@bjoernrochel/sevro 0.1.0-rc.2 → 0.1.0-rc.21
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/CONTRIBUTING.md +170 -0
- package/LICENSE +63 -136
- package/README.md +83 -66
- package/docs/README.md +44 -0
- package/docs/acknowledgements.md +14 -0
- package/docs/architecture.md +146 -0
- package/docs/assets/README.md +27 -0
- package/docs/assets/sevro-logo.png +0 -0
- package/docs/assets/sevro-symbol.svg +1 -0
- package/docs/assets/sevro-wordmark.svg +1 -0
- package/docs/builtin-graders-v1.md +41 -13
- package/docs/codex-host.md +184 -52
- package/docs/contracts.md +19 -0
- package/docs/creating-cases.md +139 -0
- package/docs/development-cli.md +190 -30
- package/docs/documentation-quality.md +61 -0
- package/docs/evaluation-workflow.md +110 -0
- package/docs/evidence/code-mode-controls-71.md +64 -0
- package/docs/evidence/code-mode-programs-85.md +64 -0
- package/docs/evidence/codex-state-73/README.md +204 -0
- package/docs/evidence/codex-state-73/darrow-migration.patch +182 -0
- package/docs/evidence/codex-state-73/results.json +661 -0
- package/docs/evidence/local-http-74.md +104 -0
- package/docs/evidence/runtime-startup-30/README.md +185 -0
- package/docs/evidence/runtime-startup-30/after.json +319 -0
- package/docs/evidence/runtime-startup-30/baseline-timeout.jsonl +2 -0
- package/docs/evidence/runtime-startup-30/before.json +403 -0
- package/docs/evidence/runtime-startup-30/consumer-smoke.json +63 -0
- package/docs/evidence/runtime-startup-30/full-policy-diagnostic.jsonl +3 -0
- package/docs/evidence/runtime-startup-30/minimal-policy-diagnostic.jsonl +3 -0
- package/docs/evidence/runtime-startup-30/plugin-template-assessment.json +20 -0
- package/docs/evidence/runtime-startup-30/seedless-timeout.jsonl +2 -0
- package/docs/evidence/skill-reads-70/results.json +823 -0
- package/docs/evidence.md +21 -0
- package/docs/extension-protocol-v1.md +50 -16
- package/docs/fixture-cleanup-validation.md +3 -0
- package/docs/getting-started.md +102 -0
- package/docs/guide-evaluation.md +89 -0
- package/docs/guide-verification.json +239 -0
- package/docs/identity-v1.md +35 -1
- package/docs/installing.md +88 -0
- package/docs/licenses/BUSL-1.1-0.1.0-rc.2.txt +159 -0
- package/docs/licensing.md +42 -0
- package/docs/native-controls.md +76 -5
- package/docs/native-hosts.md +228 -0
- package/docs/reading-results.md +141 -0
- package/docs/recipes.md +489 -0
- package/docs/releases.md +81 -40
- package/docs/report-v1.md +29 -1
- package/docs/results-v1.md +180 -0
- package/docs/run-settings-v1.md +147 -0
- package/docs/running-evaluations.md +31 -0
- package/docs/runtime-v1.md +614 -0
- package/docs/specs/repository-guide.md +58 -0
- package/docs/troubleshooting.md +30 -0
- package/docs/typescript-quality.md +234 -0
- package/docs/usability-study.md +114 -0
- package/examples/basic/host.ts +2 -2
- package/examples/recipes/candidate.ts +26 -0
- package/examples/recipes/cases/coding-change.json +21 -0
- package/examples/recipes/cases/host-tools.json +21 -0
- package/examples/recipes/cases/mise.json +21 -0
- package/examples/recipes/cases/node.json +21 -0
