ds4-context-engine 0.3.0-beta.3 → 0.3.1

package/docs/COMPACTION.md

@@ -57,6 +57,27 @@ When privacy is enabled, local summary generation may consume allowed local-only
 
 Pi 0.84.3 locates the post-compaction entry by summary text, which can surface an older entry when deterministic test summaries are identical. DS4 therefore correlates commit with its pending summary ID and resolves the matching newly appended entry from `SessionManager`, never by text equality.
 
+## Dedicated compaction model and thinking
+
+By default the active session model generates summaries, exactly as in previous releases. A dedicated model can be opted into:
+
+```json
+{
+  "compaction": {
+    "model": { "provider": "anthropic", "id": "claude-sonnet-4-5" },
+    "summary": { "thinking": "medium" }
+  }
+}
+```
+
+Semantics:
+
+- when `compaction.model` is absent, the active session model is used for budget, segmentation, requests, and record provenance;
+- when present, the model is resolved once per compaction through the model registry and used uniformly for input budget, segmentation, sanitization, segment/aggregate requests, and `provider`/`model` provenance in summary records;
+- the dedicated model must exist, be configured with auth, and accept text input; otherwise DS4 logs a warning and falls back to the session model — compaction is never blocked by configuration;
+- `compaction.summary.thinking` defaults to `off` and applies only to summary requests: `off` keeps the pre-existing request shape (no thinking fields), while other levels map per API (`thinkingEnabled`/`effort` for `anthropic-messages`, `samplingParams.reasoning_effort` for OpenAI-compatible APIs) and are ignored for unsupported providers;
+- `context.maxSummaryTokens` remains a session-level limit and does not rise for the dedicated model; the minimum with the model's `maxTokens` still applies.
+
 ## Proactive trigger
 
 After a settled turn, DS4 computes:

