@try-works/dsh-recursive-mode 0.4.5 → 0.4.7

package/lib/index.js

@@ -5390,12 +5390,30 @@ const TOOL_ERRORS = {
 		problem: "this gate has no default artifact, so one must be named",
 		next: "pass artifact: <file> so the answer has somewhere durable to land"
 	},
+	RELAY_ONLY_FOR_RUN_START: {
+		code: "RM1150",
+		klass: "input",
+		problem: "relay applies only to the run-start gate, which is the only gate that asks the user-questions channel",
+		next: "drop relay for this gate and answer it with one of its own labels, or call recursive_ask with gate: run-start when the decision is whether to start the run"
+	},
 	MISSING_RUN_ID: {
 		code: "RM1101",
 		klass: "input",
 		problem: "runId is required",
 		next: "call recursive_status with no runId to see the latest run id in this workspace"
 	},
+	/**
+	* A run id is the NAME of the run directory and is joined onto the run layer
+	* as one path segment, so a path-shaped id is refused before any directory is
+	* created. See `run-id.ts` for the rule and for why it is not the runtime that
+	* learns to accept a path.
+	*/
+	BAD_RUN_ID: {
+		code: "RM1107",
+		klass: "input",
+		problem: "runId is not a single directory name",
+		next: "pass the run directory name such as 03-something, then call recursive_init again"
+	},
 	MISSING_ARTIFACT: {
 		code: "RM1102",
 		klass: "input",
@@ -5462,17 +5480,45 @@ const TOOL_ERRORS = {
 		problem: "the run has unresolved delegated work, so this phase cannot lock yet",
 		next: "call recursive_status to see the pending delegation, have the child write its reply.md, then lock again"
 	},
+	/**
+	* THE ORDERING DEFECT, AS A CODE. `recursive_ask gate=run-start` used to raise "start this run or hold?"
+	* over a Phase 0 document that was still the scaffold `recursive_init` wrote — placeholder requirements,
+	* unchecked lists, `FAIL` gates — and nothing put that document in front of the person either. The owner:
+	* *"i was never shown the spec before that so how could i approve if i havent seen it"*. Approving an
+	* unfilled template is not a decision about a spec, so the gate refuses to be raised until there is one.
+	*/
+	RUN_START_SPEC_UNFILLED: {
+		code: "RM4404",
+		klass: "state",
+		problem: "the Phase 0 requirements document is still the unfilled template, so there is no run spec for a person to approve",
+		next: "fill the requirements document in (define the requirement ids and their acceptance criteria, and complete the TODO list) and then call recursive_ask with gate: run-start again"
+	},
 	RUN_START_NO_CHANNEL: {
 		code: "RM5502",
 		klass: "runtime",
 		problem: "the run-start gate needs an answer, and this composition mounts no user-questions channel to ask one directly",
 		next: "call recursive_ask with gate: run-start and no answer to surface the question, then retry with answer: \"Start run\""
 	},
+	/**
+	* ⚠ RM5503 USED TO LIE. Its text asserted "so no person was asked" for EVERY cause, because the caller
+	* that produced it had already thrown the cause away — and a live session showed the cost: the gate
+	* failed 22.9 s into the call with a reason nobody could see, and its `Next:` clause prescribed the very
+	* call that had just failed, so no route to start a run remained. The problem statement now claims only
+	* what the gate knows (no decision came back), and the cause travels in the `detail` the caller
+	* supplies. `RUN_START_ANSWER_UNUSABLE` carries the one case this entry must NOT cover: a person was
+	* reached and their answer was not an approval.
+	*/
 	RUN_START_UNANSWERED: {
 		code: "RM5503",
 		klass: "runtime",
-		problem: "the user-questions channel mounted in this composition refused the run-start question, so no person was asked",
-		next: "use recursive_ask without an answer to surface the question, and retry it with answer: \"Start run\" once the user has approved the run start"
+		problem: "the run-start question reached no decision: the mounted user-questions channel failed before a person answered it",
+		next: "read the cause named in the detail, fix it and call recursive_ask again, or - when this composition cannot deliver the question at all - call recursive_ask with answer: \"Start run\" and relay=true to record the person's explicit approval as a relayed one"
+	},
+	RUN_START_ANSWER_UNUSABLE: {
+		code: "RM5504",
+		klass: "runtime",
+		problem: "a person was asked to start this run and their answer was not one of the labels the run-start gate offered",
+		next: "call recursive_ask again and have the person choose exactly \"Start run\" or \"Hold\"; a skipped or custom answer is not an approval, and no relayed answer can replace a decision the person made"
 	},
 	RUNTIME_REFUSED: {
 		code: "RM5501",
@@ -6831,6 +6877,79 @@ function extractAndGroup(runner, env, options = {}) {
 	};
 }
 //#endregion
+//#region src/run-spec.ts
+/** The named evidence classes, so a reader can tell a placeholder from an unmet gate. */
+const ARTIFACT_MARKER_IDS = {
+	placeholder: "placeholder",
+	uncheckedTodo: "unchecked-todo",
+	failedGate: "failed-gate"
+};
+/** Every marker, in the order they are reported for a single line. */
+const MARKER_PATTERNS = [
+	{
+		id: ARTIFACT_MARKER_IDS.placeholder,
+		re: /(^\s*\.\.\.\s*$)|(\[[^[\]\n<>]{2,120}\])|(<[^<>\n]{2,120}>)/
+	},
+	{
+		id: ARTIFACT_MARKER_IDS.uncheckedTodo,
+		re: /^\s*[-*]\s*\[ \]/
+	},
+	{
+		id: ARTIFACT_MARKER_IDS.failedGate,
+		re: /^\s*(Coverage|Approval):\s*FAIL\b/i
+	}
+];
+/**
+* Which marker a single line carries, or null.
