co-maintainer 0.7.9 → 0.7.10-beta.2
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/ai/mermaid_loop.d.ts +6 -0
- package/dist/src/ai/mermaid_loop.js +19 -5
- package/dist/src/ai/tool_gate.d.ts +77 -0
- package/dist/src/ai/tool_gate.js +151 -0
- package/dist/src/cli/commands/registry.js +18 -0
- package/dist/src/cli/commands/review.js +2 -0
- package/dist/src/cli/commands/set.js +28 -2
- package/dist/src/cli/review_result.d.ts +6 -1
- package/dist/src/cli/review_result.js +6 -1
- package/dist/src/config.d.ts +10 -0
- package/dist/src/local/review_local.js +4 -2
- package/dist/src/pr/checkout.d.ts +5 -0
- package/dist/src/pr/checkout.js +13 -2
- package/dist/src/pr/codegraph_tools.d.ts +6 -4
- package/dist/src/pr/codegraph_tools.js +10 -7
- package/dist/src/pr/review_tools.d.ts +55 -0
- package/dist/src/pr/review_tools.js +69 -0
- package/dist/src/pr/reviewer.d.ts +12 -3
- package/dist/src/pr/reviewer.js +79 -50
- package/dist/src/pr/scope.js +2 -2
- package/dist/src/remote/client.js +10 -1
- package/dist/src/remote/push_client.js +3 -0
- package/dist/src/remote/schema.js +1 -0
- package/dist/src/services/push_review.js +2 -0
- package/dist/src/services/remote_review.js +18 -5
- package/dist/src/tools/codegraph.d.ts +7 -0
- package/dist/src/tools/codegraph.js +81 -0
- package/dist/src/tools/search_non_code.d.ts +58 -0
- package/dist/src/tools/search_non_code.js +255 -0
- package/package.json +1 -1
|
@@ -0,0 +1,55 @@
|
|
|
1
|
+
/** The tools a review offers the model, prepared in one place for every mode
|
|
2
|
+
* (issue 79): codegraph when it is on, `search-non-code` where a checkout is
|
|
3
|
+
* there to search, and `read-full-diff` from the diff plan. A later kind of
|
|
4
|
+
* tool, such as a repository's MCP servers, joins the list here. */
|
|
5
|
+
import type { ToolHandler } from "../ai/mermaid_loop.ts";
|
|
6
|
+
import type { Options } from "../types.ts";
|
|
7
|
+
import { type Run } from "./checkout.ts";
|
|
8
|
+
import { type SearchOptions } from "../tools/search_non_code.ts";
|
|
9
|
+
/** Where a review's tools run. */
|
|
10
|
+
export type ReviewToolSource =
|
|
11
|
+
/** A checkout co-maintainer makes of the repository: the pull request's
|
|
12
|
+
* worktree for CLI PR and App reviews, worktree 0 at the pushed commit for
|
|
13
|
+
* a push review, which gets no search. */
|
|
14
|
+
{
|
|
15
|
+
kind: "checkout";
|
|
16
|
+
pr: number;
|
|
17
|
+
commit: string;
|
|
18
|
+
search: boolean;
|
|
19
|
+
allowInstall?: boolean;
|
|
20
|
+
}
|
|
21
|
+
/** A tree already on disk with its codegraph tools ready: local review at
|
|
22
|
+
* the git root. */
|
|
23
|
+
| {
|
|
24
|
+
kind: "tree";
|
|
25
|
+
root: string;
|
|
26
|
+
codegraph: ToolHandler[];
|
|
27
|
+
}
|
|
28
|
+
/** Tools another process runs: a remote review, where the client runs
|
|
29
|
+
* codegraph and search in its own checkout. */
|
|
30
|
+
| {
|
|
31
|
+
kind: "bridge";
|
|
32
|
+
tools: ToolHandler[];
|
|
33
|
+
};
|
|
34
|
+
export type ReviewTools = {
|
|
35
|
+
/** Every tool to offer, codegraph first. */
|
|
36
|
+
tools: ToolHandler[];
|
|
37
|
+
/** The codegraph tools alone, which decide `codegraphState`. */
|
|
38
|
+
codegraph: ToolHandler[];
|
|
39
|
+
};
|
|
40
|
+
export type ReviewToolsDeps = {
|
|
41
|
+
run?: Run;
|
|
42
|
+
/** Overrides the codegraph setup, for tests. */
|
|
43
|
+
codegraph?: (worktree: () => Promise<string>) => Promise<ToolHandler[]>;
|
|
44
|
+
/** Overrides the repository's search settings, for tests. */
|
|
45
|
+
search?: SearchOptions;
|
|
46
|
+
};
|
|
47
|
+
export declare function isCodegraphTool(name: string): boolean;
|
|
48
|
+
/** The review's tools. `diffTools` is the diff plan's own tools
|
|
49
|
+
* (`read-full-diff`), awaited last so the plan and the checkout can be
|
|
50
|
+
* prepared at the same time. A setup that fails leaves its tools out and the
|
|
51
|
+
* review goes on with the rest. */
|
|
52
|
+
export declare function prepareReviewTools(source: ReviewToolSource, options: Pick<Options, "repo" | "useCodegraph">, diffTools: ToolHandler[] | Promise<ToolHandler[]>, deps?: ReviewToolsDeps): Promise<ReviewTools>;
