@heretek-ai/epistemic-swarm 0.6.0 → 0.7.0

package/.claude-plugin/marketplace.json

@@ -36,6 +36,26 @@
       "category": "research",
       "source": "./plugins/research-cache",
       "homepage": "https://github.com/Heretek-AI/IUMBTEMS"
+    },
+    {
+      "name": "darkharvest",
+      "description": "Product-level competitor teardown and clean-room harvest engine.",
+      "author": {
+        "name": "Heretek AI"
+      },
+      "category": "research",
+      "source": "./plugins/darkharvest",
+      "homepage": "https://github.com/Heretek-AI/IUMBTEMS"
+    },
+    {
+      "name": "factory",
+      "description": "Coding-factory Manager loop with grill-gated phased builds and dual QA.",
+      "author": {
+        "name": "Heretek AI"
+      },
+      "category": "agents",
+      "source": "./plugins/factory",
+      "homepage": "https://github.com/Heretek-AI/IUMBTEMS"
     }
   ]
 }

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "epistemic-swarm",
-  "version": "0.4.8",
+  "version": "0.7.0",
   "description": "High-Integrity Dialectic Research Agent Harness for Claude Code enforcing empirical evidence over parametric hallucination.",
   "author": {
     "name": "Heretek AI",
@@ -69,6 +69,8 @@
     "./skills/swarm_config",
     "./skills/code_audit",
     "./skills/oss_scout",
-    "./skills/brainstorming"
+    "./skills/brainstorming",
+    "./skills/darkharvest",
+    "./skills/factory"
   ]
 }

package/MARKETPLACE.md

@@ -30,6 +30,8 @@ claude plugin marketplace list
 | **`epistemic-swarm`** | `research` | **Flagship Swarm Harness**: Multi-agent dialectic research harness pairing Agent Alpha (Thesis) against Agent Beta (Red Team), audited by an Epistemic Auditor with verbatim empirical quote validation. | `claude plugin install epistemic-swarm@heretek-official` |
 | **`socratic-grilling`** | `agents` | **Socratic Ideation**: Matt Pocock-style Socratic interrogation, premise inversion, and lateral exploration skill. | `claude plugin install socratic-grilling@heretek-official` |
 | **`research-cache`** | `research` | **Source Hasher**: Content-addressed SHA256 Markdown source hashing and verbatim quote verification engine. | `claude plugin install research-cache@heretek-official` |
+| **`darkharvest`** | `research` | **Competitor Teardown**: Product-level teardown with per-feature `depend\|vendor\|clean-room\|skip` verdicts and SPDX attribution. | `claude plugin install darkharvest@heretek-official` |
+| **`factory`** | `agents` | **Coding Factory**: Manager loop with grill-gated phases, programmer spawns, and dual QA. | `claude plugin install factory@heretek-official` |
 
 ---
 
@@ -72,6 +74,32 @@ claude plugin marketplace list
   claude plugin install research-cache@heretek-official
   ```
 
+### 4. `darkharvest` (Modular Tool)
+- **Manifest**: [`plugins/darkharvest/.claude-plugin/plugin.json`](file:///plugins/darkharvest/.claude-plugin/plugin.json)
+- **Components**:
+  - **Skill**: `skills/darkharvest` (competitor × capability teardown).
+  - **Agents**: `harvest-proponent` (per-feature verdicts), `harvest-redteam` (license/Bloat/CVE vetting).
+  - **Evals**: `plugins/darkharvest/evals/teardown-verdict` (skill fires + license-line rubric).
+- **Install**:
+  ```bash
+  claude plugin install darkharvest@heretek-official
+  ```
+
+### 5. `factory` (Modular Agent Pack)
+- **Manifest**: [`plugins/factory/.claude-plugin/plugin.json`](file:///plugins/factory/.claude-plugin/plugin.json)
+- **Components**:
+  - **Skill**: `skills/factory` (Manager loop, gates, QA bounds).
+  - **Agents**: `factory-manager`, `programmer`, `qa-functional`, `qa-adversarial`.
+  - **Evals**: `plugins/factory/evals/gate-halt` (skill fires + gates-first rubric).
+- **Install**:
+  ```bash
+  claude plugin install factory@heretek-official
+  ```
+
+### Flagship agents & evals
+- **Agents** (repo-root `agents/`): `alpha-thesis`, `beta-antithesis`, `epistemic-auditor` — condensed from `prompts/agent_alpha_thesis.md`, `prompts/agent_beta_antithesis.md`, `prompts/epistemic_auditor.md`.
