akm-cli 0.9.17-alpha.3 → 0.9.17-alpha.4

package/dist/commands/proposal/validators/proposal-quality-validators.js

@@ -2,77 +2,27 @@
 // License, v. 2.0. If a copy of the MPL was not distributed with this
 // file, You can obtain one at https://mozilla.org/MPL/2.0/.
 /**
- * Shared content-quality validators consumed by the improve pipeline
- * (`distill`, `consolidate`, `reflect`) and by the `proposal accept` gate.
+ * Content-quality validators for the improve stages and `proposal accept`.
  *
- * ## Reflect size gate — calibrated blended formula (2026-05-22)
- *
- * ### Distribution baseline (n=844 reflect-eligible stash assets)
- *
- *   min=1  p10=371  p25=778  p50=1508  p75=5456  p90=11721  p99=43463  max=298010 bytes
- *   Buckets: <500=135 (16%), 500–2000=340 (40%), 2000–8000=222 (26%), >8000=147 (17%)
- *
- * ### Problem with the original fixed-ratio gate
- *
- *   - Small sources (~420 bytes, 16th pct): the 200% expansion ceiling fires at
- *     only 840 bytes proposed — one good paragraph.  Hair-trigger for a terse
- *     reference note.
- *   - Large sources (~7KB, 75th pct): 200% ceiling = 14KB; reasonable, but a hard
- *     cap prevents runaway expansion from LLM hallucinations.
- *
- * ### Blended-bound formula
- *
- *   Shrinkage floor (accept if proposed >= lower):
- *     lower = max(REFLECT_SHRINK_RATIO_MIN * sourceLen, REFLECT_ABSOLUTE_FLOOR_BYTES)
- *     → For tiny sources (sourceLen < 300), the absolute floor dominates so a
- *       genuinely tightened note still passes.
- *     → For large sources (>1KB), the ratio floor dominates (50% of 7KB = 3.5KB).
- *
- *   Expansion ceiling (accept if proposed <= upper):
- *     upper = max(REFLECT_EXPAND_RATIO_MAX * sourceLen, REFLECT_ABSOLUTE_CEILING_BYTES)
- *     …but always capped at REFLECT_ABSOLUTE_MAX_BYTES.
- *     → For small sources (≤778 bytes, p25), the absolute ceiling (2000 bytes)
- *       dominates — one substantive paragraph is always acceptable.
- *     → For medium/large sources (>1KB), the ratio ceiling dominates.
- *     → Any proposal exceeding 25000 bytes is always rejected regardless of ratio.
- *
- * ### Constant calibration rationale
- *
- *   REFLECT_ABSOLUTE_FLOOR_BYTES = 150
- *     Half of p10 (371) ≈ 185; we set 150 so even very aggressive condensation
- *     of a seed note is allowed down to roughly a two-sentence summary.
- *
- *   REFLECT_ABSOLUTE_CEILING_BYTES = 2500
- *     Raised from 2000 (2026-05-22): small-source rejections at 248–281% on
- *     900–953 byte assets were borderline false positives. 2500 gives a short
- *     lesson or command ~1.5KB of room to grow before the absolute kicks in.
- *
- *   REFLECT_ABSOLUTE_MAX_BYTES = 25000
- *     Below p99 (43463). Catches genuine LLM runaway (whole-chapter insertions)
- *     without blocking legitimate large rewrites of large sources.
- *
- *   REFLECT_EXPAND_RATIO_MAX = 2.5
- *     Raised from 2.0 (2026-05-22): 2× was too tight for dense short assets
- *     (lessons, commands) that have legitimate room to grow. 2.5× resolves
- *     248% expansion on a 900-byte lesson while still catching 281%+ on ~1KB
- *     assets where the absolute ceiling takes over.
+ * The reflect size gate bounds a rewrite's body against its source's:
+ * shrinking below max(50% of the source, 150 bytes) suggests deleted content,
+ * growing past min(max(250% of the source, 2500 bytes), 25000 bytes) suggests
+ * speculation. The absolute bounds keep small assets (a p25 source is ~780
+ * bytes) from tripping on one good paragraph, and 25000 (below p99) still
+ * catches runaway expansion. Sources under 200 bytes are too noisy to judge.
