vigiles 39.2.0 → 40.1.0
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/README.md +9 -8
- package/dist/adapter.d.ts +3 -2
- package/dist/adapters/claude-code/dialect.js +4 -0
- package/dist/adapters/claude-code/output-style.js +15 -0
- package/dist/adapters/claude-code/skill-listing.d.ts +48 -0
- package/dist/adapters/claude-code/skill-listing.js +54 -0
- package/dist/adapters/codex/dialect.js +4 -0
- package/dist/adapters/codex/layout.js +5 -0
- package/dist/adapters/codex/mock-model.d.ts +7 -0
- package/dist/adapters/codex/mock-model.js +1 -0
- package/dist/adapters/codex/skill-listing.d.ts +42 -0
- package/dist/adapters/codex/skill-listing.js +40 -0
- package/dist/audit-html.js +7 -2
- package/dist/audit-report.d.ts +51 -0
- package/dist/audit-report.js +54 -8
- package/dist/audit-report.template.html +33 -38
- package/dist/cli-main.d.ts +9 -3
- package/dist/cli-main.js +32 -17
- package/dist/{instruction-weight-ratchet.d.ts → context-creep-ratchet.d.ts} +34 -12
- package/dist/context-creep-ratchet.js +270 -0
- package/dist/core/always-loaded.d.ts +128 -0
- package/dist/core/always-loaded.js +134 -0
- package/dist/core/compile.js +1 -1
- package/dist/core/dialect.d.ts +9 -0
- package/dist/core/harness-unit.d.ts +20 -0
- package/dist/core/harness-unit.js +24 -0
- package/dist/core/instruction-baseline.d.ts +179 -11
- package/dist/core/instruction-baseline.js +349 -47
- package/dist/core/instruction-weight.d.ts +2 -11
- package/dist/core/instruction-weight.js +6 -16
- package/dist/core/layout.d.ts +40 -0
- package/dist/core/layout.js +25 -1
- package/dist/core/output-style.d.ts +20 -0
- package/dist/core/skill-listing.d.ts +49 -0
- package/dist/core/skill-listing.js +23 -0
- package/dist/core/types.d.ts +2 -2
- package/dist/deprecated-key-files.d.ts +1 -1
- package/dist/deprecated-key-files.js +1 -1
- package/dist/fs-walk.d.ts +31 -0
- package/dist/fs-walk.js +49 -0
- package/dist/local-files.js +2 -1
- package/dist/plugin-loader.js +40 -19
- package/dist/rules/index.d.ts +4 -4
- package/dist/rules/index.js +7 -7
- package/dist/scan-files.js +24 -3
- package/dist/scan.d.ts +30 -1
- package/dist/scan.js +69 -3
- package/dist/surface-discovery-fs.d.ts +15 -0
- package/dist/surface-discovery-fs.js +55 -0
- package/dist/vigilesrc.schema.json +2 -2
- package/hooks/session-start.sh +3 -1
- package/package.json +1 -1
- package/dist/instruction-weight-ratchet.js +0 -181
package/README.md
CHANGED
|
@@ -73,13 +73,14 @@
|
|
|
73
73
|
block; ONE idea per sentence; NO internal vocabulary (moat/flywheel) / NO
|
|
74
74
|
research/ links / NO enterprise/national-interest framing — name the user
|
|
75
75
|
benefit; ~220-line body cap; push depth into docs/ and LINK it. Assets: the hero
|
|
76
|
-
vigiles-audit.png is a REPRESENTATIVE report ("my-plugin", C 77) — the
|
|
77
|
-
|
|
78
|
-
enforced" preview is DEMOTED low + badged EXPERIMENTAL
|
|
79
|
-
research/rule-enforcer-design.md §8)
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
prefers-color-scheme:dark emulated + a fullPage screenshot
|
|
76
|
+
vigiles-audit.png is a REPRESENTATIVE report ("my-plugin", C 77) — the grade
|
|
77
|
+
gauge + category rings + "Do these first" + broken references + always-loaded
|
|
78
|
+
context. The "Your rules → enforced" preview is DEMOTED low + badged EXPERIMENTAL
|
|
79
|
+
(the rule map is alpha — research/rule-enforcer-design.md §8) and sits below the
|
|
80
|
+
crop. Re-render if the UI changes: `npm run build` (rebuilds
|
|
81
|
+
report/dist/index.html from the SAMPLE fixture), then headless Chromium at a
|
|
82
|
+
1100px viewport with prefers-color-scheme:dark emulated + a fullPage screenshot
|
|
83
|
+
at deviceScaleFactor 2, cropped at the bottom of the always-loaded section
|
|
83
84
|
→ vigiles-audit.png (2200px wide).
