merge: integrate narrowed MCP gap closure

Author: vicary

README.md

@@ -84,12 +84,24 @@ surface** for data-heavy work. These tools run in-process alongside the plugin
 hooks and share the same canonical root-session identity and Redis/FalkorDB hot
 tier.
 
+The `session_*` tools also write into the same local continuity model as the
+rest of the session: their bounded summaries are recorded as structured events,
+folded into the local snapshot, and preserved through compaction under the same
+`<session_memory>` envelope used for ordinary chat continuity.
+
 - **Bounded execution** (`session_execute`, `session_execute_file`,
   `session_batch_execute`) — run commands or process files locally, store full
   output in the local corpus, and return only a bounded summary to the model.
+  `session_batch_execute` supports ordered mixed steps, so one request can
+  combine bounded command execution with local corpus search.
 - **Local indexing and search** (`session_index`, `session_search`,
   `session_fetch_and_index`) — index content into a per-session local corpus in
-  Redis/FalkorDB and search it with bounded result sets.
+  Redis/FalkorDB and search it with bounded result sets. The local corpus stays
+  local-first: indexing and retrieval happen against the session's local store,
+  while any Graphiti augmentation remains asynchronous and cache-backed.
+  `session_index` accepts either inline `content` or a local `path`; when the
+  same `source` and `label` are indexed again for one root session, the prior
+  logical document is replaced instead of appended.
 - **Diagnostics** (`session_stats`, `session_doctor`) — inspect session state
   and corpus health.
 
@@ -311,8 +323,9 @@ Events are also queued for background ingestion into long-term memory:
 
 - **On idle** (`session.idle`): buffered events are sent to Graphiti and the
   priority-tiered snapshot is rebuilt.
-- **Before compaction** (`session.compacted`): all pending events are sent
-  immediately so nothing is lost.
+- **After compaction** (`session.compacted`): the compaction summary and any
+  pending continuity are scheduled for background Graphiti ingestion so nothing
+  is lost across compaction boundaries.
 
 ### Compaction Preservation
 

docs/ContextOverhaul.md

@@ -7,8 +7,15 @@
 - `docs/superpowers/plans/2026-03-20-context-mode-mcp-first-implementation.md`
   (implementation)
 
+> **Historical-only note:** This document preserves the earlier native-routing
+> overhaul proposal and its original section numbering. Any implementation
+> phases, file-change lists, or acceptance checklists below are historical notes
+> only and are **not** the active backlog for the repository. For current
+> architecture and acceptance criteria, use the two superseding MCP-first plan
+> documents above together with `README.md`.
+
 **Date:** 2026-03-20\
-**Canonical refs:** `AGENTS.md`, `README.md`, `docs/ContextOverhaulTests.md`
+**Historical refs:** `README.md`, `docs/ContextOverhaulTests.md`
 
 ---
 

docs/ContextOverhaulTests.md

@@ -11,6 +11,11 @@ superseded)\
 > design. The active architecture is now MCP-first; see the implementation plan
 > linked above for the current acceptance criteria. The suites below remain as
 > historical reference for the original hot-path invariants.
+>
+> **Historical-only note:** Checklist items in this document are not the
+> authoritative Task 7 gate. Use the implementation plan above for current
+> MCP-first acceptance, including local-first `<session_memory>`, compaction
+> continuity, and Graphiti-off-the-hot-path verification.
 
 ---
 
@@ -209,7 +214,7 @@ and within budget.
 - [ ] B-3: Total injected payload (session + persistent) does not exceed 5% of a
       128k-token model context (≈ 25 600 chars).
 - [ ] B-4: Snapshot XML conforms to the priority-tiered schema from
-      `ContextOverhaul.md` §8.3.
+      [`docs/ContextOverhaul.md` §8.3](ContextOverhaul.md#83-snapshot-policy).
 - [ ] B-5: Snapshot respects the 3 KB budget — lower-priority sections are
       truncated first.
 - [ ] B-6: Each `session_memory` always contains `last_request`; list sections
@@ -541,12 +546,13 @@ parent.
 
 **Tier:** Unit + Integration
 
-**Canonical design reference:** `docs/ContextOverhaul.md` §11.1
+**Historical design reference:**
+[`docs/ContextOverhaul.md` §11.1](ContextOverhaul.md#111-kept-divergence)
 
 **Divergence note:** This behavior intentionally differs from official
 `mksglu/context-mode`, which treats subagent work as summarized tool events
-rather than first-class session participants. See §11.1 of the design doc for
-the rationale and alignment guidance.
+rather than first-class session participants. See the historical §11.1 design
+note above for the rationale and alignment guidance.
 
 #### Checklist
 

docs/superpowers/plans/2026-03-20-context-mode-mcp-first-implementation.md

@@ -1,6 +1,6 @@
 # Context-Mode-Aligned MCP-First Replacement — Implementation Task Plan
 
-**Status:** Planned\
+**Status:** Completed\
 **Date:** 2026-03-20\
 **Primary architecture:**
 `docs/superpowers/plans/2026-03-20-context-mode-mcp-first.md`\

docs/superpowers/plans/2026-03-20-context-mode-mcp-first.md

@@ -1,6 +1,6 @@
 # Context-Mode-Aligned MCP-First Replacement Plan
 
-**Status:** Superseding plan\
+**Status:** Completed\
 **Date:** 2026-03-20\
 **Supersedes:** `plans/ContextOverhaul.md` and any in-progress Task 1 / Task 2
 work derived from that native-hook-first plan\