+*
+* Exported because the client prints the marker NAMES beside the quoted lines, and a second classifier that
+* re-derived them would be a second answer to the same question.
+*/
+function markerIdsOnLine(line) {
+	const ids = [];
+	for (const pattern of MARKER_PATTERNS) if (pattern.re.test(line)) ids.push(pattern.id);
+	return ids;
+}
+/**
+* Classify one artifact's text.
+*
+* A verdict of `unfilled` means the document still carries the template's own placeholder text — the
+* evidence travels with it, line by line, so the refusal (and the client notice) can name what is missing
+* instead of asserting a state the reader cannot check.
+*/
+function classifyArtifact(text) {
+	const lines = text.split(/\r?\n/);
+	const hits = [];
+	for (let i = 0; i < lines.length; i += 1) {
+		const line = lines[i];
+		for (const id of markerIdsOnLine(line)) hits.push({
+			id,
+			line: i + 1,
+			text: line.trim()
+		});
+	}
+	return {
+		verdict: hits.some((hit) => hit.id === ARTIFACT_MARKER_IDS.placeholder) ? "unfilled" : "filled",
+		hits
+	};
+}
+/** The unfilled evidence only (what a refusal names). */
+function unfilledEvidence(result) {
+	return result.hits.filter((hit) => hit.id === ARTIFACT_MARKER_IDS.placeholder);
+}
+/**
+* One line naming what is missing, for a refusal sentence.
+*
+* The line number and the text are both quoted: "line 12: <short title>" is a thing a reader can go and
+* look at, while "the requirements are not filled in" is an assertion they would have to take on trust.
+*/
+function describeEvidence(hits, limit = 3) {
+	const shown = hits.slice(0, limit).map((hit) => "line " + String(hit.line) + ": " + hit.text);
+	const rest = hits.length - shown.length;
+	const suffix = rest > 0 ? " (and " + String(rest) + " more)" : "";
+	return shown.join(" | ") + suffix;
+}
+//#endregion
 //#region src/run-start.ts
 /**
 * PHASE 0 — STARTING A RUN IS A HUMAN DECISION, NOT A SIDE EFFECT OF SCAFFOLDING.
@@ -6854,9 +6973,15 @@ function extractAndGroup(runner, env, options = {}) {
 *      and nothing else, so the presence of a `Run Start` line is never on its own consent.
 *   2. IT IS ASKED, NOT ASSUMED. When the composition mounts `ctx.userQuestions` — the harness's own
 *      blocking human channel, the same one plan-mode's exit uses — the question is PUT TO THE PERSON and
-*      only their selection is recorded; a caller-supplied answer cannot stand in for it, and a channel
-*      that cannot reach anyone ends the call without a decision (RM5503). Only a composition with no
-*      channel at all falls back to the relayed answer, which is the contract the other three gates have.
+*      only their selection is recorded; a caller-supplied answer cannot stand in for it. A channel that
+*      RESOLVES with an answer the gate does not recognise is a person's decision the gate cannot record
+*      and it ends the call (RM5504). A channel that FAILS ends the call too (RM5503), naming the cause the
+*      channel threw — and there the caller may take the relayed route deliberately, with `relay=true`,
+*      which the result reports as `source: "relayed"` rather than as a person's own selection, so a
+*      composition whose channel cannot deliver the question can still start a run. A failure that means
+*      the question was cancelled, aborted, or timed out is never relayable. Only a composition with no
+*      channel at all falls back to the relayed answer unconditionally, which is the contract the other
+*      three gates have.
 *   3. THE GOAL CANNOT BE CREATED WITHOUT IT. `syncRunGoal` refuses to create a goal for a run whose
 *      approval record is absent, in EVERY branch that would create one — not only the "no goal yet"
 *      branch. That is the property `tests/run-start-approval.spec.ts` asserts, because a single
@@ -6947,6 +7072,43 @@ function readRunStartApproval(root, runId) {
 * the same string instead of re-typing it (a re-typed reason is a caller that silently stops matching).
 */
 const RUN_START_NOT_APPROVED = "run not started: phase 0 approval has not been granted";
+/**
+* PHASE 0 — THE GATE CANNOT BE RAISED BEFORE THERE IS A SPEC TO DECIDE ABOUT.
+*
+* THE DEFECT. `recursive_init` scaffolds Phase 0 as a TEMPLATE, and `recursive_ask gate=run-start` raised
+* "start this run or hold?" over it immediately — while every requirement was still `<short title>`, every
+* acceptance criterion was still `[observable condition 1]`, and nothing put the document in front of the
+* person at all. The owner: *"the card ui for accepting the spec appeared, but i was never shown the spec
+* before that so how could i approve if i havent seen it"*. Approving an unfilled template is not a decision
+* about a spec; there is no spec yet, and a card that asks the question anyway teaches a person to answer
+* without reading.
+*
+* ⚠ WHAT THIS DOES *NOT* TOUCH. It does not weaken the gate's own contract, it does not add a second way to
+* start a run, and it does not make the plugin the decider: it only refuses to ASK. A person's own answer
+* still wins (`recordRunStartAnswer` is unchanged), a spec still creates no goal, and cancellation / abort /
+* timeout are still unrelayable. The check runs BEFORE the question is put to anybody, so no card is shown
+* for a document that cannot be approved meaningfully.
+*
+* ⚠ AND IT IS A CHECK ON THE DOCUMENT, NOT ON THE CALLER. A `runId` that does not resolve is not this
+* refusal's business — the ask path already reports that — so the guard says `ok: true` there and lets the
+* existing route handle it.
