@xaccefy/pi-casefile 0.6.2 → 0.7.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xaccefy/pi-casefile",
-  "version": "0.6.2",
+  "version": "0.7.0",
   "description": "Offensive security case tracker for Pi Agent — bug bounties, CTFs, security audits",
   "keywords": [
     "pi-package",
@@ -35,6 +35,7 @@
     "src/ledger.ts",
     "src/workflow.ts",
     "src/poc-runner.ts",
+    "src/scratchpad.ts",
     "src/sqlite-compat/index.ts",
     "skills",
     "README.md",

package/src/index.ts

@@ -1,7 +1,7 @@
 /**
  * Casefile — offensive security case tracker for Pi.
  *
- * Tools: CaseAdd, CaseUpdate, PromoteFinding, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseReport
+ * Tools: CaseAdd, CaseUpdate, PromoteFinding, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseReport, ScratchpadInit, ScratchpadResume, ScratchpadCheckpoint, ScratchpadWrite, ScratchpadRead, ScratchpadPhaseDone, ScratchpadClear
  * Command: /casefile — interactive dashboard
  * Event: before_agent_start — injects cyber workflow (+ active case list) once per user prompt
  */
@@ -46,6 +46,17 @@ import {
   writeCaseReport,
 } from "./ledger.ts";
 import { type PocRun, runPoc } from "./poc-runner.ts";
+import {
+  type ScratchpadPhase,
+  type ScratchpadResume,
+  scratchpad_checkpoint,
+  scratchpad_clear,
+  scratchpad_init,
+  scratchpad_phase_done,
+  scratchpad_read,
+  scratchpad_resume,
+  scratchpad_write,
+} from "./scratchpad.ts";
 import { STATIC_CYBER_WORKFLOW } from "./workflow.ts";
 
 // ── Schemas ───────────────────────────────────────────────────────────
@@ -216,6 +227,88 @@ const ReportSchema = Type.Object(
   { additionalProperties: false },
 );
 
+// ── Tool: Scratchpad ─────────────────────────────────────────────────
+//
+// The scratchpad is the pipeline's crash-recoverable artifact store.
+// The casefile owns state transitions; the scratchpad owns artifacts
+// (recon maps, trace outputs, verification logs). Resume re-reads
+// artifacts; it does not re-run completed phases (idempotent).
+
+const SCRATCHPAD_PHASES = [
+  "recon",
+  "hunt",
+  "gapfil",
+  "trace",
+  "skeptic",
+  "validate",
+  "chain",
+  "patch",
+  "report",
+] as const;
+
+const ScratchpadPhaseSchema = Type.String({
+  enum: [...SCRATCHPAD_PHASES],
+  description: "Pipeline phase: recon | hunt | gapfil | trace | skeptic | validate | chain | patch | report",
+});
+
+const ScratchpadInitSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Unique run identifier for this pipeline run" }),
+  },
+  { additionalProperties: false },
+);
+
+const ScratchpadResumeSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Run identifier to resume" }),
+  },
+  { additionalProperties: false },
+);
+
+const ScratchpadCheckpointSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Run identifier" }),
+    phase: ScratchpadPhaseSchema,
+    ids: Type.Optional(Type.Array(Type.String(), { description: "Key IDs produced by this phase (case IDs, finding IDs)" })),
+    summary: Type.Optional(Type.String({ description: "One-line summary of phase completion" })),
+  },
+  { additionalProperties: false },
+);
+
+const ScratchpadWriteSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Run identifier" }),
+    phase: ScratchpadPhaseSchema,
+    artifact_name: Type.String({ description: "Artifact filename (sanitized; path traversal is blocked)" }),
+    content: Type.String({ description: "Artifact content to write" }),
+  },
+  { additionalProperties: false },
+);
+
+const ScratchpadReadSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Run identifier" }),
+    phase: ScratchpadPhaseSchema,
+    artifact_name: Type.String({ description: "Artifact filename to read" }),
+  },
+  { additionalProperties: false },
+);
+
+const ScratchpadPhaseDoneSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Run identifier" }),
+    phase: ScratchpadPhaseSchema,
+  },
+  { additionalProperties: false },
+);
+
+const ScratchpadClearSchema = Type.Object(
+  {
+    run_id: Type.String({ description: "Run identifier to clear (deletes that run only)" }),
+  },
+  { additionalProperties: false },
+);
+
 interface Theme {
   fg(color: string, text: string): string;
   bold(text: string): string;
@@ -1066,6 +1159,366 @@ export default function casefileExtension(pi: ExtensionAPI) {
     },
   });
 
+  // ── Tool: ScratchpadInit ──
+
+  pi.registerTool({
+    name: "ScratchpadInit",
+    label: "Init Scratchpad",
+    description:
+      "Initialize a crash-recoverable artifact store for a pipeline run. Creates the directory structure and an initial state.json checkpoint. Idempotent — safe to call on resume without --fresh; returns the existing checkpoint if the run already exists.",
+    promptSnippet: "Initialize the pipeline artifact store for a run",
+    promptGuidelines: [
+      "Call ScratchpadInit once at the start of a pipeline run (or on resume before ScratchpadResume).",
+      "The run_id is arbitrary but should be unique per pipeline run — typically <target>-<timestamp>.",
+      "On resume, ScratchpadInit returns the existing checkpoint without wiping it; pair with ScratchpadResume to skip completed phases.",
+    ],
+    parameters: ScratchpadInitSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const cp = scratchpad_init(params.run_id as string);
+      return {
+        content: [
+          {
+            type: "text",
+            text: `Scratchpad initialized for run ${cp.run_id}.\nCompleted phases: ${cp.completed_phases.length ? cp.completed_phases.join(", ") : "none"}\nDone: ${cp.done}`,
+          },
+        ],
+        details: { checkpoint: cp },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadInit ")) +
+          theme.fg("dim", (args.run_id as string) ?? ""),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const cp = (result.details as { checkpoint: { run_id: string; done: boolean } } | undefined)?.checkpoint;
+      return new Text(
+        theme.fg("success", "✓ ") +
+          `ScratchpadInit ${cp?.run_id ?? ""}${cp?.done ? " (done)" : ""}`,
+        0,
+        0,
+      );
+    },
+  });
+
+  // ── Tool: ScratchpadResume ──
+
+  pi.registerTool({
+    name: "ScratchpadResume",
+    label: "Resume Scratchpad",
+    description:
+      "Read the checkpoint and artifact listing for a pipeline run to decide where to resume. Returns the next phase to run (or null if done) and which phases already completed. Returns null if the run does not exist.",
+    promptSnippet: "Check pipeline resume state — which phases are done",
+    promptGuidelines: [
+      "Call ScratchpadResume at pipeline start to determine where to resume. If it returns a checkpoint, skip completed phases (check ScratchpadPhaseDone before each dispatch) and continue from next_phase.",
+      "If ScratchpadResume returns null, the run has no checkpoint — call ScratchpadInit to start fresh.",
+      "Use ScratchpadPhaseDone before dispatching each stage to avoid re-running completed phases (idempotent resume).",
+    ],
+    parameters: ScratchpadResumeSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const resume = scratchpad_resume(params.run_id as string);
+      if (!resume) {
+        return {
+          content: [
+            {
+              type: "text",
+              text: `No scratchpad found for run ${params.run_id}. Call ScratchpadInit to start a new run.`,
+            },
+          ],
+          details: { resume: null },
+        };
+      }
+      const cp = resume.checkpoint;
+      return {
+        content: [
+          {
+            type: "text",
+            text:
+              `Resume run ${cp.run_id}:\n` +
+              `Completed phases: ${cp.completed_phases.length ? cp.completed_phases.join(", ") : "none"}\n` +
+              `Next phase: ${resume.next_phase ?? "none (run is done)"}\n` +
+              `Done: ${cp.done}`,
+          },
+        ],
+        details: { resume },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadResume ")) +
+          theme.fg("dim", (args.run_id as string) ?? ""),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const resume = (result.details as { resume: ScratchpadResume | null } | undefined)?.resume;
+      if (!resume) return new Text(theme.fg("warning", "↷ ScratchpadResume — no run found"), 0, 0);
+      return new Text(
+        theme.fg("success", "✓ ") +
+          `ScratchpadResume ${resume.checkpoint.run_id} → next: ${resume.next_phase ?? "done"}`,
+        0,
+        0,
+      );
+    },
+  });
+
+  // ── Tool: ScratchpadCheckpoint ──
+
+  pi.registerTool({
+    name: "ScratchpadCheckpoint",
+    label: "Checkpoint Phase",
+    description:
+      "Mark a pipeline phase as complete in the scratchpad state.json. Records the completion timestamp, key IDs, and an optional summary. Idempotent — re-checkpointing a phase overwrites its summary/IDs without duplicating the completed_phases entry.",
+    promptSnippet: "Record a pipeline phase as complete",
