@pmelab/gtd 27.6.0 → 27.7.1

package/claude/hooks/wording.ts

@@ -2,6 +2,9 @@
 
 import type { Stop } from "../types"
 import type { Beat } from "./drive"
+import type { Shipped } from "./ship"
+import { JOB_NAME } from "./tools"
+import type { JobName } from "./tools"
 
 const step = (s: { label?: string; state?: string }) => s.label ?? s.state ?? "the next step"
 
@@ -53,3 +56,11 @@ export function pushText(stop: Stop, repo: string, url?: string) {
   if (stop.kind !== "gate") return at
   return url ? `${at}. Open: ${url}` : at
 }
+
+const capitalised = (s: string) => s[0]!.toUpperCase() + s.slice(1)
+
+export const jobHeadline = (job: JobName, outcome: Shipped) =>
+  `${outcome.ok ? "✔" : "✘"} ${capitalised(JOB_NAME[job])} ${outcome.ok ? "finished" : "failed"}`
+
+export const jobPushText = (job: JobName, outcome: Shipped, repo: string) =>
+  `${repo}: ${JOB_NAME[job]} ${outcome.ok ? "finished" : "failed"}. ${outcome.text}`.trim()

package/dist/chunks/{reserved-Df8GeSHw.mjs → reserved-UdSuHDUJ.mjs}

@@ -1234,13 +1234,13 @@ const fixRisks = (notes) => agentWithSkills("review.fix.risks.fixing", buildRevi
 	system: reviewerSystem(),
 	allowEmpty: true
 });
-const awaitReview = (base, message = buildReviewAwaitReviewMessage(base), acceptClean = false) => human("review.await-review", {
+const awaitReview = (base, message = buildReviewAwaitReviewMessage(base)) => human("review.await-review", {
 	message,
 	label: "Awaiting your review",
 	file: REVIEW,
 	mode: "review",
 	base,
-	...acceptClean ? { acceptClean } : {}
+	acceptClean: true
 });
 /** The round at commit `round` left no `.gtd/REVIEW.md`. */
 const reviewMissing = (round) => human("review.review-missing", {

package/dist/chunks/{tail-DmIfEAI4.mjs → tail-fj60rQ3y.mjs}

@@ -1,6 +1,6 @@
 import { read, refuse, removeScript, run, scope, showScript, vars } from "../flows.mjs";
 import { fixFindings, lensesOf, qualityLap as qualityLap$1, reviewToHandOff } from "../workflow-kit.mjs";
-import { C as fixRisks, D as reviewing, E as reviewQuality, G as handOffAwaitReviewMessage, S as fixQuality, T as reviewMissing, f as guarded, g as QUALITY, p as healthy, u as escalation, v as REVIEW, x as fixCheck, y as awaitReview } from "./reserved-Df8GeSHw.mjs";
+import { C as fixRisks, D as reviewing, E as reviewQuality, G as handOffAwaitReviewMessage, S as fixQuality, T as reviewMissing, f as guarded, g as QUALITY, p as healthy, u as escalation, v as REVIEW, x as fixCheck, y as awaitReview } from "./reserved-UdSuHDUJ.mjs";
 //#region src/workflows/shared/tail.ts
 /** The reviewer's own `.gtd/REVIEW.md`, kept beside the human's notes so `fix` can tell them apart. */
 const REVIEW_ORIG = ".gtd/REVIEW.orig.md";

package/dist/gtd.bundle.mjs

@@ -41695,7 +41695,7 @@ const FLAGS = [
 		help: [
 			"(gtd land, with --cost) tag that cost's model",
 			"(gtd judge run, llm answerer) the claude model; omitted:",
-			"GTD_JUDGE_MODEL, then judge.json, then haiku"
+			"GTD_JUDGE_MODEL, then judge.json, then sonnet"
 		]
 	},
 	{
@@ -42088,7 +42088,7 @@ const COMMAND_ROWS = [
 			"--answers <path> or GTD_JUDGE_ANSWERS; questions the file",
 			"does not cover are left out. --provider jev asks TypeSafe",
 			"(TYPESAFE_API_KEY). --provider llm asks the `claude` CLI",
-			"on PATH (Claude Code only), model haiku unless --model.",
+			"on PATH (Claude Code only), model sonnet unless --model.",
 			"No --provider: GTD_JUDGE_PROVIDER, then ~/.config/gtd/",
 			"judge.json, else jev when TYPESAFE_API_KEY is set and no",
 			"model is given, else llm; --model likewise falls back to",
@@ -58805,7 +58805,7 @@ var DEFAULT_MODEL, TIMEOUT_MS, SYSTEM_PROMPT, fail$1, unreadable, levelsOf, opti
 var init_llm = __esmMin((() => {
 	init_criteria();
 	init_io();
-	DEFAULT_MODEL = "haiku";
+	DEFAULT_MODEL = "sonnet";
 	TIMEOUT_MS = 18e4;
 	SYSTEM_PROMPT = "You are a strict, impartial judge. Answer each question from the supplied evidence alone.\nReply only with the structured answer; never use tools or ask questions.";
 	fail$1 = (message) => {

