@gamaze/hicortex 0.16.0 → 0.16.1

package/README.md

@@ -226,6 +226,7 @@ Config at `~/.hicortex/config.json`. Created by `init`. Key options:
 | `recallMinSimilarity` | Relevance floor for index entries (default: 0.55; text-search matches always pass) |
 | `recallReshowTurns` | Turns before an already-shown memory may reappear in the same session (default: 30) |
 | `recallMinPromptChars` | Prompts shorter than this skip the recall index (default: 20) |
+| `recallTitleChars` | Chars of each memory's first line shown in an index entry (default: 150, range 40–400). Raised from 100 on 2026-08-02: with topic-first memory titles, 150 chars carries the subject *and* its claim, where 100 cut the claim mid-sentence. Costs roughly +74 tokens per 6-line block |
 | `sessionIntentWeight` | Blend weight of the session-intent rolling centroid in the recall search vector: `query = (1-w)·prompt + w·centroid` (default: 0.33; set 0 to disable — pure-prompt recall, the kill-switch). The first turn of a session searches with pure prompt and seeds the centroid; subsequent turns blend so recall follows the session's intent instead of being query-literal. The EMA rate (0.4) is a shipped constant, not configurable |
 | `dedupMergeThreshold` | Minimum cosine similarity for `hicortex dedup` to cluster memories as near-duplicates (default: 0.92) |
 | `supersessionMinSimilarity` | Minimum cosine similarity for a nightly supersession candidate pair (default: 0.80) |

package/dist/consolidate.js

@@ -307,7 +307,10 @@ async function stageReflection(db, memories, llm, budget, embedFn, dryRun) {
             const severity = String(lo.severity ?? "important");
             const confidence = String(lo.confidence ?? "medium");
             const sourcePattern = String(lo.source_pattern ?? "");
-            let content = `## Lesson: ${lessonText}\n\n`;
+            // No `## Lesson:` prefix: memory_type='lesson' carries the type, and the
+            // text is the topic-first first line (display reads the first line, not a
+            // header parse — see lessons-context.ts / index.ts).
+            let content = `${lessonText}\n\n`;
             content += `**Type:** ${lessonType}\n`;
             content += `**Severity:** ${severity}\n`;
             content += `**Confidence:** ${confidence}\n`;

package/dist/distiller.js

@@ -314,6 +314,14 @@ async function distillChunk(llm, transcript, projectName, date) {
         return { entries: [], dropped: [] };
     }
     const parsed = parseDistilledEntries(result);
+    // Smoke alarm (PR #218 review): the prompt enforces topic-first, but models
+    // sometimes ignore constraints (cf. the prior max-15-bullet failure). Count
+    // entries that still look actor-led or bracket-led so a format regression
+    // shows in nightly logs, not months later in the next eval. Non-blocking.
+    const offTopic = parsed.filter((e) => /^\s*(user|ai|the user|assistant)\b/i.test(e) || /^\s*\[/.test(e)).length;
+    if (parsed.length > 0 && offTopic > 0) {
+        console.log(`[hicortex]     topic-first check: ${offTopic}/${parsed.length} entries look actor/bracket-led (prompt may be ignored)`);
+    }
     const entries = [];
     const dropped = [];
     for (const entry of parsed) {
@@ -409,23 +417,22 @@ function hasMinimalSubstance(entry) {
 function parseDistilledEntries(markdown) {
     const entries = [];
     const lines = markdown.split("\n");
-    let currentSection = "";
     for (const line of lines) {
         const trimmed = line.trim();
-        // Section headers
-        if (trimmed.startsWith("### ")) {
-            currentSection = trimmed.slice(4).trim();
+        // Skip all markdown headers (session title, classification, section
+        // headings). Sections are NOT prefixed onto entries: each bullet already
+        // starts with its [SUBJECT] (topic-first, enforced by prompts.ts), and
+        // prepending "[Section]" re-introduced the category-first prefix the
+        // 2026-08-02 corpus rewrite removed. The section label is unused
+        // downstream (distilled memories all store memory_type='episode').
+        if (trimmed.startsWith("# ") ||
+            trimmed.startsWith("## ") ||
+            trimmed.startsWith("### ")) {
             continue;
         }
-        // Skip top-level headers and classification
-        if (trimmed.startsWith("# ") || trimmed.startsWith("## "))
-            continue;
-        // Bullet items are individual memories
+        // Bullet items are individual, already topic-first memories.
         if (trimmed.startsWith("- ") && trimmed.length > 5) {
-            const entry = currentSection
-                ? `[${currentSection}] ${trimmed.slice(2)}`
-                : trimmed.slice(2);
-            entries.push(entry);
+            entries.push(trimmed.slice(2));
         }
     }
     return entries;

