[argus] patch #46: tool_search.exclude — deferral opt-out for hot MCP tools#10
Merged
Merged
Conversation
…r hot MCP tools Deferral is all-or-nothing over MCP tools; the canary bench (goldset-wide-v2, 15 runs/arm) showed the promotion round-trip doubling mean wall clock because the hot knowledge tools (pythia_query, kb_query, ...) are used in most turns. exclude patterns (fnmatch on final prefixed names) keep hot tools always-bound while the long tail defers. Fail-closed guard treats all-excluded as a valid empty setup. Threaded through all four assemble call sites. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…mangling) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of the Acropolis capability rewrite P1 canary gate.
Deferral (tool_search.enabled) is all-or-nothing over MCP-tagged tools. The canary bench (goldset-wide-v2, 15 runs/arm, 2026-07-23) measured the cost: mean wall clock doubled (13.0s -> 26.3s) because the hot knowledge tools (pythia_query, kb_query, entity_search, code_search_*) are used in most turns — the tool_search promotion round-trip on a local 27B costs more latency than their schemas cost context.
ToolSearchConfig.exclude: list[str]— fnmatch patterns against FINAL (server-prefixed) tool names. Excluded MCP tools skip the deferred catalog and stay always-bound; the fail-closed guard treats all-excluded as a valid empty setup. Threaded through all four assemble_deferred_tools call sites (lead x2, embedded client, subagent executor).Config (atlas-template) will carry:
Tests: test_deferred_setup.py::TestExclude (4 new); deferred suites 34/34 green. PATCHES.md updated same-commit per the hard rule. Exit: upstream per-server/per-tool deferral control.
Canary re-bench happens from this branch before merge; pin bump + fleet follow after the gate.
🤖 Generated with Claude Code
https://claude.ai/code/session_018KmwWGKMPVxabtjafVYTLz