+    promptGuidelines: [
+      "Call ScratchpadCheckpoint after every phase completes: ScratchpadCheckpoint(run_id, phase, { ids, summary }).",
+      "ids are the key case/finding IDs the phase produced — used by resume to reconstruct state.",
+      "Keep completed_phases in pipeline order; the checkpoint sorts automatically.",
+    ],
+    parameters: ScratchpadCheckpointSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const cp = scratchpad_checkpoint(
+        params.run_id as string,
+        params.phase as ScratchpadPhase,
+        { ids: params.ids as string[] | undefined, summary: params.summary as string | undefined },
+      );
+      return {
+        content: [
+          {
+            type: "text",
+            text:
+              `Phase ${params.phase} checkpointed for run ${cp.run_id}.\n` +
+              `Completed phases: ${cp.completed_phases.join(", ")}`,
+          },
+        ],
+        details: { checkpoint: cp },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadCheckpoint ")) +
+          theme.fg("dim", `${args.run_id ?? ""} ${args.phase ?? ""}`),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const cp = (result.details as { checkpoint: { run_id: string; completed_phases: string[] } } | undefined)?.checkpoint;
+      return new Text(
+        theme.fg("success", "✓ ") +
+          `ScratchpadCheckpoint ${cp?.run_id ?? ""} — ${cp?.completed_phases.length ?? 0} phases done`,
+        0,
+        0,
+      );
+    },
+  });
+
+  // ── Tool: ScratchpadWrite ──
+
+  pi.registerTool({
+    name: "ScratchpadWrite",
+    label: "Write Artifact",
+    description:
+      "Write an intermediate artifact (recon map, trace output, verification log) to a phase's subdirectory in the scratchpad. Overwrites if the name exists. Artifact names are sanitized — path traversal is blocked.",
+    promptSnippet: "Save a pipeline artifact to the scratchpad",
+    promptGuidelines: [
+      "Agents write artifacts to the scratchpad, not to each other's output files (prevents an echo chamber).",
+      "The casefile owns state transitions; the scratchpad owns artifacts. Use ScratchpadWrite for bulky intermediate outputs, not CaseUpdate.",
+    ],
+    parameters: ScratchpadWriteSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const path = scratchpad_write(
+        params.run_id as string,
+        params.phase as ScratchpadPhase,
+        params.artifact_name as string,
+        params.content as string,
+      );
+      return {
+        content: [
+          {
+            type: "text",
+            text: `Artifact written: ${params.artifact_name} → ${path}`,
+          },
+        ],
+        details: { path, artifact_name: params.artifact_name },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadWrite ")) +
+          theme.fg("dim", `${args.run_id ?? ""}/${args.phase ?? ""}/${args.artifact_name ?? ""}`),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const name = (result.details as { artifact_name?: string } | undefined)?.artifact_name;
+      return new Text(theme.fg("success", `✓ ScratchpadWrite ${name ?? ""}`), 0, 0);
+    },
+  });
+
+  // ── Tool: ScratchpadRead ──
+
+  pi.registerTool({
+    name: "ScratchpadRead",
+    label: "Read Artifact",
+    description:
+      "Read an artifact from a phase's subdirectory in the scratchpad. Returns null if the artifact is missing. Use to resume a phase from a prior run's intermediate output.",
+    promptSnippet: "Read a pipeline artifact from the scratchpad",
+    promptGuidelines: [
+      "On resume, ScratchpadRead retrieves a prior phase's intermediate output so the next phase can proceed without re-running it.",
+      "Returns null for missing artifacts — treat as 'not yet produced' rather than an error.",
+    ],
+    parameters: ScratchpadReadSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const content = scratchpad_read(
+        params.run_id as string,
+        params.phase as ScratchpadPhase,
+        params.artifact_name as string,
+      );
+      if (content === null) {
+        return {
+          content: [
+            {
+              type: "text",
+              text: `Artifact not found: ${params.artifact_name} in ${params.phase}/`,
+            },
+          ],
+          details: { artifact_name: params.artifact_name, found: false },
+        };
+      }
+      return {
+        content: [{ type: "text", text: content }],
