Server: eib-mcp-rag-full (local stdio, Python MCP server, DB_BACKEND=cots) Host: Parallel Works COTS host (ChromaDB + Neo4j, mpnet768 embeddings) Snapshot: 2026-07-15T16:40Z Purpose: Build a robust, verifiable ground-truth context so we can evaluate what changes need to land in the top-level instruction files for GitHub Copilot (.github/copilot-instructions.md) and Kiro (.kiro/steering/*.md).

This report is generated by running every health/status/state MCP tool at maximum depth against the on-workspace COTS server, then cross-referencing the results with the CHANGELOG.md "Unreleased" entries, the 68 SDD workflow specs, and the 49 Kiro spec directories. It is not a report on the AWS agentcore-mcp-rag runtime — the July 13 sister report on the Gateway/AgentCore surface is the companion for that plane.

---

Dimension	Status	Notes
Overall Health	HEALTHY (4/4)	Base Server, Utility, Vector, Graph all pass
Functional Smoke Tests	11/11 pass, 0 fail, 0 skip	All modules exercised end-to-end
Tool Surface	53 tools across 9 active modules	1 module inactive (SPOT — parity with AgentCore)
Tenants	5 configured, 5 filesystem-reachable	gw (default), gw_sfs, gw_jedi_gfs, gw_v17, gw_gefs_v12
Vector Store (ChromaDB)	17 indices, 220,538 docs in gw, 15 collections	Healthy
Graph Store (Neo4j)	225,836 nodes / 4,051,374 rels in gw	Healthy
Knowledge Integrity	4/4 checks pass (0 SKIP)	Phase 68 fix landed cleanly
Health Trend (15 snapshots)	Neo4j stable ±0; ChromaDB +2,610 docs	No anomalies
Quality Benchmarks	MISSING	No quality_metrics.jsonl on disk
Manifest	v9.2.0, 67 sources (61 shared / 6 tenant)	Scope-aware naming live
Manifest Gap Detector	All 5 tracked collections report 0 actual	Naming-schema drift (details in §7)
SDD Framework	v6.0 Phase 31, 68 workflows, 49 sessions	35 completed, 5 abandoned
Active SDD Session	None	Last completed: Phase 68 (RAG Data-Plane Gap Closure)
Latest CHANGELOG entry	Phase 68 — RAG Data-Plane Gap Closure (Jul 10, 2026)	Verified COTS-truthfully
Active in-flight spec	cots-reingest-ralph-framework	Paused after PoC; resume in fresh session

Headline: The COTS stdio server is fully green on structure and behavior. The two open items that should shape the next instruction-file revision are (a) the manifest gap-detector's declared collection names no longer match the actual COTS collections produced by the Phase 68 resolve_collection_name namer, and (b) no RAG quality benchmark is currently on disk, so we cannot regression-detect embedding or retrieval changes.

---

Overall: HEALTHY (4/4 components)

Component	Status	Detail
Base Server	[OK]	FastMCP running
Utility Tools	[OK]	4 utility tools registered
Vector Database	[OK]	17 indices
Graph Database	[OK]	108,280 nodes, 4,229,217 relationships (unfiltered; per-tenant totals in §4)

tenant_id	branch	lifecycle	index_prefix	label_prefix	workflow_subdir	reachable
gw	develop	production	(empty)	(empty)	develop	yes
gw_sfs	dev/sfs	experimental	gw_sfs_	GW_SFS_	dev-sfs	yes
gw_jedi_gfs	dev/jedi-gfs	experimental	gw_jedi_gfs_	GW_JEDI_GFS_	dev-jedi-gfs	yes
gw_v17	dev/gfs.v17	staging	gw_v17_	GW_V17_	dev-v17	yes
gw_gefs_v12	release/gefs_v12	production	gw_gefs_v12_	GW_GEFS_V12_	gefs-v12	yes

Default tenant: gw (resolved from catalog.defaults.tenant_id). Mount root: /mcp_rag_eib/eib-mcp-rag-server/.pw_workflow_mount (mounted; all 5 subdirectories present).