package/docs/releases/0.3.0-beta.3.md

@@ -0,0 +1,83 @@
+# DS4 Context Engine 0.3.0-beta.3
+
+Status: published on 2026-09-03 under the npm `beta` dist-tag.
+
+This coordinated beta fixes storage maintenance on Windows and carries forward beta.2 without changing canonical records, SQLite schema, runtime contracts, retention limits, privacy policy, compaction validation, or fallback semantics.
+
+## Fixed since beta.2
+
+- `compactStorage()` now opens stage/state/source files for the durability flush with write access (`"r+"`). `fsyncPath` previously opened them read-only, which fails on Windows with `EPERM` (`FlushFileBuffers` requires write access), aborting `compact` with `stage=compact; category=maintenance-failed` before the backup could be verified. The failure was confirmed on Windows with a minimal probe:
+
+  ```text
+  fsync-on-readonly-handle: FAIL code=EPERM
+  ```
+
+- The flush now mirrors `fsyncDirectory`'s best-effort policy for platforms that cannot flush (`EINVAL`, `ENOTSUP`, `EBADF`, `EPERM`, `EACCES`), while still flushing normally on supported platforms. No content is written by the `"r+"` open.
+
+- Coordinated package versions moved to `0.3.0-beta.3`.
+
+There are no functional storage-format or provider-path changes relative to beta.2.
+
+## Safety and compatibility
+
+Unchanged from beta.2: Pi JSONL remains canonical and append-only; SQLite remains disposable and rebuildable; physical maintenance remains explicit, offline, local-TTY-only, and unavailable to model-callable tools; every `context_persistence` write still requires a fresh positive local UI decision.
+
+Compatibility remains unchanged:
+
+- SQLite schema: `15`;
+- configuration: `ds4-context-config-v1`;
+- runtime adapter: `runtime-adapter-v1`;
+- persistence tool: `ds4-context-persistence-tool-v1`;
+- persistence result: `ds4-context-persistence-result-v1`;
+- Pi: `0.84.3`;
+- Node.js: `>=22.19.0`.
+
+## Package/version policy
+
+The coordinated version is `0.3.0-beta.3` for:
+
+```text
+ds4-context-core
+ds4-context-reference-adapter
+ds4-context-engine
+```
+
+Both adapters depend exactly on `ds4-context-core@0.3.0-beta.3`. Publication used the explicit npm `beta` dist-tag for all three packages. npm `latest` remains stable `0.2.0`, and GitHub Actions remains validation-only with OIDC and package-write permissions denied.
+
+## Candidate validation
+
+Local candidate verification on Node.js `26.5.1`:
+
+- `npm run check`: 69 files and 341 tests passed.
+- `npm run quality:compare`: candidate quality `0.9875` versus baseline `0.808156`.
+- `npm run schema:context-persistence`: 1,266 bytes and 317 estimated tokens; below the 1,500 absolute and 320 relative limits.
+- `npm run pack:check`: verified core (223 files), reference adapter (7 files), and Pi adapter (68 files) in a clean consumer.
+- `npm pack --dry-run --json` for all three packages passed with the same inventories and no forbidden local/session/storage files.
+- `git diff --check` passed; the only pre-existing untracked path is `.serena/`, which is excluded from commits and package inventories.
+- Version, exact core dependencies, package-lock entries, extension constant, and reference-adapter constant are synchronized to `0.3.0-beta.3`.
+
+Validation-only CI run [`33771957691`](https://github.com/Alucard24/ds4-context-engine/actions/runs/33771957691) passed on Node.js `22.19.0` and `24.x` for commit `8c4abf6106b675aa1af1ee163f6ca07b5e1c0ef3`.
+
+## Registry evidence
+
+Exact post-publication verification passed for all three `0.3.0-beta.3` packages. The clean registry consumer verified exact adapter/core dependencies, public core exports, compiled reference-adapter conformance, packaged quality corpus, packaged Pi extension startup, and the installed `ds4-context-storage` CLI declaration, shim, and usage probe.
+
+Published package shasums:
+
+```text
+ds4-context-core:              0428629e03cf2d02a7033fa3e016344dff3af399
+ds4-context-reference-adapter: 53924c98d6c9672709076edb5cf70bf3f98c1915
+ds4-context-engine:            d0ae468daae4f4e4970497282fd5c18b86f847d5
+```
+
+For all three packages, npm `beta` resolves to `0.3.0-beta.3`, `alpha` remains `0.3.0-alpha.5`, `rc` remains `0.2.0-rc.1`, and `latest` remains `0.2.0`.
+
+Annotated tag `v0.3.0-beta.3` targets `8c4abf6106b675aa1af1ee163f6ca07b5e1c0ef3`. The GitHub prerelease is published at <https://github.com/Alucard24/ds4-context-engine/releases/tag/v0.3.0-beta.3>.
+
+## Documentation
+
+- [`../STORAGE_MAINTENANCE.md`](../STORAGE_MAINTENANCE.md)
+- [`../STORAGE.md`](../STORAGE.md)
+- [`../COMPACTION.md`](../COMPACTION.md)
+- [`../RELEASING.md`](../RELEASING.md)
+- [`0.3.0-beta.2.md`](0.3.0-beta.2.md)

