review: dedupe lens discipline snippets into one staged disciplines file

.changeset/review-dispatch-tax.md

@@ -0,0 +1,5 @@
+---
+"review": patch
+---
+
+Dispatch-tax trim: dedupe the specialist-lens discipline snippets. The v1.4.0 re-run's cost premium vs the v1.3.1 baseline sat at dispatch time (uncached input +51-65%, cache writes +18-34% per run), partly because the shared discipline blocks (bounded investigation, untrusted input, lens-owned skills with quote-the-rule, the schema-rules trailer, the tri-state hunt contract) were stamped verbatim into every one of the eleven specialist-lens definitions and paid on every dispatch. The shared text now lives once, in a marker-delimited "REVIEW DISCIPLINES" section of review.md's main body; Step 1 stages it to `/tmp/gh-aw/review/disciplines.md` with one whole-line-anchored `sed` extraction of the engine-provided rendered prompt (`$GH_AW_PROMPT`), verified by grep with a byte-for-byte heredoc fallback, and each lens carries only a pointer plus its own domain notes (its investigation examples, review rules, hunts, and output JSON, all unchanged). Behavior-neutral by construction: the instruction content is unchanged, the three variants that had already drifted apart are unified on the fullest wording, and the label-shape reviewers (whose variants differ materially) keep their own copies. Pinned by `lib/disciplines.test.ts`: the section extracts cleanly under the exact sed range semantics, every lens points at the staged file and carries no residual copy, and the label-shape reviewers still carry theirs.

workflows/review/eval/corpus/live.ts

@@ -72,7 +72,19 @@ export type LiveDefectSpec = {
  */
 export type CaseLive = {
     prContext: LivePrContext;
-    /** Case-dir-relative path to the post-change tree (default `tree`). */
+    /**
+     * Case-dir-relative path to the post-change tree (default `tree`).
+     *
+     * Trees are deliberately minimal: validation requires only the changed
+     * files to exist. The staged copy of this tree is the sub-agent's whole
+     * world (its cwd, with no network), so the agent WILL often try to read
+     * an imported module or caller that is not there; that read returns an
+     * ordinary not-found tool error the agent tolerates and works around,
+     * not a run failure. The cost of a missing file is realism, not a crash:
+     * include the context files whose absence would change what a reviewer
+     * concludes (e.g. the decorator module a sibling handler imports), and
+     * nothing more.
+     */
     tree: string;
     /** Labeled defects a live run must catch. */
     mustCatchSpecs?: LiveDefectSpec[];