  */
-// ── Reflect-size guard ───────────────────────────────────────────────────────
 import { parseFrontmatter } from "../../../core/asset/frontmatter.js";
 import { parseRefInput } from "../../../core/asset/resolve-ref.js";
 import { DESCRIPTION_MAX_CHARS, DESCRIPTION_MIN_CHARS, WHEN_TO_USE_MAX_CHARS, WHEN_TO_USE_MIN_CHARS, } from "../../../core/authoring-rules.js";
 import { containsRedactedContent, containsReflectPromptScaffolding, REDACTED_CONTENT_MARKER, REFLECT_AVOID_PATTERNS_HEADING, } from "../../../core/content-safety.js";
 import { proposalContent } from "../../../core/file-change.js";
-/**
- * The canonical asset NAME an inputRef names, lower-cased — the tail the
- * "just restates the ref" heuristics compare against. WI-8.5c: the ref is the
- * conceptId (`<subdir>/<name>`), so the name is parsed off the conceptId.
- */
+import { detectTruncatedDescription, TRUNCATION_TRAILING_WORDS } from "../../../core/text-truncation.js";
+import { REFLECT_TRUNCATION_MARKER } from "../../../integrations/agent/prompts.js";
+import { splitFrontmatter } from "../../improve/reflect-noise.js";
+/** The asset name a ref names, lower-cased — what the "just restates the ref" checks compare. */
 function refNameTail(inputRef) {
     return parseRefInput(inputRef).name.toLowerCase();
 }
-import { detectTruncatedDescription, TRUNCATION_TRAILING_WORDS } from "../../../core/text-truncation.js";
-import { REFLECT_TRUNCATION_MARKER } from "../../../integrations/agent/prompts.js";
 // ── Description / when_to_use shape ─────────────────────────────────────────
 export const HEADING_FRAGMENT_PATTERNS = [
     /^for example\b/i,
@@ -164,9 +114,8 @@ export function detectDoubleFrontmatter(content) {
             kind: "double-frontmatter-fence",
             message: `Content contains ${fenceLines.length} \`---\` fence lines; assets with frontmatter must have exactly 2 (one open, one close).`,
         };
-    const body = content.replace(/^---\r?\n[\s\S]*?\r?\n---\r?\n?/, "");
-    const pseudoLine = body
-        .split(/\r?\n/)
+    const pseudoLine = splitFrontmatter(content)
+        .body.split(/\r?\n/)
         .find((l) => /^\s*(\*\*|__)?\s*(description|when_to_use)\s*(\*\*|__)?\s*:/i.test(l));
     if (pseudoLine)
         return {
@@ -191,50 +140,13 @@ export function hasSupersededStatus(frontmatter) {
 export function hasHotCaptureMode(frontmatter) {
     return frontmatter?.captureMode === "hot";
 }
-// ── Consolidate merge size gate ──────────────────────────────────────────────
-// ── Reflect size gate ────────────────────────────────────────────────────────
-/** Ratio lower-bound: proposed body must be at least this fraction of source. */
 export const REFLECT_SHRINK_RATIO_MIN = 0.5;
-/** Ratio upper-bound: proposed body must not exceed this fraction of source. */
 export const REFLECT_EXPAND_RATIO_MAX = 2.5;
-/**
- * Below this byte count, ratio checks are too noisy — skip them entirely.
- * Unchanged from the original gate.
- */
 export const REFLECT_SIZE_GUARD_MIN_BYTES = 200;
-/**
- * Absolute shrinkage floor (bytes).  Even if `ratio * sourceLen` is lower, a
- * proposed body of at least this many bytes is always accepted on the shrinkage
- * side.  Protects against false positives when the source is small (<300 bytes).
- */
 export const REFLECT_ABSOLUTE_FLOOR_BYTES = 150;
-/**
- * Absolute expansion ceiling (bytes).  Even if `ratio * sourceLen` is lower, a
- * proposed body up to this many bytes is always accepted on the expansion side.
- * Protects against false positives when the source is small (≤778 bytes, p25).
- */
 export const REFLECT_ABSOLUTE_CEILING_BYTES = 2500;
-/**
- * Hard expansion cap (bytes).  Regardless of ratio, a proposed body exceeding
- * this limit is always rejected.  Guards against runaway LLM hallucinations on
- * large sources.
