@xaccefy/pi-casefile 0.7.1 → 0.7.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xaccefy/pi-casefile",
-  "version": "0.7.1",
+  "version": "0.7.2",
   "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/pipeline-submit.ts",
     "src/scratchpad.ts",
     "src/sqlite-compat/index.ts",
     "skills",

package/src/index.ts

@@ -1,9 +1,9 @@
 /**
  * Casefile — offensive security case tracker for Pi.
  *
- * Tools: CaseAdd, CaseUpdate, PromoteFinding, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseReport, ScratchpadInit, ScratchpadResume, ScratchpadCheckpoint, ScratchpadWrite, ScratchpadRead, ScratchpadPhaseDone, ScratchpadClear
+ * Tools: CaseAdd, CaseUpdate, PromoteFinding, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseReport, PipelineSubmit, 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
+ * Event: before_agent_start — injects cyber workflow once per session, refreshes the active case list per prompt
  */
 
 import { existsSync, readFileSync, writeFileSync } from "node:fs";
@@ -44,6 +44,7 @@ import {
   updateCaseResult,
   writeCaseReport,
 } from "./ledger.ts";
+import { pipeline_submit, SUBMIT_STAGES, type SubmitStage } from "./pipeline-submit.ts";
 import { type PocRun, runPoc } from "./poc-runner.ts";
 import {
   type ScratchpadPhase,
@@ -511,9 +512,15 @@ function buildCaseListContext(records: CaseRecord[]): string {
   return lines.join("\n");
 }
 
-/** Always includes cyber workflow; attaches case list when active cases exist. */
-function buildAgentInjection(active: CaseRecord[]): string {
+/**
+ * Builds the per-prompt injection. The cyber workflow is session-scope data —
+ * it never changes — so the caller passes includeWorkflow=true exactly once
+ * per session; re-injecting it on every prompt is pure token cost. The active
+ * case list DOES change as cases are added, so it is refreshed every prompt.
+ */
+function buildAgentInjection(active: CaseRecord[], includeWorkflow: boolean): string {
   const caseList = buildCaseListContext(active);
+  if (!includeWorkflow) return caseList;
   // Workflow FIRST for prominence, then case list as reference data.
   return caseList ? `${STATIC_CYBER_WORKFLOW}\n\n${caseList}` : STATIC_CYBER_WORKFLOW;
 }
@@ -1207,6 +1214,80 @@ export default function casefileExtension(pi: ExtensionAPI) {
     },
   });
 
