shapeup-sdlc 1.7.0 → 3.0.0

package/hooks/anti-rationalization.mjs

@@ -1,238 +0,0 @@
-#!/usr/bin/env node
-// Anti-rationalization — advisory Stop hook (v1.2).
-//
-// When the session's final message claims completion ("done", "all tests pass", "ready to
-// ship") while the harness's own mechanical facts disagree (unfinished board tasks, a red T0
-// verdict), this hook says so — to the user, out loud, with the facts.
-//
-// ADVISORY ONLY, by architectural invariant: "QA is a level-up, not a gate." This hook exits
-// 0 always and emits at most { systemMessage } — never { decision: "block" }, never exit 2.
-// A blocking Stop hook would be a second gate behind the single judge (spec-evaluator).
-//
-// Harness-scoped: fires only when a run is actually active (.shapeup/active-scope
-// exists, or some .shapeup/*/harness-run.md is mid-build). An always-on nag on
-// non-harness work is exactly the annoyance that gets hooks disabled.
-//
-// Contract: Stop stdin JSON { cwd, stop_hook_active, last_assistant_message, transcript_path }.
-
-import { readFileSync, readdirSync, existsSync, statSync } from "node:fs";
-import { join } from "node:path";
-import { isMain } from "../skills/tech-lead/scripts/lib/is-main.mjs";
-import { localDir, localRoot, activeScope } from "../skills/tech-lead/scripts/lib/paths.mjs";
-
-import { runHook, readStdin, settle } from "./lib/decision.mjs";
-
-function readJSON(p) {
-  try { return JSON.parse(readFileSync(p, "utf8")); } catch { return null; }
-}
-
-function frontmatter(text) {
-  const m = /^---\n([\s\S]*?)\n---/.exec(text || "");
-  if (!m) return {};
-  const fm = {};
-  for (const line of m[1].split("\n")) {
-    const kv = /^([A-Za-z_][\w-]*):\s*(.*)$/.exec(line.trim());
-    if (kv) fm[kv[1]] = kv[2].replace(/^['"]|['"]$/g, "");
-  }
-  return fm;
-}
-
-const MID_RUN = new Set(["orienting", "mapping", "building", "evaluating"]);
-
-/** The active harness slug, or null when no run is in progress. */
-export function activeSlug(cwd) {
-  const pointer = readJSON(activeScope(cwd));
-  if (pointer?.slug) return pointer.slug;
-  const root = localDir(cwd);
-  if (!existsSync(root)) return null;
-  for (const entry of readdirSync(root)) {
-    const runPath = join(root, entry, "harness-run.md");
-    if (!existsSync(runPath)) continue;
-    try {
-      const fm = frontmatter(readFileSync(runPath, "utf8"));
-      if (MID_RUN.has(fm.status)) return entry;
-    } catch { /* unreadable run file → not this one */ }
-  }
-  return null;
-}
-
-/**
- * Does the text claim the work is finished — or promise that it is about to be?
- *
- * The past-tense half is the original detector. The future-tense half was added after a transcript
- * this hook should have caught and structurally could not: the session ended on "The tech-lead
- * skill is orchestrating the full harness. It will: 1…".
- * A promise at the END of a session is a completion claim wearing different grammar — the run is
- * over, and the thing it says it will do is never going to happen. Matching only past tense meant
- * the emptiest failures were the least detectable, which is backwards.
- *
- * (The zero-work case — dispatched and never started — belongs to `gate-zerowork.mjs`, which
- * blocks on a mechanical absence rather than on phrasing. This one covers narration INSIDE a run
- * that did start.)
- */
-const FUTURE_CLAIM = /\b(will (?:now )?(?:run|orchestrate|execute|proceed|begin|start)|is orchestrating|I'?ll (?:now )?(?:run|start|begin|orchestrate)|going to (?:run|orchestrate|execute))\b/i;
-
-export function detectClaim(text) {
-  if (!text || typeof text !== "string") return null;
-  const past = /\b(done|complete(?:d)?|finished|shipped|ready to ship|all (?:tests|tasks) pass(?:ing|ed)?|everything works)\b/i.exec(text);
-  if (past) return past[1];
-  const future = FUTURE_CLAIM.exec(text);
-  return future ? future[0] : null;
-}
-
-/**
- * Is this claim a promise about work still to come, rather than an assertion that work is done?