|
|
53
|
+
/** `codegraphState` from the codegraph tools alone, never from the length of
|
|
54
|
+
* the whole tool list. */
|
|
55
|
+
export declare function codegraphState(useCodegraph: boolean | undefined, codegraph: ToolHandler[]): "used" | "disabled" | "unavailable";
|
|
@@ -0,0 +1,69 @@
|
|
|
1
|
+
import { log } from "../util/log.js";
|
|
2
|
+
import { ensureWorktree, runCommand } from "./checkout.js";
|
|
3
|
+
import { prepareCodegraphTools } from "./codegraph_tools.js";
|
|
4
|
+
import { searchNonCodeTool, searchOptionsFor, } from "../tools/search_non_code.js";
|
|
5
|
+
export function isCodegraphTool(name) {
|
|
6
|
+
return name.startsWith("codegraph-");
|
|
7
|
+
}
|
|
8
|
+
/** The review's tools. `diffTools` is the diff plan's own tools
|
|
9
|
+
* (`read-full-diff`), awaited last so the plan and the checkout can be
|
|
10
|
+
* prepared at the same time. A setup that fails leaves its tools out and the
|
|
11
|
+
* review goes on with the rest. */
|
|
12
|
+
export async function prepareReviewTools(source, options, diffTools, deps = {}) {
|
|
13
|
+
const [own, diff] = await Promise.all([
|
|
14
|
+
prepareOwnTools(source, options, deps),
|
|
15
|
+
diffTools,
|
|
16
|
+
]);
|
|
17
|
+
return { ...own, tools: [...own.tools, ...diff] };
|
|
18
|
+
}
|
|
19
|
+
async function prepareOwnTools(source, options, deps) {
|
|
20
|
+
const search = (root) => searchNonCodeTool(root, deps.search ?? searchOptionsFor(options.repo));
|
|
21
|
+
if (source.kind === "bridge") {
|
|
22
|
+
const codegraph = options.useCodegraph
|
|
23
|
+
? source.tools.filter((tool) => isCodegraphTool(tool.name))
|
|
24
|
+
: [];
|
|
25
|
+
const others = source.tools.filter((tool) => !isCodegraphTool(tool.name));
|
|
26
|
+
return { tools: [...codegraph, ...others], codegraph };
|
|
27
|
+
}
|
|
28
|
+
if (source.kind === "tree") {
|
|
29
|
+
const codegraph = options.useCodegraph ? source.codegraph : [];
|
|
30
|
+
return {
|
|
31
|
+
tools: [...codegraph, search(source.root)],
|
|
32
|
+
codegraph,
|
|
33
|
+
};
|
|
34
|
+
}
|
|
35
|
+
if (!options.useCodegraph && !source.search) {
|
|
36
|
+
return { tools: [], codegraph: [] };
|
|
37
|
+
}
|
|
38
|
+
// One checkout for both, made once, and only when something needs it.
|
|
39
|
+
const run = deps.run ?? runCommand;
|
|
40
|
+
let checkout;
|
|
41
|
+
const worktree = () => (checkout ??= ensureWorktree(options.repo, source.pr, source.commit, run));
|
|
42
|
+
let root;
|
|
43
|
+
if (source.search) {
|
|
44
|
+
try {
|
|
45
|
+
root = await worktree();
|
|
46
|
+
}
|
|
47
|
+
catch (error) {
|
|
48
|
+
log("checkout", `search unavailable · ${String(error)}`);
|
|
49
|
+
}
|
|
50
|
+
}
|
|
51
|
+
const codegraph = options.useCodegraph
|
|
52
|
+
? await (deps.codegraph
|
|
53
|
+
? deps.codegraph(worktree)
|
|
54
|
+
: prepareCodegraphTools(worktree, {
|
|
55
|
+
allowInstall: source.allowInstall,
|
|
56
|
+
}))
|
|
57
|
+
: [];
|
|
58
|
+
return {
|
|
59
|
+
tools: [...codegraph, ...(root ? [search(root)] : [])],
|
|
60
|
+
codegraph,
|
|
61
|
+
};
|
|
62
|
+
}
|
|
63
|
+
/** `codegraphState` from the codegraph tools alone, never from the length of
|
|
64
|
+
* the whole tool list. */
|
|
65
|
+
export function codegraphState(useCodegraph, codegraph) {
|
|
66
|
+
if (!useCodegraph)
|
|
67
|
+
return "disabled";
|
|
68
|
+
return codegraph.length > 0 ? "used" : "unavailable";
|
|
69
|
+
}
|
|
@@ -1,6 +1,7 @@
|
|
|
1
|
-
import { type ToolHandler } from "../ai/mermaid_loop.ts";
|
|
2
1
|
import type { ReferenceContext } from "../review/references.ts";
|
|
3
2
|
import type { Revision } from "../review/revision.ts";
|
|
3
|
+
import { type ReviewToolSource } from "./review_tools.ts";
|
|
4
|
+
import { type ToolUsage } from "../ai/tool_gate.ts";
|
|
4
5
|
import { type Coverage, type CoverageWarning } from "../review/coverage.ts";
|
|
5
6
|
import type { Snapshot } from "./snapshot.ts";
|
|
6
7
|
import type { AiProvider, AiResponse, GitHubClient, Json, Options } from "../types.ts";
|
|
@@ -13,7 +14,10 @@ export declare function modelSummary(options: Options): string;
