@xaccefy/pi-casefile 0.9.0 → 0.9.2

package/src/scratchpad.ts

@@ -16,7 +16,7 @@
  *     verify/     — PoC logs, run outputs (validate phase)
  *     chain/      — exploit-chain analysis
  *     patch/      — remediation work
- *     report/     — report-writer context
+ *     report/     — final report context
  *     state.json  — checkpoint file with phase completion + key IDs
  *
  * Resume re-reads scratchpad artifacts; it does not re-run completed phases
@@ -25,16 +25,15 @@
  */
 
 import { createHash } from "node:crypto";
+import { existsSync, readdirSync, rmSync } from "node:fs";
+import { basename, dirname, join, resolve } from "node:path";
 import {
-  existsSync,
-  mkdirSync,
-  readdirSync,
-  readFileSync,
-  renameSync,
-  rmSync,
-  writeFileSync,
-} from "node:fs";
-import { dirname, join, resolve } from "node:path";
+  assertSafeRegularFile,
+  assertSafeStateDirectory,
+  ensureSafeStateDirectory,
+  readSafeFile,
+  writeSafeFileAtomic,
+} from "./safe-state.ts";
 
 // ── Types ────────────────────────────────────────────────────────────
 
@@ -74,7 +73,9 @@ export interface ScratchpadResume {
 
 // ── Constants ────────────────────────────────────────────────────────
 
-export const PHASE_ORDER: ScratchpadPhase[] = [
+// All accepted artifact buckets. Some are legacy/manual-only and should not be
+// scheduled by ScratchpadResume for new swarm runs.
+export const SCRATCHPAD_PHASES: ScratchpadPhase[] = [
   "recon",
   "hunt",
   "gapfil",
@@ -86,6 +87,17 @@ export const PHASE_ORDER: ScratchpadPhase[] = [
   "report",
 ];
 
+// Active pipeline order for new/resumed runs.
+export const PHASE_ORDER: ScratchpadPhase[] = [
+  "recon",
+  "hunt",
+  "trace",
+  "skeptic",
+  "validate",
+  "chain",
+  "report",
+];
+
 const PHASE_DIRS: Record<ScratchpadPhase, string> = {
   recon: "recon",
   hunt: "hunt",
@@ -215,17 +227,23 @@ function emptyCheckpoint(runId: string, projectRoot: string): ScratchpadCheckpoi
 }
 
 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 });
+  const scratchpadRoot = dirname(runDir);
+  const projectRoot = dirname(scratchpadRoot);
+  const runName = basename(runDir);
+  ensureSafeStateDirectory(projectRoot, [SCRATCHPAD_DIR, runName]);
+  for (const phase of SCRATCHPAD_PHASES) {
+    ensureSafeStateDirectory(projectRoot, [SCRATCHPAD_DIR, runName, PHASE_DIRS[phase]]);
   }
 }
 
 function readCheckpointRaw(runId: string, projectRoot?: string): ScratchpadCheckpoint | null {
+  const root = projectRoot ?? detectWorkspaceRoot();
   const statePath = getStatePath(runId, projectRoot);
-  if (!existsSync(statePath)) return null;
-  const raw = readFileSync(statePath, "utf8");
+  if (existsSync(statePath)) {
+    assertSafeStateDirectory(root, [SCRATCHPAD_DIR, runDirName(runId)]);
+  }
+  if (!assertSafeRegularFile(statePath, "Scratchpad checkpoint")) return null;
+  const raw = readSafeFile(statePath, "Scratchpad checkpoint").toString("utf8");
   const cp = JSON.parse(raw) as ScratchpadCheckpoint;
   if (typeof cp !== "object" || cp === null || Array.isArray(cp)) {
     throw new Error(`Corrupt scratchpad state for ${runId}: root must be an object`);
@@ -237,7 +255,7 @@ function readCheckpointRaw(runId: string, projectRoot?: string): ScratchpadCheck
     throw new Error(`Corrupt scratchpad state for ${runId}: completed_phases must be an array`);
   }
   for (const phase of cp.completed_phases) {
-    if (!PHASE_ORDER.includes(phase)) {
+    if (!SCRATCHPAD_PHASES.includes(phase)) {
       throw new Error(`Corrupt scratchpad state for ${runId}: invalid phase ${phase}`);
     }
   }
@@ -258,9 +276,7 @@ function writeCheckpointRaw(cp: ScratchpadCheckpoint, projectRoot?: string): voi
   cp.last_updated = new Date().toISOString();
   const statePath = getStatePath(cp.run_id, projectRoot);
   ensureRunDirs(getRunDir(cp.run_id, projectRoot));
-  const tmp = `${statePath}.${process.pid}.${Date.now()}.tmp`;
-  writeFileSync(tmp, JSON.stringify(cp, null, 2), "utf8");
-  renameSync(tmp, statePath);
+  writeSafeFileAtomic(statePath, JSON.stringify(cp, null, 2));
 }
 