package/dist/testkit.mjs

@@ -10825,7 +10825,7 @@ const FLAGS = [
 		help: [
 			"(gtd land, with --cost) tag that cost's model",
 			"(gtd judge run, llm answerer) the claude model; omitted:",
-			"GTD_JUDGE_MODEL, then judge.json, then haiku"
+			"GTD_JUDGE_MODEL, then judge.json, then sonnet"
 		]
 	},
 	{
@@ -11218,7 +11218,7 @@ const COMMAND_ROWS = [
 			"--answers <path> or GTD_JUDGE_ANSWERS; questions the file",
 			"does not cover are left out. --provider jev asks TypeSafe",
 			"(TYPESAFE_API_KEY). --provider llm asks the `claude` CLI",
-			"on PATH (Claude Code only), model haiku unless --model.",
+			"on PATH (Claude Code only), model sonnet unless --model.",
 			"No --provider: GTD_JUDGE_PROVIDER, then ~/.config/gtd/",
 			"judge.json, else jev when TYPESAFE_API_KEY is set and no",
 			"model is given, else llm; --model likewise falls back to",
@@ -13076,18 +13076,22 @@ const recognizeGtdUncheck = (repo, block) => {
 * definition) keeps the `@inmem` tier green.
 */
 const recognizeDidNotRunComment = (block) => block === "# gtd emitted this and did NOT run it — pipe it into `sh` to land the turn" ? { kind: "noop" } : void 0;
-/** `combinedScript`'s optional half. ALWAYS reports `noop` whatever the inner result, mirroring bash's `( … ) || <warning>`, where the subshell's exit status is swallowed and must never fail the outer script. */
+/** `combinedScript`'s optional half. Reports `noop` whatever the inner result, mirroring bash's `( … ) || <warning>`. Exception: under `git: false` any unrecognized inner block fails the outer script (git or not), so the in-memory twin knowingly diverges from `sh` rather than run git unseen. */
 const PRESENTATION_FAILURE_TEXT = `${extractQuotedTokens(PRESENTATION_FAILURE_WARNING)[0].replace(/\\n$/, "")}\n`;
 const PRESENTATION_SUBSHELL_PREFIX = `${PRESENTATION_ONLY_COMMENT}\n(\n`;
 const PRESENTATION_SUBSHELL_SUFFIX = `\n) || ${PRESENTATION_FAILURE_WARNING}`;
-const recognizePresentationSubshell = (repo, commands, block) => {
+const recognizePresentationSubshell = (repo, commands, block, git) => {
 	if (!block.startsWith(PRESENTATION_SUBSHELL_PREFIX) || !block.endsWith(PRESENTATION_SUBSHELL_SUFFIX)) return;
 	const inner = block.slice(PRESENTATION_SUBSHELL_PREFIX.length, block.length - PRESENTATION_SUBSHELL_SUFFIX.length);
-	const result = applyEmittedScript(repo, commands, inner);
+	const result = applyEmittedScript(repo, commands, inner, { git });
 	if (result.ok) return {
 		kind: "noop",
 		output: result.output
 	};
+	if (!git && result.error?.startsWith(UNRECOGNIZED_PREFIX)) return {
+		kind: "failed",
+		error: `${result.output}${result.error}`
+	};
 	return {
 		kind: "noop",
 		output: `${result.output}${PRESENTATION_FAILURE_TEXT}`
@@ -13194,17 +13198,17 @@ const recognizeWriteLines = (repo, block) => {
 	return { kind: "applied" };
 };
 /** Every recognizer `applyEmittedScript` tries, first match wins — `recognizeScriptedCommand` stays last since it's the exact-hit fallback for anything the rest don't claim. */
-const recognizersFor = (repo, commands) => [
+const recognizersFor = (repo, commands, git) => [
 	(block) => recognizeSetFlags(block),
 	(block) => recognizeDidNotRunComment(block),
 	(block) => recognizePrecondition(repo, block),
 	(block) => recognizeFileExistsGuard(repo, block),
 	(block) => recognizeBinaryGuard(commands, block),
-	(block) => recognizeGitBuilders(repo, block),
+	(block) => git ? recognizeGitBuilders(repo, block) : void 0,
 	(block) => recognizeDeferNotes(repo, block),