- package/examples/recipes/cases/python-uv.json +21 -0
- package/examples/recipes/cases/semantic.json +17 -0
- package/examples/recipes/coding-change/check.ts +23 -0
- package/examples/recipes/extension-artifacts.ts +65 -0
- package/examples/recipes/extension-methods.ts +174 -0
- package/examples/recipes/extension.ts +37 -0
- package/examples/recipes/repository-skill/.agents/skills/startup-guide/SKILL.md +7 -0
- package/examples/recipes/repository-skill/.claude/skills/startup-guide/SKILL.md +7 -0
- package/examples/recipes/repository-skill/README.md +3 -0
- package/examples/recipes/run-inputs.ts +188 -0
- package/examples/recipes/run-options.ts +131 -0
- package/examples/recipes/run.ts +63 -0
- package/examples/recipes/runtime/empty.json +1 -0
- package/examples/recipes/runtime/goals.json +1 -0
- package/examples/recipes/runtime/mise.json +6 -0
- package/examples/recipes/runtime/path.json +1 -0
- package/examples/recipes/runtime/plugins.json +4 -0
- package/examples/recipes/runtime/python-uv.json +10 -0
- package/examples/recipes/semantic-judge.ts +29 -0
- package/examples/templates/coding/adapter.ts +17 -0
- package/examples/templates/coding/case.json +15 -0
- package/examples/templates/repository-skill/.agents/skills/startup-guide/SKILL.md +6 -0
- package/examples/templates/repository-skill/extension-data.ts +64 -0
- package/examples/templates/repository-skill/extension.ts +31 -0
- package/examples/templates/response/adapter.ts +12 -0
- package/examples/templates/response/case.json +13 -0
- package/package.json +39 -6
- package/schemas/report-v1.schema.json +9 -0
- package/schemas/run-evidence-v1.schema.json +2 -1
- package/schemas/runtime-v1.schema.json +193 -0
- package/src/advisory.ts +10 -5
- package/src/candidate-artifacts.ts +208 -0
- package/src/candidate-codex-state.ts +185 -0
- package/src/case-validation.ts +105 -0
- package/src/cli/cancellation.ts +20 -0
- package/src/cli/command.ts +81 -0
- package/src/cli/completion.ts +15 -0
- package/src/cli/demo.ts +59 -0
- package/src/cli/evaluation.ts +55 -0
- package/src/cli/extension.ts +79 -0
- package/src/cli/help.ts +222 -0
- package/src/cli/hosts.ts +59 -0
- package/src/cli/init-settings.ts +30 -0
- package/src/cli/init.ts +111 -0
- package/src/cli/invocation.ts +627 -0
- package/src/cli/loading.ts +151 -0
- package/src/cli/preparation.ts +163 -0
- package/src/cli/presentation.ts +91 -0
- package/src/cli/progress.ts +6 -0
- package/src/cli/run-manifest.ts +133 -0
- package/src/cli/run-options.ts +101 -0
- package/src/cli/run-overrides.ts +83 -0
- package/src/cli/settings.ts +160 -0
- package/src/cli/summary.ts +43 -0
- package/src/cli.ts +13 -843
- package/src/contract-types.ts +94 -0
- package/src/engine.ts +19 -2373
- package/src/evaluation/advisory.ts +150 -0
- package/src/evaluation/check-plan.ts +510 -0
- package/src/evaluation/collection.ts +88 -0
- package/src/evaluation/context.ts +10 -0
- package/src/evaluation/grading-host.ts +39 -0
- package/src/evaluation/grading.ts +258 -0
- package/src/evaluation/invocation-plan.ts +548 -0
- package/src/evaluation/plan.ts +736 -0
- package/src/evaluation/progress.ts +19 -0
