MCP budget: size get_active_queries defaults to the response budget (#4198)

[erikdarlingdata]

Part of #4198.

Why

get_active_queries had no default response-size budget. A busy production store's default call (limit 50, hours_back 1) measured 56,586 bytes on 2026-09-25, well over the shared 32 KB McpResponseBudget.DefaultBytes.

Unlike get_deadlock_detail (#4254), this tool has no single dominant wide field. Each row carries twenty-three fields. The fixed columns (collection_time, every wait/blocking/memory field, login_name, host_name, program_name) add up across the page even before query_text is counted. A synthetic 50-row page of this shape measured 81,489 bytes with the old code. Every optional field was populated, a fifth of rows carrying a long literal list. Isolating the fixed-column cost: even at query_text = "", 50 rows of this shape run close to the budget on their own.

What changes

Both Darling (DarlingMcpSessionTools.GetActiveQueries) and Lite (McpSessionTools.GetActiveQueries), matching the shape #4254 established:

• Default limit comes down from 50 to 25, sized so a full default page of wide rows fits the budget.
• query_text previews to 500 characters by default (down from a flat, unconditional 2000-character cap) with a new per-row query_text_truncated flag.
• A new full_text opt-in argument (default false) returns every row's whole text. This is the same shape get_store_query_stats' full_text and get_deadlock_detail's full_graph use.
• truncated / total_snapshots already told a caller who wants more to raise limit or narrow hours_back. That is unchanged.

Lane W2 (this update)

• The web viewer's Active Queries tab still gets 2,000-character query text. Before this PR, every caller (MCP and web) got McpHelpers.Truncate(r.QueryText, 2000). A naive drop of the MCP default to 500 also shrinks the web page's text. DarlingMcpSessionTools.GetActiveQueries now splits in two: a public MCP-facing method, and an internal overload that takes an explicit query-text preview length. This is the same split Trend tools return every raw point: get_file_io_trend sends 12,451 points / 1.4 MB for one server at defaults, more than an LLM client's context #3897's trend tools use for their budgets. The public method passes 500, or no truncation at all when full_text is set. DarlingWebEndpoints' /api/read row calls the internal overload with 2000, the pre-MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198 amount. query_text_truncated stays honest on both paths: false whenever the preview length is null, or the text does not exceed it. Lite is unaffected. Its UI reads LocalDataService directly, not this MCP tool. A grep confirms this: the only other reference to McpSessionTools in Lite/ is its MCP host registration.
• Added a no-rig pin, DarlingWebEndpointsTests.ReadEndpoints_ActiveQueries_KeepsTheTwoThousandCharacterWebPreview, that reads DarlingWebEndpoints.cs's source text and fails if the get_active_queries row stops calling the overload with the literal 2000. Verified it fails against the pre-fix row shape (temporarily reverted, ran the test, saw [FAIL], then restored the fix and reconfirmed [PASS]).
• Renamed the opt-in argument from full_query_text to full_text, on both Darling (DarlingMcpSessionTools.cs) and Lite (Lite/Mcp/McpSessionTools.cs). This matches get_store_query_stats' full_text, the MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198 precedent named in the issue, and the other MCP read tools have no default response-size budget: at default arguments 12 tools return >50 KB and 3 return >100 KB for one server, more than an agent client's per-result cap #4198 tools. Each parameter's own description text is unchanged, so its pinned byte length in McpToolsListBudget/*SessionTools.txt stays 84 on both products. The shorter property name shrinks the served tools/list JSON by 6 bytes on each product. That saving is banked by lowering TotalCeilingBytes to the newly measured total (171,906 Darling, 90,057 Lite), with a new change-log comment line on each.
• Added query_text_truncated to McpPayloadContractCensusTests' new FieldPreviewCutKeys roster. This is added here because MCP budget: preview the deadlock graph XML by default on get_deadlock_detail (#4198) #4254 had not merged yet. Expect a merge conflict with it that just combines both tools' entries.
• This tool is not on A budget pin over every MCP read tool (part of #4198) #4224's ExemptOffenders roster, so McpReadToolBudgetLiveTests (also not merged yet) was left untouched.