|
|
|
13
14
|
export type ReviewExtras = {
|
|
14
15
|
carryPrompt?: string;
|
|
15
16
|
unchangedPaths?: string[];
|
|
16
|
-
|
|
17
|
+
/** Where the review's tools run. Left out, a pull request review checks
|
|
18
|
+
* the pull request out, and a workspace review is a push review, which
|
|
19
|
+
* checks out the pushed commit only for codegraph. */
|
|
20
|
+
tools?: ReviewToolSource;
|
|
17
21
|
/** The pull request and its files, already read by the caller. The review
|
|
18
22
|
* then makes neither request, so it judges what the caller stored. */
|
|
19
23
|
prefetched?: {
|
|
@@ -49,8 +53,11 @@ export type ReviewOutcome = AiResponse & {
|
|
|
49
53
|
referenceContext: ReferenceContext;
|
|
50
54
|
codegraphState: "used" | "disabled" | "unavailable";
|
|
51
55
|
coverage: Coverage;
|
|
52
|
-
/** `diff_dropped`, `diff_withheld
|
|
56
|
+
/** `diff_dropped`, `diff_withheld`, `summary_failed` and
|
|
57
|
+
* `tool_budget_spent`. */
|
|
53
58
|
warnings: CoverageWarning[];
|
|
59
|
+
/** What the review's tools did, for `usage.tools`. */
|
|
60
|
+
tools: ToolUsage;
|
|
54
61
|
/** Every changed file of the subject, for the result's `revision` block. */
|
|
55
62
|
diffStats: {
|
|
56
63
|
files: number;
|
|
@@ -76,6 +83,8 @@ export declare function reviewGuidesBlock(guide: string, detailed: string, codeb
|
|
|
76
83
|
/** Shared PR and local/remote workspace review instructions: confirm claims
|
|
77
84
|
* against the indexed graph, not only the diff slice. */
|
|
78
85
|
export declare const CODEGRAPH_DIFF_VERIFICATION = "Examine the changes line by line, not just file by file. A single file can\ncontain more than one independent defect, and a change that looks fine in\nisolation can be wrong once you trace what calls it or what else it affects.\nWhen a finding depends on behavior outside the changed lines, confirm it with\ncodegraph-node, codegraph-callers, codegraph-callees, codegraph-impact, and\ncodegraph-affected before reporting it. Drop or correct findings that only seem\nplausible from the diff but contradict unchanged callers, callees, or the same\npattern elsewhere in the repo. Use those tools to check blast radius and whether\na test reaches the path. Do not guess coverage or impact from the diff alone\nwhen a tool can answer. A missing regression test is not a substitute for\nidentifying the concrete input or code path that misbehaves when you can.";
|
|
86
|
+
/** Said next to the codegraph instructions when the review offers search. */
|
|
87
|
+
export declare const SEARCH_NON_CODE_GUIDANCE = "Use search-non-code for docs, configuration and other files codegraph does not index. Use codegraph for code.";
|
|
79
88
|
export declare function clampToolRounds(totalDiffLines: number): number;
|
|
80
89
|
export declare function clampImproveMatrix(matrix: number): number;
|
|
81
90
|
export declare function readGuide(repo: string, name: string): Promise<string>;
|
package/dist/src/pr/reviewer.js
CHANGED
|
@@ -4,7 +4,9 @@ import { completeWithMermaidTools, } from "../ai/mermaid_loop.js";
|
|
|
4
4
|
import { loadGuides } from "../review/guides.js";
|
|
5
5
|
import { notInitializedError } from "../review/not_initialized.js";
|
|
6
6
|
import { computeScope } from "./scope.js";
|
|
7
|
-
import {
|
|
7
|
+
import { codegraphState, prepareReviewTools, } from "./review_tools.js";
|
|
8
|
+
import { ToolGate, toolBudgetChars, toolResultMaxChars, } from "../ai/tool_gate.js";
|
|
9
|
+
import { SEARCH_NON_CODE } from "../tools/search_non_code.js";
|
|
8
10
|
import { severitySection } from "./review_copy.js";
|
|
9
11
|
import { repositoryInstructionsBlock, reviewSystemPromptBlock, } from "../review/system_prompt.js";
|
|
10
12
|
import { readDiffMaxChars } from "./diff_summary.js";
|
|
@@ -96,8 +98,8 @@ If a diagram is useful, call read-mermaid-syntaxes before writing any Mermaid.
|
|
|
96
98
|
Request every type you need in a single call, because repeated tool rounds are
|
|
97
99
|
capped. Use the smallest type that matches the evidence.`;
|
|
98
100
|
const REVIEW_ROLE = `You are a precise open-source code reviewer.