package/dist/eval/relevance-eval.d.ts

@@ -0,0 +1,64 @@
+#!/usr/bin/env node
+/**
+ * Real-query relevance + SNIPPET eval — recall QUALITY on real agent prompts.
+ *
+ * v2 (spec `specs/2026-08-02-relevance-eval.md`) extends the v1 selection-only
+ * eval with the SNIPPET layer: v1 asked "did retrieve() surface the right
+ * memories?" (judge sees up to 2000 chars). Production shows the agent only a
+ * ~100-char one-liner (`recall-index.ts#memoryTitle`), so a memory can be
+ * genuinely relevant while its rendered line is useless — v1 scored that as a
+ * win. v2 grades BOTH: `full_verdict` (selection quality, judge sees full
+ * content) and `line_verdict` (snippet quality, judge sees ONLY the rendered
+ * production one-liner — imported from `recall-index.ts`, never reimplemented).
+ *
+ * v2 additions (spec §4, §5, §6, §6b) layered onto the v1 base (prompt
+ * sampling, readonly snapshot handling, embed-once + neverCalledEmbed,
+ * lenient JSON parse, distribution/CI reporting):
+ *   1. Dual verdict per surfaced memory — TWO separate, blind judge calls.
+ *   2. Snippet-length sweep (100/200/300/title+1st-sentence) on a fixed
+ *      40-prompt subset (8 per source).
+ *   3. Similarity-floor + retrieval-source analysis (near-free — logged, not
+ *      re-judged).
+ *   4. Token-cost estimate (char/4) per K and per snippet-length variant.
+ *   5. ~20 rendered ACTUAL production blocks dumped into the report.
+ *   6. Redundancy — one set-level judge call per (prompt × mode) over the
+ *      production 6.
+ *   7. Rate limiting + resumability (§6b, MANDATORY): serial calls,
+ *      `--judge-delay-ms` (default 2000), exponential backoff with jitter on
+ *      429/5xx/timeout (5→10→20→40→80s, max 5 retries, respects
+ *      `Retry-After`), checkpoint-per-call to a `.jsonl` sidecar, `--resume`,
+ *      progress logging, 10%-error-rate abort, `--max-calls` budget guard
+ *      (default 900).
+ *
+ * Prompt corpus (spec §2, owner decision §11.1): EVEN split, 20 prompts per
+ * source × 5 sources — Hermes (lenny, raider, nano) + CC (the DevOps
+ * `infrastructure` project, the `aironic-marine` project). Saved to
+ * `data/prompts.json`, stable/reused verbatim once a valid v2 set exists.
+ *
+ * Judge: GLM-5.2 via z.ai — the INSTRUMENT only. It never picks candidates;
+ * retrieve() (LLM-free) does. A dedicated raw HTTP caller (NOT `LlmClient`) is
+ * used here on purpose: `LlmClient.completeReflect` bakes in a
+ * nightly-tolerant retry policy (30s/60s/120s, unlimited rate-limit patience)
+ * that conflicts with §6b's specific real-time batch policy (5/10/20/40/80s +
+ * jitter, 5 retries, a hard call budget). Implemented directly here rather
+ * than adding a second retry mode to `llm.ts` (out of scope for this eval,
+ * and another agent is concurrently working elsewhere in this repo).
+ *
+ * Honesty invariants (non-negotiable — mirror recall-sweep.ts + spec §7):
+ *   - Snapshot opened READONLY via openSnapshot — never initDb.
+ *   - noStrengthen: true on every retrieve() call.
+ *   - Real bge-small-en-v1.5 embedder, embed-once + queryEmbedding reuse.
+ *   - neverCalledEmbed self-check ABORTS the run if retrieve() ignores
+ *     queryEmbedding (would invalidate every measured number).
+ *   - GLM-5.2 is the JUDGE only, real prompts, no synthetic queries.
+ *   - Production renderer (`formatIndexLine`/`memoryTitle`) imported from
+ *     `recall-index.ts`, never reimplemented.
+ *   - judge_error batches/units excluded from every denominator, reported
+ *     separately.
+ *
+ * Run:
+ *   npm run eval:relevance -- <snapshot.db> [prompts.json] [report.md] \
+ *       [--judge-delay-ms=2000] [--max-calls=900] [--resume] \
+ *       [--verdicts-json=path] [--verdicts-jsonl=path]
+ */
+export {};