@event4u/agent-config 9.27.0 → 9.29.0

package/src/scripts/hooks/delegation_nudge_hook.ts

@@ -0,0 +1,484 @@
+#!/usr/bin/env node
+/**
+ * Delegation nudge — `user_prompt_submit` concern (F3-lite, road-to-orchestrator-
+ * discipline-carriers Phase 4).
+ *
+ * Deterministic, no-LLM carrier for the `delegation-policy` rule's "decompose
+ * and dispatch" obligation. `classifyTask` (`_lib/auto_dispatch.ts`) has zero
+ * production callers today — its only importer is the `routing:doctor`
+ * diagnostic — so the delegate-by-default rule never actually sees a real
+ * prompt. This concern is the missing signal producer: it extracts cheap,
+ * regex-only structural signals from the SUBMITTED prompt text (enumerated
+ * file lists, "for each"/"alle …" shapes, explicit slice counts, ordered-plan
+ * markers, multi-deliverable conjunctions), feeds them to the judgment
+ * ladder (`_lib/judgment_ladder.ts`, which wraps `classifyTask` — see the
+ * JUDGMENT-LADDER CITATION note below) together with the REAL activation
+ * gate (the emergency incident switch +
+ * the host-capability manifest, resolved via `probeHostCapabilities(hostId)`
+ * — the F5 committed registry merged with observable environment facts;
+ * always-on orchestration removed the `subagents.host_capabilities` settings
+ * override — the same source `routing_doctor.collect_orchestration` uses),
+ * and — only on a positive
+ * `do-in-parallel` / `do-in-steps` verdict — injects
+ * ONE line naming the mode, the slice count, and a tier recommendation for
+ * the slices (`resolveSubagentRouting`'s first production caller, applied
+ * with a conservative "lite, downshifted" default).
+ *
+ * ANTI-CANARY CONDITION (roadmap 4.2). Silence is the default: no enumerated
+ * signal → no injection, not even a "nothing to delegate here" line. This is
+ * deliberate — a per-turn reminder that fires on every turn is exactly the
+ * canary shape that measured a 24/29 miss rate; this concern only ever speaks
+ * when it has a concrete, falsifiable verdict to report.
+ *
+ * JUDGMENT-LADDER CITATION (road-to-always-on-orchestration Phase 2.4). The
+ * verdict now comes from `_lib/judgment_ladder.ts::classifyLadder` (which
+ * WRAPS `classifyTask` rather than replacing it) instead of calling
+ * `classifyTask` directly, and the injected line names the resolved rung
+ * (`rung-2: dispatch 6 lite slices …`). This concern stays scoped to the
+ * SUBAGENT dispatch rungs (1/2) — a rung-0 script verdict, a rung-3/4
+ * team/council escalation, and the silent ∅ (ask/in-session) verdicts all
+ * resolve to silence here exactly as before this concern cited the ladder
+ * at all; team and council get their own carrier once Phase 3/4 give them
+ * one.
+ *
+ * DELIVERY PATH — fixed to mirror `language_mirror_hook.ts` (found while
+ * building this concern; the original `{decision:"allow", context:…}` shape
+ * below is superseded). `context_blocks` forwarding in `dispatch_hook.ts` is
+ * restricted to the `session_start` event by the code's own comment
+ * ("session_start context forwarding … All other events keep the swallow-
+ * stdout contract unchanged"). Empirically verified against the real
+ * dispatcher (`dispatch_entry.ts`, platform `claude`): a `user_prompt_submit`
+ * concern that returns `{decision:"allow", context: "…"}` at exit 0 produces
+ * ZERO stdout from the dispatcher — the `context` field is silently dropped,
+ * because `host_semantics.emitFor` returns `{exit:0, stdout:"", stderr:""}`
+ * unconditionally for severity `"allow"` (derived from the concern's own exit
+ * code, which is 0 here). The ONLY path that survives to
+ * `hookSpecificOutput.additionalContext` on `user_prompt_submit` is a concern
+ * that reports `reason`/`additional_context` and an exit code that reduces to
+ * `"warn"` (2) — `language_mirror_hook.ts`'s shipped pattern. This concern now
+ * follows that pattern exactly: `{decision:"warn", reason, additional_context}`
+ * at exit `EXIT_WARN` (2).
