[FEATURE] Visible confidence/risk indicator before risky or uncertain actions

Summary

Claude Code's output reads with uniform confidence regardless of how certain the underlying reasoning actually is — there's no built-in signal distinguishing "I'm confident about this" from "I'm guessing/extrapolating into unfamiliar territory," especially before a risky or hard-to-reverse action. Please add a visible confidence/risk indicator for such moments.

Current behavior

Whether Claude is working from well-established, verified context or extrapolating into an area with thin/ambiguous information, the tone and presentation of its output looks the same. There's no in-product signal that says "this next action is based on lower certainty than usual" before something risky proceeds.

Motivation

This is a widely reported UX gap across the AI-agent market, not specific to one tool: a Medium piece on agent UX anti-patterns ("10 Agent UX Mistakes Users Never Forgive") specifically calls out uniform confident-sounding output as a trust-destroying pattern, and a review of Devin notes it "fails silently" — producing plausible-but-wrong output in unfamiliar domains rather than flagging the uncertainty. Separately, an existing Claude Code bug report (#60226) documents Claude explicitly stating its own analysis is unfounded, then completing and presenting that analysis anyway in the same response — the self-identified uncertainty exists internally but isn't surfaced or gated on.

Proposed behavior

1. Before a risky/hard-to-reverse action (destructive file ops, force-pushes, schema migrations, etc.), surface a lightweight confidence signal alongside the action — not necessarily a numeric score, but a distinguishable "low confidence" flag when the model's own reasoning indicates uncertainty, missing context, or extrapolation beyond verified information.
2. When Claude has explicitly reasoned that something is uncertain or unfounded, that should visibly gate the action (at minimum, a flagged callout), not silently proceed as if it were confident — closing the exact gap #60226 documents.
3. Should be low-noise — reserved for genuinely risky/high-stakes moments, not applied to routine, low-risk responses.

Why it matters

• Addresses a documented, recurring failure mode (self-identified uncertainty not gating output) with a concrete existing repro (#60226) rather than a hypothetical.
• Matches a broader market expectation that agent tools should signal calibrated confidence, not uniform confidence, especially as agent autonomy increases.

Related issues (adjacent, not duplicates)

• #60226 Claude states the reason its current analysis is unfounded, then completes the analysis in the same response — self-identified blocking gaps do not gate output — the concrete symptom this feature would address; framed there as a bug, here as a proposed UI/gating mechanism.
• #10084 [FEATURE] Expose Claude Code Cognitive Telemetry States via API (closed) — adjacent (internal state exposure), different surface (API telemetry vs. a user-facing confidence indicator before risky actions).

Notes

• Reviewed for duplicates before filing (multiple search passes: "confidence risk indicator before risky action uncertain", "low confidence warning before executing command", "agent confidence score uncertain output display") — nothing found proposing a user-facing confidence/risk indicator specifically, as of 2026-07-19.
• Sourced from market research (Medium "10 Agent UX Mistakes Users Never Forgive"; AIToolRanked "Devin AI Review 2026") plus a concrete existing Claude Code bug (#60226) as supporting evidence the same gap exists here.