ADR 0017 — Context & content management: host budgets first, language zones later

Status: Accepted (Layer 0–1; Layer 2 language faces deferred)

Context

The blackboard accumulates without bound. Historically the judge CONTEXT blob was capped at JUDGE_CONTEXT_CHARS with a silent prefix slice (fixed: head+tail marker). Formal dual-channel zones remain a non-goal for the current language line (SPEC §11).

Decision

Stratify; implement bottom-up:

1. Layer 0 — host rendering budgets. Judge CONTEXT: head+tail with explicit …[context_truncated]… marker (never silent prefix-only). Produce-prompt char budgets and console history windowing are not claimed done yet.
2. Layer 1 — machine patterns. Explicit compress states / examples.
3. Layer 2 — deferred language faces. context_policy, pin, trusted zones.

Checklist

Item	Status
Judge CONTEXT head+tail marker	done (llm/context_view.py)
SPEC §5 wording	done
Produce-prompt char budget	done (interpolate.PROMPT_VALUE_CHARS, run(prompt_value_chars=…))
Console history window	done (history_for_brain, prompt-only; full audit kept)
Compress pattern example	done (examples/research_compress.mk + scenario test)
std_compress / language faces	deferred

Consequences

• Existing .mk files keep working.
• Authors still own semantic summarization; the host only makes judge truncation visible.