 // ── Public API ───────────────────────────────────────────────────────
@@ -309,7 +325,7 @@ export function scratchpad_write(
   const safeName = artifactFileName(artifactName);
   const dir = join(runDir, PHASE_DIRS[phase]);
   const filePath = join(dir, safeName);
-  writeFileSync(filePath, content, "utf8");
+  writeSafeFileAtomic(filePath, content);
   return filePath;
 }
 
@@ -325,8 +341,11 @@ export function scratchpad_read(
   const root = projectRoot ?? detectWorkspaceRoot();
   const safeName = artifactFileName(artifactName);
   const filePath = join(getRunDir(runId, root), PHASE_DIRS[phase], safeName);
-  if (!existsSync(filePath)) return null;
-  return readFileSync(filePath, "utf8");
+  if (existsSync(filePath)) {
+    assertSafeStateDirectory(root, [SCRATCHPAD_DIR, runDirName(runId), PHASE_DIRS[phase]]);
+  }
+  if (!assertSafeRegularFile(filePath, "Scratchpad artifact")) return null;
+  return readSafeFile(filePath, "Scratchpad artifact").toString("utf8");
 }
 
 /**
@@ -335,13 +354,16 @@ export function scratchpad_read(
 export function scratchpad_runs(projectRoot?: string): string[] {
   const root = getScratchpadRoot(projectRoot);
   if (!existsSync(root)) return [];
+  assertSafeStateDirectory(dirname(root), [SCRATCHPAD_DIR]);
   const out: string[] = [];
   for (const entry of readdirSync(root, { withFileTypes: true })) {
     if (!entry.isDirectory()) continue;
     const state = join(root, entry.name, "state.json");
-    if (!existsSync(state)) continue;
+    if (!assertSafeRegularFile(state, "Scratchpad checkpoint")) continue;
     try {
-      const cp = JSON.parse(readFileSync(state, "utf8")) as { run_id?: unknown };
+      const cp = JSON.parse(readSafeFile(state, "Scratchpad checkpoint").toString("utf8")) as {
+        run_id?: unknown;
+      };
       if (typeof cp.run_id !== "string") continue;
       if (getRunDir(cp.run_id, projectRoot) === join(root, entry.name)) {
         out.push(cp.run_id);
@@ -366,7 +388,10 @@ export function scratchpad_list(
   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");
+  assertSafeStateDirectory(root, [SCRATCHPAD_DIR, runDirName(runId), PHASE_DIRS[phase]]);
+  return readdirSync(dir).filter(
+    (f) => f !== "state.json" && assertSafeRegularFile(join(dir, f), "Scratchpad artifact"),
+  );
 }
 
 /**
@@ -387,7 +412,7 @@ export function scratchpad_checkpoint(
   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.completed_phases.sort((a, b) => SCRATCHPAD_PHASES.indexOf(a) - SCRATCHPAD_PHASES.indexOf(b));
   }
   cp.last_phase_at = new Date().toISOString();
   if (data.ids) cp.phase_ids[phase] = data.ids;
@@ -437,5 +462,8 @@ export function scratchpad_phase_done(
 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 });
+  if (existsSync(runDir)) {
+    assertSafeStateDirectory(root, [SCRATCHPAD_DIR, runDirName(runId)]);
+    rmSync(runDir, { recursive: true, force: true });
+  }
 }