- package/src/evaluation/resources.ts +241 -0
- package/src/evaluation/run.ts +206 -0
- package/src/{trial-scheduler.ts → evaluation/scheduler.ts} +3 -1
- package/src/evaluation/semantic.ts +312 -0
- package/src/evaluation/trial.ts +866 -0
- package/src/evaluation/types.ts +207 -0
- package/src/evidence/host-results.ts +92 -0
- package/src/evidence/retained-artifacts.ts +18 -0
- package/src/evidence/trial-writer.ts +203 -0
- package/src/evidence/usage.ts +17 -0
- package/src/extension-client.ts +228 -148
- package/src/extension-session.ts +326 -195
- package/src/fixtures/advisory.ts +338 -0
- package/src/fixtures/bin.ts +57 -0
- package/src/{fixture-cleanup.ts → fixtures/cleanup.ts} +13 -7
- package/src/fixtures/create.ts +187 -0
- package/src/fixtures/generated.ts +300 -0
- package/src/fixtures/git-hooks.ts +24 -0
- package/src/fixtures/preparation.ts +270 -0
- package/src/{repository-fixture.ts → fixtures/repository.ts} +207 -85
- package/src/fixtures/setup.ts +224 -0
- package/src/fixtures/tools.ts +88 -0
- package/src/fixtures/workspace-state.ts +212 -0
- package/src/generated/cli-result.cjs +2 -0
- package/src/generated/cli-result.d.cts +3 -0
- package/src/generated/extension.cjs +2 -0
- package/src/generated/extension.d.cts +3 -0
- package/src/generated/report.cjs +2 -0
- package/src/generated/report.d.cts +3 -0
- package/src/generated/run-evidence.cjs +2 -0
- package/src/generated/run-evidence.d.cts +3 -0
- package/src/generated/runtime.cjs +2 -0
- package/src/generated/runtime.d.cts +3 -0
- package/src/graders/artifact.ts +56 -23
- package/src/graders/git-head.ts +79 -75
- package/src/graders/output.ts +222 -140
- package/src/graders/semantic.ts +102 -43
- package/src/graders/shell.ts +379 -207
- package/src/host-diagnostic-redaction.ts +52 -0
- package/src/host-diagnostics.ts +319 -0
- package/src/hosts/claude-continuation.ts +63 -18
- package/src/hosts/claude-credential.ts +43 -62
- package/src/hosts/claude-events.ts +86 -51
- package/src/hosts/claude-keychain.ts +59 -0
- package/src/hosts/claude-native-goal.ts +57 -33
- package/src/hosts/claude-process.ts +111 -0
- package/src/hosts/claude-repository-invocation.ts +208 -117
- package/src/hosts/claude-settings.ts +240 -21
- package/src/hosts/claude-skill-activations.ts +166 -0
- package/src/hosts/claude-state.ts +403 -0
- package/src/hosts/claude-tool-calls.ts +171 -106
- package/src/hosts/claude.ts +561 -451
- package/src/hosts/codex-app-server.ts +326 -178
- package/src/hosts/codex-configuration.ts +44 -18
- package/src/hosts/codex-events.ts +97 -60
- package/src/hosts/codex-exec-controls.ts +516 -0
- package/src/hosts/codex-invocation.ts +70 -0
- package/src/hosts/codex-marketplace.ts +283 -0
- package/src/hosts/codex-native-calls.ts +546 -500
- package/src/hosts/codex-process.ts +202 -0
- package/src/hosts/codex-profile.ts +257 -28
- package/src/hosts/codex-session-files.ts +106 -0
- package/src/hosts/codex-state.ts +492 -0
- package/src/hosts/codex.ts +783 -832
- package/src/hosts/isolation-common.ts +94 -0
- package/src/hosts/isolation-roots.ts +54 -27
- package/src/hosts/isolation.ts +10 -0