+*/
+function runStartSpecGuard(root, runId) {
+	const content = readRunStartArtifact(root, runId);
+	if (content === null) return {
+		ok: false,
+		reason: toolError("RUN_START_SPEC_UNFILLED", "there is no Phase 0 document to approve: " + runStartArtifactPath(root, runId) + " does not exist yet")
+	};
+	const verdict = classifyArtifact(content);
+	if (verdict.verdict === "filled") return { ok: true };
+	const evidence = unfilledEvidence(verdict);
+	const context = verdict.hits.filter((hit) => hit.id !== "placeholder");
+	const contextNote = context.length === 0 ? "" : " (it also carries " + String(context.length) + " unfinished marker(s) of its own, starting at line " + String(context[0]?.line ?? 0) + ")";
+	return {
+		ok: false,
+		reason: toolError("RUN_START_SPEC_UNFILLED", "00-requirements.md for run " + JSON.stringify(runId) + " still carries the template scaffold" + contextNote + ": " + describeEvidence(evidence))
+	};
+}
 //#endregion
 //#region src/recursive_ask.tool.ts
 /**
@@ -7162,7 +7324,7 @@ function pendingGateFor(artifactFile, artifactText) {
 function createRecursiveAskTool(recursive) {
 	return defineTool({
 		name: "recursive_ask",
-		description: "Ask a human gate as a structured decision (tdd-mode, qa-signoff, gate-block), or ASK TO START A RUN (run-start: nothing runs, and no goal exists, until this gate is approved). Call without `answer` to ask; call with it to write the answer into the artifact as a durable marker. One ask per step.",
+		description: "Ask a human gate as a structured decision (tdd-mode, qa-signoff, gate-block), or ASK TO START A RUN (run-start: nothing runs, and no goal exists, until this gate is approved). Call without `answer` to ask; call with it to write the answer into the artifact as a durable marker. For run-start the mounted human channel is asked first and its own selection wins; when that channel cannot deliver the question, the refusal names the cause and `relay=true` with an explicit `answer` records the person's relayed approval. The run-start gate is REFUSED while the Phase 0 requirements document is still the unfilled template — the refusal quotes the placeholder lines, and there is nothing to approve until they are written. One ask per step.",
 		parameters: {
 			gate: {
 				type: "string",
@@ -7179,6 +7341,10 @@ function createRecursiveAskTool(recursive) {
 			answer: {
 				type: "string",
 				description: "One of the gate's option labels. Omit to ASK. For run-start, the labels are: " + RUN_START_GATE.options.map((option) => option.label).join(" | ") + "."
+			},
+			relay: {
+				type: "boolean",
+				description: "run-start only, and only after the person has approved in this conversation. Set relay=true when the mounted user-questions channel cannot deliver the run-start question: `answer` then stands in for the channel's selection and the result reports source: \"relayed\" instead of a direct selection. It is refused when the channel reports the question was cancelled, aborted, or timed out, and it is not needed when the person answers the card."
 			}
 		},
 		output: {
@@ -7193,9 +7359,20 @@ function createRecursiveAskTool(recursive) {
 			const runId = args.runId?.trim() ?? "";
 			if (runId === "") return { error: toolError("MISSING_RUN_ID") };
 			if (!askGateIds().includes(gateId)) return { error: toolError("BAD_ASK_GATE", "gate must be one of " + askGateIds().join(" | ")) };
+			if (args.relay === true && !isRunStartGate(gateId)) return { error: toolError("RELAY_ONLY_FOR_RUN_START", "gate is " + gateId) };
 			const root = await recursive.resolveWorkspaceRoot(exec.agent);
 			if (!root) return { error: toolError("NO_WORKSPACE") };
 			const artifact = (isRunStartGate(gateId) ? RUN_START_ARTIFACT : args.artifact ?? GATE_DEFAULT_ARTIFACT[gateId]).trim();
+			if (isRunStartGate(gateId)) {
+				const guard = runStartSpecGuard(root, runId);
+				if (!guard.ok) return {
+					error: guard.reason,
+					gate: RUN_START_GATE_ID,
+					runId,
+					artifact: RUN_START_ARTIFACT,
+					question: buildAskQuestionFor(RUN_START_GATE_ID)
+				};
+			}
 			let question;
 			try {
 				question = buildAskQuestionFor(gateId);
@@ -7217,7 +7394,7 @@ function createRecursiveAskTool(recursive) {
 				return { error: toolError("BAD_ASK_ANSWER", err instanceof Error ? err.message : String(err)) };
 			}
 			if (artifact === "") return { error: toolError("MISSING_ASK_ARTIFACT", "this gate needs an explicit artifact to record into") };
-			if (isRunStartGate(gateId)) return recordRunStartAnswer(recursive, root, runId, answer, exec);
+			if (isRunStartGate(gateId)) return recordRunStartAnswer(recursive, root, runId, answer, exec, args.relay === true);
 			const marker = answerMarker(gateId, answer);
 			const written = recursive.recordAskAnswer(root, runId, artifact, marker);
 			return {
@@ -7234,28 +7411,83 @@ function createRecursiveAskTool(recursive) {
 /**
 * PHASE 0 — record the answer to the run-start gate, and start the run only if it says so.
 *
-* ⚠ THIS GATE NEVER ACCEPTS A RELAYED ANSWER WHILE A HUMAN CHANNEL IS MOUNTED. That is the rule that makes
-* an approval a human act rather than an inference: when `ctx.userQuestions` is present, the question is
-* PUT TO THE PERSON and nothing else can settle it — not the caller's own `answer` argument, and not a
-* fabrication, because `ask()` resolves only with a real selection. A person's decline is likewise final
-* for that call and cannot be overridden by a model that asked for `Start run` in the same breath.
-*
-* ⚠ AND WHEN NO CHANNEL IS MOUNTED, THE RELAYED ANSWER IS THE ONLY POSSIBLE SOURCE, so it is used — that
-* is the same contract the other three gates have always had, and refusing it would leave a composition
-* without the channel unable to start any run at all. The question is surfaced first by the ASK branch
-* (the card data the host renders), and the model's `answer` is that person's selection coming back.