package/src/workflow.ts

@@ -1,5 +1,5 @@
 /**
- * Cyber workflow injected into agent context when XP mode is ON.
+ * Cyber workflow injected into agent context when XP mode is SWARM.
  *
  * Skills (cyberwf, web-pentest) cover tool usage and methodology. This file
  * adds the unique attacker discipline: state machine with preconditions,
@@ -39,42 +39,30 @@ type DispatchSpec = {
   hardGate: string;
   /** Crash-handling paragraph. */
   crash: string;
-  /** Skeptic dispatch snippet (follows "dispatch it BEFORE the exploit agent with "). */
+  /** Skeptic dispatch snippet (follows "dispatch it before main-agent validation with "). */
   skeptic: string;
-  /** Confirmer dispatch snippet (follows "dispatch the confirmer: "). */
-  confirmer: string;
-  /** Reporter dispatch snippet (follows "Dispatch the reporter subagent with "). */
-  reporter: string;
 };
 
 const PI_DISPATCH: DispatchSpec = {
   reference:
-    "**Subagent dispatch:** every launch uses `subagent({ workflowScript: \"return runs.run('stable-key', { agent: 'tracer', task: '...' })\", context: 'fresh', async: true })`. Parallel HUNT uses one workflowScript with `return runs.all([{ key: 'run-class-attempt', agent: 'auditor', task: '...' }, ...])`. Stable keys include run, stage, class/case, and attempt. Dispatch specialists; do NOT do their work yourself.",
+    "**Subagent dispatch:** every launch uses `subagent({ workflowScript: \"return runs.run('stable-key', { agent: 'tracer', task: '...' })\", context: 'fresh', async: true })`. Parallel HUNT uses one workflowScript with `return runs.all([...])`, capped at 3 auditor tasks per round. Batch related classes by surface/family instead of spawning one agent per bug class. Stable keys include run, stage, batch/case, and attempt.",
   hardGate:
-    "Your next tool call MUST launch one async workflowScript whose `runs.all([...])` dispatches HUNT auditors.",
+    "Your next tool call MUST launch one async workflowScript whose `runs.all([...])` dispatches at most 3 batched HUNT auditors.",
   crash:
-    "**Subagent crash handling:** a crash (SIGABRT, OOM, timeout) is a RETRY, not a verdict. Launch one new workflowScript with the same specialist task, a new stable attempt key, and a stronger model. Crash again → record `blocked: <agent> crashed` in the pipeline-run case and continue; never silently drop the stage.",
+    "**Subagent failure handling:** a crash, timeout, hung run, unparseable output, or schema-invalid output is a RETRY, not a verdict. Launch one new workflowScript with the same specialist task, a new stable attempt key, and a stronger model. Failure again → record `blocked: <agent> failed` in the pipeline-run case and continue; never silently drop the stage.",
   skeptic:
     "`subagent({ workflowScript: \"return runs.run('skeptic-<case>-1', { agent: 'skeptic', task: '...' })\", context: 'fresh', async: true })`",
-  confirmer:
-    "`subagent({ workflowScript: \"return runs.run('confirm-<case>-1', { agent: 'confirmer', task: 'Verify the PoC evidence for case <id> (poc_path=..., control_target=..., evidence_sha256=..., poc_sha256=...). Assume fabricated, prove real. Re-send the verify request yourself. Return the verdict.' })\", context: 'fresh', async: true })`",
-  reporter:
-    "`subagent({ workflowScript: \"return runs.run('report-<case>-1', { agent: 'reporter', task: 'Write the final report. case_id=<id>, context_path=<context path>, report_path=<report path>, program_name=<if known>.' })\", context: 'fresh', async: true })`",
 };
 