+  // ── Tool: PipelineSubmit ──
+
+  pi.registerTool({
+    name: "PipelineSubmit",
+    label: "Submit Stage Output",
+    description:
+      "Submit a pipeline stage's output (hunt, trace, skeptic, validate, chain, report) through the validation gate. Validates required fields against the stage spec (mirrors schemas/*.json), applies the deterministic pre-filter (test-path and file-existence filters on hunt findings, trivial dedup by file+class+line), and counts repair attempts (max 2, then rejected). A stage cannot advance on an invalid output — submit fixed output until accepted.",
+    promptSnippet: "Validate and submit a pipeline stage's output",
+    promptGuidelines: [
+      "Every stage output a subagent returns must go through PipelineSubmit before the next stage is dispatched. Do not eyeball schemas.",
+      "If the verdict is repair, fix the fields listed in errors and re-submit the same output. The repair budget is 2 attempts per finding — after that the submission is rejected and the stage is failed.",
+      "Unhandled skeptic output: an unparseable or schema-invalid skeptic response is UNDETERMINED, never DISPROVEN. A tracer error is UNREACHABLE. PipelineSubmit returns repair for these instead of accepting them.",
+      "Test-path findings and hallucinated files are rejected by the pre-filter, not repairable — the finding itself is noise.",
+    ],
+    parameters: Type.Object(
+      {
+        run_id: Type.String({
+          description: "Pipeline run identifier (same as the scratchpad run_id)",
+        }),
+        stage: Type.String({
+          enum: [...SUBMIT_STAGES],
+          description: "Pipeline stage: hunt | trace | skeptic | validate | chain | report",
+        }),
+        output: Type.Union([Type.String(), Type.Object({}, { additionalProperties: true })], {
+          description: "The stage output as a JSON object or JSON string (code fences tolerated)",
+        }),
+      },
+      { additionalProperties: false },
+    ),
+
+    async execute(_id, params, _signal, _onUpdate, _ctx) {
+      const result = pipeline_submit(
+        params.run_id as string,
+        params.stage as SubmitStage,
+        params.output,
+      );
+      const statusLine =
+        result.verdict === "accepted"
+          ? `ACCEPTED (${params.stage}) — artifact: ${result.artifact}`
+          : result.verdict === "repair"
+            ? `REPAIR (attempt ${result.repair_attempt}/2) — fix these and re-submit:\n  - ${result.errors.join("\n  - ")}`
+            : `REJECTED — ${result.errors.join("\n")}`;
+      return {
+        content: [{ type: "text", text: statusLine }],
+        isError: result.verdict !== "accepted",
+        details: result as unknown as Record<string, unknown>,
+      };
+    },
+
+    renderCall(args, theme) {
+      return new Text(
+        theme.fg("toolTitle", theme.bold("PipelineSubmit ")) +
+          theme.fg("dim", `${args.stage ?? ""}`),
+        0,
+        0,
+      );
+    },
+
+    renderResult(result, _opts, theme) {
+      const details = result.details as { verdict?: string; repair_attempt?: number } | undefined;
+      if (details?.verdict === "accepted") {
+        return new Text(theme.fg("success", "✓ PipelineSubmit accepted"), 0, 0);
+      }
+      if (details?.verdict === "repair") {
+        return new Text(
+          theme.fg("warning", `↷ PipelineSubmit repair ${details.repair_attempt}/2`),
+          0,
+          0,
+        );
+      }
+      return new Text(theme.fg("error", "✗ PipelineSubmit rejected"), 0, 0);
+    },
+  });
+
   // ── Tool: ScratchpadInit ──
 
   pi.registerTool({
@@ -1599,13 +1680,22 @@ export default function casefileExtension(pi: ExtensionAPI) {
 
   // ── Event: Inject cyber workflow into system prompt ──
   // XP (offensive) mode is OFF by default so normal dev work stays quiet.
-  // Only when enabled do we inject the cyber workflow (and case list) into
-  // the system prompt each turn. Injecting into event.systemPrompt (not as a
-  // conversation message) makes the attacker mindset immediate and avoids
-  // session bloat from repeated message entries.
+  // When enabled, the cyber workflow is injected ONCE per session (first
+  // prompt); the active case list refreshes every prompt because it changes
+  // as cases are added. Injecting into event.systemPrompt (not as a
+  // conversation message) avoids session bloat from repeated message entries.
+  let workflowInjected = false;
 
   pi.on("before_agent_start", async (event) => {
     if (readXpMode() === "off") return;
+    // Skip subagent child processes: pi-subagents runs each child in its own
+    // pi process (PI_SUBAGENT_CHILD=1) with this extension loaded. Injecting
+    // the workflow + entire active-case ledger into every child dispatch is a
+    // token multiplier (N subagents × workflow + growing case list per turn) —
+    // workers get what they need via their task and tool guidelines.
+    if (process.env.PI_SUBAGENT_CHILD === "1") return;
+
+    const includeWorkflow = !workflowInjected;
 
     let active: CaseRecord[] = [];
     try {
@@ -1614,7 +1704,9 @@ export default function casefileExtension(pi: ExtensionAPI) {
       // No database yet — still inject workflow.
     }
 
-    const injection = buildAgentInjection(active);
+    const injection = buildAgentInjection(active, includeWorkflow);
+    if (!injection) return; // workflow already injected, no active cases
+    workflowInjected = true;
 
     // Inject workflow FIRST (before skills) so the attacker mindset is
     // prominent, not buried at the end of a long system prompt.

