@yagni-app/code-staging 1.2.3-staging.1621.1 → 1.2.3-staging.1627.1

package/dist/extension/pipeline/findings.js

@@ -13,8 +13,12 @@
  * after MAX_REVIEW_ROUNDS rounds, always with an explicit reason.
  */
 import { MAX_REVIEW_ROUNDS } from "./types.js";
-const SEVERITIES = ["critical", "high", "medium", "low"];
+const SEVERITIES = ["critical", "high", "medium", "low", "needs_person"];
 const DEFAULT_LENS = "correctness";
+/** `NEEDS PERSON`, `needs-person` and `needs_person` all read as `needs_person`. */
+function normalizeSeverity(value) {
+    return value.trim().toLowerCase().replace(/[\s-]+/g, "_");
+}
 function isSeverity(value) {
     return SEVERITIES.includes(value);
 }
@@ -33,7 +37,7 @@ function parsePipeLine(line, lens) {
     const parts = line.split("|").map((p) => p.trim());
     if (parts.length < 2)
         return null;
-    const severity = parts[0].toLowerCase();
+    const severity = normalizeSeverity(parts[0]);
     if (!isSeverity(severity))
         return null;
     let finding;
@@ -81,7 +85,7 @@ function parseJsonFindings(raw, lens) {
         if (!item || typeof item !== "object")
             continue;
         const obj = item;
-        const severity = typeof obj.severity === "string" ? obj.severity.toLowerCase() : "";
+        const severity = typeof obj.severity === "string" ? normalizeSeverity(obj.severity) : "";
         if (!isSeverity(severity))
             continue;
         const finding = {

package/dist/extension/pipeline/goCommand.js

@@ -906,7 +906,7 @@ export function registerGoCommand(pi, deps = {}) {
                                     })()
                                     : {}),
                                 openPr: parsed.flags.pr,
-                                remainingFindings: result.findings.filter((f) => f.severity === "medium" || f.severity === "low"),
+                                remainingFindings: result.findings.filter((f) => f.severity === "medium" || f.severity === "low" || f.severity === "needs_person"),
                                 ...(finishBaselinePaths.length > 0 ? { baselinePaths: finishBaselinePaths } : {}),
                                 signal: runSignal,
                             });

package/dist/extension/pipeline/headlessGo.js

@@ -323,10 +323,12 @@ export async function runHeadlessGo(argv, deps = {}) {
         },
         onStage: (s) => {
             // `output` carries the plan stage's full text; the stream reports the
-            // boundary, never the body.
+            // boundary, never the body. `at` is when the boundary happened, so a
+            // host that reads the line later (a reattach replays the log) still
+            // dates it right.
             const { output: _output, ...boundary } = s;
             if (stream)
-                emit({ type: "pipeline_stage", ...boundary });
+                emit({ type: "pipeline_stage", ...boundary, at: Date.now() });
         },
         logger: (event, data) => {
             deps.logger?.(event, data);