-*
-* ⚠ WHAT THE GATE THEREFORE DOES *NOT* CLAIM, stated rather than implied: in a composition with no
-* `userQuestions` channel, a plugin cannot verify that a person was really asked, so a model could in
-* principle relay a label nobody gave. That is a property of the relay, not of this gate — and it is the
-* reason the channel is consulted in preference whenever it exists. See the header of `run-start.ts`.
-*/
-async function recordRunStartAnswer(recursive, root, runId, answer, exec) {
+* ⚠ THREE OUTCOMES, AND TELLING THEM APART IS THE FIX. The first version of this function collapsed all of
+* them into one `null`: "the channel threw", "the channel resolved with something unrecognisable", and
+* "nobody answered" produced the same refusal, whose text asserted a cause ("so no person was asked") the
+* plugin had already thrown away. A live session paid for that: the call failed after 22.9 s, the operator
+* could not be told why, and the refusal's own advice prescribed the call that had just failed. So:
+*
+*   1. A PERSON WAS REACHED (`unusable`): the channel resolved, so somebody answered, and their answer is
+*      not a label this gate offered — a skip, a custom value, several labels at once. That is a DECISION
+*      the gate cannot record, and it is final: neither the caller's `answer` nor `relay=true` may replace
+*      it. (RM5504)
+*   2. THE CHANNEL FAILED (`unavailable`): no decision came back at all, and the refusal NAMES THE CAUSE
+*      from the error the channel threw. Here the run can still be started, because a composition whose
+*      channel cannot deliver the question would otherwise be unable to start any run — but only by the
+*      caller asking for the relay in so many words (`relay=true`), which the result reports as
+*      `source: "relayed"` rather than as a person's own selection. (RM5503)
+*   3. NO CHANNEL IS MOUNTED: the relayed answer is the only possible source, exactly as before. (RM5502
+*      when there is no answer either)
+*
+* ⚠ AND A CANCELLED OR CLOSED QUESTION IS NEVER RELAYABLE. `ASK_CANCELLED`, `ASK_ABORTED` and
+* `ASK_TIMED_OUT` are the codes that mean the question was settled from outside this gate — the card was
+* dismissed, the turn was cancelled, or a foreground window ended. The relay is refused for those, so a
+* question the operator stopped cannot be turned into an approval by asking again in the same breath.
+* Every other failure is a composition or capability failure — the question reached nobody — which is the
+* class the relay exists for.
+*
+* ⚠ WHAT THE GATE STILL DOES *NOT* CLAIM: a relayed approval is a relayed approval. The plugin cannot
+* verify that a person gave the label, and it does not pretend otherwise — the result's `source` and
+* `channel` fields say where the decision came from, and a direct selection is preferred whenever the
+* channel can produce one.
+*/
+async function recordRunStartAnswer(recursive, root, runId, answer, exec, relay = false) {
+	const question = buildAskQuestionFor(RUN_START_GATE_ID);
 	const channel = recursive.userQuestionsChannel;
-	const fromChannel = channel ? await askRunStartDirectly(channel, exec) : null;
-	if (channel !== null && fromChannel === null) return { error: toolError("RUN_START_UNANSWERED") };
+	const channelOutcome = channel ? await askRunStartDirectly(channel, exec) : null;
+	if (channelOutcome !== null && channelOutcome.kind === "unusable") return {
+		error: toolError("RUN_START_ANSWER_UNUSABLE", channelOutcome.detail),
+		gate: RUN_START_GATE_ID,
+		runId,
+		artifact: RUN_START_ARTIFACT,
+		question
+	};
+	if (channelOutcome !== null && channelOutcome.kind === "unavailable") {
+		const blocked = !relay ? "the caller did not ask for the relay" : "the channel reports the question was cancelled, aborted, or timed out, so it is not relayable";
+		if (!relay || !channelOutcome.relayable) return {
+			error: toolError("RUN_START_UNANSWERED", channelOutcome.detail + " (" + blocked + ")"),
+			gate: RUN_START_GATE_ID,
+			runId,
+			artifact: RUN_START_ARTIFACT,
+			question,
+			channel: {
+				outcome: "unavailable",
+				cause: channelOutcome.cause,
+				relayable: channelOutcome.relayable
+			}
+		};
+		if (answer === void 0) return {
+			error: toolError("RUN_START_UNANSWERED", channelOutcome.detail + " (the relay was authorised but no answer was supplied, so there is no decision to record)"),
+			gate: RUN_START_GATE_ID,
+			runId,
+			artifact: RUN_START_ARTIFACT,
+			question,
+			channel: {
+				outcome: "unavailable",
+				cause: channelOutcome.cause,
+				relayable: channelOutcome.relayable
+			}
+		};
+	}
+	const fromChannel = channelOutcome !== null && channelOutcome.kind === "answered" ? channelOutcome.answer : null;
 	const final = fromChannel ?? answer;
-	if (final === void 0) return { error: toolError("RUN_START_NO_CHANNEL") };
+	if (final === void 0) return {
+		error: toolError("RUN_START_NO_CHANNEL"),
+		gate: RUN_START_GATE_ID,
+		runId,
+		artifact: RUN_START_ARTIFACT,
+		question
+	};
 	let decided;
 	try {
 		decided = validateAskAnswerFor(RUN_START_GATE_ID, final);
@@ -7268,6 +7500,12 @@ async function recordRunStartAnswer(recursive, root, runId, answer, exec) {
 		answer: decided,
 		artifact: RUN_START_ARTIFACT,
 		source: fromChannel === null ? "relayed" : "user-questions",
+		...channelOutcome !== null && channelOutcome.kind === "unavailable" ? { channel: {
+			outcome: "unavailable",
+			cause: channelOutcome.cause,
+			relayable: channelOutcome.relayable,
+			relayed: true
+		} } : {},
 		path: outcome.path,
 		replaced: outcome.replaced,
 		armed: outcome.ok && outcome.goal.ok,
@@ -7275,16 +7513,66 @@ async function recordRunStartAnswer(recursive, root, runId, answer, exec) {
 	};
 }
 /**
-* Ask the run-start question through the blocking channel and return the selection, or null when the
-* person was not reachable (no answerer, no live root agent, a dismissal, an abort).