+        details: { artifact_name: params.artifact_name, found: true, length: content.length },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadRead ")) +
+          theme.fg("dim", `${args.run_id ?? ""}/${args.phase ?? ""}/${args.artifact_name ?? ""}`),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const found = (result.details as { found?: boolean } | undefined)?.found;
+      return new Text(
+        found ? theme.fg("success", "✓ ScratchpadRead") : theme.fg("warning", "↷ ScratchpadRead — not found"),
+        0,
+        0,
+      );
+    },
+  });
+
+  // ── Tool: ScratchpadPhaseDone ──
+
+  pi.registerTool({
+    name: "ScratchpadPhaseDone",
+    label: "Phase Done?",
+    description:
+      "Check whether a phase has already been checkpointed in the scratchpad — for idempotent re-run. Returns true if the phase is complete; skip re-dispatching it on resume.",
+    promptSnippet: "Check if a pipeline phase is already complete",
+    promptGuidelines: [
+      "Call ScratchpadPhaseDone before dispatching each stage to avoid re-running completed phases on resume.",
+      "A completed phase with a checkpoint is a no-op on re-run — skip it and continue to the next incomplete phase.",
+    ],
+    parameters: ScratchpadPhaseDoneSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const done = scratchpad_phase_done(
+        params.run_id as string,
+        params.phase as ScratchpadPhase,
+      );
+      return {
+        content: [
+          {
+            type: "text",
+            text: `Phase ${params.phase} for run ${params.run_id}: ${done ? "DONE (skip on resume)" : "not done"}`,
+          },
+        ],
+        details: { phase: params.phase, done },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadPhaseDone ")) +
+          theme.fg("dim", `${args.run_id ?? ""} ${args.phase ?? ""}`),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const done = (result.details as { done?: boolean } | undefined)?.done;
+      return new Text(
+        done ? theme.fg("success", "✓ ScratchpadPhaseDone — done") : theme.fg("warning", "↷ ScratchpadPhaseDone — not done"),
+        0,
+        0,
+      );
+    },
+  });
+
+  // ── Tool: ScratchpadClear ──
+
+  pi.registerTool({
+    name: "ScratchpadClear",
+    label: "Clear Run",
+    description:
+      "Clear a single pipeline run's scratchpad directory. Used by --fresh for one run. Does not touch other runs. The run must be re-initialized with ScratchpadInit afterward.",
+    promptSnippet: "Clear one pipeline run's artifacts",
+    promptGuidelines: [
+      "Use ScratchpadClear to force a fresh start for a single run (--fresh). It deletes that run's directory only.",
+      "After clearing, call ScratchpadInit to recreate the directory structure before writing artifacts.",
+    ],
+    parameters: ScratchpadClearSchema,
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      scratchpad_clear(params.run_id as string);
+      return {
+        content: [
+          {
+            type: "text",
+            text: `Scratchpad cleared for run ${params.run_id}. Call ScratchpadInit to start a new run.`,
+          },
+        ],
+        details: { run_id: params.run_id, cleared: true },
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("ScratchpadClear ")) +
+          theme.fg("dim", (args.run_id as string) ?? ""),
+        0,
+        0,
+      );
+    },
+
+    renderResult(_result, _opts, theme) {
+      return new Text(theme.fg("success", "✓ ScratchpadClear"), 0, 0);
+    },
+  });
+
   // ── Command: /casefile ──
 
   pi.registerCommand("casefile", {

package/src/scratchpad.ts

@@ -0,0 +1,359 @@
+/**
+ * Scratchpad — intermediate artifact store for pipeline runs.
+ *
+ * The casefile owns state transitions; the scratchpad owns artifacts.
+ * Agents write their outputs here (recon maps, trace outputs, verification
+ * logs) instead of stuffing everything into casefile text fields or relying
+ * on each other's output streams (which creates an echo chamber).
+ *
+ * Directory layout per pipeline run:
+ *   {project_root}/.scratchpad/{run_id}/
+ *     recon/      — fingerprints, tech detection, surface maps
+ *     trace/      — per-finding reachability traces
+ *     verify/     — PoC logs, run outputs
+ *     state.json  — checkpoint file with phase completion + key IDs
+ *
+ * Resume re-reads scratchpad artifacts; it does not re-run completed phases
+ * (idempotent). The `.scratchpad/` directory is preserved between runs;
+ * `--fresh` clears it via scratchpad_clear().
