MCP budget: get_blocking / get_blocked_process_reports default under 32 KB (#4198)

[erikdarlingdata]

Part of #4198.

Why

get_blocking (Darling) / get_blocked_process_reports (Lite) had no response-size budget. Each row
carries about 37 fields. That includes isolation levels, client app/host/login for both sides, six
last-tran/last-batch stamps, and a dedup_key, before either SQL text column is counted. So the default 30-row page was already heavy.

A live test seeded 30 rows with realistic-but-modest SQL text: sub-2000 characters, so never truncated under the old cap. It measured 89,096 bytes for blocked/blocking text combined, 2.7x the shared 32 KB McpResponseBudget.DefaultBytes. The wide row shape, not the two text columns alone, was most of that weight.

What changes

• Darling (DarlingMcpBlockingTools.cs, get_blocking): default row limit halves, 30 -> 15
  (DefaultLimit). blocked_sql_text / blocking_sql_text preview to 150 chars
  (SqlTextPreviewLength) instead of the old fixed 2000, each with its own *_truncated flag. A new
  full_text opt-in returns both columns whole. A dedup_key call, which already names one incident,
  always gets the whole text too, the same exemption get_deadlock_detail uses for full_graph.
• The public MCP method is now a thin wrapper. It calls a new internal overload that takes an explicit
  sqlTextPreviewLength, 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 TrendBudget.Chart.
• Web viewer: the /api/read row for get_blocking now calls that internal overload directly. It
  passes the OLD row limit (30, already hardcoded there) and the OLD 2000-char preview
  (WebSqlTextPreviewLength). The web page's result is unchanged. I traced both call sites through
  panels.js's readTool(), which hits GET /api/read/get_blocking: wwwroot/js/pages/server-tabs.js's
  Blocking table and wwwroot/js/view-templates.js's Custom Views template. Both already send
  limit: 30 explicitly and neither sends full_text, so both keep today's result with no JS changes.
• Lite (McpBlockingTools.cs, get_blocked_process_reports): same default-limit halving, same
  150-char preview, same full_text opt-in. Lite has no dedup_key on this tool, so it needs no
  exemption for it. Lite has no web viewer to pin.
• Census: added blocked_sql_text_truncated / blocking_sql_text_truncated to
  McpPayloadContractCensusTests.FieldPreviewCutKeys on both SKUs.
• tools/list budget: DarlingMcpBlockingTools.txt and Lite's McpBlockingTools.txt get the new
  full_text param line and the tool's grown byte count. TotalCeilingBytes is raised by the exact
  measured growth on each SKU (+391 Darling, +271 Lite), with a change-log line above each constant.
• Lite parity test (this lane): Lite.Tests/McpPageContractTests.cs gained
  GetBlockedProcessReports_Default_StaysUnderResponseBudget_WithThirtyWideReports, next to lane TB's
  GetDeadlockDetail_Default_StaysUnderResponseBudget_WithFiveWideGraphs precedent. Plants 30 reports
  with both text fields near 700-870 characters and every other field populated. Asserts the default call
  previews under McpResponseBudget.DefaultBytes, and that full_text still returns the whole text on
  every row.
• Head-length fix (this lane): the full Darling suite (step 4) found a defect in lane TI's own
  change. It had pushed get_blocking's served head (the part before <<GUIDE>>) from 620 to 781
  characters, over McpToolGuideTests' D3 cap. Adding the new preview sentence there was the cause.
  Moved that sentence into the tail instead, the same way get_plan_corrections' full_text note lives
  in its tail and not its head. Every HeadFacts substring McpToolGuideHeadsBlockingSlotsTests requires
  for get_blocking stays verbatim in the head. The head now measures 576 served characters. Banked
  DarlingMcpBlockingTools.txt's per-tool ceiling from 781 down to 576 to hold the saving.

#4224 note

Your tool (get_blocking) is on #4224's ExemptOffenders roster. Per this lane's brief, #4224 has not
merged and its wait was lifted by the coordinator. McpReadToolBudgetLiveTests is not on dev, so this
PR does not create or edit it. Whichever of #4224 or this PR merges second must remove this tool's row
from that roster. After this PR, get_blocking is no longer an offender.

Test plan

• Live test Darling/Darling.Tests/DarlingMcpBlockingBudgetLiveTests.cs (own file, own seed, not the
  shared DarlingMcpBlockingToolsLivePostgresTests seeding). It plants 30 rows with realistic 701/872-char
  blocked/blocking SQL text.
  • Before this fix: 89,096 bytes (measured, then reverted to confirm before writing the fix).
  • After this fix: 26,745 bytes, under the 32,768-byte budget. full_text: true and a dedup_key call
    both return the whole text. The default call previews with *_truncated: true.
• Lite.Tests/McpPageContractTests.cs's new GetBlockedProcessReports_Default_... test: plants 30
  reports with ~700-870-char text on both fields. Measured 23,969 bytes at the default call, under the
  32,768-byte budget. full_text: true returns all 30 rows' text whole and untruncated. Proved once that
  the test fails against the old shape. Temporarily reverted SqlTextPreviewLength to 2000 and limit's
  default to 30, rebuilt, and confirmed it then returns 30 rows instead of 15. Reverted back and confirmed
  green. 24/24 passed in McpPageContractTests.
• McpPayloadContractCensusTests (Darling.Tests): 68/68 passed after adding the two new keys and
  merging in origin/dev's own new top_query_text_truncated key alongside them.
• McpToolsListBudgetTests (Darling.Tests and Lite.Tests): both 5/5 passed. Resolved the merge
  conflict on TotalCeilingBytes by keeping every change-log comment from both sides. Set each constant
  to the measured total: Darling 172,758 bytes, Lite 90,689 bytes.
• git merge origin/dev: three conflicts (McpPayloadContractCensusTests.cs,
  McpToolsListBudgetTests.cs on both SKUs), resolved as above.
• Full Darling.Tests suite, once, against a fresh darlingtest on a UTC rig: 13,887 total, 2 failed,
  47 skipped, 1 not run. Both failures are pre-existing, not this branch's:
  • EventWindowedReadsAreBoundedLivePostgresTests.TheFlooredReads_PlanAtMostThreeChunksForA24HourWindow_AgainstDevPostgres
    fails the same way alone on a freshly created database. Dev's own latest CI Build run (commit
    f5e001fe, the exact commit this branch merged) fails it too: Darling PG tests (0), run
    36115696476. It is unrelated TimescaleDB chunk-floor planning for job_history, not blocking.
  • CaptureDownChunkOrderTests.TheShippedRead_ExecutesOnlyTheNewestChunk_AndTheNewestRunDecides_AgainstDevPostgres
    passed cleanly when re-run alone on a fresh database. It only failed inside the full 13,887-test run.
    That points at shared-state or timing inside that one big run, not a defect. Neither file is one this
    PR touches.
• Full Lite.Tests suite, once: 5,346 total, 0 failed.

Deferred

Nothing filed as a separate issue. The two full-suite failures above are pre-existing on dev. They are
outside this PR's subsystem: TimescaleDB chunk-floor planning for job_history/collection_log, not
blocking. I verified them against dev's own latest CI rather than filing anything, since diagnosing that
subsystem is past this lane's brief. Worth a census entry at the next wave boundary.

CHANGELOG entry

SECTION: Fixed
ENTRY:

• MCP get_blocking (Darling) and get_blocked_process_reports (Lite) cap their response size ([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]): A default call measured 89,096 bytes on a 30-row test seed, over the 32 KB limit. The default row limit drops from 30 to 15. The blocked and blocking SQL text come back as a 150-character preview instead of a 2,000-character cut, and each has its own *_truncated flag. A new full_text option returns both texts whole. On Darling, a call with dedup_key names one incident, so it always gets the whole text. The web viewer's Blocking table and Custom Views keep 30 rows and the 2,000-character text.
  REF:
  [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]: 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

Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3