A clarification request that also carries an intent clause ("Let me know what you're after and I'll dig in.") matched INTENT_SIGNAL, so the tool loop re-prompted it and the user saw the same question twice. INTENT_SIGNAL already exempts "let me know", but the lookahead clears only that one alternative and the trailing "I'll" still fired.

is_short_intent_without_action now returns False whenever _CLARIFICATION_REQUEST matches, so a turn waiting on the user is never treated as a stall. An earlier draft only suppressed an ask that preceded the intent match. Over the 600 real turns in tests/data/plan_vs_answer.jsonl (both the text and retry_text fields) that comparison cost two correct suppressions and gained nothing, so it is gone. The helper is shared, so the GGUF, safetensors and MLX loops all pick it up.

_should_suppress_forced_no_tool_output needed the same gate. It early-returns on _FORCED_PLAN_INTENT before consulting the helper, so a re-prompted retry reading "I need to use the web tool for this. Could you clarify which site you mean?" was discarded whole and the user lost the question entirely.

Three tests cover the three paths, and each fails if the gate is removed: the shared classifier in test_safetensors_tool_loop.py, the discard branch in test_forced_turn_keeps_a_retry_that_asks_the_user_a_question, and the GGUF loop itself in test_clarification_request_is_not_nudged, which asserts a single generation.

On the 300-turn corpus of finished answers, wasted nudges drop from 5 to 4 and discarded retries from 1 to 0. tests/data/refactor_guard baselines carry the new module symbol; only those entries changed, since a full snapshot would sweep in unrelated drift already present on main.

Fixes unslothai#8907

for more information, see https://pre-commit.ci

for more information, see https://pre-commit.ci

Two gaps in _CLARIFICATION_REQUEST. A question that carries the ask on its own ("Which repository should I inspect?", "Do you mean the package or the UI?") went unmatched, so the duplicate-question regression survived for those phrasings. And "let me know whether ..." reads as a closing remark exactly like "let me know if ...", so a plan behind one skipped its nudge.

"let me know when ..." keeps matching: it waits on the user rather than closing the turn.

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

Patching one question word at a time kept leaving gaps, most recently "Where should I search?" and "How would you like me to proceed?". The what/which alternatives collapse into a single interrogative rule covering what, which, where, when, who and how, which is both shorter than the two it replaces and closed to the same class of gap.

Word order keeps advice out: "what you should do" needs the pronoun before the modal, while the rule wants the modal first.

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

"What version are you using?" is a blocking question the interrogative arm could not reach, because the auxiliary list held only should, would, shall, do and did. Added is, are, was, were, can, could and have.

Word order still keeps statements out: "I'll tell you what the results are" has no pronoun after the auxiliary.

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

"Are you using Windows or Linux?" and "Is this about the package or the UI?" ask for a detail without a wh-word, so the interrogative arm could not reach them.

Scoped to is/are/was/were with a sentence-start anchor. Auxiliaries like do and can are left out because "Do you want me to run it?" asks permission to act, which is the stall the nudge exists to recover, and the anchor keeps ordinary prose out ("Here is that file you wanted").

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

A clarification under a heading ("One question:\nAre you using Windows or Linux?") slipped past the copular arm: the anchor accepted only . ! ? and `^` is not multiline, and model output is routinely formatted this way.

The anchor now also takes a colon, a newline and a list marker. The leading \b moved onto each word-initial arm, since a \b ahead of the group can never match at a newline preceded by punctuation, which left the newline anchor dead on arrival.

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

"Why do you need this information?" is the same class as the "what ... do you" forms already covered, so why joins the interrogative list.

The copular arm was too loose. "I'll search the advisories now. Is this package vulnerable to CVE-2026-1234?" restates the lookup the model just promised rather than asking the user for anything, and suppressing it left the plan unfulfilled. The arm now takes only "is/are/was/were you" and "is/are/was/were this|that|it about", both of which can only be about the user.