- package/src/hosts/linux-sandbox.ts +159 -0
- package/src/hosts/mac-sandbox.ts +187 -61
- package/src/hosts/native-controls.ts +152 -62
- package/src/hosts/process-lifecycle.ts +37 -0
- package/src/hosts/protected-worktrees.ts +51 -19
- package/src/hosts/skill-hook-catalog.ts +131 -0
- package/src/hosts/skill-hook-handler.ts +374 -0
- package/src/hosts/skill-hooks.ts +580 -0
- package/src/hosts/workspace-fingerprint.ts +55 -35
- package/src/identity.ts +102 -48
- package/src/instrumentation.ts +109 -48
- package/src/native-transcript-state.ts +131 -0
- package/src/native-transcript-view.ts +90 -0
- package/src/native-transcripts.ts +286 -0
- package/src/process-lifecycle.ts +150 -0
- package/src/provenance.ts +161 -81
- package/src/prune.ts +215 -0
- package/src/report.ts +172 -140
- package/src/resolved-case.ts +33 -0
- package/src/results-lock.ts +29 -0
- package/src/results.ts +133 -86
- package/src/run-owner.ts +154 -63
- package/src/runtime/config.ts +456 -0
- package/src/runtime/discovery-apple.ts +90 -0
- package/src/runtime/discovery-providers.ts +63 -0
- package/src/runtime/discovery.ts +70 -0
- package/src/runtime/hook-processes.ts +97 -0
- package/src/runtime/hooks.ts +476 -0
- package/src/runtime/paths.ts +313 -0
- package/src/runtime/query.ts +85 -0
- package/src/runtime/seeds.ts +134 -0
- package/src/runtime/shared-cache.ts +143 -0
- package/src/runtime/state.ts +143 -0
- package/src/schema-types.ts +198 -0
- package/src/schema.ts +15 -19
- package/src/transcript-retention.ts +146 -0
- package/src/value-guards.ts +20 -0
- package/src/advisory-fixture.ts +0 -282
- package/src/fixture-bin.ts +0 -87
- package/src/fixture-setup.ts +0 -163
- package/src/generated-fixture.ts +0 -220
- package/src/git-hooks.ts +0 -63
- package/src/hosts/claude-nested-skills.ts +0 -218
- package/src/hosts/codex-command-output.ts +0 -399
- package/src/hosts/codex-executor-command.ts +0 -195
- package/src/hosts/codex-skill-reads.ts +0 -391
- package/src/preparation.ts +0 -211
|
@@ -0,0 +1,374 @@
|
|
|
1
|
+
import { randomUUID } from "node:crypto";
|
|
2
|
+
import { open, readFile, stat, writeFile } from "node:fs/promises";
|
|
3
|
+
import { join } from "node:path";
|
|
4
|
+
import { isRecord } from "../value-guards";
|
|
5
|
+
|
|
6
|
+
const MAX_INPUT_BYTES = 8 * 1024 * 1024;
|
|
7
|
+
const MAX_RECEIPT_BYTES = 4 * 1024 * 1024;
|
|
8
|
+
|
|
9
|
+
export interface HookSkill {
|
|
10
|
+
invocation: string;
|
|
11
|
+
sha256: string;
|
|
12
|
+
body: string;
|
|
13
|
+
}
|
|
14
|
+
export interface SkillHookConfiguration {
|
|
15
|
+
host: "codex";
|
|
16
|
+
receipts: string;
|
|
17
|
+
skills: HookSkill[];
|
|
18
|
+
catalogComplete: boolean;
|
|
19
|
+
}
|
|
20
|
+
export interface SkillHookReceipt {
|
|
21
|
+
event: string;
|
|
22
|
+
sessionId: string | null;
|
|
23
|
+
turnId: string | null;
|
|
24
|
+
toolUseId: string | null;
|
|
25
|
+
toolName: string | null;
|
|
26
|
+
invocations: string[];
|
|
27
|
+
ambiguousInvocations: string[];
|
|
28
|
+
bodyRanges?: SkillBodyRanges[];
|
|
29
|
+
partial: boolean;
|
|
30
|
+
}
|
|
31
|
+
export interface SkillBodyRanges {
