UNPKG

paqad-ai

Version:

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

github.com/Eliyce/paqad-ai

Eliyce/paqad-ai

59 lines (43 loc) • 2.41 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
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
--- name: map-verification description: "Refutes or confirms the claims the deterministic checks could not settle, the Tier-B pass over the site map, grounding each verdict in resolving evidence. Use when assembly is done and the engine's inconclusive-claim table has entries." metadata: model_tier: reasoning cacheable: 'false' output_format: json paqad-meta: skill.meta.yaml --- ## What It Does Provides the Tier-B judgment layer: the engine's deterministic checks settle what they can and hand over a table of claims they could not: an evidence pointer that half-resolves, a transition whose runtime behaviour is ambiguous. This skill refutes or confirms each, grounded in the code, so an inconclusive claim becomes either a finding or a fact, never a shrug. ## Use This When After assembly, over the engine's inconclusive-claim table. It settles only what the deterministic checks left open; it never re-litigates a claim the engine already proved. ## Inputs - The machine-built claim digest from the run bundle (each claim with its evidence and why it is inconclusive). - The code the claims cite. - Read `references/refutation-discipline.md` before recording a verdict. ## Procedure The deterministic checks are the engine's; you settle only the inconclusive residue. 1. Read each inconclusive claim and open the evidence it cites. 2. Try to **refute** it first: does the code actually show what the claim asserts? Default to `inconclusive` when the evidence does not settle it, never to `confirmed`. 3. Record a verdict (`confirmed | refuted | inconclusive`) with the resolving `file:line`, and let `site-map-gap-analysis` turn confirmed problems into findings. ## Output Contract - A JSON object `{ verdicts: [{ claim_id, verdict, evidence, rationale }] }`. - `verdict` ∈ `confirmed | refuted | inconclusive`. - Every `confirmed` or `refuted` verdict carries a resolving `file:line`; an `inconclusive` one states what evidence is missing. ## Escalate / Stop Conditions - Default to `inconclusive` when the evidence does not settle the claim. Do not confirm to look thorough or refute to look clean. - Do not invent a new claim here; new problems are a fresh `site-map` run, not a verification verdict. - A claim whose evidence no longer resolves is `confirmed` drift, never quietly dropped. ## Resources - `references/refutation-discipline.md` - `agents/openai.yaml`