package/docs/releases/0.3.0.md

@@ -0,0 +1,82 @@
+# DS4 Context Engine 0.3.0
+
+Status: promoted 2026-09-04 from beta.3; printed under the npm `latest` dist-tag.
+
+This is the first stable release of the 0.3 line. It carries exactly the beta.3 candidate: no canonical record format, SQLite schema, runtime contract, retention limit, privacy policy, compaction validation, or fallback semantics changed during promotion.
+
+## Promoted from
+
+- [`0.3.0-beta.3.md`](0.3.0-beta.3.md) is the exact promoted candidate; beta.3 itself fixed Windows storage-maintenance flush handling on top of beta.2.
+
+## Promotion evidence
+
+The promotion criteria in [`../DOGFOODING_0.3.0_BETA.md`](../DOGFOODING_0.3.0_BETA.md) are satisfied as follows:
+
+- Full mode matrix with synthetic data on 2026-09-03: TUI, RPC (with and without responding UI bridge), print, JSON stream, no-session, canonical append audit, branch switch, restart, projection rebuild, two concurrent client leases, untrusted RPC, isolated offline maintenance with local TTY confirmation, and online retention convergence. Evidence: `/tmp/ds4-beta-dogfood/evidence/`.
+- Configured remote provider (real): the RPC/branch/untrusted matrix ran against `openai-codex` / `gpt-5.4-mini` with real authentication on 2026-09-03; the disposable database records 21 `openai-codex` / `gpt-5.4-mini` context manifests with provider-reported usage. The same matrix covered accepted, cancelled, and provider-policy-denied persistence writes.
+- Additional real-provider TUI soak on 2026-09-04 against `deepseek` / `deepseek-v4-flash-vision-exp` (Pi 0.85.0): normal model call with provider-reported usage, `/context status`, `/context storage`, and a session pin persisted with scope `session`, status `active`, and a canonical source entry. Evidence: `/tmp/ds4-real-dogfood/evidence/` and the disposable database at `/tmp/ds4-real-agent/ds4-context/context.db` (3 real manifests, pin `b5828a50-689b-4c59-8a90-241eca30217c`).
+- Verified local provider: explicitly recorded as unavailable — no local provider (Ollama/LM Studio or equivalent) is configured in the environment; the runbook allows the missing run to remain recorded rather than inferred.
+- Live maintenance authorization: completed 2026-09-03 (2.42 GiB → 377 MiB, `quick_check=ok`, FK `0`, manifests 1,630 → 128); the fixed backup `context.db.precompact.bak` (including `-shm`/`-wal` residues) was removed 2026-09-04 after a post-removal read-only inspection confirmed `quick_check=ok`, FK `0`, 128 manifests, and `Maintenance recommended: no`.
+
+## Safety and compatibility
+
+Unchanged from beta.3: Pi JSONL remains canonical and append-only; SQLite remains disposable and rebuildable; physical maintenance remains explicit, offline, local-TTY-only, and unavailable to model-callable tools; every `context_persistence` write still requires a fresh positive local UI decision.
+
+Compatibility surface:
+
+- SQLite schema: `15`;
+- configuration: `ds4-context-config-v1`;
+- runtime adapter: `runtime-adapter-v1`;
+- persistence tool: `ds4-context-persistence-tool-v1`;
+- persistence result: `ds4-context-persistence-result-v1`;
+- Pi: `0.84.3` (validated; 0.85.0 additionally exercised by the 2026-09-04 TUI soak);
+- Node.js: `>=22.19.0`.
+
+## Package/version policy
+
+The coordinated version is `0.3.0` for:
+
+```text
+ds4-context-core
+ds4-context-reference-adapter
+ds4-context-engine
+```
+
+Both adapters depend exactly on `ds4-context-core@0.3.0`. Publication used npm's default `latest` tag for all three packages: `latest` now resolves to `0.3.0` (previously `0.2.0`), `beta` remains `0.3.0-beta.3`, `alpha` remains `0.3.0-alpha.5`, and `rc` remains `0.2.0-rc.1`. GitHub Actions remains validation-only with OIDC and package-write permissions denied.
+
+## Candidate validation
+
+Local candidate verification on Node.js `26.5.1`:
+
+- `npm run check`: 69 files and 341 tests passed on the promoted tree.
+- `npm run quality:compare`: candidate quality versus the frozen baseline passed.
+- `npm run schema:context-persistence`: within the 1,500 / 320 absolute and relative limits.
+- `npm run pack:check` and `npm pack --dry-run --json` for all three packages passed with no forbidden local/session/storage paths.
+- `git diff --check` passed; the only pre-existing untracked path is `.serena/`, which is excluded from commits and package inventories.
+- Version, exact core dependencies, package-lock entries, extension constant, and reference-adapter constant are synchronized to `0.3.0`.
+
+Exact registry verification and the annotated tag are recorded below.
+
+## Registry evidence
+
+Exact post-publication verification passed for all three `0.3.0` packages on 2026-09-04. The clean registry consumer verified exact adapter/core dependencies, public core exports, compiled reference-adapter conformance, packaged quality corpus, packaged Pi extension startup, and the installed `ds4-context-storage` CLI declaration, shim, and usage probe.
+
+Published package shasums:
+
+```text
+ds4-context-core:              7e3c6407996f263db3f6159e5de8b30de6dfabe8
+ds4-context-reference-adapter: ce133b184b6dbb7ed2a7337d639dca13337d7df1
+ds4-context-engine:            6d265de2a8e6c904482ab0ff89acbadbde862854
+```
+
+For all three packages, npm `latest` now resolves to `0.3.0`, `beta` remains `0.3.0-beta.3`, `alpha` remains `0.3.0-alpha.5`, and `rc` remains `0.2.0-rc.1`.
+
+Annotated tag `v0.3.0` targets `608a6238325ed8627dc16c0c32ad10697977e0e0`. The GitHub release is published at <https://github.com/Alucard24/ds4-context-engine/releases/tag/v0.3.0>.
+
+## Documentation
+
+- [`../STORAGE_MAINTENANCE.md`](../STORAGE_MAINTENANCE.md)
+- [`../STORAGE.md`](../STORAGE.md)
+- [`../COMPACTION.md`](../COMPACTION.md)
+- [`../RELEASING.md`](../RELEASING.md)
+- [`0.3.0-beta.3.md`](0.3.0-beta.3.md)

