feat(compaction): make conversation compaction configurable per model preset #338

Merged

rwachters merged 4 commits from feat/model-preset-compaction-config into master

2026-10-08 18:41:45 +00:00

rwachters commented

2026-10-08 18:17:22 +00:00

(Migrated from github.com)

Summary

Makes conversation compaction configurable per model preset. A preset gains a compaction_enabled flag and an optional compaction_threshold_tokens override, and the effective configuration is resolved and validated once per turn, inside the turn-preparation transaction, instead of being read mid-turn from the user preference alone. A preset can only restrain compaction (its flag is ANDed with the user preference) and override the threshold; the auxiliary summarization model, settings, instruction and summary label stay user-preference-owned.

Changes

Added the additive V35__model_preset_compaction_config.sql migration plus the table, entity, mapper, DAO and service plumbing for the two new preset columns, defaulting existing rows to enabled with no override so current behaviour is preserved.
Added compactionEnabled (default true) and compactionThresholdTokens (default null) to ModelPresetDto, the create/update preset requests and the built-in preset tool schemas, and mapped create/update failures to typed errors.
Added the app-side preset form controls (a switch and an optional threshold where blank means "use the user preference threshold"), the detail-panel/list-item rendering of the resulting configuration, and their tests.
Added ResolvedCompactionConfig / EffectiveCompactionSettings and DefaultEffectiveCompactionConfigResolver, which resolve and validate the preset and the stored preference inside the turn's transaction; the value is carried through PreparedConversationTurn → ConversationTurnRequest → ConversationCompactionService.beginTurn.
Effective enablement is preset.compactionEnabled AND a present preference row AND preference.enabled; a preset with compaction disabled short-circuits before the preference is read or decoded. Effective threshold is the preset override when set, otherwise the preference threshold.

Behaviour changes

B1: A stored non-positive preference thresholdTokens now fails the turn at validation time instead of only when compaction ran, and it can no longer be masked by a preset threshold override.
B2: A compaction-enabled preset combined with a preference whose auxiliary modelId/settingsId is null now fails every turn of that session instead of running until the window exceeded the threshold.
B3: Non-positive ids and a blank instruction now fail up front rather than mid-turn; this can only affect legacy or hand-edited preference rows.
B4: A malformed stored preference is reported at validation time as MODEL_CONFIGURATION_ERROR with nothing persisted, instead of a mid-turn CompactionFailed after the user message had already been saved.

API contract change

PUT /api/v1/me/preferences/conversation_compaction now rejects null auxiliary ids with 400 invalid_argument while enabled == true; null ids remain valid while enabled == false.
Because enabled defaults to true, a PUT that omits enabled and sends null ids is now rejected.

Implementation notes

The configuration is intentionally snapshotted once per turn: chunks already persisted keep the threshold that was in effect when they were created.
Only the preset's enablement flag and threshold cross into the turn; the auxiliary summarizer configuration remains a user preference, so there is a single place to change the summarization model.
ConversationCompactionService.beginTurn no longer returns an Either: the configuration was already resolved and validated during preparation, so the remaining failure modes surface from preparePrimaryContext only when compaction is actually needed.

Validation

./gradlew :server:test — focused filters for the new resolver, the compaction configuration service and resolver, turn preparation, the orchestrator suites, the preset routes/DAO/service/tools and the migrator test were run green.
./gradlew :app:desktopTest — preset form state, form dialog and API client suites were run green.
./gradlew :common:desktopTest — preset request serialization and built-in tool catalog suites were run green.

Risks and follow-up

The preference PUT is the only intentionally breaking part of this change; clients that stored null auxiliary ids with enabled omitted or true must now set enabled: false or provide ids.
The two new columns are additive with defaults, so pre-existing rows need no data update; downgrade is not supported.

## Summary Makes conversation compaction configurable per model preset. A preset gains a `compaction_enabled` flag and an optional `compaction_threshold_tokens` override, and the effective configuration is resolved and validated once per turn, inside the turn-preparation transaction, instead of being read mid-turn from the user preference alone. A preset can only restrain compaction (its flag is ANDed with the user preference) and override the threshold; the auxiliary summarization model, settings, instruction and summary label stay user-preference-owned. ## Changes - Added the additive `V35__model_preset_compaction_config.sql` migration plus the table, entity, mapper, DAO and service plumbing for the two new preset columns, defaulting existing rows to enabled with no override so current behaviour is preserved. - Added `compactionEnabled` (default `true`) and `compactionThresholdTokens` (default `null`) to `ModelPresetDto`, the create/update preset requests and the built-in preset tool schemas, and mapped create/update failures to typed errors. - Added the app-side preset form controls (a switch and an optional threshold where blank means "use the user preference threshold"), the detail-panel/list-item rendering of the resulting configuration, and their tests. - Added `ResolvedCompactionConfig` / `EffectiveCompactionSettings` and `DefaultEffectiveCompactionConfigResolver`, which resolve and validate the preset and the stored preference inside the turn's transaction; the value is carried through `PreparedConversationTurn` → `ConversationTurnRequest` → `ConversationCompactionService.beginTurn`. - Effective enablement is `preset.compactionEnabled` AND a present preference row AND `preference.enabled`; a preset with compaction disabled short-circuits before the preference is read or decoded. Effective threshold is the preset override when set, otherwise the preference threshold. ## Behaviour changes - B1: A stored non-positive preference `thresholdTokens` now fails the turn at validation time instead of only when compaction ran, and it can no longer be masked by a preset threshold override. - B2: A compaction-enabled preset combined with a preference whose auxiliary `modelId`/`settingsId` is null now fails every turn of that session instead of running until the window exceeded the threshold. - B3: Non-positive ids and a blank instruction now fail up front rather than mid-turn; this can only affect legacy or hand-edited preference rows. - B4: A malformed stored preference is reported at validation time as `MODEL_CONFIGURATION_ERROR` with nothing persisted, instead of a mid-turn `CompactionFailed` after the user message had already been saved. ## API contract change - `PUT /api/v1/me/preferences/conversation_compaction` now rejects null auxiliary ids with `400 invalid_argument` while `enabled == true`; null ids remain valid while `enabled == false`. - Because `enabled` defaults to `true`, a `PUT` that omits `enabled` and sends null ids is now rejected. ## Implementation notes - The configuration is intentionally snapshotted once per turn: chunks already persisted keep the threshold that was in effect when they were created. - Only the preset's enablement flag and threshold cross into the turn; the auxiliary summarizer configuration remains a user preference, so there is a single place to change the summarization model. - `ConversationCompactionService.beginTurn` no longer returns an `Either`: the configuration was already resolved and validated during preparation, so the remaining failure modes surface from `preparePrimaryContext` only when compaction is actually needed. ## Validation - `./gradlew :server:test` — focused filters for the new resolver, the compaction configuration service and resolver, turn preparation, the orchestrator suites, the preset routes/DAO/service/tools and the migrator test were run green. - `./gradlew :app:desktopTest` — preset form state, form dialog and API client suites were run green. - `./gradlew :common:desktopTest` — preset request serialization and built-in tool catalog suites were run green. ## Risks and follow-up - The preference `PUT` is the only intentionally breaking part of this change; clients that stored null auxiliary ids with `enabled` omitted or `true` must now set `enabled: false` or provide ids. - The two new columns are additive with defaults, so pre-existing rows need no data update; downgrade is not supported.