- */
 export const REFLECT_ABSOLUTE_MAX_BYTES = 25000;
-/**
- * Calibrated size check: compare proposed body length against source body
- * length using a blended-bound formula.
- *
- * **Shrinkage** — accept if:
- *   `proposedLen >= max(REFLECT_SHRINK_RATIO_MIN * sourceLen, REFLECT_ABSOLUTE_FLOOR_BYTES)`
- *
- * **Expansion** — accept if:
- *   `proposedLen <= min(max(REFLECT_EXPAND_RATIO_MAX * sourceLen, REFLECT_ABSOLUTE_CEILING_BYTES), REFLECT_ABSOLUTE_MAX_BYTES)`
- *
- * Returns `{ ok: true }` when:
- *   - `sourceBody` is absent or `undefined`
- *   - source body is shorter than {@link REFLECT_SIZE_GUARD_MIN_BYTES}
- *   - the proposed length is within the blended bounds
- */
+/** The reflect size gate (see the module note); no source or a tiny one passes. */
 export function checkReflectSize(sourceBody, proposedBody) {
     if (typeof sourceBody !== "string")
         return { ok: true };
@@ -243,19 +155,14 @@ export function checkReflectSize(sourceBody, proposedBody) {
         return { ok: true };
     const proposedLen = proposedBody.trim().length;
     const ratio = proposedLen / sourceLen;
-    // Shrinkage check: lower bound = max(ratio floor, absolute floor)
-    const shrinkFloor = Math.max(REFLECT_SHRINK_RATIO_MIN * sourceLen, REFLECT_ABSOLUTE_FLOOR_BYTES);
-    if (proposedLen < shrinkFloor) {
+    if (proposedLen < Math.max(REFLECT_SHRINK_RATIO_MIN * sourceLen, REFLECT_ABSOLUTE_FLOOR_BYTES)) {
         return { ok: false, code: "EXCESSIVE_SHRINKAGE", ratio };
     }
-    // Expansion check: upper bound = min(max(ratio ceiling, absolute ceiling), hard cap)
     const expandCeiling = Math.min(Math.max(REFLECT_EXPAND_RATIO_MAX * sourceLen, REFLECT_ABSOLUTE_CEILING_BYTES), REFLECT_ABSOLUTE_MAX_BYTES);
-    if (proposedLen > expandCeiling) {
+    if (proposedLen > expandCeiling)
         return { ok: false, code: "EXCESSIVE_EXPANSION", ratio };
-    }
     return { ok: true };
 }
-// ── ProposalValidator entries (registered with proposal-validators.ts) ──────
 const descriptionQualityValidator = {
     name: "description-quality",
     appliesTo(_proposal, ctx) {
@@ -282,6 +189,46 @@ const descriptionQualityValidator = {
         ];
     },
 };
+/**
+ * The lesson checks distill and accept share: a valid description and
+ * when_to_use that differ, and no pseudo-frontmatter in the body. `text`
+ * continues a caller-specific subject.
