@dzhechkov/harness-cli 0.3.259 → 0.3.260

package/README.md

@@ -999,7 +999,7 @@ dz teach             "<pattern>" [--reward <0-1>] [--domain <name>] [--type rule
 dz teach --reinforce "<dzId-or-exact-text>" [--project <dir>] # bump an existing learned pattern instead of writing a near-duplicate
 dz teach --from-json <file> [--project <dir>] [--no-mirror] [--harmonize]   # bulk-import a `dz recall --all --json` export; prints a harmonize dry-run advisory
 dz consolidate       [--sessions-dir <dir>] [--project <dir>] [--no-mirror]
-dz recall            "<query>" [--limit <N>] [--semantic | --no-semantic] [--project <dir>]   # hybrid lexical+vector when the vector tier is enabled
+dz recall            "<query>" [--limit <N>] [--domain <name>] [--semantic | --no-semantic] [--project <dir>]   # hybrid lexical+vector when the vector tier is enabled; --domain BOOSTS same-domain lessons (never filters)
 dz recall --all      [--json] [--stats]             # export the WHOLE learned store, or inspect learning stats/top uses
 dz recall --books    "<query>" [--book <slug>]      # digitized-book KUs; --book narrows to one book
 dz vector status     [--project <dir>] [--json]     # semantic tier: engine availability, mirrored vs lexical counts, pending queue
@@ -2347,6 +2347,36 @@ guard, `--loc-cap <n>` tunes, default 700), **secret-scan** (the same high-signa
 edit caps/paths/waivers in `guards.config.json`; the project-critic role should then *not* re-flag these
 rules, only waivers without a reason.
 
+### Domain-scoped recall — `--domain <name>` (a boost, never a filter)
+
+**When to use it.** One brain store, more than one *kind* of work. `dz teach --domain` has always
+recorded a domain, but until now the ranking ignored it — so medical-research lessons and coding
+lessons came back interleaved, and the dilution grew with the store.
+
+```bash
+dz recall "signature" --limit 3 --domain security
+#   dz recall "signature"  —  3 hit(s), SQLite FTS5 + vector (agentdb) ranking
+#     [0.90] (security) ⟨both⟩ Crypto anti-hallucination anti-pattern: verifying a signature …
+#     [0.85] (publishing) ⟨both⟩ ⚠q 'Re-sign LAST' applies to EVERY signed pack a sweep touches …
+#     [0.80] (testing) ⟨vector⟩ A structural registry-coverage test that only checks key-SET …
+#     domain "security": 1 exact match(es); the order was already correct — nothing moved;
+#     foreign-domain lessons kept (a boost, not a filter)
+
+dz recall "signature" --limit 3 --domain health-research
+#     domain "health-research": no lesson in this result carries it — order unchanged, nothing was hidden
+```
+
+**Why a boost and not a filter.** Cross-domain transfer is the reason one shared store beats two
+separate ones: *"a reviewer's evidence needs the same execute-don't-describe discipline as your own
+claims"* was learned reviewing code and applies verbatim to medical sources. A filter would cut
+exactly those transfers. So matching hits move UP — at most **2 places** for an exact domain, **1**
+for a hierarchical relative (`health` ↔ `health-research`) — and nothing is ever dropped: the hit
+count with `--domain` equals the hit count without it.
+
+The bound matters: a domain tag is a hint about relevance, not evidence of it, so it breaks near-ties
+instead of overruling lexical ranking. And the note is honest in both directions — it says when
+nothing matched, and it distinguishes *matched* from *actually moved*.
+
 ### Semantic recall (vector tier)
 
 `dz recall` is **hybrid** when the vector tier is available and **exactly the old lexical command** when it is not — enabling it never changes behavior for projects that skip it.

package/dist/cli.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AA0XH,2EAA2E;AAC3E,MAAM,WAAW,KAAK;IACpB,QAAQ,CAAC,GAAG,CAAC,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,CAAC;IACxC;;;;OAIG;IACH,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;;;;;OAKG;IACH,QAAQ,CAAC,aAAa,CAAC,EAAE,iBAAiB,CAAC;CAC5C;AAED,yFAAyF;AACzF,MAAM,MAAM,iBAAiB,GAAG,CAC9B,GAAG,EAAE,MAAM,EACX,IAAI,EAAE;IAAE,QAAQ,CAAC,GAAG,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAA;CAAE,KACvD;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,OAAO,CAAA;CAAE,CAAC;AA+0O9E,wBAAsB,MAAM,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,GAAE,KAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CA8J5E"}
+{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AA8XH,2EAA2E;AAC3E,MAAM,WAAW,KAAK;IACpB,QAAQ,CAAC,GAAG,CAAC,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,CAAC;IACxC;;;;OAIG;IACH,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;;;;;OAKG;IACH,QAAQ,CAAC,aAAa,CAAC,EAAE,iBAAiB,CAAC;CAC5C;AAED,yFAAyF;AACzF,MAAM,MAAM,iBAAiB,GAAG,CAC9B,GAAG,EAAE,MAAM,EACX,IAAI,EAAE;IAAE,QAAQ,CAAC,GAAG,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAA;CAAE,KACvD;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,OAAO,CAAA;CAAE,CAAC;AAk3O9E,wBAAsB,MAAM,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,GAAE,KAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CA8J5E"}

