fix(cursor): native adapter hardening

CHANGELOG.md

@@ -4,6 +4,12 @@ All notable changes to `@fusengine/harness`. Format: [Keep a Changelog](https://
 
 ## [Unreleased]
 
+## [0.1.92] - 2026-09-02
+
+### Fixed
+
+- **Cursor native adapter hardening** (`src/adapters/cursor/`, `src/runtime/handle.ts`, `src/runtime/lifecycle/{rules-root,aipilot/dispatch-aipilot,failure-lesson}.ts`) — plugin-root discovery now follows an explicit precedence (`CURSOR_PLUGIN_ROOT` → `CLAUDE_PLUGIN_ROOT` → cwd marker) with a stderr diagnostic on miss, and project cwd resolution honors `CURSOR_PROJECT_DIR`/`CLAUDE_PROJECT_DIR`. Scope dispatchers now receive a canonical payload projection (`tool_name`, `cwd`, `tool_input` coerced string→object), and `MCP:<tool>` names are canonicalized via a closed server table. The doc-cache gate is now reachable on `beforeMCPExecution` under a cursor-only guard. Added per-event native response schemas with an exhaustive renderer, plus an `additional_context` 10,000-char cap backed by a session-level budget ledger (keyed `session|event|generation|tool_use_id`, fail-open, idempotent truncation). Covered by 23 provenance-labelled fixtures and byte-level stdout tests; test hygiene improved (tmp cwd/HOME, env restore, hex-free nonce for a pre-existing flake).
+
 ## [0.1.91] - 2026-09-01
 
 ### Fixed

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@fusengine/harness",
-	"version": "0.1.91",
+	"version": "0.1.92",
 	"description": "Harness-agnostic toolkit for AI coding agents: runtime harness detection (Claude Code, Codex, Cursor, Cline, Gemini, Aider...), pure policy core (env config, project/framework detection, SOLID/file-size limits, APEX freshness, guard patterns, portable prompts), cache, project memory, ref routing, state/locks, statusline, per-harness adapters (Claude/Cursor/Cline/Gemini) and a cli-mode harness-check binary. Bun-native, with a built dist for Node + bundlers.",
 	"type": "module",
 	"module": "src/index.ts",