- * The two are checked against DIFFERENT facts — see contradictions().
- * @param {string} claim - The matched claim fragment from detectClaim().
- * @returns {boolean} True for a future-tense promise.
- */
-export function isFutureClaim(claim) {
-  return typeof claim === "string" && FUTURE_CLAIM.test(claim);
-}
-
-/** Read-only mechanical facts about the run — the evidence the claim is checked against. */
-export function gatherFacts(cwd, slug) {
-  const root = localRoot(cwd, slug);
-  const facts = { unfinished: [], red_t0: null, run_status: null, final_verdict: null };
-
-  const tasksDir = join(root, "tasks");
-  if (existsSync(tasksDir)) {
-    for (const f of readdirSync(tasksDir)) {
-      if (!/^TASK-.*\.md$/.test(f)) continue;
-      try {
-        const fm = frontmatter(readFileSync(join(tasksDir, f), "utf8"));
-        if (fm.status && fm.status !== "done") facts.unfinished.push(fm.id || f.replace(/\.md$/, ""));
-      } catch { /* unreadable task file → no fact */ }
-    }
-  }
-
-  const verdictsDir = join(root, "t0", "verdicts");
-  if (existsSync(verdictsDir)) {
-    let latest = null;
-    for (const f of readdirSync(verdictsDir)) {
-      const m = /^r(\d+)-a(\d+)\.json$/.exec(f);
-      if (!m) continue;
-      const key = [Number(m[1]), Number(m[2])];
-      if (!latest || key[0] > latest.key[0] || (key[0] === latest.key[0] && key[1] > latest.key[1])) {
-        latest = { key, file: f };
-      }
-    }
-    if (latest) {
-      const verdict = readJSON(join(verdictsDir, latest.file));
-      if (verdict?.overall === "red") facts.red_t0 = latest.file.replace(/\.json$/, "");
-    }
-  }
-
-  const runPath = join(root, "harness-run.md");
-  if (existsSync(runPath)) {
-    try {
-      const fm = frontmatter(readFileSync(runPath, "utf8"));
-      facts.run_status = fm.status || null;
-      facts.final_verdict = fm.final_verdict || null;
-    } catch { /* no fact */ }
-  }
-  return facts;
-}
-
-/**
- * The facts that contradict a claim, as human-readable fragments.
- *
- * TENSE DECIDES WHICH FACTS COUNT, and getting this wrong is what makes a hook get disabled.
- * An unfinished board and a red T0 both contradict "it is done". NEITHER of
- * them contradict "I am about to run the evaluator" — they are that sentence's PREMISE. Checking
- * a promise against them meant every healthy build round ended with the hook announcing that the
- * turn's own plan "disagrees with the facts", naming as evidence the very work the plan exists to
- * do. This file's header calls an always-on nag "exactly the annoyance that gets hooks disabled",
- * and the sibling future-tense detector in gate-zerowork.mjs is marked advisory for the same
- * reason. So a promise is contradicted by one thing only: a run that is already CLOSED, where the
- * promised work is never going to happen.
- *
- * The zero-work case that motivated future-tense detection — a session that dispatches, narrates
- * "it will: 1…", and stops having written nothing — is gate-zerowork.mjs's, which blocks on a
- * mechanical absence rather than on phrasing. It is not this hook's to double-report.
- *
- * @param {string} claim - The matched claim fragment from detectClaim().
- * @param {object} facts - Mechanical facts from gatherFacts().
- * @returns {string[]} Human-readable contradiction fragments; empty means the claim stands.