|
|
84
85
|
-->
|
|
85
86
|
|
|
@@ -127,7 +128,7 @@ It's free and open-source, runs entirely on your machine, and never bills per to
|
|
|
127
128
|
## What it caught
|
|
128
129
|
|
|
129
130
|
<p align="center">
|
|
130
|
-
<img src="vigiles-audit.png" width="760" alt="vigiles audit report for my-plugin: a
|
|
131
|
+
<img src="vigiles-audit.png" width="760" alt="vigiles audit report for my-plugin: a C grade gauge (77/100) beside the verdict 'Two one-line fixes away from a B.' and '3 points to B', score rings for Truthfulness, Triggering, Structure and Safety plus dashed advisory rings for Tested and Evaluated, the findings in full, a numbered 'Do these first' list with '+N pts' gains, broken-reference findings, and the always-loaded context total split into three shares" />
|
|
131
132
|
</p>
|
|
132
133
|
|
|
133
134
|
**Like Google's Lighthouse, but for your agent harness.** One command grades it A–F across six categories, leads with a plain-English verdict — _"two one-line fixes away from a B"_ — and ranks every fix by the points it buys back:
|
package/dist/adapter.d.ts
CHANGED
|
@@ -17,8 +17,9 @@
|
|
|
17
17
|
*/
|
|
18
18
|
export type { HarnessAdapter, AdapterCapabilities } from "./core/adapter.js";
|
|
19
19
|
export type { HarnessDialect } from "./core/dialect.js";
|
|
20
|
-
export type { PluginLayout, SurfaceDirs, SurfaceKind } from "./core/layout.js";
|
|
21
|
-
export type { OutputStyle, OutputStyleRules } from "./core/output-style.js";
|
|
20
|
+
export type { PluginLayout, SkillDiscovery, SurfaceDirs, SurfaceKind, } from "./core/layout.js";
|
|
21
|
+
export type { OutputStyle, OutputStyleRules, StyleSelection, } from "./core/output-style.js";
|
|
22
|
+
export type { ListedSkill, SkillListingRules } from "./core/skill-listing.js";
|
|
22
23
|
export type { ModelRequest } from "./core/harness-driver.js";
|
|
23
24
|
export { executableSourceDirs, materializePrefix, surfaceDirs, } from "./core/layout.js";
|
|
24
25
|
/**
|
|
@@ -2,6 +2,7 @@
|
|
|
2
2
|
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
3
|
exports.claudeCodeDialect = exports.claudeCodeSideEffectingTools = exports.claudeCodeBuiltinAgentTools = void 0;
|
|
4
4
|
const event_capability_js_1 = require("./event-capability.js");
|
|
5
|
+
const skill_listing_js_1 = require("./skill-listing.js");
|
|
5
6
|
const vocabulary_js_1 = require("./vocabulary.js");
|
|
6
7
|
/**
|
|
7
8
|
* The Claude Code built-in subagent tool catalog as a `const` tuple, so a typed
|
|
@@ -92,6 +93,9 @@ exports.claudeCodeDialect = {
|
|
|
92
93
|
// re-read on every request whatever the window. Read from the 2.1.296 binary.
|
|
93
94
|
capturedFrom: "claude-code 2.1.296 large-file warning, floor of max(40000, window*0.05*charsPerToken); the rules dir measured 2026-09-16 in a consumer repo",
|
|
94
95
|
},
|
|
96
|
+
// The session-start skill listing — part of the always-loaded context; see
|
|
97
|
+
// ./skill-listing.ts for each rule and its source.
|
|
98
|
+
skillListing: skill_listing_js_1.claudeCodeSkillListing,
|
|
95
99
|
// The capability table — what each event CARRIES and HONOURS. Nine of the 31
|
|
96
100
|
// events, each with its basis; see ./event-capability.ts. The three flat lists
|
|
97
101
|
// above are kept for now and are ASSERTED against this table in
|
|
@@ -68,6 +68,20 @@ const reached = (style, request) => {
|
|
|
68
68
|
const probe = squash(`${HEADER}${style.name}\n${style.body}`);
|
|
69
69
|
return [request.system, ...request.messages.map((m) => m.text)].some((text) => squash(text).includes(probe));
|
|
70
70
|
};
|
|
71
|
+
/**
|
|
72
|
+
* `outputStyle` in one settings file. `default` (any case) is Claude Code's own
|
|
73
|
+
* style, which adds no text; anything else is a name, matched by the harness
|
|
74
|
+
* against each style's `name` in exact case — a wrong case silently loads the
|
|
75
|
+
* default, which is why the resolution against files stays exact.