 	(block) => recognizeWriteLines(repo, block),
 	(block) => recognizeFailurePromptWrapper(repo, commands, block),
-	(block) => recognizePresentationSubshell(repo, commands, block),
+	(block) => recognizePresentationSubshell(repo, commands, block, git),
 	(block) => recognizeModeContradictionSkipNotice(block),
 	(block) => recognizeModeContradictionCheck(repo, commands, block),
 	(block) => recognizeGtdCheck(repo, block),
@@ -13227,7 +13231,7 @@ const recognizeBlock = (recognizers, block) => {
 * non-zero scripted-command exit — exactly like `set -euo pipefail` stops a
 * real script. No block after the failing one is applied.
 */
-const applyEmittedScript = (repo, commands, script) => {
+const applyEmittedScript = (repo, commands, script, { git = true } = {}) => {
 	const flow = applyFlowScript(repo, commands, script);
 	if (flow !== void 0) return flow.ok ? {
 		ok: true,
@@ -13239,10 +13243,10 @@ const applyEmittedScript = (repo, commands, script) => {
 		output: flow.output,
 		error: flow.error ?? ""
 	};
-	return script.startsWith("#!") ? unrecognized(script) : applyBlocks(repo, commands, script);
+	return script.startsWith("#!") ? unrecognized(script) : applyBlocks(repo, commands, script, git);
 };
-const applyBlocks = (repo, commands, script) => {
-	const recognizers = recognizersFor(repo, commands);
+const applyBlocks = (repo, commands, script, git) => {
+	const recognizers = recognizersFor(repo, commands, git);
 	let output = "";
 	for (const block of splitBlocks(script)) {
 		const outcome = recognizeBlock(recognizers, block);
@@ -13266,11 +13270,12 @@ const failedAt = (outcome, output) => {
 		error: outcome.error
 	};
 };
+const UNRECOGNIZED_PREFIX = "unrecognized script block: ";
 const unrecognized = (block, output = "") => ({
 	ok: false,
 	status: 1,
 	output,
-	error: `unrecognized script block: ${block.split("\n")[0]}`
+	error: `${UNRECOGNIZED_PREFIX}${block.split("\n")[0]}`
 });
 //#endregion
 //#region src/testkit/TestRepo.ts
@@ -13322,10 +13327,17 @@ const LINES = [
 		repo.writeFile(dest, content);
 	}]
 ];
-const runLine = (repo, line) => {
-	const known = LINES.find(([pattern]) => pattern.test(line));
-	if (known === void 0) throw new Error(`testkit: cannot run \`${line}\` in memory — give the step an actor instead`);
-	known[1](repo, words(line));
+const knownLine = (line) => LINES.find(([pattern]) => pattern.test(line));
+const cannotRun = (error) => /* @__PURE__ */ new Error(`testkit: cannot run script in memory — give the step an actor instead\n${error}`);
+const runLines = (repo, script) => {
+	const lines = script.split("\n").map((line) => line.trim());
+	const known = lines.map(knownLine);
+	if (known.every(Boolean)) {
+		known.forEach((k, i) => k[1](repo, words(lines[i])));
+		return;
+	}
+	const applied = applyEmittedScript(repo, /* @__PURE__ */ new Map(), script, { git: false });
+	if (!applied.ok) throw cannotRun(applied.error ?? "");
 };
 const failed = (what, result) => /* @__PURE__ */ new Error(`${what} exited ${result.exitCode}\n${result.stderr}${result.stdout}`);
 const testRepo = (options) => {
@@ -13407,7 +13419,7 @@ const testRepo = (options) => {
 		show: (ref, path) => repo.fileAtRef(ref, path) ?? void 0,
 		runScript: (script) => {
 			const flow = applyFlowScript(repo, /* @__PURE__ */ new Map(), script);
-			if (flow === void 0) for (const line of script.split("\n")) runLine(repo, line.trim());
+			if (flow === void 0) runLines(repo, script);
 			else if (!flow.ok) throw new Error(`testkit: ${flow.error ?? ""}`);
 		},
 		gtd,

