diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
index 3583dc9..d10262a 100644
--- a/.claude-plugin/marketplace.json
+++ b/.claude-plugin/marketplace.json
@@ -11,7 +11,7 @@
       "name": "data-engineering-knowledge-capture",
       "source": "./",
       "description": "Capture data-platform knowledge \u2014 schemas, lineage, medallion, semantic layer, business glossary \u2014 into OKF. Extends PKC; depends on OKF.",
-      "version": "0.5.1",
+      "version": "0.5.2",
       "author": {
         "name": "Rick Hightower"
       },
diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index aac5654..efd040f 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "data-engineering-knowledge-capture",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "Data Engineering Knowledge Capture (DEKC) \u2014 extends PKC + OKF for schemas, lineage, medallion layers, SQL/DAX, semantic models, business objects, and an indexed second brain. Agents walk data lakes. Works in Claude Code, Grok Build, Codex, and OpenCode.",
   "author": {
     "name": "Rick Hightower",
diff --git a/.codex-plugin/plugin.json b/.codex-plugin/plugin.json
index 58df265..af17f5f 100644
--- a/.codex-plugin/plugin.json
+++ b/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "data-engineering-knowledge-capture",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "Data Engineering Knowledge Capture (DEKC) \u2014 schemas, lineage, medallion layers, semantic models, business glossary. Extends PKC; depends on OKF. Codex-native skills + hooks port.",
   "author": {
     "name": "Rick Hightower",
diff --git a/.cursor-plugin/plugin.json b/.cursor-plugin/plugin.json
index 399c5be..a30822d 100644
--- a/.cursor-plugin/plugin.json
+++ b/.cursor-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "data-engineering-knowledge-capture",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "Data Engineering Knowledge Capture: schemas, lineage, medallion, semantic models, and glossary into a durable OKF knowledge graph. Multi-host bindings and write isolation.",
   "author": {
     "name": "Rick Hightower"
diff --git a/.grok-plugin/marketplace.json b/.grok-plugin/marketplace.json
index a4a7ddb..9f0341b 100644
--- a/.grok-plugin/marketplace.json
+++ b/.grok-plugin/marketplace.json
@@ -1,13 +1,13 @@
 {
   "name": "dekc-plugin-marketplace",
   "description": "Optional native Grok marketplace metadata. Grok Build loads Claude plugins with zero config; this pins identity for listings.",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "plugins": [
     {
       "name": "data-engineering-knowledge-capture",
       "source": ".",
       "description": "DEKC \u2014 data engineering knowledge capture on PKC + OKF. Claude-compatible.",
-      "version": "0.5.1",
+      "version": "0.5.2",
       "compatibility": {
         "claude_plugin": true,
         "zero_config": true
diff --git a/.opencode/plugin/dekc.json b/.opencode/plugin/dekc.json
index 838184f..9fc5b17 100644
--- a/.opencode/plugin/dekc.json
+++ b/.opencode/plugin/dekc.json
@@ -1,6 +1,6 @@
 {
   "name": "data-engineering-knowledge-capture",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "OpenCode port of DEKC \u2014 same skills/agents/scripts as Claude & Codex. Policy via AGENTS.md.",
   "skills": "../../skills",
   "agents": "../../agents",
diff --git a/AGENTS.md b/AGENTS.md
index 0d9dd53..598f61c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -79,6 +79,16 @@ Plugin root: `${CLAUDE_PLUGIN_ROOT}`.
 | **data-lake-walker** | Default orchestrator: walk → produce → adversarial grade → index |
 | **reverse-engineering-orchestrator** | Multi-cloud RE (Fabric/AWS/GCP), strict LoopPolicy + fan-out |
 
+## Query-time retrieval
+
+Orthogonal to reverse engineering. Search, score, pack, and deepen stay in the sub-agent. The parent sees a **summary card only**.
+
+| Agent | Role |
+|-------|------|
+| **data-retriever** | Query-time data-plane context. Spawn via `/dekc-retrieve`. Never dump hit lists or pack markdown into the parent. |
+
+When the question also needs project decisions or system topology, fan out PKC **knowledge-retriever** and/or SAC **architecture-retriever** in parallel. Do not run `dekc_search.py` / `dekc_pack.py` / `dekc_brain.py` in the parent for Q&A.
+
 ## Producer workers
 
 | Agent | Role |
@@ -131,7 +141,8 @@ python3 tests/test_dekc.py
 | medallion health | layer-auditor / dekc-doctor |
 | wireframe / architecture / ERD / job diagrams | dekc-diagram / dekc_diagram.py |
 | lake / mart / catalog / stream / DQ / ingestion job | dekc-platform / dekc_platform.py |
-| search the second brain | dekc-search / dekc-index |
+| retrieve / what's this table / metric / impact Q&A | data-retriever / dekc-retrieve |
+| search the second brain (raw hits, not parent Q&A) | dekc-search / dekc-index |
 | multi-agent loop authoring | okf-agent-graph `/ager-*` + DEKC KnowledgeBind |
 
