co-maintainer 0.7.2 → 0.7.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/package.json +1 -1
- package/dist/src/cli/command_json.d.ts +3 -1
- package/dist/src/cli/command_json.js +1 -0
- package/dist/src/cli/commands/review.js +2 -0
- package/dist/src/cli/commands/set.js +19 -2
- package/dist/src/cli/pr_carry.d.ts +2 -0
- package/dist/src/cli/pr_carry.js +2 -2
- package/dist/src/cli/review_carry.d.ts +3 -1
- package/dist/src/cli/review_carry.js +5 -3
- package/dist/src/cli/review_result.d.ts +10 -1
- package/dist/src/cli/review_result.js +14 -1
- package/dist/src/config.d.ts +9 -0
- package/dist/src/github/collect.d.ts +7 -1
- package/dist/src/github/collect.js +41 -16
- package/dist/src/knowledge/guide.d.ts +24 -1
- package/dist/src/knowledge/guide.js +39 -9
- package/dist/src/knowledge/guide_ids.d.ts +9 -0
- package/dist/src/knowledge/guide_ids.js +26 -7
- package/dist/src/knowledge/rule_changes.d.ts +18 -0
- package/dist/src/knowledge/rule_changes.js +41 -0
- package/dist/src/knowledge/rules.d.ts +11 -0
- package/dist/src/knowledge/rules.js +10 -1
- package/dist/src/knowledge/stable_ids.d.ts +29 -0
- package/dist/src/knowledge/stable_ids.js +90 -0
- package/dist/src/knowledge/synthesis.d.ts +13 -1
- package/dist/src/knowledge/synthesis.js +89 -21
- package/dist/src/knowledge/types.d.ts +16 -0
- package/dist/src/local/review_local.js +3 -2
- package/dist/src/pr/findings.d.ts +3 -0
- package/dist/src/pr/findings.js +21 -1
- package/dist/src/pr/findings_json.d.ts +1 -1
- package/dist/src/pr/findings_json.js +14 -1
- package/dist/src/pr/reviewer.d.ts +4 -0
- package/dist/src/pr/reviewer.js +2 -0
- package/dist/src/remote/client.js +1 -0
- package/dist/src/review/carry_over.d.ts +8 -0
- package/dist/src/review/carry_over.js +8 -0
- package/dist/src/review/references.d.ts +68 -0
- package/dist/src/review/references.js +176 -0
- package/dist/src/server/pages/pr.js +2 -0
- package/dist/src/server/pages/references.d.ts +2 -0
- package/dist/src/server/pages/references.js +55 -0
- package/dist/src/server/pages/repo.js +26 -0
- package/dist/src/services/dashboard.d.ts +2 -0
- package/dist/src/services/dashboard.js +2 -0
- package/dist/src/services/push_review.js +4 -1
- package/dist/src/services/remote_review.js +5 -2
- package/dist/src/services/review.js +35 -8
- package/dist/src/services/rules_build.d.ts +20 -1
- package/dist/src/services/rules_build.js +85 -6
- package/dist/src/services/setup.js +27 -2
- package/dist/src/store/findings.d.ts +2 -0
- package/dist/src/store/findings.js +3 -3
- package/dist/src/store/migrations.js +17 -0
- package/dist/src/store/rows.d.ts +2 -0
- package/dist/src/store/rule_events.d.ts +15 -0
- package/dist/src/store/rule_events.js +20 -0
- package/dist/src/types.d.ts +4 -0
- package/package.json +1 -1
|
@@ -32,6 +32,18 @@ function takeSourceBracket(line) {
|
|
|
32
32
|
function setKey(numbers) {
|
|
33
33
|
return numbers.join(",");
|
|
34
34
|
}
|
|
35
|
+
/** `(overturns R4, R7)` right before the source bracket. */
|
|
36
|
+
const OVERTURNS = /\s*\(overturns\s+(R[1-9]\d*(?:\s*,\s*R[1-9]\d*)*)\)\s*$/;
|
|
37
|
+
/** The line without its overturns note, and the ids it named. */
|
|
38
|
+
export function takeOverturns(line) {
|
|
39
|
+
const match = OVERTURNS.exec(line);
|
|
40
|
+
if (!match)
|
|
41
|
+
return { line, ids: [] };
|
|
42
|
+
return {
|
|
43
|
+
line: line.slice(0, match.index),
|
|
44
|
+
ids: [...new Set(match[1].split(",").map((id) => id.trim()))],
|
|
45
|
+
};
|
|
46
|
+
}
|
|
35
47
|
function ruleText(line) {
|
|
36
48
|
return line.replace(/^- /, "").replaceAll("`", "").trim();
|
|
37
49
|
}
|
|
@@ -54,7 +66,8 @@ function rewrite(text, file, findingCount, idFor) {
|
|
|
54
66
|
if (valid.length === 0)
|
|
55
67
|
return taken.line;
|
|
56
68
|
withIds++;
|
|
57
|
-
|
|
69
|
+
const note = takeOverturns(taken.line);
|
|
70
|
+
return `${note.line} [${idFor(valid, note.line, note.ids)}]`;
|
|
58
71
|
});
|
|
59
72
|
return { text: lines.join("\n"), count: { file, rules, withIds } };
|
|
60
73
|
}
|
|
@@ -65,22 +78,28 @@ function rewrite(text, file, findingCount, idFor) {
|
|
|
65
78
|
export function assignGuideIds(guide, detailed, findingCount) {
|
|
66
79
|
const rules = [];
|
|
67
80
|
const bySet = new Map();
|
|
68
|
-
const add = (file, numbers, line) => {
|
|
81
|
+
const add = (file, numbers, line, overturns = []) => {
|
|
69
82
|
const id = `R${rules.length + 1}`;
|
|
70
|
-
rules.push({
|
|
83
|
+
rules.push({
|
|
84
|
+
id,
|
|
85
|
+
file,
|
|
86
|
+
text: ruleText(line),
|
|
87
|
+
findings: numbers,
|
|
88
|
+
...(overturns.length ? { overturns } : {}),
|
|
89
|
+
});
|
|
71
90
|
return id;
|
|
72
91
|
};
|
|
73
92
|
if (detailed === undefined) {
|
|
74
|
-
const single = rewrite(guide, GUIDE_FILE, findingCount, (numbers, line) => add(GUIDE_FILE, numbers, line));
|
|
93
|
+
const single = rewrite(guide, GUIDE_FILE, findingCount, (numbers, line, overturns) => add(GUIDE_FILE, numbers, line, overturns));
|
|
75
94
|
return { guide: single.text, rules, counts: [single.count] };
|
|
76
95
|
}
|
|
77
|
-
const full = rewrite(detailed, DETAILED_GUIDE_FILE, findingCount, (numbers, line) => {