|
|
99
|
-
Evidence must come from the supplied diff
|
|
100
|
-
concise by default.`;
|
|
101
|
+
Evidence must come from the supplied diff, the review guide and the results
|
|
102
|
+
of the tools you call. Keep findings concise by default.`;
|
|
101
103
|
const REVIEW_DIAGRAM_RULES = `People generally find it easier to understand the
|
|
102
104
|
problem you've identified when it's presented in diagrams. When a multi-step
|
|
103
105
|
flow, lifecycle, dependency, data model, protocol, or architecture change is
|
|
@@ -156,6 +158,24 @@ pattern elsewhere in the repo. Use those tools to check blast radius and whether
|
|
|
156
158
|
a test reaches the path. Do not guess coverage or impact from the diff alone
|
|
157
159
|
when a tool can answer. A missing regression test is not a substitute for
|
|
158
160
|
identifying the concrete input or code path that misbehaves when you can.`;
|
|
161
|
+
/** Said next to the codegraph instructions when the review offers search. */
|
|
162
|
+
export const SEARCH_NON_CODE_GUIDANCE = `Use search-non-code for docs, configuration and other files codegraph does not index. Use codegraph for code.`;
|
|
163
|
+
/** The codegraph instructions, with the search sentence when search is one
|
|
164
|
+
* of the tools the model can call. A provider without tool support gets
|
|
165
|
+
* none, so it is passed no tools. */
|
|
166
|
+
function toolGuidance(tools) {
|
|
167
|
+
return tools.some((tool) => tool.name === SEARCH_NON_CODE)
|
|
168
|
+
? `${CODEGRAPH_DIFF_VERIFICATION}\n${SEARCH_NON_CODE_GUIDANCE}`
|
|
169
|
+
: CODEGRAPH_DIFF_VERIFICATION;
|
|
170
|
+
}
|
|
171
|
+
/** One review's tool gate, from the repository's settings. */
|
|
172
|
+
function reviewGate(repo) {
|
|
173
|
+
const repoConfig = readConfig().repos?.[repo];
|
|
174
|
+
return new ToolGate({
|
|
175
|
+
resultMaxChars: toolResultMaxChars(repoConfig),
|
|
176
|
+
budgetChars: toolBudgetChars(repoConfig),
|
|
177
|
+
});
|
|
178
|
+
}
|
|
159
179
|
/** An AI call a Stop can cut between tool rounds. The calls made before the
|
|
160
180
|
* Stop are paid, so their cost is recorded before the stop goes on up. */
|
|
161
181
|
async function stoppable(call, usage) {
|
|
@@ -174,7 +194,7 @@ async function stoppable(call, usage) {
|
|
|
174
194
|
* explicit reminder, if that also fails, the raw text is returned so the
|
|
175
195
|
* legacy Markdown parser can still read it. This keeps a provider that ignores
|
|
176
196
|
* `response_format` working, and never drops a review on the floor. */
|
|
177
|
-
async function normalizeReviewResponse(provider, request, response, extraTools, maxToolRounds, usage, report, options, signal) {
|
|
197
|
+
async function normalizeReviewResponse(provider, request, response, extraTools, maxToolRounds, usage, report, options, gate, signal) {
|
|
178
198
|
const direct = findingsMarkdownFromJson(response.text);
|
|
179
199
|
if (direct !== undefined)
|
|
180
200
|
return direct;
|
|
@@ -183,7 +203,7 @@ async function normalizeReviewResponse(provider, request, response, extraTools,
|
|
|
183
203
|
...request,
|
|
184
204
|
job: "review_pull_request_retry",
|
|
185
205
|
prompt: `${request.prompt}\n\nThe previous answer was not valid JSON. Return only the JSON object described above, with no prose and no code fence.`,
|
|
186
|
-
}, 1, extraTools, maxToolRounds, { signal }), usage);
|
|
206
|
+
}, 1, extraTools, maxToolRounds, { signal, gate }), usage);
|
|
187
207
|
if (usage)
|
|
188
208
|
await usage(retry);
|
|
189
209
|
if (options.debug) {
|
|
@@ -301,28 +321,40 @@ export async function reviewPullRequest(client, options, usage, snapshot, ai, pr
|
|
|
301
321
|
}
|
|
302
322
|
const provider = ai ?? reviewProvider(options, options.highModel, "high");
|
|
303
323
|
const diagrams = provider.supportsTools !== false;
|
|
304
|
-
const
|
|
305
|
-
|
|
306
|
-
|
|
307
|
-
|
|
308
|
-
|
|
309
|
-
|
|
310
|
-
|
|
311
|
-
|
|
312
|
-
|
|
313
|
-
|
|
314
|
-
|
|
315
|
-
|
|
316
|
-
|
|
317
|
-
|
|
324
|
+
const preparing = prepareReviewDiff({
|
|
325
|
+
files: ownFiles.map(({ file }) => diffFileFromGithub(file)),
|
|
326
|
+
unchangedPaths: extras?.unchangedPaths,
|
|
327
|
+
catchesUp: extras?.catchesUp ?? false,
|
|
328
|
+
tools: diagrams,
|
|
329
|
+
concurrency: options.aiConcurrent,
|
|
330
|
+
readDiffMaxChars: readDiffMaxChars(readConfig().repos?.[options.repo]),
|
|
331
|
+
summaryProvider: () => summaryProvider(options, provider, report, extras),
|
|
332
|
+
usage,
|
|
333
|
+
report,
|
|
334
|
+
});
|
|
335
|
+
// The pull request's checkout serves codegraph and search alike, and is
|
|
336
|
+
// made while the diff is prepared.
|
|
337
|
+
const source = extras?.tools ??