+- **Evals** (repo-root `evals/`): `grill-fires`, `darkharvest-fires`, `factory-gate` — each `prompt.md` plus `tool_used: Skill` and `llm` graders. Run `claude plugin eval .` (billable model calls); CI gates via `.github/workflows/plugin-evals.yml` on release/dispatch.
+
 ---
 
 ## 💡 Lateral Brainstorming (`/brainstorming`)

package/install.sh

@@ -21,7 +21,7 @@ echo "✅ Core prerequisites detected (Python $(python3 --version | cut -d' ' -f
 mkdir -p "$CLAUDE_DIR/skills"
 
 echo "🔗 Linking skills into $CLAUDE_DIR/skills/..."
-for skill in grilling research_cache epistemic_search swarm_config code_audit oss_scout brainstorming; do
+for skill in grilling research_cache epistemic_search swarm_config code_audit oss_scout brainstorming darkharvest factory; do
   # legacy research-cache dir name kept as alias for older configs
   ln -sfn "$REPO_DIR/skills/$skill" "$CLAUDE_DIR/skills/$skill"
   echo "   - $CLAUDE_DIR/skills/$skill -> $REPO_DIR/skills/$skill"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@heretek-ai/epistemic-swarm",
-  "version": "0.6.0",
+  "version": "0.7.0",
   "description": "IUMBTEMS: I Use My Brain To Express My Self — High-Integrity Dialectic Research Agent Harness for Claude Code, OpenCode V2, Pi, OMP (oh-my-pi), Gemini CLI, Codex CLI, and AntiGravity",
   "main": "bin/cli.js",
   "bin": {

package/plugins/darkharvest/.claude-plugin/plugin.json

@@ -0,0 +1,15 @@
+{
+  "name": "darkharvest",
+  "version": "0.1.0",
+  "description": "Product-level competitor teardown and clean-room harvest engine for Claude Code.",
+  "author": {
+    "name": "Heretek AI",
+    "email": "dev@heretek.ai"
+  },
+  "license": "Apache-2.0",
+  "repository": "https://github.com/Heretek-AI/IUMBTEMS",
+  "homepage": "https://github.com/Heretek-AI/IUMBTEMS#readme",
+  "skills": [
+    "./skills/darkharvest"
+  ]
+}

package/plugins/darkharvest/agents/harvest-proponent.md

@@ -0,0 +1,22 @@
+---
+name: harvest-proponent
+description: Competitor-teardown proponent. Inventories what a competitor does well and proposes per-feature harvest verdicts. Use per competitor scope in darkharvest runs.
+model: sonnet
+---
+
+You are the Harvest Proponent in the darkharvest plugin.
+
+Per assigned competitor: build the capability inventory (present / missing /
+partial per capability with `[VERIFIED: <hash>]` evidence from
+`.research/sources/<sha256>.md`), then propose per-feature verdicts:
+`depend | vendor | clean-room-rebuild | skip(reason)`, ranked by
+Impact x Effort x Differentiation.
+
+Legal guardrails (final): permissive licenses only (MIT, Apache-2.0, BSD, ISC)
+may be `depend`/`vendor`. GPL / AGPL / UNKNOWN license means
+`clean-room-rebuild` spec only — never copy code. Workflows clonable;
+copy-text, UI assets, and brand are never copied. Every `vendor` item emits an
+SPDX attribution block (license + upstream URL + files). Closed targets get
+metadata-only rows plus `[NEGATIVE_KNOWLEDGE: <query>]`, never a failed run.
+
+Full specification: `skills/darkharvest/SKILL.md` in the plugin root.

package/plugins/darkharvest/agents/harvest-redteam.md

@@ -0,0 +1,20 @@
+---
+name: harvest-redteam
+description: Competitor-teardown red team. Vets license contamination, bloat, CVEs, and staleness; maps white-space gaps both ways. Use per competitor scope alongside the proponent.
+model: sonnet
+---
+
+You are the Harvest Red Team in the darkharvest plugin.
+
+Per assigned competitor: determine SPDX from LICENSE file plus package
+metadata; flag transitive bloat, CVE/takeover surface, and staleness
+(`STALE` when inactive over 12 months or solo-maintained). Gates are
+warn-only — badge inline per matrix cell plus a risks section, never
+auto-skip. Map white-space gaps in BOTH directions: what the competitor lacks
+that we own or could own, and what we lack.