The tool's served head description is unchanged at 616 bytes. The new preview/opt-in note went after the <<GUIDE>> marker instead. A separate pin, McpToolGuideTests.EveryConvertedHead_StaysAtOrUnder620Characters, caps converted heads at 620 characters, tighter than the general 1,000-byte cap this repo's skill notes describe.

Test plan

New live test Darling/Darling.Tests/DarlingMcpActiveQueriesBudgetLiveTests.cs plants 50 rows shaped like a busy server. It is its own file, not the shared seeding helpers, because many #4198 lanes ran against this rig tonight.

• Before the fix: 81,489 bytes (over budget).
• After the fix: 29,821 bytes (under the 32,768-byte budget, about 2,900 bytes of headroom).
• full_query_text: true opts back into the whole text, and query_text_truncated is never true on that call.

New Lite test GetActiveQueries_Default_StaysUnderResponseBudget_WithFiftyRealisticRows in Lite.Tests/McpPageContractTests.cs asserts the same shape and the same opt-in behavior (DuckDB, no rig needed).

• Darling.Tests build: 0 Warning(s), 0 Error(s).
• Lite.Tests build: 0 Warning(s), 0 Error(s).
• Full Darling.Tests suite (rig on port 55983, UTC, seeded fresh darlingtest): 13,849 total, 5 failed, 47 skipped, 1 not run. Every failure traced back to this change. All fixed in a follow-up commit: a param-order pin, the 620-char head cap noted above, and a <see cref> type/file error.
• Targeted re-run after the fixes (DarlingMcpSessionToolsSurfaceAndSqlTests, McpToolGuideTests, McpToolGuideHeadsSqlCoreActiveQueriesTests, DocCommentHygieneTests, McpToolsListBudgetTests, McpPayloadContractCensusTests, DarlingMcpSessionToolsTests, DarlingMcpActiveQueriesBudgetLiveTests, McpFilterSemanticsLivePostgresTests): 181 total, 0 failed.
• The full Darling.Tests suite was not re-run a second time after the follow-up fixes, because of the context budget. The targeted re-run above covers every class that failed plus the directly related ones. Worth a final full run before merge.
• Full Lite.Tests suite (after the same fixes): 5,340 total, 0 failed.
• git merge origin/dev done. One conflict, in McpToolsListBudgetTests.cs's ceiling constant, resolved by keeping both change-log comments and re-measuring the constant.

Lane W2 (this update, no rig: DARLING_TEST_PG unset, live classes skip)

• Darling.Tests build: 0 Warning(s), 0 Error(s).
• Lite.Tests build: 0 Warning(s), 0 Error(s).
• New pin (ReadEndpoints_ActiveQueries_KeepsTheTwoThousandCharacterWebPreview) proven to fail on the pre-fix row shape, then confirmed passing again after restoring the fix.
• Targeted Darling.Tests run: every class this update touches, every class reading DarlingWebEndpoints.cs (50 classes, grep-derived), plus McpToolsListBudgetTests, McpPayloadContractCensusTests, and DocCommentHygiene. 1,506 total, 0 failed, 33 skipped (all live, no rig).
• Full Lite.Tests suite, run once: 5,340 total, 0 failed, 0 skipped.
• Full Darling.Tests live suite against a real rig: CI decides (no rig in this lane).

CHANGELOG entry

SECTION: Fixed
ENTRY:

• get_active_queries stays under its response-size budget by default ([MCP budget: size get_active_queries defaults to the response budget (#4198) #4261]) - A default call on a busy server returned well over the 32 KB target. The number of fields per row and an uncapped query text preview were the cause. The default page size is now 25 rows (was 50). Query text now previews to 500 characters with a query_text_truncated flag. A new full_text argument opts back into the whole text. The web viewer's Active Queries tab keeps its pre-existing 2,000-character preview.
  REF:
  [MCP budget: size get_active_queries defaults to the response budget (#4198) #4261]: MCP budget: size get_active_queries defaults to the response budget (#4198) #4261

---

Generated with Claude Code

https://claude.ai/code/session_01FVjn4PBJN71NQXdFo6ZxNQ