 const OMP_DISPATCH: DispatchSpec = {
   reference:
-    "**Subagent dispatch (OMP):** every launch uses `task({ context: 'fresh', tasks: [{ name: 'stable-key', agent: 'tracer', task: '...' }] })`. Parallel HUNT dispatches ONE task call whose `tasks` array carries one entry per attack class: `task({ context: 'fresh', tasks: [{ name: 'hunt-sqli-1', agent: 'auditor', task: '...' }, { name: 'hunt-xss-1', agent: 'auditor', task: '...' }] })`. Stable names include run, stage, class/case, and attempt. Results deliver automatically; steer with `hub`. Dispatch specialists; do NOT do their work yourself.",
+    "**Subagent dispatch (OMP):** every launch uses `task({ context: 'fresh', tasks: [{ name: 'stable-key', agent: 'tracer', task: '...' }] })`. Parallel HUNT dispatches ONE task call whose `tasks` array carries at most 3 auditor tasks per round, batching related classes by surface/family instead of spawning one agent per bug class. Stable names include run, stage, batch/case, and attempt. Results deliver automatically; steer with `hub`.",
   hardGate:
-    "Your next tool call MUST launch one async `task` call whose `tasks` array dispatches HUNT auditors (one entry per attack class). When their results are delivered, submit each output through PipelineSubmit.",
+    "Your next tool call MUST launch one async `task` call whose `tasks` array dispatches at most 3 batched HUNT auditors. When their results are delivered, submit each output through PipelineSubmit.",
   crash:
-    "**Subagent crash handling:** a failed or hung task (SIGABRT, OOM, timeout) is a RETRY, not a verdict. Re-dispatch the same specialist task with a new attempt name and a stronger model. Crash again → record `blocked: <agent> crashed` in the pipeline-run case and continue; never silently drop the stage.",
+    "**Subagent failure handling:** a failed or hung task, timeout, unparseable output, or schema-invalid output is a RETRY, not a verdict. Re-dispatch the same specialist task with a new attempt name and a stronger model. Failure again → record `blocked: <agent> failed` in the pipeline-run case and continue; never silently drop the stage.",
   skeptic:
     "`task({ context: 'fresh', tasks: [{ name: 'skeptic-<case>-1', agent: 'skeptic', task: '...' }] })`",
-  confirmer:
-    "`task({ context: 'fresh', tasks: [{ name: 'confirm-<case>-1', agent: 'confirmer', task: 'Verify the PoC evidence for case <id> (poc_path=..., control_target=..., evidence_sha256=..., poc_sha256=...). Assume fabricated, prove real. Re-send the verify request yourself. Return the verdict.' }] })`",
-  reporter:
-    "`task({ context: 'fresh', tasks: [{ name: 'report-<case>-1', agent: 'reporter', task: 'Write the final report. case_id=<id>, context_path=<context path>, report_path=<report path>, program_name=<if known>.' }] })`",
 };
 
 /** Build the full cyber workflow for a host's dispatch convention. */
@@ -94,14 +82,28 @@ Think like a real external attacker, not a code reviewer. Technical bugs are che
 
 ${d.reference}
 
+**Swarm delegation boundary:** only auditor (HUNT rounds), tracer (TRACE), skeptic (high-confidence challenge), and chain (CHAIN) run as subagents. You, the main coordinator, own RECON, VALIDATE/PoC writing, ConfirmFinding, patching, final reports, state decisions, and all orchestration.
+
 ## Stage Machine (run in order — you are the coordinator)
 