package/dist/extension/pipeline/orchestrator.d.ts

@@ -30,7 +30,17 @@ import { type ResilienceAttemptRecord, type RunStageFn } from "./resilience.js";
 import { runStage as defaultRunStage } from "./runner.js";
 import { type VerifyOutcome, type WorkstreamCheckResult, type WorkstreamCheckTarget } from "./verify.js";
 import { snapshotWorkspace as defaultSnapshotWorkspace } from "./workspace.js";
-import { type CheckpointStore, type FanoutBudgetVerdict, type FanoutMode, type PipelineStage, type JsonEvent, type PipelineProgress, type PipelineResult, type ResiliencePolicy, type ResumePlan, type StageId, type StageResult, type StageTag } from "./types.js";
+import { type CheckpointStore, type FanoutBudgetVerdict, type FanoutMode, type Finding, type PipelineStage, type JsonEvent, type PipelineProgress, type PipelineResult, type ResiliencePolicy, type ResumePlan, type StageId, type StageResult, type StageTag } from "./types.js";
+/** Each review round's findings as the host stores them: bounded, messages clipped. */
+export declare const MAX_ROUND_ITEMS = 20;
+export interface RoundFindingItem {
+    severity: Finding["severity"];
+    lens: Finding["lens"];
+    message: string;
+    file?: string;
+    line?: number;
+}
+export declare function roundFindingItems(findings: readonly Finding[]): RoundFindingItem[];
 export interface RunPipelineDeps {
     /** Bounded external design preparation; never runs inside the read-only planner or approved missions. */
     prepareDesign?: (input: {
@@ -112,6 +122,10 @@ export interface RunPipelineDeps {
      * `output` carries a stage's `finalOutput` ONLY on the plan FINISH boundary, so
      * /go can record the plan onto the work item without streaming every stage's
      * full text. It is absent on every other boundary.
+     *
+     * `items` carries each review round's own findings on the review FINISH
+     * boundary (at most {@link MAX_ROUND_ITEMS}, messages clipped), so the host
+     * keeps every round and not only the last one.
      */
     onStage?: (s: {
         stage: StageId;
@@ -119,6 +133,7 @@ export interface RunPipelineDeps {
         round?: number;
         findings?: number;
         output?: string;
+        items?: RoundFindingItem[];
     }) => void;
     /**
      * Optional, fail-soft durable journal seam (spec: 2026-06-29-go-resilience).

package/dist/extension/pipeline/orchestrator.js

@@ -35,6 +35,20 @@ import { makeRunVerify, makeWorkstreamCheck, parseChangedPaths, } from "./verify
 import { builderStage, fixerStage, orchestratorStage, partitionReaskStage, PARTITION_CALLER_LABEL, REQUIRED_LENSES, REVIEW_LENSES, reaskStage, reviewStage, selectStages, synthesizerFixStage, synthesizerStage, SYNTHESIZER_CALLER_LABEL, workstreamCallerLabel, } from "./stages.js";
 import { snapshotWorkspace as defaultSnapshotWorkspace, workspaceChanged } from "./workspace.js";
 import { DEFAULT_RESILIENCE_POLICY, MAX_CONCURRENCY, MAX_FANOUT_CONCURRENCY, MAX_FANOUT_CONCURRENCY_ULTRA, MAX_FIX_TURNS, MIN_TOOL_CALLS_FOR_HEALTH, TOOL_ERROR_FAIL_RATE, } from "./types.js";
+/** Each review round's findings as the host stores them: bounded, messages clipped. */
+export const MAX_ROUND_ITEMS = 20;
+const MAX_ROUND_ITEM_MESSAGE = 300;
+export function roundFindingItems(findings) {
+    return findings.slice(0, MAX_ROUND_ITEMS).map((finding) => ({
+        severity: finding.severity,
+        lens: finding.lens,
+        message: finding.message.length > MAX_ROUND_ITEM_MESSAGE
+            ? `${finding.message.slice(0, MAX_ROUND_ITEM_MESSAGE - 1)}…`
+            : finding.message,
+        ...(finding.file ? { file: finding.file } : {}),
+        ...(finding.line !== undefined ? { line: finding.line } : {}),
+    }));
+}
 /** Error thrown when a build stage fails; carries the partial run for the caller. */
 export class PipelineStageError extends Error {
     stageId;
@@ -1026,7 +1040,7 @@ export async function runPipeline(ticket, deps) {
             ...(degradedLenses ? { degradedLenses } : {}),
         });
         progress({ kind: "findings", round, total: findings.length, blocking: blockingCount(findings) });
-        stageEvent("review", "finish", { round, findings: findings.length });
+        stageEvent("review", "finish", { round, findings: findings.length, items: roundFindingItems(findings) });
         // Honest health check BEFORE trusting the verdict: an aborted or crashed
         // review must NOT be read as a clean review (empty output → [] findings →
         // would otherwise stop 'clean'). Mirrors the build half's isFailed guard.

