rulereceipt 0.1.48 → 0.1.50

package/README.md

@@ -393,6 +393,34 @@ report to the path you name, and `rules --include/--exclude` records a
 correction in `.rulereceipt/overrides.json`. Plain `rulereceipt check` writes
 nothing and makes no network calls.
 
+**Severity, per rule.** A committed, team-shared `.rulereceipt/config.json`
+sets how hard each rule bites in CI, by its stable handle (from `rulereceipt
+rules --list`):
+
+```json
+{
+  "rules": {
+    "a1b2c3": "off",     // hidden from the report, never gates
+    "d4e5f6": "warn",    // shown, but does not fail the build
+    "97h8i9": "error"    // shown, FAILS the build — the default for a checkable rule
+  },
+  "checks": {
+    "emoji": "off",      // silence a whole check type by name
+    "git": "warn"        // emoji, attribution, approval, git, files, code, claim, tests, judgment
+  }
+}
+```
+
+A per-rule `rules` entry wins over a per-check `checks` entry, which wins over
+the default. No config means today's behaviour: every checkable FAIL is an
+`error`. This is
+the one place severity lives — a team marks the must-not-break rules `error`
+and the nice-to-haves `warn`, so CI gates on what matters instead of going red
+on day one. (Refusing a command *before* it runs is separate, and stays with
+the guard's `rules --forbid` clause-mark — a config that could block on any
+rule would refuse far too much.) The older `{"warn": ["a1b2c3"]}` list still
+works and means the same as `"warn"` above.
+
 **You can verify the package came from this source.** Every release from
 0.1.19 on is built and published by GitHub Actions and signed with
 [npm provenance](https://docs.npmjs.com/generating-provenance-statements),

package/dist/cli.js

@@ -6,7 +6,7 @@ import { join, dirname, resolve, isAbsolute } from "node:path";
 import { existsSync, readFileSync, writeFileSync } from "node:fs";
 import { fileURLToPath } from "node:url";
 import { parseClaudeMd } from "./parsers/readClaudeMd.js";
-import { readLatestTranscript, readTranscriptFromFile, findLatestSessionFile } from "./parsers/transcriptParser.js";
+import { readLatestTranscript, readTranscriptFromFile, findLatestSessionFile, subagentNote } from "./parsers/transcriptParser.js";
 import { loadRules } from "./rules.js";
 import { adviseRules } from "./checkability.js";
 import { shadowedAgentsMd } from "./shadowedAgents.js";
@@ -33,7 +33,7 @@ import { maybeShowWhatsNew } from "./whatsNew.js";
 import { verifyReceipt, parseReceipt } from "./receipt.js";
 import { buildBadge } from "./badge.js";
 import { buildInitGuidance } from "./init.js";
-import { loadProjectConfig, handleMap, blockingFailures, warningFailures, PROJECT_CONFIG_PATH } from "./projectConfig.js";
+import { loadProjectConfig, handleMap, blockingFailures, warningFailures, visibleResults, PROJECT_CONFIG_PATH } from "./projectConfig.js";
 import { maybeCheckUpdates, isUpdateCheckEnabled } from "./updateCheck.js";
 import { generateDigest } from "./digest.js";
 import { enableSchedule, disableSchedule, scheduleStatus } from "./schedule.js";
@@ -264,12 +264,14 @@ async function runCheck(opts) {
     // sends only a random install ID, never rule text or transcript content,
     // regardless of --llm.
     const judgmentResults = llm ? await runJudgmentChecks(judgment, events) : judgment.map(({ rule }) => needsLlmResult(rule));
-    const results = [...deterministicResults, ...judgmentResults];
-    // Severity: rules a team marked as warnings in .rulereceipt/config.json are
-    // still reported but do not fail the build. handleFor maps a result back to
-    // its stable handle so the mark survives edits that renumber rule ids.
+    const rawResults = [...deterministicResults, ...judgmentResults];
+    // Severity ladder from .rulereceipt/config.json (per rule handle): `off`
+    // rules are hidden entirely, `warn` rules are shown but do not fail the
+    // build, everything else is `error` (the default). handleFor maps a result
+    // back to its stable handle so the mark survives edits that renumber ids.
     const projectConfig = loadProjectConfig(cwd);
     const handleFor = handleMap(rules);
+    const results = visibleResults(rawResults, projectConfig, handleFor);
     const blockingFails = blockingFailures(results, projectConfig, handleFor);
     const warnedFails = warningFailures(results, projectConfig, handleFor);
     const meta = { sessionFilePath, ruleCount: results.length };
@@ -281,6 +283,9 @@ async function runCheck(opts) {
     }
     else {
         console.log(reportText);
+        const subNote = subagentNote(sessionFilePath);
+        if (subNote)
+            console.log(`\n${subNote}`);
     }
     // Shown only to someone who has just read their own broken rules, and only
     // if they have not already wired it up. See report/gateOffer.ts.