- */
-export function contradictions(claim, facts) {
-  const out = [];
-
-  if (isFutureClaim(claim)) {
-    if (facts.run_status === "shipped" || facts.run_status === "escalated") {
-      out.push(`run status is already "${facts.run_status}" — the run is closed, so the promised work cannot happen in it`);
-    }
-    return out;
-  }
-
-  if (facts.unfinished.length > 0) {
-    const named = facts.unfinished.slice(0, 5).join(", ");
-    out.push(`${facts.unfinished.length} board task(s) not done (${named}${facts.unfinished.length > 5 ? ", …" : ""})`);
-  }
-  if (facts.red_t0) out.push(`latest T0 verdict ${facts.red_t0} is red`);
-  if (facts.final_verdict === "fail") out.push("harness-run records final_verdict: fail");
-  if (/ship/i.test(claim || "") && MID_RUN.has(facts.run_status)) out.push(`run status is still "${facts.run_status}"`);
-  return out;
-}
-
-function lastAssistantFromTranscript(transcriptPath) {
-  try {
-    if (!transcriptPath || !existsSync(transcriptPath)) return null;
-    if (statSync(transcriptPath).size > 20 * 1024 * 1024) return null; // stay cheap
-    const lines = readFileSync(transcriptPath, "utf8").trim().split("\n");
-    for (let i = lines.length - 1; i >= 0 && i >= lines.length - 64; i--) {
-      let entry;
-      try { entry = JSON.parse(lines[i]); } catch { continue; }
-      const msg = entry?.message;
-      if (entry?.type === "assistant" || msg?.role === "assistant") {
-        const content = msg?.content ?? entry?.content;
-        if (typeof content === "string") return content;
-        if (Array.isArray(content)) {
-          return content.filter((b) => b?.type === "text").map((b) => b.text).join("\n");
-        }
-      }
-    }
-  } catch { /* unreadable transcript → no claim source */ }
-  return null;
-}
-
-async function main() {
-  await runHook("anti-rationalization", async () => {
-  const raw = await readStdin();
-  let p;
-  /** Stay silent — with the reason on the record (hooks/lib/decision.mjs). */
-  const defer = (reason, rule) => settle({ verdict: "allow", event: "Stop", cwd: p?.cwd, reason, rule });
-  try { p = JSON.parse(raw || "{}"); }
-  catch (e) { settle({ verdict: "error", event: "Stop", reason: `unparseable payload: ${e.message}` }); }
-
-  if (p.stop_hook_active) defer("stop_hook_active — never participate in a stop-hook loop", "loop-guard");
-
-  const cwd = p.cwd || process.cwd();
-  const slug = activeSlug(cwd);
-  if (!slug) defer("no active run — nothing to check the claim against", "no-run");
-
-  const message = typeof p.last_assistant_message === "string" && p.last_assistant_message
-    ? p.last_assistant_message
-    : lastAssistantFromTranscript(p.transcript_path);
-  const claim = detectClaim(message);
-  if (!claim) defer("the final message makes no completion claim", "no-claim");
-
-  const facts = gatherFacts(cwd, slug);
-  const contra = contradictions(claim, facts);
-  if (contra.length === 0) defer(`claim "${claim}" agrees with the run facts — inspected and permitted`, "claim-supported");
-
-  return {
-    verdict: "allow", event: "Stop", cwd, subject: slug, rule: "claim-contradicted", emit: true,
-    reason: `claim "${claim}" contradicted by ${contra.length} fact(s) — advisory note emitted, not a block`,
-    payload: {
-      systemMessage:
-        `anti-rationalization (advisory): the last message claims "${claim}" but run "${slug}" facts disagree — ` +
-        `${contra.join("; ")}. Not blocking (QA is a level-up, not a gate) — verify against the board and T0 before shipping.`,
-    },
-  };
-  });
-}
-
-if (isMain(import.meta.url)) {
-  main();
-}

package/hooks/compact-snapshot.mjs

@@ -1,47 +0,0 @@
-#!/usr/bin/env node
-// Compact snapshot — PreCompact hook (v1.2).
-//
-// PreCompact provably CANNOT inject context (no additionalContext/systemMessage channel), so
-// this hook is a pure side effect: when a harness run is mid-flight, freeze a RunSnapshot to
-// .shapeup/<slug>/run-snapshot.json before the conversation is compacted. It is the
-// audit anchor ("what did the files say the moment the summary was made?") and the fallback
-// hooks/session-rehydrate.mjs reads if live derivation ever throws post-compact.
-//
-// Never blocks compaction: the whole body is fail-open, exit 0 always.