package/dist/extension/pipeline/personas.d.ts

@@ -7,7 +7,8 @@
  *   - scout / planner: call `ask_yagni` before inferring a convention,
  *   - worker: call `record_decision` only for a consequential product-intent or architecture call,
  *   - reviewer (business-fit lens): call `review_business_match` and treat a
- *     conflict with a recorded decision as at least High.
+ *     conflict with a recorded decision as at least High, and a disagreement
+ *     with the plan the person approved as needs_person.
  *
  * The implement diamond adds two more roles: `orchestrator` (the read-only
  * partitioner, carrying the ```partition output contract `parsePartition` reads)

package/dist/extension/pipeline/personas.js

@@ -7,7 +7,8 @@
  *   - scout / planner: call `ask_yagni` before inferring a convention,
  *   - worker: call `record_decision` only for a consequential product-intent or architecture call,
  *   - reviewer (business-fit lens): call `review_business_match` and treat a
- *     conflict with a recorded decision as at least High.
+ *     conflict with a recorded decision as at least High, and a disagreement
+ *     with the plan the person approved as needs_person.
  *
  * The implement diamond adds two more roles: `orchestrator` (the read-only
  * partitioner, carrying the ```partition output contract `parsePartition` reads)
@@ -331,9 +332,9 @@ export const BLIND_PERSONA_BODIES = {
 };
 /** The lens-specific clause appended to the reviewer body, one per review angle. */
 const LENS_CLAUSES = {
-    correctness: "Lens: CORRECTNESS. Hunt bugs, broken logic, unhandled edge cases, race conditions, and incorrect error handling. A real defect that can ship is at least High.",
-    business_fit: "Lens: BUSINESS-FIT. This is the only-YAGNI lens. Call review_business_match and consult the decision corpus: does this change match the recorded decisions, conventions, and current priorities of this company? Right code doing the wrong thing is exactly the failure you exist to catch. A conflict with a recorded decision is at least High.",
-    does_it_hold: "Lens: DOES-IT-HOLD. Does the change actually accomplish the ticket, and does it build/test as far as read-only bash lets you verify? Missing tests for new behavior, or a change that does not do the task, is at least High.",
+    correctness: "Lens: CORRECTNESS. Hunt bugs, broken logic, unhandled edge cases, race conditions, and incorrect error handling. A defect that breaks behavior or data at runtime is at least High. Copy, naming, dead code, and style are Medium or Low.",
+    business_fit: "Lens: BUSINESS-FIT. This is the only-YAGNI lens. Call review_business_match and consult the decision corpus: does this change match the recorded decisions, conventions, and current priorities of this company? Right code doing the wrong thing is exactly the failure you exist to catch. A conflict with a recorded decision is at least High. A disagreement with the plan the person approved is needs_person: the person decided it, so only the person can change it.",
+    does_it_hold: "Lens: DOES-IT-HOLD. Does the change actually accomplish the ticket, and does it build/test as far as read-only bash lets you verify? A change that does not do the task, or new behavior with no test at all, is at least High. Something only a person can do (a named sign-off, an account, a credential, a step outside the repository) is needs_person, never High.",
 };
 /**
  * The machine-readable findings contract every reviewer must emit so the
@@ -346,7 +347,7 @@ End your review with a fenced block in EXACTLY this form, one line per finding:
 SEVERITY | file:line | message
 \`\`\`
 
-SEVERITY is one of: critical, high, medium, low. Use \`file:line\` when you can point to a location; otherwise give a short location or omit it. Put one finding per line and nothing else inside the block. If you found no problems, emit an empty \`\`\`findings block. Only critical and high findings block the change.
+SEVERITY is one of: critical, high, medium, low, needs_person. needs_person means only a person can do it (a named sign-off, an account, a credential, a call outside the code); it does not block, and the person sees it on the pull request. Use \`file:line\` when you can point to a location; otherwise give a short location or omit it. Put one finding per line and nothing else inside the block. If you found no problems, emit an empty \`\`\`findings block. Only critical and high findings block the change.
 
