@dzhechkov/harness-cli 0.7.7 → 0.7.9

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/harness-cli",
-  "version": "0.7.7",
+  "version": "0.7.9",
   "description": "The dz CLI — install AI skills for Claude Code, Codex, OpenCode, Hermes, OpenClaude, GitHub Copilot. 67 commands, 14 presets, 10 platform targets.",
   "type": "module",
   "license": "MIT",
@@ -41,22 +41,22 @@
     "sbom.json"
   ],
   "dependencies": {
-    "@dzhechkov/harness-core": "^0.7.8",
+    "@dzhechkov/harness-core": "^0.7.12",
     "@dzhechkov/harness-presets": "^0.5.14",
     "@dzhechkov/scout": "^0.8.7",
     "@dzhechkov/skills-devops": "^0.3.13",
     "@dzhechkov/skills-mcp": "^0.3.6",
     "@dzhechkov/skills-web3": "^0.2.5",
     "@dzhechkov/skills-qe": "^0.3.10",
-    "@dzhechkov/skills-academic": "^0.3.3",
     "@dzhechkov/skills-meta": "^0.9.51",
+    "@dzhechkov/skills-academic": "^0.3.3",
     "@dzhechkov/skills-ecc": "^0.1.8",
-    "@dzhechkov/skills-pm": "^0.1.1",
+    "@dzhechkov/skills-news": "^0.1.12",
     "@dzhechkov/skills-taste": "^0.1.1",
+    "@dzhechkov/skills-pm": "^0.1.1",
     "@dzhechkov/skills-idea2prd": "^0.1.11",
-    "@dzhechkov/skills-reverse-engineering": "^0.1.9",
     "@dzhechkov/skills-presentation-storyteller": "^0.1.10",
-    "@dzhechkov/skills-news": "^0.1.12",
+    "@dzhechkov/skills-reverse-engineering": "^0.1.9",
     "@dzhechkov/skills-website-cloner": "^0.1.1"
   },
   "devDependencies": {

package/sbom.json

@@ -25,7 +25,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "3d621db3f7811b174ee46b19fa462316ef13679786a910ac64c65fb05bfd5ef7"
+          "content": "07e1f395043324fe72c1170406340df3ebdb1dbabde64d80dc8124269a0cb5a6"
         }
       ]
     },
@@ -125,7 +125,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "c7856b79fdf6683fe855268191805ed07b1c68f2ff7ac544d46ac3d93e4c94ef"
+          "content": "947a13984fe9e992098c4d7b63303b6048f30d7ff0f509b666022e1c8c67cd54"
         }
       ]
     },
@@ -135,7 +135,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "27741d68ac996a060df7161b49e372f68e31b990a406f0792c79c474ff17eeca"
+          "content": "d539a0084dced536df3edadd79bcce94d1cb7205d1e534922fd8a1ca37c7fc7d"
         }
       ]
     },
@@ -145,7 +145,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "bea844c256e8d947464a4690d2224e4f28d61f433d6e50e5fe26bed3b415ae26"
+          "content": "b3b37c4d27ea28e00da2f6d1bc438e22ebb48da8dec365e9f4128071647c3dd0"
         }
       ]
     },
@@ -335,7 +335,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "b7c47af8811ca86ee8dd5d4d98cab76c4cceddaafbe09676141bacae92b69b9d"
+          "content": "98583e52265eb3951c2596f652f5046cc2044e5a4c12353ea1a380a037819778"
         }
       ]
     },
@@ -365,7 +365,7 @@
       "hashes": [
         {
           "alg": "SHA-256",
-          "content": "d8288240b5420ddca2cabdb7b358c7c18ba59724defda951d0dbc61439b6fc4c"
+          "content": "e13c80580e477748ab072078f3b90e9301ad60b01106e228501f16bc753f994a"
         }
       ]
     },

package/src/cli.ts

@@ -302,6 +302,9 @@ import {
   renderContentProbe,
   findNonRegistrableSkillDirs,
   assembleCompoundingReport,
+  banditStats,
+  narrowBanditReport,
+  renderBanditHealth,
   // Cold-vs-warm EPOCH RUNNER (feature epoch-replay) — orchestrates + scores, never calls a model.
   replayableInstances,
   buildWorkOrder,
@@ -518,6 +521,11 @@ Usage:
   dz tg-post --draft <file.html> [--manifest <sources.json>] [--channel <@name|id>] [--send --yes] [--night] [--preview] [--json]   (the sender for an APPROVED channel post, per the accepted genai-tweets-channel ADRs: HTML mode only — never MarkdownV2; link preview OFF by default (x.com previews in Telegram are broken); the 00:00-06:00 MSK quiet window refuses without an explicit --night. DEFAULT IS A DRY-RUN: it validates the draft (tag balance, allowed tags, bare &/<, the 4096 visible-character limit with the overshoot counted) and runs the provenance gate over --manifest IN-PROCESS — a draft with no manifest is refused as unchecked, and anything but ALLOWED refuses. A real send needs --send --yes, stating ADR-004's manual-publishing decision out loud each time. The token comes from TELEGRAM_BOT_TOKEN or telegram.tokenFile in .dz/config.json and is never printed. exit 0 sent or clean dry-run / 1 refused or Telegram error / 2 usage)
   dz name-check [--command <n>] [--module <basename>] [--export <a,b>] [--project <dir>] [--json]   (is this name free, BEFORE a line of code? Scans workspace SOURCE — never dist, because a stale build answers 'free' confidently. Checks a dz command name against the dispatcher AND the help block, a module basename against every package's src/, and exported identifiers against every declaration in the workspace. exit 0 all free / 1 at least one taken, naming where / 2 nothing asked or the scan did not run — an empty sweep is never a clean bill. Honest limit, printed on the passing path: it reads declarations, so a re-export under a different name stays the build's job)
   dz provenance-check --manifest <sources.json> [--project <dir>] [--json]   (nothing goes out citing a source that may not leave this machine. Checks PROVENANCE, not words: every claim names its source, and only a KNOWN kind that resolves safely is cleared. Repo paths go through 'git -C <root> check-ignore' over the RESOLVED path — a symlink into an ignored directory is REFUSED (git classifies the string and never dereferences, MEASURED), and the verdict does not change with your working directory. Store records must be named in the git-TRACKED provenance-public.json, so declaring one public is a reviewable commit rather than a field inside an ignored store. An undeclared kind is refused, never inferred from the path's shape. exit 0 allowed / 1 blocked / 3 NOT ESTABLISHED — an empty manifest, an unreadable one, or an oracle that did not run is never a pass. It proves what was CITED: it cannot see a paraphrase with no citation, nor confidential text pasted by hand into an allowed file)
+  dz project-skills [--project <dir>] [--json] [--stages-json]   (polymorphic feature-adr: resolve architecture/project-skills.json — fixed roles product-vision/critic/brand/impl-bar plus an open extra[] — into per-stage guidance. READ-ONLY. --project names the root explicitly, so it works from any cwd; without it the manifest is read from the current repo. No manifest ⇒ a byte-identical generic run)
+  dz discrimination-check --slug <slug> [--base <ref>] [--json]   (does the ADR's named test actually DISCRIMINATE? Re-runs it on a worktree at the pre-feature commit, where it MUST go red. A test that passes with the feature removed proves nothing; dz amendment-check proves the test exists, this proves it bites)
+  dz guard [check|promote|init] [--json] [--force]   (HARD/SOFT repo rules — readme-first, lockfile-in-sync, claim tagging — run automatically as a pre-flight inside dz publish. HARD blocks, SOFT warns)
+  dz architecture [--check --slug <s> --desc <text>] [--project <dir>] [--revise]   (the live product map + vision: --check is the soft Step-0 сверка of a new feature against them, reporting {signal,confidence} rather than blocking)
+  dz sbom [--pack <name>] [--out <file>]   (CycloneDX software bill of materials for the workspace, or for one pack with --pack)
   dz amendment-check --slug <slug> | --feature-dir <dir> | --all [--json]   (the deterministic Step-8 amendment gate: every AM-N row must resolve to a test found INSIDE the file the row names; the PLAN is authoritative when it carries rows, and an ideation amendment the plan drops is a failure. exit 0 pass/skip, 1 fail, 3 NOT-ESTABLISHED — a section that parsed ZERO rows is never a pass. --all is a CENSUS and always exits 0. Does NOT prove non-vacuity — that is dz discrimination-check)
   dz feature-adr-record --kind ledger|training-pair --stage <s> [--slug <s>] [--row|--pair <json>] [--mark <n>] [--once] [--json]   (the witnessed writer for the run-cost ledger and training pairs: the payload arrives as an ARGUMENT, never as shell; a malformed or wrong-kind payload is REFUSED before any write; the timestamp is stamped before serialising; the append is verified by re-reading the tail. exit 0 written|duplicate|skipped, 2 refused, 3 not-verified — a record failure is never blocking)
   dz feature-adr-checkpoint (--slug <feature> | --feature-dir <abs>) --stage <s> --input-hash <h> --result <json> [--artifact a,b] [--json]   (record a pipeline stage ONLY after measuring its artifacts on disk; refuses a null result, an absent artifact, or a stage that declares none — the subagent runs a COMMAND instead of hand-writing durable state)
@@ -3572,8 +3580,8 @@ async function cmdRecall(options: Map<string, string>, flags: Set<string>, cwd:
   // second occurrence. Foreign stdout is routed to stderr for the duration of the engine call — our
   // own output is written after it returns.
   const result = asJson
-    ? await withForeignStdoutOnStderr(() => recallHybrid(projectRoot, query, { limit: fetchLimit, mode }))
-    : await recallHybrid(projectRoot, query, { limit: fetchLimit, mode });
+    ? await withForeignStdoutOnStderr(() => recallHybrid(projectRoot, query, { limit: fetchLimit, mode, deferExposures: true, ...(wantedDomain !== undefined ? { domain: wantedDomain } : {}) }))
+    : await recallHybrid(projectRoot, query, { limit: fetchLimit, mode, deferExposures: true, ...(wantedDomain !== undefined ? { domain: wantedDomain } : {}) });
 
   if (mode === 'semantic' && result.vectorEngine === 'none') {
     // --semantic is an explicit ask — degrading it silently would be dishonest (FR-3).
@@ -3605,6 +3613,23 @@ async function cmdRecall(options: Map<string, string>, flags: Set<string>, cwd:
   // "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;
+  // lesson-bandit-rerank FR-8/AC-11: the bandit block must describe the POST-cut list the reader
+  // actually sees. `dz recall` over-fetches under --domain and truncates AGAIN here, so reporting
+  // the core's own (already post-merge-cut) count would still describe a list one cut too early.
+  // ABSENT while disarmed — its presence is what tells a reader the feature ran.
+  const shownDzIds = hits.map((h) => patternRecordId(h.pattern));
+  const banditBlock = result.bandit === undefined
+    ? undefined
+    : narrowBanditReport(result.bandit, shownDzIds);
+  // Exposures were DEFERRED (`deferExposures: true`) precisely so this cut happens first: `dz recall`
+  // over-fetches under --domain and truncates again here, and counting an over-fetched hit as "seen"
+  // both inflates the health metrics and mislabels a candidate the reader never laid eyes on.
+  result.commitExposures?.(shownDzIds);
+  const renderBanditNote = (b: NonNullable<typeof banditBlock>): string =>
+    `  ℹ bandit payoff: ${b.moved} of ${b.armsConsidered} shown hit(s) moved (ctx ${b.contextKey}` +
+    `, ${b.unknownArms} with no measured payoff yet, ${b.quarantinedExcluded} quarantined excluded` +
+    `${b.exploration ? `, exploration ON — ${b.explored} trial impression(s)` : ''}` +
+    `${b.reason !== null ? `, state ${b.reason}` : ''})`;
   if (asJson) {
     // Portable contract UNCHANGED (I-7/AC-6): a plain PatternRecord[] — round-trips through
     // `dz teach --from-json` regardless of which backend ranked each hit.
@@ -3635,6 +3660,10 @@ async function cmdRecall(options: Map<string, string>, flags: Set<string>, cwd:
       const cutNoteJson = renderDomainCutNote(displaced, limit);
       if (cutNoteJson !== '') process.stderr.write(`${cutNoteJson}\n`);
     }
+    // STDERR, not stdout: `--json`'s stdout contract is a plain PatternRecord[] that round-trips
+    // through `dz teach --from-json`, and wrapping it in an object to make room for one report
+    // would break every existing consumer. The block travels beside the other honesty notes.
+    if (banditBlock !== undefined) process.stderr.write(`${JSON.stringify({ bandit: banditBlock })}\n`);
     return semanticUnserved ? 1 : 0;
   }
 
@@ -3733,6 +3762,12 @@ async function cmdRecall(options: Map<string, string>, flags: Set<string>, cwd:
     // 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 (banditBlock !== undefined) {
+    // Say what the payoff term did — INCLUDING when it did nothing. `moved: 0` over many queries
+    // means the feature is armed and inert, which is precisely the outcome nobody would notice
+    // without this line.
+    write(renderBanditNote(banditBlock));
+  }
   if (boost !== null && shownDomain !== 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.
@@ -8041,11 +8076,16 @@ async function cmdDownloads(cwd: string, write: Write): Promise<number> {
  * The pure engine lives in harness-core; this is the thin I/O shell (scan disk + git, render).
  */
 function cmdArchitecture(options: Map<string, string>, flags: Set<string>, cwd: string, write: Write): number {
-  // Resolve repo root: prefer git toplevel, fall back to cwd (matches cmdRoam / cmdStats duality).
-  let repoRoot = cwd;
+  // Resolve repo root: an explicit --project wins, then git toplevel, then cwd (matches cmdRoam /
+  // cmdStats duality). The --project leg is the same fix as cmdProjectSkills got for field report
+  // doc-25b, applied to its neighbour before it costs a second report: this command also reads an
+  // `architecture/` tree, and on a feature-adr run against a target checkout the Step-0 product-vision
+  // сверка silently no-ops for exactly that reason.
+  const explicitRoot = (options.get('project') ?? '').trim();
+  let repoRoot = explicitRoot !== '' ? resolve(cwd, explicitRoot) : cwd;
   try {
-    repoRoot = execSync('git rev-parse --show-toplevel', { cwd, encoding: 'utf-8' }).trim() || cwd;
-  } catch { /* not a git repo — use cwd */ }
+    repoRoot = execSync('git rev-parse --show-toplevel', { cwd: repoRoot, encoding: 'utf-8', stdio: ['ignore', 'pipe', 'ignore'] }).trim() || repoRoot;
+  } catch { /* not a git repo — use the root as given */ }
 