package/src/pipeline-submit.ts

@@ -0,0 +1,509 @@
+/**
+ * PipelineSubmit — the stage-output gate.
+ *
+ * The coordinator (model) dispatches stage subagents and submits their output
+ * here. This module is where the pipeline stops trusting prose: it validates
+ * stage output against the field specs mirrored from schemas/*.json, applies
+ * the deterministic pre-filter (test paths, hallucinated files, trivial dedup),
+ * and counts repair attempts. A stage cannot advance on an invalid output —
+ * the answer is REPAIR (with field-level errors) or REJECTED, in code.
+ *
+ * KEEP IN SYNC with schemas/*.json at the repo root. The JSON schemas are the
+ * canonical data contract for documentation; the SPECS table here is the
+ * executable gate (a focused validator for exactly these six shapes — no
+ * general JSON Schema engine).
+ *
+ * Persistence: .scratchpad/{run_id}/pipeline-submit.json
+ *   { repairs: { "<stage>:<key>": n }, accepted_findings: FindingRef[] }
+ */
+
+import { createHash } from "node:crypto";
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import { dirname, isAbsolute, join, relative, resolve } from "node:path";
+import {
+  getRunDir,
+  getScratchpadRoot,
+  type ScratchpadPhase,
+  scratchpad_write,
+} from "./scratchpad.ts";
+
+// ── Types ────────────────────────────────────────────────────────────
+
+export const SUBMIT_STAGES = ["hunt", "trace", "skeptic", "validate", "chain", "report"] as const;
+export type SubmitStage = (typeof SUBMIT_STAGES)[number];
+
+export type SubmitVerdict = "accepted" | "repair" | "rejected";
+
+export type SubmitResult = {
+  verdict: SubmitVerdict;
+  stage: SubmitStage;
+  /** Field-level validation errors (repair) or rejection reason (rejected). */
+  errors: string[];
+  /** Repair attempt number (1-based) when verdict is repair. */
+  repair_attempt?: number;
+  /** Stable key identifying this finding's repair bucket. */
+  key?: string;
+  /** Set when hunt-stage dedup matched an accepted finding. */
+  duplicate_of?: string;
+  /** Scratchpad path the accepted output was written to. */
+  artifact?: string;
+};
+
+type StageSpec = {
+  /** Fields that must be present and non-empty. */
+  required: {
+    name: string;
+    type: "string" | "integer" | "array" | "object";
+    enum?: readonly string[];
+    minItems?: number;
+  }[];
+  /** Exactly one of these locator field-sets must be fully present. */
+  locatorXor?: [string[], string[]];
+  /** Conditional requirements: when field equals value, these must be non-empty. */
+  conditional?: { when: { field: string; equals: string }; require: string[] }[];
+};
+
+// ── Stage specs (mirror of schemas/*.json semantics) ─────────────────
+
+const VULN_CLASSES = [
+  "injection",
+  "xss",
+  "idor",
+  "bola",
+  "path-traversal",
+  "ssrf",
+  "command-injection",
+  "deserialization",
+  "auth-bypass",
+  "privilege-escalation",
+  "business-logic",
+  "race-condition",
+  "xxe",
+  "ssti",
+  "open-redirect",
+  "information-disclosure",
+  "crypto-weakness",
+  "other",
+] as const;
+
+const KILL_REASONS = [
+  "unreachable",
+  "framework_protection",
+  "input_validation_blocks",
+  "requires_privilege_attacker_lacks",
+  "poc_failed_3x",
+  "no_real_impact",
+  "intended_behavior",
+  "duplicate",
+] as const;
+
+const SPECS: Record<SubmitStage, StageSpec> = {
+  // schemas/stage-finding.json
+  hunt: {
+    required: [
+      { name: "vuln_class", type: "string", enum: VULN_CLASSES },
+      { name: "sink", type: "string" },
+      { name: "entry_point", type: "string" },
+      { name: "confidence", type: "string", enum: ["low", "medium", "high"] },
+      { name: "evidence", type: "string" },
+    ],
+    // Source targets: file + line. Live targets: endpoint.
+    locatorXor: [["file", "line"], ["endpoint"]],
+  },