|
|
338
|
+
(headSha
|
|
339
|
+
? {
|
|
340
|
+
kind: "checkout",
|
|
341
|
+
pr: number,
|
|
342
|
+
commit: headSha,
|
|
343
|
+
search: true,
|
|
318
344
|
allowInstall: options.allowToolInstall,
|
|
319
|
-
}
|
|
320
|
-
:
|
|
345
|
+
}
|
|
346
|
+
: undefined);
|
|
347
|
+
const [prepared, reviewTools] = await Promise.all([
|
|
348
|
+
preparing,
|
|
349
|
+
source
|
|
350
|
+
? prepareReviewTools(source, options, preparing.then((plan) => plan.tools))
|
|
351
|
+
: preparing.then((plan) => ({ tools: plan.tools, codegraph: [] })),
|
|
321
352
|
]);
|
|
322
353
|
if (options.useCodegraph) {
|
|
323
|
-
report(`codegraph tools · ${
|
|
354
|
+
report(`codegraph tools · ${reviewTools.codegraph.length > 0 ? "ready" : "unavailable"}`);
|
|
324
355
|
}
|
|
325
|
-
const extraTools =
|
|
356
|
+
const extraTools = reviewTools.tools;
|
|
357
|
+
const gate = reviewGate(options.repo);
|
|
326
358
|
// Bigger own-scope diffs need more codegraph round-trips to trace, fixed
|
|
327
359
|
// at 3 rounds regardless of size caused the model to run out mid-review on
|
|
328
360
|
// large PRs (verified via a tokio benchmark run: it hallucinated fake
|
|
@@ -368,7 +400,7 @@ When the DIFF section below has an UPSTREAM CONTEXT part, that code arrived
|
|
|
368
400
|
through a merge and was not authored by this pull request. Do not raise a
|
|
369
401
|
finding located only there, and never mark blocking a finding whose only
|
|
370
402
|
support is upstream context.
|
|
371
|
-
${
|
|
403
|
+
${toolGuidance(diagrams ? extraTools : [])}
|
|
372
404
|
Keep the default finding's "body" under 120 words excluding an optional
|
|
373
405
|
suggestion. Write prose only: do not emit Markdown headings, a Location line,
|
|
374
406
|
backticks around the path, or the sentence "If you'd like me to explain it in
|
|
@@ -413,6 +445,7 @@ ${diff}`;
|
|
|
413
445
|
}
|
|
414
446
|
let response = await stoppable(completeWithMermaidTools(provider, request, 1, extraTools, maxToolRounds, {
|
|
415
447
|
signal,
|
|
448
|
+
gate,
|
|
416
449
|
}), usage);
|
|
417
450
|
if (usage)
|
|
418
451
|
await usage(response);
|
|
@@ -427,7 +460,7 @@ ${diff}`;
|
|
|
427
460
|
// The model returns JSON. We render the Markdown. A reply
|
|
428
461
|
// that is not usable JSON is retried once and then falls back to the legacy
|
|
429
462
|
// Markdown parser, so a provider that ignores the schema still works.
|
|
430
|
-
let reviewText = await normalizeReviewResponse(provider, request, response, extraTools, maxToolRounds, usage, report, options, signal);
|
|
463
|
+
let reviewText = await normalizeReviewResponse(provider, request, response, extraTools, maxToolRounds, usage, report, options, gate, signal);
|
|
431
464
|
for (let pass = 2; pass <= matrix; pass++) {
|
|
432
465
|
const improvementRequest = {
|
|
433
466
|
...request,
|
|
@@ -448,7 +481,7 @@ ${reviewText}`,
|
|
|
448
481
|
console.log(`[debug] improvement ${pass - 1} request · prompt=${improvementRequest.prompt.length} chars · maxTokens=${improvementRequest.maxTokens}`);
|
|
449
482
|
}
|
|
450
483
|
report(`AI improvement pass ${pass - 1} of ${matrix - 1}`);
|
|
451
|
-
response = await stoppable(completeWithMermaidTools(provider, improvementRequest, 1, extraTools, maxToolRounds, { signal }), usage);
|
|
484
|
+
response = await stoppable(completeWithMermaidTools(provider, improvementRequest, 1, extraTools, maxToolRounds, { signal, gate }), usage);
|
|
452
485
|
if (usage)
|
|
453
486
|
await usage(response);
|
|
454
487
|
report(`AI improvement response · input=${response.tokensIn} tokens · ` +
|
|
@@ -456,7 +489,7 @@ ${reviewText}`,
|
|
|
456
489
|
if (!response.text.trim()) {
|
|
457
490
|
throw new Error(`${aiProviderName(options)} returned an empty review improvement at pass ${pass - 1}`);
|
|
458
491
|
}
|
|
459
|
-
reviewText = await normalizeReviewResponse(provider, improvementRequest, response, extraTools, maxToolRounds, usage, report, options, signal);
|
|
492
|
+
reviewText = await normalizeReviewResponse(provider, improvementRequest, response, extraTools, maxToolRounds, usage, report, options, gate, signal);
|
|
460
493
|
if (options.debug) {
|
|
461
494
|
console.log(`[debug] improvement ${pass - 1} response · input=${response.tokensIn} tokens · output=${response.tokensOut} tokens`);
|
|
462
495
|
console.log(`\n----- IMPROVED REVIEW ${pass - 1} -----\n${reviewText}\n`);
|
|
@@ -468,13 +501,10 @@ ${reviewText}`,
|
|
|
468
501
|
visiblePaths: prepared.seenPaths,
|
|
469
502
|
guideBuiltAt: guides.guideBuiltAt,
|
|
470
503
|
referenceContext: { rules: guides.rules, codebase: guides.codebase },
|
|
471
|
-
codegraphState:
|
|
472
|
-
? "disabled"
|
|
473
|
-
: codegraphTools.length > 0
|
|
474
|
-
? "used"
|
|
475
|
-
: "unavailable",
|
|
504
|
+
codegraphState: codegraphState(options.useCodegraph, reviewTools.codegraph),
|
|
476
505
|
coverage: prepared.coverage,
|
|
477
|
-
warnings: prepared.warnings,
|
|
506
|
+
warnings: [...prepared.warnings, ...gate.warnings()],
|
|
507
|
+
tools: gate.usage(),
|
|
478
508
|
diffStats: diffStats(named.map(({ file }) => diffFileFromGithub(file))),
|
|
479
509
|
};
|
|
480
510
|
}
|
|
@@ -505,12 +535,14 @@ export async function reviewWorkspaceRevision(revision, options, headSha, usage,
|
|
|
505
535
|
usage,
|
|
506
536
|
report,
|
|
507
537
|
});
|
|
508
|
-
const
|
|
509
|
-
|
|
510
|
-
|
|
511
|
-
|
|
512
|
-
:
|
|
513
|
-
|
|
538
|
+
const reviewTools = await prepareReviewTools(extras?.tools ?? {
|
|
539
|
+
kind: "checkout",
|
|
540
|
+
pr: 0,
|
|
541
|
+
commit: headSha,
|
|
542
|
+
search: false,
|
|
543
|
+
}, options, prepared.tools);
|
|
544
|
+
const extraTools = reviewTools.tools;
|
|
545
|
+
const gate = reviewGate(options.repo);
|
|
514
546
|
const totalDiffLines = ownFiles.reduce((sum, file) => sum + file.changes, 0);
|
|
515
547
|
const maxToolRounds = clampToolRounds(totalDiffLines);
|
|
516
548
|
const ownDiff = prepared.diff;
|
|
@@ -527,7 +559,7 @@ the diff and either the guide or the codebase conventions.
