kind. Do not work around the ratchet by deleting unrelated code to keep a total flat; bring the measure back, or open the exception.