[BUG] "workflow" keyword in user message forces model to invoke Workflow tool even in purely conversational/investigative context

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When the word \"workflow\" appears anywhere in a user message, a <system-reminder> is injected into the conversation that instructs the model:

\"The user included the keyword 'workflow' or 'workflows', which means you should use the Workflow tool to fulfill their request.\"

This model-level instruction fires on any occurrence of the word, regardless of conversational intent. As a result, a message like:

\"Can you check the Temporal workflow for this job ID?\"

causes Claude to launch a multi-agent Workflow execution instead of simply running a temporal workflow show command. The user's intent was investigative/conversational; the model was forced into tool invocation.

This is distinct from the TUI highlight issue tracked in #63957 (macOS dismissal shortcuts broken). That issue is about the visual affordance. This issue is about the model being instructed to act — even when the user only meant the word in its plain technical sense (Temporal, Airflow, GitHub Actions, n8n, business process workflows, etc.).

Key difference from #63957: Even if the TUI highlight is dismissed, the <system-reminder> injection may still occur, causing the model to invoke the Workflow tool. The two surfaces (TUI hint vs. model instruction) appear to be coupled to the same keyword match but may have separate fix paths.

Reproduction Steps

1. In a Claude Code session, type a message that contains the word \"workflow\" in a non-workflow-invocation context, e.g.:

I want to investigate the Temporal workflow ao:12345

1. Send the message.
2. Observe that Claude interprets the message as a request to run a Workflow tool, rather than performing the stated investigative action.

What Should Happen?

The keyword detection should distinguish between:

• Workflow invocation intent: \"run a workflow\", \"use a workflow\", \"start a workflow\", \"create a workflow\" — explicit action verbs paired with the noun
• Contextual reference: \"check the workflow\", \"look at this workflow\", \"the Temporal/Airflow/GitHub Actions workflow\" — referring to an external artifact, not requesting Claude to run one

At minimum, the model-level instruction should not use normative language (\"should use the Workflow tool\") when the keyword appears in an ambiguous context. It should instead surface a clarification or require more explicit opt-in phrasing, consistent with how slash commands work.

Related

• #63957 — macOS per-prompt TUI dismissals (alt+w, backspace) don't work; overlapping population of affected users (Temporal / Airflow / GitHub Actions users)
• #63784 — Settings docs missing workflowKeywordTriggerEnabled toggle

Environment

• Platform: macOS (Darwin 25.3.0)
• Claude Code CLI (latest)