|
|
76
|
+
*/
|
|
77
|
+
const selected = (settings) => {
|
|
78
|
+
const value = settings["outputStyle"];
|
|
79
|
+
if (typeof value !== "string" || value === "")
|
|
80
|
+
return { kind: "unset" };
|
|
81
|
+
return value.toLowerCase() === "default"
|
|
82
|
+
? { kind: "default" }
|
|
83
|
+
: { kind: "named", name: value };
|
|
84
|
+
};
|
|
71
85
|
exports.claudeCodeOutputStyles = {
|
|
72
86
|
dir: "output-styles",
|
|
73
87
|
isStyleFile: (pathInDir) => pathInDir.endsWith(".md"),
|
|
@@ -77,5 +91,6 @@ exports.claudeCodeOutputStyles = {
|
|
|
77
91
|
// project; the project's own `default` wins over it (measured on 2.1.292).
|
|
78
92
|
selectNone: { outputStyle: "default" },
|
|
79
93
|
reached,
|
|
94
|
+
selected,
|
|
80
95
|
};
|
|
81
96
|
//# sourceMappingURL=output-style.js.map
|
|
@@ -0,0 +1,48 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* How Claude Code renders a skill into the listing it loads at session start.
|
|
3
|
+
*
|
|
4
|
+
* MODELLED ON THE PINNED VERSION, `VALIDATED_CC_VERSION` in `dialect-drift.ts`
|
|
5
|
+
* (2.1.187) — the version the dialect is validated against and the one CI
|
|
6
|
+
* installs, so `skill-listing-delivery.test.ts` proves these rules against the
|
|
7
|
+
* binary on every run. Read from https://code.claude.com/docs/en/skills
|
|
8
|
+
* (fetched 2026-10-31) and measured against the mock model:
|
|
9
|
+
*
|
|
10
|
+
* - the LABEL is the skill's directory name, whatever the frontmatter `name`;
|
|
11
|
+
* - the text is the `description`, else the first non-empty line of the
|
|
12
|
+
* body, else the word `Skill`, trimmed; `when_to_use` "is appended to
|
|
13
|
+
* `description` … and counts toward the 1,536-character cap", joined by
|
|
14
|
+
* ` - `;
|
|
15
|
+
* - nothing else is rewritten: `<`, `>`, `&` and control characters reach
|
|
16
|
+
* the model as written;
|
|
17
|
+
* - the CAP applies to the combined text: over 1,536 UTF-16 units (or
|
|
18
|
+
* `skillListingMaxDescChars`), the first cap−1 units are kept and `…`
|
|
19
|
+
* appended — a cut that can split a surrogate pair, as Claude Code's does;
|
|
20
|
+
* - `disable-model-invocation: true` "removes the skill from Claude's context
|
|
21
|
+
* entirely";
|
|
22
|
+
* - `skillOverrides` is keyed by the DIRECTORY name (a key equal to the
|
|
23
|
+
* frontmatter name does nothing): `"name-only"` lists `- dir` alone,
|
|
24
|
+
* `"off"` and `"user-invocable-only"` do not list it.
|
|
25
|
+
*
|
|
26
|
+
* ⚠️ NEWER RELEASES RENDER MORE, so on them this is an UNDER-count. Measured on
|
|
27
|
+
* 2.1.316, against the same fixture:
|
|
28
|
+
*
|
|
29
|
+
* - `<` and `>` become `<` and `>` BEFORE the cap — +3 characters
|
|
30
|
+
* each (`- alpha: Run <cmd> now & then - When the user says <go>`, 55
|
|
31
|
+
* characters on 2.1.187, is 67 on 2.1.316);
|
|
32
|
+
* - a frontmatter `name` that differs from the directory follows it in
|
|
33
|
+
* parentheses, `- dir (name): …` — +3 plus the name's length;
|
|
34
|
+
* - each control, format, private-use or separator code point becomes one
|
|
35
|
+
* space — the same length, a different text.
|
|
36
|
+
*
|
|
37
|
+
* Moving to those rules is a bump of `VALIDATED_CC_VERSION`, which moves CI's
|
|
38
|
+
* binary with it; modelling them while CI runs the pinned release would leave
|
|
39
|
+
* the delivery test proving nothing.
|
|
40
|
+
*
|
|
41
|
+
* The listing as a WHOLE also has a budget (1% of the model's context window),
|
|
42
|
+
* past which descriptions are dropped by least use. That budget depends on the
|
|
43
|
+
* model and on this machine's usage, so it is not a property of a commit and
|
|
44
|
+
* is not modelled here; every entry is counted at its full capped size.