Module	Status	Latency	Notes
semantic_search	[OK]	903 ms	Slowest — expected (embed + vector search)
code_analysis	[OK]	7 ms	Neo4j hot
graph_rag	[OK]	5 ms	Neo4j hot
ee2_compliance	[OK]	41 ms
operational	[OK]	43 ms
sdd_workflow	[OK]	0 ms	Filesystem
workflow_info	[OK]	0 ms	Filesystem
github_tools	[OK]	120 ms	GitHub API round-trip
utility	[OK]	0 ms	Built-in
code_awareness	[OK]	72 ms
branch_isolation	[OK]	108 ms	Multi-tenant probe

Health snapshot persisted to health_history.jsonl (feeds §6).

---

• Version: MDC MCP/RAG Server v1.0.0
• Total tools: 53
• Active modules: 9 of 10 — semantic_search, code_analysis, graph_rag, ee2_compliance, operational, sdd_workflow, workflow_info, github_tools, utility
• Tenants: 5 (default gw)
• Capabilities: Data Access connected; Vector Search available; Graph Queries available; Utility Tools always-on (4)

analyze_code_structure, analyze_ee2_compliance, analyze_repository_structure, analyze_workflow_dependencies, check_knowledge_integrity, checkpoint_state, complete_sdd_session, describe_component, explain_with_context, explain_workflow_component, extract_ci_error_signal, extract_code_for_analysis, find_callers_callees, find_dependencies, find_env_dependencies, find_related_files, find_similar_code, generate_compliance_report, get_change_impact, get_code_context, get_health_trend, get_ingested_urls_array, get_job_details, get_knowledge_base_status, get_operational_guidance, get_pull_requests, get_quality_metrics, get_sdd_execution_history, get_sdd_framework_status, get_sdd_session, get_sdd_workflow, get_server_info, get_session_context, get_system_configs, get_workflow_structure, list_all_sources, list_ingested_urls, list_job_scripts, list_sdd_workflows, mark_as_modified, mcp_health_check, record_sdd_step, restore_checkpoint, scan_repository_compliance, search_architecture, search_documentation, search_ee2_standards, search_issues, start_sdd_session, trace_data_flow, trace_execution_path, trace_full_execution_chain, validate_sdd_compliance.

---

• Collections: 15
• Total documents: 220,538
• Status: [OK] Healthy

Collection	Documents
global-workflow-docs-v8-1-0	20,511
code-with-context-v8-0-0	60,574
phase48-scratch	3,630
jjobs-v8-0-0	700
mdc-jjobs-mpnet768	700
global-workflow-docs-v8-2-0	23,624
mdc-ee2-standards-mpnet768	34
community-summaries	2,113
global-workflow-docs-v8-0-0	22,498
ci-test-cases-v1-0-0	74
ee2-standards-v5-0-0-enhanced	34
jjobs-v8-1-0	859
mdc-code-context-mpnet768	60,576
mdc-workflow-docs-mpnet768	22,498
mdc-community-summaries-mpnet768	2,113

Note the co-existence of two naming families: the legacy versioned collections (*-v8-*-*) still serving as the reference set, and the new scope-aware mdc-*-mpnet768 collections introduced by Phase 68's collection_namer.py. This is the deliberate alongside pattern from the cots-reingest-ralph-* spec chain — cutover is a separate human-gated step.

• Files: 17,273
• Functions: 113,571
• Classes: 648
• Total nodes: 225,836
• Total relationships: 4,051,374
• Status: [OK] Healthy

Relationship types (top 10):

Type	Count
CALLS	3,306,540
USES	679,698
DEPENDS_ON_ENV	28,504
DEFINES	11,687
IMPORTS	9,141
EXPORTS	7,276
DEPENDS_ON	4,032
INVOKES	2,579
SOURCES	1,788
EXECUTES	129

Label breakdown (top 10):

Label	Count
Function	89,595
FortranSubroutine	80,745
File	17,273
FortranFunction	16,849
FortranModule	9,014
PythonFunction	7,127
PythonModule	1,480
Module	1,330
FortranProgram	1,186
ShellScript	589

---

Overall: All checks passed — this is the direct outcome of Phase 68's fix that unblocked the previously-[SKIP] checks.

Check	Status	Details
Path Consistency	[OK]	0/50 randomly sampled docs carry a checkout-specific prefix
Orphaned Graph Nodes	[OK]	17,273 File nodes in graph; 0/20 sampled lack identity
Stale Embeddings	[OK]	50/50 sampled docs appear current vs. their git source
Coverage Gap	[OK]	106,608 Fortran symbols in graph; 7,242 files on disk (1,472.1% coverage)

The "1472% coverage" is the expected symbols-per-file ratio (many symbols per Fortran file) — it is a healthy signal that the graph is denser than the file inventory, not a sign of double-ingestion.

