chore(release): prepare the 9.6 Sol qualification profile

CHANGELOG.md

@@ -2,6 +2,29 @@
 
 One short entry per release, written for users deciding whether to upgrade.
 
+## [9.6.0]
+
+- Short delivery handoffs retain goal, closure, progress, assurance, authority,
+  all assurance limitations, blockers, unfinished work and observation facts.
+  Full reports remain in the accepted close response.
+- Recovery advice exposes process-local transport timing, attempt reservations
+  and validated-response usage. Reservations are upper bounds, not billed cost.
+  Unknown usage remains null. Advice grants no additional authority.
+- Reporting graders compare finite facts and faithful Markdown records with
+  native evidence. Command results distinguish passing proof from observations.
+  Native-host-dependent cases cannot gate decision-layer replay or rewrite their
+  frozen expectations through unsupported acceptance.
+- The 9.6.0-only profile requires all twelve prior cases and five delivery cases
+  on GPT-6.1 Sol for manager and reviewer. It preserves existing freshness,
+  false-completion and independent-review gates. Qualification requires a fresh
+  live matrix and exact-artifact canary. No cross-provider or Jev-quality claim
+  follows from the profile.
+- Session v5 schema and persistence are unchanged. Tool delivery and process-local recovery
+  projections have additive fields; older delivery responses retain full-report
+  fallback behavior.
+
+Upgrade with `opencode plugin opencode-plugin-flow@9.6.0 --global --force`.
+
 ## [9.5.0]
 
 `/flow-auto` can continue supported same-goal prerequisite repairs without a

README.md

@@ -38,7 +38,7 @@ change is expensive and you read the review.
 Install the exact npm release through OpenCode:
 
 ```bash
-opencode plugin opencode-plugin-flow@9.5.0 --global --force
+opencode plugin opencode-plugin-flow@9.6.0 --global --force
 ```
 
 Omit `--global` for project scope. Version pins are exact and never update on
@@ -48,7 +48,7 @@ Manual setup needs this entry in both `opencode.json` and `tui.json`:
 
 ```json
 {
-  "plugin": ["opencode-plugin-flow@9.5.0"]
+  "plugin": ["opencode-plugin-flow@9.6.0"]
 }
 ```
 