+ *
+ * WHY EXIT 2 NEVER BLOCKS. `host_semantics.emitFor`'s `severity === "warn"`
+ * branch returns `{exit: 0, stdout: claudeAdditionalContext(event, reason)}`
+ * UNCONDITIONALLY — it does not consult `CLAUDE_BLOCK_CAPABLE_EVENTS` at all
+ * (that check only gates the `"block"` severity branch). So on the verified
+ * `claude` platform the real process exit handed back to the host is always
+ * 0 for a warn verdict, on `user_prompt_submit` exactly as on every other
+ * event — the internal `EXIT_WARN` is dispatcher-internal bookkeeping, never
+ * the byte that reaches Claude Code. This is the same mechanism that makes
+ * `language_mirror_hook.ts` (which returns the same exit code on the same
+ * slot, on effectively every prompt carrying language markers) run in
+ * production today without ever blocking a turn.
+ *
+ * PLATFORM SCOPE — bound only on `claude` in the manifest (not the full
+ * `language-mirror` platform list). `host_semantics.VERIFIED_PLATFORMS`
+ * covers only `claude`, so `claude` is the one platform where the "never
+ * blocks" claim above is verified against documented, testable behaviour.
+ * `cowork` is DELIBERATELY NOT bound (F10, review) — `scripts/hooks/
+ * cowork-dispatcher.sh` discards the dispatcher's exit code and stdout
+ * unconditionally (`>/dev/null 2>&1 || true; exit 0`), and this concern's
+ * entire delivery depends on `additional_context` reaching the model via
+ * stdout at the verified exit-2 warn path — a delivery that trampoline makes
+ * structurally impossible regardless of platform. Binding it there would be
+ * dead weight: the concern still runs (settings read, classifier,
+ * activation resolution), still costs the dispatch, and its entire output
+ * is thrown away for nothing. Cursor / Cline / Windsurf / Gemini are
+ * deliberately NOT added here either: their trampolines were not inspected
+ * for the same discard property, and extending the exit-2 pattern to an
+ * unverified propagation path is exactly the speculative mapping
+ * `host_semantics.ts`'s own header calls out as "the same class of bug this
+ * module exists to remove". `language-mirror`'s existing broader binding is
+ * a pre-existing fact this change does not touch or relitigate.
+ */
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath, pathToFileURL } from "node:url";
+import { readHookStdin } from "./hook_stdin.js";
+import { load_agent_settings } from "../_lib/agent_settings.js";
+import { probeHostCapabilities } from "../_lib/host_capability.js";
+import type { ActivationInputs, Classification, TaskSignals } from "../_lib/auto_dispatch.js";
+import { classifyLadder, type LadderRung } from "../_lib/judgment_ladder.js";
+import { resolveSubagentRouting, type Tier } from "../_lib/subagent_routing.js";
+import { isSyntheticPrompt } from "../_lib/prompt_shape.js";
+
+const EXIT_ALLOW = 0;
+// Severity is taken from the EXIT CODE, not from the `decision` field in the
+// stdout payload — mirrors `language_mirror_hook.ts`'s own note (found the
+// same way: by tracing delivery, not by re-reading the unit tests). A warn
+// verdict is reported at exit 2 so `host_semantics.emitFor` reduces it to
+// severity `"warn"` and forwards `additional_context`; see the file header
+// for why this exit code never actually blocks the turn.
+const EXIT_WARN = 2;
+
+type JsonValue = string | number | boolean | null | JsonValue[] | { [k: string]: JsonValue };
+type JsonObject = { [k: string]: JsonValue };
+
+function _isObject(v: unknown): v is JsonObject {
+  return typeof v === "object" && v !== null && !Array.isArray(v);
+}
+
+// ── Signal extraction (cheap, regex-only — no LLM call) ──────────────────
+
+/**
+ * Enumerated file-like tokens: backtick-quoted or bare `name.ext` shapes.