+  // schemas/stage-trace.json
+  trace: {
+    required: [
+      { name: "trace_result", type: "string", enum: ["REACHABLE", "UNREACHABLE"] },
+      { name: "entry_point", type: "string" },
+      { name: "call_chain", type: "array", minItems: 1 },
+      { name: "defenses_checked", type: "array" },
+      { name: "attacker_model", type: "string" },
+    ],
+    conditional: [
+      { when: { field: "trace_result", equals: "REACHABLE" }, require: ["impact_if_reachable"] },
+      { when: { field: "trace_result", equals: "UNREACHABLE" }, require: ["unreachable_reason"] },
+    ],
+  },
+  // schemas/stage-skeptic.json
+  skeptic: {
+    required: [
+      { name: "finding_id", type: "string" },
+      { name: "verdict", type: "string", enum: ["CONFIRMED", "DISPROVEN"] },
+      { name: "reasoning", type: "string" },
+      { name: "evidence_reviewed", type: "array", minItems: 1 },
+    ],
+    conditional: [
+      { when: { field: "verdict", equals: "DISPROVEN" }, require: ["disproval_reason"] },
+    ],
+  },
+  // schemas/stage-validation.json
+  validate: {
+    required: [
+      { name: "finding_id", type: "string" },
+      { name: "status", type: "string", enum: ["confirmed", "killed", "reported"] },
+      { name: "technique_used", type: "string" },
+      { name: "detection_method", type: "string" },
+    ],
+    conditional: [
+      {
+        when: { field: "status", equals: "confirmed" },
+        require: ["poc_path", "run_log", "evidence_extracted"],
+      },
+      { when: { field: "status", equals: "killed" }, require: ["kill_reason"] },
+    ],
+  },
+  // schemas/stage-chain.json
+  chain: {
+    required: [
+      { name: "chains", type: "array" },
+      { name: "summary", type: "string" },
+    ],
+  },
+  // schemas/stage-report.json
+  report: {
+    required: [
+      { name: "target", type: "string" },
+      { name: "pipeline_status", type: "string", enum: ["complete", "partial", "aborted"] },
+      { name: "findings", type: "array" },
+      { name: "coverage", type: "object" }, // patternProperties object, not array
+      { name: "summary", type: "string" },
+    ],
+  },
+};
+
+const MAX_REPAIR_ATTEMPTS = 2;
+
+// Segment-based test-path detection: matches "test", "__tests__", "specs",
+// "e2e", "test-utils", "fixtures", ... anchored per path segment so
+// "latest"/"contest"/"attest" do NOT match. A regex-only version missed
+// leading underscores ("__tests__").
+const TEST_SEGMENT_RE =
+  /^[._-]*(tests?|specs?|e2e|fixtures?|mocks?|stubs?|examples?|samples?|test[-_]?data|test[-_]?utils)[._-]*$/i;
+const TEST_FILE_RE =
+  /([._-](test|spec|mock|fixture|stub|example|sample)\.[a-z0-9]+$|^test[-_]utils\.[a-z0-9]+$)/i;
+
+/** Chain items: each must have title, severity, steps (≥2), narrative. */
+const CHAIN_SEVERITIES = ["low", "medium", "high", "critical"] as const;
+
+// ── Pre-filter constants (hunt stage only) ───────────────────────────
+
+/**
+ * Test/mock/example paths carry no real findings (mirrors VVAH S5). Exception
+ * from VVAH deliberately not copied: hardcoded-creds-in-test-files — the
+ * auditor can submit those under vuln_class "other"+bugClass documentation;
+ * the gate errs on filtering noise.
+ */
+
+/** Trivial dedup: same file + vuln_class + line within this tolerance. */
+const DEDUP_LINE_TOLERANCE = 10;
+
+// ── Persistence ─────────────────────────────────────────────────────
+
+type FindingRef = {
+  key: string;
+  file: string;
+  line?: number;
+  vuln_class: string;
+};
+
+type SubmitState = {
+  repairs: Record<string, number>;
+  accepted_findings: FindingRef[];
+};
+
+function statePath(runId: string): string {
+  return join(getRunDir(runId), "pipeline-submit.json");
+}
+
+function readState(runId: string): SubmitState {
+  const p = statePath(runId);
+  if (!existsSync(p)) return { repairs: {}, accepted_findings: [] };