 <!-- worklog:policy:start -->
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2c99225..9a1ab76 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,27 @@
 
 ## Unreleased
 
+## 0.5.2 — 2026-09-12
+
+Query-time retrieval sub-agent, matching the PKC/SAC parity pattern: search +
+score + pack + deepen stay off the parent; the parent gets a summary card only.
+
+### Added
+
+- **`data-retriever`** agent — retrieval-only. No capture, walk, reverse
+  engineer, or knowledge writes. Returns a retrieval card (seed, fit, engine,
+  pack stats, lead nodes, lineage note, gaps, next).
+- **`/dekc-retrieve`** skill + slash command. Parent must spawn `data-retriever`
+  and must not run `dekc_search.py` / `dekc_pack.py` in-process for Q&A.
+  Orthogonal fan-out: PKC `knowledge-retriever`, SAC `architecture-retriever`.
+- `dekc_pack.py --summary` — compact card-friendly output, bodies off (root
+  included), fail-closed token budget. Retriever prefers `--tiny --summary`.
+
+### Notes
+
+- Reverse-engineering agents (`data-lake-walker`, `lineage-tracer`, …) are
+  unchanged. Query-time is a separate section on the RE orchestrator.
+
 ## 0.5.1 — 2026-08-31
 
 Correctness patch for catalog rendering and the rg-backed reverse index. Both
diff --git a/README.md b/README.md
index a1bf393..704aaa4 100644
--- a/README.md
+++ b/README.md
@@ -8,7 +8,7 @@ DEKC **extends [Project Knowledge Capture (PKC)](https://github.com/SpillwaveSol
 |---|---|
 | **Plugin name** | `data-engineering-knowledge-capture` |
 | **Repo** | [SpillwaveSolutions/data-engineering-knowledge-capture](https://github.com/SpillwaveSolutions/data-engineering-knowledge-capture) |
-| **Version** | 0.5.1 |
+| **Version** | 0.5.2 |
 | **License** | MIT |
 | **Hosts** | Claude Code · Grok Build · Codex · OpenCode · Agent Plugins 1.0 · Grok Bot · LangChain Deep Agents |
 
@@ -208,12 +208,14 @@ python3 scripts/dekc_doctor.py --repo . --bundle knowledge
 python3 scripts/dekc_search.py "revenue" --repo . --bundle knowledge
 ```
 
-Slash / skill entry points: `/dekc-init` · `/dekc-walk` · `/dekc-lineage` · `/dekc-business-object` · `/dekc-glossary` · `/dekc-semantic` · `/dekc-context` · `/dekc-search` · `/dekc-index` · `/dekc-doctor`
+Slash / skill entry points: `/dekc-init` · `/dekc-walk` · `/dekc-lineage` · `/dekc-business-object` · `/dekc-glossary` · `/dekc-semantic` · `/dekc-retrieve` · `/dekc-context` · `/dekc-search` · `/dekc-index` · `/dekc-doctor`
 
 ## Agent loop (AGER-shaped): producers + adversarial judges
 
 **Orchestrators:** `data-lake-walker` (default) · `reverse-engineering-orchestrator` (Fabric/AWS/GCP).
 
+**Query-time:** `data-retriever` (`/dekc-retrieve`) — search + pack stay in the sub-agent; the parent gets a summary card only. Not an RE worker.
+
 Producer workers fan out, then **adversarial skeptics grade reverse engineering with rubrics** before index/pack. Lead judge: **re-adversary-judge** (threshold **0.75**). Fail → re-plan or **retract** unproven claims — never invent edges to pass.
 