|
|
45
|
+
*/
|
|
46
|
+
import type { SkillListingRules } from "../../core/skill-listing.js";
|
|
47
|
+
export declare const claudeCodeSkillListing: SkillListingRules;
|
|
48
|
+
//# sourceMappingURL=skill-listing.d.ts.map
|
|
@@ -0,0 +1,54 @@
|
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.claudeCodeSkillListing = void 0;
|
|
4
|
+
/** The documented per-entry cap, in characters (UTF-16 units, as measured). */
|
|
5
|
+
const ENTRY_CAP = 1536;
|
|
6
|
+
/** What replaces the cut tail; the cap includes it. */
|
|
7
|
+
const ELLIPSIS = "…";
|
|
8
|
+
/** The text an entry carries when the skill gives none. */
|
|
9
|
+
const NO_DESCRIPTION = "Skill";
|
|
10
|
+
/** The trimmed text, or `undefined` when nothing is left. */
|
|
11
|
+
function trimmed(v) {
|
|
12
|
+
if (typeof v !== "string")
|
|
13
|
+
return undefined;
|
|
14
|
+
const text = v.trim();
|
|
15
|
+
return text === "" ? undefined : text;
|
|
16
|
+
}
|
|
17
|
+
const isRecord = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
|
|
18
|
+
/** The configured cap when it is a positive integer, else the documented one. */
|
|
19
|
+
const capFrom = (settings) => {
|
|
20
|
+
const v = settings["skillListingMaxDescChars"];
|
|
21
|
+
return typeof v === "number" && Number.isInteger(v) && v > 0 ? v : ENTRY_CAP;
|
|
22
|
+
};
|
|
23
|
+
const capped = (text, cap) => text.length <= cap ? text : text.slice(0, cap - 1) + ELLIPSIS;
|
|
24
|
+
const firstLine = (body) => body
|
|
25
|
+
.split("\n")
|
|
26
|
+
.map(trimmed)
|
|
27
|
+
.find((l) => l !== undefined);
|
|
28
|
+
/** `skillOverrides[dir]`, when the settings carry one for this skill. */
|
|
29
|
+
const overrideFor = (settings, dirName) => {
|
|
30
|
+
const overrides = settings["skillOverrides"];
|
|
31
|
+
return isRecord(overrides) ? overrides[dirName] : undefined;
|
|
32
|
+
};
|
|
33
|
+
const entry = (skill, settings) => {
|
|
34
|
+
const fm = skill.frontmatter;
|
|
35
|
+
// A skill whose frontmatter is not valid YAML is not registered by the
|
|
36
|
+
// harness (observed in a dogfood session: fixing the YAML made the skill
|
|
37
|
+
// appear in the listing it had been missing from).
|
|
38
|
+
if (fm === null || fm["disable-model-invocation"] === true)
|
|
39
|
+
return null;
|
|
40
|
+
const override = overrideFor(settings, skill.dirName);
|
|
41
|
+
if (override === "off" || override === "user-invocable-only")
|
|
42
|
+
return null;
|
|
43
|
+
if (override === "name-only")
|
|
44
|
+
return `- ${skill.dirName}`;
|
|
45
|
+
const description = trimmed(fm["description"]) ?? firstLine(skill.body) ?? NO_DESCRIPTION;
|
|
46
|
+
const whenToUse = trimmed(fm["when_to_use"]);
|
|
47
|
+
const combined = whenToUse === undefined ? description : `${description} - ${whenToUse}`;
|
|
48
|
+
return `- ${skill.dirName}: ${capped(combined, capFrom(settings))}`;
|
|
49
|
+
};
|
|
50
|
+
exports.claudeCodeSkillListing = {
|
|
51
|
+
entry,
|
|
52
|
+
capturedFrom: "code.claude.com/docs/en/skills (fetched 2026-10-31); measured on Claude Code 2.1.187 (VALIDATED_CC_VERSION)",
|
|
53
|
+
};
|
|
54
|
+
//# sourceMappingURL=skill-listing.js.map
|
|
@@ -1,6 +1,7 @@
|
|
|
1
1
|
"use strict";
|
|
2
2
|
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
3
|
exports.codexDialect = void 0;
|
|
4
|
+
const skill_listing_js_1 = require("./skill-listing.js");
|
|
4
5
|
exports.codexDialect = {
|
|
5
6
|
name: "codex",
|
|
6
7
|
// Codex model-facing tools (shell/exec, file edits, plan, web), + MCP tools.
|
|
@@ -44,6 +45,9 @@ exports.codexDialect = {
|
|
|
44
45
|
// DIRECTORY along root→cwd, so at the root it is one file. A monorepo with
|
|
45
46
|
// twelve package-level AGENTS.md files was told it was 12x over a budget no
|
|
46
47
|
// session approaches. Which files load is now `codexLayout.instructionChain`.
|
|
48
|
+
// The session-start skill listing — part of the always-loaded context, counted in
|
|
49
|
+
// the same bytes as the instructions; see ./skill-listing.ts.