---

Timestamp (UTC)	Neo4j Nodes	Neo4j Rels	ChromaDB Docs	Collections	Node Drift	Doc Drift
2026-06-24 23:43	108,280	4,220,211	220,538	15	0	0
2026-06-25 22:00	108,280	4,220,211	220,538	15	0	0
2026-06-26 13:08	108,280	4,220,211	220,538	15	0	0
2026-06-26 20:31	108,280	4,220,211	220,538	15	0	0
2026-06-26 20:40	108,280	4,220,211	220,538	15	0	0
2026-06-26 23:23	108,280	4,220,211	220,538	15	0	0
2026-07-02 13:37	108,280	4,220,211	220,538	15	0	0
2026-07-02 16:01	108,280	4,220,211	220,538	15	0	0
2026-07-02 21:58	108,280	4,220,211	220,538	15	0	0
2026-07-02 22:03	108,280	4,220,211	220,538	15	0	0
2026-07-03 03:18	108,280	4,220,211	220,538	15	0	0
2026-07-09 22:07	108,280	4,229,217	223,148	17	0	+2,610
2026-07-09 23:28	108,280	4,229,217	223,148	17	0	0
2026-07-10 02:26	108,280	4,229,217	223,148	17	0	0
2026-07-14 17:43	108,280	4,229,217	223,148	17	0	0

Trends

• Neo4j nodes: stable (+0 over 15 snapshots)
• ChromaDB docs: increasing (+2,610 over the window — attributable to the July 8-10 Ralph-loop PoC that produced the two new mdc-* collections + phase48-scratch; see CHANGELOG "COTS Full Re-Ingest via Ralph Loop")

Anomalies: none (all consecutive changes within the 10% threshold).

The current get_knowledge_base_status in §4 shows 220,538 docs / 15 collections while the trend rows since 2026-07-09 report 223,148 / 17. The delta (+2,610 / +2 collections) is the July PoC snapshot; the KB-status view is scoped to the gw (default) tenant per Phase 68's _filter_indices_by_tenant fix, so the two collections created for gw_v17 (jjobs + partial documentation) are correctly excluded from the gw tenant total. This is expected behavior, not drift.

---

• Manifest version: 9.2.0
• Generated: 2026-07-15T16:40:13Z
• Total sources: 67 (65 enabled)

Type	Sources	Enabled	Declared docs
code_parse	3	3	88,614
community_summary	1	1	2,113
config_parse	2	2	1,484
jjob_docs	1	1	751
on_disk_submodule	1	1	1,759
standards	1	1	34
url_crawl	58	56	19,489

Scope	Sources	Enabled	Note
shared	61	59	NWS-wide, ingested once (unprefixed collection)
tenant	6	6	per (repo, branch) — tenant-prefixed collection

Collection	Status	Declared	Actual	Coverage
code-with-context-v8-0-0	gap	90,098	0	0.0%
community-summaries	gap	2,113	0	0.0%
ee2-standards-v5-0-0-enhanced	gap	34	0	0.0%
global-workflow-docs-v8-0-0	gap	21,248	0	0.0%
jjobs-v8-0-0	gap	751	0	0.0%

Finding — naming-schema drift: the gap detector reports 0 actual for every collection even though §4.1 shows those collections populated with real document counts. This is a detector-side lookup mismatch, not a data-plane loss. Phase 68's resolve_collection_name moved the naming authority to collection_namer.py, but the gap-detector step of list_all_sources is evidently comparing manifest target_collection strings against a resolver that now produces mdc-*-mpnet768-style names on COTS. This is the first candidate item for the next Kiro spec.

---

Result: No benchmark results found. Expected at sdd_framework/execution_state/quality_metrics.jsonl — the file does not exist on this host.

Implications:

• We cannot detect P@k or MRR regressions from Phase 68 or the July Ralph-loop PoC.
• The 2026-07-02 comparative report MCP-Health-Status-Comparative-Report flagged this exact gap on AgentCore; it now applies to the COTS stdio server as well.
• Running the benchmark harness is a prerequisite for validating the eventual mdc-*-mpnet768 cutover.