package/dist/guard.d.ts

@@ -1,3 +1,8 @@
+import type { Rule } from "./types.js";
+interface Block {
+    rule: Rule;
+    why: string;
+}
 /**
  * A PreToolUse hook that refuses a command before it runs.
  *
@@ -39,4 +44,27 @@
  * 3. It says which rule and why, in the refusal itself, because a block with
  *    no reason is indistinguishable from a broken tool.
  */
+export interface GuardDecision {
+    /** True when the proposed call breaks a rule and should be refused. */
+    deny: boolean;
+    /** The message for the model, or "" when allowed. */
+    reason: string;
+    /** The rules that would refuse it, empty when allowed. */
+    blocks: Block[];
+}
+/**
+ * The allow/deny decision for one proposed tool call, with no I/O.
+ *
+ * Extracted from runGuard 2026-09-25 so the decision is a pure function the
+ * shell hook calls today and any future host — an in-process function hook,
+ * a different runtime — calls tomorrow without re-deriving the logic. The
+ * function-hook interface is still an unshipped Anthropic proposal (#91870),
+ * so nothing here targets it; this is only the seam that keeps the adapter
+ * thin when it lands. Same reasoning as evaluateSession: one body of code so
+ * two callers can never disagree about whether a rule was broken.
+ */
+export declare function guardDecision(cwd: string, toolName: string, toolInput: {
+    command?: unknown;
+} & Record<string, unknown>): GuardDecision;
 export declare function runGuard(): Promise<void>;
+export {};