 ```text
diff --git a/agents/data-retriever.md b/agents/data-retriever.md
new file mode 100644
index 0000000..3b29d68
--- /dev/null
+++ b/agents/data-retriever.md
@@ -0,0 +1,72 @@
+---
+name: data-retriever
+description: Retrieve DEKC data-plane context without contaminating the parent. Use for Table, Metric, LineagePath, IngestionJob, Transformation, Dashboard, DataProduct, GlossaryTerm, BusinessObject queries. Search, score fit, pack, optionally deepen/lineage. Returns a summary card only.
+---
+
+You are **Data Retriever** — the query-time retrieval sub-agent for DEKC (Data Engineering Knowledge Capture).
+
+You are **not** a reverse-engineering agent. Do not confuse this role with `data-lake-walker`, `lineage-tracer`, `schema-scout`, or other capture/RE workers. You read the knowledge graph. You do not grow it.
+
+This is the PKC `knowledge-retriever` / SAC `architecture-retriever` parity role for the data plane.
+
+## Contract
+
+- **Retrieval-only.** No capture, walk, reverse-engineer, promote, materialize, index rebuild, or knowledge writes. No `brain_session`.
+- **Never** return full hit lists or full pack markdown to the parent.
+- Search, score, pack, and deepen stay **inside this sub-agent**.
+- The parent receives a **summary card only**.
+- Fiction samples only in public trees (Northstar / Lumenfield). Never name a private remote.
+
+## Card shape (only thing the parent sees)
+
+```markdown
+## Retrieval card
+- Query: …
+- Seed: `/path` (`Type`) — why chosen
+- Fit: high|medium|low — one sentence
+- Engine: index|rg|scan
+- Pack: hops=N nodes=N tokens=N/budget
+- Lead nodes: (5–8 bullets)
+- Lineage note: (upstream/downstream one-liner) or none
+- Open gaps: or none
+- Next: stay|deepen-2hop|lineage|try-alt-seed `/other`
+```
+
+## Scripts (read-only)
+
+```bash
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_search.py" "<query>" --repo . --bundle knowledge --limit 5
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_search.py" "<query>" --type Table,Metric,LineagePath,DataProduct --limit 5 --json
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" <seed> --repo . --bundle knowledge --tiny --summary
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" <seed> --repo . --bundle knowledge --hops 2 --summary
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_lineage.py" --repo . --bundle knowledge upstream <seed> --hops 2
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_lineage.py" --repo . --bundle knowledge downstream <seed> --hops 2
+```
+
+Prefer `--summary` on every pack. Fall back to `--tiny` without `--summary` only if the flag is missing. Do **not** run `dekc_brain.py` for the parent — those packs stay here if you need an intent hint, and they still must not leak.
+
+Forbidden: `dekc_capture.py`, `dekc_walk.py`, `dekc_lineage.py materialize`, `dekc_business.py`, `dekc_pack.py --write`, `brain_session.py`.
+
+## Workflow
+
+1. **Seed or search.** If the caller already named a concept path, use it. Otherwise `dekc_search.py --limit 5`.
+2. **Score fit** (no LLM in Python — you score). Prefer **Table / Metric / LineagePath / DataProduct**. Then IngestionJob, Transformation, Dashboard, BusinessObject, GlossaryTerm.
+   - **high** — preferred type and title/path match the question
+   - **medium** — preferred type *or* a strong keyword hit
+   - **low** — glossary-only, wrong layer, or weak lexical match
+3. **Pack** `--tiny --summary`. Bodies stay off. Read hops / nodes / tokens / engine / lead nodes from the summary.
+4. **Deepen or lineage** only if the card would otherwise be thin, or the question is impact / blast-radius. Use `--hops 2 --summary` or the lineage upstream/downstream helpers. **Max 2 deepen steps.**
+5. Return the **card only**. One card. No appendix.
+
+## Fit and Next
+
+- `stay` — seed answers the question
+- `deepen-2hop` — neighborhood is thin; parent may re-spawn you with a 2-hop ask
+- `lineage` — impact / upstream / downstream still open
+- `try-alt-seed /other` — better seed exists; name it
+
+Open gaps: missing gold BO, no landing job, broken link, empty search. Write `none` when the pack is enough.
+
+## Orthogonal fan-out
+
+If the question also needs project decisions or system topology, tell the parent to spawn PKC **knowledge-retriever** and/or SAC **architecture-retriever** in parallel. Do not impersonate those agents. Do not pack their nouns.
diff --git a/agents/reverse-engineering-orchestrator.md b/agents/reverse-engineering-orchestrator.md
index 7ef7dc7..d11581e 100644
--- a/agents/reverse-engineering-orchestrator.md
+++ b/agents/reverse-engineering-orchestrator.md
@@ -88,3 +88,16 @@ Full topologies: `docs/designs/current_design_doc.md`.
 - Pass/fail vs RE threshold **0.75**
 - Explicit revise list if fail
 - Never claim reverse engineering complete without judge pass
+
+## Query-time (not reverse engineering)
+
+**data-retriever** is query-time only. It is **not** an RE sub-agent. Do not put it in producer or skeptic fan-out.
+
+When someone asks about an existing Table, Metric, LineagePath, IngestionJob, Transformation, Dashboard, DataProduct, GlossaryTerm, or BusinessObject:
+
+1. Spawn **data-retriever** (via `/dekc-retrieve`).
+2. Do **not** run `dekc_search.py`, `dekc_pack.py`, or `dekc_brain.py` in this orchestrator for that Q&A.
+3. Keep the **retrieval card** only — never a hit list or full pack.
+4. If the question also needs project decisions or system topology, fan out PKC **knowledge-retriever** and/or SAC **architecture-retriever** in parallel.
+
+See `agents/data-retriever.md` and `skills/dekc-retrieve/SKILL.md`.
diff --git a/commands/dekc-retrieve.md b/commands/dekc-retrieve.md
new file mode 100644
index 0000000..d8a9e7a
--- /dev/null
+++ b/commands/dekc-retrieve.md
@@ -0,0 +1,10 @@
+---
+name: dekc-retrieve
+description: Query-time retrieval. Spawn data-retriever; parent keeps a summary card only.
+---
+
+Run the **dekc-retrieve** skill.
+
+User request: `$ARGUMENTS`
+
+Follow `${CLAUDE_PLUGIN_ROOT}/skills/dekc-retrieve/SKILL.md` completely. Spawn **data-retriever**. Do not run `dekc_search.py` or `dekc_pack.py` in the parent. Return the retrieval card only.
diff --git a/docs/GROK_BOT.md b/docs/GROK_BOT.md
index 194c962..d748f82 100644
--- a/docs/GROK_BOT.md
+++ b/docs/GROK_BOT.md
@@ -58,7 +58,9 @@ Writes fail closed without `--author` or `SECOND_BRAIN_IDENTITY`. Successful kno
 
 Default ContextPack: **2 hops / ~20 nodes**.
 