-//
-// Contract: PreCompact stdin JSON { cwd, trigger: "manual"|"auto" }.
-
-// RECEIPTS (v1.5). This hook is the sharpest case for `hooks/lib/decision.mjs`: it is routinely
-// scored `Unfired` — 0 `PreCompact` events observed — and that score is UNOBTAINABLE, because
-// "never had to fire" and "never ran" produce identical evidence (exit 0, no output). With a
-// decision row per invocation the two become separable facts.
-
-import { deriveSnapshot, writeSnapshot } from "../skills/tech-lead/scripts/run-snapshot.mjs";
-import { isMain } from "../skills/tech-lead/scripts/lib/is-main.mjs";
-import { runHook, readStdin } from "./lib/decision.mjs";
-
-async function main() {
-  await runHook("compact-snapshot", async () => {
-    const raw = await readStdin();
-    let p;
-    try { p = JSON.parse(raw || "{}"); }
-    catch (e) { return { verdict: "error", event: "PreCompact", reason: `unparseable payload: ${e.message}` }; }
-    const cwd = p.cwd || process.cwd();
-    try {
-      const snapshot = deriveSnapshot(cwd);
-      if (!snapshot) return { verdict: "allow", event: "PreCompact", cwd, reason: "no active run — nothing to freeze" };
-      writeSnapshot(cwd, snapshot);
-      return {
-        verdict: "allow", event: "PreCompact", cwd, rule: "snapshot-written",
-        subject: snapshot.slug ?? null, reason: "RunSnapshot frozen before compaction",
-      };
-    } catch (e) {
-      // A snapshot failure must never block compaction — but it is now a fact rather than silence.
-      return { verdict: "error", event: "PreCompact", cwd, reason: `snapshot failed: ${e.message}` };
-    }
-  });
-}
-
-if (isMain(import.meta.url)) {
-  main();
-}

package/hooks/gate-deadline.mjs

@@ -1,147 +0,0 @@
-#!/usr/bin/env node
-// GATE D — DEADLINE. PreToolUse hook. The wall-clock breaker, enforced rather than requested.
-//
-// WHY THIS EXISTS (and it corrects an earlier diagnosis).
-//
-// A run killed at an external time cap looks exactly like a stall from outside: no acceptance, no
-// verdict, nothing to show. The natural reading is "it hung at a gate". Often the opposite is
-// true — the run was working, steadily, and was still working when the clock ran out.
-//
-// Both existing breakers count EVENTS (`round_budget` per round, `attempt_budget` per T0
-// attempt), so neither can observe that a single round has been running for half an hour. A run
-// can burn its whole budget with both breakers untouched.
-//
-// The cost of that is not the missing verdict. It is that an externally killed run ships NOTHING, not
-// even the scopes that were already green. A breaker that trips from the inside routes to GATE H
-// instead: census, baseline comparison, ship the part that works. Same clock, different ending.
-//
-// WHAT IT DENIES, AND WHAT IT DELIBERATELY DOES NOT. Past the deadline this denies dispatches
-// that START NEW WORK — `task-executor`. It never denies `spec-evaluator`, `scope-hammer` or
-// `qa-edge-hunter`, because a run past its deadline must still be able to
-// judge, hammer and close. A breaker that also blocked the exit would strand the run with green
-// scopes it could not ship, which is the failure it exists to prevent.
-//
-// Non-regression: with no `wall_clock_budget_s` in the receipt (the default) this defers
-// instantly. Everything ambiguous — no receipt, unparseable payload, missing timestamp — fails
-// open, because a gate that blocks legitimate runs gets disabled, and a disabled gate enforces
-// nothing.
-//
-// Contract: PreToolUse stdin JSON { tool_name, tool_input:{skill_name|skill, skill_args|args}, cwd }.
-// Deny via { hookSpecificOutput: { hookEventName, permissionDecision:"deny", permissionDecisionReason } }.