+ */
+export function lessonQualityIssues(fm, content, inputRef) {
+    const issues = [];
+    const descCheck = isValidDescription(fm.description, inputRef);
+    if (!descCheck.ok) {
+        issues.push({
+            kind: "invalid-description",
+            field: "description",
+            text: ` has an invalid description: ${descCheck.reason}.`,
+            ...(descCheck.severity ? { severity: descCheck.severity } : {}),
+        });
+    }
+    const wtuCheck = isValidWhenToUse(fm.when_to_use, inputRef);
+    if (!wtuCheck.ok) {
+        issues.push({
+            kind: "invalid-when_to_use",
+            field: "when_to_use",
+            text: ` has an invalid when_to_use: ${wtuCheck.reason}.`,
+        });
+    }
+    if (descCheck.ok &&
+        wtuCheck.ok &&
+        typeof fm.description === "string" &&
+        typeof fm.when_to_use === "string" &&
+        fm.description.trim().toLowerCase() === fm.when_to_use.trim().toLowerCase()) {
+        issues.push({
+            kind: "description-equals-when_to_use",
+            field: "description",
+            text: " has identical description and when_to_use.",
+        });
+    }
+    const dfm = detectDoubleFrontmatter(content);
+    if (dfm)
+        issues.push({ kind: dfm.kind, field: "body", text: `: ${dfm.message}` });
+    return issues;
+}
 const lessonContentQualityValidator = {
     name: "lesson-content-quality",
     appliesTo(_proposal, ctx) {
@@ -297,34 +244,11 @@ const lessonContentQualityValidator = {
         catch {
             return [];
         }
-        const findings = [];
-        const descCheck = isValidDescription(fm.description, proposal.ref);
-        if (!descCheck.ok)
-            findings.push({
-                kind: "invalid-description",
-                message: `Lesson proposal ${proposal.id} (${proposal.ref}) has an invalid description: ${descCheck.reason}.`,
-                ...(descCheck.severity ? { severity: descCheck.severity } : {}),
-            });
-        const wtuCheck = isValidWhenToUse(fm.when_to_use, proposal.ref);
-        if (!wtuCheck.ok)
-            findings.push({
-                kind: "invalid-when_to_use",
-                message: `Lesson proposal ${proposal.id} (${proposal.ref}) has an invalid when_to_use: ${wtuCheck.reason}.`,
-            });
-        if (descCheck.ok &&
-            wtuCheck.ok &&
-            typeof fm.description === "string" &&
-            typeof fm.when_to_use === "string" &&
-            fm.description.trim().toLowerCase() === fm.when_to_use.trim().toLowerCase()) {
-            findings.push({
-                kind: "description-equals-when_to_use",
-                message: `Lesson proposal ${proposal.id} (${proposal.ref}) has identical description and when_to_use.`,
-            });
-        }
-        const dfm = detectDoubleFrontmatter(proposalContent(proposal));
-        if (dfm)
-            findings.push({ kind: dfm.kind, message: `Lesson proposal ${proposal.id} (${proposal.ref}): ${dfm.message}` });
-        return findings;
+        return lessonQualityIssues(fm, proposalContent(proposal), proposal.ref).map((issue) => ({
+            kind: issue.kind,
+            message: `Lesson proposal ${proposal.id} (${proposal.ref})${issue.text}`,
+            ...(issue.severity ? { severity: issue.severity } : {}),
+        }));
     },
 };
 const sourceNotSupersededValidator = {
@@ -344,18 +268,14 @@ const sourceNotSupersededValidator = {
         return [];
     },
 };
-/** Strip an opening frontmatter block (`---\n…\n---`) from `content`, returning the body. */
-function stripFrontmatterBody(content) {
-    return content.replace(/^---\r?\n[\s\S]*?\r?\n---\r?\n?/, "");
-}
 const reflectSizeGuardValidator = {
     name: "reflect-size-guard",
     appliesTo(proposal, ctx) {
         return proposal.source === "reflect" && typeof ctx.source?.content === "string";
     },
     validate(proposal, ctx) {
-        const sourceBody = stripFrontmatterBody(ctx.source?.content ?? "");
-        const proposedBody = typeof proposal.payload?.content === "string" ? stripFrontmatterBody(proposalContent(proposal)) : "";
+        const sourceBody = splitFrontmatter(ctx.source?.content ?? "").body;
+        const proposedBody = typeof proposal.payload?.content === "string" ? splitFrontmatter(proposalContent(proposal)).body : "";
         const outcome = checkReflectSize(sourceBody, proposedBody);
         if (outcome.ok)
             return [];
@@ -373,16 +293,10 @@ const reflectSizeGuardValidator = {
     },
 };
 /**
- * Accept-time data-loss guard (#952 Addendum). `sanitizeReflectPayload`
- * already defers a proposal whose body echoes {@link REFLECT_TRUNCATION_MARKER}
- * (the notice appended when the source asset was too large to send in full)
- * with `reflect-truncation-leak` — but that is a creation-time check, and a
- * proposal can reach `proposal accept` / drain promotion without ever going
- * through it (e.g. a defer that a human then accepts anyway, or a future
- * reflect code path that mints proposals directly). A leaked marker replacing
- * real asset content on disk is data loss, not a quality nit, so unlike the
- * rest of this file's validators this one is NOT wrapped by {@link advisory}
- * — it blocks acceptance the same way the generic/canonical validators do.