+* ⚠ THE CODES THAT MEAN THE QUESTION WAS CANCELLED OR CLOSED rather than never delivered: the person
+* dismissed the card, their turn was cancelled, or a foreground window ended. A caller may not convert any
+* of those into an approval by asking for the relay in the same breath. Every other failure means the
+* question reached nobody — a composition or capability failure, which is the class the relay exists for.
+*/
+const NON_RELAYABLE_CHANNEL_CODES = [
+	"ASK_CANCELLED",
+	"ASK_ABORTED",
+	"ASK_TIMED_OUT"
+];
+/**
+* Name the failure of one `ask()` call, without inventing anything about it.
+*
+* The cause is the error's own `code` when it has one (the harness's `UserQuestionError` carries
+* `NO_PROVIDER`, `CALLER_NOT_LIVE`, `DELEGATED_CALLER`, `ASK_ABORTED`, …), else its `name`, else its
+* JavaScript type. `detail` keeps the message verbatim so a reader sees the channel's own words rather
+* than this plugin's paraphrase — the paraphrase is exactly how the previous version came to assert a
+* cause nobody had.
+*/
+function classifyChannelFailure(err) {
+	const code = err?.code;
+	const name = err instanceof Error ? err.name : typeof err;
+	const message = err instanceof Error ? err.message : String(err);
+	const hasCode = typeof code === "string" && code.trim() !== "";
+	const cause = hasCode ? code : name;
+	const relayable = !(hasCode && NON_RELAYABLE_CHANNEL_CODES.includes(code));
+	return {
+		cause,
+		detail: "channel threw " + name + "[" + cause + "]: " + message,
+		relayable
+	};
+}
+/** Describe an answer that arrived but is not a decision this gate can record. */
+function describeUnusableAnswer(item) {
+	const offered = RUN_START_GATE.options.map((option) => option.label);
+	const list = (values) => JSON.stringify(values.join(" | "));
+	if (item === void 0) return "the channel resolved with no answer for question " + JSON.stringify(RUN_START_GATE.id) + " at all";
+	const raw = item.selected ?? [];
+	const custom = item.custom?.trim() ?? "";
+	if (raw.length === 0 && custom === "") return "the person skipped the question, and a skip is not an approval";
+	if (raw.length === 0) return "the person answered " + JSON.stringify(custom) + " as free text rather than one of " + list(offered);
+	const recognised = raw.filter((label) => offered.includes(label));
+	if (recognised.length === 0) return "the person selected " + list(raw) + ", and none of those name a label this gate offered (" + offered.join(" | ") + ")";
+	if (recognised.length === raw.length) return "the person selected " + list(raw) + ", and an approval is exactly one of " + list(offered);
+	return "the person selected " + list(raw) + ", of which only " + list(recognised) + " name this gate's labels " + list(offered);
+}
+/**
+* Ask the run-start question through the blocking channel and report WHAT HAPPENED.
 *
-* The selection is filtered to the gate's OWN labels before it is returned: a question a UI answered with
-* a free-text custom value must not become an approval just because it arrived on the right channel.
+* ⚠ THE CATCH IS THE POINT. It used to be `catch { return null }` — a blocking human question whose
+* failure cause was erased at the exact moment the cause was the only thing worth knowing. Every path out
+* of this function now says which path it was.
+*
+* The selection is filtered to the gate's OWN labels: a question a UI answered with a free-text custom
+* value must not become an approval just because it arrived on the right channel.
 */
 async function askRunStartDirectly(channel, exec) {
 	const known = RUN_START_GATE.options.map((option) => option.label);
 	try {
-		const selected = (await channel.ask({
+		const item = (await channel.ask({
 			questions: [{
 				id: RUN_START_GATE.id,
 				header: RUN_START_GATE.header,
@@ -7294,11 +7582,21 @@ async function askRunStartDirectly(channel, exec) {
 			agent: exec.agent,
 			signal: exec.signal,
 			wait: { callId: exec.callId }
-		})).answers.find((entry) => entry.id === RUN_START_GATE.id)?.selected?.filter((label) => known.includes(label)) ?? [];
-		if (selected.length !== 1) return null;
-		return selected[0];
-	} catch {
-		return null;
+		})).answers.find((entry) => entry.id === RUN_START_GATE.id);
+		const selected = item?.selected?.filter((label) => known.includes(label)) ?? [];
+		if (selected.length !== 1) return {
+			kind: "unusable",
+			detail: describeUnusableAnswer(item)
+		};
+		return {
+			kind: "answered",
+			answer: selected[0]
+		};
+	} catch (err) {
+		return {
+			kind: "unavailable",
+			...classifyChannelFailure(err)
+		};
 	}
 }
 //#endregion
@@ -11347,6 +11645,75 @@ function createRecursiveStatusTool(recursive) {
 	});
 }
 //#endregion
+//#region src/run-id.ts
+/**
+* A RUN ID IS A NAME, NOT A PATH.
+*
+* WHY THIS MODULE EXISTS. Every consumer of a run id JOINS it onto a directory
+* that already carries the meaning "the run layer":
+*
+*     join(root, '.recursive', 'run', runId)   // runtime.ts, run.ts, handoff.ts, scratch.ts
+*     join(repoRoot, '.worktrees', runId)      // worktree.ts (a linked worktree)
+*     'recursive/' + runId                     // worktree.ts (the run's git branch)
+*
+* `join` is a PATH operation: absolute paths, drive specifiers and `..` segments
+* are all legal input to it, and each one silently changes what the call means.