+ * Counts UNIQUE matches — "a.ts a.ts" is one file mentioned twice, not two.
+ */
+const FILE_TOKEN_RE =
+  /`?\b[\w][\w.\-]{0,80}\.(ts|tsx|js|jsx|mjs|cjs|py|php|rb|go|rs|java|kt|swift|md|mdx|json|ya?ml|css|scss|less|html|vue|svelte|sql|sh|bash|c|cpp|h|hpp|cs|toml|xml)\b`?/gi;
+
+export function detectEnumeratedFiles(text: string): number {
+  const matches = text.match(FILE_TOKEN_RE) ?? [];
+  const unique = new Set(matches.map((m) => m.replace(/`/g, "").toLowerCase()));
+  return unique.size;
+}
+
+/**
+ * Explicit slice counts: "6 files", "3 modules", "5 Dateien" — a number
+ * immediately followed by a plural noun the classifier recognises as a
+ * countable unit of work.
+ */
+const EXPLICIT_COUNT_RE =
+  /\b(\d+)\s+(files?|modules?|components?|endpoints?|scripts?|skills?|rules?|tickets?|tests?|screens?|pages?|repos?|services?|classes?|functions?|methods?|slices?|tasks?|items?|targets?|routes?|controllers?|Dateien|Module|Tests|Komponenten|Skripte|Regeln|Tickets|Aufgaben|Klassen|Funktionen|Endpunkte)\b/i;
+
+export function detectExplicitSliceCount(text: string): number | null {
+  const m = EXPLICIT_COUNT_RE.exec(text);
+  if (!m) return null;
+  const n = Number.parseInt(m[1] ?? "", 10);
+  return Number.isFinite(n) && n > 0 ? n : null;
+}
+
+/**
+ * "for each" / "for every" / "alle …" / "jede(n|s|r) …" — a same-shape-repeated
+ * signal with no explicit count attached. Contributes a minimal slice count
+ * (2) when nothing more specific was found, since the shape itself implies
+ * "more than one" without saying how many.
+ */
+const FOR_EACH_RE =
+  /\bfor\s+each\b|\bfor\s+every\b|\beach\s+of\s+the\b|\bevery\s+one\s+of\b|\ball(?:e|en)\b|\bjed(?:e|er|es|en)\b|\bjeweils\b/i;
+
+export function detectForEachShape(text: string): boolean {
+  return FOR_EACH_RE.test(text);
+}
+
+/**
+ * Multi-deliverable conjunction: "X, Y, and Z" / "X, Y, und Z" — an
+ * Oxford-comma-style list of three or more items closed by "and"/"und"
+ * immediately after the LAST comma in a sentence. Deliberately requires the
+ * comma before the conjunction (>= 2 commas total): a looser, single-comma
+ * match ("A, B and C") would also match a stray "<item>, and <note>" aside
+ * that is not an enumeration at all. Known gap (documented, not fixed
+ * here): a German list that skips the comma before "und" ("Docs, Changelog
+ * und README") has only 1 comma and is NOT detected by this function —
+ * `detectForEachShape`'s "alle"/"jede…" patterns are the German-first signal
+ * for that shape instead.
+ */
+export function detectMultiDeliverableConjunction(text: string): number {
+  for (const sentence of text.split(/[.!?\n]+/)) {
+    const commaCount = (sentence.match(/,/g) ?? []).length;
+    if (commaCount < 2) continue;
+    const tail = sentence.slice(sentence.lastIndexOf(",") + 1);
+    if (/^\s*(and|und)\b/i.test(tail)) {
+      return commaCount + 1;
+    }
+  }
+  return 0;
+}
+
+/**
+ * Ordered-plan markers: "Step 1"/"Schritt 2", or lines opening with `1. `/`2)`.
+ * Two or more distinct markers is treated as an explicit ordered plan — one
+ * numbered reference alone ("see step 3") is not a plan shape.