-Pack before answering or writing. Do not dump the entire tree.
+For Q&A, spawn **data-retriever** (`/dekc-retrieve`). Search, score, pack, and deepen stay in that sub-agent. The parent keeps a **retrieval card** only — never a hit list, full pack, or `dekc_brain.py` dump. Pack `--summary` (bodies off) is the card-friendly path.
+
+Do not dump the entire tree. When the question also needs project or architecture context, fan out PKC **knowledge-retriever** and/or SAC **architecture-retriever** in parallel.
 
 ## Skill binding
 
diff --git a/docs/ONBOARDING.md b/docs/ONBOARDING.md
index beaefcd..543ce8f 100644
--- a/docs/ONBOARDING.md
+++ b/docs/ONBOARDING.md
@@ -37,7 +37,7 @@ See [GROK_BOT.md](GROK_BOT.md) for the binding contract.
 
 1. State your identity: `Grok Bot: Data Engineering Knowledge Capture`.
 2. Confirm the knowledge root (`SECOND_BRAIN_ROOT` or the target bundle).
-3. Pack the relevant subgraph (2 hops) before answering or writing.
+3. For Q&A, spawn **data-retriever** (`/dekc-retrieve`) and keep the retrieval card only. Do not dump a 2-hop pack or `dekc_brain.py` output into the parent. Pack before writing.
 4. Persist only through skills + deterministic scripts inside an isolation session when writing a shared brain.
 5. Report path + validation result, not a dumped graph.
 
diff --git a/docs/designs/current_design_doc.md b/docs/designs/current_design_doc.md
index 1b6be67..221228f 100644
--- a/docs/designs/current_design_doc.md
+++ b/docs/designs/current_design_doc.md
@@ -314,6 +314,8 @@ Compensation is rare (knowledge is additive); “undo” = revert Git commit of
 
 Orchestrator and workers load **2-hop packs** (~20 nodes) via RetrievalBinding, not the full lake graph. Judge sees doctor JSON + gap list only. This matches OKF progressive disclosure and AGER ContextIsolationPolicy for subagents.
 
+Query-time Q&A is a different path: spawn **data-retriever** (`/dekc-retrieve`). Search, score, pack (`--summary`), and deepen stay in that sub-agent; the parent keeps a retrieval card only. Do not confuse this with RE producers.
+
 ### 3.10 Optional formal AGER bundle
 
 Teams can author an explicit AGER bundle (via `/ager-init`) that **KnowledgeBinds** the DEKC `knowledge/` root and **uses** DEKC scripts as Tools. DEKC agents remain usable without that bundle; the AGER bundle is the portable runtime config for LangGraph/CrewAI adapters.
diff --git a/docs/designs/retrieval-ladder.md b/docs/designs/retrieval-ladder.md
index b05f22f..281c66f 100644
--- a/docs/designs/retrieval-ladder.md
+++ b/docs/designs/retrieval-ladder.md
@@ -55,6 +55,7 @@ the tree. The curate hook does not install rg and does not build the index.
 |---|---|---|
 | `dekc_search.py` | yes | candidate prefilter |
 | `dekc_pack.py` inbound | yes | reverse index via literal path + self SQL |
+| `dekc_pack.py --summary` | n/a | card-friendly render; bodies off; fail-closed budget |
 | `dekc_validate.py` / `dekc-curate.sh` | **no** | must parse frontmatter + resolve links |
 | orphans / doctor listings | **no** | need the parsed graph |
 | mermaid (`build_graph`) | **no** | leftover full scan; accepted |
diff --git a/docs/user_guide/user-guide.md b/docs/user_guide/user-guide.md
index a3636c8..a74074d 100644
--- a/docs/user_guide/user-guide.md
+++ b/docs/user_guide/user-guide.md
@@ -132,7 +132,9 @@ python3 scripts/dekc_brain.py "silver.orders" --intent impact --repo . --bundle
 
 Each intent returns: design checklist, ranked schema-typed concepts, progressive disclosure pack, lineage snippet.
 
-Skills: `dekc-second-brain`, `dekc-design-report`, `dekc-land-data`.
+For Q&A, spawn **data-retriever** (`/dekc-retrieve`) instead of dumping a brain pack into the parent.
+
+Skills: `dekc-retrieve`, `dekc-second-brain`, `dekc-design-report`, `dekc-land-data`.
 
 Patterns: `patterns/design-report-from-gold.md`, `patterns/land-stream-to-bronze.md`.
 