src/adapters/cursor/context-budget.ts

@@ -0,0 +1,144 @@
+/**
+ * @module context-budget
+ * Cursor-only shared `additional_context` budget registry. Cursor 3.18.25
+ * runs every hook plugin configured on an event in its OWN process, then
+ * merges their `additional_context` outputs with a 9-char `"\n\n---\n\n"`
+ * separator and drops the WHOLE merge past 10,000 UTF-16 units — so no
+ * single process can know the total by itself. This module gives every
+ * plugin's process a shared, best-effort view of that total via a small
+ * JSON registry file under the project's state dir (see `../../runtime/paths.ts`),
+ * keyed by `${sessionId}|${event}|${generationId ?? ""}|${toolUseId ?? ""}`
+ * (one key per merge group — Cursor merges preToolUse/postToolUse/
+ * postToolUseFailure PER TOOL CALL, so `toolUseId` joins the key on those
+ * three events), with entries older than 10s ignored (concurrent hooks on one
+ * event fire within the same second). Best-effort, fail-open throughout: any
+ * I/O or JSON error degrades to "no shared budget", i.e. the flat
+ * per-response cap in `./context-limit.ts` alone — never a thrown error, and
+ * never a Cursor-side regression.
+ */
+import { existsSync, readFileSync } from "node:fs";
+import { join } from "node:path";
+import { atomicWrite } from "../../util/json-io";
+import {
+  ADDITIONAL_CONTEXT_LIMIT, TRUNCATION_MARKER, additionalContextLength, capAdditionalContext, omitAdditionalContext,
+} from "./context-limit";
+import type { CursorBudgetContext } from "./interfaces/context-budget";
+
+const REGISTRY_FILE = "cursor-context-budget.json";
+/** Matches Cursor 3.18.25's observed `"\n\n---\n\n"` merge separator length. */
+const SEPARATOR_LENGTH = 9;
+const ENTRY_WINDOW_MS = 10_000;
+/** Below this, a truncated value would carry more marker than budget — omit the field instead. */
+const OMIT_THRESHOLD = TRUNCATION_MARKER.length + 100;
+
+interface BudgetEntry {
+  at: number;
+  length: number;
+}
+type BudgetRegistry = Record<string, BudgetEntry[]>;
+
+function isRegistry(value: unknown): value is BudgetRegistry {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+function registryPath(stateDir: string): string {
+  return join(stateDir, REGISTRY_FILE);
+}
+
+function budgetKey(ctx: Pick<CursorBudgetContext, "sessionId" | "event" | "generationId" | "toolUseId">): string {
+  return `${ctx.sessionId}|${ctx.event}|${ctx.generationId ?? ""}|${ctx.toolUseId ?? ""}`;
+}
+
+function loadRegistry(path: string): BudgetRegistry {
+  try {
+    if (!existsSync(path)) return {};
+    const parsed: unknown = JSON.parse(readFileSync(path, "utf8"));
+    return isRegistry(parsed) ? parsed : {};
+  } catch {
+    return {};
+  }
+}
+
+function freshEntries(entries: BudgetEntry[] | undefined, now: number): BudgetEntry[] {
+  return (entries ?? []).filter((entry) => now - entry.at <= ENTRY_WINDOW_MS);
+}
+
+/** Sum of a key's fresh entry lengths plus the separators already joining them. */
+function consumed(entries: BudgetEntry[]): number {
+  return entries.reduce((total, entry) => total + entry.length, 0) + SEPARATOR_LENGTH * Math.max(0, entries.length - 1);
+}
+
+/** {@link reserveAdditionalContext} input: budget context plus the length the caller wants to emit. */
+export type ReserveInput = CursorBudgetContext & { wanted: number };
+/** {@link recordAdditionalContext} input: budget context plus the length actually emitted. */
+export type RecordInput = CursorBudgetContext & { emitted: number };
+
+/**
+ * Reserve room in the shared budget for one hook's `additional_context`
+ * contribution to one (session, event, generation) merge group. `wanted` is
+ * accepted for a symmetric call shape with {@link recordAdditionalContext}
+ * but does not shrink `allowed` itself — the ceiling only depends on what
+ * OTHER entries already hold; a smaller `wanted` simply means the caller
+ * won't need all of it. Best-effort, fail-open: any I/O/JSON error returns
+ * the full flat ceiling, as if no other plugin had run.
+ * @param input - Registry location, reservation key, and the wanted length.
+ */
+export function reserveAdditionalContext(input: ReserveInput): { allowed: number } {
+  try {
+    const now = input.now ?? Date.now();
+    const registry = loadRegistry(registryPath(input.stateDir));
+    const fresh = freshEntries(registry[budgetKey(input)], now);
+    const separator = fresh.length > 0 ? SEPARATOR_LENGTH : 0;
+    return { allowed: Math.max(0, ADDITIONAL_CONTEXT_LIMIT - consumed(fresh) - separator) };
+  } catch {
+    return { allowed: ADDITIONAL_CONTEXT_LIMIT };
+  }
+}
+
+/**
+ * Record the length actually emitted for one reservation, best-effort.
+ * Prunes every key's stale entries while it holds the write so the registry
+ * file stays bounded. Silently no-ops on any I/O error (fail-open).
+ * @param input - Registry location, reservation key, and the emitted length.
+ */
+export function recordAdditionalContext(input: RecordInput): void {
+  try {
+    const now = input.now ?? Date.now();
+    const path = registryPath(input.stateDir);
+    const registry = loadRegistry(path);
+    const pruned: BudgetRegistry = {};
+    for (const [key, entries] of Object.entries(registry)) {
+      const fresh = freshEntries(entries, now);
+      if (fresh.length > 0) pruned[key] = fresh;
+    }
+    const key = budgetKey(input);
+    pruned[key] = [...(pruned[key] ?? []), { at: now, length: input.emitted }];
+    atomicWrite(path, JSON.stringify(pruned));
+  } catch {
+    // Best-effort: a lost entry only makes the NEXT reservation over-generous
+    // (never under), which is the safe direction to fail in.
+  }
+}
+
+/**
+ * Cap a Cursor stdout JSON's `additional_context` against the shared budget
+ * instead of the flat per-response ceiling alone. Falls back to the plain
+ * cap (`./context-limit.ts`), unbudgeted, when `budget` is `undefined` or
+ * the stdout carries no `additional_context` at all.
+ * @param stdout - A native Cursor JSON stdout candidate.
+ * @param budget - Shared budget context, or `undefined` to skip it.
+ */
+export function capAdditionalContextWithBudget(stdout: string, budget: CursorBudgetContext | undefined): string {
+  if (!budget) return capAdditionalContext(stdout);
+  const wanted = additionalContextLength(stdout);
+  if (wanted === 0) return stdout;
+  const { allowed } = reserveAdditionalContext({ ...budget, wanted });
+  if (allowed < OMIT_THRESHOLD) {
+    process.stderr.write(`[fuse-harness] cursor: additional_context budget exhausted for ${budget.event} (allowed=${allowed})\n`);
+    return omitAdditionalContext(stdout);
+  }
+  const limit = Math.min(ADDITIONAL_CONTEXT_LIMIT, allowed);
+  const capped = capAdditionalContext(stdout, limit);
+  recordAdditionalContext({ ...budget, emitted: additionalContextLength(capped) });
+  return capped;
+}