|
|
78
|
-
const id = add(DETAILED_GUIDE_FILE, numbers, line);
|
|
96
|
+
const full = rewrite(detailed, DETAILED_GUIDE_FILE, findingCount, (numbers, line, overturns) => {
|
|
97
|
+
const id = add(DETAILED_GUIDE_FILE, numbers, line, overturns);
|
|
79
98
|
if (!bySet.has(setKey(numbers)))
|
|
80
99
|
bySet.set(setKey(numbers), id);
|
|
81
100
|
return id;
|
|
82
101
|
});
|
|
83
|
-
const short = rewrite(guide, GUIDE_FILE, findingCount, (numbers, line) => bySet.get(setKey(numbers)) ?? add(GUIDE_FILE, numbers, line));
|
|
102
|
+
const short = rewrite(guide, GUIDE_FILE, findingCount, (numbers, line, overturns) => bySet.get(setKey(numbers)) ?? add(GUIDE_FILE, numbers, line, overturns));
|
|
84
103
|
return {
|
|
85
104
|
guide: short.text,
|
|
86
105
|
detailed: full.text,
|
|
@@ -40,3 +40,21 @@ export declare function compareRules(previous: Record<string, RuleEntry> | undef
|
|
|
40
40
|
/** `2 rules added (#402, #407) · 1 rewritten (R12, #340 joined) · 1 dropped
|
|
41
41
|
* (#101 left the window)`, or `no rule changed`. */
|
|
42
42
|
export declare function ruleChangeLine(changes: RuleChanges): string;
|
|
43
|
+
export type Contradiction = {
|
|
44
|
+
/** The old rule. */
|
|
45
|
+
id: string;
|
|
46
|
+
/** The new rule that goes against it. */
|
|
47
|
+
by: string;
|
|
48
|
+
/** The new rule's pull requests, `#340`. */
|
|
49
|
+
sources: string[];
|
|
50
|
+
/** The newest merge date of the new rule, `YYYY-MM-DD`. */
|
|
51
|
+
newest: string;
|
|
52
|
+
/** False when the old id is still in this build. */
|
|
53
|
+
dropped: boolean;
|
|
54
|
+
};
|
|
55
|
+
/** The old rules the guide AI said a new rule goes against, kept only when
|
|
56
|
+
* the old rule existed and the new rule's newest pull request was merged
|
|
57
|
+
* after the old rule's newest one. Without both dates nothing is claimed. */
|
|
58
|
+
export declare function contradictions(previous: Record<string, RuleEntry> | undefined, next: Record<string, RuleEntry>, overturns: Map<string, string[]>): Contradiction[];
|
|
59
|
+
/** `R4 dropped, R9 goes against it (#340, 2026-03)`. */
|
|
60
|
+
export declare function contradictionLine(item: Contradiction): string;
|
|
@@ -96,3 +96,44 @@ export function ruleChangeLine(changes) {
|
|
|
96
96
|
}
|
|
97
97
|
return parts.length > 0 ? parts.join(" · ") : "no rule changed";
|
|
98
98
|
}
|
|
99
|
+
function newestDate(entry) {
|
|
100
|
+
const dates = entry.sources.flatMap((source) => "pr" in source && source.mergedAt ? [source.mergedAt] : []);
|
|
101
|
+
return dates.length ? dates.sort().at(-1) : undefined;
|
|
102
|
+
}
|
|
103
|
+
/** The old rules the guide AI said a new rule goes against, kept only when
|
|
104
|
+
* the old rule existed and the new rule's newest pull request was merged
|
|
105
|
+
* after the old rule's newest one. Without both dates nothing is claimed. */
|
|
106
|
+
export function contradictions(previous, next, overturns) {
|
|
107
|
+
const found = [];
|
|
108
|
+
for (const [by, ids] of overturns) {
|
|
109
|
+
const entry = next[by];
|
|
110
|
+
if (!entry)
|
|
111
|
+
continue;
|
|
112
|
+
const newest = newestDate(entry);
|
|
113
|
+
if (!newest)
|
|
114
|
+
continue;
|
|
115
|
+
for (const id of ids) {
|
|
116
|
+
const old = previous?.[id];
|
|
117
|
+
if (!old || id === by)
|
|
118
|
+
continue;
|
|
119
|
+
const oldNewest = newestDate(old);
|
|
120
|
+
if (!oldNewest || newest <= oldNewest)
|
|
121
|
+
continue;
|
|
122
|
+
found.push({
|
|
123
|
+
id,
|
|
124
|
+
by,
|
|
125
|
+
sources: labels(entry),
|
|
126
|
+
newest,
|
|
127
|
+
dropped: !(id in next),
|
|
128
|
+
});
|
|
129
|
+
}
|
|
130
|
+
}
|
|
131
|
+
return found;
|
|
132
|
+
}
|
|
133
|
+
/** `R4 dropped, R9 goes against it (#340, 2026-03)`. */
|
|
134
|
+
export function contradictionLine(item) {
|
|
135
|
+
const evidence = `${listed(item.sources, 3)}, ${item.newest.slice(0, 7)}`;
|
|
136
|
+
return item.dropped
|
|
137
|
+
? `${item.id} dropped, ${item.by} goes against it (${evidence})`
|
|
138
|
+
: `${item.by} goes against ${item.id} (${evidence})`;
|
|
139
|
+
}
|
|
@@ -16,6 +16,10 @@ export declare const DETAILED_GUIDE_FILE = "PR_REVIEW_DETAILED_GUIDE.md";
|
|
|
16
16
|
export type RuleSource = {
|
|
17
17
|
pr: number;
|
|
18
18
|
mergedAt?: string;
|
|
19
|
+
/** The review comment the rule came from. Missing when the pull
|
|
20
|
+
* request was read before links were kept, or the rule came from its
|
|
21
|
+
* body or diff. */
|
|
22
|
+
url?: string;
|
|
19
23
|
} | {
|
|
20
24
|
path: string;
|
|
21
25
|
line: number;
|
|
@@ -35,6 +39,13 @@ export type RulesFile = {
|
|
|
35
39
|
schemaVersion: 1;
|
|
36
40
|
builtAt: string;
|
|
37
41
|
rules: Record<string, RuleEntry>;
|
|
42
|
+
/** The highest number ever given per kind, so a dropped id is never given
|
|
43
|
+
* to another rule. Missing in files written before ids were kept stable:
|
|
44
|
+
* the highest id in `rules` stands in. */
|
|
45
|
+
lastIds?: {
|
|
46
|
+
R: number;
|
|
47
|
+
I: number;
|
|
48
|
+
};
|
|
38
49
|
};
|
|
39
50
|
/** Ids are `R` for guide rules and `I` for instruction rules, then a number.