|
|
32
|
+
invocation: string;
|
|
33
|
+
sha256: string;
|
|
34
|
+
ranges: [number, number][];
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
const MIN_PAGE_CHARACTERS = 32;
|
|
38
|
+
interface MatchBudget {
|
|
39
|
+
comparisons: number;
|
|
40
|
+
anchors: number;
|
|
41
|
+
}
|
|
42
|
+
|
|
43
|
+
function spend(budget: MatchBudget, count: number) {
|
|
44
|
+
budget.comparisons -= count;
|
|
45
|
+
if (budget.comparisons < 0)
|
|
46
|
+
throw new Error("skill page matching exceeds limit");
|
|
47
|
+
}
|
|
48
|
+
|
|
49
|
+
function spendAnchor(budget: MatchBudget) {
|
|
50
|
+
budget.anchors--;
|
|
51
|
+
if (budget.anchors < 0) throw new Error("skill page anchors exceed limit");
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
function matchingLength(
|
|
55
|
+
body: string,
|
|
56
|
+
text: string,
|
|
57
|
+
bodyStart: number,
|
|
58
|
+
textStart: number,
|
|
59
|
+
direction: number,
|
|
60
|
+
budget: MatchBudget,
|
|
61
|
+
): number {
|
|
62
|
+
const available =
|
|
63
|
+
direction === 1
|
|
64
|
+
? Math.min(body.length - bodyStart, text.length - textStart)
|
|
65
|
+
: Math.min(bodyStart + 1, textStart + 1);
|
|
66
|
+
for (let length = 0; length < available; length++) {
|
|
67
|
+
spend(budget, 1);
|
|
68
|
+
if (
|
|
69
|
+
body[bodyStart + direction * length] !==
|
|
70
|
+
text[textStart + direction * length]
|
|
71
|
+
)
|
|
72
|
+
return length;
|
|
73
|
+
}
|
|
74
|
+
return available;
|
|
75
|
+
}
|
|
76
|
+
|
|
77
|
+
function lineOffsets(body: string) {
|
|
78
|
+
const offsets = new Map<string, number[]>();
|
|
79
|
+
for (const line of body.matchAll(/[^\n]*(?:\n|$)/g)) {
|
|
80
|
+
if (!line[0].trim()) continue;
|
|
81
|
+
const positions = offsets.get(line[0]) ?? [];
|
|
82
|
+
positions.push(line.index);
|
|
83
|
+
offsets.set(line[0], positions);
|
|
84
|
+
}
|
|
85
|
+
return offsets;
|
|
86
|
+
}
|
|
87
|
+
|
|
88
|
+
function anchoredRange(
|
|
89
|
+
body: string,
|
|
90
|
+
text: string,
|
|
91
|
+
bodyStart: number,
|
|
92
|
+
textStart: number,
|
|
93
|
+
anchorLength: number,
|
|
94
|
+
budget: MatchBudget,
|
|
95
|
+
) {
|
|
96
|
+
spendAnchor(budget);
|
|
97
|
+
const before = matchingLength(
|
|
98
|
+
body,
|
|
99
|
+
text,
|
|
100
|
+
bodyStart - 1,
|
|
101
|
+
textStart - 1,
|
|
102
|
+
-1,
|
|
103
|
+
budget,
|
|
104
|
+
);
|
|
105
|
+
const after = matchingLength(
|
|
106
|
+
body,
|
|
107
|
+
text,
|
|
108
|
+
bodyStart + anchorLength,
|
|
109
|
+
textStart + anchorLength,
|
|
110
|
+
1,
|
|
111
|
+
budget,
|
|
112
|
+
);
|
|
113
|
+
return {
|
|
114
|
+
start: bodyStart - before,
|
|
115
|
+
end: bodyStart + anchorLength + after,
|
|
116
|
+
outputEnd: textStart + anchorLength + after,
|
|
117
|
+
};
|
|
118
|
+
}
|
|
119
|
+
|
|
120
|
+
function uniqueRange(
|
|
121
|
+
body: string,
|
|
122
|
+
start: number,
|
|
123
|
+
end: number,
|
|
124
|
+
budget: MatchBudget,
|
|
125
|
+
) {
|
|
126
|
+
spend(budget, body.length * 2);
|
|
127
|
+
const fragment = body.slice(start, end);
|
|
128
|
+
return (
|
|
129
|
+
body.indexOf(fragment) === start && body.indexOf(fragment, start + 1) === -1