src/adapters/cursor/context-limit.ts

@@ -0,0 +1,115 @@
+/**
+ * @module context-limit
+ * Cursor 3.18.25's `hooks-carriers` drops an `additional_context` carrier
+ * once `o.length>1e4` — but `o` is the MERGED text of every hook's
+ * `additional_context` for that event (concatenated with `"\n\n---\n\n"`
+ * before the 10,000-char check), not this harness's response in isolation.
+ * Capping our own contribution at {@link ADDITIONAL_CONTEXT_LIMIT} is
+ * therefore the LAST-RESORT guard, not the real protection: on its own it
+ * only proves OUR piece stays under 10,000, while the total across every
+ * hook plugin configured on the same event can still exceed it and get
+ * dropped wholesale — measured at ~8,400 chars on `sessionStart` from core
+ * plugins alone, close enough to the ceiling that one more plugin tips it
+ * over. The actual protection is the cross-process shared budget registry
+ * in `./context-budget.ts` (Cursor id only), which reserves a slice of the
+ * 10,000 ceiling per (session, event, generation) key BEFORE calling
+ * {@link truncateAdditionalContext} here with the reserved amount instead of
+ * the flat {@link ADDITIONAL_CONTEXT_LIMIT} — this module stays a pure,
+ * budget-agnostic primitive so it keeps working unbudgeted (its historical,
+ * still-correct behavior) wherever no budget context is available. The
+ * limit unit is UTF-16 code units (`String.prototype.length`), matching
+ * `value.length` here exactly. Only 5 events carry `additional_context`
+ * through this carrier — sessionStart, beforeSubmitPrompt, preToolUse,
+ * postToolUse, postToolUseFailure — subagentStart/subagentStop use a
+ * different, unlimited channel. "Drops silently" also only holds when no
+ * `failClosed: true` hook is declared on that step/tool: with one declared,
+ * an oversized carrier REJECTS the tool call instead of being dropped quiet.
+ */
+
+/** Cursor's hard `additional_context` character ceiling. */
+export const ADDITIONAL_CONTEXT_LIMIT = 10_000;
+
+/** Suffix appended by {@link truncateAdditionalContext} once a value is cut. */
+export const TRUNCATION_MARKER = "\n[fuse-harness] additional_context truncated to Cursor's 10000-char limit";
+
+function isPlainObject(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+/**
+ * Truncate a string to end with {@link TRUNCATION_MARKER} once its length
+ * exceeds `limit`. Leaves shorter values untouched. Idempotent under a
+ * SHRINKING `limit` across repeated calls (e.g. an unbudgeted flat-cap pass
+ * followed by a budgeted re-cap of the same stdout — see `./respond.ts`'s
+ * `toCursorLifecycleResponse` doc): when `value` already ends with
+ * {@link TRUNCATION_MARKER}, that marker is stripped BEFORE re-slicing so the
+ * result carries exactly one marker instead of risking a duplicated/cut one.
+ * @param value - Candidate `additional_context` body.
+ * @param limit - Effective ceiling for this call (defaults to the flat
+ * {@link ADDITIONAL_CONTEXT_LIMIT}; a shared-budget caller passes a smaller,
+ * per-reservation value instead).
+ */
+export function truncateAdditionalContext(value: string, limit: number = ADDITIONAL_CONTEXT_LIMIT): string {
+  const alreadyMarked = value.endsWith(TRUNCATION_MARKER);
+  if (!alreadyMarked && value.length <= limit) return value;
+  if (limit <= TRUNCATION_MARKER.length) return TRUNCATION_MARKER.slice(0, Math.max(0, limit));
+  const base = alreadyMarked ? value.slice(0, value.length - TRUNCATION_MARKER.length) : value;
+  return base.slice(0, limit - TRUNCATION_MARKER.length) + TRUNCATION_MARKER;
+}
+
+/**
+ * Length of a Cursor stdout JSON's `additional_context` string field, or 0
+ * when the stdout is not JSON, has no such field, or that field isn't a
+ * string.
+ * @param stdout - A native Cursor JSON stdout candidate.
+ */
+export function additionalContextLength(stdout: string): number {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(stdout);
+  } catch {
+    return 0;
+  }
+  return isPlainObject(parsed) && typeof parsed.additional_context === "string" ? parsed.additional_context.length : 0;
+}
+
+/**
+ * Re-serialize a Cursor stdout string with its `additional_context` field
+ * dropped entirely — used once the shared budget has no room left even for
+ * a truncated marker. Returns the input byte-for-byte unchanged when it is
+ * not JSON or has no string `additional_context` field.
+ * @param stdout - A native Cursor JSON stdout candidate.
+ */
+export function omitAdditionalContext(stdout: string): string {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(stdout);
+  } catch {
+    return stdout;
+  }
+  if (!isPlainObject(parsed) || typeof parsed.additional_context !== "string") return stdout;
+  const { additional_context: _omitted, ...rest } = parsed;
+  return JSON.stringify(rest);
+}
+
+/**
+ * Re-serialize a Cursor stdout string with its `additional_context` field
+ * capped at `limit` characters. Returns the input byte-for-byte unchanged
+ * when it is not JSON, has no string `additional_context` field, or that
+ * field is already within the limit — so callers can wrap every return path
+ * unconditionally.
+ * @param stdout - A native Cursor JSON stdout candidate.
+ * @param limit - Effective ceiling for this call (see {@link truncateAdditionalContext}).
+ */
+export function capAdditionalContext(stdout: string, limit: number = ADDITIONAL_CONTEXT_LIMIT): string {
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(stdout);
+  } catch {
+    return stdout;
+  }
+  if (!isPlainObject(parsed) || typeof parsed.additional_context !== "string") return stdout;
+  const truncated = truncateAdditionalContext(parsed.additional_context, limit);
+  if (truncated === parsed.additional_context) return stdout;
+  return JSON.stringify({ ...parsed, additional_context: truncated });
+}