package/docs/releases/0.3.1.md

@@ -0,0 +1,72 @@
+# DS4 Context Engine 0.3.1
+
+Status: published 2026-09-04 as stable, printed under the npm `latest` dist-tag.
+
+This release adds the opt-in dedicated compaction model and per-request reasoning control on top of the 0.3.0 stable candidate. It carries forward 0.3.0 without changing canonical records, SQLite schema, runtime contracts, retention limits, privacy policy, compaction validation, or fallback semantics.
+
+## Added since 0.3.0
+
+- **Dedicated compaction model** (`compaction.model.provider` / `compaction.model.id`, opt-in):
+  - when absent, compaction keeps using the active session model exactly as before, including preflight budget, segmentation, sanitization, and record provenance;
+  - when present, the model is resolved once per compaction through Pi's model registry and used uniformly for the input budget, segmentation, sanitization, segment/aggregate requests, and the `provider`/`model` provenance of summary records;
+  - automatic fallback with a `compaction.dedicated_model_fallback` warning when the model is not registered, has no configured authentication, or does not accept text input: a configuration mistake can never block compaction, and the hook still returns the session-model path;
+  - the resolved model is recorded in diagnostics as `compaction.dedicated_model` at `debug`.
+- **Summary reasoning control** (`compaction.summary.thinking`, default `off`):
+  - applies only to summary requests, never to normal session traffic;
+  - `off` (default) leaves the request shape unchanged from 0.3.0;
+  - levels `minimal`/`low`/`medium`/`high`/`xhigh`/`max` map per API: `thinkingEnabled`/`effort` for `anthropic-messages`, `samplingParams.reasoning_effort` for OpenAI-compatible APIs (`openai-responses`, `openai-completions`, `azure-openai-responses`, `openai-codex-responses`);
+  - unsupported APIs ignore the option without errors; `xhigh`/`max` clamp to the highest supported effort where the API has no equivalent.
+- Configuration validation in the loader recognizes the two opt-in object paths and rejects malformed provider/id tokens or unknown thinking levels, generating a `configuration ignored` warning and leaving the default config in force.
+- Coordinated package versions moved to `0.3.1`.
+
+There are no functional storage-format or provider-path changes relative to 0.3.0.
+
+## Safety and compatibility
+
+Unchanged from 0.3.0: Pi JSONL remains canonical and append-only; SQLite remains disposable and rebuildable; physical maintenance remains explicit, offline, local-TTY-only, and unavailable to model-callable tools; every `context_persistence` write still requires a fresh positive local UI decision.
+
+Compatibility remains unchanged:
+
+- SQLite schema: `15`;
+- configuration: `ds4-context-config-v1` (two opt-in object keys added; default shape and all existing keys unchanged);
+- runtime adapter: `runtime-adapter-v1`;
+- persistence tool: `ds4-context-persistence-tool-v1`;
+- persistence result: `ds4-context-persistence-result-v1`;
+- Pi: `0.84.3`;
+- Node.js: `>=22.19.0`.
+
+## Package/version policy
+
+The coordinated version is `0.3.1` for:
+
+```text
+ds4-context-core
+ds4-context-reference-adapter
+ds4-context-engine
+```
+
+Both adapters depend exactly on `ds4-context-core@0.3.1`. Publication uses npm's default `latest` tag for all three packages: `latest` now resolves to `0.3.1` (previously `0.3.0`), `beta` remains `0.3.0-beta.3`, `alpha` remains `0.3.0-alpha.5`, and `rc` remains `0.2.0-rc.1`. GitHub Actions remains validation-only with OIDC and package-write permissions denied.
+
+## Candidate validation
+
+Local candidate verification on Node.js `26.5.1`:
+
+- `npm run check`: 70 files and 352 tests passed (including five new dedicated-model/thinking integration tests and the thinking-mapping unit tests).
+- `npm run quality:compare`: candidate quality versus the frozen baseline passed.
+- `npm run schema:context-persistence`: within the 1,500 / 320 absolute and relative limits.
+- `npm run latency:check` against the frozen baseline passed.
+- `npm run pack:check` and `npm pack --dry-run --json` for all three packages passed with no forbidden local/session/storage paths.
+- `git diff --check` passed; the only pre-existing untracked path is `.serena/`, which is excluded from commits and package inventories.
+- Version, exact core dependencies, package-lock entries, extension constant, and reference-adapter constant are synchronized to `0.3.1`.
+
+Validation-only CI is recorded below with the release commit. Exact registry verification and the annotated tag are recorded after execution.
+
+## Registry evidence
+
+To be completed after publication.
+
+## Documentation
+
+- [`../COMPACTION.md`](../COMPACTION.md)
+- [`../RELEASING.md`](../RELEASING.md)
+- [`0.3.0.md`](0.3.0.md)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ds4-context-engine",
-  "version": "0.3.0-beta.3",
+  "version": "0.3.1",
   "description": "Non-destructive, provider-independent context management for Pi.",
   "type": "module",
   "license": "MIT",
@@ -62,7 +62,7 @@
     ]
   },
   "dependencies": {
-    "ds4-context-core": "0.3.0-beta.3"
+    "ds4-context-core": "0.3.1"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "0.84.3",