-RECON (you, inline) → **HUNT** (auditor subagents, one per attack class, parallel) → TRACE (tracer) → SKEPTIC (high-confidence only) → VALIDATE (exploit) → CHAIN (chain) → REPORT (reporter)
+RECON (you, inline) → **HUNT** (2-3 batched auditor subagents) → TRACE (tracer for prioritized findings) → SKEPTIC (bounded high-risk review) → VALIDATE (you, inline) → CHAIN (chain subagent) → REPORT (you, inline)
+
+### Blackbox recon — attack-surface mapping (no source access)
+
+Live web target, CTF, or bounty box: RECON aggressively gathers high-signal intel and turns it into the map HUNT will use: entry-point inventory, attacker model, auth/role boundaries, trust boundaries, likely vuln-class batches, and known gaps. Aim for the richest useful map, not the largest raw pile. Choose the next recon move from the target and current unknowns. Use JS/source maps, \`robots.txt\`, \`sitemap.xml\`, \`/.well-known/\`, OpenAPI/Swagger, GraphQL introspection, passive archives, and exposed backup/VCS checks when they are likely to change class selection, target selection, or attacker modeling. Stop when additional collection is unlikely to change the HUNT plan; re-enter RECON when HUNT/TRACE exposes missing surface.
+- **Client-side code can be high signal** — pull JS bundles/source maps when the app is SPA/API-heavy or routes are hidden, then bank discovered endpoints, params, and secrets as leads.
+- **Zero-traffic intel is optional, not ritual** — check public metadata, schemas, and passive archives when scope allows and the result can change target selection, auth modeling, or class selection.
+- **Fingerprint for decisions** — stack + version confidence should drive \`exploit_search\` and HUNT class selection; record uncertainty instead of forcing a guess.
+- **Bank useful leads** — write the entry-point map, selected HUNT class batches, and open gaps to the scratchpad; file high-value leaks (source map, origin IP, exposed schema, leaked creds) as \`EvidenceAdd role=observation\`. Tactical commands: web-pentest skill §2.
+
+**Observe behavior, then analyze — static intel is only half.** Interrogate the target empirically and infer its internals from how it *reacts*; the differential (vary one input, watch what changes) is the signal. **Web/API:** status vs length vs timing vs body vs error across crafted inputs; how auth actually gates (401 vs 302 vs 200-with-error); reflected vs stored; timing oracles for blind bugs; state changes across a request sequence. **Binary/local target:** map the I/O contract, trace syscalls + library calls (\`strace\`/\`ltrace\`), feed malformed/boundary input and watch crashes, signals, and return codes, and diff behavior across inputs to expose the parse/branch logic. **Protocol/service:** walk the handshake + state machine, then replay and mutate one field and observe the divergence and side effects. Loop: stimulus → observe → infer the internal model → craft a discriminating probe → repeat. Every observed anomaly (crash, error leak, timing gap, unexpected 200, state change) is a HYPOTHESIS — \`CaseAdd\` it with its \`disproveIf\`, don't just note it.
 
 **HARD GATE — after RECON:** record the entry-point inventory, then STOP all inline reading/probing. ${d.hardGate} When its completion is delivered, submit each output through PipelineSubmit. If you catch yourself mapping a sink, reading a handler, or probing an endpoint beyond the recon inventory, stop and add it to a HUNT task.
 
 ${d.crash}
 
+**TRACE verdicts:** only schema-valid \`trace_result: "REACHABLE"\` advances toward validation. \`UNREACHABLE\` requires a concrete blocker for the stated attacker model; \`UNDETERMINED\` means missing context/auth/WAF/source ambiguity and blocks or re-dispatches, never kills.
+
 ## Case Lifecycle (State Machine)
 ${LIFECYCLE_DIAGRAM}
 
@@ -111,18 +113,19 @@ ${LIFECYCLE_DIAGRAM}
 |-------|-----------|-------------|
 | RECON | (none) | Map attack surface, fingerprint, search CVEs. Something interesting → HYPOTHESIS. |
 | HUNT | HYPOTHESIS | Document the lead (impact not required yet). Clear intended-behavior/artifact → KILLED; else INVESTIGATING. |
-| CHAIN | INVESTIGATING | Test the hypothesis, chain primitives, build PoC. Explore combinations (open redirect + SSRF, leak + endpoint, …). |
-| VALIDATE | CONFIRMED | Prove impact, adversarial review, root-cause trace. Survive the gates below or fall back to INVESTIGATING / KILLED. |
-| REPORT | REPORTED | CaseContext → reporter agent → report-readiness gate. |
+| TRACE / SKEPTIC / VALIDATE | INVESTIGATING | Trace reachability, attempt disconfirmation, and produce the pending PoC evidence bundle. Failure stays INVESTIGATING or becomes KILLED. |
+| MAIN REVIEW | CONFIRMED | The main agent judges whether the machine differential actually establishes the vulnerability and impact, then commits through ConfirmFinding. |
+| CHAIN | CONFIRMED | Link confirmed findings and evaluate multi-step exploit paths; this stage does not confirm new cases. |
+| REPORT | REPORTED | CaseContext → main-agent report writing → report-readiness gate. |
 