---

Active platforms detected in env/: AWS-EC2, AWSPW, AZUREPW, CONTAINER, DERECHO, GAEAC6, GOOGLEPW, HERA, HERCULES, ORION, URSA, WCOSS2.

Path SPOT (tenant gw):

• HOMEgfs → /mcp_rag_eib/eib-mcp-rag-server/.pw_workflow_mount/develop
• Jobs, Scripts, Utilities, Parameters, Source resolve under ${HOMEgfs}/... — the Phase 67 _develop rename is transparent to the tenant, as intended.

---

• Version: 6.0 Phase 31
• Status: Operational
• Execution model: Session-Oriented Tracking
• Available workflows: 68
• Total sessions: 49
  • Completed: 35
  • Abandoned: 5

None — get_session_context returns "No Active Session". Every response you get during this conversation is currently un-attributed to a phase.

• [..] ISD approval — dormant (reserved for Phase 4C USD)
• [..] WorkflowExecutor — filesystem-backed; degraded on AgentCore microVM
• [OK] SpecificationParser — active
• [OK] SelfModificationEngine — available

Status	Phase	Steps	Note
[!!] abandoned	phase60_code_awareness_tool_parity_branch_validation	1	early exit
[OK] completed	phase60_code_awareness_tool_parity_branch_validation	6	superseded the abandoned attempt
[!!] abandoned	phase63_python_container_parity_and_backend_rename	3	rescoped as 63a + 63b
[!!] abandoned	phase63b_python_container_gateway_parity	2	later completed via CHANGELOG Phase 63b
[OK] completed	Phase 68 — RAG Data-Plane Gap Closure	8	last engagement

The 5-item abandonment pattern is not a health problem — in every case the abandoned session was replaced by a re-scoped follow-up (usually with a letter suffix), which is the intended behavior under .kiro/steering/07-feature-branch-spec-workflow.md.

---

The .kiro/specs/ directory now holds 49 spec directories. Grouped by state (inferred from CHANGELOG + wiki cross-reference):

• rag-data-plane-gap-closure — Phase 68 (Jul 10) — the item that produced the "all integrity checks pass" line in §5.
• python-container-gateway-parity — Phase 63b (Jul, per CHANGELOG) — replaced legacy Node.js eib-mcp-rag:latest with eib-mcp-rag-python:latest.
• backend-agnostic-refactor (Phase 63a) — DB_BACKEND=cots|aws|legacy rename; legacy accepted for one release with a WARN.
• manifest-status-writeback (Phase 57), url-crawl-gap-closure (Phase 58), pdf-ingestion-pipeline (Phase 59).

• cots-reingest-ralph-framework — the Ralph-loop driver adopted from mreferre/ralph-loop-kiro-specs, with Slurm sbatch dispatch. PoC ran to v9-0-0 (15 done / 3 skipped / 44 pending in .reingest_state/v9-0-0/state.json). Work paused pending a fresh session — this is where the "next engagement" should resume.
• cots-reingest-ralph-loop — Tasks 1-5, 9 complete; Tasks 6-8 carried into the framework spec.
• gemini-embedding-provider — Phase 66, planning artifacts only, gated on GEMINI_API_KEY.

bedrock-embedding-reingestion, bedrock-native-embedding-swap, graceful-missing-index-handling, neptune-recovery-reingestion, omd-tenants-3-routing, tenant-status-honesty, tenant-source-drift-detection (Phase 69), search-deduplication, functional-smoke-tests.

---

1. Phase 68 — RAG Data-Plane Gap Closure & Tenant-Scope Clarification (Jul 10, 2026) — verified COTS-truthfully on this exact server; 11/11 health tests pass, 4/4 integrity checks run (0 SKIP), KB-status total docs 220,538 [OK] Healthy, list_all_sources gained a By Scope breakdown. Manifest bumped v9.2.0 with mandatory scope: tenant|shared field.
2. COTS Full Re-Ingest via Ralph Loop (Jul 8-9, 2026) — orchestration + State_Manager landed; PoC produced fresh v9-0-0 collections in gw_v17 (gw_v17_mdc-jjobs-titan1024-v9-0-0 + partial documentation) plus new graph nodes stamped v9-0-0. State durable in .reingest_state/v9-0-0/.
3. Phase 67 — supported_repos Rename Path Conformance (Jul 8, 2026) — 44 files updated to route supported_repos/global-workflow → supported_repos/global-workflow_develop. Tenant-aware ingesters required no path change (they use tenant.workflow_root from tenants.yaml).