package/src/pi-adapter/compaction-coordinator.ts

@@ -5,7 +5,8 @@ import type {
   SessionCompactEvent,
   SessionEntry,
 } from "@earendil-works/pi-coding-agent";
-import type { Ds4ContextConfig } from "ds4-context-core/config/config";
+import type { Api, Model } from "@earendil-works/pi-ai";
+import type { CompactionThinkingLevel, Ds4ContextConfig } from "ds4-context-core/config/config";
 import { calculateContextBudget, type ContextBudget } from "ds4-context-core/core/budget-manager";
 import { createModelProfile, type ModelDescriptor } from "ds4-context-core/core/model-profile";
 import type { ContextManifest } from "ds4-context-core/manifest/context-manifest";
@@ -212,7 +213,7 @@ export class CompactionCoordinator {
     ctx: ExtensionContext,
   ): Promise<{ compaction?: CompactionResult } | undefined> {
     const config = this.dependencies.config;
-    const model = ctx.model;
+    const model = this.resolveCompactionModel(ctx) ?? ctx.model;
     if (!config.compaction.enabled || !config.enabled || !model || config.context.maxSummaryTokens <= 0) {
       return undefined;
     }
@@ -278,6 +279,8 @@ export class CompactionCoordinator {
           modifiedFiles: plan.modifiedFiles,
           event,
           ctx,
+          model,
+          thinking: this.dependencies.config.compaction.summary?.thinking,
         });
         usages.push(generated.usage);
         segmentNodes.push({
@@ -324,6 +327,7 @@ export class CompactionCoordinator {
         nextId,
         createdNodes,
         usages,
+        modelObject: model,
       });
       const activeNode = aggregated.activeNode;
       const embeddedNodes = createdNodes.filter((node) => node.id !== activeNode.id);
@@ -664,6 +668,45 @@ export class CompactionCoordinator {
     };
   }
 
+  private resolveCompactionModel(ctx: ExtensionContext): Model<Api> | undefined {
+    const configured = this.dependencies.config.compaction.model;
+    if (!configured || !ctx.model || !ctx.modelRegistry) return undefined;
+    const dedicated = ctx.modelRegistry.find(configured.provider, configured.id);
+    if (!dedicated) {
+      this.dependencies.logger.warn("compaction.dedicated_model_fallback", {
+        provider: configured.provider,
+        id: configured.id,
+        reason: "not-found",
+        sessionModel: ctx.model.id,
+      });
+      return undefined;
+    }
+    if (!ctx.modelRegistry.hasConfiguredAuth(dedicated)) {
+      this.dependencies.logger.warn("compaction.dedicated_model_fallback", {
+        provider: configured.provider,
+        id: configured.id,
+        reason: "no-auth",
+        sessionModel: ctx.model.id,
+      });
+      return undefined;
+    }
+    if (!dedicated.input?.includes("text")) {
+      this.dependencies.logger.warn("compaction.dedicated_model_fallback", {
+        provider: configured.provider,
+        id: configured.id,
+        reason: "no-text-input",
+        sessionModel: ctx.model.id,
+      });
+      return undefined;
+    }
+    this.dependencies.logger.debug("compaction.dedicated_model", {
+      provider: dedicated.provider,
+      id: dedicated.id,
+      sessionModel: ctx.model.id,
+    });
+    return dedicated;
+  }
+
   private inputBudgetTokens(model: ModelDescriptor): number {
     const config = this.dependencies.config;
     const resolved = this.dependencies.resolveModelBudget?.(model)
@@ -841,6 +884,7 @@ export class CompactionCoordinator {
     ctx: ExtensionContext;
     provider: string;
     model: string;
+    modelObject: Model<Api>;
     readFiles: readonly string[];
     modifiedFiles: readonly string[];
     inputBudgetTokens: number;
@@ -914,6 +958,8 @@ export class CompactionCoordinator {
           modifiedFiles: plan.modifiedFiles,
           event: input.event,
           ctx: input.ctx,
+          model: input.modelObject,
+          thinking: this.dependencies.config.compaction.summary?.thinking,
         });
         input.usages.push(generated.usage);
         const aggregateNode: EmbeddedSummaryNode = {
@@ -952,6 +998,8 @@ export class CompactionCoordinator {
     modifiedFiles: readonly string[];
     event: SessionBeforeCompactEvent;
     ctx: ExtensionContext;
+    model: Model<Api>;
+    thinking?: CompactionThinkingLevel;
   }): Promise<GeneratedSummary> {
     return generateValidatedSummary({
       ...input,