|
|
40
51
|
* `C` is kept for built-in checks. */
|
|
@@ -18,7 +18,8 @@ function isSource(value) {
|
|
|
18
18
|
return false;
|
|
19
19
|
const record = value;
|
|
20
20
|
if (typeof record.pr === "number") {
|
|
21
|
-
return record.mergedAt === undefined || typeof record.mergedAt === "string"
|
|
21
|
+
return ((record.mergedAt === undefined || typeof record.mergedAt === "string") &&
|
|
22
|
+
(record.url === undefined || typeof record.url === "string"));
|
|
22
23
|
}
|
|
23
24
|
return typeof record.path === "string" && typeof record.line === "number";
|
|
24
25
|
}
|
|
@@ -58,10 +59,18 @@ export function parseRules(text) {
|
|
|
58
59
|
sources,
|
|
59
60
|
};
|
|
60
61
|
}
|
|
62
|
+
const last = record.lastIds;
|
|
63
|
+
const lastIds = last &&
|
|
64
|
+
typeof last === "object" &&
|
|
65
|
+
Number.isInteger(last.R) &&
|
|
66
|
+
Number.isInteger(last.I)
|
|
67
|
+
? { R: Number(last.R), I: Number(last.I) }
|
|
68
|
+
: undefined;
|
|
61
69
|
return {
|
|
62
70
|
schemaVersion: 1,
|
|
63
71
|
builtAt: typeof record.builtAt === "string" ? record.builtAt : "",
|
|
64
72
|
rules,
|
|
73
|
+
...(lastIds ? { lastIds } : {}),
|
|
65
74
|
};
|
|
66
75
|
}
|
|
67
76
|
function idNumber(id) {
|
|
@@ -0,0 +1,29 @@
|
|
|
1
|
+
/** Rule ids that stay the same across builds.
|
|
2
|
+
*
|
|
3
|
+
* A build numbers its rules from 1 (`R1`, `I1`, ...). A review that cited
|
|
4
|
+
* `R4` must still point at the same rule after the next `sync`, so the fresh
|
|
5
|
+
* numbers are mapped onto the last build's ids before anything is written:
|
|
6
|
+
* a rule that shares a source with an old rule keeps that rule's id (it is the
|
|
7
|
+
* same rule, maybe reworded), a rule with no match gets the next number ever
|
|
8
|
+
* given, and a dropped id is never given again. */
|
|
9
|
+
import { type RuleEntry, type RuleSource } from "./rules.ts";
|
|
10
|
+
export type FreshRule = {
|
|
11
|
+
id: string;
|
|
12
|
+
text: string;
|
|
13
|
+
sources: RuleSource[];
|
|
14
|
+
/** Old ids this rule may not take, such as the rule it overturns. */
|
|
15
|
+
avoid?: string[];
|
|
16
|
+
};
|
|
17
|
+
/** Maps each fresh id to its stable id, and returns the new highest number.
|
|
18
|
+
* A fresh rule may keep an old id when it shares a source with that rule, or
|
|
19
|
+
* else has the same text. The ids are matched as a whole, not first come
|
|
20
|
+
* first served: a rule that could keep two ids gives way to a later rule that
|
|
21
|
+
* can keep only one of them, so as many ids as possible stay the same. Among
|
|
22
|
+
* its options a rule tries the one with the most shared sources first. */
|
|
23
|
+
export declare function stableIds(kind: "R" | "I", fresh: FreshRule[], previous: Record<string, RuleEntry> | undefined, lastId: number | undefined): {
|
|
24
|
+
ids: Map<string, string>;
|
|
25
|
+
lastId: number;
|
|
26
|
+
};
|
|
27
|
+
/** Rewrites the `[R1]` / `[I1]` id at the end of each line in one pass, so a
|
|
28
|
+
* rename such as R1 to R2 and R2 to R5 never collides. */
|
|
29
|
+
export declare function renameIdsInMarkdown(markdown: string, ids: Map<string, string>): string;
|
|
@@ -0,0 +1,90 @@
|
|
|
1
|
+
/** Rule ids that stay the same across builds.
|
|
2
|
+
*
|
|
3
|
+
* A build numbers its rules from 1 (`R1`, `I1`, ...). A review that cited
|
|
4
|
+
* `R4` must still point at the same rule after the next `sync`, so the fresh
|
|
5
|
+
* numbers are mapped onto the last build's ids before anything is written:
|
|
6
|
+
* a rule that shares a source with an old rule keeps that rule's id (it is the
|
|
7
|
+
* same rule, maybe reworded), a rule with no match gets the next number ever
|
|
8
|
+
* given, and a dropped id is never given again. */
|
|
9
|
+
import { sourceLabel } from "./rules.js";
|
|
10
|
+
/** The highest id number of one kind in `rules`. */
|
|
11
|
+
function highest(rules, kind) {
|
|
12
|
+
let max = 0;
|
|
13
|
+
for (const id of Object.keys(rules)) {
|
|
14
|
+
if (id.startsWith(kind))
|
|
15
|
+
max = Math.max(max, Number(id.slice(1)) || 0);
|
|
16
|
+
}
|
|
17
|
+
return max;
|
|
18
|
+
}
|
|
19
|
+
/** Maps each fresh id to its stable id, and returns the new highest number.