-
-import { readFileSync, readdirSync, existsSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import { evaluateBudget } from "../skills/tech-lead/scripts/budget-check.mjs";
-import { runHook, readStdin, settle } from "./lib/decision.mjs";
-import { localDir, localRoot } from "../skills/tech-lead/scripts/lib/paths.mjs";
-
-await runHook("gate-deadline", async () => {
-/** Fail-open, with the reason on the record (hooks/lib/decision.mjs). */
-const defer = (reason, rule) => settle({ verdict: "allow", event: "PreToolUse", tool: p?.tool_name ?? null, reason, rule });
-
-// Only these start new build work. Everything else in the harness is how a run ENDS, and must
-// stay reachable after the deadline.
-const STARTS_NEW_WORK = new Set(["task-executor"]);
-
-const raw = await readStdin();
-
-let p;
-try { p = JSON.parse(raw || "{}"); }
-catch (e) { settle({ verdict: "error", event: "PreToolUse", reason: `unparseable payload: ${e.message}` }); }
-
-if (p.tool_name !== "Skill") defer(`not a Skill call (${p.tool_name ?? "no tool_name"}) — out of scope`);
-
-const skillRaw = p.tool_input?.skill_name ?? p.tool_input?.skill ?? "";
-const skill = String(skillRaw).split(":").pop();
-if (!STARTS_NEW_WORK.has(skill)) defer(`Skill(${skill || "?"}) does not start new build work — a run past its deadline must still be able to close`, "not-new-work");
-
-const cwd = p.cwd || process.cwd();
-
-/** The active run's receipt, or null. Mirrors budget-check.mjs's discovery. */
-function findRun() {
-  const root = localDir(cwd);
-  if (!existsSync(root)) return null;
-  let slug = null;
-  try { slug = JSON.parse(readFileSync(join(root, "active-scope"), "utf8"))?.slug || null; } catch { /* scan instead */ }
-  const candidates = slug ? [slug] : (() => { try { return readdirSync(root); } catch { return []; } })();
-  for (const entry of candidates) {
-    const f = join(root, entry, "receipt.json");
-    if (!existsSync(f)) continue;
-    try { return { slug: entry, receipt: JSON.parse(readFileSync(f, "utf8")) }; } catch { /* not a run */ }
-  }
-  return null;
-}
-
-const run = findRun();
-if (!run) defer("no run receipt — nothing to time", "no-run");
-
-const budget = Number(run.receipt.config?.wall_clock_budget_s || 0) || null;
-if (!budget) defer("no wall_clock_budget_s in the receipt — breaker off (the default)", "breaker-off");
-
-const startedAt = Date.parse(run.receipt.started_at || "");
-if (Number.isNaN(startedAt)) defer("receipt has no parseable started_at — no clock, no claim", "no-clock");
-
-const state = evaluateBudget((Date.now() - startedAt) / 1000, budget);
-if (state.status !== "trip") defer(`within budget (${state.status}) — inspected and permitted`, "within-budget");
-
-// STICKY TRIP. The first version of this hook denied and said what to do instead; the orchestrator
-// then re-dispatched task-executor and was denied again, THIRTEEN times in one measured run. Each
-// denial costs a turn out of the budget the breaker exists to protect, so a re-triable denial
-// actively makes the failure it is preventing worse.
-//
-// The fix is to make the trip a FACT rather than an event: record it once, count the retries, and
-// escalate the language so the second denial cannot read like the first. A model that ignored
-// "route to GATE H" phrased gently is not helped by receiving the identical text again.