package/dist/guard.js

@@ -143,46 +143,42 @@ function reason(blocks) {
         `\n\nIf the rule should not apply here, say so to the user and let them decide. Do not work around the rule by rephrasing the command.`);
 }
 /**
- * A PreToolUse hook that refuses a command before it runs.
- *
- * The Stop hook added on 2026-09-14 catches a finished session. That is too
- * late for the case it most needs to cover: on 2026-04-25 a Cursor agent
- * running Claude Opus 4.6 deleted PocketOS's production database and every
- * volume-level backup in nine seconds, using a Railway token it found that
- * had been created for managing domains. The agent had a rule — "NEVER run
- * destructive/irreversible git commands...unless the user explicitly
- * requests them" — and afterwards quoted it back, observing that what it had
- * done was "far worse than a force push". A report would have described a
- * database that was already gone.
- *
- * What it does NOT do is the thing it was built for. Blocking on a rule's
- * banned command literal was measured before shipping and cut: see the note
- * above `reason`. It refused 62% of 16,336 real commands, and the residue
- * after two rounds of narrowing was still wrong in a way no matcher fixes.
- * PocketOS would not have been stopped by this hook, and saying otherwise
- * would be the exact failure this tool exists to catch.
- *
- * What remains is real and narrower: rules that name a FILE or a BRANCH.
- * "Never modify `.env`", "never touch `migrations/`", "never commit to
- * `main`". The classifier identified those as a path or a ref rather than
- * guessing which backtick was the prohibition, so a refusal can be stated
- * with a reason that holds up.
- *
- * Three properties, in the order they matter:
- *
- * 1. FORBIDDING rules only, answered by a structured checker. Never a
- *    judgment rule, never an LLM opinion, never a requirement, and never a
- *    bare command literal. Blocking someone's terminal on a guess is not a
- *    trade worth making at any hit rate.
- *
- * 2. It fails OPEN. Any error allows the command and writes to stderr. The
- *    opposite choice means a bug in this file stops someone from running
- *    anything at all, and they would remove the hook within the hour — which
- *    leaves them with no guard rather than an imperfect one.
- *
- * 3. It says which rule and why, in the refusal itself, because a block with
- *    no reason is indistinguishable from a broken tool.
+ * The allow/deny decision for one proposed tool call, with no I/O.
+ *
+ * Extracted from runGuard 2026-09-25 so the decision is a pure function the
+ * shell hook calls today and any future host — an in-process function hook,
+ * a different runtime — calls tomorrow without re-deriving the logic. The
+ * function-hook interface is still an unshipped Anthropic proposal (#91870),
+ * so nothing here targets it; this is only the seam that keeps the adapter
+ * thin when it lands. Same reasoning as evaluateSession: one body of code so
+ * two callers can never disagree about whether a rule was broken.
  */
+export function guardDecision(cwd, toolName, toolInput) {
+    const allow = { deny: false, reason: "", blocks: [] };
+    if (loadRules(cwd).length === 0)
+        return allow;
+    let blocks = [];
+    if (toolName === "Bash" && typeof toolInput.command === "string") {
+        const event = {
+            role: "assistant", kind: "tool_use", toolName: "Bash",
+            input: { command: toolInput.command }, timestamp: "",
+        };
+        blocks = [...structuredBlocks(cwd, event), ...ratifiedLiteralBlocks(cwd, toolInput.command)];
+    }
+    else if (toolName === "Write" || toolName === "Edit" || toolName === "NotebookEdit") {
+        const event = {
+            role: "assistant", kind: "tool_use", toolName,
+            input: toolInput, timestamp: "",
+        };
+        blocks = structuredBlocks(cwd, event);
+    }
+    else {
+        return allow;
+    }
+    if (blocks.length === 0)
+        return allow;
+    return { deny: true, reason: reason(blocks), blocks };
+}
 export async function runGuard() {
     const allow = () => {
         process.stdout.write(JSON.stringify({}));
@@ -193,29 +189,10 @@ export async function runGuard() {
         const cwd = input.cwd || process.cwd();
         const tool = input.tool_name ?? "";
         const toolInput = input.tool_input ?? {};
-        if (loadRules(cwd).length === 0)
-            return allow();
-        let blocks = [];
-        if (tool === "Bash" && typeof toolInput.command === "string") {
-            const event = {
-                role: "assistant", kind: "tool_use", toolName: "Bash",
-                input: { command: toolInput.command }, timestamp: "",
-            };
-            blocks = [...structuredBlocks(cwd, event), ...ratifiedLiteralBlocks(cwd, toolInput.command)];
-        }
-        else if (tool === "Write" || tool === "Edit" || tool === "NotebookEdit") {
-            const event = {
-                role: "assistant", kind: "tool_use", toolName: tool,
-                input: toolInput, timestamp: "",
-            };
-            blocks = structuredBlocks(cwd, event);
-        }
-        else {
-            return allow();
-        }
-        if (blocks.length === 0)
+        const decision = guardDecision(cwd, tool, toolInput);
+        if (!decision.deny)
             return allow();
-        const why = reason(blocks);
+        const why = decision.reason;
         process.stdout.write(JSON.stringify({
             hookSpecificOutput: {
                 hookEventName: "PreToolUse",

package/dist/parsers/transcriptParser.d.ts

@@ -36,4 +36,10 @@ export declare function readTranscriptFromFile(filePath: string): TranscriptEven
  * false negative — the safe direction for a tool that must not over-accuse.
  */
 export declare function findSubagentFiles(sessionFile: string): string[];
+/**
+ * A one-line note for the report so a user knows their subagents were included
+ * in the check — otherwise the coverage is invisible and they might think a
+ * violation a subagent committed went unseen. Null when there were none.
+ */
+export declare function subagentNote(sessionFile: string | null): string | null;
 export declare function readLatestTranscript(cwd: string): TranscriptEvent[];

package/dist/parsers/transcriptParser.js

@@ -197,6 +197,19 @@ export function findSubagentFiles(sessionFile) {
     const subagentDir = join(dirname(sessionFile), sessionId, "subagents");
     return listSessionFiles(subagentDir);
 }
+/**
+ * A one-line note for the report so a user knows their subagents were included
+ * in the check — otherwise the coverage is invisible and they might think a
+ * violation a subagent committed went unseen. Null when there were none.
+ */
+export function subagentNote(sessionFile) {
+    if (!sessionFile)
+        return null;
+    const n = findSubagentFiles(sessionFile).length;
+    if (n === 0)
+        return null;
+    return `Checked ${n} subagent session${n === 1 ? "" : "s"} alongside the main one — their actions are held to the same rules.`;
+}
 export function readLatestTranscript(cwd) {
     const filePath = findLatestSessionFile(cwd);
     if (!filePath)