|
|
20
|
+
* A fresh rule may keep an old id when it shares a source with that rule, or
|
|
21
|
+
* else has the same text. The ids are matched as a whole, not first come
|
|
22
|
+
* first served: a rule that could keep two ids gives way to a later rule that
|
|
23
|
+
* can keep only one of them, so as many ids as possible stay the same. Among
|
|
24
|
+
* its options a rule tries the one with the most shared sources first. */
|
|
25
|
+
export function stableIds(kind, fresh, previous, lastId) {
|
|
26
|
+
const old = Object.entries(previous ?? {})
|
|
27
|
+
.filter(([id]) => id.startsWith(kind))
|
|
28
|
+
.map(([id, entry]) => ({
|
|
29
|
+
id,
|
|
30
|
+
text: entry.text,
|
|
31
|
+
sources: new Set(entry.sources.map(sourceLabel)),
|
|
32
|
+
}));
|
|
33
|
+
const rules = fresh.filter((rule, index) => fresh.findIndex((item) => item.id === rule.id) === index);
|
|
34
|
+
// Each rule's old ids it may keep, best first.
|
|
35
|
+
const options = rules.map((rule) => {
|
|
36
|
+
const labels = rule.sources.map(sourceLabel);
|
|
37
|
+
return old
|
|
38
|
+
.flatMap((item, order) => {
|
|
39
|
+
if (rule.avoid?.includes(item.id))
|
|
40
|
+
return [];
|
|
41
|
+
const shared = labels.filter((label) => item.sources.has(label)).length;
|
|
42
|
+
// Same text with no shared source still counts, below any shared
|
|
43
|
+
// source: an instruction line that moved down a file is the same rule.
|
|
44
|
+
const score = shared > 0 ? shared : item.text === rule.text ? 0.5 : 0;
|
|
45
|
+
return score > 0 ? [{ id: item.id, score, order }] : [];
|
|
46
|
+
})
|
|
47
|
+
.sort((a, b) => b.score - a.score || a.order - b.order)
|
|
48
|
+
.map((option) => option.id);
|
|
49
|
+
});
|
|
50
|
+
// Augmenting paths: a rule takes a free id, or moves the rule holding one
|
|
51
|
+
// of its options to another id that rule may keep.
|
|
52
|
+
const holder = new Map();
|
|
53
|
+
const kept = new Map();
|
|
54
|
+
const place = (index, seen) => {
|
|
55
|
+
for (const id of options[index]) {
|
|
56
|
+
if (seen.has(id))
|
|
57
|
+
continue;
|
|
58
|
+
seen.add(id);
|
|
59
|
+
const current = holder.get(id);
|
|
60
|
+
if (current === undefined || place(current, seen)) {
|
|
61
|
+
holder.set(id, index);
|
|
62
|
+
kept.set(index, id);
|
|
63
|
+
return true;
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
return false;
|
|
67
|
+
};
|
|
68
|
+
for (let index = 0; index < rules.length; index++)
|
|
69
|
+
place(index, new Set());
|
|
70
|
+
let next = Math.max(lastId ?? 0, highest(previous ?? {}, kind));
|
|
71
|
+
const ids = new Map();
|
|
72
|
+
for (const [index, rule] of rules.entries()) {
|
|
73
|
+
const id = kept.get(index);
|
|
74
|
+
if (id)
|
|
75
|
+
ids.set(rule.id, id);
|
|
76
|
+
else {
|
|
77
|
+
next++;
|
|
78
|
+
ids.set(rule.id, `${kind}${next}`);
|
|
79
|
+
}
|
|
80
|
+
}
|
|
81
|
+
return { ids, lastId: next };
|
|
82
|
+
}
|
|
83
|
+
/** Rewrites the `[R1]` / `[I1]` id at the end of each line in one pass, so a
|
|
84
|
+
* rename such as R1 to R2 and R2 to R5 never collides. */
|
|
85
|
+
export function renameIdsInMarkdown(markdown, ids) {
|
|
86
|
+
return markdown.replace(/\[([RI][1-9]\d*)\](?=[ \t]*$)/gm, (whole, id) => {
|
|
87
|
+
const renamed = ids.get(id);
|
|
88
|
+
return renamed ? `[${renamed}]` : whole;
|
|
89
|
+
});
|
|
90
|
+
}
|
|
@@ -1,6 +1,18 @@
|
|
|
1
1
|
import type { AiProvider, AiResponse, Options } from "../types.ts";
|
|
2
|
-
import type { Fact, Source } from "./types.ts";
|
|
2
|
+
import type { Fact, PullRequest, Source } from "./types.ts";
|
|
3
3
|
type UsageSink = (job: string, response: AiResponse) => Promise<void>;
|
|
4
|
+
/** Whether the pull request was read with comment links. A row cached before
|
|
5
|
+
* links were kept has none, and keeps the old extract prompt so its cached
|
|
6
|
+
* answer is still used. */
|
|
7
|
+
export declare function hasCommentLinks(pr: PullRequest): boolean;
|
|
8
|
+
/** The comment ids of a linked pull request (`c1` issue comment, `r1` inline
|
|
9
|
+
* review comment, `v1` review) and the link behind each. */
|
|
10
|
+
export declare function commentIds(pr: PullRequest): Map<string, string>;
|
|
11
|
+
/** The extract prompt for one pull request. A pull request read before
|
|
12
|
+
* comment links were kept gets the exact old text, so its cached answer still
|
|
13
|
+
* hits. A linked one numbers its comments and asks which ones a finding rests
|
|
14
|
+
* on. */
|
|
15
|
+
export declare function extractPrompt(pr: PullRequest): string;
|
|
4
16
|
export type SynthesisResult = Record<string, string>;