|
|
130
|
+
);
|
|
131
|
+
}
|
|
132
|
+
|
|
133
|
+
function lineRanges(
|
|
134
|
+
body: string,
|
|
135
|
+
text: string,
|
|
136
|
+
line: RegExpExecArray,
|
|
137
|
+
positions: number[],
|
|
138
|
+
budget: MatchBudget,
|
|
139
|
+
) {
|
|
140
|
+
const result = {
|
|
141
|
+
found: false,
|
|
142
|
+
ranges: [] as [number, number][],
|
|
143
|
+
matchedUntil: 0,
|
|
144
|
+
};
|
|
145
|
+
for (const offset of positions) {
|
|
146
|
+
const range = anchoredRange(
|
|
147
|
+
body,
|
|
148
|
+
text,
|
|
149
|
+
offset,
|
|
150
|
+
line.index,
|
|
151
|
+
line[0].length,
|
|
152
|
+
budget,
|
|
153
|
+
);
|
|
154
|
+
if (range.end - range.start < MIN_PAGE_CHARACTERS) continue;
|
|
155
|
+
result.found = true;
|
|
156
|
+
if (uniqueRange(body, range.start, range.end, budget))
|
|
157
|
+
result.ranges.push([range.start, range.end]);
|
|
158
|
+
result.matchedUntil = Math.max(result.matchedUntil, range.outputEnd);
|
|
159
|
+
}
|
|
160
|
+
return result;
|
|
161
|
+
}
|
|
162
|
+
|
|
163
|
+
function textRanges(body: string, text: string, budget: MatchBudget) {
|
|
164
|
+
const offsets = lineOffsets(body);
|
|
165
|
+
const result = { found: false, ranges: [] as [number, number][] };
|
|
166
|
+
let matchedUntil = 0;
|
|
167
|
+
for (const line of text.matchAll(/[^\n]*(?:\n|$)/g)) {
|
|
168
|
+
if (line.index < matchedUntil) continue;
|
|
169
|
+
const match = lineRanges(
|
|
170
|
+
body,
|
|
171
|
+
text,
|
|
172
|
+
line,
|
|
173
|
+
offsets.get(line[0]) ?? [],
|
|
174
|
+
budget,
|
|
175
|
+
);
|
|
176
|
+
result.found ||= match.found;
|
|
177
|
+
result.ranges.push(...match.ranges);
|
|
178
|
+
matchedUntil = Math.max(matchedUntil, match.matchedUntil);
|
|
179
|
+
}
|
|
180
|
+
return result;
|
|
181
|
+
}
|
|
182
|
+
|
|
183
|
+
function skillRanges(
|
|
184
|
+
skill: HookSkill,
|
|
185
|
+
texts: string[],
|
|
186
|
+
budget: MatchBudget,
|
|
187
|
+
): SkillBodyRanges[] {
|
|
188
|
+
const matches = texts.map((text) => textRanges(skill.body, text, budget));
|
|
189
|
+
if (!matches.some((match) => match.found)) return [];
|
|
190
|
+
return [
|
|
191
|
+
{
|
|
192
|
+
invocation: skill.invocation,
|
|
193
|
+
sha256: skill.sha256,
|
|
194
|
+
ranges: matches.flatMap((match) => match.ranges),
|
|
195
|
+
},
|
|
196
|
+
];
|
|
197
|
+
}
|
|
198
|
+
|
|
199
|
+
function label(value: unknown): string | null {
|
|
200
|
+
return typeof value === "string" && /^[A-Za-z0-9_.:-]{1,256}$/.test(value)
|
|
201
|
+
? value
|
|
202
|
+
: null;
|
|
203
|
+
}
|
|
204
|
+
|
|
205
|
+
function responseTexts(value: unknown, depth = 0): string[] {
|
|
206
|
+
if (depth > 16) throw new Error("skill hook response nesting exceeds limit");
|
|
207
|
+
if (typeof value === "string") return [value];
|
|
208
|
+
return responseValues(value).flatMap((child) =>
|
|
209
|
+
responseTexts(child, depth + 1),
|
|
210
|
+
);
|
|
211
|
+
}
|
|
212
|
+
|
|
213
|
+
function responseValues(value: unknown): unknown[] {
|
|
214
|
+
if (Array.isArray(value)) return value;
|
|
215
|
+
return isRecord(value) ? Object.values(value) : [];
|
|
216
|
+
}
|
|
217
|
+
|
|
218
|
+
function completeBodyRanges(