package/dist/projectConfig.d.ts

@@ -1,25 +1,59 @@
-import type { CheckResult, Rule } from "./types.js";
+import type { CheckResult, CheckMethod, Rule } from "./types.js";
 /**
  * A committed, team-shared config at .rulereceipt/config.json.
  *
- * Today it holds one thing: rule handles to treat as WARNINGS — a broken
- * "warning" rule is still reported, but it does not fail the build. This is
- * the honest version of "severity": a team marks the must-not-break rules as
- * errors (the default) and the nice-to-have ones as warnings, so CI gates on
- * what actually matters instead of going red on day one.
+ * ONE place for "how hard should this rule bite", a ladder per rule handle:
+ *
+ *   off    hidden from the report, never gates CI
+ *   warn   shown, does not fail the build
+ *   error  shown, FAILS the build (the default for a checkable rule)
+ *
+ * This is the honest version of "severity": a team marks must-not-break rules
+ * as errors (the default) and nice-to-haves as warnings, and silences the
+ * irrelevant ones, so CI gates on what actually matters instead of going red
+ * on day one. It replaces three scattered mechanisms — the old `warn` list,
+ * `rules --exclude` (now `off`), and the plain default — with one field.
+ *
+ * Pre-run BLOCKING is deliberately NOT a mode here: refusing a command before
+ * it runs still goes through the guard's clause-mark (`rules --forbid`),
+ * because a config that could block on any rule would refuse the 62.8% of
+ * commands the measured guard already showed it must not. This file governs
+ * the report and the CI gate; the guard governs refusal.
  *
  * Handles, not rule ids: an id is positional and renumbers when the file is
  * edited above it; a handle is a content hash, so it survives edits. Get one
  * from `rulereceipt rules --list`.
+ *
+ * Backward compatible: the old top-level `warn: [handle, ...]` list still
+ * works and means the same as `rules: { <handle>: "warn" }`.
  */
+export type RuleMode = "off" | "warn" | "error";
 export interface ProjectConfig {
     warn: string[];
+    rules: Record<string, RuleMode>;
+    /** A whole check type, by friendly name (see CHECK_ALIASES) or raw method. */
+    checks: Record<string, RuleMode>;
 }
+/**
+ * Friendly names for a check type, so `checks: { "emoji": "off" }` silences
+ * every emoji verdict without listing each rule. A result carries a `method`
+ * (e.g. "emoji_output"); these are the human-facing aliases for those.
+ */
+export declare const CHECK_ALIASES: Record<string, CheckMethod>;
 export declare const PROJECT_CONFIG_PATH: string;
 export declare function loadProjectConfig(cwd: string): ProjectConfig;
+/**
+ * The mode for one result. Precedence: a per-rule `rules` entry wins over a
+ * per-check `checks` entry, which wins over the legacy `warn` list; anything
+ * unlisted is `error`, so the default is unchanged and no config means today's
+ * behaviour exactly.
+ */
+export declare function modeForResult(result: CheckResult, config: ProjectConfig, handleFor: (r: CheckResult) => string): RuleMode;
+/** Results the report should show — everything except rules set to `off`. */
+export declare function visibleResults(results: CheckResult[], config: ProjectConfig, handleFor: (r: CheckResult) => string): CheckResult[];
 /** A lookup from a result back to its stable rule handle, built from the loaded rules. */
 export declare function handleMap(rules: Rule[]): (r: CheckResult) => string;
-/** FAILs that are NOT configured as warnings — these fail the build. */
+/** FAILs at `error` mode — these fail the build. (`off` never reaches here.) */
 export declare function blockingFailures(results: CheckResult[], config: ProjectConfig, handleFor: (r: CheckResult) => string): CheckResult[];
-/** FAILs that ARE configured as warnings — shown, but they do not fail the build. */
+/** FAILs at `warn` mode — shown, but they do not fail the build. */
 export declare function warningFailures(results: CheckResult[], config: ProjectConfig, handleFor: (r: CheckResult) => string): CheckResult[];

