pi-pr-review 1.15.13 → 1.15.15

package/CHANGELOG.md

@@ -1,5 +1,19 @@
 # Changelog
 
+## [1.15.15](https://github.com/10ego/pi-pr-review/compare/v1.15.14...v1.15.15) (2026-08-28)
+
+
+### Bug Fixes
+
+* **review:** eliminate large-review retry amplification ([#116](https://github.com/10ego/pi-pr-review/issues/116)) ([9aa0b82](https://github.com/10ego/pi-pr-review/commit/9aa0b82c3dbf80b3bb707634acd32c4a0c61b7df))
+
+## [1.15.14](https://github.com/10ego/pi-pr-review/compare/v1.15.13...v1.15.14) (2026-08-28)
+
+
+### Bug Fixes
+
+* **review:** route oversized shards through read-only files ([#114](https://github.com/10ego/pi-pr-review/issues/114)) ([e300264](https://github.com/10ego/pi-pr-review/commit/e3002640e5d9060245a11a142a334652370598c0))
+
 ## [1.15.13](https://github.com/10ego/pi-pr-review/compare/v1.15.12...v1.15.13) (2026-08-28)
 
 

package/README.md

@@ -70,7 +70,7 @@ A review uses five focused passes by default:
 4. security (`heavy`);
 5. performance and resource ownership (`heavy`).
 
-`--full` adds a convention and maintainability pass (`medium`). Large multi-file diffs are split by whole-file boundaries and reviewed in parallel. The host enforces this before model dispatch: multi-file diffs from 200,000–399,999 bytes use two shards, and diffs of at least 400,000 bytes with at least three files use three shards, even if an orchestrator accidentally supplied the unified diff inline without requesting sharding. This avoids spending model attempts on payloads already known to exceed ordinary context budgets. Failed registered lanes remain visible and cannot be erased by model-named reruns. If the extension is unavailable, the prompt falls back to the current Pi session model.
+`--full` adds a convention and maintainability pass (`medium`). Large multi-file diffs are split by whole-file boundaries and reviewed in parallel. The host enforces this before model dispatch: multi-file diffs from 200,000–399,999 bytes use two shards, and diffs of at least 400,000 bytes with at least three files use three shards, even if an orchestrator accidentally supplied the unified diff inline without requesting sharding. This avoids spending model attempts on payloads already known to exceed ordinary context budgets. If a whole-file shard still exceeds 1 MiB and has a captured `context_file`, the host sends only a bounded diff-header manifest and forces configured read-only tools so the reviewer inspects the file-backed shard instead of embedding megabytes in its prompt. Failed registered lanes remain visible and cannot be erased by model-named reruns. If the extension is unavailable, the prompt falls back to the current Pi session model.
 
 When host lane evidence degrades or is incomplete, TUI and automation output starts with the deterministic host-owned `Comment` verdict and coverage disclosure. Any contradictory model verdict retained for evidence is relabeled as non-authoritative.
 

package/extensions/pr-review-subagent.ts

@@ -127,6 +127,7 @@ const TOOL_POLICIES: ToolPolicy[] = ["configured", "none"];
 const TOOLS_PRESETS = ["read,bash,grep,find,ls", "read,grep,find,ls", "read"];
 const DEFAULT_BATCH_PARALLEL = 4;
 const MAX_BATCH_PARALLEL = 18;
+const MAX_EMBEDDED_REVIEW_SHARD_BYTES = 1024 * 1024;
 const TIER_PURPOSE: Record<Tier, string> = {
 	light: "overview / strengths / high-level risk scan",
 	medium: "convention compliance + readability / maintainability",
@@ -814,6 +815,8 @@ interface SubagentPassRequest {
 	generation?: number;
 	artifactKey?: string;
 	requestedPassOrdinal?: number;
+	toolNames?: readonly string[];
+	fileBackedContext?: boolean;
 }
 
 interface ModelAttemptReport {
@@ -928,7 +931,7 @@ async function runSubagentAttempt(
 		const args = buildReviewBaseArgs();
 		if (attempt.spec) args.push("--model", attempt.spec);
 		appendTierThinkingArgs(args, attempt.spec, config.thinkingLevels[pass.tier]);
-		appendToolPolicyArgs(args, toolPolicy, config.tools);
+		appendToolPolicyArgs(args, toolPolicy, pass.toolNames ?? config.tools);
 