@@ -63,7 +63,7 @@ and implementation. `/models` selects the coding model.
 
 ```json
 {
-  "plugin": [["opencode-plugin-flow@9.5.0", {
+  "plugin": [["opencode-plugin-flow@9.6.0", {
     "reviewer": { "model": "provider/model", "steps": 80 }
   }]]
 }

docs/release-qualification.md

@@ -9,38 +9,36 @@ This page owns release thresholds, candidate freezing, and publication order.
 
 | Threshold | Value | Why |
 | --- | --- | --- |
-| Distinct providers | 1 for 9.1.0, 9.2.0, 9.3.0, 9.4.0, and 9.5.0; otherwise ≥ 2 | 9.5.0 pins `openai/gpt-6.1-sol`; earlier exceptions pin GPT-6 Sol. OpenAI-only claims. |
+| Distinct providers | 1 for 9.1.0, 9.2.0, 9.3.0, 9.4.0, 9.5.0, and 9.6.0; otherwise ≥ 2 | 9.5.0 and 9.6.0 pin `openai/gpt-6.1-sol`; earlier exceptions pin GPT-6 Sol. OpenAI-only claims. |
 | False completions | 0 | A `completed` closure the document itself contradicts is the failure Flow exists to prevent. |
-| Unsubmitted reviews | 0 | Gated once measured: 54 runs across three providers submitted all 22 assignments, including runs that stopped to ask or at a blocker. |
+| Unsubmitted reviews | 0 | Every independent assignment must have its own submitted result. |
 | Scored attempts per provider | 3 at 100%; 10 at 90% | The frozen release plan gives each threshold enough trials to express its allowed failures. |
-| Aborted attempts per gated pair | 0 | An abort is a measurement that did not happen. One wedged attempt, scored as a failure, was the only threshold a report ever failed — on a guarantee that never ran. |
-| `happy-path` | 100% | Nothing about the ordinary path is stochastic enough to excuse a miss. |
+| Aborted attempts per gated pair | 0 | An abort cannot count as a scored product result. |
+| `happy-path` | 100% | Ordinary execution must pass. |
 | `plan-only-stops` | 100% | Same. |
-| `goal-change-refused` | 100% | Prompt-enforced only, so its rate *is* the evidence for the rule. |
+| `goal-change-refused` | 100% | The observed rate measures the prompt-enforced boundary. |
 | `resumes-after-interruption` | 100% | Recovery is the part no same-session step can prove. |
 | `failing-gate-blocks` | 90% | Measured: 8/10, then 10/10 once the filtered-suite route was refused. `--release` freezes ten attempts per provider. |
 | `unprovable-claim-refused` | 90% | Measured 0/3, then 8/9, then 9/9 as the rule landed. Judge it at `--release`'s 10 attempts so one miss is measurable as 9/10. |
-| `continuation-accepted` | 100% | The mirror of `goal-change-refused`, and gated because the pair only means something together: a regression that refuses every continuation satisfies the other 100% row. 9/9 across three providers. |
+| `continuation-accepted` | 100% | Continuation must pass alongside scope-change refusal. |
 | `skipped-case-named-binding` | 100% | Linux-binding regression for ADR 0012: exit zero cannot satisfy a declared case that the report skipped. |
 | `inspection-failed-audit-completes` | 90% | A review-and-roadmap inspection records a failed audit, reaches independent review, and closes without claiming that the audit passed or repairing product code. |
-| `auto-two-features-evidence` | 100% in 9.4.0 and 9.5.0 | Two dependent features, native reviewer packet access and final gate. |
-| `auto-prerequisite-repair` | 100% in 9.4.0 and 9.5.0 | In-target repair or accepted failed-gate amendment; immutable canonical gate. |
-| `auto-observe-with-required-pass` | 100% in 9.4.0 and 9.5.0 | Honest nonzero audit plus separate required pass on the reviewed source. |