|
|
5
17
|
export declare function extractAiFacts(provider: AiProvider, repo: string, source: Source, options: Options, usage?: UsageSink): Promise<Fact[]>;
|
|
6
18
|
export declare function enrichFacts(provider: AiProvider, repo: string, base: Fact[], source: Source, options: Options, usage?: UsageSink): Promise<Fact[]>;
|
|
@@ -97,7 +97,7 @@ function parseJson(text) {
|
|
|
97
97
|
}
|
|
98
98
|
return [];
|
|
99
99
|
}
|
|
100
|
-
function factsFromResponse(text, evidence, defaultScope, origin) {
|
|
100
|
+
function factsFromResponse(text, evidence, defaultScope, origin, pr, ids) {
|
|
101
101
|
const parsed = parseJson(text);
|
|
102
102
|
if (!Array.isArray(parsed))
|
|
103
103
|
return [];
|
|
@@ -128,6 +128,7 @@ function factsFromResponse(text, evidence, defaultScope, origin) {
|
|
|
128
128
|
const trigger = String(value.trigger ?? "").trim();
|
|
129
129
|
if (isReview && (!impact || !action))
|
|
130
130
|
return [];
|
|
131
|
+
const links = isReview ? linksFor(value.sources, pr, ids) : [];
|
|
131
132
|
return [
|
|
132
133
|
{
|
|
133
134
|
id: `${sectionKey}:${slug(claim)}`,
|
|
@@ -146,6 +147,7 @@ function factsFromResponse(text, evidence, defaultScope, origin) {
|
|
|
146
147
|
...(trigger ? { trigger } : {}),
|
|
147
148
|
impact,
|
|
148
149
|
action,
|
|
150
|
+
...(links.length ? { links } : {}),
|
|
149
151
|
}
|
|
150
152
|
: {}),
|
|
151
153
|
},
|
|
@@ -187,19 +189,88 @@ function hasUnsupportedPath(claim, source) {
|
|
|
187
189
|
![...paths].some((path) => wildcard ? wildcard.test(path) : path.startsWith(`${normalized}/`)));
|
|
188
190
|
});
|
|
189
191
|
}
|
|
190
|
-
|
|
192
|
+
/** Whether the pull request was read with comment links. A row cached before
|
|
193
|
+
* links were kept has none, and keeps the old extract prompt so its cached
|
|
194
|
+
* answer is still used. */
|
|
195
|
+
export function hasCommentLinks(pr) {
|
|
196
|
+
return (pr.commentUrls !== undefined ||
|
|
197
|
+
pr.reviewCommentUrls !== undefined ||
|
|
198
|
+
pr.reviewUrls !== undefined);
|
|
199
|
+
}
|
|
200
|
+
/** The comment ids of a linked pull request (`c1` issue comment, `r1` inline
|
|
201
|
+
* review comment, `v1` review) and the link behind each. */
|
|
202
|
+
export function commentIds(pr) {
|
|
203
|
+
const ids = new Map();
|
|
204
|
+
const add = (prefix, urls, count) => {
|
|
205
|
+
for (let index = 0; index < count; index++) {
|
|
206
|
+
const url = urls?.[index] ?? "";
|
|
207
|
+
if (url)
|
|
208
|
+
ids.set(`${prefix}${index + 1}`, url);
|
|
209
|
+
}
|
|
210
|
+
};
|
|
211
|
+
add("c", pr.commentUrls, pr.comments.length);
|
|
212
|
+
add("r", pr.reviewCommentUrls, pr.reviewComments.length);
|
|
213
|
+
add("v", pr.reviewUrls, pr.reviews.length);
|
|
214
|
+
return ids;
|
|
215
|
+
}
|
|
216
|
+
function pullRequestUnit(pr, numbered = false) {
|
|
217
|
+
const texts = (prefix, values) => values.map((value, index) => {
|
|
218
|
+
const text = promptText(value, 2_000);
|
|
219
|
+
return numbered ? `[${prefix}${index + 1}] ${text}` : text;
|
|
220
|
+
});
|
|
191
221
|
return JSON.stringify({
|
|
192
222
|
number: pr.number,
|
|
193
223
|
title: pr.title,
|
|
194
224
|
body: promptText(pr.body),
|
|
195
225
|
labels: pr.labels,
|
|
196
|
-
comments: pr.comments
|
|
197
|
-
reviewComments: pr.reviewComments
|
|
198
|
-
reviews: pr.reviews
|
|
226
|
+
comments: texts("c", pr.comments),
|
|
227
|
+
reviewComments: texts("r", pr.reviewComments),
|
|
228
|
+
reviews: texts("v", pr.reviews),
|
|
199
229
|
changedFiles: pr.changedFiles,
|
|
200
230
|
diff: promptText(pr.diff, 8_000),
|
|
201
231
|
});
|
|
202
232
|
}
|
|
233
|
+
/** The extract prompt for one pull request. A pull request read before
|
|
234
|
+
* comment links were kept gets the exact old text, so its cached answer still
|
|
235
|
+
* hits. A linked one numbers its comments and asks which ones a finding rests
|
|
236
|
+
* on. */
|
|
237
|
+
export function extractPrompt(pr) {
|
|
238
|
+
const numbered = hasCommentLinks(pr);
|
|
239
|
+
const sourcesField = numbered
|
|
240
|
+
? ',"sources":["the ids of the comments or reviews the finding comes from, such as r2 or v1, empty when it comes from the body or the diff"]'
|
|
241
|
+
: "";
|
|
242
|
+
const sourcesRule = numbered
|
|
243
|
+
? "\n- Every comment and review starts with its id in square brackets. List in sources only ids that appear in this pull request."
|
|
244
|
+
: "";
|
|
245
|
+
return `Analyze this pull request's reviews, comments, code and diff, and return only a JSON array of review findings.
