[BUG] System reminder causes Claude to create a workflow for a one-line bugfix.

Additional repro (v2.1.219, Windows desktop / local agent mode): the keyword triggers even when it appears only inside quoted text — e.g. pasting a report that discusses this injection itself. In one session the reminder ("The user included the keyword "ultracode", opting this turn into multi-agent orchestration — use the Workflow tool to fulfill the request.") fired three times in a row, each on a message that merely quoted or discussed the word — including one quoting the injected reminder text verbatim.

Note the v2.1.210 origin gate (keyword only activates from direct human-originated input, cf. #77633) does not cover this vector: a human pasting a report or quoting a prior session IS human-originated input, but it is a mention, not a request. Detection appears to be a plain substring match (workflow_keyword_request reminder), so use vs. mention cannot be distinguished.

Suggestion: treat the keyword as opt-in only when it is part of the user's own imperative request (or at minimum exclude quoted/code-fenced content), rather than any occurrence anywhere in the prompt text.

Still reproduces on 2.1.220 (Desktop app, Windows 11 Pro, model claude-fable-5-1, session effort xhigh, Auto permission mode). Two new data points and a measured cost.

Repro 1 - the word as a noun in a list. The prompt (in Chinese) asked to edit a project skill so that its per-card recommendation lists effort levels: "high, extra, max and ultracode". The word appears once, as the name of a level, in a request to change ~40 lines of a markdown file. The harness injected The user included the keyword "ultracode", opting this turn into multi-agent orchestration - use the Workflow tool to fulfill the request. and loaded the workflow-authoring skill.

Repro 2 - quoted text. Later the user replied with a > quote of the assistant's own sentence that explained the trigger (the sentence contained the word), followed by a one-line reaction. The same system-reminder fired again on that turn. Quoted/attached text is never a request.

Impact. The model followed the reminder and ran four dynamic workflows over the turn (draft panel, judges, two verification rounds; 22 subagent transcripts). Weighting cache reads at 0.1x, cache writes at 2x and output at 5x an input token, the transcripts total about 27.4M weighted tokens, 19.2M of them in workflow subagents; the edit itself needs well under 1M. No per-run approval prompt appeared - Auto mode records consent on the first launch and later launches start silently - so nothing stood between the false trigger and the spend. The user first noticed when told why it happened.

Expected. A bare mention should not opt the turn in - only a directive (ultracode: <task>, "use a workflow"). Concretely:

1. Do not fire on quoted or attached text.
2. Require the keyword as a standalone directive (e.g. leading the prompt, as the docs example shows), not anywhere in the text.
3. The reminder is worded as an instruction ("use the Workflow tool to fulfill the request"), so the model complies even for a trivial edit. Wording it as an opt-in the model may decline for small tasks would have stopped this case.

Same class as #63725 for the old workflow keyword.

Apparently there's a setting to opt-out of this behavior: https://code.claude.com/docs/en/settings-reference#workflowkeywordtriggerenabled

{
  "workflowKeywordTriggerEnabled": false
}