package/dist/cli.js

@@ -15,7 +15,7 @@ assembleCandidates, renderPromotionReport, renderPromotionAdr, normalizePromotio
 // Cold-vs-warm EPOCH RUNNER (feature epoch-replay) — orchestrates + scores, never calls a model.
 replayableInstances, buildWorkOrder, buildJudgePrompts, unblindJudgments, verifyWorkOrder, isValidMargin, DIGEST_HONEST_SCOPE, scoreEpochReplay, generateMockOutcomes, renderEpochReplayResult, renderWorkOrderSummary, renderJudgePromptsSummary, WORK_ORDER_KIND, DEFAULT_MOCK_N, DEFAULT_MOCK_SEED, scoreRun, renderScorecard, renderCompoundingReport, readReinforcementState, readQuarantineState, registrationExitCode, renderRegistrationReport, 
 // Smart Backlog (feature smart-backlog) — goal-directed idea pipeline over the Brain vector engine.
-readBacklogConfig, readIdeas, writeIdeas, ideaId, dedupIdea, readGoalMap, readGoalMapDetailed, parseEffort, ensureBacklogGitignored, isSafeId, alignIdea, mirrorIdeaVector, snapshotIdeas, spinRoulette, rankRoulette, seededRng, eligibleIdeas, stageEnrichment, buildJiraDraft, resolveJiraAdapter, makeBacklogIO, harmonizeBacklog, BACKLOG_BACKENDS, parseReqeDebt, buildReqeBrief, settleReqeDebt, renderReqeList, REQE_SCOPE, } from '@dzhechkov/harness-core';
+readBacklogConfig, readIdeas, writeIdeas, ideaId, dedupIdea, readGoalMap, readGoalMapDetailed, parseEffort, ensureBacklogGitignored, isSafeId, alignIdea, mirrorIdeaVector, snapshotIdeas, spinRoulette, rankRoulette, seededRng, eligibleIdeas, stageEnrichment, buildJiraDraft, resolveJiraAdapter, makeBacklogIO, harmonizeBacklog, BACKLOG_BACKENDS, applyDomainBoost, countDisplacedByCut, renderDomainBoostNote, renderDomainCutNote, parseReqeDebt, buildReqeBrief, settleReqeDebt, renderReqeList, REQE_SCOPE, } from '@dzhechkov/harness-core';
 import { getPreset, PRESET_NAMES } from '@dzhechkov/harness-presets';
 import { scanGitHub, analyzeRepo, generateReport, deepAnalyze, scanAllSources, ScoutMemory } from '@dzhechkov/scout';
 const USAGE = `dz - DZ cross-platform harness CLI
@@ -61,7 +61,7 @@ Usage:
   dz teach "<pattern>" [--reward <0-1>] [--domain <name>] [--type rule|success-pattern|lesson-learned] [--project <dir>] [--no-mirror]   (--project pins the learned store to <dir>/.dz, not the cwd — pin to a canonical brain)
   dz teach --from-json <file> [--project <dir>] [--no-mirror]   (bulk-import a 'dz recall --all --json' export — share a learned store across machines)
   dz consolidate [--sessions-dir <dir>] [--project <dir>] [--no-mirror] [--prune-noise [--apply]] [--prune-quarantine [--apply]]   (both prunes: DRY-RUN by default; --apply snapshots then deletes; prune-quarantine = expired unproven lessons ONLY, never coupled to noise)
-  dz recall "<query>" [--limit <N>] [--semantic | --no-semantic] [--books [--book <slug>]] [--project <dir>] | dz recall --all [--json] | dz recall --usage [--json] | dz recall --forget <dzId>[,<dzId>] [--apply] | dz recall --promote <dzId>[,<dzId>] [--apply]   (forget/promote: dry-run default; forget snapshots before removing; promote lifts lesson-quarantine)
+  dz recall "<query>" [--limit <N>] [--domain <name>] [--semantic | --no-semantic] [--books [--book <slug>]] [--project <dir>] | dz recall --all [--json] | dz recall --usage [--json] | dz recall --forget <dzId>[,<dzId>] [--apply] | dz recall --promote <dzId>[,<dzId>] [--apply]   (--domain <name> BOOSTS lessons of that domain without dropping foreign ones — a shared store keeps its cross-domain transfers; forget/promote: dry-run default; forget snapshots before removing; promote lifts lesson-quarantine)
   dz vector status [--project <dir>] [--json]                          (semantic tier: engine, mirrored vs lexical counts, pending queue)
   dz vector reindex [--project <dir>] [--json]                         (snapshot, re-embed learned-pattern vectors, stamp current model)
   dz vector export <path> [--project <dir>]                            (portable VECTOR form (.rvf, opt-in RVF engine); patterns ship via recall --all --json)
@@ -1842,13 +1842,30 @@ async function cmdRecall(options, flags, cwd, write) {
         : flags.has('no-semantic') || flags.has('lexical')
             ? 'lexical'
             : 'hybrid';
-    const result = await recallHybrid(projectRoot, query, { limit, mode });
+    const wantedDomain = options.get('domain');
+    // OVER-FETCH before boosting (Codex QE #5): the boost used to run on hits ALREADY
+    // truncated to `limit`, so an exact-domain lesson sitting at rank limit+1 could
+    // never receive its promised lift — the feature was weakest in exactly the case it