|
|
219
|
+
text: string,
|
|
220
|
+
skill: HookSkill,
|
|
221
|
+
budget: MatchBudget,
|
|
222
|
+
): [number, number][] {
|
|
223
|
+
spend(budget, text.length);
|
|
224
|
+
const ranges: [number, number][] = [];
|
|
225
|
+
let start = text.indexOf(skill.body);
|
|
226
|
+
while (start !== -1) {
|
|
227
|
+
spendAnchor(budget);
|
|
228
|
+
ranges.push([start, start + skill.body.length]);
|
|
229
|
+
start = text.indexOf(skill.body, start + 1);
|
|
230
|
+
}
|
|
231
|
+
return ranges;
|
|
232
|
+
}
|
|
233
|
+
|
|
234
|
+
function textOutsideRanges(text: string, ranges: [number, number][]) {
|
|
235
|
+
const remaining: string[] = [];
|
|
236
|
+
let cursor = 0;
|
|
237
|
+
for (const [start, end] of ranges.sort((a, b) => a[0] - b[0])) {
|
|
238
|
+
if (start > cursor) remaining.push(text.slice(cursor, start));
|
|
239
|
+
cursor = Math.max(cursor, end);
|
|
240
|
+
}
|
|
241
|
+
remaining.push(text.slice(cursor));
|
|
242
|
+
return remaining;
|
|
243
|
+
}
|
|
244
|
+
|
|
245
|
+
function textsOutsideCompleteBodies(
|
|
246
|
+
texts: string[],
|
|
247
|
+
skills: HookSkill[],
|
|
248
|
+
budget: MatchBudget,
|
|
249
|
+
) {
|
|
250
|
+
return texts.flatMap((text) =>
|
|
251
|
+
textOutsideRanges(
|
|
252
|
+
text,
|
|
253
|
+
skills.flatMap((skill) => completeBodyRanges(text, skill, budget)),
|
|
254
|
+
),
|
|
255
|
+
);
|
|
256
|
+
}
|
|
257
|
+
|
|
258
|
+
function invocations(
|
|
259
|
+
event: Record<string, unknown>,
|
|
260
|
+
configuration: SkillHookConfiguration,
|
|
261
|
+
): {
|
|
262
|
+
invocations: string[];
|
|
263
|
+
ambiguousInvocations: string[];
|
|
264
|
+
bodyRanges?: SkillBodyRanges[];
|
|
265
|
+
} {
|
|
266
|
+
if (event.hook_event_name !== "PostToolUse")
|
|
267
|
+
return { invocations: [], ambiguousInvocations: [] };
|
|
268
|
+
const texts = responseTexts(event.tool_response);
|
|
269
|
+
const matched = configuration.skills.filter((skill) =>
|
|
270
|
+
texts.some((text) => text.includes(skill.body)),
|
|
271
|
+
);
|
|
272
|
+
const ambiguous = matched.filter((skill) =>
|
|
273
|
+
matched.some(
|
|
274
|
+
(other) =>
|
|
275
|
+
other.sha256 === skill.sha256 && other.invocation !== skill.invocation,
|
|
276
|
+
),
|
|
277
|
+
);
|
|
278
|
+
const ambiguousNames = new Set(ambiguous.map((skill) => skill.invocation));
|
|
279
|
+
const unambiguous = matched.filter(
|
|
280
|
+
(skill) => !ambiguousNames.has(skill.invocation),
|
|
281
|
+
);
|
|
282
|
+
const budget = { comparisons: 4 * MAX_INPUT_BYTES, anchors: 4096 };
|
|
283
|
+
const remainingTexts = textsOutsideCompleteBodies(texts, unambiguous, budget);
|
|
284
|
+
const bodyRanges = configuration.skills
|
|
285
|
+
.filter((skill) => !matched.includes(skill))
|
|
286
|
+
.flatMap((skill) => skillRanges(skill, remainingTexts, budget));
|
|
287
|
+
return {
|
|
288
|
+
invocations: [...new Set(unambiguous.map((skill) => skill.invocation))],
|
|
289
|
+
ambiguousInvocations: [...ambiguousNames],
|
|
290
|
+
...(bodyRanges.length ? { bodyRanges } : {}),
|
|
291
|
+
};
|
|
292
|
+
}
|
|
293
|
+
|
|
294
|
+
export function skillHookReceipt(
|
|
295
|
+