|
|
|
527
559
|
Do not invent low-value findings. If the diff contains a truncation marker or a
|
|
528
560
|
file is summarized, limit claims to the supplied text and use read-full-diff or
|
|
529
561
|
codegraph tools before asserting behavior outside what was shown.
|
|
530
|
-
${
|
|
562
|
+
${toolGuidance(diagrams ? extraTools : [])}
|
|
531
563
|
${diagrams ? (plantumlPromptRules(options) ?? DIAGRAM_PROMPT_RULES) : NO_DIAGRAM_RULES}
|
|
532
564
|
${FINDINGS_JSON_INSTRUCTIONS}
|
|
533
565
|
|
|
@@ -553,13 +585,14 @@ ${ownDiff}${unchangedListing}`;
|
|
|
553
585
|
report(`AI request · prompt=${prompt.length} chars`);
|
|
554
586
|
let response = await stoppable(completeWithMermaidTools(provider, request, 1, extraTools, maxToolRounds, {
|
|
555
587
|
signal,
|
|
588
|
+
gate,
|
|
556
589
|
}), usage);
|
|
557
590
|
if (usage)
|
|
558
591
|
await usage(response);
|
|
559
592
|
if (!response.text.trim()) {
|
|
560
593
|
throw new Error(`${aiProviderName(options)} returned an empty review`);
|
|
561
594
|
}
|
|
562
|
-
let reviewText = await normalizeReviewResponse(provider, request, response, extraTools, maxToolRounds, usage, report, options, signal);
|
|
595
|
+
let reviewText = await normalizeReviewResponse(provider, request, response, extraTools, maxToolRounds, usage, report, options, gate, signal);
|
|
563
596
|
for (let pass = 2; pass <= matrix; pass++) {
|
|
564
597
|
const improvementRequest = {
|
|
565
598
|
...request,
|
|
@@ -577,25 +610,21 @@ ${prompt}
|
|
|
577
610
|
DRAFT REVIEW (Markdown rendering of the previous JSON answer):
|
|
578
611
|
${reviewText}`,
|
|
579
612
|
};
|
|
580
|
-
response = await stoppable(completeWithMermaidTools(provider, improvementRequest, 1, extraTools, maxToolRounds, { signal }), usage);
|
|
613
|
+
response = await stoppable(completeWithMermaidTools(provider, improvementRequest, 1, extraTools, maxToolRounds, { signal, gate }), usage);
|
|
581
614
|
if (usage)
|
|
582
615
|
await usage(response);
|
|
583
|
-
reviewText = await normalizeReviewResponse(provider, improvementRequest, response, extraTools, maxToolRounds, usage, report, options, signal);
|
|
616
|
+
reviewText = await normalizeReviewResponse(provider, improvementRequest, response, extraTools, maxToolRounds, usage, report, options, gate, signal);
|
|
584
617
|
}
|
|
585
|
-
const codegraphState = !options.useCodegraph
|
|
586
|
-
? "disabled"
|
|
587
|
-
: codegraphTools.length > 0
|
|
588
|
-
? "used"
|
|
589
|
-
: "unavailable";
|
|
590
618
|
return {
|
|
591
619
|
...response,
|
|
592
620
|
text: `${severitySection()}\n${renderDiagrams(reviewText, options)}`,
|
|
593
621
|
visiblePaths: prepared.seenPaths,
|
|
594
622
|
guideBuiltAt: guides.guideBuiltAt,
|
|
595
623
|
referenceContext: { rules: guides.rules, codebase: guides.codebase },
|
|
596
|
-
codegraphState,
|
|
624
|
+
codegraphState: codegraphState(options.useCodegraph, reviewTools.codegraph),
|
|
597
625
|
coverage: prepared.coverage,
|
|
598
|
-
warnings: prepared.warnings,
|
|
626
|
+
warnings: [...prepared.warnings, ...gate.warnings()],
|
|
627
|
+
tools: gate.usage(),
|
|
599
628
|
diffStats: diffStats(files),
|
|
600
629
|
};
|
|
601
630
|
}
|
package/dist/src/pr/scope.js
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { ensureCloneUnlocked, ensureCommit, runCommand, withCloneLock, } from "./checkout.js";
|
|
1
|
+
import { ensureCloneUnlocked, withGitAuth, ensureCommit, runCommand, withCloneLock, } from "./checkout.js";
|
|
2
2
|
import { detectDefaultBranchRef } from "../git/default_branch.js";
|
|
3
3
|
import { log } from "../util/log.js";
|
|
4
4
|
function lines(output) {
|
|
@@ -79,6 +79,6 @@ export async function computeScope(repo, base, head, options = {}) {
|
|
|
79
79
|
log("scope", `unavailable · ${String(error)}`);
|
|
80
80
|
return undefined;
|
|
81
81
|
}
|
|
82
|
-