package/dist/workflows/cleanup.mjs

@@ -1,6 +1,6 @@
 import { agent, changes, glob, read, refuse, scope, start, vars } from "../flows.mjs";
-import { a as flagShipped, b as coder, c as envDefaults, i as skills, n as summary, r as access, s as defaults, t as steering } from "../chunks/reserved-Df8GeSHw.mjs";
-import { i as refusePendingHandOff, r as handOffTail } from "../chunks/tail-DmIfEAI4.mjs";
+import { a as flagShipped, b as coder, c as envDefaults, i as skills, n as summary, r as access, s as defaults, t as steering } from "../chunks/reserved-UdSuHDUJ.mjs";
+import { i as refusePendingHandOff, r as handOffTail } from "../chunks/tail-fj60rQ3y.mjs";
 import { i as requireFlag, r as registryEntry, t as flagReferences } from "../chunks/registry-CP3YuV9j.mjs";
 //#region src/workflows/cleanup/cleanup.ts
 const removingPrompt = (name, registry) => `Clean up the release flag "${name}", which is default-on: delete its entry in ${registry}, and remove every \`GTD_FLAG_${name.toUpperCase()}\` check from the shipped sources. Keep the flag-on path; delete code only the off path reached, including tests that pin the off path.`;

package/dist/workflows/feature.mjs

@@ -1,7 +1,7 @@
 import { answered, changes, changesSince, glob, hasThreadFor, head, human, judge, judgeDiffShape, kept, mark, numeric, quote, read, refuse, removeScript, requireAnswers, requireProgress, requireReplies, requireRevert, requireThreadsClosed, restoreScript, revertScript, run, scope, seamGate, start, vars } from "../flows.mjs";
 import { review as review$1 } from "../workflow-kit.mjs";
-import { $ as refactorFixPrompt, A as architectureAuthorPrompt, B as designGateAnswerMessage, C as fixRisks, D as reviewing, F as buildReviewCollectingPrompt, H as designTriagePrompt, I as buildReviewFixNitsPrompt, J as implementFixPrompt, K as idleMessage, L as buildReviewTriageMessage, M as architectureGateAnswerMessage, N as buildReviewAnswerQuestionsPrompt, O as agentWithSkills, P as buildReviewAwaitAfterDeferMessage, Q as refactorBuildingPrompt, R as builderSystem, T as reviewMissing, V as designSystem, X as packagesItemBuildingPrompt, Y as naivePassPrompt, Z as packagesItemFixSuitePrompt, _ as REQUIREMENTS, at as seamByHandMessage, b as coder, c as envDefaults, ct as unwindFailedMessage, d as gate, et as reviewEditNotesCapture, h as FEEDBACK, i as skills, it as reviewerSystem, j as architectureDecomposePrompt, k as architectSystem, lt as unwindFailure, m as ARCHITECTURE, n as summary, o as shipped, ot as seamSteerMessage, p as healthy, q as implementBuildingPrompt, r as access, rt as reviewNotesCapture, s as defaults, st as startGateBlockedMessage, t as steering, tt as reviewEditsCapture, v as REVIEW, w as planner, x as fixCheck, y as awaitReview, z as deferNotesScript } from "../chunks/reserved-Df8GeSHw.mjs";
-import { a as tail, i as refusePendingHandOff, n as closeReview } from "../chunks/tail-DmIfEAI4.mjs";
+import { $ as refactorFixPrompt, A as architectureAuthorPrompt, B as designGateAnswerMessage, C as fixRisks, D as reviewing, F as buildReviewCollectingPrompt, H as designTriagePrompt, I as buildReviewFixNitsPrompt, J as implementFixPrompt, K as idleMessage, L as buildReviewTriageMessage, M as architectureGateAnswerMessage, N as buildReviewAnswerQuestionsPrompt, O as agentWithSkills, P as buildReviewAwaitAfterDeferMessage, Q as refactorBuildingPrompt, R as builderSystem, T as reviewMissing, V as designSystem, X as packagesItemBuildingPrompt, Y as naivePassPrompt, Z as packagesItemFixSuitePrompt, _ as REQUIREMENTS, at as seamByHandMessage, b as coder, c as envDefaults, ct as unwindFailedMessage, d as gate, et as reviewEditNotesCapture, h as FEEDBACK, i as skills, it as reviewerSystem, j as architectureDecomposePrompt, k as architectSystem, lt as unwindFailure, m as ARCHITECTURE, n as summary, o as shipped, ot as seamSteerMessage, p as healthy, q as implementBuildingPrompt, r as access, rt as reviewNotesCapture, s as defaults, st as startGateBlockedMessage, t as steering, tt as reviewEditsCapture, v as REVIEW, w as planner, x as fixCheck, y as awaitReview, z as deferNotesScript } from "../chunks/reserved-UdSuHDUJ.mjs";
+import { a as tail, i as refusePendingHandOff, n as closeReview } from "../chunks/tail-fj60rQ3y.mjs";
 //#region src/workflows/feature/tier.ts
 /** GTD_FLAG_NAIVE_PASS, GTD_FLAG_RISK_TIER: both pinned at process start, so a mid-process flip never moves a process. */
 const riskTierOn = () => vars.flag_naive_pass === "1" && vars.flag_risk_tier === "1";
@@ -355,7 +355,7 @@ const review = (base, escalations = { rounds: 0 }) => {
 		awaitReview: (b) => {
 			const notes = deferred;
 			deferred = [];
-			return notes.length > 0 ? awaitReview(b, buildReviewAwaitAfterDeferMessage(b, notes), true) : awaitReview(b);
+			return notes.length > 0 ? awaitReview(b, buildReviewAwaitAfterDeferMessage(b, notes)) : awaitReview(b);
 		},
 		reviewMissing,
 		collecting,