+  try {
+    const raw = JSON.parse(readFileSync(p, "utf8")) as Partial<SubmitState>;
+    return {
+      repairs: raw.repairs ?? {},
+      accepted_findings: raw.accepted_findings ?? [],
+    };
+  } catch {
+    return { repairs: {}, accepted_findings: [] };
+  }
+}
+
+function writeState(runId: string, state: SubmitState): void {
+  writeFileSync(statePath(runId), JSON.stringify(state, null, 2), "utf8");
+}
+
+/** Project root containing the scratchpad (file-existence checks resolve here). */
+function projectRoot(): string {
+  return dirname(getScratchpadRoot());
+}
+
+// ── Parsing ──────────────────────────────────────────────────────────
+
+function parseOutput(output: unknown): { obj?: Record<string, unknown>; error?: string } {
+  if (typeof output === "object" && output !== null && !Array.isArray(output)) {
+    return { obj: output as Record<string, unknown> };
+  }
+  if (typeof output !== "string") {
+    return { error: "output must be a JSON object or a JSON string" };
+  }
+  let text = output.trim();
+  // Tolerate markdown code fences around the payload.
+  text = text.replace(/^```(?:json)?\s*\n?/i, "").replace(/\n?```\s*$/i, "");
+  try {
+    const parsed = JSON.parse(text);
+    if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) {
+      return { error: "output must parse to a JSON object" };
+    }
+    return { obj: parsed as Record<string, unknown> };
+  } catch (e) {
+    return { error: `output is not valid JSON: ${(e as Error).message.slice(0, 120)}` };
+  }
+}
+
+/** Stable repair-bucket key for a submission. */
+function submissionKey(stage: SubmitStage, obj: Record<string, unknown>): string {
+  const id =
+    (typeof obj.finding_id === "string" && obj.finding_id) ||
+    (typeof obj.title === "string" && obj.title) ||
+    (typeof obj.id === "string" && obj.id);
+  const tail = id ?? createHash("sha1").update(JSON.stringify(obj)).digest("hex").slice(0, 8);
+  return `${stage}:${tail}`;
+}
+
+// ── Validation ──────────────────────────────────────────────────────
+
+function isNonEmptyString(v: unknown): v is string {
+  return typeof v === "string" && v.trim().length > 0;
+}
+
+function validateStage(stage: SubmitStage, obj: Record<string, unknown>): string[] {
+  const spec = SPECS[stage];
+  const errors: string[] = [];
+
+  for (const field of spec.required) {
+    const v = obj[field.name];
+    if (field.type === "string") {
+      if (!isNonEmptyString(v)) {
+        errors.push(`${field.name}: missing or empty string`);
+        continue;
+      }
+    } else if (field.type === "object") {
+      if (typeof v !== "object" || v === null || Array.isArray(v)) {
+        errors.push(`${field.name}: missing or not an object`);
+        continue;
+      }
+    } else if (field.type === "integer") {
+      if (typeof v !== "number" || !Number.isInteger(v)) {
+        errors.push(`${field.name}: missing or not an integer`);
+        continue;
+      }
+    } else {
+      if (!Array.isArray(v)) {
+        errors.push(`${field.name}: missing or not an array`);
+        continue;
+      }
+      if (field.minItems !== undefined && v.length < field.minItems) {
+        errors.push(`${field.name}: needs at least ${field.minItems} item(s), got ${v.length}`);
+        continue;
+      }
+    }
+    if (field.enum && !field.enum.includes(v as never)) {
+      errors.push(`${field.name}: "${String(v)}" not in { ${field.enum.join(" | ")} }`);
+    }
+  }
+
+  if (spec.locatorXor) {
+    const [a, b] = spec.locatorXor;
+    const hasSet = (set: string[]) =>
+      set.every((f) => (f === "line" ? Number.isInteger(obj[f]) : isNonEmptyString(obj[f])));
+    const hasA = hasSet(a);
+    const hasB = hasSet(b);
+    if (hasA === hasB) {
+      errors.push(