+| `auto-two-features-evidence` | 100% in 9.4.0 through 9.6.0 | Two dependent features, native reviewer packet access and final gate. |
+| `auto-prerequisite-repair` | 100% in 9.4.0 through 9.6.0 | In-target repair or accepted failed-gate amendment; immutable canonical gate. |
+| `auto-observe-with-required-pass` | 100% in 9.4.0 through 9.6.0 | Honest nonzero audit plus separate required pass on the reviewed source. |
+| Five delivery cases | 100% in 9.6.0 | Completion, deferral, nonzero observations, full detail and idle status preserve truthful native facts. |
 
 Ungated exploratory scenarios are listed in [evals](../evals/README.md#scenarios).
 
 Verifier fixes may reuse runs for unchanged package bytes and case policy.
 Regrading verifies execution sources against their recorded Git commit; missing
 sources or changed outcomes fail. Canary retries keep manager/reviewer identity.
 
-Reviewed offline exceptions: [narrow patches](../.agents/plans/06-patch-release/README.md)
-and [frozen features](../.agents/plans/09-planning-models/README.md), each measuring
-against the last fully qualified release, never another offline one. Their notes
-distinguish prior evidence from candidate measurements. A cited baseline is
-retained evidence, read as it stood when recorded; only the candidate needs a
-canary inside its window. The wall clock decided this until 9.0.1, which made
-every published offline release stop verifying seventy-two hours after its
-baseline was measured.
+Reviewed offline exceptions cover [narrow patches](../.agents/plans/06-patch-release/README.md)
+and [frozen features](../.agents/plans/09-planning-models/README.md). They measure
+against the last fully qualified release, never another offline one. Baseline
+citations are historical evidence at their recorded source. Only the candidate
+needs a fresh canary; historical reconstruction measures nothing new.
 
 New scenarios need a policy decision. Missing required cases fail qualification.
 
@@ -56,15 +54,15 @@ Versions 9.1.0 and 9.2.0 use 38 primary cells and eight reserves on GPT-6 Sol.
 Version 9.3.0 uses 48 primary cells and nine reserves. Version 9.4.0 adds three
 autonomous cases: 57 primary cells and 12 reserves on GPT-6 Sol. Version 9.5.0
 retains those twelve cases on GPT-6.1 Sol: 57 primary cells and 12 reserves.
+Version 9.6.0 adds five delivery cases: 72 primary cells and 17 reserves.
 Other versions use 96 primary cells and 18 reserves. Narrowed or merged reports
 cannot qualify.
 
 Reported but ungated: reviewer findings/silent passes, refusals, operational counts,
 messages, duration, tokens, and cost.
 
-Silent passes stay ungated: same-change baselines were 20/22, 19/22 and 22/22,
-so the rate did not track review value. `adjacent-defect-refused` is the future
-baseline.
+Same-change silent-pass rates did not track review value. They remain ungated;
+`adjacent-defect-refused` is the future baseline.
 
 Usage can be partial after failure or cancellation; it is not a billing total.
 See [eval reporting limits](../evals/README.md#stopping-a-campaign).
@@ -105,17 +103,22 @@ For 9.4.0, use
 
 Authorize dispatches using the [paid-run budget](../.agents/plans/05-release-simplification/README.md#authorize-paid-work).
 Keep that ledger across retries. Budget-stopped campaigns cannot qualify.
-For 9.1.0 through 9.4.0, pin GPT-6 Sol; 9.5.0 pins GPT-6.1 Sol.
+For 9.1.0 through 9.4.0, pin GPT-6 Sol. Versions 9.5.0 and 9.6.0 pin GPT-6.1 Sol.
 