+* A caller who passes `E:\tmp\rm-live-diagnostics\01-calculator-lib` is asking
+* for a run "on another drive"; what they get is a `mkdir` of
+*
+*     <workspace>\.recursive\run\E:\tmp\rm-live-diagnostics\01-calculator-lib
+*
+* which is not drive-qualified at all — on POSIX and Windows alike the colon is
+* just another character in a relative component. The result is a bogus nested
+* folder INSIDE the workspace, created before anything can refuse it, surfacing
+* far away as an ENOENT-shaped runtime failure (RM5501) with the operator's
+* filesystem already dirty.
+*
+* SO THE RULE IS ENFORCED WHERE THE NAME ENTERS, and NOT by teaching the runtime
+* to accept a path. The joins in `runtime.ts` are CORRECT for a name; what was
+* missing was a gate on the name. Do not "fix" this back: a run on another drive
+* or in a worktree is reached through the session's control-plane root
+* (`recursive_worktree`, `00-worktree.md`) — the run layer is never relocated by
+* smuggling a path into the id.
+*
+* The charset below is deliberately the SAME one the read path already uses
+* (`live-route.ts` `DOC_SAFE_RE`) so a name this gate accepts is a name that
+* route can serve.
+*/
+/**
+* The accepted shape, as prose that can be embedded in a model-facing parameter
+* description and in a refusal detail, so the rule is stated once.
+*/
+const RUN_ID_RULE = "letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment";
+/** Directory-name charset — the read path's `DOC_SAFE_RE`, verbatim. */
+const RUN_ID_CHARS = /^[A-Za-z0-9._-]+$/;
+/**
+* Why a run id is refused, or `null` when it is a usable NAME.
+*
+* The returned string is the SPECIFIC problem (which rule the id broke), with no
+* trailing punctuation and no sentence of its own, so a caller can hand it to
+* `toolError('BAD_RUN_ID', …)` as the detail. `RUN_ID_RULE` states the shape.
+*
+* The order of the checks is part of the message quality: a Windows absolute
+* path is reported as a drive-qualified path (what the caller passed) rather
+* than as a separator complaint (what that path is made of).
+*/
+function runIdProblem(raw) {
+	if (raw === "") return "runId is empty";
+	if (raw.length > 100) return "runId is " + raw.length + " characters, over the 100 allowed";
+	if (/^[A-Za-z]:/.test(raw)) return "runId is a Windows drive-qualified path, starting with \"" + raw.slice(0, 2) + "\"";
+	if (raw.includes("/") || raw.includes("\\")) return "runId contains the path separator \"" + (raw.includes("/") ? "/" : "\\") + "\"";
+	if (raw.includes(":")) return "runId contains a colon (\":\"), which is a drive and stream separator on Windows";
+	if (raw.includes("..")) return "runId contains a \"..\" segment, which escapes the run directory";
+	if (raw.startsWith(".")) return "runId starts with \".\", which makes it a hidden name or a relative path segment";
+	if (raw.endsWith(".")) return "runId ends with \".\"";
+	if (!RUN_ID_CHARS.test(raw)) {
+		if (/\s/.test(raw)) return "runId contains a space or other whitespace character inside the name";
+		return "runId contains a character outside the allowed set";
+	}
+	return null;
+}
+//#endregion
 //#region src/recursive_init.tool.ts
 /**
 * PHASE 0 — SCAFFOLDING IS NOT STARTING, AND THE TOOL SAYS SO AT THE MOMENT IT MATTERS.
@@ -11362,15 +11729,23 @@ function createRecursiveStatusTool(recursive) {
 * rather than "this call created something": re-initialising an APPROVED run reports `approved: true`
 * (and the run keeps its goal), which is what a caller re-scaffolding a run it already started needs to
 * see.
+*
+* AND A RUN ID IS A NAME, NOT A PATH. This is the boundary where the name enters, so it is the boundary
+* that refuses a path-shaped one — loudly, and before `initRun` can mkdir anything. The check lives here
+* (through the shared `run-id.ts` rule) rather than in `runtime.ts` because `initRun` is not the only
+* caller and because the runtime's `join(root, '.recursive', 'run', runId)` is CORRECT for a name; what
+* was missing was a gate on the name. Teaching the runtime to accept a path would silently relocate the
+* run layer instead of rejecting the call. See `run-id.ts` for the rule, the evidence behind it, and the
+* "do not fix this back" note.
 */
 function createRecursiveInitTool(recursive) {
 	return defineTool({
 		name: "recursive_init",
-		description: "Scaffold a new recursive-mode run directory (or ensure an existing one) with stub artifact headers. Delegates to the RecursiveRuntime service (no duplicated scaffolding logic). When createWorktree is true, a linked worktree is created first and the run is scaffolded inside it. THIS DOES NOT START THE RUN: no goal exists until the user approves phase 0 through recursive_ask gate=run-start, so the result carries runStartApproval — read it and ask.",
+		description: "Scaffold a new recursive-mode run directory (or ensure an existing one) with stub artifact headers. Delegates to the RecursiveRuntime service (no duplicated scaffolding logic). When createWorktree is true, a linked worktree is created first and the run is scaffolded inside it. THIS DOES NOT START THE RUN: no goal exists until the user approves phase 0 through recursive_ask gate=run-start, so the result carries runStartApproval — read it and ask. The runId is the NAME of the run directory under .recursive/run/ and is never a path (see the parameter description): a path-shaped runId is refused before anything is written.",
 		parameters: {
 			runId: {
 				type: "string",
-				description: "Run id (e.g. 03-something). Required."
+				description: "Run id — the NAME of the run directory under .recursive/run/ (e.g. 03-something, 01-calculator-lib), never a path: " + RUN_ID_RULE + ". A run on another drive or inside a worktree is reached with recursive_worktree, not by passing a path here. Required."