+ * A body still carrying {@link REFLECT_TRUNCATION_MARKER} (its source was too
+ * large to send in full) would overwrite the asset with an incomplete rewrite:
+ * data loss, so this blocks at accept even for a proposal that never passed
+ * reflect's own creation-time check (#952).
  */
 const reflectTruncationMarkerValidator = {
     name: "reflect-truncation-marker",
@@ -435,19 +349,9 @@ const reflectPromptScaffoldingValidator = {
     },
 };
 /**
- * Report a validator's findings as advisory.
- *
- * These validators judge prose quality — a description that reads like a
- * heading, an odd backtick count, a body that grew more than the reflect
- * ratio allows. They used to BLOCK `proposal accept`, which a human types
- * after reading the diff, and the error told that human to "fix the proposal
- * payload and try again" — but there is no `akm proposal edit` and `accept`
- * takes no `--force`, so the only way out was hand-editing the proposals
- * database. A blocking check whose remedy does not exist is not a check.
- *
- * Structural findings stay blocking: an empty body, an unparseable ref,
- * malformed frontmatter and a broken workflow shape genuinely cannot be
- * written, and they live in {@link defaultProposalValidators}.
+ * Prose-quality findings only advise: a human accepting after reading the diff
+ * has no way to edit the proposal, so blocking on them left no remedy.
+ * Structural findings (in the default validators) still block.
  */
 function advisory(validator) {
     return {
@@ -455,14 +359,7 @@ function advisory(validator) {
         validate: (proposal, ctx) => validator.validate(proposal, ctx).map((finding) => ({ ...finding, severity: "warn" })),
     };
 }
-/**
- * Full set of quality validators in registration order. Appended onto
- * {@link defaultProposalValidators} so they run inside `validateProposal` on
- * `proposal accept` automatically. All prose-quality checks report without
- * blocking (see {@link advisory}). The truncation-marker, redacted-content,
- * and reflected-prompt-scaffolding validators block because they protect
- * durable content rather than judging prose quality.
- */
+/** The quality validators `validateProposal` runs; the last three protect durable content and block. */
 export const defaultProposalQualityValidators = [
     ...[
         descriptionQualityValidator,

package/dist/commands/proposal/validators/proposal-validators.js

@@ -6,7 +6,7 @@ import { parseRefInput } from "../../../core/asset/resolve-ref.js";
 import { proposalContent } from "../../../core/file-change.js";
 import { lintLessonContent } from "../../../core/lesson-lint.js";
 import { parseTaskSource } from "../../../tasks/source/parse-task-source.js";
-import { compileWorkflowSource } from "../../../workflows/source-ir/compile.js";
+import { compileWorkflowSource } from "../../../workflows/compile.js";
 import { defaultProposalQualityValidators } from "./proposal-quality-validators.js";
 const genericProposalValidator = {
     name: "generic-proposal-validator",

package/dist/commands/proposal/validators/proposals.js

@@ -1,105 +1,29 @@
 // This Source Code Form is subject to the terms of the Mozilla Public
 // License, v. 2.0. If a copy of the MPL was not distributed with this
 // file, You can obtain one at https://mozilla.org/MPL/2.0/.
-/**
- * Proposal validation and content repair.
- *
- * The proposal repository and domain service moved to `../repository.ts` (#578
- * storage consolidation). This module keeps only the two proposal *validators* — {@link validateProposal} and
- * {@link repairProposalContent}.
- */
+/** Proposal validation and the one content repair applied before promotion. */
 import { repairTruncatedDescription } from "../../../core/text-truncation.js";
+import { splitFrontmatter } from "../../improve/reflect-noise.js";
 import { runProposalValidators } from "./proposal-validators.js";
 /**
- * Validate a proposal payload before promotion. Generic by default — any
- * proposal must parse cleanly and carry a non-empty body. Asset types in the
- * fail-closed canonical registry run their existing parser or validator; types
- * without one remain on generic validation.
+ * Validate a proposal before promotion: it must parse and carry a body, and a
+ * type with a canonical validator runs it.
  */
 export function validateProposal(proposal) {
     return runProposalValidators(proposal);
 }
-// ── Content repair ──────────────────────────────────────────────────────────
 /**
- * Attempt bounded, deterministic repair of mechanically-fixable defects in a
- * proposal's markdown content. NEVER fabricates text — only strips known-bad
- * structure and applies {@link repairTruncatedDescription} to a truncated
- * description when one is detected.