package/dist/projectConfig.js

@@ -1,19 +1,81 @@
 import { readFileSync } from "node:fs";
 import { join } from "node:path";
 import { ruleFingerprint } from "./overrides.js";
+const VALID_MODES = ["off", "warn", "error"];
+/**
+ * Friendly names for a check type, so `checks: { "emoji": "off" }` silences
+ * every emoji verdict without listing each rule. A result carries a `method`
+ * (e.g. "emoji_output"); these are the human-facing aliases for those.
+ */
+export const CHECK_ALIASES = {
+    emoji: "emoji_output",
+    attribution: "attribution_scan",
+    approval: "approval_gate",
+    git: "git_events",
+    files: "file_events",
+    code: "code_content",
+    claim: "claim_vs_evidence",
+    tests: "edit_test_pairing",
+    text: "text_scan",
+    judgment: "model_judgment",
+};
 export const PROJECT_CONFIG_PATH = join(".rulereceipt", "config.json");
 export function loadProjectConfig(cwd) {
     try {
         const parsed = JSON.parse(readFileSync(join(cwd, PROJECT_CONFIG_PATH), "utf-8"));
-        const warn = parsed?.warn;
-        return { warn: Array.isArray(warn) ? warn.filter((x) => typeof x === "string") : [] };
+        const warn = Array.isArray(parsed?.warn) ? parsed.warn.filter((x) => typeof x === "string") : [];
+        const readModes = (raw) => {
+            const out = {};
+            if (raw && typeof raw === "object" && !Array.isArray(raw)) {
+                for (const [key, mode] of Object.entries(raw)) {
+                    if (typeof mode === "string" && VALID_MODES.includes(mode))
+                        out[key] = mode;
+                }
+            }
+            return out;
+        };
+        return { warn, rules: readModes(parsed?.rules), checks: readModes(parsed?.checks) };
     }
     catch {
         // Missing or malformed config means no severities configured, never an
         // error — same fail-open discipline as the rest of the tool.
-        return { warn: [] };
+        return { warn: [], rules: {}, checks: {} };
     }
 }
+/** The mode a `checks` entry sets for a result's method, if any. */
+function checkMode(method, config) {
+    if (!method || !config.checks || Object.keys(config.checks).length === 0)
+        return undefined;
+    // A raw method key ("emoji_output") wins over its friendly alias ("emoji").
+    if (config.checks[method])
+        return config.checks[method];
+    for (const [alias, m] of Object.entries(CHECK_ALIASES)) {
+        if (m === method && config.checks[alias])
+            return config.checks[alias];
+    }
+    return undefined;
+}
+/**
+ * The mode for one result. Precedence: a per-rule `rules` entry wins over a
+ * per-check `checks` entry, which wins over the legacy `warn` list; anything
+ * unlisted is `error`, so the default is unchanged and no config means today's
+ * behaviour exactly.
+ */
+export function modeForResult(result, config, handleFor) {
+    const handle = handleFor(result);
+    if (config.rules[handle])
+        return config.rules[handle];
+    const byCheck = checkMode(result.method, config);
+    if (byCheck)
+        return byCheck;
+    if (config.warn.includes(handle))
+        return "warn";
+    return "error";
+}
+/** Results the report should show — everything except rules set to `off`. */
+export function visibleResults(results, config, handleFor) {
+    return results.filter((r) => modeForResult(r, config, handleFor) !== "off");
+}
 /** A lookup from a result back to its stable rule handle, built from the loaded rules. */
 export function handleMap(rules) {
     const m = new Map();
@@ -21,11 +83,11 @@ export function handleMap(rules) {
         m.set(`${rule.source}:${rule.id}`, ruleFingerprint(rule));
     return (r) => m.get(`${r.ruleSource}:${r.ruleId}`) ?? "";
 }
-/** FAILs that are NOT configured as warnings — these fail the build. */
+/** FAILs at `error` mode — these fail the build. (`off` never reaches here.) */
 export function blockingFailures(results, config, handleFor) {
-    return results.filter((r) => r.status === "FAIL" && !config.warn.includes(handleFor(r)));
+    return results.filter((r) => r.status === "FAIL" && modeForResult(r, config, handleFor) === "error");
 }
-/** FAILs that ARE configured as warnings — shown, but they do not fail the build. */
+/** FAILs at `warn` mode — shown, but they do not fail the build. */
 export function warningFailures(results, config, handleFor) {
-    return results.filter((r) => r.status === "FAIL" && config.warn.includes(handleFor(r)));
+    return results.filter((r) => r.status === "FAIL" && modeForResult(r, config, handleFor) === "warn");
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rulereceipt",
-  "version": "0.1.48",
+  "version": "0.1.50",
   "description": "Checks whether a Claude Code session actually followed your CLAUDE.md / AGENTS.md rules, with evidence.",
   "repository": {
     "type": "git",