|
|
50
|
+
skillListing: skill_listing_js_1.codexSkillListing,
|
|
47
51
|
instructionTargets: ["AGENTS.md"],
|
|
48
52
|
pluginRootToken: "${PLUGIN_ROOT}",
|
|
49
53
|
// Codex SKILL.md frontmatter is name + description ONLY — the richer keys
|
|
@@ -16,6 +16,11 @@ exports.codexLayout = {
|
|
|
16
16
|
// NOT share a parent (`.agents/` vs `.codex/`), so no single root can name
|
|
17
17
|
// both; spelling each dir in full is the only shape that keeps the reported key
|
|
18
18
|
// equal to the real on-disk path.
|
|
19
|
+
// Repository skill roots are walked recursively: a SKILL.md directly in the
|
|
20
|
+
// folder up to six directories down, `.`-directories pruned
|
|
21
|
+
// (`codex-rs/ext/skills/src/loader/discovery.rs`; measured on 0.153.0 — see
|
|
22
|
+
// `SkillDiscovery`).
|
|
23
|
+
skillDiscovery: { minDepth: 0, maxDepth: 6, skipHidden: true },
|
|
19
24
|
surfaces: {
|
|
20
25
|
skill: ".agents/skills",
|
|
21
26
|
// No `agent` key: Codex `[agents]` is a TOML concurrency table, not a
|
|
@@ -49,5 +49,12 @@ export declare function parseResponsesRequest(body: string): {
|
|
|
49
49
|
*/
|
|
50
50
|
export declare function startCodexMock(script: readonly CodexTurn[], opts?: {
|
|
51
51
|
onRequest?: (req: CodexRequest) => void;
|
|
52
|
+
/**
|
|
53
|
+
* The raw `/v1/responses` body, before {@link parseResponsesRequest}
|
|
54
|
+
* narrows it to the prompt. For a test that must see what the parse drops
|
|
55
|
+
* — the instructions and developer messages, where Codex puts its skill
|
|
56
|
+
* listing.
|
|
57
|
+
*/
|
|
58
|
+
onBody?: (body: string) => void;
|
|
52
59
|
}): Promise<CodexMockHandle>;
|
|
53
60
|
//# sourceMappingURL=mock-model.d.ts.map
|
|
@@ -0,0 +1,42 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* How Codex renders a skill into the listing it loads at session start.
|
|
3
|
+
*
|
|
4
|
+
* Read from the Codex source (`openai/codex`, main, fetched 2026-10-31) and
|
|
5
|
+
* measured on codex-cli 0.153.0 against the mock model
|
|
6
|
+
* (`skill-listing-delivery.test.ts` pins it):
|
|
7
|
+
*
|
|
8
|
+
* - WHICH SKILLS ARE LISTED (`codex-rs/skills/src/parser.rs`,
|
|
9
|
+
* `parse_skill_frontmatter_metadata`). A SKILL.md with no frontmatter
|
|
10
|
+
* block, with a block that is not valid YAML, with an empty description,
|
|
11
|
+
* or with a name over 64 characters fails to parse, is pushed to the
|
|
12
|
+
* loader's errors, and is not listed at all.
|
|
13
|
+
* - THE TEXT. Name and description each pass through `sanitize_single_line`
|
|
14
|
+
* (`split_whitespace().join(" ")`), so a multi-line description is one
|
|
15
|
+
* line. A missing name falls back to the skill's directory name.
|
|
16
|
+
* - THE ENTRY (`codex-rs/ext/skills/src/render.rs`): one line,
|
|
17
|
+
* `- {name}: {description} ({locator_kind}: {locator})`, where a skill on
|
|
18
|
+
* disk prints `file`.
|
|
19
|
+
* - THE CAP (`truncate_catalog_skill_description`): a description longer
|
|
20
|
+
* than 1,024 characters keeps its first 1,021 and gains `...`. Characters
|
|
21
|
+
* are Unicode scalar values, so an emoji is one character, not two.
|
|
22
|
+
*
|
|
23
|
+
* ⚠️ NOT MIRRORED, each an UNDER-count:
|
|
24
|
+
*
|
|
25
|
+
* - Codex repairs one kind of invalid YAML before giving up: an unquoted
|
|
26
|
+
* scalar holding `: `, as in `description: Build for AWS: ECS`, is quoted
|
|
27
|
+
* line by line and parsed again. vigiles reads such a block as malformed
|
|
28
|
+
* and does not list the skill, where Codex does.
|
|
29
|
+
* - THE LOCATOR is per machine. Codex prints either the absolute path or a
|
|
30
|
+
* short alias (`r1/<path under the skills folder>`, beside a table that
|
|
31
|
+
* maps `r1` to the absolute folder), whichever renders cheaper. Neither can
|
|
32
|
+
* be committed and compared in CI, so this module writes the repo-relative
|
|
33
|
+
* path; the size is reproducible from a commit, and each entry is off by
|
|
34
|
+
* the difference between that path and the one Codex printed.
|
|
35
|
+
*
|
|
36
|
+
* The listing as a whole has a budget (2% of the model's context window, or a
|
|
37
|
+
* configured token count), past which descriptions are dropped. It depends on
|
|
38
|
+
* the model, so it is not a property of a commit and is not modelled here.
