Response-style profiles¶

Response-style profiles give per-role control over worker output verbosity. A reviewer role can run terse while an investigator role runs verbose, without hand-editing role prompt templates. The style is resolved deterministically at spawn time and rendered into the worker's system prompt as an addendum.

Source:

src/bernstein/core/agents/response_style.py - resolution + rendering
src/bernstein/core/agents/spawner_core.py - spawn-time wiring
src/bernstein/core/config/config_schema.py - response_style field

The three styles¶

Style	Maps to mode profile	Effect
`terse`	`fast`	Compresses prose; shortest worker output.
`balanced`	`smart`	Neutral house style (the default).
`verbose`	`deep`	Fuller narration and reasoning in prose.

balanced is the neutral default and renders an empty addendum, so a spawn with no declared profile is byte-identical to a pre-feature spawn. Only terse and verbose inject an addendum.

The addendum body is the mode-profile preamble from templates/mode_profiles/{fast,smart,deep}.yaml. No new prompt dialect is introduced; the vocabulary reuses the existing conciseness scale.

Style carve-outs¶

The style applies to prose only. Every non-empty addendum carries a fixed carve-out block instructing the worker never to compress, truncate, or restyle code blocks, commit messages, error text, or completion-contract JSON payloads. Those are always reproduced in full.

Configuring it¶

Declare response_style per role under role_model_policy in your seed config (bernstein.yaml). The special default entry sets the fallback for roles that do not declare their own:

role_model_policy:
  default:
    response_style: balanced   # seed-level fallback
  reviewer:
    cli: claude
    model: sonnet
    response_style: terse      # reviewers run terse
  investigator:
    cli: claude
    model: opus
    response_style: verbose    # investigators run verbose

The value is validated against the closed vocabulary (verbose / balanced / terse) at parse time. A style whose mapped template file is missing raises a typed ResponseStyleTemplateError during config validation - not at spawn time - so a dangling template reference fails fast.

Resolution order¶

The style for a given spawn is resolved in this fixed order; the first source that supplies a valid style wins:

Task.metadata['mode'] - accepts a style name (verbose / balanced / terse) or the equivalent mode-profile name (fast / smart / deep).
The role's own role_model_policy.<role>.response_style.
The seed default role_model_policy.default.response_style.
The built-in default, balanced.

Unknown values fall through to the next source rather than raising, so a stale Task.metadata['mode'] value cannot block a spawn. Config-level typos are rejected earlier by the seed parser.

The same inputs always resolve to the same style and the same rendered addendum, so the resolution is snapshot-testable and reproducible across machines.

Identity, ledger, and audit¶

When a profile is explicitly declared (anything other than the built-in default), the SHA-256 of the rendered addendum is:

folded into the task identity used by the deterministic schedule projection, and
stamped onto the task's cost-ledger entry (response_profile plus profile_content_sha256) and audit trail.

That stamping is what makes per-profile cost attribution possible - see Per-profile cost attribution. A spawn that falls through to the built-in balanced default is not folded into identity, keeping unset-profile runs byte-identical to before.