 		tmp = await writeTempPrompt(
 			pass.tier,
@@ -1868,41 +1871,65 @@ export default function registerPrReviewSubagents(
 				? shardUnifiedDiff(diffForSharding!, requestedShardCount)
 				: [];
 			const sharded = requestedShards.length > 1;
+			const topUnshardedFileBacked = !sharded && loadedContext.contextFileBytes > MAX_EMBEDDED_REVIEW_SHARD_BYTES && !!loadedContext.contextFile && !!loadedContext.contextFileText;
 			const selectedFileDiffIndex = loadedDiff ? inlineContext.indexOf(loadedDiff) : -1;
 			const compactInlineContext = selectedFileDiffIndex >= 0
 				? [inlineContext.slice(0, selectedFileDiffIndex).trim(), inlineContext.slice(selectedFileDiffIndex + loadedDiff!.length).trim()].filter(Boolean).join("\n\n")
 				: inlineDiffStart >= 0 && !loadedDiff
 					? inlineContext.slice(0, inlineDiffStart).trim()
 					: inlineContext;
-			const sharedContext = sharded ? compactInlineContext || undefined : loadedContext.context;
+			const sharedContext = sharded || topUnshardedFileBacked ? compactInlineContext || undefined : loadedContext.context;
 			const majorOnly = params.major_only === true;
 			const minorHygiene = params.minor_hygiene === true;
-			const passesWithoutFocus: SubagentPassRequest[] = rawPasses.flatMap((pass, index) => {
+			const exactHeaderManifest = (text: string) => {
+				const headers = text.match(/^diff --git .+$/gm) ?? [], bytes = Buffer.byteLength(headers.join("\n"), "utf8");
+				return headers.length > 0 && headers.length <= 2_000 && bytes <= 256 * 1024 && headers.every((line) => Buffer.byteLength(line, "utf8") <= 2_048 && !/[\u0000-\u001f\u007f]/.test(line)) ? headers : undefined;
+			};
+			const topShardHeaders = requestedShards.map((shard) => loadedContext.contextFile && Buffer.byteLength(shard, "utf8") > MAX_EMBEDDED_REVIEW_SHARD_BYTES ? exactHeaderManifest(shard) : []);
+			if (topShardHeaders.some((headers) => headers === undefined)) {
+				return { content: [{ type: "text", text: "Review batch context failed: oversized file-backed shard manifest is unsafe or exceeds deterministic bounds." }], isError: true, details: { passCount: rawPasses.length, contextFileBytes: loadedContext.contextFileBytes, shardCount: 0 } };
+			}
+			const fileBackedInstructions = (contextFile: string, label: string, headers?: readonly string[]) => [
+				`--- File-backed ${label} ---`,
+				`The complete captured unified diff is at ${JSON.stringify(contextFile)}. Use only the host-enforced read-only tools (read, grep, find, ls) to inspect it; the oversized diff is intentionally not embedded in this prompt.`,
+				...(headers ? ["Review only these exact diff headers:", ...headers] : ["Review the complete captured diff in this file."]),
+			].join("\n");
+			const passesWithoutFocus = rawPasses.flatMap((pass, index) => {
 				const tier = pass.tier as Tier;
 				const baseId = typeof pass.id === "string" && pass.id.trim() ? pass.id.trim() : `${index + 1}-${tier}`;
-				const baseContext = combineContexts(sharedContext, loadedPassContexts[index]!.context);
-				const makePass = (shard: string | undefined, shardIndex: number): SubagentPassRequest => ({
-					// Every actual shard receives an explicit identity, including shard 1.
-					// The unsharded compatibility path retains the caller's exact base ID.
-					id: sharded ? `${baseId}-shard-${shardIndex + 1}` : baseId,
-					expectedOutput: pass.expected_output,
-					tier,
-					objective: shard
-						? `${pass.objective}\nReview every changed line in diff shard ${shardIndex + 1}/${requestedShards.length} under this objective. Other concurrent shard passes cover the remaining changed files.`
-						: pass.objective,
-					context: shard
-						? combineContexts(
-							baseContext,
-							`--- Complete diff shard ${shardIndex + 1}/${requestedShards.length} ---\n${shard}`,
-						)
-						: baseContext,
-					toolPolicy: normalizeToolPolicy(pass.tool_policy),
-					majorOnly,
-					minorHygiene: minorHygiene && tier === "light" && baseId === "overview",
-				});
-				return sharded
-					? requestedShards.map((shard, shardIndex) => makePass(shard, shardIndex))
-					: [makePass(undefined, 0)];
+				const loadedPass = loadedPassContexts[index]!;