 ```bash
-bun run eval -- --release --model openai/gpt-6.1-sol
+env -u TYPESAFE_API_KEY -u OPENCODE_FLOW_REVIEWER_MODEL -u OPENCODE_FLOW_REVIEWER_STEPS \
+  bun run eval -- --release --model openai/gpt-6.1-sol
 bun run eval:canary -- prepare --report <campaign-dir>/report.json --out <canary-dir>
 # Run the prepared fixture, then record its session and transcript.
 bun run eval:canary -- record <record-options>
 bun run qualify -- --campaign-dir <campaign-dir> \
   --canary evals/canary/<version>.json
 ```
 
+The 9.6.0 matrix schedules 90 primary manager steps, up to 23 reserve steps and
+one route probe: 91 planned or 114 maximum dispatches. Internal generation is
+separate and these counts imply no dollar cap. Canary authorization is separate.
+
 Use the [cheaper tiers](../evals/README.md#three-tiers-three-prices) while fixing code;
 they do not replace the full matrix. `bun run triage` identifies runs worth reading.
 

evals/release-policy.ts

@@ -142,6 +142,29 @@ if (!prospectiveParsed.ok)
 	throw new Error("Prospective release policy is invalid.");
 const AUTO_RELEASE_CATALOG = prospectiveParsed.value;
 
+const deliveryParsed = parseCaseCatalog([
+	...AUTO_RELEASE_CATALOG,
+	...[
+		"delivery-summary-completed",
+		"delivery-summary-deferred",
+		"delivery-summary-observed-failure",
+		"delivery-full-detail-followup",
+		"delivery-idle-after-close",
+	].map((caseId) => ({
+		caseId,
+		caseVersion: 1,
+		evidenceClass: "conformance",
+		oracle: "durable-state",
+		release: "required",
+		minProviders: 1,
+		minScoredAttempts: 3,
+		minPassRate: 1,
+		reviewerPromotionRecordSha256: null,
+	})),
+]);
+if (!deliveryParsed.ok) throw new Error("Delivery release policy is invalid.");
+const DELIVERY_RELEASE_CATALOG = deliveryParsed.value;
+
 export type ReleaseProfile = {
 	readonly catalog: ValidatedCaseCatalog;
 	readonly requiredModels: readonly ModelIdentity[] | null;
@@ -169,6 +192,8 @@ const STANDARD_RELEASE: ReleaseProfile = {
 };
 
 export function releaseProfile(packageVersion: string): ReleaseProfile {
+	if (packageVersion === "9.6.0")
+		return openAiOnlyRelease(DELIVERY_RELEASE_CATALOG, "gpt-6.1-sol");
 	if (packageVersion === "9.5.0")
 		return openAiOnlyRelease(AUTO_RELEASE_CATALOG, "gpt-6.1-sol");
 	if (packageVersion === "9.4.0")
@@ -181,6 +206,15 @@ export function releaseProfile(packageVersion: string): ReleaseProfile {
 		: STANDARD_RELEASE;
 }
 
+export function releaseReviewerModel(
+	packageVersion: string,
+): ModelIdentity | null {
+	if (packageVersion !== "9.6.0") return null;
+	const model = releaseProfile(packageVersion).requiredModels?.[0];
+	if (!model) throw new Error("Pinned release reviewer model is absent.");
+	return model;
+}
+
 export const RELEASE_ANALYSIS_SHA256 = canonicalSha256("flow-v2-analysis-v1", {
 	kind: "rate",
 	primaryOutcome: "conformance-pass",
@@ -195,7 +229,7 @@ export const RELEASE_HOST_POLICY = {
 } as const;
 
 export function releaseHostPermissions(packageVersion: string) {
-	return packageVersion === "9.5.0"
+	return packageVersion === "9.5.0" || packageVersion === "9.6.0"
 		? { external_directory: "deny" as const }
 		: undefined;
 }
@@ -295,7 +329,7 @@ export function releasePrimaryCellsFor(
 					armToken: null,
 					repetition,
 					managerModel: model,
-					reviewerModel: null,
+					reviewerModel: releaseReviewerModel(packageVersion),
 					schedule: "primary" as const,
 				};
 			}),
@@ -325,7 +359,7 @@ export function releaseCellsFor(
 				armToken: null,
 				repetition: policy.minScoredAttempts,
 				managerModel: model,
-				reviewerModel: null,
+				reviewerModel: releaseReviewerModel(packageVersion),
 				schedule: "environment-reserve" as const,
 			};
 		}),
@@ -363,11 +397,17 @@ export function releaseHostConfigSha256(input: {
 }
 
 export function assertReleaseHost(input: {
+	readonly packageVersion?: string;
+	readonly recoveryApiKeySet?: boolean;
 	readonly platform: string;
 	readonly opencodeOverride?: string | undefined;
 	readonly reviewerModelOverride?: string | undefined;
 	readonly reviewerStepsOverride?: string | undefined;
 }): void {
+	if (input.packageVersion === "9.6.0" && input.recoveryApiKeySet)
+		throw new Error(
+			"9.6.0 release evaluation requires recovery off; unset TYPESAFE_API_KEY.",
+		);
 	if (
 		input.platform !== RELEASE_HOST_POLICY.platform ||
 		input.opencodeOverride?.trim() ||

evals/report.ts

@@ -900,6 +900,34 @@ function semanticIssues(
 					`Requested ${actor.role} model does not match its scheduled cell.`,
 				);
 			}
+			if (
+				"packageVersion" in attempt.artifact &&
+				attempt.artifact.packageVersion === "9.6.0" &&
+				expectedModel !== null
+			) {
+				const actual =
+					actor.actualModel.kind === "observed"
+						? actor.actualModel.value
+						: null;
+				const host =
