fix(prompt): Restore SOUL voice authority in system prompt header

[dcramer]

Problem

dcramer noticed Junior's voice has been diminished — "the SOUL got diminished in my improvements too" / "Junior no longer sounds Gen z".

After tracing the prompt pipeline:

• SOUL.md is still mechanically loaded into the static system prompt and rendered inside <personality> tags. That part is fine.
• The regression is in the framing introduced by 22d60c0 (the prompt simplification).
• Pre-22d60c0 the prompt opened with: "You are a Slack-based helper assistant. Identity, tone, and domain defaults are defined in the personality block." and the personality block began with "Always follow the personality guidance for tone/style unless safety or policy constraints require otherwise."
• Post-22d60c0 the header became: "You are a Slack-based helper assistant. The behavior and output blocks below are authoritative; the personality block sets voice only." and the affirmative directive inside <personality> was dropped.

The new framing tells the model the personality block is decorative ("voice only") while elevating the other blocks as "authoritative". The behavior block is ~50 imperative bullets; the personality is a few lines from SOUL.md. With that hierarchy plus that volume disparity, the model reads the framing as permission to discount the SOUL, and the deployment-configured voice gets washed out.

Fix

Replace the dismissive HEADER with an affirmative one that:

1. Tells the model to follow the personality block for voice and tone in every reply (restores SOUL authority).
2. Keeps the spec contract intact: platform mechanics live in behavior/output and override personality only when those mechanics conflict.

const HEADER =
  "You are a Slack-based helper assistant. Follow the personality block for voice and tone in every reply. The behavior and output blocks define platform mechanics and override personality only when those mechanics conflict.";

This stays consistent with specs/agent-prompt-spec.md (SOUL is voice-only; platform behavior must work even with an empty SOUL) and keeps the system prompt byte-stable across turns.

Verification

• pnpm --filter @sentry/junior typecheck ✓
• pnpm --filter @sentry/junior exec vitest run tests/unit/prompt.test.ts ✓ (existing assertions still hold; HEADER text is not asserted)
• Verified via probe that <personality> continues to receive the deployment's SOUL.md content unchanged.

Notes

• No new evals here. There is no eval today that catches voice/tone regressions in the SOUL — that's a real gap, but a separate piece of work. Filing a follow-up to add a personality-voice eval would be the right next step if we want to prevent this kind of drift.
• Behavior change is limited to header wording; no runtime, capability, or context plumbing was touched.

Slack Thread

  

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
junior-docs	Ready	Preview, Comment	May 7, 2026 7:45pm