|
|
246
|
+
Each item must have exactly:
|
|
247
|
+
{"sectionKey":"review","class":"a short area name","claim":"the invariant a reviewer would enforce, imperative and general","location":"path or path:line the finding is about, or empty when none","trigger":"the diff surface that should trigger this rule: a path, glob, or symbol, or empty when none","impact":"why this matters here, from the review's reasoning","action":"what to change or verify","scope":"historical-example","confidence":"medium"${sourcesField}}
|
|
248
|
+
Rules:
|
|
249
|
+
- Include only substantive findings: a reviewer identifies something to fix, change, verify or decide.
|
|
250
|
+
- Never write an approval, a merge, "LGTM", thanks, a reviewer name, a pull-request number, a commit id, or quoted comment text.
|
|
251
|
+
- The claim is a general invariant, not one past fix: "When X changes, Y must also change", or "Every Z uses W", never "change this one line".
|
|
252
|
+
- trigger is what a reviewer can see in a diff: a path, a glob, or a symbol. It is how the rule applies to a future change, so a rule with no trigger is a one-off and is not useful.
|
|
253
|
+
- impact and action are mandatory and concrete. Paraphrase the reviewer's reasoning for impact. Never quote it.
|
|
254
|
+
- class is a short area label chosen from a small, consistent set. Reuse the same label for findings of the same kind so related findings can be grouped.
|
|
255
|
+
- location is a repository path when the review names one, otherwise leave it empty.${sourcesRule}
|
|
256
|
+
- Return every distinct finding the reviews support, not a summary. Do not explain the answer.
|
|
257
|
+
|
|
258
|
+
PULL REQUEST:
|
|
259
|
+
${pullRequestUnit(pr, numbered)}`;
|
|
260
|
+
}
|
|
261
|
+
/** The links a finding's `sources` name. An id the pull request does not have
|
|
262
|
+
* is dropped, so a made-up id never becomes a link. */
|
|
263
|
+
function linksFor(value, pr, ids) {
|
|
264
|
+
if (pr === undefined || !ids || !Array.isArray(value))
|
|
265
|
+
return [];
|
|
266
|
+
const urls = new Set();
|
|
267
|
+
for (const item of value) {
|
|
268
|
+
const url = ids.get(String(item).trim().toLowerCase());
|
|
269
|
+
if (url)
|
|
270
|
+
urls.add(url);
|
|
271
|
+
}
|
|
272
|
+
return [...urls].map((url) => ({ pr, url }));
|
|
273
|
+
}
|
|
203
274
|
function queueFor(provider, repo, ai, model, concurrency, validate) {
|
|
204
275
|
return new AiBatch(provider, repo, Math.max(1, concurrency), `${ai}:v3:${model}`, validate);
|
|
205
276
|
}
|
|
@@ -291,23 +362,15 @@ export async function extractAiFacts(provider, repo, source, options, usage) {
|
|
|
291
362
|
const requests = [];
|
|
292
363
|
const evidence = [];
|
|
293
364
|
const scopes = [];
|
|
365
|
+
// Per request: the comment ids of a linked pull request and its number,
|
|
366
|
+
// `undefined` for the repository files request and unlinked pull requests.
|
|
367
|
+
const linkIds = [];
|
|
368
|
+
const prNumbers = [];
|
|
294
369
|
const prs = options.includePullRequests ? source.pullRequests : [];
|
|
295
370
|
for (const pr of prs) {
|
|
296
|
-
requests.push(extractRequest(
|
|
297
|
-
|
|
298
|
-
|
|
299
|
-
Rules:
|
|
300
|
-
- Include only substantive findings: a reviewer identifies something to fix, change, verify or decide.
|
|
301
|
-
- Never write an approval, a merge, "LGTM", thanks, a reviewer name, a pull-request number, a commit id, or quoted comment text.
|
|
302
|
-
- The claim is a general invariant, not one past fix: "When X changes, Y must also change", or "Every Z uses W", never "change this one line".
|
|
303
|
-
- trigger is what a reviewer can see in a diff: a path, a glob, or a symbol. It is how the rule applies to a future change, so a rule with no trigger is a one-off and is not useful.
|
|
304
|
-
- impact and action are mandatory and concrete. Paraphrase the reviewer's reasoning for impact. Never quote it.
|
|
305
|
-
- class is a short area label chosen from a small, consistent set. Reuse the same label for findings of the same kind so related findings can be grouped.
|
|
306
|
-
- location is a repository path when the review names one, otherwise leave it empty.
|
|
307
|
-
- Return every distinct finding the reviews support, not a summary. Do not explain the answer.
