4.6 KiB
ADR-0022: Use native continuation and application assessment
Status: Accepted Date: 2026-11-22
Context
Private retry machinery conflated transient failures, structural repair, semantic defects, and validator availability. Latest-response replacement also discarded conversation history while requiring an application replay protocol. Native execution can own those mechanics without owning domain acceptance.
Decision
One small typed producer helper connects native execution to module functions for fixed input preparation, interpretation, correction requests, and supported fallback construction. The pipeline supplies the operation's typed assessor. PromptKit owns transient retries, structural repair, growing conversation history, and semantic turn iteration. Notarius selects the semantic-turn allowance but does not pin omitted native repair/retry counts or extend exhausted native work with an outer retry loop. Deterministic operations have no generic retry budget.
Each correctable producer has initial and correction prompts using the same response schema and bundled default profile. Preserve explicit request selections and run overrides. Enduring principles and source/reference material are supplied initially; the correction template adds current findings and a new assignment. Native coordination carries the exchange, without manual assistant replay or rerendering original inputs. Each response is a complete replacement, not a patch. Normalization applies each proposal afresh to fixed originals and candidate handles; never union groups across turns.
Check native errors and structural validation status before interpretation, including failed validation returned with nil Go error. A usable typed candidate receives the complete applicable validator chain once, in configured order. Earlier rejection does not suppress later guidance. Revisions, unchanged candidates, and the final permitted response receive the whole chain; there is no duplicate assessment after the helper returns.
The detecting validator or module authors actionable guidance separately from reason codes and operator messages. Unavailable assessment is not a semantic defect; provider failures do not become instructions. LLM validators execute once per assessment, without their own semantic continuation loop. Native success, stopped, and exhausted statuses do not imply application acceptance.
Strict is the application default. Explicit permissive mode may construct a module-supported fallback after terminal nonacceptance. A fallback satisfies intrinsic artifact and downstream invariants without ordinary semantic approval or rerunning the normal validator chain. Failure to construct it fails normally. Caller cancellation and undelivered responses never initiate fallback. A result-less failed continuation cannot recover an earlier candidate. Reuse exclusions and disclosure follow ADRs 0023 and 0024. Exact configuration precedence and defaults belong in Configuration.
This supersedes ADR-0014 and ADR-0016's budgets, private candidate transport, latest-only replay, and terminal policies. It retains their useful separation of actionable guidance from diagnostic prose and of execution failure from detected defects. Diagnostic persistence is governed by ADR-0024. Contextual identity and request-local reconciliation safety under ADRs 0012 and 0013 remain in force.
Alternatives considered
- Retain blanket producer/validator retries. These can extend native budgets and repeat deterministic work without information capable of fixing it.
- Keep latest-only manual replay. This duplicates conversation mechanics and discards earlier application-visible turns.
- Let validators decide terminal disposition or revalidate every fallback. That distributes acceptance policy and confuses qualified recovery with ordinary validation.
Consequences
Native execution and application acceptance are separate facts. Accumulated history uses more context, bounded by the selected allowance. Not every failure has a supported fallback: an empty scene description, for example, cannot meet its required artifact shape. Permissive output is explicitly qualified and reuse-ineligible rather than silently accepted. Paid-model quality and exact provider-attempt parity with the retired loops are not promised.