- *
- * Repairs performed:
- *   1. Apply {@link repairTruncatedDescription} to a truncated/hanging
- *      `description` field in the frontmatter.
- *
- * It deliberately does NOT delete body lines. Two earlier repairs dropped
- * every body line that restated a frontmatter key and every `---` in a body
- * with frontmatter. Both fired inside fenced code blocks, so any asset
- * documenting frontmatter — a note about akm, Claude Code skills, Jekyll,
- * Hugo — was silently gutted on `proposal accept`, and the repaired bytes
- * were written back over the original in the proposals database. A repair
- * that can destroy content is not a repair.
- *
- * Returns the repaired content string. When no repairs apply the input is
- * returned byte-identical so callers can use strict equality to detect
- * whether a repair actually happened.
- *
- * CRITICAL: This function is CONTENT-PRESERVING. Callers MUST re-validate the
- * repaired output via {@link validateProposal} / {@link runProposalValidators}
- * before promotion — a repair that makes things *worse* (or is simply
- * insufficient) must be caught by the existing gate.
+ * Normalize line endings and complete a truncated frontmatter `description`
+ * (`repairTruncatedDescription`, with the body as context). Nothing else: an
+ * earlier repair that deleted body lines gutted any asset documenting
+ * frontmatter. Callers re-validate the result.
  */
 export function repairProposalContent(content) {
     if (typeof content !== "string" || content.trim() === "")
         return content;
-    // Determine whether the content has a frontmatter block so we know how
-    // many `---` fence lines are expected.
-    const hasFrontmatter = /^---\r?\n[\s\S]*?\r?\n---/.test(content);
-    // Split into lines for structural repairs.
-    const lines = content.split(/\r?\n/);
-    // Track whether we are inside the opening frontmatter block so we can
-    // leave it untouched and only repair the body.
-    let inFrontmatter = false;
-    // Frontmatter fence index tracking: first fence opens FM, second closes it.
-    let fmOpenSeen = false;
-    let fmCloseSeen = false;
-    const repairedLines = [];
-    for (const line of lines) {
-        const isFence = /^---\s*$/.test(line);
-        // Track frontmatter fences (first two `---` fences delimit the FM block).
-        if (isFence && !fmCloseSeen) {
-            if (!fmOpenSeen) {
-                fmOpenSeen = true;
-                inFrontmatter = true;
-                repairedLines.push(line);
-                continue;
-            }
-            if (inFrontmatter) {
-                fmCloseSeen = true;
-                inFrontmatter = false;
-                repairedLines.push(line);
-                continue;
-            }
-        }
-        // We are now in the body (past the frontmatter or no frontmatter).
-        if (inFrontmatter) {
-            // Still inside the frontmatter — keep as-is.
-            repairedLines.push(line);
-            continue;
-        }
-        repairedLines.push(line);
-    }
-    let repaired = repairedLines.join("\n");
-    // Repair 3: Apply repairTruncatedDescription to the description field.
-    // We operate on the raw text rather than re-parsing YAML to avoid
-    // reformatting unrelated frontmatter keys.
-    if (hasFrontmatter) {
-        // Extract the body text (after the second `---`) so we can pass it to
-        // repairTruncatedDescription as context for the swap-in heuristic.
-        const bodyMatch = repaired.match(/^---\r?\n[\s\S]*?\r?\n---\r?\n?([\s\S]*)$/);
-        const bodyText = bodyMatch?.[1] ?? "";
-        repaired = repaired.replace(/^(description:\s*)(.*?)(\r?\n)/m, (_match, prefix, rawDesc, nl) => {
-            const fixed = repairTruncatedDescription(rawDesc.trim(), bodyText);
-            return `${prefix}${fixed}${nl}`;
-        });
-    }
-    return repaired;
+    const repaired = content.replace(/\r\n/g, "\n");
+    const { fmText, body } = splitFrontmatter(repaired);
+    if (fmText === null)
+        return repaired;
+    return repaired.replace(/^(description:\s*)(.*?)(\r?\n)/m, (_match, prefix, rawDesc, nl) => `${prefix}${repairTruncatedDescription(rawDesc.trim(), body)}${nl}`);
 }