|
|
308
|
-
|
|
309
|
-
PULL REQUEST:
|
|
310
|
-
${pullRequestUnit(pr)}`, 2_400));
|
|
371
|
+
requests.push(extractRequest(extractPrompt(pr), 2_400));
|
|
372
|
+
linkIds.push(hasCommentLinks(pr) ? commentIds(pr) : undefined);
|
|
373
|
+
prNumbers.push(pr.number);
|
|
311
374
|
evidence.push(`PR #${pr.number}`);
|
|
312
375
|
scopes.push("historical-example");
|
|
313
376
|
}
|
|
@@ -333,6 +396,8 @@ Return at most 10 items. Keep each claim under 180 characters. Do not explain th
|
|
|
333
396
|
FILES:
|
|
334
397
|
${files}`, 2_200));
|
|
335
398
|
evidence.push("repository files");
|
|
399
|
+
linkIds.push(undefined);
|
|
400
|
+
prNumbers.push(undefined);
|
|
336
401
|
scopes.push("current");
|
|
337
402
|
}
|
|
338
403
|
}
|
|
@@ -349,7 +414,7 @@ ${files}`, 2_200));
|
|
|
349
414
|
const response = responses[index];
|
|
350
415
|
if (!response)
|
|
351
416
|
continue;
|
|
352
|
-
facts.push(...factsFromResponse(response.text, evidence[index], scopes[index] ?? "historical-example", evidence[index] === "repository files" ? "repository" : "pull-request").filter((item) => {
|
|
417
|
+
facts.push(...factsFromResponse(response.text, evidence[index], scopes[index] ?? "historical-example", evidence[index] === "repository files" ? "repository" : "pull-request", prNumbers[index], linkIds[index]).filter((item) => {
|
|
353
418
|
if (hasUnsupportedIdentifier(item.claim, source) ||
|
|
354
419
|
hasUnsupportedPath(item.claim, source)) {
|
|
355
420
|
return false;
|
|
@@ -388,6 +453,9 @@ function mergeFacts(base, extra) {
|
|
|
388
453
|
existing.evidence = [
|
|
389
454
|
...new Set([...existing.evidence, ...item.evidence]),
|
|
390
455
|
];
|
|
456
|
+
const links = [...(existing.links ?? []), ...(item.links ?? [])].filter((link, index, all) => all.findIndex((other) => other.url === link.url) === index);
|
|
457
|
+
if (links.length)
|
|
458
|
+
existing.links = links;
|
|
391
459
|
}
|
|
392
460
|
}
|
|
393
461
|
return [...merged.values()];
|
|
@@ -19,6 +19,13 @@ export type PullRequest = {
|
|
|
19
19
|
* them otherwise. */
|
|
20
20
|
reviewComments: string[];
|
|
21
21
|
reviews: string[];
|
|
22
|
+
/** The GitHub link of each entry in `comments`, `reviewComments` and
|
|
23
|
+
* `reviews`, at the same index. Missing on rows cached before links were
|
|
24
|
+
* kept, and an entry is empty when GitHub gave no link. Never sent to a
|
|
25
|
+
* model: a fact names the numbered comment and code looks the link up. */
|
|
26
|
+
commentUrls?: string[];
|
|
27
|
+
reviewCommentUrls?: string[];
|
|
28
|
+
reviewUrls?: string[];
|
|
22
29
|
/** Set once reviews have been read. Missing means an older checkpoint
|
|
23
30
|
* never stored the review state. */
|
|
24
31
|
changesRequested?: boolean;
|
|
@@ -72,6 +79,15 @@ export type Fact = {
|
|
|
72
79
|
* also do Y" obligations listed under a pass's exit check, so a reviewer that
|
|
73
80
|
* only sees a diff can still apply the rule. */
|
|
74
81
|
trigger?: string;
|
|
82
|
+
/** Review findings only: the GitHub link of each comment or review the
|
|
83
|
+
* finding came from, with its pull request. The model names a numbered
|
|
84
|
+
* comment, code turns it into this link. Missing when the finding came from
|
|
85
|
+
* the body or the diff, or from a pull request cached before links were
|
|
86
|
+
* kept. */
|
|
87
|
+
links?: {
|
|
88
|
+
pr: number;
|
|
89
|
+
url: string;
|
|
90
|
+
}[];
|
|
75
91
|
};
|
|
76
92
|
export type State = {
|
|
77
93
|
version: 1;
|
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
import { parseArgs } from "../cli/args.js";
|
|
2
|
+
import { readConfig } from "../config.js";
|
|
2
3
|
import { filterReviewConfigArgs, loadReviewSystemPromptFlags, reviewSystemPromptLimitFor, } from "../cli/review_args.js";
|
|
3
4
|
import { buildLocalReviewJson, formatHumanLocalReview, humanFindingsFromResolved, reviewExitCodeFromResolved, } from "../cli/review_result.js";
|
|
4
5
|
import { addAiMetrics, emptyAiMetrics, recordAiCost, runInitOrRemake, } from "../services/setup.js";
|
|
@@ -245,7 +246,7 @@ export async function runLocalReview(cli) {
|
|
|
245
246
|
return;
|
|
246
247
|
const codegraphState = codegraphPrep.state;
|
|
247
248
|
warnings.push(...response.warnings);
|
|
248
|
-
const { resolved: allResolved, visiblePaths } = finishCarryRun(carryRun, revision, response);
|
|
249
|
+
const { resolved: allResolved, visiblePaths } = finishCarryRun(carryRun, revision, response, options.repo);
|
|
249
250
|
const findingsToStore = storedFindings(allResolved);
|
|
250
251
|
const afterBuilt = await buildLocalRevision(root, baseSha, base.label, runCommand);
|
|
251
252
|
const revisionHashAfter = await revisionHash(afterBuilt.revision);
|
|
@@ -315,7 +316,7 @@ export async function runLocalReview(cli) {
|
|
|
315
316
|
}
|
|
316
317
|
else {
|
|
317
318
|
console.log("\n" +
|
|
318
|
-
formatHumanLocalReview(header, revision, response.guideBuiltAt, codegraphState, allResolved, warnings, options.reviewBlocking, response.coverage, extensions, Boolean(intentLoad.intent)) +
|
|
319
|
+
formatHumanLocalReview(header, revision, response.guideBuiltAt, codegraphState, allResolved, warnings, options.reviewBlocking, response.coverage, extensions, Boolean(intentLoad.intent), readConfig().repos?.[options.repo]?.reviewReferences === true) +
|
|
319
320
|
"\n");
|
|
320
321
|
printRunSummary(run);
|
|
321
322
|
}
|
|
@@ -10,5 +10,8 @@ export type ParsedFinding = Span & {
|
|
|
10
10
|
blocking?: boolean;
|
|
11
11
|
symbol?: string;
|
|
12
12
|
summary?: string;
|
|
13
|
+
/** The ids from the finding's `References:` line, as the model wrote them.