-const tripPath = join(localRoot(cwd, run.slug), "deadline-tripped.json");
-let denials = 0;
-try { denials = JSON.parse(readFileSync(tripPath, "utf8"))?.denials || 0; } catch { /* first trip */ }
-denials += 1;
-try {
-  writeFileSync(tripPath, JSON.stringify({
-    type: "deadline-trip", slug: run.slug, denials,
-    first_tripped_at: denials === 1 ? new Date().toISOString() : undefined,
-    elapsed_s: state.elapsed_s, budget_s: budget,
-  }, null, 2) + "\n");
-} catch { /* recording is best-effort; the denial still stands */ }
-
-const repeat = denials > 1
-  ? [
-      "",
-      `⛔ THIS IS DENIAL #${denials}. Re-dispatching task-executor will be denied every time, and each`,
-      "attempt spends wall-clock the run does not have. STOP TRYING TO BUILD. The only forward move is",
-      "GATE H, below. If you have already dispatched scope-hammer, wait for it and close the run.",
-    ]
-  : [];
-
-const reason = [
-  `✋ GATE D — DEADLINE BREAKER. Run "${run.slug}" has used ${state.elapsed_s}s of its ${budget}s wall-clock budget.`,
-  "",
-  "Dispatching task-executor would start work this run cannot finish. Denied — not because the work",
-  "is wrong, but because a run that is killed from outside ships nothing, including the scopes that",
-  "are already green.",
-  "",
-  "Go to GATE H instead:",
-  "",
-  "  Skill(scope-hammer, \"--breaker deadline --feature " + run.slug + "\")",
-  "",
-  "Scope-hammer runs the must-have census, compares the shippable subset against the BASELINE (never",
-  "against the ideal), and produces a cut list. Ship what is green; the rest becomes a raw idea for",
-  "the next Betting Table.",
-  "",
-  "spec-evaluator, scope-hammer and qa-edge-hunter are all still permitted — a run past its",
-  "deadline must still be able to judge, hammer, and close.",
-  ...repeat,
-].join("\n");
-
-return {
-  verdict: "deny", event: "PreToolUse", tool: "Skill", subject: skill, rule: "deadline-tripped",
-  reason: `wall-clock budget of ${budget}s exhausted — routing to GATE H`,
-  payload: {
-    hookSpecificOutput: {
-      hookEventName: "PreToolUse",
-      permissionDecision: "deny",
-      permissionDecisionReason: reason,
-    },
-  },
-};
-});

package/hooks/gate-l2.mjs

@@ -1,161 +0,0 @@
-#!/usr/bin/env node
-// GATE L2 — board-green check on the once-per-round EVAL. PreToolUse hook.
-//
-// ADVISORY SINCE ADR-0001, and the downgrade was a deliberate product decision rather than a
-// retreat. This hook used to hard-DENY the EVAL delegation (tech-lead → spec-evaluator) while the
-// task board was not green. It now permits the dispatch and says what it found.
-//
-// WHY. The board is LOCAL and per-machine (`.shapeup/<slug>/tasks/`), and the harness runs only on
-// the machine that invoked it — a teammate reads the committed design, they never resume someone
-// else's run. So this gate never protected a team boundary; it protected the operator from their
-// own agent, at the cost of denying a call the operator had asked for. The project chose the
-// signal over the denial.
-//
-// WHAT IS LOST, stated plainly: nothing now mechanically prevents an EVAL on a half-green board.
-// That is a defect this project has measured before (a run reached EVAL with 16/20 task files
-// still `status: ready`). The warning names the offending tasks; it cannot stop the call.
-//
-// WHAT IS RETAINED: the detection is unchanged. Both independent reads still run — per-task
-// frontmatter AND the board table — and the verdict is recorded as `warn`, which is its own row
-// in `decisions.jsonl` precisely so "permitted because green" and "permitted despite not green"
-// never collapse into the same fact (see hooks/lib/decision.mjs).
-//
-// Design (deliberate, conservative):
-//   • Scope — only ever looks at `Skill` → `spec-evaluator` in ROUND mode (`--single-pass`/
-//     `--feature`, no `--task`). A per-task eval (`--task TASK-NNN`) grades one task and is not
-//     in scope: the board-green rule is about the round. Anything else defers instantly.
-//   • Fail-OPEN whenever there is nothing to verify (no --spec, no board file, unparseable input,
-//     zero discoverable tasks) — it only speaks when it can prove the board is partial.
-//
-// Contract: PreToolUse stdin JSON { tool_name, tool_input:{skill_name, skill_args}, cwd, ... }.
-// Advises via { systemMessage } — never { permissionDecision: "deny" }.
-
-//
-// RECEIPTS (v1.5). Every `defer()` below now names WHICH of its fail-open conditions was met, and
-// records it (hooks/lib/decision.mjs). The fail-open direction is unchanged and defended above;
-// what changes is that "read the board and deferred" is no longer byte-identical to "did not run".
-
-import { readFileSync, existsSync, readdirSync } from "node:fs";
-import { resolve, join, basename, dirname } from "node:path";
-import { runHook, readStdin, settle } from "./lib/decision.mjs";
-// Aliased: the body binds a local `tasksDir` to whichever candidate actually holds a board.
-import { tasksDir as localTasksDir } from "../skills/tech-lead/scripts/lib/paths.mjs";
-
-await runHook("gate-l2", async () => {
-/** Allow normal permission flow — with the reason on the record. */
-const defer = (reason, rule) => settle({ verdict: "allow", event: "PreToolUse", tool: p?.tool_name ?? null, reason, rule });
-
-// 1. Read & parse the PreToolUse payload.
-const raw = await readStdin();
-let p;
-try { p = JSON.parse(raw || "{}"); }
-catch (e) { settle({ verdict: "error", event: "PreToolUse", reason: `unparseable payload: ${e.message}` }); }
-
-// 2. Only Skill → spec-evaluator is in scope.
-if (p.tool_name !== "Skill") defer(`not a Skill call (${p.tool_name ?? "no tool_name"}) — out of scope`);
-const skill = p.tool_input?.skill_name || "";
-const args = p.tool_input?.skill_args || "";
-if (skill !== "spec-evaluator") defer(`Skill(${skill || "?"}) is not the judge — out of scope`);
-
-// 3. Round mode only. Two shapes qualify:
-//    (a) legacy flags: --single-pass / --feature, without --task (per-task eval is not gated);
-//    (b) pure-skill envelope (v1.0): --order <WorkOrder> whose operation is "evaluate" — the
-//        round dispatch tech-lead compiles. Slug/spec come from the order itself.
-let orderSlug = null, orderSpec = null;
-const om = args.match(/--order(?:\s+|=)(?:"([^"]+)"|'([^']+)'|(\S+))/);
-if (om) {
-  try {
-    const order = JSON.parse(readFileSync(resolve(p.cwd || process.cwd(), om[1] || om[2] || om[3]), "utf8"));
-    if (order.worker === "spec-evaluator" && (order.operation || "evaluate") === "evaluate") {
-      orderSlug = String(order.order_id || "").split("/")[0] || order.payload?.feature || null;
-      orderSpec = order.payload?.spec_folder || null;
-    } else defer("order is for another job — not the round EVAL", "--order"); // not the round EVAL
-  } catch (e) {
-    if (e?.name === "HookDecision") throw e;
-    /* unreadable order → validate-envelope denies it; nothing to gate here */
-    defer(`order unreadable (${e.message}) — validate-envelope owns that denial`, "--order");
-  }
-}
-const hasTask = /--task(?:\s|=)/.test(args);
-const roundMode = orderSlug !== null || (!hasTask && (/--single-pass\b/.test(args) || /--feature(?:\s|=)/.test(args)));
-if (!roundMode) defer(hasTask ? "per-task eval — the board-green rule is about the round" : "not a round dispatch", "round-mode");
-
-// 4. Locate the board. Since v0.4.0 (Local Tasks Architecture) it lives under the LOCAL
-//    gitignored root `.shapeup/<slug>/tasks/`, NOT the committed spec dir — resolving only
-//    `<spec>/tasks/` made this hook silently fail-open on every v0.4.0+ run — EVAL proceeded with
-//    most of the board still `status: ready`.
-//    <slug> comes from --feature (the round invocation always carries it: tech-lead's eval plan
-//    is `--spec <path> --feature <slug> --single-pass`), falling back to the spec-path
-//    convention shapeup/<slug>/spec → parent dir name. `<spec>/tasks/` is kept as the
-//    legacy fallback so pre-v0.4.0 boards stay gated. A missing board on every candidate is a
-//    legitimate state — spec-evaluator v0.9 grades from the committed spec on machines that
-//    never generated a local board — so it stays fail-open.
-const m = args.match(/--spec(?:\s+|=)(?:"([^"]+)"|'([^']+)'|(\S+))/);
-const specPath = orderSpec || (m ? (m[1] || m[2] || m[3]) : null);
-if (!specPath && !orderSlug) defer("no --spec and no order slug — nothing to locate a board from", "no-spec");
-const cwd = p.cwd || process.cwd();
-const specDir = specPath ? resolve(cwd, specPath) : null;
-const fm = args.match(/--feature(?:\s+|=)(?:"([^"]+)"|'([^']+)'|(\S+))/);
-const slug = orderSlug || (fm && (fm[1] || fm[2] || fm[3])) ||
-  (basename(specDir) === "spec" ? basename(dirname(specDir)) : basename(specDir));
-const tasksDir = [localTasksDir(cwd, slug), ...(specDir ? [join(specDir, "tasks")] : [])]
-  .find((d) => existsSync(join(d, "_index.md")));
-// no board on this machine → nothing to verify, don't break the run
-if (!tasksDir) defer(`no board for "${slug}" on this machine — nothing to verify`, "no-board");
-const board = join(tasksDir, "_index.md");
-
-// 5. Assert the board is green, from two independent reads; fail-closed if EITHER shows unfinished
-//    work. (a) per-task frontmatter `status:` (the authoritative field); (b) the board table's
-//    status cell (what GATE L2 literally reads). Done = `status: done` / a ✅ in the row.
-const DONE_FRONTMATTER = /^status:\s*done\s*$/im;
-const NOT_DONE_MARK = /⬜|🔄|🚫|\b(ready|in-progress|blocked)\b/i;
-const unfinished = new Set();
-
-// (a) task files
-let sawTaskFile = false;
-try {
-  for (const f of readdirSync(tasksDir)) {
-    if (!/^TASK-[\w.-]+\.md$/i.test(f)) continue; // skip _index.md and non-task files
-    sawTaskFile = true;
-    const body = readFileSync(join(tasksDir, f), "utf8");
-    const fm = body.match(/^---\r?\n([\s\S]*?)\r?\n---/);
-    const id = (body.match(/^id:\s*(TASK-[\w.-]+)/im) || [])[1] || f.replace(/\.md$/, "");
-    if (!fm || !DONE_FRONTMATTER.test(fm[1])) unfinished.add(id);
-  }
-} catch { /* unreadable tasks dir → fall back to board table below */ }
-
-// (b) board table rows — any row naming a task whose status cell is not ✅/done.
-let sawBoardRow = false;
-for (const line of readFileSync(board, "utf8").split(/\r?\n/)) {
-  const idm = line.match(/\bTASK-[\w.-]+/);
-  if (!idm || !line.includes("|")) continue;
-  sawBoardRow = true;
-  const done = line.includes("✅") || /\bdone\b/i.test(line);
-  if (!done || NOT_DONE_MARK.test(line)) unfinished.add(idm[0]);
-}
-
-// 6. If neither source yielded a single task, there's nothing to assert → defer.
-if (!sawTaskFile && !sawBoardRow) defer("board directory holds zero tasks — nothing to assert", "empty-board");
-
-// 7. Verdict.
-// Board fully green → allow the EVAL. THIS is the row that used to be indistinguishable from the
-// gate never having run: same exit code, same empty stdout, opposite meaning.
-if (unfinished.size === 0) defer(`board green — ${sawTaskFile ? "task files" : "board rows"} all done, EVAL permitted`, "board-green");
-
-// Board is partial. Advise, and record it as its own verdict — a `warn` row is what keeps
-// "evaluated a non-green board" countable after the denial was removed (ADR-0001).
-const list = [...unfinished].sort().join(", ");
-return {
-  verdict: "warn", event: "PreToolUse", tool: "Skill", subject: slug, rule: "board-not-green",
-  reason: `${unfinished.size} unfinished task(s): ${list}`,
-  payload: {
-    systemMessage:
-      `⚠ GATE L2 — the board is NOT green and the EVAL is proceeding anyway.\n` +
-      `Unfinished (${unfinished.size}): ${list}\n` +
-      `EVAL is designed to run once per round, after every task is done. A verdict taken now grades ` +
-      `a partial board, so a PASS does not mean the feature is complete — it means the finished part ` +
-      `passed. Route back to BUILD (task-executor) to close these, or use --task for a deliberate ` +
-      `single-task check.`,
-  },
-};
-});