 Be terse. Do not narrate your process, restate the diff, or quote code back at length — spend your output on the findings themselves, at most a few short paragraphs before the block. Cap the block at the 12 most important findings, most severe first; a review cut off by its own length limit helps nobody.`;
 /**
@@ -357,7 +358,7 @@ Be terse. Do not narrate your process, restate the diff, or quote code back at l
  */
 const BLIND_LENS_CLAUSES = {
     correctness: LENS_CLAUSES.correctness,
-    business_fit: "Lens: BUSINESS-FIT. Does this change match the apparent product intent and the conventions visible in the code? Right code doing the wrong thing is exactly the failure you exist to catch. A clear mismatch is at least High.",
+    business_fit: "Lens: BUSINESS-FIT. Does this change match the apparent product intent and the conventions visible in the code? Right code doing the wrong thing is exactly the failure you exist to catch. A clear mismatch is at least High. A disagreement with the plan the person approved is needs_person.",
     does_it_hold: LENS_CLAUSES.does_it_hold,
 };
 /**

package/dist/extension/pipeline/types.d.ts

@@ -35,8 +35,13 @@ export type StageId = "map" | "plan" | "implement" | "review" | "fix";
 export type FeedStageId = StageId | "finish";
 /** The three adversarial review angles (spec §3). */
 export type ReviewLens = "correctness" | "business_fit" | "does_it_hold";
-/** Finding severities; only `critical` + `high` block the review→fix loop. */
-export type Severity = "critical" | "high" | "medium" | "low";
+/**
+ * Finding severities; only `critical` + `high` block the review→fix loop.
+ * `needs_person` is something only a person can do (a named sign-off, an
+ * account, a credential, a call outside the code): the reviewer assigns it,
+ * the fix stage never sees it, and it rides the pull request for the person.
+ */
+export type Severity = "critical" | "high" | "medium" | "low" | "needs_person";
 /**
  * One declarative stage. `model` + `tools` are the tuning surface; `taskTemplate`
  * carries `{ticket}` / `{previous}` placeholders the invocation builder fills.
@@ -496,8 +501,12 @@ export type PipelineProgress = {
     kind: "done";
     stopReason: StopReason;
 };
-/** Stop the review→fix loop after at most this many rounds (spec §7.3). */
-export declare const MAX_REVIEW_ROUNDS = 3;
+/**
+ * Stop the review→fix loop after at most this many rounds (spec §7.3). A third
+ * round never helped: every build that reached it ended with as many findings
+ * or more, and hit the cap anyway.
+ */
+export declare const MAX_REVIEW_ROUNDS = 2;
 /**
  * Hard cap on the implement diamond's fix turns (spec decision 7). Verification
  * may re-engage the builders that broke the tree, but only this many times: past

package/dist/extension/pipeline/types.js

@@ -6,8 +6,12 @@
  * scattered string literals. The stage list is a declarative contract (spec
  * §7.2) so future complexity-lanes select a subset, never a rewrite.
  */
-/** Stop the review→fix loop after at most this many rounds (spec §7.3). */
-export const MAX_REVIEW_ROUNDS = 3;
+/**
+ * Stop the review→fix loop after at most this many rounds (spec §7.3). A third
+ * round never helped: every build that reached it ended with as many findings
+ * or more, and hit the cap anyway.
+ */
+export const MAX_REVIEW_ROUNDS = 2;
 /**
  * Hard cap on the implement diamond's fix turns (spec decision 7). Verification
  * may re-engage the builders that broke the tree, but only this many times: past

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yagni-app/code-staging",
-  "version": "1.2.3-staging.1621.1",
+  "version": "1.2.3-staging.1627.1",
   "description": "YAGNI Code: a terminal coding agent that already knows your company. One YAGNI login routes the model and grounds the agent in your team's context.",
   "license": "SEE LICENSE IN LICENSE.md",
   "author": "YAGNI, Inc. <jack@yagni.app> (https://yagni.app)",
@@ -58,5 +58,5 @@
     "turndown": "^7.2.4",
     "typebox": "^1.3.15"
   },
-  "yagniSourceSha": "7d2396135b54ce5b4461bd57dcdb99292eb52115"
+  "yagniSourceSha": "a0b27bcdb66fa5dfeca35883d6786433cbb85e8e"
 }