+
+Challenge every proponent harvest proposal: any `vendor` verdict on
+copyleft or unknown-licensed code must be rewritten as `clean-room-rebuild`
+with reasoning. Benchmarks are optional but, when present, must be VERIFIED.
+
+Full specification: `skills/darkharvest/SKILL.md` in the plugin root.

package/plugins/darkharvest/evals/teardown-verdict/graders/license-line.md

@@ -0,0 +1,6 @@
+---
+type: llm
+---
+
+PASS if the reply distinguishes what may be borrowed directly (permissive licenses) from what must be re-implemented clean-room (copyleft or unknown licenses), or asks which seed competitors to tear down before verdicts.
+FAIL if the reply recommends vendoring GPL/AGPL-licensed code, copying UI assets, or gives harvest verdicts with no license reasoning at all.

package/plugins/darkharvest/evals/teardown-verdict/graders/skill-fired.md

@@ -0,0 +1,5 @@
+---
+type: tool_used
+tool: Skill
+input_match: '"skill"\s*:\s*"(?:[\w-]+:)?darkharvest"'
+---

package/plugins/darkharvest/evals/teardown-verdict/prompt.md

@@ -0,0 +1,6 @@
+---
+max_turns: 10
+allowed_tools: [Read, Glob, Grep, Skill]
+---
+
+I'm building a competitor to a desktop app that orchestrates multiple coding agents. What do similar open-source projects do well, and what could I borrow clean-room?

package/plugins/darkharvest/skills/darkharvest/SKILL.md

@@ -0,0 +1,70 @@
+---
+name: darkharvest
+description: Product-level competitor teardown and clean-room harvest engine. Use when user wants to compete with or learn from existing products (e.g. Paseo, OpenChambers). Seed with inspiration URLs plus prompt, auto-expand to adjacents, clone-scan competitors, compare product plus code, and emit per-feature depend/vendor/clean-room/skip verdicts with SPDX attribution. Never copies GPL/AGPL code or UI assets.
+---
+
+# Darkharvest — Competitor Teardown & Clean-Room Harvest
+
+Product-level teardown, not library scouting. `oss_scout` answers "which Raft lib do I depend on"; `darkharvest` answers "I'm building a Paseo competitor — what do Paseo-likes do well, what white space exists, and what may I harvest clean-room?"
+
+## 1. Invocation
+
+```bash
+# Full dialectic teardown swarm (mock = zero token cost)
+python3 runner/research_swarm.py --mode darkharvest --objective "Paseo-class agent harness competitor" --mock-claude
+
+# Competitor fetch helper (read-only scan, allowlisted, capped)
+python3 skills/darkharvest/scripts/harvest.py --repo https://github.com/owner/repo --show-context
+
+# Via CLI
+iumbtems darkharvest "Paseo-class agent harness competitor" --seeds https://github.com/a/b,https://github.com/c/d --max-repos 6 --mock-claude
+```
+
+In OpenCode V2: `iumbtems_darkharvest` tool. Slash: `/darkharvest <objective>`.
+In Pi / OMP: `/darkharvest <objective>`. In Gemini: skill auto-activates on competitor-teardown intent.
+
+## 2. Input contract (hybrid, merged)
+
+1. Seeds: GitHub/GitLab URLs plus free-text prompt guidance. Unlimited seeds accepted; runner caps at `--max-repos` (default 10, recommended 6 for live runs).
+2. Local baseline: full `brainstorm.py` domain model (tree + README + stack + TODOs + git log/status). Manual prompt ADDS to auto facts; conflicts record both, manual tagged `[HYPOTHESIS: <how to check>]`.
+3. Gaps: open-loops + user wishlist combined.
+
+## 3. Discovery (seed + expand)
+
+- Seeds + 5–8 auto adjacents; direct competitors and adjacent inspirations SPLIT in the report.
+- Similarity: hybrid pre-filter (README topics + feature keywords + dep overlap) then LLM rerank. Never LLM-vibe alone.
+- Rank relevance first, stars second. Any stack allowed with porting-effort note. Stale (>12mo or solo-maintainer) flagged with `⚠️ STALE`, never auto-skipped.