@@ -221,6 +223,9 @@ python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/gold-order-daily.md
 # tiny pack for chat focus
 python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/gold-order-daily.md \
   --repo . --bundle knowledge --tiny
+# card-friendly summary (bodies off) — prefer this from data-retriever
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/gold-order-daily.md \
+  --repo . --bundle knowledge --tiny --summary
 ```
 
 If `okf-graph-eng` is available, prefer its pack/impact for the same paths.
@@ -237,8 +242,9 @@ If `okf-graph-eng` is available, prefer its pack/impact for the same paths.
 | `dekc-business-object` | Promote technical → business + glossary |
 | `dekc-glossary` | Author glossary terms |
 | `dekc-semantic` | Semantic models, metrics, dashboards |
-| `dekc-context` | Context packs |
-| `dekc-search` | Second-brain search |
+| `dekc-retrieve` | Query-time retrieval — spawn `data-retriever`; card only |
+| `dekc-context` | Context packs (retriever internals / known seed) |
+| `dekc-search` | Second-brain search (raw hits; not parent Q&A) |
 | `dekc-index` | Rebuild index |
 | `dekc-grade` | Rubric grade + adversarial RE protocol |
 | `dekc-doctor` | Health: coverage, orphans, validation |
@@ -247,6 +253,7 @@ If `okf-graph-eng` is available, prefer its pack/impact for the same paths.
 
 | Agent | Use when |
 |-------|----------|
+| **data-retriever** | Query-time Q&A — spawn via `/dekc-retrieve`; parent keeps a card only |
 | **data-lake-walker** | Full reverse-engineer loop (orchestrator) |
 | **schema-scout** | Only structure (schemas/tables/columns) |
 | **lineage-tracer** | Only SQL/job lineage and promotions |
diff --git a/hosts/grok-bot/SKILL.md b/hosts/grok-bot/SKILL.md
index 985a124..1011d80 100644
--- a/hosts/grok-bot/SKILL.md
+++ b/hosts/grok-bot/SKILL.md
@@ -9,6 +9,6 @@ Read `docs/ONBOARDING.md` first, then follow `docs/GROK_BOT.md`.
 
 1. Identity: `grok-bot/data-engineering-knowledge-capture`
 2. Open an isolation session before knowledge writes (`scripts/brain_session.py open`) unless the human already pointed `SECOND_BRAIN_ROOT` at a session worktree.
-3. Pack 2 hops, then write owned types only via this plugin's scripts.
+3. For Q&A, spawn **data-retriever** (`/dekc-retrieve`); keep the card only. Pack 2 hops before writing owned types via this plugin's scripts.
 4. Close the session to PR. Report path + validation result.
 5. Never document a private remote. Never write raw Markdown into the tree.
diff --git a/marketplace.json b/marketplace.json
index 7ba8dee..8e90829 100644
--- a/marketplace.json
+++ b/marketplace.json
@@ -11,7 +11,7 @@
       "name": "data-engineering-knowledge-capture",
       "source": "./",
       "description": "Data lake second brain: schemas, lineage, medallion, semantic layer, glossary. Depends on PKC + OKF.",
-      "version": "0.5.1",
+      "version": "0.5.2",
       "author": {
         "name": "Rick Hightower"
       },
diff --git a/package.json b/package.json
index f9d1764..0e509cd 100644
--- a/package.json
+++ b/package.json
@@ -99,6 +99,6 @@
     "vite": "^8.2.0"
   },
   "description": "Data Engineering Knowledge Capture (DEKC) \u2014 Claude/Grok/Codex/OpenCode plugin + explorer",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "license": "MIT"
 }
diff --git a/plugin.json b/plugin.json
index 03a5f82..cd3c341 100644
--- a/plugin.json
+++ b/plugin.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
   "name": "data-engineering-knowledge-capture",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "Data Engineering Knowledge Capture: schemas, lineage, medallion, semantic models, and glossary into a durable OKF knowledge graph. Multi-host bindings and write isolation.",
   "author": {
     "name": "Rick Hightower",
diff --git a/public/data/catalog.json b/public/data/catalog.json
index 3f9fe41..bbf6417 100644
--- a/public/data/catalog.json
+++ b/public/data/catalog.json
@@ -1,6 +1,6 @@
 {
   "title": "Retail Lake Sample Knowledge",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "depends_on": [
     "project-knowledge-capture",
     "okf-graph-eng"
diff --git a/scripts/dekc_pack.py b/scripts/dekc_pack.py
index 9a4e8f0..8db8189 100755
--- a/scripts/dekc_pack.py
+++ b/scripts/dekc_pack.py
@@ -377,17 +377,72 @@ def render_markdown(
     return "\n".join(md)
 
 
+LEAD_NODE_LIMIT = 8
+
+
+def bodies_off(result: dict) -> dict:
+    """Card path: never include concept bodies (root included)."""
+    out = dict(result)
+    out["nodes"] = [{**n, "body": ""} for n in result.get("nodes", [])]
+    return out
+
+
+def _lead_why(node: dict) -> str:
+    desc = str(node.get("description") or "").strip()
+    title = str(node.get("title") or Path(str(node.get("path") or "")).stem)
+    why = desc.splitlines()[0] if desc else title
+    if len(why) > 140:
+        return why[:137] + "..."
+    return why
+
+
+def render_summary(
+    result: dict,
+    *,
+    tokens: int | None = None,
+    budget: int | None = None,
+) -> str:
+    """Compact card-friendly pack. Bodies off. No mermaid, no excerpts."""
+    focus = result["focus"]
+    seed = next((n for n in result.get("nodes", []) if n.get("path") == focus), None)
+    seed_type = (seed or {}).get("type") or "?"
+    token_bit = ""
+    if tokens is not None and budget is not None:
+        token_bit = f" tokens={tokens}/{budget}"
+    lines = [
+        "# Pack summary",
+        "",
+        f"- Seed: `{focus}` (`{seed_type}`)",
+        f"- Engine: {result.get('reverse_index') or 'scan'}",
+        f"- Pack: hops={result['hops']} nodes={result['node_count']}{token_bit}",
+        "- Lead nodes:",
+    ]
+    for n in result.get("nodes", [])[:LEAD_NODE_LIMIT]:
+        lines.append(
+            f"  - `{n['path']}` (`{n.get('type') or '?'}`) — {_lead_why(n)}"
+        )
+    lines.append("")
+    return "\n".join(lines)
+
+
 def finalize_markdown(
     result: dict,
     *,
     bundle: Path | None = None,
     max_tokens: str | int | None = None,
     window_tokens: str | int | None = None,
+    summary: bool = False,
 ) -> tuple[str, dict[str, int]]:
     window, budget = resolve_pack_budget(max_tokens, window_tokens)
-    draft = render_markdown(result, bundle=bundle, tokens=0, budget=budget)
-    tokens = estimate_tokens(draft)
-    md = render_markdown(result, bundle=bundle, tokens=tokens, budget=budget)
+    if summary:
+        card = bodies_off(result)
+        draft = render_summary(card, tokens=0, budget=budget)
+        tokens = estimate_tokens(draft)
+        md = render_summary(card, tokens=tokens, budget=budget)
+    else:
+        draft = render_markdown(result, bundle=bundle, tokens=0, budget=budget)
+        tokens = estimate_tokens(draft)
+        md = render_markdown(result, bundle=bundle, tokens=tokens, budget=budget)
     tokens = estimate_tokens(md)
     meta = {"tokens": tokens, "budget": budget, "window": window}
     if tokens > budget:
@@ -405,6 +460,11 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--max-tokens", default="")
     parser.add_argument("--window-tokens", default="")
     parser.add_argument("--tiny", action="store_true")
+    parser.add_argument(
+        "--summary",
+        action="store_true",
+        help="Compact card-friendly summary; bodies off (query-time retriever)",
+    )
     parser.add_argument("--json", action="store_true")
     parser.add_argument("--mermaid", action="store_true")
     parser.add_argument("--write", action="store_true", help="Write pack under packs/")
@@ -443,11 +503,14 @@ def main(argv: list[str] | None = None) -> int:
                 raise PackBudgetError(tokens, budget, window, [n["path"] for n in result["nodes"]])
             print(diagram)
             return 0
+        if args.summary:
+            result = bodies_off(result)
         md, meta = finalize_markdown(
             result,
             bundle=bundle,
             max_tokens=args.max_tokens,
             window_tokens=args.window_tokens,
+            summary=args.summary,
         )
     except PackBudgetError as exc:
         payload = {
@@ -495,7 +558,12 @@ def main(argv: list[str] | None = None) -> int:
         append_log(bundle, f"Wrote context pack for {result['focus']}")
 
     if args.json:
-        print(json.dumps({k: v for k, v in result.items() if k != "markdown"}, indent=2))
+        payload = {k: v for k, v in result.items() if k != "markdown"}
+        if args.summary:
+            payload["summary"] = md
+            for node in payload.get("nodes") or []:
+                node.pop("body", None)
+        print(json.dumps(payload, indent=2))
     elif not args.write:
         print(md)
     return 0
diff --git a/skills/dekc-context/SKILL.md b/skills/dekc-context/SKILL.md
index acc1429..7e2bd26 100644
--- a/skills/dekc-context/SKILL.md
+++ b/skills/dekc-context/SKILL.md
@@ -8,7 +8,10 @@ description: Build progressive-disclosure context packs for a data concept.
 ```bash
 python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/<slug>.md --repo . --bundle knowledge --hops 2
 python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/<slug>.md --repo . --bundle knowledge --tiny