input: unknown,
|
|
296
|
+
configuration: SkillHookConfiguration,
|
|
297
|
+
): SkillHookReceipt {
|
|
298
|
+
if (!isRecord(input) || typeof input.hook_event_name !== "string")
|
|
299
|
+
throw new Error("invalid skill hook input");
|
|
300
|
+
const receipt = {
|
|
301
|
+
event: input.hook_event_name,
|
|
302
|
+
sessionId: label(input.session_id),
|
|
303
|
+
turnId: label(input.turn_id),
|
|
304
|
+
toolUseId: label(input.tool_use_id),
|
|
305
|
+
toolName: label(input.tool_name),
|
|
306
|
+
...invocations(input, configuration),
|
|
307
|
+
partial: false,
|
|
308
|
+
};
|
|
309
|
+
receipt.partial =
|
|
310
|
+
!validReceiptIdentity(receipt) || receipt.ambiguousInvocations.length > 0;
|
|
311
|
+
return receipt;
|
|
312
|
+
}
|
|
313
|
+
|
|
314
|
+
function validReceiptIdentity(receipt: SkillHookReceipt): boolean {
|
|
315
|
+
if (!receipt.sessionId) return false;
|
|
316
|
+
return (
|
|
317
|
+
!receipt.event.includes("ToolUse") ||
|
|
318
|
+
(receipt.toolUseId !== null && receipt.toolName !== null)
|
|
319
|
+
);
|
|
320
|
+
}
|
|
321
|
+
|
|
322
|
+
async function boundedInput(): Promise<unknown> {
|
|
323
|
+
const chunks: Uint8Array[] = [];
|
|
324
|
+
let bytes = 0;
|
|
325
|
+
for await (const chunk of Bun.stdin.stream()) {
|
|
326
|
+
bytes += chunk.byteLength;
|
|
327
|
+
if (bytes > MAX_INPUT_BYTES) throw new Error("skill hook input too large");
|
|
328
|
+
chunks.push(chunk);
|
|
329
|
+
}
|
|
330
|
+
return JSON.parse(Buffer.concat(chunks).toString("utf8")) as unknown;
|
|
331
|
+
}
|
|
332
|
+
|
|
333
|
+
async function appendReceipt(root: string, receipt: SkillHookReceipt) {
|
|
334
|
+
const path = join(root, "events.jsonl");
|
|
335
|
+
const size = await stat(path).then(
|
|
336
|
+
(file) => file.size,
|
|
337
|
+
() => 0,
|
|
338
|
+
);
|
|
339
|
+
if (size > MAX_RECEIPT_BYTES) {
|
|
340
|
+
await writeFile(join(root, "partial"), "limit\n", { mode: 0o600 });
|
|
341
|
+
return;
|
|
342
|
+
}
|
|
343
|
+
const file = await open(path, "a", 0o600);
|
|
344
|
+
try {
|
|
345
|
+
await file.write(JSON.stringify(receipt) + "\n");
|
|
346
|
+
} finally {
|
|
347
|
+
await file.close();
|
|
348
|
+
}
|
|
349
|
+
}
|
|
350
|
+
|
|
351
|
+
/** A silent observer: errors affect evidence availability, never tool decisions. */
|
|
352
|
+
export async function runSkillHook(configurationPath: string): Promise<void> {
|
|
353
|
+
const configuration = JSON.parse(
|
|
354
|
+
await readFile(configurationPath, "utf8"),
|
|
355
|
+
) as SkillHookConfiguration;
|
|
356
|
+
try {
|
|
357
|
+
await appendReceipt(
|
|
358
|
+
configuration.receipts,
|
|
359
|
+
skillHookReceipt(await boundedInput(), configuration),
|
|
360
|
+
);
|
|
361
|
+
} catch {
|
|
362
|
+
await writeFile(join(configuration.receipts, "partial"), randomUUID(), {
|
|
363
|
+
mode: 0o600,
|
|
364
|
+
});
|
|
365
|
+
}
|
|
366
|
+
}
|
|
367
|
+
|
|
368
|
+
if (import.meta.main) {
|
|
369
|
+
try {
|
|
370
|
+
await runSkillHook(process.argv[2] ?? "");
|
|
371
|
+
} catch {
|
|
372
|
+
// Missing receipts remain unavailable; hook failures never alter the task.
|
|
373
|
+
}
|
|
374
|
+
}
|