+ */
+const STEP_WORD_RE = /\b(?:step|schritt)\s*\d+\b/gi;
+const STEP_LINE_RE = /^\s*\d+[.)]\s+\S/gm;
+
+/**
+ * Minimum bare numbered-list lines ("1. … \n2. …") before the LINE shape
+ * counts as an ordered plan (F7, review): a 2-line numbered list is common,
+ * ordinary prose structure ("1. rename a.ts\n2. rename b.ts" is not itself
+ * an enumerated work plan) and was firing as a false positive at the
+ * previous threshold of 2. The explicit "Step N"/"Schritt N" WORD marker
+ * (`STEP_WORD_RE`) is a much stronger, author-stated signal even at 2 — an
+ * author who writes "Step 1 … Step 2 …" is naming a plan on purpose — and
+ * keeps its own, lower threshold.
+ */
+const ORDERED_LINE_FLOOR = 3;
+const ORDERED_WORD_FLOOR = 2;
+
+export function detectOrderedPlan(text: string): { ordered: boolean; stepCount: number } {
+  const wordMatches = text.match(STEP_WORD_RE) ?? [];
+  const lineMatches = text.match(STEP_LINE_RE) ?? [];
+  const ordered = wordMatches.length >= ORDERED_WORD_FLOOR || lineMatches.length >= ORDERED_LINE_FLOOR;
+  const count = Math.max(wordMatches.length, lineMatches.length);
+  return { ordered, stepCount: count };
+}
+
+export interface ExtractedSignals {
+  signals: TaskSignals;
+  /** Slice count to surface in the injected line, matching whichever shape fired. */
+  sliceCountForLine: number;
+}
+
+/**
+ * Turn a raw submitted-prompt string into `TaskSignals` for `classifyTask`.
+ * Pure and total — never throws; an unparseable/empty prompt resolves to
+ * "no signal" (`size_estimate: 0`), which the size floor alone already
+ * routes to `in-session`.
+ */
+export function extractTaskSignals(text: string): ExtractedSignals {
+  const ordered = detectOrderedPlan(text);
+  if (ordered.ordered) {
+    return {
+      signals: {
+        parallelizable: null,
+        ordered_plan: true,
+        independent_slices: 0,
+        size_estimate: Math.max(ordered.stepCount, 2),
+      },
+      sliceCountForLine: ordered.stepCount,
+    };
+  }
+
+  const fileCount = detectEnumeratedFiles(text);
+  const explicit = detectExplicitSliceCount(text) ?? 0;
+  const conjunction = detectMultiDeliverableConjunction(text);
+  const forEach = detectForEachShape(text);
+
+  // F7 (review): a bare file-token count below 3 is common and NOT itself
+  // evidence of independent, parallelizable work — "rename a.ts to b.ts"
+  // names 2 files but is one mechanical edit, not two delegable slices. The
+  // explicit "N files/Dateien" phrase (`detectExplicitSliceCount`) is a
+  // stronger, author-stated signal and keeps its existing floor of N>=2.
+  const fileSignal = fileCount >= 3 ? fileCount : 0;
+
+  let slices = Math.max(fileSignal, explicit, conjunction);
+  if (slices === 0 && forEach) {
+    slices = 2;
+  }
+
+  const parallelizable: TaskSignals["parallelizable"] =
+    fileSignal >= 3 ? "files" : forEach ? "independent" : null;
+
+  return {
+    signals: {
+      parallelizable,
+      ordered_plan: false,
+      independent_slices: slices,
+      size_estimate: slices > 0 ? Math.max(slices, 2) : 0,
+    },
+    sliceCountForLine: slices,
+  };
+}
+
+// ── Activation resolution (mirrors routing_doctor.collect_orchestration) ──
+
+/**
+ * `hostId` is the caller-supplied platform identifier (the envelope's own
+ * `platform` field, e.g. `"claude"`) — `probeHostCapabilities` resolves the
+ * committed registry row for it, merged with observable environment facts.
+ * Always-on orchestration (road-to-always-on-orchestration Phase 1):
+ * capability is a fact about the host, never a settings decision, so there
+ * is no more `subagents.host_capabilities` override to read here.