workflows/review/lib/disciplines.test.ts

@@ -0,0 +1,164 @@
+import {readFileSync} from "node:fs";
+import {join} from "node:path";
+
+import {describe, it, expect} from "vitest";
+
+import {SPECIALIST_LENSES} from "./router.ts";
+
+/**
+ * Dispatch-tax dedupe smoke tests.
+ *
+ * The measured cost premium of the v1.4.0 re-run sat at dispatch time
+ * (uncached input +51-65%, cache writes +18-34% per run), partly because
+ * discipline snippets (bounded investigation, quote-the-rule via lens-owned
+ * skills, the schema-rules trailer, ...) were stamped verbatim into every one
+ * of the eleven specialist-lens definitions and paid on every dispatch. The
+ * shared text now lives once, in the marker-delimited REVIEW DISCIPLINES
+ * section of review.md's main body, which Step 1 stages to
+ * `/tmp/gh-aw/review/disciplines.md` with a mechanical `sed` range extraction.
+ *
+ * These tests pin the contract: the section exists and extracts cleanly (the
+ * same line-range semantics the sed command uses), it carries every section
+ * the lens pointer names, each lens definition points at the staged file and
+ * no longer carries its own copy, and the label-shape reviewers (whose
+ * variants differ materially) still carry theirs.
+ */
+
+const reviewMd = readFileSync(join(__dirname, "..", "review.md"), "utf8");
+
+const BEGIN = "<!-- BEGIN REVIEW DISCIPLINES -->";
+const END = "<!-- END REVIEW DISCIPLINES -->";
+
+/**
+ * The sed range extraction from review.md Step 1, replicated line-for-line.
+ * The Step 1 command anchors both patterns to whole lines (`^...$`), so only
+ * the marker lines themselves open/close the range — never prose that mentions
+ * a marker, and never the sed command's own text.
+ */
+const sedRangeExtract = (text: string): string => {
+    const lines = text.split("\n");
+    const out: string[] = [];
+    let inRange = false;
+    for (const line of lines) {
+        if (!inRange && line === BEGIN) {
+            inRange = true;
+        }
+        if (inRange) {
+            out.push(line);
+            if (line === END) {
+                break;
+            }
+        }
+    }
+    return out.join("\n");
+};
+
+/** The main body (the orchestrator prompt): everything before the first agent. */
+const mainBody = reviewMd.slice(0, reviewMd.indexOf("\n## agent: `"));
+
+/** One lens's definition section, as the sub-agent extractor would cut it. */
+const lensSection = (lens: string): string => {
+    const start = reviewMd.indexOf(`## agent: \`${lens}\``);
+    expect(start).toBeGreaterThan(-1);
+    const rest = reviewMd.slice(start);
+    const next = rest.indexOf("\n## agent: `");
+    return next === -1 ? rest : rest.slice(0, next);
+};
+
+const DISCIPLINE_HEADINGS = [
+    "## Staged inputs",
+    "## Untrusted input",
+    "## Read every line",
+    "## Bounded investigation",
+    "## Lens-owned skills",
+    "## Out-of-lane handoff",
+    "## Structured finding schema and hunts",
+];
+
+describe("the shared disciplines section", () => {
+    it("lives in the main body (the staged prompt), not in an agent section", () => {
+        expect(mainBody).toContain(BEGIN);
+        expect(mainBody).toContain(END);
+    });
+
+    it("extracts cleanly with the Step 1 sed range semantics", () => {
+        const extracted = sedRangeExtract(reviewMd);
+        expect(extracted.startsWith(BEGIN)).toBe(true);
+        expect(extracted.endsWith(END)).toBe(true);
+        for (const heading of DISCIPLINE_HEADINGS) {
+            expect(extracted).toContain(`\n${heading}\n`);
+        }
+    });
+
+    it("keeps the quote-the-rule discipline verbatim", () => {
+        const extracted = sedRangeExtract(reviewMd);
+        expect(extracted).toContain(
+            "Flag a skill violation only when you can quote **both** the exact rule",
+        );
+        expect(extracted).toContain("no spirit-of-the-doc inference");
+    });
+
+    it("keeps the tri-state hunt contract", () => {
+        const extracted = sedRangeExtract(reviewMd);
+        for (const state of ["`found`", "`ran`", "`not-applicable`"]) {
+            expect(extracted).toContain(state);
+        }
+    });
+
+    it("is staged by Step 1 with the sed extraction and a fallback guard", () => {
+        expect(mainBody).toContain("$GH_AW_PROMPT");
+        expect(mainBody).toContain("/tmp/gh-aw/review/disciplines.md");
+        expect(mainBody).toContain(
+            "grep -q '## Structured finding schema and hunts'",
+        );
+    });
+});
+
+describe("each specialist lens definition", () => {
+    for (const lens of SPECIALIST_LENSES) {
+        it(`${lens}: points at the staged disciplines and carries no copy`, () => {
+            const section = lensSection(lens);
+            expect(section).toContain("**Shared disciplines first.**");
+            expect(section).toContain("/tmp/gh-aw/review/disciplines.md");
+            // The deduped blocks must be gone from the lens definition.
+            expect(section).not.toContain("**Bounded investigation.**");
+            expect(section).not.toContain("**Untrusted input.**");
+            expect(section).not.toContain("Read from disk:");
+            expect(section).not.toContain("Schema rules");
+            expect(section).not.toContain(
+                "**Hand off, never drop, an out-of-lane observation.**",
+            );
+            // The domain-specific content stays.
+            expect(section).toContain("### Review rules");
+            expect(section).toContain("### Incident-derived hunts (tri-state)");
+            expect(section).toContain("### Output");
+            expect(section).toContain(
+                "{{#runtime-import .github/aw/review/skills.md}}",
+            );
+            expect(section).toContain(`\`lens\` is exactly \`${lens}\``);
+            expect(section).toContain(
+                "Domain notes for §Bounded investigation",
+            );
+        });
+    }
+});
+
+describe("the label-shape reviewers still carry their own disciplines", () => {
+    // Their variants differ materially (CLI cap invocation with per-agent id
+    // semantics, `discussion` instead of `evidence_trace`), so they were
+    // deliberately left out of the dedupe.
+    for (const agent of [
+        "correctness-reviewer",
+        "skill-auditor",
+        "claim-validator",
+        "holistic",
+        "completeness",
+        "test-adequacy",
+        "first-principles",
+        "conventions",
+    ]) {
+        it(`${agent}: keeps its own bounded-investigation block`, () => {
+            expect(lensSection(agent)).toContain("**Bounded investigation.**");
+        });
+    }
+});