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

A turn that echoes the prompt before promising work (`You asked, "Can you provide the latest CUDA release?" I'll search the web now.`) matched on the quoted words, so the gate read the user's question as the model's own and the promised call never got its nudge.

Both call sites now blank double-quoted spans before searching. Quotes are common in real output, 69 of the 597 turns in tests/data/plan_vs_answer.jsonl carry one, and stripping them changes no classification there, so the counts stay at 4 nudged and 0 discarded.

…l credits to the user

Two gaps. "Which version of the package are you using?" has four words before its auxiliary and the filler bound stopped at two. And the quote strip added last round erased the model's own emphasised question (`One question: "Which repository should I inspect?"`), classifying what was left as a stalled plan.

_ECHOED_USER_QUOTE replaces the blanket span: a quote is skipped only when the model credits it to the user ("you asked", "your question"), which is the case that fix was for. It matches no turn in tests/data/plan_vs_answer.jsonl, and the wider filler changes no count either, so both stay at 4 nudged and 0 discarded.

…reach

Three gaps. "What's your operating system?" contracts the auxiliary and the interrogative arm required a space. "I'll search the web now. How can I find the latest release notes?" is the model narrating its own task, but `can i` read as a request for user input. And `To answer your question, one clarification: "Which repository should I inspect?"` let the attribution wildcard cross a clause boundary and erase a question the model wrote itself.

Only `should` now takes `i`, since "Which repository should I inspect?" defers the choice to the user while "How can I ..." does not. The attribution gap is down to light punctuation from a 40-character wildcard.

Corpus counts in tests/data/plan_vs_answer.jsonl are unchanged at 4 nudged and 0 discarded.

Ten review rounds each closed one phrasing and opened the next, and twice the new finding was a defect in the previous round's fix. Measuring the run explains why: rounds two onward grew the added pattern from 121 to 496 characters and moved the 300-turn corpus in tests/data/plan_vs_answer.jsonl by nothing at all. Every one of those was fitted to a constructed sentence.

The cause was a second regex, _CLARIFICATION_REQUEST, vetoing the whole response. INTENT_SIGNAL does not work that way: it scopes each exclusion as a lookahead on the alternative it qualifies, which is why _HELP_OFFER only applies after "I'll" or "let me". A whole-response veto has unbounded surface, so there was always another hole.

_CLARIFICATION_REQUEST, _ECHOED_USER_QUOTE and the llama_cpp.py changes are gone. _HELP_OFFER gains "dig in" and "help analyze", both guarded to fire only when no work follows, mirroring how the neighbouring "help you" arm defers to _ACTION_VERB. The "let me" arm reaches the same sign-offs, since "let me dig in" closes a clarification just as "I'll dig in" does.

Counts stay at 4 nudged and 0 discarded, from 5 and 1 before the PR. Every counter-example raised in rounds two through eleven passes without special handling, since all of them were regressions the veto introduced. The interrogative and copular suppressions are not carried over: a trailing lookahead cannot express them and none moved the corpus, so they stay the long tail of the original behaviour.

The fix now sits in the existing parity test's phrase lists, because INTENT_SIGNAL and the helper agree again.

Making the "to" optional on the let-me arm let the whole of _HELP_OFFER through, so "Let me assist by searching the web now." stopped being nudged even though it names work. Only the two sign-offs were meant to skip that gate.

_SIGN_OFF now carries them on its own lookahead, leaving the "to" requirement intact for the rest of _HELP_OFFER.

Corpus counts in tests/data/plan_vs_answer.jsonl stay at 4 nudged and 0 discarded.

"I'll dig in. Starting with a web search." satisfied the old guard, which only looked past whitespace and commas, so a turn that carried on with real work lost its only intent signal. The sign-offs now have to reach the end of the response.

\Z rather than $, since INTENT_SIGNAL is compiled with re.M and a plan continuing on the next line is still a plan.

Checked by sweep rather than by example: 486 opener, sign-off and continuation combinations agree with intent, and across 360 combinations of the pre-existing _HELP_OFFER members this PR changes no behaviour at all. Corpus counts stay at 4 nudged and 0 discarded.