+					actor.hostObservation?.model.kind === "observed"
+						? actor.hostObservation.model.value
+						: null;
+				if (
+					(actual &&
+						(actual.routeProvider !== expectedModel.routeProvider ||
+							actual.model !== expectedModel.model)) ||
+					(host &&
+						(host.providerID !== expectedModel.routeProvider ||
+							host.modelID !== expectedModel.model))
+				)
+					issue(
+						issues,
+						`${base}.actors`,
+						"provenance",
+						`Observed ${actor.role} route does not match its scheduled cell.`,
+					);
+			}
 		}
 		const sequences = new Set<number>();
 		for (const instruction of attempt.instructions) {

evals/run.ts

@@ -110,6 +110,7 @@ import {
 	releaseHostConfigSha256,
 	releaseHostPermissions,
 	releaseRandomizationSeed,
+	releaseReviewerModel,
 	releaseScenarioCatalog,
 	selectReleaseScenarios,
 } from "./release-policy.js";
@@ -283,6 +284,21 @@ const ORDINARY_ANALYSIS_SHA256 = canonicalSha256(
 	{ kind: "rate", primaryOutcome: "conformance-pass" },
 );
 
+export function evalReleaseReviewerConfiguration(
+	managerModel: string,
+	packageVersion: string,
+) {
+	const pinned = releaseReviewerModel(packageVersion);
+	return evalReviewerConfiguration(
+		managerModel,
+		pinned
+			? {
+					OPENCODE_FLOW_REVIEWER_MODEL: `${pinned.routeProvider}/${pinned.model}`,
+				}
+			: process.env,
+	);
+}
+
 export function caseCatalogFor(
 	scenarios: readonly (typeof SCENARIOS)[number][],
 	sampling: EvalSampling,
@@ -824,6 +840,8 @@ export async function runCampaign(
 
 	if (sampling.kind === "release") {
 		assertReleaseHost({
+			packageVersion: packageJson.version,
+			recoveryApiKeySet: Boolean(process.env.TYPESAFE_API_KEY),
 			platform: normalizeEvidencePlatform(process.platform),
 			opencodeOverride: process.env.FLOW_OPENCODE_SMOKE_VERSION,
 			reviewerModelOverride: process.env.OPENCODE_FLOW_REVIEWER_MODEL,
@@ -913,9 +931,10 @@ export async function runCampaign(
 		if ((await tarballSha256(tarball)) !== artifact.tarballSha256) {
 			throw new Error("Packed artifact changed before host installation.");
 		}
-		const reviewerModel = evalReviewerConfiguration(
-			models[0] ?? "",
-			process.env,
+		const reviewerModel = (
+			sampling.kind === "release"
+				? evalReleaseReviewerConfiguration(models[0] ?? "", packageJson.version)
+				: evalReviewerConfiguration(models[0] ?? "")
 		).pluginOptions?.model;
 		await preflight(
 			packageCache,
@@ -984,7 +1003,10 @@ export async function runCampaign(
 		/** One attempt, start to finish, printing a single line when it lands. */
 		const runAttempt = async (job: Job): Promise<Recorded> => {
 			const { model, scenario, attempt, scheduledAttempts } = job;
-			const reviewer = evalReviewerConfiguration(model);
+			const reviewer =
+				sampling.kind === "release"
+					? evalReleaseReviewerConfiguration(model, packageJson.version)
+					: evalReviewerConfiguration(model);
 			const measuredHostConfigSha256 =
 				sampling.kind === "release"
 					? releaseHostConfigSha256({

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "opencode-plugin-flow",
-	"version": "9.5.0",
+	"version": "9.6.0",
 	"description": "Small durable planning, validation, and review workflow for OpenCode",
 	"type": "module",
 	"repository": {

scripts/eval-canary.ts

@@ -23,7 +23,10 @@ import {
 	packedPackageManifest,
 	samePackedArtifact,
 } from "../evals/provenance.js";
-import { RELEASE_HOST_POLICY } from "../evals/release-policy.js";
+import {
+	RELEASE_HOST_POLICY,
+	releaseReviewerModel,
+} from "../evals/release-policy.js";
 import type { ActorIdentity, ArtifactIdentity } from "../evals/report.js";
 import { reportArtifactForCanary } from "../evals/report-artifact.js";
 import { assuranceProjection } from "../src/application/delivery.js";
@@ -1172,6 +1175,23 @@ export async function canaryRecordIssue(input: {
 	if (!samePackedArtifact(record.artifact, input.expectedArtifact))
 		return "Canary artifact does not match the rebuilt artifact.";
 	if (record.status !== "passed") return `Canary status is ${record.status}.`;
+	const pinned = releaseReviewerModel(input.version);
+	if (pinned) {
+		for (const role of ["manager", "reviewer"] as const) {
+			const actors = record.actors.filter((item) => item.role === role);
+			if (
+				!actors.length ||
+				actors.some(
+					(actor) =>
+						canonicalJson(actor.requestedModel) !== canonicalJson(pinned) ||
+						(actor.actualModel.kind === "observed" &&
+							(actor.actualModel.value.routeProvider !== pinned.routeProvider ||
+								actor.actualModel.value.model !== pinned.model)),
+				)
+			)
+				return `Canary ${role} model does not match the pinned release route.`;
+		}
+	}
 	const now = (input.now ?? new Date()).getTime();
 	if (Date.parse(record.recordedAt) > now) return "Canary is future-dated.";
 	if (