 			},
 			createWorktree: {
 				type: "boolean",
@@ -11389,9 +11764,12 @@ function createRecursiveInitTool(recursive) {
 			}]
 		},
 		async execute(args, exec) {
-			if (!args.runId || args.runId.trim() === "") return { error: toolError("MISSING_RUN_ID") };
+			const runId = args.runId?.trim() ?? "";
+			if (runId === "") return { error: toolError("MISSING_RUN_ID") };
+			const problem = runIdProblem(runId);
+			if (problem !== null) return { error: toolError("BAD_RUN_ID", problem + " - run ids allow letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment - e.g. 01-calculator-lib, fixture-run") };
 			try {
-				return await recursive.initRun(args.runId.trim(), exec.agent, {
+				return await recursive.initRun(runId, exec.agent, {
 					createWorktree: args.createWorktree === true,
 					baseBranch: args.baseBranch?.trim() || void 0
 				});
@@ -11605,6 +11983,16 @@ function createRecursiveLintTool(recursive) {
 * SESSION's workspace only (R1 workspace-scoping invariant). The run is resolved
 * via the session agent's cwd -> workspace registry; a runId outside the current
 * workspace is rejected.
+*
+* A RUN ID IS A NAME, NOT A PATH HERE TOO, and "outside the current workspace is rejected" is NOT enough
+* on its own: `closeoutRun` joins the id onto the run layer and then writes a receipt under it, and its
+* scoping check is `runDir.startsWith(runRoot)` — a string prefix test, not containment. A `..\` segment
+* that lands on a SIBLING of the run layer passes that check whenever the sibling's name begins with
+* `run`, so a path-shaped id can still receive a write. MEASURED pre-fix: `..\run-away` and `../run-away`
+* were accepted and reached the report; the other shapes below were stopped only by the sibling not
+* existing, which is the operator's filesystem deciding, not the tool. The rule is `run-id.ts`; this
+* boundary is where the name enters, so this is where it is refused, with `recursive_init`'s refusal shape
+* — `BAD_RUN_ID` (RM1107), same detail sentence.
 */
 function createRecursiveCloseoutTool(recursive) {
 	return defineTool({
@@ -11617,7 +12005,7 @@ function createRecursiveCloseoutTool(recursive) {
 			},
 			runId: {
 				type: "string",
-				description: "Run id (e.g. 03-something). Required. Must resolve inside the current workspace."
+				description: "Run id — the NAME of the run directory under .recursive/run/ (e.g. 03-something), never a path: " + RUN_ID_RULE + ". Required. Must resolve inside the current workspace."
 			}
 		},
 		output: {
@@ -11629,9 +12017,12 @@ function createRecursiveCloseoutTool(recursive) {
 		},
 		async execute(args, exec) {
 			if (!args.phase || !args.runId || args.runId.trim() === "") return { error: toolError("MISSING_PHASE_AND_RUN") };
+			const runId = args.runId.trim();
+			const problem = runIdProblem(runId);
+			if (problem !== null) return { error: toolError("BAD_RUN_ID", problem + " - run ids allow letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment - e.g. 01-calculator-lib, fixture-run") };
 			const root = await recursive.resolveWorkspaceRoot(exec.agent);
 			if (!root) return { error: toolError("NO_WORKSPACE") };
-			return await recursive.closeoutRun(root, args.runId.trim(), args.phase.trim(), exec.agent);
+			return await recursive.closeoutRun(root, runId, args.phase.trim(), exec.agent);
 		}
 	});
 }
@@ -11641,6 +12032,16 @@ function createRecursiveCloseoutTool(recursive) {
 * `recursive_scratch` — read/write/append the run-scoped disposable scratchpad
 * (R5) under the CURRENT session workspace only (R1). Scratch is git-ignored
 * and never citable as an Input.
+*
+* A RUN ID IS A NAME, NOT A PATH HERE TOO. `scratchRun` joins it onto the run layer, and it then WRITES
+* (write/append) into the directory it landed on, so a path-shaped id does not merely fail to find a run:
+* with a `..\` segment it can find and write into a SIBLING of the run layer, because the runtime's
+* workspace-scoping check is `runDir.startsWith(runRoot)` — a string prefix test, not containment — and a
+* sibling directory whose name begins with `run` passes it. MEASURED pre-fix: `..\run-away` was accepted
+* and `scratchRun` wrote `scratch.md` into `<workspace>\.recursive\run-away\scratch\`. The rule is
+* `run-id.ts`; the gate sits here, at the boundary where the name enters, rather than in `scratchRun`, for
+* the same reason `recursive_init`'s does. Refusal shape is `recursive_init`'s, `BAD_RUN_ID` (RM1107),
+* composed identically: one rule, one message.
 */
 function createRecursiveScratchTool(recursive) {
 	return defineTool({
@@ -11653,7 +12054,7 @@ function createRecursiveScratchTool(recursive) {
 			},
 			runId: {
 				type: "string",
-				description: "Run id (e.g. 03-something). Required; must resolve inside the current workspace."
+				description: "Run id — the NAME of the run directory under .recursive/run/ (e.g. 03-something), never a path: " + RUN_ID_RULE + ". Required; must resolve inside the current workspace."
 			},
 			target: {
 				type: "string",
@@ -11676,6 +12077,8 @@ function createRecursiveScratchTool(recursive) {
 			const runId = args.runId?.trim() ?? "";
 			const target = args.target ?? "";
 			if (!action || !runId || !target) return { error: toolError("MISSING_SCRATCH_ARGS") };
+			const problem = runIdProblem(runId);
+			if (problem !== null) return { error: toolError("BAD_RUN_ID", problem + " - run ids allow letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment - e.g. 01-calculator-lib, fixture-run") };
 			if (target !== "md" && target !== "ts") return { error: toolError("BAD_TARGET") };
 			const root = await recursive.resolveWorkspaceRoot(exec.agent);
 			if (!root) return { error: toolError("NO_WORKSPACE") };
@@ -11689,6 +12092,22 @@ function createRecursiveScratchTool(recursive) {
 * `recursive_worktree` — create a linked git worktree for a run and/or
 * promote a branch up the dev/stage/main chain. Workspace-scoped: the
 * operations run under the SESSION's control-plane root only.
