Repository navigation
MCP reads window only backwards from now, so no caller can ask about a past incident #2495
Description
Split out of #2494 rather than grown into it.
What
get_collection_log's reader is bounded on both sides — it takes a window start and a window end — and the viewer's Collection Log tab uses that to honour a custom From/To range. The MCP tool exposes only hours_back, always anchored to end = now.
So the capability exists in the read and is unreachable through the tool. An agent asked "what were the collectors doing during the incident last Tuesday?" can only say "the last N hours" and widen until Tuesday falls inside — dragging along everything since.
I claimed the past-incident case worked in #2494's description. It does not, and this is the correction.
Scope, because it is not one tool
get_collection_log is where it was noticed, not where it starts. The same hours_back-from-now shape is the norm across the read surface — get_wait_stats, get_query_store_top, get_blocking, get_alert_history and others all anchor to now. Any read whose stored data outlives its default window has the same gap.
That makes this a surface-wide parameter convention question, not a bug in one tool:
- Add optional
from/toto reads that already window, keepinghours_backas the default path. Backward compatible; the widest change. - Add an
as_ofanchor — keephours_backas the span and let a caller move the endpoint off "now". Smaller and covers the incident case, which is nearly always "N hours around a known time". - Leave it. Callers widen and filter client-side. Cheapest, and gets worse the more history the store keeps — which is the direction it is going.
Option 2 is the smaller change and covers the actual use case, but this wants deciding once for the surface rather than per tool, which is why it is an issue and not a patch.
Not a workaround
Widening hours_back until the window reaches back far enough is not equivalent: it changes what a top-N returns, what an average is taken over, and how much a capped read truncates. For an aggregate read, a wider window is a different answer, not the same answer with more rows.
Activity
Refuse an ambiguous as_of, and fix the pins as_of moved (#2495)
Make the anchor pins fail loudly when reflection comes up empty (#2495)
Anchor a windowed read at a past incident, not only at now (#2495)
Refuse an ambiguous as_of, and fix the pins as_of moved (#2495)
Make the anchor pins fail loudly when reflection comes up empty (#2495)
Correct the read counts the rebase moved (#2495)
Say in the README that a windowed tool can be anchored (#2495)
Say why three reads resolve the anchor without ValidateWindow (#2495)
Anchor a windowed read at a past incident, not only at now (#2495)
Refuse an ambiguous as_of, and fix the pins as_of moved (#2495)
Make the anchor pins fail loudly when reflection comes up empty (#2495)
Correct the read counts the rebase moved (#2495)
Say in the README that a windowed tool can be anchored (#2495)
Say why three reads resolve the anchor without ValidateWindow (#2495)
Anchor the four reads #2505 added, and re-pin their contracts (#2495)
Make eight tools that ignored their own anchor use it, and pin the ca…
Stop the server-local anchor test passing at offset zero (#2495)
Reunite a stranded summary with the member it describes (#2495)
Merge pull request #2504 from erikdarlingdata/feat/2495-as-of
erikdarlingdata commentedon Aug 22, 2026
Done in #2504.
as_of— an optional ISO-8601 UTC anchor that moves the window's END off "now" whilehours_backstays its LENGTH. Applied to 57 Darling / 50 Lite reads with identical name, semantics and description on both, plus all 57 web catalog descriptors and dispatch lambdas, so it is reachable from/api/read/*and not only from MCP.Option 2 was right, and for a better reason than this issue gave.
hours_backwas never a window — it was half a window plus an implicitend = now.as_ofdoes not add a knob so much as make the existing one honest. Option 1 would have given ~100 reads three time parameters with a validity rule between them, including a state (from+to+hours_backall set) with no defensible answer.Left unanchored, deliberately
analyze_serveralso persists. Filed as The analysis family cannot be anchored at a past window, and it is the one an incident starts with #2506.get_pvs_statsandget_fleet_overview— each mixes a latest-snapshot measurement with a windowed one, so anchoring half would return a result whose two halves describe different instants.Both exclusions are named in the instructions and pinned by a test that lists them, so the set cannot quietly grow.
Two findings worth more than the feature
Eight tools advertised
as_of, validated it, refused bad input correctly — and then queriedDateTime.UtcNowanyway. All eight came from resolving a rebase indev's favour. All compiled and passed a green 5,600-test suite. One was worse than a dropped anchor: it passed already-UTC instants into server-local slots, shifting the window by the monitored server's offset even when unanchored. A tool that validates an anchor and then ignores it is worse than one that never offered it, because the caller is told the window moved.The pin written to catch that was itself vacuous at first. Reintroducing one of the bugs left it green, because the body still contained
windowEndfrom its ownout var. It is now an absolute — an anchored tool may not nameDateTime.UtcNowat all — verified by reintroducing one bug of each form and watching both fail. It also asserts it examined at least 40 tools, so a broken scan fails instead of passing for free, and derives its list of anchorable service methods from shipped source rather than a transcribed list.Closing.
Merge dev, and take the window anchor the heatmap needs most (#2484)
Bound the findings window only when it is anchored, and pin why (#2506)
Thread the window anchor into the analysis engine, and refuse to pers…