vigiles 29.1.0 → 30.0.0

package/dist/adapters/claude-code/dialect.js

@@ -69,22 +69,24 @@ exports.claudeCodeDialect = {
     // budget here costs money and attention — not rules. (Codex is the opposite;
     // see its dialect, and see why `onExceed` is reported at all.)
     //
-    // 🔴 `alwaysLoaded` IS THE POINT, and `.claude/rules/**` is in it on a
-    // MEASUREMENT, not a doc: a consumer repo moved 225 837 characters out of
-    // CLAUDE.md into that directory and the request cost did not move, because
-    // the harness loads it either way. A per-file check would have called that
-    // split a success.
+    // 🔴 WHICH FILES THE SUM IS TAKEN OVER IS NO LONGER A FIELD HERE. It was
+    // `alwaysLoaded: ["CLAUDE.md", "CLAUDE.local.md", ".claude/CLAUDE.md",
+    // ".claude/rules/**"]` — globs this dialect wrote and the CORE expanded by
+    // walking the repository, which is the defect zernie/vigiles#262 records.
+    // The set now comes from `claudeCodeLayout.instructionChain`, which can say
+    // what a glob cannot: that a `paths:`-scoped rule loads on demand, and that
+    // `CLAUDE.local.md` is a per-machine file that is read and never SCORED.
+    //
+    // What stays here is the harness's own NUMBER, which is a format fact: the
+    // rules directory counts on a MEASUREMENT, not a doc — a consumer repo moved
+    // 225 837 characters out of CLAUDE.md into it and the request cost did not
+    // move, because the harness loads it either way. A per-file check would have
+    // called that split a success.
     instructionBudget: {
         unit: "chars",
         limit: 40000,
         onExceed: "warns",
-        capturedFrom: "claude-code /doctor large-file warning; .claude/rules/** measured 2026-09-16 in a consumer repo",
-        alwaysLoaded: [
-            "CLAUDE.md",
-            "CLAUDE.local.md",
-            ".claude/CLAUDE.md",
-            ".claude/rules/**",
-        ],
+        capturedFrom: "claude-code /doctor large-file warning; the rules dir measured 2026-09-16 in a consumer repo",
     },
     // The capability table — what each event CARRIES and HONOURS. Nine of the 31
     // events, each with its basis; see ./event-capability.ts. The three flat lists
@@ -95,16 +97,80 @@ exports.claudeCodeDialect = {
     // `hookSpecificOutput.permissionDecision:"deny"`; the legacy top-level
     // `decision` field is ignored there.
     permissionDecisionHookEvents: ["PreToolUse"],
-    // Claude Code natively reads CLAUDE.md only — it does NOT auto-load AGENTS.md
-    // (anthropics/claude-code#34235 is open; AGENTS.md works solely via an
-    // `@AGENTS.md` import inside CLAUDE.md or a symlink). AGENTS.md is the
-    // cross-tool standard (Codex's native target), not a CC dialect fact; vigiles's
-    // tool-agnostic recognition of it lives in validate.ts's INSTRUCTION_FILES.
-    instructionTargets: ["CLAUDE.md"],
+    // 🔴 THIS LIST GAINED `AGENTS.md` ON 2026-09-21, REVERSING WHAT STOOD HERE.
+    // The comment it replaces read "Claude Code natively reads CLAUDE.md only — it
+    // does NOT auto-load AGENTS.md (anthropics/claude-code#34235 is open; AGENTS.md
+    // works solely via an `@AGENTS.md` import inside CLAUDE.md or a symlink)".
+    // That is no longer true. Vendor, `https://code.claude.com/docs/en/memory`,
+    // read 2026-09-21:
+    //
+    //   "Claude Code can read `AGENTS.md` as your project instructions, so a
+    //    repository already set up for other coding agents works without adding a
+    //    `CLAUDE.md`, an import, or a setting."
+    //
+    //   "Reading `AGENTS.md` directly requires Claude Code v2.1.277 or later."
+    //
+    // The conditions under which it actually loads are NOT a dialect fact and are
+    // not restated here — they depend on sibling files and on settings, which is
+    // exactly why they live in `./instruction-chain.ts` as a method. This field
+    // answers the narrower question its docblock asks: which filenames does this
+    // harness READ as project instructions.
+    //
+    // 🔴 WHAT CHANGED WHEN `AGENTS.md` WAS ADDED — MEASURED, NOT EXPECTED, AND THE
+    // EXPECTED ANSWER WAS WRONG. The obvious worry is `detect`: a second target
+    // ought to make this adapter score a bare `AGENTS.md` repository and start
+    // fighting Codex for it. It does not, and cannot, because `detect` never reads
+    // this field — it asks the LAYOUT (`claudeCodeLayout.instructionFile`).
+    // Recorded by instrumenting `adapter.detect` with a call-recording predicate,
+    // before and after the edit:
+    //
+    //     detect asked: [".claude-plugin/plugin.json", ".claude/settings.json",
+    //                    "CLAUDE.md"]        — IDENTICAL both ways
+    //     claims("AGENTS.md"): false         — IDENTICAL both ways
+    //
+    // That separation is load-bearing rather than incidental: `claims` is derived
+    // from `layoutLocations`, and `adapter-properties.test.ts` refuses a `detect`
+    // that asks about a path `claims` does not cover. So a Claude Code adapter
+    // that detected on `AGENTS.md` would have to CLAIM it, and two registered
+    // adapters claiming one path is the collision `claims` exists to prevent. The
+    // file stays owned by Codex and merely READ here — which is the same reason
+    // `instruction-chain.ts` keeps `AGENTS.md` as a local constant instead of
+    // putting it on `PluginLayout`.
+    //
+    // THE ONE CONSUMER THAT DID CHANGE is `core/validate.ts:300`
+    // (`recognized = dialect?.instructionTargets ?? INSTRUCTION_FILES`), and the
+    // change fixes a real inconsistency. Measured on an `AGENTS.md` path:
+    //
+    //     before — with the CC dialect injected: []
+    //     before — with NO dialect:              ["require-instructions-spec"]
+    //     after  — either way:                   ["require-instructions-spec"]
+    //
+    // i.e. injecting this dialect used to make vigiles recognise FEWER instruction
+    // files than its own no-dialect default, so an `AGENTS.md` that Claude Code
+    // really does read was linted as if it were an ordinary markdown file. Full
+    // unit suite with this edit alone: 4 failed / 4185 passed — the same four
+    // `spec.test.ts` pylint/rubocop CLI tests that fail on this branch regardless.
+    //
+    // ORDER IS PART OF THE CONTRACT: `instructionTargets[0]` is the default
+    // COMPILE target (`core/compile.ts:792`), and what `vigiles init` writes for
+    // Claude Code is still `CLAUDE.md`. `AGENTS.md` is second because this harness
+    // reads it, not because it authors it. `docs/adapter-api.md` already documented
+    // `["CLAUDE.md","AGENTS.md"]` for this adapter; the code now agrees with it.
+    instructionTargets: ["CLAUDE.md", "AGENTS.md"],
     pluginRootToken: "${CLAUDE_PLUGIN_ROOT}",
-    // Claude Code reads the full SKILL.md frontmatter set (description,
-    // disable-model-invocation, argument-hint, …).
-    skillFrontmatter: "claude-code",
+    // Claude Code reads the full SKILL.md frontmatter set. Spelled out rather than
+    // aliased to the compiler's RENDERABLE_SKILL_FRONTMATTER_KEYS: that constant is
+    // "what vigiles can write", this list is "what Claude Code reads", and a new
+    // renderable key must not silently become a claim about the vendor.
+    skillFrontmatterKeys: [
+        "name",
+        "description",
+        "disable-model-invocation",
+        "context",
+        "argument-hint",
+        "allowed-tools",
+        "disallowed-tools",
+    ],
     // Tools that produce side effects in Claude Code. The complement — the
     // read-only tools — are: Read, Grep, Glob, ToolSearch, LSP, ListAgents,
     // TaskGet, TaskList, CronList. Bash (and PowerShell) are side-effecting