|
|
39
|
+
*/
|
|
40
|
+
import type { SkillListingRules } from "../../core/skill-listing.js";
|
|
41
|
+
export declare const codexSkillListing: SkillListingRules;
|
|
42
|
+
//# sourceMappingURL=skill-listing.d.ts.map
|
|
@@ -0,0 +1,40 @@
|
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.codexSkillListing = void 0;
|
|
4
|
+
/** Longer descriptions are cut, in Unicode scalar values. */
|
|
5
|
+
const DESCRIPTION_CAP = 1024;
|
|
6
|
+
/** What replaces the cut tail; the cap includes it. */
|
|
7
|
+
const TRUNCATION_SUFFIX = "...";
|
|
8
|
+
/** A longer name fails to parse, in Unicode scalar values. */
|
|
9
|
+
const NAME_CAP = 64;
|
|
10
|
+
/** `sanitize_single_line`: every run of whitespace becomes one space. */
|
|
11
|
+
const singleLine = (v) => v.split(/\s+/u).filter(Boolean).join(" ");
|
|
12
|
+
/** A frontmatter value as Codex reads it, or `""` when it is not text. */
|
|
13
|
+
const field = (v) => typeof v === "string" ? singleLine(v) : "";
|
|
14
|
+
/** `truncate_catalog_skill_description`, counted in code points. */
|
|
15
|
+
function capped(description) {
|
|
16
|
+
// Code points, as Rust's `chars()` counts — not UTF-16 units, not graphemes.
|
|
17
|
+
const chars = Array.from(description);
|
|
18
|
+
return chars.length <= DESCRIPTION_CAP
|
|
19
|
+
? description
|
|
20
|
+
: chars.slice(0, DESCRIPTION_CAP - TRUNCATION_SUFFIX.length).join("") +
|
|
21
|
+
TRUNCATION_SUFFIX;
|
|
22
|
+
}
|
|
23
|
+
const entry = (skill) => {
|
|
24
|
+
const fm = skill.frontmatter;
|
|
25
|
+
// `{}` is a file with no block, `null` a block that is not valid YAML; both
|
|
26
|
+
// are parse errors to Codex, and an empty description is one too.
|
|
27
|
+
if (fm === null)
|
|
28
|
+
return null;
|
|
29
|
+
const name = field(fm["name"]) || singleLine(skill.dirName);
|
|
30
|
+
const description = field(fm["description"]);
|
|
31
|
+
if (description === "" || name === "" || Array.from(name).length > NAME_CAP) {
|
|
32
|
+
return null;
|
|
33
|
+
}
|
|
34
|
+
return `- ${name}: ${capped(description)} (file: ${skill.path})`;
|
|
35
|
+
};
|
|
36
|
+
exports.codexSkillListing = {
|
|
37
|
+
entry,
|
|
38
|
+
capturedFrom: "openai/codex codex-rs/skills/src/parser.rs + codex-rs/ext/skills/src/render.rs (main, fetched 2026-10-31); measured on codex-cli 0.153.0",
|
|
39
|
+
};
|
|
40
|
+
//# sourceMappingURL=skill-listing.js.map
|
package/dist/audit-html.js
CHANGED
|
@@ -46,11 +46,16 @@ function injectReportData(template, report, serve) {
|
|
|
46
46
|
if (!re.test(template)) {
|
|
47
47
|
throw new Error("audit report template is missing the data placeholder");
|
|
48
48
|
}
|
|
49
|
-
|
|
49
|
+
// Every replacement goes through a function: a replacement STRING reads `$1`,
|
|
50
|
+
// `$&`, `$'` and the like as patterns, and report text quotes files that say
|
|
51
|
+
// "$120/h". A function's return value is used verbatim.
|
|
52
|
+
const data = escapeForScript(JSON.stringify(report));
|
|
53
|
+
let out = template.replace(re, () => data);
|
|
50
54
|
if (serve) {
|
|
51
55
|
// Prepend the serve global onto the same statement that carries the data, so
|
|
52
56
|
// it's set before the app bundle reads it. The token is hex, but escape anyway.
|
|
53
|
-
|
|
57
|
+
const prefix = `window.__VIGILES_SERVE__=${escapeForScript(JSON.stringify(serve))};`;
|
|
58
|
+
out = out.replace("window.__VIGILES_DATA__", () => `${prefix}window.__VIGILES_DATA__`);
|
|
54
59
|
}
|
|
55
60
|
return out;
|
|
56
61
|
}
|
package/dist/audit-report.d.ts
CHANGED
|
@@ -22,6 +22,7 @@ import type { RuleInventoryItem } from "./rule-inventory.js";
|
|
|
22
22
|
import type { RuleRouting } from "./rule-routing.js";
|
|
23
23
|
import type { EvidenceCounts } from "./coverage-evidence.js";
|
|
24
24
|
import type { DeprecatedKeyFilesFinding } from "./deprecated-key-files.js";
|
|
25
|
+
import type { LoadedPartName } from "./core/always-loaded.js";
|
|
25
26
|
import { type FilteredScan, type UngradedGroup } from "./core/apply-rules.js";
|
|
26
27
|
/**
|
|
27
28
|
* The current schema version. Bump only on a BREAKING change to the shape.