|
|
14
|
+
* Missing when the line is absent. */
|
|
15
|
+
references?: string[];
|
|
13
16
|
};
|
|
14
17
|
export declare function parseFindings(markdown: string): ParsedFinding[];
|
package/dist/src/pr/findings.js
CHANGED
|
@@ -22,8 +22,24 @@ function location(block) {
|
|
|
22
22
|
to: found.groups.to ? Number(found.groups.to) : from,
|
|
23
23
|
};
|
|
24
24
|
}
|
|
25
|
+
/** The `References:` line, only right under the Location line where
|
|
26
|
+
* `findingMarkdown` writes it, so a body sentence starting with the same word
|
|
27
|
+
* is never taken. */
|
|
28
|
+
const REFERENCES_LINE = /^(\s*(?:Location|File)\s*:[^\n]*)\n\s*References\s*:\s*([^\n]+?)\s*$/im;
|
|
29
|
+
function references(block) {
|
|
30
|
+
const match = REFERENCES_LINE.exec(block);
|
|
31
|
+
if (!match)
|
|
32
|
+
return undefined;
|
|
33
|
+
return match[2]
|
|
34
|
+
.split(",")
|
|
35
|
+
.map((id) => id.trim())
|
|
36
|
+
.filter(Boolean);
|
|
37
|
+
}
|
|
25
38
|
function bodyWithoutLocation(block) {
|
|
26
|
-
return block
|
|
39
|
+
return block
|
|
40
|
+
.replace(REFERENCES_LINE, "$1")
|
|
41
|
+
.replace(/^\s*(?:Location|File)\s*:\s*`?.+?`?\s*$/im, "")
|
|
42
|
+
.trim();
|
|
27
43
|
}
|
|
28
44
|
function newFinding(header, block) {
|
|
29
45
|
const meta = NEW_HEADING.exec(header);
|
|
@@ -51,6 +67,10 @@ function newFinding(header, block) {
|
|
|
51
67
|
blocking: impact === "blocking",
|
|
52
68
|
symbol: symbol || undefined,
|
|
53
69
|
summary,
|
|
70
|
+
...(() => {
|
|
71
|
+
const ids = references(block);
|
|
72
|
+
return ids ? { references: ids } : {};
|
|
73
|
+
})(),
|
|
54
74
|
};
|
|
55
75
|
}
|
|
56
76
|
export function parseFindings(markdown) {
|
|
@@ -13,7 +13,7 @@ import type { ParsedFinding } from "./findings.ts";
|
|
|
13
13
|
* get the prompt's instruction to return a fenced JSON block. */
|
|
14
14
|
export declare const FINDINGS_JSON_SCHEMA: Json;
|
|
15
15
|
/** The prompt's contract, kept next to the schema so the two cannot drift. */
|
|
16
|
-
export declare const FINDINGS_JSON_INSTRUCTIONS = "Return a single JSON object, and nothing else, with this shape:\n{\"findings\":[{\"severity\":\"P1\",\"blocking\":true,\"path\":\"src/a.ts\",\"lineFrom\":42,\"lineTo\":42,\"symbol\":\"helper()\",\"title\":\"The helper ignores its argument\",\"body\":\"One sentence on what is wrong and its impact, then a short evidence paragraph.\",\"suggestion\":\"the exact replacement lines, or omit this field\"}],\"previousFindings\":[{\"id\":\"F1\",\"state\":\"open\"}]}\n\nRules for the JSON:\n- \"severity\" is exactly one of P0, P1, P2, P3. \"blocking\" is true or false.\n- \"lineFrom\" and \"lineTo\" are numbers in the new file, copied from the DIFF\n column, never counted from the @@ header. \"lineTo\" may equal \"lineFrom\".\n- \"title\" is one sentence naming the defect, without the severity or the path.\n- \"body\" is the explanation. Do not repeat the title, the path, or the location\n line inside it. Our code renders those. Do not add a closing sentence asking\n whether to explain more, and do not use the words Mechanism, Symptom,\n Scenario, Verified, Repro, Options, or Scope as labels.\n- \"suggestion\" holds only the replacement source lines, with their original\n indentation, when the fix replaces the exact lines lineFrom-lineTo in one\n hunk of the same file. Omit it otherwise. Never wrap it in a code fence.\n- When a fix needs removed lines, another file, or more than one hunk, omit\n \"suggestion\".\n- Return every independently actionable finding, including none: use\n {\"findings\":[]} when the change is clean. Never invent a finding to fill the\n array.";
|
|
16
|
+
export declare const FINDINGS_JSON_INSTRUCTIONS = "Return a single JSON object, and nothing else, with this shape:\n{\"findings\":[{\"severity\":\"P1\",\"blocking\":true,\"path\":\"src/a.ts\",\"lineFrom\":42,\"lineTo\":42,\"symbol\":\"helper()\",\"title\":\"The helper ignores its argument\",\"body\":\"One sentence on what is wrong and its impact, then a short evidence paragraph.\",\"suggestion\":\"the exact replacement lines, or omit this field\",\"references\":[\"R12\"]}],\"previousFindings\":[{\"id\":\"F1\",\"state\":\"open\"}]}\n\nRules for the JSON:\n- \"severity\" is exactly one of P0, P1, P2, P3. \"blocking\" is true or false.\n- \"lineFrom\" and \"lineTo\" are numbers in the new file, copied from the DIFF\n column, never counted from the @@ header. \"lineTo\" may equal \"lineFrom\".\n- \"title\" is one sentence naming the defect, without the severity or the path.\n- \"body\" is the explanation. Do not repeat the title, the path, or the location\n line inside it. Our code renders those. Do not add a closing sentence asking\n whether to explain more, and do not use the words Mechanism, Symptom,\n Scenario, Verified, Repro, Options, or Scope as labels.\n- \"suggestion\" holds only the replacement source lines, with their original\n indentation, when the fix replaces the exact lines lineFrom-lineTo in one\n hunk of the same file. Omit it otherwise. Never wrap it in a code fence.\n- When a fix needs removed lines, another file, or more than one hunk, omit\n \"suggestion\".\n- \"references\" lists what the finding rests on: the bracketed id at the end\n of a guide or repository instruction line (such as R12 or I3),\n \"codebase:<section title>\" for a CODEBASE CONVENTIONS section, or \"system\"\n when it rests only on these review instructions. Name only ids you used.\n- Return every independently actionable finding, including none: use\n {\"findings\":[]} when the change is clean. Never invent a finding to fill the\n array.";
|
|
17
17
|
/** Reads the structured output and renders it as Markdown. `undefined` means
|
|
18
18
|
* the text was not usable JSON, so the caller falls back to the legacy
|
|
19
19
|
* Markdown parser rather than dropping the review. */
|