docs(rfd): Add RFD 107 discoverable tools

docs/.vitepress/rfd-summaries.json

@@ -312,15 +312,15 @@
     "summary": "Plugin configuration integration with AppConfig enabling trust policies, checksum pinning, execution policies, and per-plugin options."
   },
   "078-tool-config-mutation.md": {
-    "hash": "a7d017ca934de55cbfe69ca9d6becc4d98bdbadeb4cc5282444450238ed62e49",
+    "hash": "c208380ff1f5206c9c825d3dec0f28a436897a3f64846eee4003133637c92125",
     "summary": "Scoped tool config access via `access.config` rules, with re-invocation on delta rejection and per-cycle commit buffering."
   },
   "080-editor-as-a-config-source.md": {
     "hash": "c7d2e8dbb1e8983e879540c135ccc343a92cd321de3c7f72cf6f595abad0bc03",
     "summary": "Move editor invocation into startup pipeline to treat it as a config source, fixing phantom deltas."
   },
   "081-decompose-tool-enable-into-state-and-allow_toggle.md": {
-    "hash": "c3fd1931d5ac8f9fa534b5f16bae3fe8cb05ee28e43be7c4b617768c326f248c",
+    "hash": "1de1aa963258c8ad189bbc1c23bf438d8d51a6f7232c745ce43293c8e592a761",
     "summary": "Split tool enable into state and allow_toggle to fix directive bugs and eliminate variant proliferation."
   },
   "082-unified-inquiry-event-recording.md": {
@@ -420,7 +420,11 @@
     "summary": "Allocate collision-free conversation IDs by scanning for occupancy and claiming the first unoccupied slot under a stable lock."
   },
   "105-mid-conversation-operator-directives.md": {
-    "hash": "0ac17903569c11f2c6e09ccfc1a349aa708c2762fe815d1cf195a56fbd240726",
+    "hash": "84bf892d29776fa9b7be919075187351874e85215673ebf94bcd4e62bf504eef",
     "summary": "Mid-conversation system and tool changes anchored in message stream preserve Anthropic's prompt cache via positional directives."
+  },
+  "107-discoverable-tools.md": {
+    "hash": "9de2d1264370d0bfac642d3ab503c3c7c3cd79b1262d9644270356b0f765acd3",
+    "summary": "JP adds discoverable tool state: tools in catalogue but not offered until model searches, keeping large MCP catalogues out of context."
   }
 }

docs/rfd/078-tool-config-mutation.md

@@ -5,6 +5,7 @@
 - **Authors**: Jean Mertz <git@jeanmertz.com>
 - **Date**: 2026-04-17
 - **Requires**: [RFD 076], [RFD 070]
+- **Extended by**: [RFD 107]
 
 ## Summary
 
@@ -1298,3 +1299,4 @@ Depends on: Phase 2.
 [RFD 042]: 042-tool-options.md
 [RFD 070]: 070-negative-config-deltas.md
 [RFD 076]: 076-tool-access-grants.md
+[RFD 107]: 107-discoverable-tools.md

docs/rfd/081-decompose-tool-enable-into-state-and-allow_toggle.md

@@ -5,6 +5,7 @@
 - **Authors**: Jean Mertz <git@jeanmertz.com>
 - **Date**: 2026-05-11
 - **Extends**: [RFD 008]
+- **Extended by**: [RFD 107]
 
 ## Summary
 
@@ -883,3 +884,4 @@ Phase 4 (tests and RFD updates) lands in the same PR.
 [RFD 078]: 078-tool-config-mutation.md
 [RFD 079]: 079-config-sources-and-load-order.md
 [RFD 083]: 083-built-in-ask_user-tool-for-assistant-initiated-inquiries.md
+[RFD 107]: 107-discoverable-tools.md

docs/rfd/105-mid-conversation-operator-directives.md

@@ -207,6 +207,14 @@ but it cannot *withdraw* a tool, and it answers a different question: which
 tools should the model pick from a large uncurated catalogue.
 Deferred to its own RFD.
 
+That deferral is not neutral, and the follow-up should know why.
+This RFD leaves discovery nowhere to operate: the initial directive offers every
+enabled tool and withdraws every disabled one, so a reference to an enabled tool
+is a no-op and a reference to a disabled one bypasses the user's intent.
+Discovery becomes meaningful only once tool enablement grows a third state — in
+the catalogue, permitted, not offered until the model asks for it.
+Designing that state, not the search itself, is the follow-up's real work.
+
 **Unify on `tool_reference` and drop directives.** Rejected: there is no
 "unreference" block, so mid-conversation disabling would still rewrite the
 prefix.