package/dist/workflows/fix.mjs

@@ -1,7 +1,7 @@
 import { read, refuse, removeScript, run, scope, start } from "../flows.mjs";
 import { d as reviewNotes, f as reviewRisks } from "../chunks/steering-DVILanVf.mjs";
-import { O as agentWithSkills, U as finisherSystem, W as fixReviewNotesPrompt, b as coder, c as envDefaults, i as skills, l as baseline, n as summary, o as shipped, r as access, s as defaults, t as steering, v as REVIEW } from "../chunks/reserved-Df8GeSHw.mjs";
-import { r as handOffTail, t as REVIEW_ORIG } from "../chunks/tail-DmIfEAI4.mjs";
+import { O as agentWithSkills, U as finisherSystem, W as fixReviewNotesPrompt, b as coder, c as envDefaults, i as skills, l as baseline, n as summary, o as shipped, r as access, s as defaults, t as steering, v as REVIEW } from "../chunks/reserved-UdSuHDUJ.mjs";
+import { r as handOffTail, t as REVIEW_ORIG } from "../chunks/tail-fj60rQ3y.mjs";
 //#region src/workflows/fix/fix.ts
 /**
 * Fix every note a human left in a handed-off `.gtd/REVIEW.md` and every

package/dist/workflows/review.mjs

@@ -1,6 +1,6 @@
 import { start } from "../flows.mjs";
-import { c as envDefaults, d as gate, i as skills, n as summary, nt as reviewGateBlockedMessage, o as shipped, r as access, s as defaults, t as steering } from "../chunks/reserved-Df8GeSHw.mjs";
-import { i as refusePendingHandOff, r as handOffTail } from "../chunks/tail-DmIfEAI4.mjs";
+import { c as envDefaults, d as gate, i as skills, n as summary, nt as reviewGateBlockedMessage, o as shipped, r as access, s as defaults, t as steering } from "../chunks/reserved-UdSuHDUJ.mjs";
+import { i as refusePendingHandOff, r as handOffTail } from "../chunks/tail-fj60rQ3y.mjs";
 //#region src/workflows/review/review.ts
 /**
 * Pure review of everything since `reviewBase`: a green gate, the quality

package/dist/workflows/rollout.mjs

@@ -1,5 +1,5 @@
 import { agent, changesSince, head, read, refuse, scope, vars } from "../flows.mjs";
-import { a as flagShipped, b as coder, c as envDefaults, d as gate, i as skills, n as summary, r as access, s as defaults, t as steering } from "../chunks/reserved-Df8GeSHw.mjs";
+import { a as flagShipped, b as coder, c as envDefaults, d as gate, i as skills, n as summary, r as access, s as defaults, t as steering } from "../chunks/reserved-UdSuHDUJ.mjs";
 import { i as requireFlag, n as flipProblems } from "../chunks/registry-CP3YuV9j.mjs";
 //#region src/workflows/rollout/rollout.ts
 const flippingPrompt = (name, registry) => `Roll out the release flag "${name}": in ${registry}, set \`defaultOn: true\` on its entry. Change nothing else — no other entry, no other field, no other file, no comments.`;

package/dist/workflows.mjs

@@ -17995,13 +17995,13 @@ const fixRisks = (notes) => agentWithSkills("review.fix.risks.fixing", buildRevi
 	system: reviewerSystem(),
 	allowEmpty: true
 });
-const awaitReview = (base, message = buildReviewAwaitReviewMessage(base), acceptClean = false) => human("review.await-review", {
+const awaitReview = (base, message = buildReviewAwaitReviewMessage(base)) => human("review.await-review", {
 	message,
 	label: "Awaiting your review",
 	file: REVIEW,
 	mode: "review",
 	base,
-	...acceptClean ? { acceptClean } : {}
+	acceptClean: true
 });
 /** The round at commit `round` left no `.gtd/REVIEW.md`. */
 const reviewMissing = (round) => human("review.review-missing", {
@@ -37911,7 +37911,7 @@ const review$1 = (base, escalations = { rounds: 0 }) => {
 		awaitReview: (b) => {
 			const notes = deferred;
 			deferred = [];
-			return notes.length > 0 ? awaitReview(b, buildReviewAwaitAfterDeferMessage(b, notes), true) : awaitReview(b);
+			return notes.length > 0 ? awaitReview(b, buildReviewAwaitAfterDeferMessage(b, notes)) : awaitReview(b);
 		},
 		reviewMissing,
 		collecting,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pmelab/gtd",
-  "version": "27.6.0",
+  "version": "27.7.1",
   "private": false,
   "description": "Git-aware CLI that emits the next prompt for an autonomous coding agent based on the current repository state",
   "bin": {