+python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/<slug>.md --repo . --bundle knowledge --tiny --summary
 python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_pack.py" tables/<slug>.md --repo . --bundle knowledge --write
 ```
 
+`--summary` is card-friendly (bodies off, fail-closed budget). Query-time Q&A must spawn **data-retriever** via `/dekc-retrieve` instead of running this skill in the parent.
+
 When okf-plugin is installed, `okf pack` / impact also work on DEKC concepts.
diff --git a/skills/dekc-retrieve/SKILL.md b/skills/dekc-retrieve/SKILL.md
new file mode 100644
index 0000000..5a14b82
--- /dev/null
+++ b/skills/dekc-retrieve/SKILL.md
@@ -0,0 +1,52 @@
+---
+name: dekc-retrieve
+description: Query-time DEKC retrieval. Spawn data-retriever; parent gets a summary card only. Use for Table, Metric, LineagePath, IngestionJob, Transformation, Dashboard, DataProduct, GlossaryTerm, BusinessObject questions.
+---
+
+# Retrieve (query-time)
+
+The parent **must spawn `data-retriever`**. Do **not** run `dekc_search.py` or `dekc_pack.py` in the parent for retrieval. Search, score, pack, and deepen stay in the sub-agent so hit lists and pack markdown never contaminate the parent.
+
+This is **not** reverse engineering. Do not spawn `data-lake-walker`, `lineage-tracer`, or other RE/capture agents for a question about existing knowledge.
+
+## Spawn
+
+Ask the host to run the `data-retriever` sub-agent with the user's question (and a seed path if the user already named one). The only thing you keep is the retrieval card:
+
+```markdown
+## Retrieval card
+- Query: …
+- Seed: `/path` (`Type`) — why chosen
+- Fit: high|medium|low — one sentence
+- Engine: index|rg|scan
+- Pack: hops=N nodes=N tokens=N/budget
+- Lead nodes: (5–8 bullets)
+- Lineage note: (upstream/downstream one-liner) or none
+- Open gaps: or none
+- Next: stay|deepen-2hop|lineage|try-alt-seed `/other`
+```
+
+If `Next` is `deepen-2hop`, `lineage`, or `try-alt-seed /other`, spawn `data-retriever` again with that instruction. Do not unpack the graph yourself.
+
+## Orthogonal fan-out
+
+When the question also needs project decisions or system topology, spawn in parallel:
+
+| Plane | Sub-agent |
+|-------|-----------|
+| Data platform (this pack) | **data-retriever** |
+| Project / product knowledge | PKC **knowledge-retriever** |
+| System architecture | SAC **architecture-retriever** |
+
+Do not merge their packs into the parent. Each returns its own card.
+
+## Not this skill
+
+| Need | Use |
+|------|-----|
+| Walk / capture / reverse-engineer | `data-lake-walker` / `/dekc-walk` |
+| Raw keyword search (developer) | `/dekc-search` — still not for parent Q&A |
+| Full pack for a known seed (retriever internals) | `/dekc-context` |
+| Intent checklists (`dekc_brain.py`) | `/dekc-second-brain` — do **not** dump those packs inline for Q&A |
+
+See `agents/data-retriever.md`.
diff --git a/skills/dekc-search/SKILL.md b/skills/dekc-search/SKILL.md
index ce95054..646f9f2 100644
--- a/skills/dekc-search/SKILL.md
+++ b/skills/dekc-search/SKILL.md
@@ -22,3 +22,5 @@ uses FTS5 MATCH (prefix tokens; not score-identical). Missing index or rg
 is not an error. See `/dekc-index`.
 
 After hits on a table, offer `/dekc-context` (2-hop pack).
+
+For parent Q&A, do **not** run this skill in the parent. Spawn **data-retriever** via `/dekc-retrieve`. Search + pack stay in that sub-agent; the parent gets a retrieval card only.
diff --git a/skills/dekc-second-brain/SKILL.md b/skills/dekc-second-brain/SKILL.md
index d7c2b2c..5709d61 100644
--- a/skills/dekc-second-brain/SKILL.md
+++ b/skills/dekc-second-brain/SKILL.md
@@ -37,7 +37,7 @@ python3 "${CLAUDE_PLUGIN_ROOT}/scripts/dekc_brain.py" "gold order_daily" \
 
 ## Workflow
 
-1. **Retrieve** with `dekc_brain.py` (intent + query) → checklist + ranked concepts + 2-hop pack  
+1. **Retrieve** with `dekc_brain.py` (intent + query) → checklist + ranked concepts + 2-hop pack. For parent Q&A, spawn **data-retriever** (`/dekc-retrieve`) and keep the card only — do not dump the brain pack inline.  
 2. **Design** using only evidence from the pack (do not invent lineage)  
 3. **Capture** new assets with `dekc_capture.py` matching `schemas/okf-concepts/`  
 4. **Index** is optional — search/pack refresh `knowledge/.dekc/index.sqlite` themselves. `dekc_index.py build` is `refresh --force`.  
diff --git a/tests/test_dekc.py b/tests/test_dekc.py
index 17eb70a..40a14f2 100644
--- a/tests/test_dekc.py
+++ b/tests/test_dekc.py
@@ -3,6 +3,8 @@
 
 from __future__ import annotations
 
+import contextlib
+import io
 import json
 import re
 import subprocess