+        `locator: provide exactly one of { ${a.join("+")} } (source) or { ${b.join("+")} } (live)`,
+      );
+    }
+    if (hasA && typeof obj.line === "number" && obj.line < 1) {
+      errors.push("line: must be >= 1");
+    }
+  }
+
+  for (const cond of spec.conditional ?? []) {
+    if (obj[cond.when.field] === cond.when.equals) {
+      for (const name of cond.require) {
+        if (!isNonEmptyString(obj[name])) {
+          errors.push(`${name}: required when ${cond.when.field} = ${cond.when.equals}`);
+        }
+      }
+    }
+  }
+
+  // Chain items have their own inner contract (≥2 steps, severity enum).
+  if (stage === "chain" && Array.isArray(obj.chains)) {
+    obj.chains.forEach((c, i) => {
+      const chain = c as Record<string, unknown>;
+      if (!isNonEmptyString(chain.title)) errors.push(`chains[${i}].title: missing or empty`);
+      if (
+        !isNonEmptyString(chain.severity) ||
+        !(CHAIN_SEVERITIES as readonly string[]).includes(chain.severity)
+      ) {
+        errors.push(`chains[${i}].severity: must be one of { ${CHAIN_SEVERITIES.join(" | ")} }`);
+      }
+      if (!Array.isArray(chain.steps) || chain.steps.length < 2) {
+        errors.push(`chains[${i}].steps: needs at least 2 case IDs`);
+      }
+      if (!isNonEmptyString(chain.narrative))
+        errors.push(`chains[${i}].narrative: missing or empty`);
+    });
+  }
+
+  return errors;
+}
+
+// ── Pre-filter + dedup (hunt only) ──────────────────────────────────
+
+function prefilterHunt(obj: Record<string, unknown>): string | null {
+  const file = typeof obj.file === "string" ? obj.file : undefined;
+  if (!file) return null; // live target: endpoint locator, nothing to filter
+  const normalized = file.replace(/^\.?\//, "");
+  const segments = normalized.split("/");
+  if (segments.some((s) => TEST_SEGMENT_RE.test(s)) || TEST_FILE_RE.test(normalized)) {
+    return (
+      `test-path filter: "${file}" matches test/fixture/mock paths — findings in ` +
+      `test code are noise. If this is a deliberately-shipped test credential, ` +
+      `re-submit documenting why it ships to production.`
+    );
+  }
+  const root = projectRoot();
+  const abs = isAbsolute(normalized) ? resolve(normalized) : resolve(root, normalized);
+  // Containment: resolved path must stay inside the project, otherwise a
+  // "finding" can point at ../ or absolute files outside the target repo.
+  const rel = relative(root, abs);
+  if (rel.startsWith("..") || isAbsolute(rel)) {
+    return (
+      `containment filter: "${file}" resolves outside the project root (${root}). ` +
+      `Findings must reference files inside the target repository.`
+    );
+  }
+  if (!existsSync(abs)) {
+    return (
+      `file-existence filter: "${file}" does not exist under the project root ` +
+      `(${root}). Hallucinated paths are rejected outright.`
+    );
+  }
+  return null;
+}
+
+function dedupHunt(state: SubmitState, obj: Record<string, unknown>): { duplicateOf?: string } {
+  const file = typeof obj.file === "string" ? obj.file.replace(/^\.?\//, "") : undefined;
+  const vulnClass = typeof obj.vuln_class === "string" ? obj.vuln_class : undefined;
+  const line = typeof obj.line === "number" ? obj.line : undefined;
+  if (!file || !vulnClass) return {};
+  for (const accepted of state.accepted_findings) {
+    if (accepted.vuln_class !== vulnClass) continue;
+    if (accepted.file !== file) continue;
+    if (
+      line !== undefined &&
+      accepted.line !== undefined &&
+      Math.abs(line - accepted.line) > DEDUP_LINE_TOLERANCE
+    ) {
+      continue;
+    }
+    return { duplicateOf: accepted.key };
+  }
+  return {};
+}
+
+// ── Public API ───────────────────────────────────────────────────────
+
+const STAGE_TO_PHASE: Record<SubmitStage, ScratchpadPhase> = {
+  hunt: "hunt",
+  trace: "trace",
+  skeptic: "skeptic",