+				const passFileBacked = loadedPass.contextFileBytes > MAX_EMBEDDED_REVIEW_SHARD_BYTES && !!loadedPass.contextFile && !!loadedPass.contextFileText;
+				const passContext = passFileBacked ? (typeof pass.context === "string" ? pass.context.trim() : undefined) : loadedPass.context;
+				const baseContext = combineContexts(sharedContext, passContext);
+				const makePass = (shard: string | undefined, shardIndex: number) => {
+					const topFileBacked = (!!shard && Buffer.byteLength(shard, "utf8") > MAX_EMBEDDED_REVIEW_SHARD_BYTES || topUnshardedFileBacked) && !!loadedContext.contextFile;
+					const fileBacked = passFileBacked || topFileBacked;
+					const label = sharded ? `diff shard ${shardIndex + 1}/${requestedShards.length}` : passFileBacked ? `pass context for ${baseId}` : "complete diff";
+					const fileInstructions = passFileBacked
+						? fileBackedInstructions(loadedPass.contextFile!, label)
+						: topFileBacked
+							? fileBackedInstructions(loadedContext.contextFile!, label, shard ? topShardHeaders[shardIndex] as readonly string[] : undefined)
+							: undefined;
+					return {
+						// Every actual shard receives an explicit identity, including shard 1.
+						// The unsharded compatibility path retains the caller's exact base ID.
+						id: sharded ? `${baseId}-shard-${shardIndex + 1}` : baseId,
+						expectedOutput: pass.expected_output,
+						tier,
+						objective: shard
+							? `${pass.objective}\nReview every changed line in diff shard ${shardIndex + 1}/${requestedShards.length} under this objective. Other concurrent shard passes cover the remaining changed files.`
+							: pass.objective,
+						context: fileInstructions ? combineContexts(baseContext, fileInstructions) : shard
+							? combineContexts(baseContext, `--- Complete diff shard ${shardIndex + 1}/${requestedShards.length} ---\n${shard}`)
+							: baseContext,
+						toolPolicy: fileBacked ? "configured" as const : normalizeToolPolicy(pass.tool_policy),
+						...(fileBacked ? { toolNames: ["read", "grep", "find", "ls"] } : {}),
+						majorOnly,
+						minorHygiene: minorHygiene && tier === "light" && baseId === "overview",
+						fileBackedContext: fileBacked,
+					};
+				};
+				return sharded ? requestedShards.map((shard, shardIndex) => makePass(shard, shardIndex)) : [makePass(undefined, 0)];
 			});
 			if (!loopCoordinator.isLeaseActive(lease, ctx)) return reviewLoopDeniedResult("review_subagents");
 			const artifactPublisher = loopCoordinator.createArtifactPublisher(lease, ctx);
@@ -1986,6 +2013,7 @@ export default function registerPrReviewSubagents(
 					shardingSource: automaticShardCount > explicitShardCount ? "automatic-size-preflight" : explicitShardCount > 1 ? "requested" : "none",
 					diffBytes,
 					changedFileCount,
+					fileBackedPassCount: passes.filter((pass) => pass.fileBackedContext === true).length,
 					sharedContextBytes: Buffer.byteLength(sharedContext ?? "", "utf8"),
 					contextFileBytes:
 						loadedContext.contextFileBytes +

package/lib/pr-review-artifacts.ts

@@ -462,7 +462,11 @@ function expectedLaneSections(input: ReviewLaneCompletionInput): boolean {
 		const fields = input.minorHygiene ? ["overview", "strengths", "minor candidates"] : ["overview", "strengths"];
 		return fields.every((field) => hasMeaningfulLightSection(text, field, fields));
 	}
-	if (text === NO_FINDINGS_SENTINEL) return true;
+	// Ordinary lanes use the sentinel only as an unambiguous clean result; a
+	// missing final period carries no semantic uncertainty and must not amplify
+	// into expensive replacement passes. Integrated/deep `nonempty` output keeps
+	// its exact byte contract in parseIntegratedCompletion().
+	if (/^NO FINDINGS\.?$/.test(text)) return true;
 	return CANDIDATE_FIELDS.every((field) => hasMeaningfulField(text, field));
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-pr-review",
-  "version": "1.15.13",
+  "version": "1.15.15",
   "description": "Parallel AI code review for GitHub pull requests in the Pi coding agent, with model-agnostic tiered subagents, structured findings, optional verification, and host-gated COMMENT or qualified APPROVE publishing.",
   "keywords": [
     "pi-package",