The most recent [Unreleased] block is Phase 68; nothing has landed since Jul 10 on the CHANGELOG.

---

These are the concrete signals in this snapshot that suggest updates to .github/copilot-instructions.md and .kiro/steering/*.md:

1. Manifest v9.2.0 + scope-awareness is now a load-bearing SPOT, but the Copilot instructions still describe the ingest surface at Phase 63 vintage. The scope: tenant|shared required field, the collection_namer.resolve_collection_name(domain, scope, tenant, version, profile) authority, and the tenant vs. shared collection-naming rules (mdc-{domain}-{profile}{suffix} vs. {index_prefix}mdc-{domain}-{profile}{suffix}) belong in the "Key Conventions → SPOT" section.

2. Gap-detector drift (§7): the list_all_sources gap detector reports 0/0 for all five collections while the underlying data is present. A follow-up spec (candidate name: manifest-gap-detector-namer-alignment) is warranted; until it lands, the instructions should tell agents to not interpret list_all_sources gap rows as data-loss without cross-checking get_knowledge_base_status.

3. Session tracking is off by default. get_session_context reports "No Active Session" and the last engagement (Phase 68) completed cleanly. The Superpowers assessment (SDD-Framework-Comparative-Assessment-Superpowers) already recommended auto-triggering start_sdd_session — this is the right moment to promote that from a "recommendation" to an operating rule in .kiro/steering/02-development-workflow.md and mirror it in .github/copilot-instructions.md under a new "Session Discipline" subsection.

4. quality_metrics.jsonl is missing (§8). Instructions should call out running the benchmark harness after any RAG-plane change (embeddings, ingesters, resolver logic). Today it is only mentioned obliquely in the test-command block.

5. Ralph-loop is the active in-flight orchestration (cots-reingest-ralph-framework). Neither instruction set names it. Agents landing here without context could accidentally launch competing ingest runs. A one-line "Currently in-flight: cots-reingest-ralph-framework (paused; resume via its tasks.md)" note in 04-phase48-progress.md (or a new 14-active-in-flight-specs.md steering doc) would prevent this.

6. The 5 abandoned SDD sessions are all rescopes (§10). They are noise unless we tag them as superseded_by: <phase>. 07-feature-branch-spec-workflow.md could gain a paragraph on the abandon-and-rescope pattern so it stops looking like failure in get_sdd_framework_status output.

7. The default-tenant filter is the correct behavior (§6 footnote). This is worth a one-liner in .github/copilot-instructions.md — because the multi-tenant call surface can otherwise cause agents to conclude the KB shrank when a tool actually just changed its tenant scope.

8. Container-image staleness rule is well-covered (Docker MCP Gateway section of Copilot instructions) — no change needed. This snapshot confirms the current stdio server matches source; there is nothing to reconcile against the Python container image today.

---

• Did not modify .github/copilot-instructions.md or any .kiro/steering/*.md. Per project policy this report only stages the findings; the actual edits happen under a new Kiro spec after human review.
• Did not touch supported_repos/global-workflow_develop or any other read-only submodule. The only submodule touched by this session is supported_repos/global-workflow.wiki, and only to publish this report + update Home.md.
• Did not run git commit or git push outside the wiki submodule — no changes were made outside it.

---

Every table in this report is the direct return of one of the following calls against eib-mcp-rag-full (COTS stdio, DB_BACKEND=cots, MCP_DEFAULT_TENANT=gw) at 2026-07-15T16:40Z:

• mcp_health_check(deep=true, detailed=true, functional=true) → §2
• get_server_info(include_capabilities=true) → §3
• get_knowledge_base_status(include_graph=true, include_vector=true) → §4
• check_knowledge_integrity(sample_size=50) → §5
• get_health_trend(limit=15) → §6
• list_all_sources(format=summary, include_gaps=true) → §7
• get_quality_metrics(compare=true) → §8
• get_system_configs(platform=all, config_type=all) → §9
• get_sdd_framework_status(detailed=true) → §10
• get_sdd_execution_history(analytics=true, limit=20) → §10
• get_session_context(include_dirty=true) → §10

Companion reports on the AWS surface: agentcore-mcp-rag-Health-Status-Report-2026-07-13 and MCP-Health-Status-Comparative-Report.