ADR 0018 — Output truncation detection and host recovery policies

Status: Accepted (detect + report/halt + surfaces; continue deferred)

Context

When a provider stops for length / max tokens, partial text was deposited as if complete. Gates and parse: list saw silently cut-off content.

Decision

1. Detect always. Produced.truncated + finish_reason (provider-normalized).
2. Trace always. Step records truncated: true (+ finish_reason when known). Fan-out parent steps get the same when any branch truncated.
3. Live events. state-done events carry truncated / finish_reason.
4. Recovery host policy, default report:
5. report — annotate and continue.
6. halt — state-error: output-truncated.
7. continue — deferred (stitch N pieces under cost budget; never default).
8. Surfaces. engine.run(..., on_truncate=…), CLI --on-truncate, MCP on_truncate, console ConsoleTools.on_truncate, scripttest run: { on_truncate: halt }.
9. parse: list. Truncated produce → state-error: parse-list-truncated (clearer than a generic JSON parse failure). Under halt, output-truncated wins first.

Checklist

Item	Status
Detect + trace + events	done
report / halt policies	done
CLI / MCP / console / scripttest parity	done
Unit tests (mock produce)	done
Adapter unit tests with fake responses	done (Anthropic + OpenAI-compat)
continue stitching	deferred

Consequences

• Observability improves with zero .mk changes.
• Auto-continue stays out so token budgets and traces stay honest.