|
|
@@ -107,6 +108,49 @@ export interface Adoptable {
|
|
|
107
108
|
/** The one command that adopts every surface at once. */
|
|
108
109
|
readonly createAllCommand: string;
|
|
109
110
|
}
|
|
111
|
+
/** One file's share of an always-loaded part, in the report's unit. */
|
|
112
|
+
export interface AlwaysLoadedFile {
|
|
113
|
+
readonly path: string;
|
|
114
|
+
readonly size: number;
|
|
115
|
+
}
|
|
116
|
+
/**
|
|
117
|
+
* One part of the always-loaded context. Only a `counted` part has a size: every other
|
|
118
|
+
* state is a reason there is none, kept apart because each reads differently
|
|
119
|
+
* (nothing selected · the harness has no such thing · vigiles does not model
|
|
120
|
+
* it · a style selected by a name no repository file declares).
|
|
121
|
+
*/
|
|
122
|
+
export type AlwaysLoadedPart = {
|
|
123
|
+
readonly part: "instructions" | LoadedPartName;
|
|
124
|
+
readonly kind: "counted";
|
|
125
|
+
readonly total: number;
|
|
126
|
+
readonly files: readonly AlwaysLoadedFile[];
|
|
127
|
+
} | {
|
|
128
|
+
readonly part: LoadedPartName;
|
|
129
|
+
readonly kind: "none" | "not-supported" | "not-modelled";
|
|
130
|
+
} | {
|
|
131
|
+
readonly part: "output-style";
|
|
132
|
+
readonly kind: "unresolved";
|
|
133
|
+
/** The name the committed settings select. */
|
|
134
|
+
readonly name: string;
|
|
135
|
+
};
|
|
136
|
+
/**
|
|
137
|
+
* What the harness loads at session start before any decision: the
|
|
138
|
+
* instructions, the output style the committed settings select and the skill
|
|
139
|
+
* listing, in the harness's own unit (characters or bytes, never tokens). The
|
|
140
|
+
* part of every request the REPOSITORY decides — the harness's own prompt,
|
|
141
|
+
* user and plugin skills and the conversation come on top.
|
|
142
|
+
*/
|
|
143
|
+
export interface AlwaysLoadedReport {
|
|
144
|
+
readonly unit: "chars" | "bytes";
|
|
145
|
+
/** The instructions plus every counted part. */
|
|
146
|
+
readonly total: number;
|
|
147
|
+
/** Always `instructions`, `output-style`, `skill-listing`, in that order. */
|
|
148
|
+
readonly parts: readonly AlwaysLoadedPart[];
|
|
149
|
+
}
|
|
150
|
+
/** How much context the harness carries before the conversation starts. */
|
|
151
|
+
export interface ContextCostReport {
|
|
152
|
+
readonly alwaysLoaded: AlwaysLoadedReport;
|
|
153
|
+
}
|
|
110
154
|
/**
|
|
111
155
|
* The full audit, as the dashboard / CI / HTML all consume it. Self-describing
|
|
112
156
|
* and versioned; additive-only within a `schemaVersion`.
|
|
@@ -190,6 +234,13 @@ export interface AuditReport {
|
|
|
190
234
|
* Present only when at least one atomic rule was segmented. Additive/optional.
|
|
191
235
|
*/
|
|
192
236
|
readonly ruleRouting?: RuleRouting;
|
|
237
|
+
/**
|
|
238
|
+
* The context the harness carries before any decision, per part
|
|
239
|
+
* ({@link ContextCostReport}). Present whenever the harness publishes an
|
|
240
|
+
* instruction budget — the same condition as the terminal's weight block.
|
|
241
|
+
* Additive/optional — schema version unchanged.