+*
+* A RUN ID IS A NAME, NOT A PATH HERE TOO, and this is the worst place to be without the rule: a `create`
+* builds TWO things out of the id — the linked worktree directory `.worktrees/<runId>` AND the git branch
+* `recursive/<runId>` (git accepts '/' inside a ref) — so a path-shaped id used to leave a worktree and a
+* ref behind, not just a folder. MEASURED pre-fix, per id, against a fresh repo: `nested/child-run`
+* returned ok:true and created BOTH `.worktrees/nested/child-run` and
+* `refs/heads/recursive/nested/child-run`, while the shapes git itself refuses as ref syntax
+* (`recursive//tmp/x`, `recursive/C:…`, `.hidden-run`, a trailing space) failed the worktree add and
+* created neither. The rule is `run-id.ts` and is not restated here; the gate sits at this boundary, ahead
+* of `createRunWorktree`, so the refusal no longer depends on git happening to dislike the ref name.
+*
+* The refusal is `BAD_RUN_ID` (RM1107), composed exactly as `recursive_init` composes it — same code, same
+* detail, same sentence. One message for one rule is what keeps a caller from having to learn a second
+* vocabulary for the same defect, and the shared remedy it carries ("call recursive_init again") is right
+* for this tool as well: a `create` for a run that does not exist yet is exactly what `recursive_init`
+* with `createWorktree: true` does.
 */
 function createRecursiveWorktreeTool(recursive) {
 	return defineTool({
@@ -11697,7 +12116,7 @@ function createRecursiveWorktreeTool(recursive) {
 		parameters: {
 			runId: {
 				type: "string",
-				description: "Run id the worktree is created for (e.g. 03-something). Required for create."
+				description: "Run id the worktree is created for — the NAME of the run (e.g. 03-something), never a path: " + RUN_ID_RULE + ". A path-shaped id is refused for create. Required for create."
 			},
 			action: {
 				type: "string",
@@ -11729,7 +12148,10 @@ function createRecursiveWorktreeTool(recursive) {
 			if (!root) return { error: toolError("NO_WORKSPACE") };
 			if (action === "create") {
 				if (!args.runId || args.runId.trim() === "") return { error: toolError("MISSING_CREATE_RUN_ID") };
-				return recursive.createRunWorktree(root, args.runId.trim(), args.baseBranch?.trim() || void 0);
+				const runId = args.runId.trim();
+				const problem = runIdProblem(runId);
+				if (problem !== null) return { error: toolError("BAD_RUN_ID", problem + " - run ids allow letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment - e.g. 01-calculator-lib, fixture-run") };
+				return recursive.createRunWorktree(root, runId, args.baseBranch?.trim() || void 0);
 			}
 			if (action === "promote") {
 				if (!args.fromBranch || !args.toBranch) return { error: toolError("MISSING_PROMOTE_BRANCHES") };
@@ -11748,6 +12170,14 @@ function createRecursiveWorktreeTool(recursive) {
 * (runtime.phaseRules -> phaseRulesFor) as the once-per-phase pre-step
 * reminder, so the agent can re-ask for the rules without re-injecting them on
 * every step. Returns { error } when no active phase is found.
+*
+* A RUN ID IS A NAME, NOT A PATH HERE TOO, INCLUDING WHEN IT IS OMITTED. Omitted and path-shaped are
+* different cases and must stay different: omitted means "the latest run by mtime", which `resolveRunDir`
+* answers by DISCOVERY rather than by joining anything, and that case is untouched below. A path-shaped id,
+* by contrast, is joined by `phaseRules` -> `resolveRunDir` and then read, and `recordInjection` WRITES
+* `memory-injections.json` under whatever directory it resolved to — with no scoping check at all on this
+* path. So the gate fires only on an id that was actually supplied, and `run-id.ts` owns the rule. The
+* refusal is `recursive_init`'s, `BAD_RUN_ID` (RM1107), composed identically.
 */
 function createRecursivePhaseTool(recursive) {
 	return defineTool({
@@ -11755,7 +12185,7 @@ function createRecursivePhaseTool(recursive) {
 		description: "Return the lint rules + instructions for the current recursive-mode phase (required sections, gates, TDD/QA notes). Call once when entering a new phase; the same rules are also auto-injected once per phase transition.",
 		parameters: { runId: {
 			type: "string",
-			description: "Optional run id (defaults to the latest run by mtime)"
+			description: "Optional run id — the NAME of the run directory under .recursive/run/ (e.g. 03-something), never a path: " + RUN_ID_RULE + ". Omit it for the latest run by mtime."
 		} },
 		output: {
 			schema: { type: "json" },
@@ -11765,7 +12195,12 @@ function createRecursivePhaseTool(recursive) {
 			}]
 		},
 		async execute(args, exec) {
-			const result = await recursive.phaseRules(args.runId, exec.agent);
+			const runId = args.runId?.trim();
+			if (runId !== void 0) {
+				const problem = runId === "" ? "runId is empty" : runIdProblem(runId);
+				if (problem !== null) return { error: toolError("BAD_RUN_ID", problem + " - run ids allow letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment - e.g. 01-calculator-lib, fixture-run") };
+			}
+			const result = await recursive.phaseRules(runId, exec.agent);
 			if (!result) return { error: toolError("NO_PHASE") };
 			return result;
 		}