+  validate: "validate",
+  chain: "chain",
+  report: "report",
+};
+
+export function pipeline_submit(runId: string, stage: SubmitStage, output: unknown): SubmitResult {
+  const parsed = parseOutput(output);
+  if (parsed.error || !parsed.obj) {
+    const state = readState(runId);
+    const key = `${stage}:unparseable`;
+    state.repairs[key] = (state.repairs[key] ?? 0) + 1;
+    const attempt = state.repairs[key];
+    // Persist before BOTH returns — otherwise unparseable output bypasses the
+    // repair budget forever (counter never hits disk on the rejected path).
+    writeState(runId, state);
+    if (attempt > MAX_REPAIR_ATTEMPTS) {
+      return { verdict: "rejected", stage, errors: [parsed.error ?? "unparseable"], key };
+    }
+    return {
+      verdict: "repair",
+      stage,
+      errors: [parsed.error ?? "unparseable"],
+      repair_attempt: attempt,
+      key,
+    };
+  }
+
+  const obj = parsed.obj;
+  const key = submissionKey(stage, obj);
+
+  const errors = validateStage(stage, obj);
+  if (errors.length > 0) {
+    const state = readState(runId);
+    state.repairs[key] = (state.repairs[key] ?? 0) + 1;
+    const attempt = state.repairs[key];
+    if (attempt > MAX_REPAIR_ATTEMPTS) {
+      writeState(runId, state);
+      return {
+        verdict: "rejected",
+        stage,
+        errors: [...errors, `repair budget exhausted (${MAX_REPAIR_ATTEMPTS} attempts)`],
+        key,
+      };
+    }
+    writeState(runId, state);
+    return { verdict: "repair", stage, errors, repair_attempt: attempt, key };
+  }
+
+  // Hunt stage: deterministic noise gates before acceptance.
+  if (stage === "hunt") {
+    const filtered = prefilterHunt(obj);
+    if (filtered) {
+      return { verdict: "rejected", stage, errors: [filtered], key };
+    }
+    const state = readState(runId);
+    const { duplicateOf } = dedupHunt(state, obj);
+    if (duplicateOf) {
+      return {
+        verdict: "rejected",
+        stage,
+        errors: [
+          `trivial dedup: same file + vuln_class within ${DEDUP_LINE_TOLERANCE} lines of accepted finding ${duplicateOf}`,
+        ],
+        key,
+        duplicate_of: duplicateOf,
+      };
+    }
+    if (typeof obj.file === "string" && typeof obj.vuln_class === "string") {
+      state.accepted_findings.push({
+        key,
+        file: obj.file.replace(/^\.?\//, ""),
+        line: typeof obj.line === "number" ? obj.line : undefined,
+        vuln_class: obj.vuln_class,
+      });
+      writeState(runId, state);
+    }
+  }
+
+  const artifact = scratchpad_write(
+    runId,
+    STAGE_TO_PHASE[stage],
+    `${key.replace(/[^a-zA-Z0-9._:-]/g, "_")}.json`,
+    JSON.stringify(obj, null, 2),
+  );
+  return { verdict: "accepted", stage, errors: [], key, artifact };
+}

package/src/workflow.ts

@@ -1,7 +1,7 @@
 /**
  * Cyber workflow injected into agent context when XP mode is ON.
  *
- * Skills (pipeline, web-pentest) already cover tool usage and methodology.
+ * Skills (cyberwf, web-pentest) already cover tool usage and methodology.
  * This file adds the unique attacker discipline: state machine with
  * preconditions, attacker model, impact validation, adversarial review,
  * kill checklist, and report-readiness criteria.
@@ -19,7 +19,7 @@ Every lead starts HYPOTHESIS. Nothing reaches CONFIRMED without a proven attacke
 
 ## Tool Reference
 
-**Casefile (state tracking):** CaseAdd, CaseUpdate, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseReport, PromoteFinding
+**Casefile (state tracking):** CaseAdd, CaseUpdate, CaseGet, CaseList, CaseSearch, CaseLink, CaseUnlink, CaseReport, PromoteFinding, PipelineSubmit
 
 **Scratchpad (pipeline artifacts):** ScratchpadInit, ScratchpadResume, ScratchpadCheckpoint, ScratchpadWrite, ScratchpadRead, ScratchpadPhaseDone, ScratchpadClear