+ */
+
+import { existsSync, mkdirSync, readdirSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { dirname, join, resolve } from "node:path";
+
+// ── Types ────────────────────────────────────────────────────────────
+
+export type ScratchpadPhase =
+  | "recon"
+  | "hunt"
+  | "gapfil"
+  | "trace"
+  | "skeptic"
+  | "validate"
+  | "chain"
+  | "patch"
+  | "report";
+
+export interface ScratchpadCheckpoint {
+  run_id: string;
+  project_root: string;
+  created_at: string;
+  last_updated: string;
+  /** Ordered list of phases that have completed (in pipeline order). */
+  completed_phases: ScratchpadPhase[];
+  /** ISO timestamp of the last phase completion. */
+  last_phase_at: string | null;
+  /** Key IDs produced by each phase — case IDs, finding IDs, etc. */
+  phase_ids: Record<ScratchpadPhase, string[]>;
+  /** Free-form summary per phase, set by checkpoint(). */
+  phase_summaries: Record<ScratchpadPhase, string>;
+  /** Whether the run is fully complete. */
+  done: boolean;
+}
+
+export interface ScratchpadResume {
+  checkpoint: ScratchpadCheckpoint;
+  /** The next phase to run (or null if the run is done). */
+  next_phase: ScratchpadPhase | null;
+  /** Artifact references per phase: { trace: ["finding-abc.json", ...], ... } */
+  artifacts: Record<string, string[]>;
+}
+
+// ── Constants ────────────────────────────────────────────────────────
+
+const PHASE_ORDER: ScratchpadPhase[] = [
+  "recon",
+  "hunt",
+  "gapfil",
+  "trace",
+  "skeptic",
+  "validate",
+  "chain",
+  "patch",
+  "report",
+];
+
+const PHASE_DIRS: Record<ScratchpadPhase, string> = {
+  recon: "recon",
+  hunt: "hunt",
+  gapfil: "gapfil",
+  trace: "trace",
+  skeptic: "skeptic",
+  validate: "verify",
+  chain: "chain",
+  patch: "patch",
+  report: "report",
+};
+
+const SCRATCHPAD_DIR = ".scratchpad";
+
+// ── Helpers ──────────────────────────────────────────────────────────
+
+let scratchpadRootOverride: string | undefined;
+
+/**
+ * Detect the workspace root by walking up for a .git dir or package.json,
+ * matching the ledger's detectWorkspaceRoot() heuristic.
+ */
+function detectWorkspaceRoot(): string {
+  if (scratchpadRootOverride) return scratchpadRootOverride;
+
+  const envs = ["XPI_SCRATCHPAD_ROOT", "PI_WORKSPACE_ROOT", "GITHUB_WORKSPACE", "PWD"];
+  for (const e of envs) {
+    const v = process.env[e];
+    if (v) return resolve(v);
+  }
+
+  let curr = resolve(process.cwd());
+  for (let i = 0; i < 20; i++) {
+    if (existsSync(join(curr, ".git")) || existsSync(join(curr, "package.json"))) return curr;
+    const parent = dirname(curr);
+    if (parent === curr) break;
+    curr = parent;
+  }
+  return resolve(process.cwd());
+}
+
+/** Override the scratchpad root (for testing). Pass undefined to reset. */
+export function setScratchpadRoot(path: string | undefined): void {
+  scratchpadRootOverride = path ? resolve(path) : undefined;
+}
+
+/** The top-level scratchpad directory for a given project root. */
+export function getScratchpadRoot(projectRoot?: string): string {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  return join(root, SCRATCHPAD_DIR);
+}
+
+/** The directory for a specific run. */
+export function getRunDir(runId: string, projectRoot?: string): string {
+  return join(getScratchpadRoot(projectRoot), runId);
+}
+
+/** The state.json path for a run. */
+export function getStatePath(runId: string, projectRoot?: string): string {
+  return join(getRunDir(runId, projectRoot), "state.json");
+}
+
+function emptyCheckpoint(runId: string, projectRoot: string): ScratchpadCheckpoint {
+  const now = new Date().toISOString();
+  return {
+    run_id: runId,
+    project_root: projectRoot,
+    created_at: now,
+    last_updated: now,
+    last_phase_at: null,
+    completed_phases: [],
+    phase_ids: {} as Record<ScratchpadPhase, string[]>,
+    phase_summaries: {} as Record<ScratchpadPhase, string>,
+    done: false,
+  };
+}
+
+function ensureRunDirs(runDir: string): void {
+  if (!existsSync(runDir)) mkdirSync(runDir, { recursive: true });
+  for (const phase of PHASE_ORDER) {
+    const dir = join(runDir, PHASE_DIRS[phase]);
+    if (!existsSync(dir)) mkdirSync(dir, { recursive: true });
+  }
+}
+
+function readCheckpointRaw(runId: string, projectRoot?: string): ScratchpadCheckpoint | null {
+  const statePath = getStatePath(runId, projectRoot);
+  if (!existsSync(statePath)) return null;
+  try {
+    const raw = readFileSync(statePath, "utf8");
+    const cp = JSON.parse(raw) as ScratchpadCheckpoint;
+    // Backfill maps for phases not yet checkpointed (defensive).