+
+## 4. Fetch & scan budgets (defaults, all flag-overridable)
+
+- Defaults: `--max-repos 10 --depth 3 --per-repo-mb 100 --per-repo-timeout 300s`. Prefer `--max-repos 6 --depth 2` for live runs.
+- `harvest.py` clones to temp then drops; allowlist scan only: tree + README + LICENSE + manifests + key source headers. Skips `.git, node_modules, dist, build, target, .next, .venv, .research`. Caps file count and bytes — full dump without caps will OOM and is banned.
+- Closed or non-cloneable targets: try Firecrawl/docs fetch; when unavailable emit metadata-only row + `[NEGATIVE_KNOWLEDGE: <query>]`, never fail the run.
+- Evidence: every matrix cell needs a SHA-256 cached source (`.research/sources/<sha256>.md`) with verbatim quote, or a `file://<path>#L<start>-L<end>` pointer. Unverified cells are purged to NEGATIVE_KNOWLEDGE.
+
+## 5. Dialectic roles (one scope per competitor)
+
+- Orchestrator decomposes the candidate list into one scope per competitor (DAG default, auction optional).
+- Alpha (harvest proponent): what competitor does well + per-feature harvest proposals.
+- Beta (red-team): license contamination, transitive bloat, CVE/takeover surface, staleness, both-ways missing (white space).
+- Auditor: enforces strict-cells bar, warn-only gates surface as inline `⚠️ LICENSE/CVE/STALE` badges plus a risks section.
+
+## 6. Legal guardrails (final)
+
+- Permissive only (MIT / Apache-2.0 / BSD / ISC) may be `depend` or `vendor`.
+- GPL / AGPL / unknown license → `clean-room-rebuild` spec only, never copy.
+- Workflows clonable; copy-text, UI assets, brand never copied.
+- Every vendored item emits an SPDX attribution block (license + upstream URL + files).
+
+## 7. Output
+
+- `.research/darkharvest_report.md`: competitor × capability matrix (present / missing / partial + evidence) + white-space gaps both directions + harvest backlog (per-feature `depend|vendor|clean-room-rebuild|skip(reason)`, Impact×Effort×Differentiation ranked, top 5) + risks + epistemic audit totals.
+- Machine dossier: `.research/darkharvest_dossier.json` for rerun diffs (changed cells + new/removed candidates).
+- Errors: structured log + NEGATIVE_KNOWLEDGE, never silent. Writes only under `.research/`.
+
+## 8. Anti-patterns (hard bans)
+
+- No auto-install of dependencies, no writes outside `.research/`.
+- No pixel-level UX cloning, no GPL/AGPL vendoring.
+- No parametric repo claims as `[VERIFIED]` — metadata suffices only for rank hints, never for harvest verdicts.
+- No unbounded expansion: seeds + expansion must respect `--max-repos` and per-repo caps.

package/plugins/darkharvest/skills/darkharvest/scripts/harvest.py

@@ -0,0 +1,252 @@
+#!/usr/bin/env python3
+"""
+Darkharvest fetch helper: clone-to-temp + allowlisted scan + SHA-256 cache.
+
+Read-only. Never installs dependencies. Never writes outside .research/
+(except the temp clone, which is dropped). Enforces per-repo caps so
+`--max-repos 10 --depth 3` style defaults cannot OOM the host.
+
+Usage:
+  python3 skills/darkharvest/scripts/harvest.py --repo <url> [--show-context]
+  python3 skills/darkharvest/scripts/harvest.py --repo <url> --export <path>
+
+Closed / non-cloneable targets: prints a metadata-only stub with a
+[NEGATIVE_KNOWLEDGE] gap note instead of failing.
+"""
+
+import argparse
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+from datetime import datetime, timezone
+from pathlib import Path
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent.parent
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+from skills.research_cache.hasher import SourceHasher  # noqa: E402
+
+SKIP_DIRS = {
+    ".git",
+    "node_modules",
+    "__pycache__",
+    ".venv",
+    "venv",
+    "dist",
+    "build",
+    "target",
+    ".next",
+    ".research",
+    "coverage",
+}
+MANIFEST_FILES = [
+    "package.json",
+    "pyproject.toml",
+    "Cargo.toml",
+    "go.mod",
+    "requirements.txt",
+    "LICENSE",
+    "LICENSE.md",
+    "README.md",
+]
+TEXT_EXTS = {".md", ".json", ".toml", ".yaml", ".yml", ".txt", ".py", ".js", ".ts"}
+
+DEFAULT_MAX_FILES = 120
+DEFAULT_MAX_BYTES = 400_000
+DEFAULT_TIMEOUT = 120
+
+SPDX_MAP = {
+    "mit": "MIT",
+    "apache-2.0": "Apache-2.0",
+    "apache 2.0": "Apache-2.0",
+    "bsd-3-clause": "BSD-3-Clause",
+    "bsd 3-clause": "BSD-3-Clause",
+    "isc": "ISC",
+    "gpl": "GPL",
+    "agpl": "AGPL",
+    "mpl": "MPL",
+    "lgpl": "LGPL",
+}
+
+
+def sh(cmd, cwd=None, timeout=DEFAULT_TIMEOUT):
+    try:
+        r = subprocess.run(
+            cmd,
+            capture_output=True,
+            text=True,
+            cwd=str(cwd or PROJECT_ROOT),
+            timeout=timeout,
+        )
+        return r.returncode, (r.stdout or "").strip()
+    except Exception as exc:  # noqa: BLE001
+        return 1, f"error: {exc}"
+
+
+def detect_license(text):
+    low = (text or "").lower()
+    for key, spdx in SPDX_MAP.items():
+        if key in low:
+            return spdx
+    return "UNKNOWN"
+
+
+def scan_tree(root, max_files=DEFAULT_MAX_FILES, max_bytes=DEFAULT_MAX_BYTES):
+    entries = []
+    total_bytes = 0
+    manifests = {}
+    license_text = ""
+    readme_head = ""
+    for dirpath, dirnames, filenames in os.walk(root):
+        dirnames[:] = [d for d in dirnames if d not in SKIP_DIRS]
+        rel = Path(dirpath).relative_to(root)
+        depth = len(rel.parts) if str(rel) != "." else 0
+        if depth > 3:
+            dirnames[:] = []
+            continue
+        for f in sorted(filenames):
+            if len(entries) >= max_files or total_bytes >= max_bytes:
+                return entries, manifests, license_text, readme_head, True
+            fp = Path(dirpath) / f
+            try:
+                size = fp.stat().st_size
+            except OSError:
+                continue
+            if size > 100_000:
+                continue
+            rel_p = str(rel / f) if str(rel) != "." else f
+            entries.append(rel_p)
+            if f in MANIFEST_FILES or Path(f).suffix in TEXT_EXTS:
+                try:
+                    content = fp.read_text(encoding="utf-8", errors="replace")
+                except OSError:
+                    continue
+                total_bytes += len(content.encode("utf-8", errors="replace"))
+                if f in MANIFEST_FILES and f not in manifests:
+                    manifests[f] = content[:4000]
+                if f.lower().startswith("license") and not license_text:
+                    license_text = content[:4000]
+                if f.lower() == "readme.md" and not readme_head:
+                    readme_head = content[:4000]
+    return entries, manifests, license_text, readme_head, False
+
+
+def harvest_repo(
+    url,
+    max_files=DEFAULT_MAX_FILES,
+    max_bytes=DEFAULT_MAX_BYTES,
+    timeout=DEFAULT_TIMEOUT,
+    base_dir=".research",
+):
+    started = datetime.now(timezone.utc).isoformat()
+    tmp = tempfile.mkdtemp(prefix="darkharvest-")
+    try:
+        code, out = sh(["git", "clone", "--depth", "1", url, tmp], timeout=timeout)
+        if code != 0:
+            return {
+                "url": url,
+                "status": "metadata-only",
+                "generated_at": started,
+                "negative_knowledge": (
+                    f"[NEGATIVE_KNOWLEDGE: clone failed for {url}; "
+                    f"used metadata-only row. Detail: {out[:200]}]"
+                ),
+            }
+        entries, manifests, license_text, readme_head, truncated = scan_tree(
+            Path(tmp), max_files=max_files, max_bytes=max_bytes
+        )
+        spdx = detect_license(license_text + manifests.get("package.json", ""))
+        harvestable = (
+            "depend-or-vendor"
+            if spdx in ("MIT", "Apache-2.0", "BSD-3-Clause", "ISC")
+            else "clean-room-rebuild-only"
+        )
+        hasher = SourceHasher(Path(base_dir))
+        cache_blob = (
+            f"# {url}\n\nSPDX: {spdx}\n\n"
+            f"## README head\n{readme_head[:2000]}\n\n"
+            f"## Tree sample ({len(entries)} entries)\n" + "\n".join(entries[:60])
+        )
+        source_hash = hasher.store_source(url, cache_blob, f"Darkharvest {url}")
+        warnings = []
+        if spdx in ("GPL", "AGPL"):
+            warnings.append(
+                "⚠️ LICENSE: strong copyleft — spec rebuild only, never vendor"
+            )
+        elif spdx == "UNKNOWN":
+            warnings.append("⚠️ LICENSE: unknown — treat as clean-room-rebuild-only")
+        if truncated:
+            warnings.append(
+                "⚠️ SCAN: truncated by file/byte caps; treat as partial evidence"
+            )
+        return {
+            "url": url,
+            "status": "scanned",
+            "generated_at": started,
+            "spdx": spdx,
+            "harvest_policy": harvestable,
+            "tree_entries": len(entries),
+            "truncated": truncated,
+            "source_hash": source_hash,
+            "readme_head": readme_head[:1500],
+            "manifests_seen": sorted(manifests.keys()),
+            "warnings": warnings,
+        }
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Darkharvest competitor fetch helper")
+    ap.add_argument("--repo", required=True, help="Competitor repo URL to scan")
+    ap.add_argument("--show-context", action="store_true")
+    ap.add_argument("--export", default="")
+    ap.add_argument("--max-files", type=int, default=DEFAULT_MAX_FILES)
+    ap.add_argument("--max-bytes", type=int, default=DEFAULT_MAX_BYTES)
+    ap.add_argument("--timeout", type=int, default=DEFAULT_TIMEOUT)
+    ap.add_argument("--dir", default=".research")
+    args = ap.parse_args()
+
+    if not re.match(r"^https?://", args.repo):
+        print(f"WARN: {args.repo} is not an http(s) URL; metadata-only stub emitted.")
+        result = {
+            "url": args.repo,
+            "status": "metadata-only",
+            "generated_at": datetime.now(timezone.utc).isoformat(),
+            "negative_knowledge": (
+                f"[NEGATIVE_KNOWLEDGE: non-cloneable target {args.repo}]"
+            ),
+        }
+    else:
+        result = harvest_repo(
+            args.repo,
+            max_files=args.max_files,
+            max_bytes=args.max_bytes,
+            timeout=args.timeout,
+            base_dir=args.dir,
+        )
+
+    if args.show_context or not args.export:
+        print(f"\n🌑 [Darkharvest] {result['url']} → {result['status']}")
+        for key in ("spdx", "harvest_policy", "tree_entries", "source_hash"):
+            if key in result:
+                print(f"  {key}: {result[key]}")
+        for w in result.get("warnings", []):
+            print(f"  {w}")
+        if result.get("negative_knowledge"):
+            print(f"  {result['negative_knowledge'][:200]}")
+
+    if args.export:
+        out = Path(os.path.realpath(args.export))
+        out.parent.mkdir(parents=True, exist_ok=True)
+        out.write_text(json.dumps(result, indent=2), encoding="utf-8")
+        print(f"✅ Harvest scan exported to {out}")
+
+
+if __name__ == "__main__":
+    main()

package/plugins/factory/.claude-plugin/plugin.json

@@ -0,0 +1,15 @@
+{
+  "name": "factory",
+  "version": "0.1.0",
+  "description": "Coding-factory Manager loop with grill-gated phased builds and dual QA for Claude Code.",
+  "author": {
+    "name": "Heretek AI",
+    "email": "dev@heretek.ai"
+  },
+  "license": "Apache-2.0",
+  "repository": "https://github.com/Heretek-AI/IUMBTEMS",
+  "homepage": "https://github.com/Heretek-AI/IUMBTEMS#readme",
+  "skills": [
+    "./skills/factory"
+  ]
+}

package/plugins/factory/agents/factory-manager.md

@@ -0,0 +1,22 @@
+---
+name: factory-manager
+description: Coding-factory Manager. Owns grill gates, swarm dispatch, roadmap synthesis, and QA tiebreaks. Never writes code. Use to drive /factory runs.
+model: sonnet
+---
+
+You are the Factory Manager in the factory plugin.
+
+Loop: grill the user until `.factory/frontier.json` is settled (max 5
+brainstorm+darkharvest swarm cycles per gate); explicit user approve advances
+each gate. Per gate, dispatch both swarms (mock-first on fixtures), then
+synthesize `.roadmap/<phase>/` as GOAL.md plus dossier.json
+(goal/evidence/acceptance/brief/verdict/hashes — every claim needs a VERIFIED
+hash or file pointer). Spawn the programmer subagent per phase with the phase
+dossier (cite phase hashes). Run qa-a and qa-b per phase; track retries with
+`skills/factory/scripts/factory.py` (3 failures escalate back to you with both
+QA reports). Tiebreak QA disagreements; require explicit user sign-off per phase.
+
+Never write implementation code yourself. Never bypass gates outside an
+active `/domainexpansion` count. Never accept unverified harvest verdicts.
+
+Full specification: `skills/factory/SKILL.md` in the plugin root.

package/plugins/factory/agents/programmer.md

@@ -0,0 +1,16 @@
+---
+name: programmer
+description: Factory implementer. Implements exactly one phase brief per spawn and cites phase evidence hashes. Never invokes swarms or other programmers. Use per .roadmap phase.
+model: sonnet
+---
+
+You are the Factory Programmer in the factory plugin.
+
+You receive one phase dossier (GOAL.md + dossier.json). Implement exactly what
+the brief specifies — no scope expansion, no drive-by refactors. Cite the
+phase evidence hashes for any borrowed pattern. If the brief is ambiguous or
+its acceptance criteria are untestable, stop and return questions to the
+manager instead of guessing. Never invoke brainstorm/darkharvest swarms, never
+spawn other agents, never edit outside the phase scope.
+
+Full specification: `skills/factory/SKILL.md` in the plugin root.

package/plugins/factory/agents/qa-adversarial.md

@@ -0,0 +1,17 @@
+---
+name: qa-adversarial
+description: Factory QA, adversarial seat. Hunts edge cases, regressions, and acceptance loopholes the functional pass missed. Read-only plus test execution. Use alongside qa-functional per phase.
+model: sonnet
+---
+
+You are QA-B (adversarial seat) in the factory plugin.
+
+Assume the functional pass missed something. Attack the phase: edge-case
+inputs, regressions in neighboring behavior, acceptance criteria that are
+vacuously satisfiable, error paths, and resource/timeout boundaries. Read-only
+plus test execution — never edit code. Verdicts: `pass | fail(reason) |
+conditional(note)`, each with a reproduction. Your prompt is deliberately
+diverged from qa-functional: it proves what works, you hunt what breaks.
+Disagreements go to the factory-manager tiebreak.
+
+Full specification: `skills/factory/SKILL.md` in the plugin root.

package/plugins/factory/agents/qa-functional.md

@@ -0,0 +1,17 @@
+---
+name: qa-functional
+description: Factory QA, functional seat. Verifies phase acceptance criteria pass on the real surface. Read-only plus test execution. Use alongside qa-adversarial per phase.
+model: sonnet
+---
+
+You are QA-A (functional seat) in the factory plugin.
+
+Verify each acceptance criterion in the phase dossier by executing it on the
+real surface (run the tests, exercise the feature, inspect the output).
+Read-only plus test execution — never edit code to make a check pass. Verdicts:
+`pass | fail(reason) | conditional(note)`, each tied to a specific criterion.
+Record via the manager; three failures on one phase escalate it back to the
+manager with both QA reports attached. Your prompt is deliberately diverged
+from qa-adversarial: you prove what works, it hunts what breaks.
+
+Full specification: `skills/factory/SKILL.md` in the plugin root.

package/plugins/factory/evals/gate-halt/graders/gates-first.md

@@ -0,0 +1,6 @@
+---
+type: llm
+---
+
+PASS if the reply proposes phased scope with explicit approval gates (or asks gate/phase-clarifying questions first) instead of writing implementation code immediately.
+FAIL if the reply starts implementing code with no phase breakdown and no user sign-off step.

package/plugins/factory/evals/gate-halt/graders/skill-fired.md

@@ -0,0 +1,5 @@
+---
+type: tool_used
+tool: Skill
+input_match: '"skill"\s*:\s*"(?:[\w-]+:)?factory"'
+---

package/plugins/factory/evals/gate-halt/prompt.md

@@ -0,0 +1,6 @@
+---
+max_turns: 10
+allowed_tools: [Read, Glob, Grep, Skill]
+---
+
+I want to build a small session-handoff feature across agent harnesses, broken into reviewable phases with QA on each phase. Set up the gated build loop.