+ */
+export function resolveActivation(
+  workspace_root: string,
+  hostId: string | null | undefined,
+): {
+  activation: ActivationInputs;
+  downshift: boolean;
+  separate_quota_pool: boolean;
+  /** Host-capability manifest's `agent_teams` field — the ladder's rung-3 precondition. */
+  agentTeams: boolean;
+} {
+  const settings = load_agent_settings({ cwd: workspace_root }) as Record<string, unknown>;
+  const emergency = (settings["emergency"] ?? {}) as Record<string, unknown>;
+  const halted = emergency["orchestration_halt"] === true;
+  const host_manifest = probeHostCapabilities(hostId);
+  const sub = (settings["subagents"] ?? {}) as Record<string, unknown>;
+  const downshift = sub["downshift"] !== false;
+  return {
+    activation: { halted, subagent_spawn: host_manifest.subagent_spawn },
+    downshift,
+    separate_quota_pool: host_manifest.separate_quota_pool,
+    agentTeams: host_manifest.agent_teams,
+  };
+}
+
+/**
+ * Tier recommendation for the (not-yet-inspected) slices: `resolveSubagentRouting`'s
+ * first production caller. Conservative "lite, downshifted" default — the
+ * concern has no per-slice type classification available at prompt-submit
+ * time, so it recommends the cheapest starting point and lets the downshift
+ * setting (shipped default: on) decide whether that survives.
+ */
+export function recommendSliceTier(downshift: boolean, separate_quota_pool: boolean): Tier {
+  const decision = resolveSubagentRouting({
+    task_tier: "lite",
+    session_tier: "high",
+    downshift,
+    quota_arbitrage: false,
+    model_map: {},
+    separate_quota_pool,
+  });
+  return decision.tier;
+}
+
+/**
+ * `rung` names the resolved judgment-ladder rung (2.4): `null` renders as
+ * `rung-?` — this concern only ever calls it for a `subagent` verdict
+ * (rung 1 or 2), so `null` is unreachable in production and exists only so
+ * the function stays total for direct unit-test callers.
+ */
+export function buildNudgeLine(
+  rung: LadderRung,
+  classification: Classification,
+  sliceCount: number,
+  tier: Tier,
+): string {
+  const unit = sliceCount === 1 ? "slice" : "slices";
+  const rungLabel = rung === null ? "rung-?" : `rung-${rung}`;
+  // Rung 1 (single bounded slice) carries no do-in-steps/do-in-parallel
+  // mode of its own — `classification.mode` is `null` for that shape.
+  const modeLabel = classification.mode ?? "single-slice";
+  return (
+    `<delegation-nudge>${rungLabel}: dispatch ${modeLabel} ` +
+    `(${sliceCount} ${unit}, ${tier} tier recommended). Consider dispatching via ` +
+    `subagent-orchestration instead of doing every slice in-session — ` +
+    `${classification.reason}.</delegation-nudge>`
+  );
+}
+
+export interface NudgeResult {
+  classification: Classification;
+  sliceCount: number;
+  tier: Tier;
+  /** The judgment-ladder rung this verdict resolved to (2.4) — 1 or 2 only, this concern's scope. */
+  rung: LadderRung;
+}
+
+/**
+ * Full pipeline: prompt text + workspace root + host id → a positive nudge
+ * verdict, or `null` when nothing should be injected (no signal, a rung-0/3/4
+ * verdict out of this concern's scope, activation gate closed, or an
+ * internal error — all degrade to the same silent outcome).
+ */
+export function classifyPrompt(
+  prompt: string,
+  workspace_root: string,
+  hostId: string | null | undefined,
+): NudgeResult | null {
+  try {
+    const { signals, sliceCountForLine } = extractTaskSignals(prompt);
+    const { activation, downshift, separate_quota_pool, agentTeams } = resolveActivation(
+      workspace_root,
+      hostId,
+    );
+    const ladder = classifyLadder({ taskText: prompt, signals, activation, agentTeams });
+    if (ladder.verdict !== "subagent") {
+      return null; // rung-0 script, rung-3/4 team/council, or ∅ ask/in-session — out of scope here
+    }
+    const tier = recommendSliceTier(downshift, separate_quota_pool);
+    // Rung 1 (single bounded slice) has no explicit slice count of its own —
+    // extractTaskSignals never produces one for that shape — so the line
+    // always names it as exactly one slice.