package/src/cli/Cli.ts

@@ -241,7 +241,7 @@ const FLAGS: readonly FlagRow[] = [
     help: [
       "(gtd land, with --cost) tag that cost's model",
       "(gtd judge run, llm answerer) the claude model; omitted:",
-      "GTD_JUDGE_MODEL, then judge.json, then haiku",
+      "GTD_JUDGE_MODEL, then judge.json, then sonnet",
     ],
   },
   {
@@ -673,7 +673,7 @@ const COMMAND_ROWS: readonly CommandRow[] = [
       "--answers <path> or GTD_JUDGE_ANSWERS; questions the file",
       "does not cover are left out. --provider jev asks TypeSafe",
       "(TYPESAFE_API_KEY). --provider llm asks the `claude` CLI",
-      "on PATH (Claude Code only), model haiku unless --model.",
+      "on PATH (Claude Code only), model sonnet unless --model.",
       "No --provider: GTD_JUDGE_PROVIDER, then ~/.config/gtd/",
       "judge.json, else jev when TYPESAFE_API_KEY is set and no",
       "model is given, else llm; --model likewise falls back to",

package/src/testkit/EmittedScriptRecognizer.test.ts

@@ -630,6 +630,18 @@ describe("applyEmittedScript — Emit.ts's combinedScript (the whole stdout of a
     expect(repo.lastCommitMessage()).toBe("gtd(human): idle")
   })
 
+  it("fails an unrecognized optional block under git: false", () => {
+    const repo = new InMemRepo()
+    const result = applyEmittedScript(
+      repo,
+      NO_COMMANDS,
+      combinedScript("set -eu", "totally-unrecognized-command"),
+      { git: false },
+    )
+    expect(result.ok).toBe(false)
+    expect(result.error).toContain("totally-unrecognized-command")
+  })
+
   it("still fails loudly on an unrecognized block outside the presentation-only subshell", () => {
     const repo = new InMemRepo()
     const script = `${combinedScript("totally-unrecognized-required-command", "")}`
@@ -691,3 +703,32 @@ describe("applyEmittedScript — commitClean", () => {
     expect(snapshot(repo)).toEqual(before)
   })
 })
+
+describe("applyEmittedScript — git: false reaches through the presentation subshell", () => {
+  it("refuses a git builder wrapped in combinedScript and changes nothing", () => {
+    const repo = new InMemRepo()
+    repo.writeFile("a.txt", "1")
+    const before = snapshot(repo)
+    const result = applyEmittedScript(
+      repo,
+      NO_COMMANDS,
+      combinedScript("set -eu", commitAll("chore: sneaky")),
+      { git: false },
+    )
+    expect(result.ok).toBe(false)
+    expect(result.error).toMatch(/unrecognized script block/)
+    expect(snapshot(repo)).toEqual(before)
+  })
+
+  it("still applies a non-git write-lines optional half", () => {
+    const repo = new InMemRepo()
+    const result = applyEmittedScript(
+      repo,
+      NO_COMMANDS,
+      combinedScript("set -eu", "mkdir -p d\nprintf '%s\\n' 'x' > 'd/a.txt'"),
+      { git: false },
+    )
+    expect(result).toMatchObject({ ok: true })
+    expect(repo.readFile("d/a.txt")).toBe("x\n")
+  })
+})

package/src/testkit/EmittedScriptRecognizer.ts

@@ -563,7 +563,7 @@ const recognizeGtdUncheck = (repo: InMemRepo, block: string): BlockOutcome | und
 const recognizeDidNotRunComment = (block: string): BlockOutcome | undefined =>
   block === DID_NOT_RUN_COMMENT ? { kind: "noop" } : undefined
 
-/** `combinedScript`'s optional half. ALWAYS reports `noop` whatever the inner result, mirroring bash's `( … ) || <warning>`, where the subshell's exit status is swallowed and must never fail the outer script. */
+/** `combinedScript`'s optional half. Reports `noop` whatever the inner result, mirroring bash's `( … ) || <warning>`. Exception: under `git: false` any unrecognized inner block fails the outer script (git or not), so the in-memory twin knowingly diverges from `sh` rather than run git unseen. */
 const PRESENTATION_FAILURE_TEXT = `${extractQuotedTokens(PRESENTATION_FAILURE_WARNING)[0]!.replace(/\\n$/, "")}\n`
 const PRESENTATION_SUBSHELL_PREFIX = `${PRESENTATION_ONLY_COMMENT}\n(\n`
 const PRESENTATION_SUBSHELL_SUFFIX = `\n) || ${PRESENTATION_FAILURE_WARNING}`
@@ -572,6 +572,7 @@ const recognizePresentationSubshell = (
   repo: InMemRepo,
   commands: ReadonlyMap<string, ScriptedCommand>,
   block: string,
+  git: boolean,
 ): BlockOutcome | undefined => {
   if (
     !block.startsWith(PRESENTATION_SUBSHELL_PREFIX) ||
@@ -583,8 +584,11 @@ const recognizePresentationSubshell = (
     PRESENTATION_SUBSHELL_PREFIX.length,
     block.length - PRESENTATION_SUBSHELL_SUFFIX.length,
   )
-  const result = applyEmittedScript(repo, commands, inner)
+  const result = applyEmittedScript(repo, commands, inner, { git })
   if (result.ok) return { kind: "noop", output: result.output }
+  if (!git && result.error?.startsWith(UNRECOGNIZED_PREFIX)) {
+    return { kind: "failed", error: `${result.output}${result.error}` }
+  }
   return { kind: "noop", output: `${result.output}${PRESENTATION_FAILURE_TEXT}` }
 }
 