|
|
242
|
+
*/
|
|
243
|
+
readonly contextCost?: ContextCostReport;
|
|
193
244
|
}
|
|
194
245
|
export interface BuildAuditReportOptions extends AuditScoreOptions {
|
|
195
246
|
readonly harness: string;
|
package/dist/audit-report.js
CHANGED
|
@@ -75,6 +75,47 @@ function inventoryOf(report) {
|
|
|
75
75
|
function outputStylesField(n) {
|
|
76
76
|
return n === undefined || n === 0 ? {} : { outputStyles: n };
|
|
77
77
|
}
|
|
78
|
+
/** One scan-side part as the report carries it. */
|
|
79
|
+
function loadedPartReport(part, value) {
|
|
80
|
+
switch (value.kind) {
|
|
81
|
+
case "counted":
|
|
82
|
+
return { part, kind: "counted", total: value.total, files: value.files };
|
|
83
|
+
case "unresolved":
|
|
84
|
+
return { part: "output-style", kind: "unresolved", name: value.name };
|
|
85
|
+
case "none":
|
|
86
|
+
case "not-supported":
|
|
87
|
+
case "not-modelled":
|
|
88
|
+
return { part, kind: value.kind };
|
|
89
|
+
}
|
|
90
|
+
}
|
|
91
|
+
/** The report's `contextCost`, or nothing when the scan weighed no instructions. */
|
|
92
|
+
function contextCostField(report) {
|
|
93
|
+
const weight = report.instructionWeight;
|
|
94
|
+
const loaded = report.alwaysLoaded;
|
|
95
|
+
if (weight === null || loaded === undefined)
|
|
96
|
+
return {};
|
|
97
|
+
const files = weight.files
|
|
98
|
+
.filter((f) => f.scope === "repo")
|
|
99
|
+
.map((f) => ({ path: f.path, size: f.size }));
|
|
100
|
+
return {
|
|
101
|
+
contextCost: {
|
|
102
|
+
alwaysLoaded: {
|
|
103
|
+
unit: loaded.unit,
|
|
104
|
+
total: loaded.total,
|
|
105
|
+
parts: [
|
|
106
|
+
{
|
|
107
|
+
part: "instructions",
|
|
108
|
+
kind: "counted",
|
|
109
|
+
total: loaded.instructions,
|
|
110
|
+
files,
|
|
111
|
+
},
|
|
112
|
+
loadedPartReport("output-style", loaded.outputStyle),
|
|
113
|
+
loadedPartReport("skill-listing", loaded.skillListing),
|
|
114
|
+
],
|
|
115
|
+
},
|
|
116
|
+
},
|
|
117
|
+
};
|
|
118
|
+
}
|
|
78
119
|
/**
|
|
79
120
|
* The scan's concrete finding lists, each present only when non-empty — and the
|
|
80
121
|
* ungraded groups' blocks, present only when some rule of one is on.
|
|
@@ -101,6 +142,17 @@ function gradedPart(report, opts) {
|
|
|
101
142
|
const verdict = (0, audit_verdict_js_1.computeVerdict)({ report: graded, score, recommendations });
|
|
102
143
|
return { score, verdict, recommendations };
|
|
103
144
|
}
|
|
145
|
+
/** The report's self-description; `generatedAt` is the CLI's, never set here. */
|
|
146
|
+
function auditMeta(dir, opts) {
|
|
147
|
+
return {
|
|
148
|
+
schemaVersion: exports.AUDIT_SCHEMA_VERSION,
|
|
149
|
+
tool: "vigiles",
|
|
150
|
+
kind: "audit",
|
|
151
|
+
vigilesVersion: opts.vigilesVersion,
|
|
152
|
+
harness: opts.harness,
|
|
153
|
+
dir,
|
|
154
|
+
};
|
|
155
|
+
}
|
|
104
156
|
/**
|
|
105
157
|
* Assemble the versioned {@link AuditReport} from a scan report — pure, no clock.
|
|
106
158
|
* The CLI attaches `meta.generatedAt` when it writes the JSON artifact; the
|
|
@@ -112,19 +164,13 @@ report, opts) {
|
|
|
112
164
|
const adoptable = buildAdoptable(opts.adoptableSurfaces);
|
|
113
165
|
const { score, verdict, recommendations } = gradedPart(report, opts);
|
|
114
166
|
return {
|
|
115
|
-
meta:
|
|
116
|
-
schemaVersion: exports.AUDIT_SCHEMA_VERSION,
|
|
117
|
-
tool: "vigiles",
|
|
118
|
-
kind: "audit",
|
|
119
|
-
vigilesVersion: opts.vigilesVersion,
|
|
120
|
-
harness: opts.harness,
|
|
121
|
-
dir: report.dir,
|
|
122
|
-
},
|
|
167
|
+
meta: auditMeta(report.dir, opts),
|
|
123
168
|
score,
|
|
124
169
|
verdict,
|
|
125
170
|
recommendations,
|
|
126
171
|
inventory: inventoryOf(report),
|
|
127
172
|
...scanFindingLists(report),
|
|
173
|
+
...contextCostField(report),
|
|
128
174
|
...(adoptable ? { adoptable } : {}),
|
|
129
175
|
...(opts.observations ? { observations: opts.observations } : {}),
|
|
130
176
|
...(opts.rulesInventory && opts.rulesInventory.length
|