package/src/pi-adapter/summary-generator.ts

@@ -1,9 +1,10 @@
 import { randomUUID } from "node:crypto";
-import type { Usage } from "@earendil-works/pi-ai";
+import type { Api, Model, Usage } from "@earendil-works/pi-ai";
 import type {
   ExtensionContext,
   SessionBeforeCompactEvent,
 } from "@earendil-works/pi-coding-agent";
+import type { CompactionThinkingLevel } from "ds4-context-core/config/config";
 import {
   analyzeUnsupportedExactValueBullets,
   groundSummaryFileSections,
@@ -33,6 +34,10 @@ export interface GenerateValidatedSummaryInput {
   maxSummaryTokens: number;
   event: SessionBeforeCompactEvent;
   ctx: ExtensionContext;
+  /** Dedicated compaction model; falls back to `ctx.model` when absent. */
+  model?: Model<Api>;
+  /** Reasoning level for the summary request; `off` (default) keeps the pre-existing request shape. */
+  thinking?: CompactionThinkingLevel;
   now: () => number;
   onTransportRetry?: (diagnostic: CompactionTransportRetryDiagnostic) => void;
 }
@@ -47,6 +52,36 @@ function unique(values: readonly string[]): string[] {
   return [...new Set(values)];
 }
 
+/**
+ * Maps a compaction thinking level to provider-specific request options.
+ * Unknown/unsupported APIs return an empty object so the request keeps the
+ * pre-existing shape (no thinking fields). `off` also returns an empty object.
+ */
+export function compactionThinkingOptions(
+  api: string,
+  level: CompactionThinkingLevel | undefined,
+): Record<string, unknown> {
+  if (!level || level === "off") return {};
+  if (api === "anthropic-messages") {
+    const effort = level === "minimal" || level === "low" ? "low" : level;
+    return { thinkingEnabled: true, effort };
+  }
+  if ([
+    "openai-completions",
+    "openai-responses",
+    "azure-openai-responses",
+    "openai-codex-responses",
+  ].includes(api)) {
+    const effort = level === "minimal" || level === "low"
+      ? "low"
+      : level === "xhigh" || level === "max"
+        ? "high"
+        : level;
+    return { samplingParams: { reasoning_effort: effort } };
+  }
+  return {};
+}
+
 function responseText(response: unknown): string {
   if (!response || typeof response !== "object" || !("content" in response)) return "";
   const content = (response as { content?: unknown }).content;
@@ -124,7 +159,7 @@ function transportFailureSuffix(category: ProviderFailureCategory, attempts: num
 export async function generateValidatedSummary(
   input: GenerateValidatedSummaryInput,
 ): Promise<GeneratedSummary> {
-  const model = input.ctx.model;
+  const model = input.model ?? input.ctx.model;
   if (!model) throw new Error("Compaction summary generation requires an active model");
   const maxTokens = Math.max(1, Math.min(input.maxSummaryTokens, model.maxTokens ?? input.maxSummaryTokens));
   const retryUsages: Usage[] = [];
@@ -148,7 +183,8 @@ export async function generateValidatedSummary(
           signal: input.event.signal,
           cacheRetention: "none",
           sessionId: randomUUID(),
-        },
+          ...compactionThinkingOptions(model.api, input.thinking),
+        } as NonNullable<Parameters<typeof input.ctx.modelRegistry.complete>[2]>,
       );
     } catch (error) {
       if (input.event.signal.aborted) throw abortedError();

package/src/pi-adapter/version.ts

@@ -1,4 +1,4 @@
-export const EXTENSION_VERSION = "0.3.0-beta.3";
+export const EXTENSION_VERSION = "0.3.1";
 export const SUPPORTED_PI_VERSION = "0.84.3";
 export const OBSERVER_PLANNER_VERSION = "observer-model-aware-v1";
 export const PLANNER_VERSION = "managed-learned-ranking-v1";