return await scopeInClone(clone, base, head, run);
|
|
82
|
+
return await scopeInClone(clone, base, head, withGitAuth(run, clone));
|
|
83
83
|
});
|
|
84
84
|
}
|
|
@@ -12,6 +12,8 @@ import { prepareLocalCodegraph } from "../local/codegraph_prepare.js";
|
|
|
12
12
|
import { canPrompt } from "../tools/codegraph.js";
|
|
13
13
|
import { runRemoteToolCalls, toolHandlerMap, } from "./client_tools.js";
|
|
14
14
|
import { REMOTE_KNOWN_TOOL_NAMES } from "./schema.js";
|
|
15
|
+
import { searchNonCodeTool, searchOptionsFor, } from "../tools/search_non_code.js";
|
|
16
|
+
import { toolUsageFromJson } from "../ai/tool_gate.js";
|
|
15
17
|
import { runCommand } from "../pr/checkout.js";
|
|
16
18
|
import { assertGitQuiet, currentBranch, detectRemoteRepo, gitRoot, } from "../local/git_ops.js";
|
|
17
19
|
import { buildLocalRevision, mergeBase, resolveBaseRef, } from "../local/git_revision.js";
|
|
@@ -165,7 +167,12 @@ export async function runRemoteReview(cli) {
|
|
|
165
167
|
// `--json` must stay machine-readable, `canPrompt` adds TTY and CI.
|
|
166
168
|
interactive: !cli.json && canPrompt(),
|
|
167
169
|
});
|
|
168
|
-
|
|
170
|
+
// The server asks for search the way it asks for codegraph, and the
|
|
171
|
+
// client runs it in this checkout. An older server ignores the name.
|
|
172
|
+
const localTools = toolHandlerMap([
|
|
173
|
+
...codegraphPrep.tools,
|
|
174
|
+
searchNonCodeTool(root, searchOptionsFor(repo)),
|
|
175
|
+
]);
|
|
169
176
|
const capabilityTools = [...localTools.keys()]
|
|
170
177
|
.filter((name) => REMOTE_KNOWN_TOOL_NAMES.has(name))
|
|
171
178
|
.map((name) => ({ name }));
|
|
@@ -324,6 +331,8 @@ export async function runRemoteReview(cli) {
|
|
|
324
331
|
extensions,
|
|
325
332
|
intent: intent !== null,
|
|
326
333
|
showReferences: readConfig().repos?.[repo]?.reviewReferences === true,
|
|
334
|
+
tools: toolUsageFromJson(payload.result.usage
|
|
335
|
+
?.tools),
|
|
327
336
|
}) +
|
|
328
337
|
"\n");
|
|
329
338
|
if (run)
|
|
@@ -13,6 +13,7 @@
|
|
|
13
13
|
* - a server that cannot be reached or rejects the token is red. */
|
|
14
14
|
import { VERSION } from "../version.js";
|
|
15
15
|
import { exitWith } from "../cli/error.js";
|
|
16
|
+
import { toolUsageFromJson } from "../ai/tool_gate.js";
|
|
16
17
|
import { formatHumanReview, humanFindingsFromJson, } from "../cli/review_result.js";
|
|
17
18
|
import { coverageFromResult, warningsFromJson } from "../review/coverage.js";
|
|
18
19
|
import { baseUrl, die, handshakeFailure, readApiError, remoteFetch, remoteFromEnv, } from "./http.js";
|
|
@@ -273,6 +274,8 @@ export async function runRemotePushReview(cli) {
|
|
|
273
274
|
findings: human,
|
|
274
275
|
warnings: warningsFromJson(payload.result.warnings),
|
|
275
276
|
coverage: coverageFromResult(payload.result),
|
|
277
|
+
tools: toolUsageFromJson(payload.result.usage
|
|
278
|
+
?.tools),
|
|
276
279
|
}) +
|
|
277
280
|
"\n");
|
|
278
281
|
}
|
|
@@ -16,6 +16,7 @@ import { withResolvedReferences } from "../review/references.js";
|
|
|
16
16
|
import { parseFindings } from "../pr/findings.js";
|
|
17
17
|
import { githubFileToRevisionFile } from "../pr/github_revision.js";
|
|
18
18
|
import { reviewWorkspaceRevision } from "../pr/reviewer.js";
|
|
19
|
+
import { emptyToolUsage } from "../ai/tool_gate.js";
|
|
19
20
|
import { anchorTextFromPatch, parsePreviousVerdicts, } from "../review/carry_over.js";
|
|
20
21
|
import { normalizeAnchor } from "../review/revision.js";
|
|
21
22
|
import { revisionFromSubmitJson } from "../remote/revision_from_submit.js";
|
|
@@ -749,6 +750,7 @@ async function runPushReview(job, log, signal) {
|
|
|
749
750
|
tokensOut,
|
|
750
751
|
...costUsage(costTally),
|
|
751
752
|
billedTo: "server",
|
|
753
|
+
tools: response?.tools ?? emptyToolUsage(),
|
|
752
754
|
},
|
|
753
755
|
