[Gintm]

Why

Every pinned tool definition is sent on every turn of every session: the 53 pinned tools cost 87.8K characters (~22K tokens). Several are almost never called, while three hidden tools are reached through search_tools by 6-13% of sessions, each search costing a model turn and ~14.7K characters (Axiom zernio-mcp, 2 h, 1,226 sessions; Sentry, 7 d).

What

• Pins messages_list_inbox_conversations (13% of sessions), media_get_media_presigned_url (8%), messages_get_inbox_conversation_messages (6%).
• Unpins 13 tools with almost no use: the 6 queue_*, the 2 tracking_tags_*, usage_get_usage, validate_post_length, mentions_reply_to_mention, posts_bulk_upload_posts, accounts_move_account_to_profile, account_groups_list_account_groups. They stay reachable through search_tools / call_tool and by direct name.
• Clients now see 44 tools and 75,530 characters instead of 55 and 89,417 (-15.5%).
• An unknown tool name now also suggests pinned tools that share its words (search only indexes hidden tools), so a guess like inbox_list_conversations still points to the now-pinned tool.
• call_tool still reaches newly pinned tools, so agents that already call them by name through call_tool keep working (test).

🤖 Generated with Claude Code

[Gintm]

Review (code-review high --fix)

Finding	Result
Pinned-name suggestions only split guesses on underscores (camelCase, kebab-case never matched)	Fixed, with two guessed-name test cases
The pinned-list comment quoted measured shares that go stale	Fixed: the comment keeps the constraint, the numbers live in this PR
A pinned word match can push a spelling-only suggestion out of the top 3	Skipped: it needs a guess sharing two words with a pinned tool, which is then the likelier intended tool
The call_tool reachability test reads a validation message	Skipped: it asserts on the argument name, the stable part

QA

Check	Result
Real tool catalog (develop + this branch): what clients list	55 tools / 89,417 chars → 44 tools / 75,530 chars
Every pinned name is registered	Yes (test)
call_tool still dispatches to a newly pinned tool (agents that call messages_* by name)	Yes: reaches the API, no Unknown tool (test and local run)
An unpinned tool is still found by search_tools and callable	Yes: queue_list_queue_slots (test and local run)
Full suite	1,805 passed

Metrics

Metric	Before (Axiom zernio-mcp, 2 h)	Expected after
Tool list sent every turn	89,417 chars	75,530 chars (-15.5%)
search_tools calls per 1,000 sessions	386 (473 searches / 1,226 sessions)	lower; sessions that only needed the inbox or presigned URL reads stop searching
Sessions reaching messages_* / presigned URL through call_tool	158 / 69 / 94	move to via=direct

Post-deploy check: same 2 h window a day later in zernio-mcp: searches per 1,000 sessions, and the via split for the three newly pinned tools; control: total sessions and calls of untouched pinned tools (posts_list, accounts_list).