+    // exists for (foreign-domain dilution pushing a relevant lesson just past the cut).
+    // Fetch a bounded surplus, re-rank, then trim to the limit the caller asked for.
+    const fetchLimit = wantedDomain !== undefined ? Math.min(limit * 3, limit + 20) : limit;
+    const result = await recallHybrid(projectRoot, query, { limit: fetchLimit, mode });
     if (mode === 'semantic' && result.vectorEngine === 'none') {
         // --semantic is an explicit ask — degrading it silently would be dishonest (FR-3).
         write(`dz recall --semantic: ${result.vectorReason ?? 'no vector engine available — run: dz setup --memory agentdb'}`);
         return 1;
     }
-    const hits = result.hits;
+    // Domain-aware re-ranking (health-advisor slice H): `--domain <name>` lifts lessons
+    // tagged with that domain WITHOUT dropping foreign ones — a boost, not a filter, so a
+    // shared store keeps the cross-domain transfers that make it worth more than two stores.
+    const boost = wantedDomain !== undefined ? applyDomainBoost(result.hits, wantedDomain) : null;
+    const hits = (boost ? boost.hits : result.hits).slice(0, limit);
+    // The boost never drops a hit, but the CUT still can: promoting a match into the top
+    // `limit` pushes the last one out, so a lesson visible WITHOUT --domain can vanish
+    // WITH it. Cross-model review called this out as a lie by omission — the note said
+    // "foreign-domain lessons kept" about the pre-cut list while the printed list was
+    // missing one. Count it and say so; the reader can act on it (raise --limit).
+    const displaced = boost !== null ? countDisplacedByCut(result.hits, boost.hits, limit) : 0;
     if (asJson) {
         // Portable contract UNCHANGED (I-7/AC-6): a plain PatternRecord[] — round-trips through
         // `dz teach --from-json` regardless of which backend ranked each hit.
@@ -1858,6 +1875,11 @@ async function cmdRecall(options, flags, cwd, write) {
     if (hits.length === 0) {
         write(`dz recall "${query}"`);
         write(`  No matching patterns (teach some with \`dz teach\`, or consolidate sessions).`);
+        // The domain note must print here too (Codex QE #10): a --domain run with zero hits
+        // silently said nothing about the domain, so the reader could not tell whether the
+        // boost had been applied and found nothing, or had not run at all.
+        if (boost !== null && wantedDomain !== undefined)
+            write(renderDomainBoostNote(boost, wantedDomain));
         return 0;
     }
     const vectorOn = result.vectorEngine !== 'none' && result.vectorError === undefined && mode !== 'lexical';
@@ -1870,12 +1892,26 @@ async function cmdRecall(options, flags, cwd, write) {
         const qTag = h.quarantined === true ? ' ⚠q' : '';
         if (h.quarantined === true)
             sawQuarantined = true;
-        write(`  [${h.pattern.reward.toFixed(2)}] (${h.pattern.domain})${backendTag}${qTag} ${h.pattern.pattern.slice(0, 80)}`);
+        // ONE line per hit, always. A lesson may contain newlines, and printing them raw
+        // let stored CONTENT forge lines that look like the tool's own output — including
+        // the domain-boost note that learning_bridge.py reads as a capability probe. Any
+        // consumer that parses this output line-wise has the same exposure, so the fix
+        // belongs at the point of rendering rather than in each reader.
+        const oneLine = h.pattern.pattern.replace(/[\r\n]+/g, ' ⏎ ').slice(0, 80);
+        write(`  [${h.pattern.reward.toFixed(2)}] (${h.pattern.domain})${backendTag}${qTag} ${oneLine}`);
     }
     if (sawQuarantined) {
         // The loop stays VISIBLE (ADR D2): a quarantined hit is shown, marked, and explained.
         write('  ⚠q = quarantined (unproven hypothesis, rank damped) — confirm with dz teach --reinforce, or dz recall --promote <dzId> --apply');
     }
+    if (boost !== null && wantedDomain !== undefined) {
+        // Say what the boost did — INCLUDING when it did nothing. A silent reorder would
+        // let the reader believe the ranking was domain-aware when it had no match to work with.
+        write(renderDomainBoostNote(boost, wantedDomain));
+        const cutNote = renderDomainCutNote(displaced, limit);
+        if (cutNote !== '')
+            write(cutNote);
+    }
     if (result.vectorError !== undefined && mode !== 'lexical') {
         // Engine present but the semantic leg failed/timed out — one honest line, exit 0 (05 §2.3).
         write(`  ℹ vector search degraded: ${result.vectorError} — showing lexical ranking`);