src/adapters/cursor/context.ts

@@ -30,17 +30,36 @@ function contains(root: string, filePath: string): boolean {
   return rel === "" || (!rel.startsWith("..") && !isAbsolute(rel));
 }
 
-/** Select Cursor's project scope without replacing a valid payload cwd. */
+/**
+ * Select Cursor's project scope without replacing a valid payload cwd. Order:
+ * payload `cwd` -> longest workspace root containing `filePath` ->
+ * `workspaceRoots[0]` -> `CURSOR_PROJECT_DIR` env -> `CLAUDE_PROJECT_DIR` env
+ * -> `fallback`. Both env vars are validated the same way as any other Cursor
+ * path (`cursorAbsolutePath`: absolute, NUL-free, realpath-resolved), so an
+ * unset or malformed value is silently skipped rather than trusted.
+ * @param cwd - Cursor payload `cwd`, when present.
+ * @param workspaceRoots - Validated, deduped Cursor `workspace_roots`.
+ * @param filePath - The file the current event targets, when present.
+ * @param fallback - Caller-supplied last resort (never `process.cwd()`).
+ * @param env - Environment (defaults to `process.env`).
+ * @returns The resolved project root.
+ */
 export function cursorProjectCwd(
   cwd: string | undefined,
   workspaceRoots: readonly string[],
   filePath: string | undefined,
   fallback: string,
+  env: Record<string, string | undefined> = process.env,
 ): string {
   if (cwd) return cwd;
   if (filePath) {
     const matches = workspaceRoots.filter((root) => contains(root, filePath));
     if (matches.length > 0) return matches.sort((a, b) => b.length - a.length)[0]!;
   }
-  return workspaceRoots[0] ?? fallback;
+  if (workspaceRoots[0]) return workspaceRoots[0];
+  const fromCursorEnv = cursorAbsolutePath(env.CURSOR_PROJECT_DIR);
+  if (fromCursorEnv) return fromCursorEnv;
+  const fromClaudeEnv = cursorAbsolutePath(env.CLAUDE_PROJECT_DIR);
+  if (fromClaudeEnv) return fromClaudeEnv;
+  return fallback;
 }

src/adapters/cursor/interfaces/context-budget.ts

@@ -0,0 +1,25 @@
+/**
+ * Reservation key + registry location for one hook invocation's slice of
+ * Cursor's shared, cross-process `additional_context` budget (see
+ * `../context-budget.ts`). `undefined` at a call site means "no shared
+ * budget available" — callers then fall back to the flat per-response cap.
+ */
+export interface CursorBudgetContext {
+  /** Project state directory the registry file lives under (see `defaultStateDir`). */
+  stateDir: string;
+  /** Cursor `session_id` (its `conversation_id`). */
+  sessionId: string;
+  /** Raw Cursor `hook_event_name` (e.g. `"sessionStart"`). */
+  event: string;
+  /** Cursor `generation_id`; absent on `sessionStart`/`workspaceOpen`. */
+  generationId?: string;
+  /**
+   * Cursor `tool_use_id`; present on preToolUse/postToolUse/postToolUseFailure
+   * — Cursor merges `additional_context` PER TOOL CALL for these events, not
+   * once per (session, event, generation), so this must join the key or
+   * concurrent tool calls in the same generation would wrongly share one slice.
+   */
+  toolUseId?: string;
+  /** Test seam: injectable clock (defaults to `Date.now()`). */
+  now?: number;
+}