+    if (!cp.phase_ids) cp.phase_ids = {} as Record<ScratchpadPhase, string[]>;
+    if (!cp.phase_summaries) cp.phase_summaries = {} as Record<ScratchpadPhase, string>;
+    return cp;
+  } catch {
+    return null;
+  }
+}
+
+function writeCheckpointRaw(cp: ScratchpadCheckpoint, projectRoot?: string): void {
+  cp.last_updated = new Date().toISOString();
+  const statePath = getStatePath(cp.run_id, projectRoot);
+  ensureRunDirs(getRunDir(cp.run_id, projectRoot));
+  writeFileSync(statePath, JSON.stringify(cp, null, 2), "utf8");
+}
+
+// ── Public API ───────────────────────────────────────────────────────
+
+/**
+ * Initialize a new scratchpad run. Creates the directory structure and writes
+ * an initial state.json. If the run already exists, returns the existing
+ * checkpoint (idempotent — safe to call on resume without --fresh).
+ */
+export function scratchpad_init(runId: string, projectRoot?: string): ScratchpadCheckpoint {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const runDir = getRunDir(runId, root);
+  ensureRunDirs(runDir);
+
+  const existing = readCheckpointRaw(runId, root);
+  if (existing) return existing;
+
+  const cp = emptyCheckpoint(runId, root);
+  writeCheckpointRaw(cp, root);
+  return cp;
+}
+
+/**
+ * Write an artifact to a phase's subdirectory. Overwrites if the name exists.
+ * Returns the full path to the written artifact.
+ */
+export function scratchpad_write(
+  runId: string,
+  phase: ScratchpadPhase,
+  artifactName: string,
+  content: string,
+  projectRoot?: string,
+): string {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const runDir = getRunDir(runId, root);
+  ensureRunDirs(runDir);
+
+  // Sanitize artifact name: no path traversal.
+  const safeName = artifactName.replace(/[^a-zA-Z0-9._-]/g, "_");
+  const dir = join(runDir, PHASE_DIRS[phase]);
+  const filePath = join(dir, safeName);
+  writeFileSync(filePath, content, "utf8");
+  return filePath;
+}
+
+/**
+ * Read an artifact. Returns null if missing.
+ */
+export function scratchpad_read(
+  runId: string,
+  phase: ScratchpadPhase,
+  artifactName: string,
+  projectRoot?: string,
+): string | null {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const safeName = artifactName.replace(/[^a-zA-Z0-9._-]/g, "_");
+  const filePath = join(getRunDir(runId, root), PHASE_DIRS[phase], safeName);
+  if (!existsSync(filePath)) return null;
+  return readFileSync(filePath, "utf8");
+}
+
+/**
+ * List all artifacts written for a phase.
+ */
+export function scratchpad_list(
+  runId: string,
+  phase: ScratchpadPhase,
+  projectRoot?: string,
+): string[] {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const dir = join(getRunDir(runId, root), PHASE_DIRS[phase]);
+  if (!existsSync(dir)) return [];
+  return readdirSync(dir).filter((f) => f !== "state.json");
+}
+
+/**
+ * Mark a phase as complete. Records the completion timestamp, key IDs, and an
+ * optional summary in state.json. Idempotent: re-checkpointing a phase
+ * overwrites its previous summary/IDs but does not duplicate the entry in
+ * completed_phases.