+    const sliceCount = ladder.rung === 1 ? 1 : sliceCountForLine;
+    const classification: Classification = {
+      delegable: true,
+      action: "dispatch",
+      mode: ladder.mode ?? null,
+      reason: ladder.reason,
+    };
+    return { classification, sliceCount, tier, rung: ladder.rung };
+  } catch {
+    return null; // classifier/settings error → silence, never a crash-to-block
+  }
+}
+
+function _extractPrompt(payload: JsonObject): string {
+  for (const key of ["prompt", "userPrompt", "user_prompt", "message", "text"]) {
+    const v = payload[key];
+    if (typeof v === "string" && v.trim()) {
+      return v;
+    }
+  }
+  return "";
+}
+
+function _workspaceRoot(envelope: JsonObject): string {
+  const v = envelope["workspace_root"];
+  if (typeof v === "string" && v) {
+    return v;
+  }
+  return process.cwd();
+}
+
+export function main(): number {
+  let envelope: JsonValue = {};
+  try {
+    const raw = readHookStdin();
+    envelope = raw.trim() ? (JSON.parse(raw) as JsonValue) : {};
+  } catch {
+    return EXIT_ALLOW; // malformed envelope — never block
+  }
+  const env: JsonObject = _isObject(envelope) ? envelope : {};
+
+  const event = env["event"];
+  const slot = typeof event === "string" && event !== "" ? event : "user_prompt_submit";
+  if (slot !== "user_prompt_submit") {
+    return EXIT_ALLOW; // this concern only reads the submitted prompt
+  }
+
+  const payload = _isObject(env["payload"]) ? (env["payload"] as JsonObject) : env;
+  const prompt = _extractPrompt(payload);
+  if (!prompt) {
+    return EXIT_ALLOW;
+  }
+  if (isSyntheticPrompt(prompt)) {
+    return EXIT_ALLOW; // harness-generated turn, not a human-authored task
+  }
+
+  const root = _workspaceRoot(env);
+  const platform = env["platform"];
+  const hostId = typeof platform === "string" && platform ? platform : null;
+  const result = classifyPrompt(prompt, root, hostId);
+  if (!result) {
+    return EXIT_ALLOW; // anti-canary default: no verdict → no output
+  }
+
+  process.stdout.write(
+    `${JSON.stringify({
+      decision: "warn",
+      reason: `delegation-nudge: rung-${result.rung} verdict (${result.sliceCount} slices, ${result.tier} tier)`,
+      additional_context: buildNudgeLine(result.rung, result.classification, result.sliceCount, result.tier),
+    })}\n`,
+  );
+  return EXIT_WARN;
+}
+
+// Bundle-safety: never auto-run when inlined into an esbuild bundle, where
+// every module shares the bundle's `import.meta.url` (see cmd_migrate.ts).
+declare const __AGENT_CONFIG_BUNDLE__: boolean | undefined;
+function _isCliEntry(): boolean {
+  if (typeof __AGENT_CONFIG_BUNDLE__ !== "undefined" && __AGENT_CONFIG_BUNDLE__) {
+    return false;
+  }
+  if (process.argv[1] === undefined) {
+    return false;
+  }
+  const argvUrl = pathToFileURL(path.resolve(process.argv[1])).href;
+  if (import.meta.url === argvUrl) {
+    return true;
+  }
+  try {
+    const here = fs.realpathSync(fileURLToPath(import.meta.url));
+    const argv = fs.realpathSync(path.resolve(process.argv[1]));
+    return here === argv;
+  } catch {
+    return false;
+  }
+}
+
+if (_isCliEntry()) {
+  process.exit(main());
+}