@@ -703,17 +707,18 @@ const recognizeWriteLines = (repo: InMemRepo, block: string): BlockOutcome | und
 const recognizersFor = (
   repo: InMemRepo,
   commands: ReadonlyMap<string, ScriptedCommand>,
+  git: boolean,
 ): ReadonlyArray<(block: string) => BlockOutcome | undefined> => [
   (block) => recognizeSetFlags(block),
   (block) => recognizeDidNotRunComment(block),
   (block) => recognizePrecondition(repo, block),
   (block) => recognizeFileExistsGuard(repo, block),
   (block) => recognizeBinaryGuard(commands, block),
-  (block) => recognizeGitBuilders(repo, block),
+  (block) => (git ? recognizeGitBuilders(repo, block) : undefined),
   (block) => recognizeDeferNotes(repo, block),
   (block) => recognizeWriteLines(repo, block),
   (block) => recognizeFailurePromptWrapper(repo, commands, block),
-  (block) => recognizePresentationSubshell(repo, commands, block),
+  (block) => recognizePresentationSubshell(repo, commands, block, git),
   (block) => recognizeModeContradictionSkipNotice(block),
   (block) => recognizeModeContradictionCheck(repo, commands, block),
   (block) => recognizeGtdCheck(repo, block),
@@ -746,6 +751,7 @@ export const applyEmittedScript = (
   repo: InMemRepo,
   commands: ReadonlyMap<string, ScriptedCommand>,
   script: string,
+  { git = true }: { readonly git?: boolean } = {},
 ): AppliedScriptResult => {
   const flow = applyFlowScript(repo, commands, script)
   if (flow !== undefined) {
@@ -753,15 +759,16 @@ export const applyEmittedScript = (
       ? { ok: true, status: 0, output: flow.output }
       : { ok: false, status: 1, output: flow.output, error: flow.error ?? "" }
   }
-  return script.startsWith("#!") ? unrecognized(script) : applyBlocks(repo, commands, script)
+  return script.startsWith("#!") ? unrecognized(script) : applyBlocks(repo, commands, script, git)
 }
 
 const applyBlocks = (
   repo: InMemRepo,
   commands: ReadonlyMap<string, ScriptedCommand>,
   script: string,
+  git: boolean,
 ): AppliedScriptResult => {
-  const recognizers = recognizersFor(repo, commands)
+  const recognizers = recognizersFor(repo, commands, git)
   let output = ""
   for (const block of splitBlocks(script)) {
     const outcome = recognizeBlock(recognizers, block)
@@ -786,9 +793,11 @@ const failedAt = (
   }
 }
 
+const UNRECOGNIZED_PREFIX = "unrecognized script block: "
+
 const unrecognized = (block: string, output = ""): AppliedScriptResult => ({
   ok: false,
   status: 1,
   output,
-  error: `unrecognized script block: ${block.split("\n")[0]}`,
+  error: `${UNRECOGNIZED_PREFIX}${block.split("\n")[0]}`,
 })

package/src/testkit/TestRepo.test.ts

@@ -1,4 +1,17 @@
 import { describe, expect, it } from "vitest"
+import {
+  commitAll,
+  combinedScript,
+  commitAsIs,
+  commitClean,
+  deleteRef,
+  discardPending,
+  hardResetTo,
+  keepTurn,
+  mixedResetTo,
+  softResetTo,
+  updateRef,
+} from "../emit/index.js"
 import { agent, human, moveScript, read, removeScript, run } from "../flows/index.js"
 import { runsItsScript, testRepo } from "./TestRepo.js"
 