+ */
+export function scratchpad_checkpoint(
+  runId: string,
+  phase: ScratchpadPhase,
+  data: { ids?: string[]; summary?: string },
+  projectRoot?: string,
+): ScratchpadCheckpoint {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const cp = readCheckpointRaw(runId, root) ?? scratchpad_init(runId, root);
+
+  if (!cp.completed_phases.includes(phase)) {
+    cp.completed_phases.push(phase);
+    // Keep completed_phases in pipeline order for predictable resume.
+    cp.completed_phases.sort((a, b) => PHASE_ORDER.indexOf(a) - PHASE_ORDER.indexOf(b));
+  }
+  cp.last_phase_at = new Date().toISOString();
+  if (data.ids) cp.phase_ids[phase] = data.ids;
+  if (data.summary) cp.phase_summaries[phase] = data.summary;
+
+  writeCheckpointRaw(cp, root);
+  return cp;
+}
+
+/**
+ * Read the checkpoint + all artifact references for resume.
+ * Returns null if the run doesn't exist.
+ */
+export function scratchpad_resume(runId: string, projectRoot?: string): ScratchpadResume | null {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const cp = readCheckpointRaw(runId, root);
+  if (!cp) return null;
+
+  // Find the next phase: the first phase in order not in completed_phases.
+  const next = PHASE_ORDER.find((p) => !cp.completed_phases.includes(p)) ?? null;
+
+  // Gather artifact listing per completed phase.
+  const artifacts: Record<string, string[]> = {};
+  for (const phase of cp.completed_phases) {
+    artifacts[phase] = scratchpad_list(runId, phase, root);
+  }
+
+  return { checkpoint: cp, next_phase: next, artifacts };
+}
+
+/**
+ * Check whether a phase has already been checkpointed (for idempotent re-run).
+ */
+export function scratchpad_phase_done(
+  runId: string,
+  phase: ScratchpadPhase,
+  projectRoot?: string,
+): boolean {
+  const cp = readCheckpointRaw(runId, projectRoot);
+  return cp?.completed_phases.includes(phase) ?? false;
+}
+
+/**
+ * Clear a specific run's scratchpad directory. Used by `--fresh` for a single
+ * run. Does not touch other runs.
+ */
+export function scratchpad_clear(runId: string, projectRoot?: string): void {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const runDir = getRunDir(runId, root);
+  if (existsSync(runDir)) rmSync(runDir, { recursive: true, force: true });
+}
+
+/**
+ * Clear the entire scratchpad directory (all runs). Used by `--fresh` with no
+ * run ID. Use with care.
+ */
+export function scratchpad_clear_all(projectRoot?: string): void {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const dir = getScratchpadRoot(root);
+  if (existsSync(dir)) rmSync(dir, { recursive: true, force: true });
+}
+
+/**
+ * List all run IDs in the scratchpad (for resume selection).
+ */
+export function scratchpad_list_runs(projectRoot?: string): string[] {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const dir = getScratchpadRoot(root);
+  if (!existsSync(dir)) return [];
+  return readdirSync(dir, { withFileTypes: true })
+    .filter((e) => e.isDirectory())
+    .map((e) => e.name)
+    .sort();
+}
+
+/**
+ * Mark the run as fully done. Prevents resume from re-entering.
+ */
+export function scratchpad_finish(runId: string, projectRoot?: string): ScratchpadCheckpoint {
+  const root = projectRoot ?? detectWorkspaceRoot();
+  const cp = readCheckpointRaw(runId, root) ?? scratchpad_init(runId, root);
+  cp.done = true;
+  writeCheckpointRaw(cp, root);
+  return cp;
+}

package/src/workflow.ts

@@ -147,6 +147,8 @@ is considered disproven and promotion is blocked. If you cannot write a
 meaningful disconfirmation script, you may not understand the finding well
 enough to promote it.
 
+**Adversarial disconfirmation (skeptic subagent):** For findings at severity >= high, a dedicated skeptic subagent independently re-reads the source and tries to disprove the finding BEFORE the exploit agent runs. The skeptic's \`disconfirmation_attempt\` is written into this \`disconfirmation\` field by the harness — it satisfies this gate and is stronger than self-disconfirmation because a separate agent produced it. If the skeptic says DISPROVEN, the finding is killed directly. Self-disconfirmation still applies for findings below high severity.
+
 ### 2. Production Path Verification (must be in impact field)
 
 The **impact** field for CONFIRMED must explicitly answer: