From 440bff7bdd585d4b058bc2c691f0288fc8657f7f Mon Sep 17 00:00:00 2001 From: steve Date: Fri, 10 Jul 2026 16:33:31 +0000 Subject: [PATCH] test(agent): guard the no-preamble-veto-for-long-answers case; reword comments (gadfly) Add "long answer opening with a conversational word is still recovered" (regression guard); retarget the skipped-preamble case to the borderline band; drop the external "#N" annotations in favor of self-explanatory comments. Co-Authored-By: Claude Opus 4.8 (1M context) --- agent/finalize_test.go | 45 ++++++++++++++++++++++++++++++------------ 1 file changed, 32 insertions(+), 13 deletions(-) diff --git a/agent/finalize_test.go b/agent/finalize_test.go index a96fac5..a8948cb 100644 --- a/agent/finalize_test.go +++ b/agent/finalize_test.go @@ -45,22 +45,22 @@ func TestIsCitationsOnly(t *testing.T) { {"sources-md-links", "Sources: [pcprice.watch](https://pcprice.watch/x), [ebay](https://ebay.com/1).", true}, {"lowercase-bare-url", "sources: see https://example.com/a", true}, {"bold-label-colon-inside", "**Sources:** [a](https://a), [b](https://b)", true}, - {"bold-label-colon-outside", "**Sources**: [a](https://a), [b](https://b)", true}, // #4: colon after the bold + {"bold-label-colon-outside", "**Sources**: [a](https://a), [b](https://b)", true}, // colon after the closing ** {"references-dash", "References — [a](https://a)", true}, {"citations-label", "Citations: https://x/y", true}, {"leading-list-marker", "- Sources: [a](https://a)", true}, - {"atx-heading", "## Sources: [a](https://a), [b](https://b)", true}, // #8: heading + its space + {"atx-heading", "## Sources: [a](https://a), [b](https://b)", true}, // ATX heading marker + its trailing space {"further-reading", "Further reading: https://example.com/deep-dive", true}, - {"annotated-multi-source", "Sources: [pcprice.watch](https://a) (tracker), [eBay](https://b) (sold), [bestvaluegpu](https://c) (retail), [resaleprices](https://d) (asking).", true}, // the real #1418 shape - {"backref-plus-links-is-citations", "References: as noted above, [pcprice.watch](https://pcprice.watch/x).", true}, // #13: still a citations addendum + {"annotated-multi-source", "Sources: [pcprice.watch](https://a) (tracker), [eBay](https://b) (sold), [bestvaluegpu](https://c) (retail), [resaleprices](https://d) (asking).", true}, // the reported issue-1418 shape + {"backref-plus-links-is-citations", "References: as noted above, [pcprice.watch](https://pcprice.watch/x).", true}, // a back-ref phrase inside a real sources list is still citations {"empty", "", false}, {"label-but-no-link", "Source: internal analysis, no URL here", false}, {"prose-then-sources", "It sells for ~$2,700. Sources: [a](https://a)", false}, // answer first → not a pure addendum - {"source-led-prose-answer", "Source: According to https://cdc.gov the flu vaccine is 40-60% effective, and the CDC recommends annual vaccination for everyone over six months old.", false}, // #5: prose that merely opens with a label + {"source-led-prose-answer", "Source: According to https://cdc.gov the flu vaccine is 40-60% effective, and the CDC recommends annual vaccination for everyone over six months old.", false}, // a prose answer that merely opens with a "Source:" label {"mentions-sources-midsentence", "The sources of the leak were never confirmed.", false}, {"link-without-label", "Here is the link you asked for: [a](https://a)", false}, - {"bare-domains-out-of-scope", "Sources: pcprice.watch (used ~$200), ebay.com (sold listings)", false}, // #7: no scheme/link signal + {"bare-domains-out-of-scope", "Sources: pcprice.watch (used ~$200), ebay.com (sold listings)", false}, // bare domains: no scheme or markdown link to key on {"crisp-number", "42", false}, } for _, c := range cases { @@ -96,8 +96,13 @@ func TestFinalOutput(t *testing.T) { longSources := "Sources: [pcprice.watch](https://pcprice.watch/gpu/rtx5090) (tracker), [ebay](https://www.ebay.com/sch/rtx5090) (sold), [newegg](https://newegg.com/rtx5090) (retail), [pcpartpicker](https://pcpartpicker.com/rtx5090) (history)." // A substantive answer that merely OPENS with "Source:" (not a bare list). sourceLedAnswer := "Source: https://nvd.nist.gov/vuln/detail/CVE-2024-1234 — this is the authoritative NVD entry for the vulnerability, rated CVSS 9.8 critical." - // A >=200-byte planning preamble (must be skipped during recovery). - longPreamble := "I'll look up the current eBay sold listings, then cross-reference the pcprice.watch tracker and a couple of retail sources, compare the medians across all of them, and put together a clear price range for you before I give the final number." + // A borderline-band (80–200 byte) turn that opens like a planning preamble: + // it clears the floor, but the preamble filter still vetoes it (the filter + // applies only in the borderline band; a >=200-byte turn is accepted as-is). + preambleTurn := "Let me look that up across a few different sites and then compile the full comparison for you here." + // A >=200-byte real answer that merely OPENS with a conversational word + // ("Sure,"). The preamble filter must NOT veto it (gadfly regression guard). + longConversationalAnswer := "Sure, here's the rundown: it currently sells for about $2,700 used on eBay, typically $2,400 to $2,900 depending on condition and bundle, with the sealed Founders Edition commanding the top of that range while used AIB cards go a bit lower." tests := []struct { name string @@ -222,21 +227,35 @@ func TestFinalOutput(t *testing.T) { want: sourceLedAnswer, }, { - // A long (>=200) planning preamble must be skipped during recovery - // (preamble check applies before the length shortcut); the older - // real answer is recovered instead. - name: "long preamble is skipped; older real answer recovered", + // A borderline-length turn that opens like a preamble is vetoed + // during recovery; the older real answer is recovered instead. (A + // >=200-byte turn would be accepted verbatim — see the next case.) + name: "borderline preamble is skipped; older real answer recovered", msgs: []llm.Message{ llm.UserText("q?"), asst(conciseAnswer, cite...), llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}), - asst(longPreamble, cite...), + asst(preambleTurn, cite...), llm.ToolResultsMessage(llm.ToolResult{ID: "c2", Name: "cite", Content: "ok"}), asst(sources), }, terminal: sources, want: conciseAnswer + "\n\n" + sources, }, + { + // Guards the gadfly regression: a LONG (>=200-byte) front-loaded + // answer that merely opens with a conversational word ("Sure, …") + // must still be recovered — the preamble filter must not veto it. + name: "long answer opening with a conversational word is still recovered", + msgs: []llm.Message{ + llm.UserText("q?"), + asst(longConversationalAnswer, cite...), + llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}), + asst(sources), + }, + terminal: sources, + want: longConversationalAnswer + "\n\n" + sources, + }, } for _, tc := range tests { t.Run(tc.name, func(t *testing.T) {