package/dist/adapters/claude-code/hook-protocol.js

@@ -2,6 +2,7 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.claudeCodeHookProtocol = void 0;
 const hook_condition_js_1 = require("./hook-condition.js");
+const hook_install_js_1 = require("../../hook-install.js");
 exports.claudeCodeHookProtocol = {
     name: "claude-code",
     blockExitCode: 2,
@@ -41,5 +42,20 @@ exports.claudeCodeHookProtocol = {
     // at all. See ./hook-condition.ts — without it a conditional guard was reported
     // as blocking every disaster in the battery.
     condition: hook_condition_js_1.claudeCodeHookCondition,
+    // Claude Code NESTS: one matcher block holds a list of commands.
+    registration(on, matcher, command) {
+        const entry = matcher === undefined
+            ? { hooks: [{ type: "command", command }] }
+            : { matcher, hooks: [{ type: "command", command }] };
+        return { hooks: { [on]: [entry] } };
+    },
+    // Delegates to the JSON merge, which is where the measured behaviour and its
+    // test suite live (`hook-install.ts` / `hook-install.test.ts`). The port owns
+    // WHICH merge; the module owns HOW. The cast is the structural-to-concrete
+    // step the core cannot take: `CompiledHooks` is the application layer's type,
+    // and a core port may not name it.
+    mergeRegistrations(existing, compiled, managedBy) {
+        return (0, hook_install_js_1.mergeHooksJson)(existing, compiled, managedBy);
+    },
 };
 //# sourceMappingURL=hook-protocol.js.map

package/dist/adapters/claude-code/instruction-chain.d.ts

@@ -0,0 +1,25 @@
+import type { InstructionChain, SettingsSource } from "../../core/instruction-chain.js";
+/** Inputs the layout supplies so this module names no path of its own. */
+export interface ClaudeCodeChainInput {
+    /** `CLAUDE.md`. */
+    readonly instructionFile: string;
+    /** `.claude`. */
+    readonly userSurfaceRoot: string;
+    /**
+     * The settings files, in precedence order (repo, then the local sibling),
+     * each with its SCOPE — a pattern out of the gitignored sibling may not
+     * change what a teammate on this commit is scored for.
+     */
+    readonly settingsSources: readonly SettingsSource[];
+    /** The layout's own codec — this module does not know the encoding. */
+    readonly parseSettings: (text: string) => Record<string, unknown>;
+    /** {@link RULE_FILE_LEAF_RE}, compiled against the rules dir. */
+    readonly ruleRe: RegExp;
+}
+/**
+ * `CLAUDE.md` → `CLAUDE.local.md`: the per-machine sibling, DERIVED so the two
+ * names cannot drift apart the way a second constant would.
+ */
+export declare function localSiblingOf(instructionFile: string): string;
+export declare function claudeCodeInstructionChain(files: Readonly<Record<string, string>>, input: ClaudeCodeChainInput): InstructionChain;
+//# sourceMappingURL=instruction-chain.d.ts.map