   const manifest = loadSubsystemManifest(repoRoot);
   if (!manifest) {
@@ -8091,11 +8131,21 @@ function cmdArchitecture(options: Map<string, string>, flags: Set<string>, cwd:
  *   --stages-json   { hasManifest, design, code, qe, report } — ready to thread into prompts
  * No manifest ⇒ empty guidance strings ⇒ byte-identical run (FR-7).
  */
-function cmdProjectSkills(flags: Set<string>, cwd: string, write: Write): number {
-  let repoRoot = cwd;
+function cmdProjectSkills(options: Map<string, string>, flags: Set<string>, cwd: string, write: Write): number {
+  // `--project <dir>` names the root EXPLICITLY. Until 2026-08-25 this command took only `cwd`, so
+  // the manifest was reachable exclusively from the workspace you happened to stand in — and because
+  // the known-flag list is deliberately FLAT, `--project` passed validation on a command that never
+  // read it: exit 0, no warning, no manifest. MEASURED that day: a feature-adr run against a target
+  // repo probed with `cd REPO`, found nothing, and fell open to a generic run — the honest
+  // `polymorphism:null` was recorded and not one project lens reached any stage.
+  const explicitRoot = (options.get('project') ?? '').trim();
+  let repoRoot = explicitRoot !== '' ? resolve(cwd, explicitRoot) : cwd;
   try {
-    repoRoot = execSync('git rev-parse --show-toplevel', { cwd, encoding: 'utf-8' }).trim() || cwd;
-  } catch { /* not a git repo — use cwd */ }
+    // stderr is SWALLOWED: a non-repo directory is an ordinary case here, and letting git print
+    // "fatal: not a git repository" onto the operator's terminal made a working command look broken.
+    // The sibling cmdFeatureAdrSetup already redirects; this one did not.
+    repoRoot = execSync('git rev-parse --show-toplevel', { cwd: repoRoot, encoding: 'utf-8', stdio: ['ignore', 'pipe', 'ignore'] }).trim() || repoRoot;
+  } catch { /* not a git repo — use the root as given */ }
 
   const plan = planProjectSkills(repoRoot);
   const hasManifest = plan.injections.length > 0 || plan.skipped.length > 0;
@@ -12036,8 +12086,13 @@ function cmdCompounding(options: Map<string, string>, flags: Set<string>, cwd: s
   }
 
   const report = assembleCompoundingReport({ lessons, usage, guard, nowTs: new Date().toISOString(), evidenceLogs });
-  if (json) write(JSON.stringify({ ...report, exitCode: 0 }, null, 2));
-  else write(renderCompoundingReport(report));
+  // lesson-bandit-rerank §11: the payoff axis joins THIS report rather than growing a private
+  // dashboard — the `rewardEvents : exposureEvents` row asks exactly the question this command
+  // already asks of the reinforcement loop (is the apply leg alive, or is it a write-only log?).
+  // Read-only, and INSUFFICIENT_DATA on an absent state file — never a fake verdict.
+  const bandit = banditStats(root);
+  if (json) write(JSON.stringify({ ...report, bandit, exitCode: 0 }, null, 2));
+  else write(`${renderCompoundingReport(report)}\n\n${renderBanditHealth(bandit)}`);
   return 0;
 }
 
@@ -13674,7 +13729,7 @@ export async function runCli(argv: string[], io: CliIo = {}): Promise<number> {
       case 'architecture':
         return cmdArchitecture(options, flags, cwd, write);
       case 'project-skills':
-        return cmdProjectSkills(flags, cwd, write);
+        return cmdProjectSkills(options, flags, cwd, write);
       case 'mr-rakes':
         return await cmdMrRakes(options, flags, cwd, write);
       case 'retro':