@@ -710,6 +712,120 @@ def test_over_budget_fails_closed(self):
             shutil.rmtree(tmp)
 
 
+class TestPackSummary(unittest.TestCase):
+    def _tiny_bundle(self) -> Path:
+        tmp = Path(tempfile.mkdtemp())
+        tables = tmp / "tables"
+        tables.mkdir()
+        (tmp / "index.md").write_text(
+            '---\nokf_version: "0.2"\ntitle: t\n---\n', encoding="utf-8"
+        )
+        (tables / "root.md").write_text(
+            "---\ntype: Table\ntitle: Lumenfield Root\n"
+            "description: root-desc for the seed table\n"
+            "links:\n  - target: /tables/neighbor.md\n    rel: feeds\n"
+            "---\n# Lumenfield Root\n\nROOT_BODY_MARKER secret-of-root\n",
+            encoding="utf-8",
+        )
+        (tables / "neighbor.md").write_text(
+            "---\ntype: Table\ntitle: Neighbor\n"
+            "description: neighbor-frontmatter-only\n---\n"
+            "# Neighbor\n\nNEIGHBOR_BODY_MARKER must-not-pack\n",
+            encoding="utf-8",
+        )
+        self.addCleanup(shutil.rmtree, tmp)
+        return tmp
+
+    def test_summary_bodies_off_even_for_root(self):
+        from dekc_pack import finalize_markdown, pack
+
+        tmp = self._tiny_bundle()
+        result = pack(tmp, "tables/root.md", hops=1, max_nodes=8)
+        md, meta = finalize_markdown(result, bundle=tmp, summary=True)
+        self.assertIn("# Pack summary", md)
+        self.assertIn("`/tables/root.md`", md)
+        self.assertIn("(`Table`)", md)
+        self.assertIn("hops=1", md)
+        self.assertIn("nodes=", md)
+        self.assertIn("tokens=", md)
+        self.assertIn("Lead nodes:", md)
+        self.assertIn("root-desc for the seed table", md)
+        self.assertIn("neighbor-frontmatter-only", md)
+        self.assertNotIn("ROOT_BODY_MARKER", md)
+        self.assertNotIn("NEIGHBOR_BODY_MARKER", md)
+        self.assertNotIn("```mermaid", md)
+        self.assertGreater(meta["budget"], 0)
+        self.assertLessEqual(meta["tokens"], meta["budget"])
+
+    def test_summary_smaller_than_full_pack(self):
+        from dekc_pack import finalize_markdown, pack
+
+        tmp = self._tiny_bundle()
+        result = pack(tmp, "tables/root.md", hops=1, max_nodes=8)
+        full, _ = finalize_markdown(result, bundle=tmp, summary=False)
+        card, _ = finalize_markdown(result, bundle=tmp, summary=True)
+        self.assertLess(len(card), len(full))
+        self.assertIn("ROOT_BODY_MARKER", full)
+        self.assertNotIn("ROOT_BODY_MARKER", card)
+
+    def test_summary_over_budget_fails_closed(self):
+        from dekc_pack import PackBudgetError, finalize_markdown, pack
+        from dekc_pack import main as pack_main
+
+        tmp = self._tiny_bundle()
+        result = pack(tmp, "tables/root.md", hops=1, max_nodes=8)
+        with self.assertRaises(PackBudgetError) as ctx:
+            finalize_markdown(result, bundle=tmp, max_tokens=8, summary=True)
+        self.assertGreater(ctx.exception.tokens, ctx.exception.budget)
+        self.assertEqual(ctx.exception.budget, 8)
+        rc = pack_main(
+            [
+                "tables/root.md",
+                "--repo",
+                str(tmp),
+                "--bundle",
+                str(tmp),
+                "--tiny",
+                "--summary",
+                "--max-tokens",
+                "8",
+                "--write",
+                "--json",
+                "--author",
+                "grok-bot/northstar-console",
+            ]
+        )
+        self.assertNotEqual(rc, 0)
+        self.assertEqual(list(tmp.rglob("pack-root.md")), [])
+
+    def test_summary_cli_json_has_no_bodies(self):
+        from dekc_pack import main as pack_main
+
+        tmp = self._tiny_bundle()
+        buf = io.StringIO()
+        with contextlib.redirect_stdout(buf):
+            rc = pack_main(
+                [
+                    "tables/root.md",
+                    "--repo",
+                    str(tmp),
+                    "--bundle",
+                    str(tmp),
+                    "--tiny",
+                    "--summary",
+                    "--json",
+                ]
+            )
+        self.assertEqual(rc, 0)
+        payload = json.loads(buf.getvalue())
+        self.assertIn("summary", payload)
+        self.assertIn("# Pack summary", payload["summary"])
+        self.assertNotIn("ROOT_BODY_MARKER", payload["summary"])
+        self.assertNotIn("markdown", payload)
+        for node in payload["nodes"]:
+            self.assertFalse(node.get("body"))
+
+
 if __name__ == "__main__":
     unittest.main()