remote: { jobId: job.id, reviewId: review.id },
|
|
754
756
|
};
|
|
@@ -4,6 +4,8 @@ import { CliError } from "../cli/error.js";
|
|
|
4
4
|
import { resolvedFromFirstReview, reviewCoverageFields, sortResolvedFindings, summaryCounts, toJsonFinding, } from "../cli/review_result.js";
|
|
5
5
|
import { parseFindings } from "../pr/findings.js";
|
|
6
6
|
import { codegraphToolHandlersForBridge } from "../pr/codegraph_tools.js";
|
|
7
|
+
import { isCodegraphTool } from "../pr/review_tools.js";
|
|
8
|
+
import { SEARCH_NON_CODE, SEARCH_NON_CODE_TOOL, } from "../tools/search_non_code.js";
|
|
7
9
|
import { reviewWorkspaceRevision } from "../pr/reviewer.js";
|
|
8
10
|
import { buildCarryPromptSection, classifyCarryItems, guideRebuiltSince, incrementalDiffPaths, parsePreviousVerdicts, resolveCarryOutcomes, } from "../review/carry_over.js";
|
|
9
11
|
import { visiblePathsToSave } from "../review/visible_paths.js";
|
|
@@ -301,7 +303,9 @@ export function registerRemoteReviewHandler() {
|
|
|
301
303
|
const allowed = new Set((caps.tools ?? [])
|
|
302
304
|
.map((tool) => tool.name)
|
|
303
305
|
.filter((name) => typeof name === "string" && name));
|
|
304
|
-
|
|
306
|
+
// Codegraph is on when the client offers a codegraph tool. Search
|
|
307
|
+
// alone does not turn it on.
|
|
308
|
+
const useCodegraph = [...allowed].some(isCodegraphTool);
|
|
305
309
|
const options = reviewOptionsForRepo(job.repo, useCodegraph, byokKey, (text) => log("info", text));
|
|
306
310
|
// The client's one-run instructions travel with the job, so a retry
|
|
307
311
|
// keeps them. They apply on top of the server's per-repo prompt.
|
|
@@ -356,10 +360,18 @@ export function registerRemoteReviewHandler() {
|
|
|
356
360
|
extras.carryPrompt = buildCarryPromptSection(carryItems, revision);
|
|
357
361
|
}
|
|
358
362
|
}
|
|
359
|
-
|
|
360
|
-
|
|
361
|
-
|
|
362
|
-
|
|
363
|
+
// The client runs codegraph and search in its own checkout.
|
|
364
|
+
extras.tools = {
|
|
365
|
+
kind: "bridge",
|
|
366
|
+
tools: remoteBridgeToolHandlers(job.id, signal, allowed, [
|
|
367
|
+
...codegraphToolHandlersForBridge(),
|
|
368
|
+
{
|
|
369
|
+
name: SEARCH_NON_CODE,
|
|
370
|
+
tool: SEARCH_NON_CODE_TOOL,
|
|
371
|
+
run: () => "",
|
|
372
|
+
},
|
|
373
|
+
]),
|
|
374
|
+
};
|
|
363
375
|
const started = performance.now();
|
|
364
376
|
let tokensIn = 0;
|
|
365
377
|
let tokensOut = 0;
|
|
@@ -404,6 +416,7 @@ export function registerRemoteReviewHandler() {
|
|
|
404
416
|
tokensOut,
|
|
405
417
|
...costUsage(costTally),
|
|
406
418
|
billedTo: billed,
|
|
419
|
+
tools: result.tools,
|
|
407
420
|
},
|
|
408
421
|
}, extensions, true);
|
|
409
422
|
setRemoteSyncResult(job.id, {
|
|
@@ -5,6 +5,13 @@ import { type UserConfig } from "../config.ts";
|
|
|
5
5
|
* someone else's machine. Raise it, adjust whatever reads the output, done. */
|
|
6
6
|
export declare const CODEGRAPH_VERSION = "1.6.0";
|
|
7
7
|
export declare const CODEGRAPH_PACKAGE = "@colbymchenry/codegraph";
|
|
8
|
+
/** The file extensions codegraph {@link CODEGRAPH_VERSION} extracts symbols
|
|
9
|
+
* from, lower case. `search-non-code` searches every other tracked text file,
|
|
10
|
+
* so update this list whenever the version is raised. Taken from
|
|
11
|
+
* `EXTENSION_MAP` in the installed package's `extraction/grammars.js`, without
|
|
12
|
+
* the languages it tracks only as files and extracts nothing from (`.yml`,
|
|
13
|
+
* `.yaml`, `.twig`, `.properties`, and `.xml` outside MyBatis mappers). */
|
|
14
|
+
export declare const CODEGRAPH_EXTENSIONS: readonly string[];
|
|
8
15
|
/** A version that is safe in a directory name and an npm argument: a bare
|
|
9
16
|
* semver, optionally with a prerelease suffix. Anything else (a path, a
|
|
10
17
|
* leading `-`, whitespace) is refused, so a bad config value can never reach
|