@@ -93,7 +106,72 @@ describe("testRepo", () => {
 
   it("refuses to fake a script it cannot run", () => {
     const repo = testRepo({ workflow: { default: tiny } })
-    expect(() => repo.runScript("npm test\n")).toThrow(/cannot run `npm test` in memory/)
+    expect(() => repo.runScript("npm test\n")).toThrow(/cannot run script in memory[\s\S]*npm test/)
+  })
+
+  it("carries the recognizer's error when a git-free fallback fails", () => {
+    const repo = testRepo({ workflow: { default: tiny } })
+    expect(() => repo.runScript("printf 'x' > 'y'\n")).toThrow(
+      "testkit: cannot run script in memory — give the step an actor instead\nunrecognized script block: printf 'x' > 'y'",
+    )
+  })
+
+  it("names the block that failed, not an earlier line that ran", () => {
+    const repo = testRepo({ workflow: { default: tiny } })
+    let message = ""
+    try {
+      repo.runScript("mkdir -p d\nprintf '%s\\n' 'x' > 'd/a.txt'\n\nnpm test\n")
+    } catch (e) {
+      message = (e as Error).message
+    }
+    expect(message).toContain("unrecognized script block: npm test")
+    expect(message).not.toContain("printf")
+  })
+
+  it.each([
+    ["commitAll", commitAll("chore: sneaky")],
+    ["commitClean", commitClean("chore: sneaky")],
+    ["commitAsIs", commitAsIs("chore: sneaky")],
+    ["keepTurn", keepTurn("refs/x", "chore: sneaky")],
+    ["discardPending", discardPending()],
+    ["softResetTo", softResetTo("HEAD~1")],
+    ["mixedResetTo", mixedResetTo("HEAD~1")],
+    ["hardResetTo", hardResetTo("HEAD~1")],
+    ["updateRef", updateRef("refs/x", "HEAD")],
+    ["deleteRef", deleteRef("refs/x")],
+  ])("refuses a %s script instead of mutating git in memory", (_, script) => {
+    const repo = testRepo({ workflow: { default: tiny } })
+    repo.write("a.txt", "dirty\n")
+    const before = repo.subjects()
+    expect(() => repo.runScript(script)).toThrow(/unrecognized script block/)
+    expect(repo.subjects()).toEqual(before)
+  })
+
+  it.each([
+    ["commitAll", commitAll("chore: sneaky")],
+    ["commitClean", commitClean("chore: sneaky")],
+    ["commitAsIs", commitAsIs("chore: sneaky")],
+    ["keepTurn", keepTurn("refs/x", "chore: sneaky")],
+    ["discardPending", discardPending()],
+    ["softResetTo", softResetTo("HEAD~1")],
+    ["mixedResetTo", mixedResetTo("HEAD~1")],
+    ["hardResetTo", hardResetTo("HEAD~1")],
+    ["updateRef", updateRef("refs/x", "HEAD")],
+    ["deleteRef", deleteRef("refs/x")],
+  ])("refuses a wrapped %s script instead of mutating git in memory", (_, script) => {
+    const repo = testRepo({ workflow: { default: tiny } })
+    repo.write("a.txt", "dirty\n")
+    const before = repo.subjects()
+    expect(() => repo.runScript(combinedScript("set -eu", script))).toThrow(
+      /cannot run script in memory[\s\S]*unrecognized script block/,
+    )
+    expect(repo.subjects()).toEqual(before)
+  })
+
+  it("applies a write block `LINES` lacks through the git-free recognizers", () => {
+    const repo = testRepo({ workflow: { default: tiny } })
+    repo.runScript("mkdir -p d\nprintf '%s\\n' 'echo $(git rev-parse HEAD)' > 'd/a.txt'\n")
+    expect(repo.read("d/a.txt")).toBe("echo $(git rev-parse HEAD)\n")
   })
 
   it("reads a file as committed", () => {

package/src/testkit/TestRepo.ts

@@ -150,12 +150,23 @@ const LINES: ReadonlyArray<readonly [RegExp, (repo: InMemRepo, words: string[])
   ],
 ]
 
-const runLine = (repo: InMemRepo, line: string): void => {
-  const known = LINES.find(([pattern]) => pattern.test(line))
-  if (known === undefined) {
-    throw new Error(`testkit: cannot run \`${line}\` in memory — give the step an actor instead`)
+const knownLine = (line: string) => LINES.find(([pattern]) => pattern.test(line))
+
+const cannotRun = (error: string): Error =>
+  new Error(`testkit: cannot run script in memory — give the step an actor instead\n${error}`)
+
+// A script with a line `LINES` lacks goes to the contract-tested recognizers
+// whole, minus the git builders: they would apply a commit/reset/update-ref in
+// memory, so any git-mutating block is refused whatever its spelling.
+const runLines = (repo: InMemRepo, script: string): void => {
+  const lines = script.split("\n").map((line) => line.trim())
+  const known = lines.map(knownLine)
+  if (known.every(Boolean)) {
+    known.forEach((k, i) => k![1](repo, words(lines[i]!)))
+    return
   }
-  known[1](repo, words(line))
+  const applied = applyEmittedScript(repo, new Map(), script, { git: false })
+  if (!applied.ok) throw cannotRun(applied.error ?? "")
 }
 
 const failed = (what: string, result: CapturedCliResult): Error =>
@@ -251,7 +262,7 @@ export const testRepo = (options: TestRepoOptions): TestRepo => {
     show: (ref, path) => repo.fileAtRef(ref, path) ?? undefined,
     runScript: (script) => {
       const flow = applyFlowScript(repo, new Map(), script)
-      if (flow === undefined) for (const line of script.split("\n")) runLine(repo, line.trim())
+      if (flow === undefined) runLines(repo, script)
       else if (!flow.ok) throw new Error(`testkit: ${flow.error ?? ""}`)
     },
     gtd,