 ### Preconditions Per State Transition (MANDATORY)
 
 | Advance To | Required Case Fields | On Disk |
 |-----------|---------------------|---------|
 | HYPOTHESIS → INVESTIGATING | evidence (observations), confidence | Notes on what was observed |
-| INVESTIGATING → **CONFIRMED** | evidence, poc, **impact** (content below), severity, **target**, **disconfirmation** (the confirmer's documented disprove attempt) | PromoteFinding phase 1: PoC runs 2× against target + 1× against a distinct \`control_target\` (same script, sha256-enforced); every run completes with output fully captured and writes nonce-bound \`evidence.json\` to \`$PI_POC_EVIDENCE_DIR\`; target runs are deterministic and the control evidence differs from the target's (machine-checked). Then dispatch the **confirmer** subagent and commit its verdict with **ConfirmFinding**: CONFIRMED requires the confirmer to have re-sent the verify request itself (\`re_executed: true\`), a \`target_only\` differential, and its own \`disconfirmation_attempt\` (becomes the case's disconfirmation). Exit codes and output markers are diagnostics, not gates. |
+| INVESTIGATING → **CONFIRMED** | evidence, poc, **impact** (content below), severity, **target**, **disconfirmation** (the main agent's documented disprove attempt) | PromoteFinding phase 1: PoC runs 2× against target + 1× against an operator-approved \`control_target\` (same script, sha256-enforced); every run completes at exit zero with output fully captured and writes nonce-bound \`evidence.json\` with a response-body predicate; the harness obtains conclusive target/control responses and requires \`target_only\`. Then the **main/coordinator agent itself** reviews and calls **ConfirmFinding**, which captures a fresh second harness replay before commit. Worker agents cannot submit phase 2. Zero exit is necessary run integrity, never vulnerability proof; output markers are diagnostic only. |
 | Any → KILLED | assumptions (why it died) | — |
-| CONFIRMED → REPORTED | CaseContext(id) succeeded (records report path) AND the reporter agent wrote the report file | Context bundle + report file |
+| CONFIRMED → REPORTED | CaseContext(id) succeeded (records report path) AND you wrote the report file | Context bundle + report file |
 
 **Empty required field = you cannot advance.** The fields ARE the gates.
 
@@ -174,7 +177,7 @@ If you cannot name a concrete attacker who gains something they should not have
 
 The finding must survive an attempt to disprove it. Two tiers, gated on \`confidence\` (severity comes later, from the PoC):
 
-**\`confidence: high\` → skeptic subagent (MANDATORY):** dispatch it BEFORE the exploit agent with ${d.skeptic}. It independently re-reads the source (or re-probes live), verifies scope, tries to disprove, and audits the PoC file for cheats. Its schema-validated verdict must carry its own \`disconfirmation_attempt\` (CONFIRMED verdicts without one are rejected by PipelineSubmit). DISPROVEN → add EvidenceAdd role=refutation, then killed directly, no tie-breaker. Do NOT skip; do NOT self-disconfirm high-confidence findings.
+**\`confidence: high\` → skeptic subagent (MANDATORY):** dispatch it before main-agent validation with ${d.skeptic}. It independently re-reads the source (or re-probes live), verifies scope, tries to disprove, and audits any PoC file you already have for cheats. Its schema-validated CONFIRMED verdict must carry its own \`disconfirmation_attempt\` (CONFIRMED verdicts without one are rejected by PipelineSubmit). DISPROVEN → add EvidenceAdd role=refutation, then killed directly, no tie-breaker. UNDETERMINED → block/re-dispatch; do not validate yet. Do NOT skip; do NOT self-disconfirm high-confidence findings.
 
 **Below high → self-disconfirmation:** actively try to disprove your own finding; document it (see the strong/weak example below). Not a formality.
 
@@ -183,15 +186,17 @@ An attempt: reproduce under different conditions (auth/config/network position);
 Strong example: "Read /api/users/123 as user B after confirming user A owns 123 → 403. Repeated with X-Override-User header (seen in admin traffic) → user A's data returned. Protection bypassed via the admin header."
 Weak: "Tried to disprove. Could not." — insufficient.
 
-**The CONFIRMED disconfirmation comes from the confirmer, not a script.** There is no \`disconfirmation_path\` gate: the confirmer subagent (fresh context, different model, dispatched between PromoteFinding and ConfirmFinding) must re-send the verify request itself and write its own failed disproof attempt, which becomes the case's \`disconfirmation\`. A case whose promotion reached CONFIRMED without the confirmer's disconfirmation_attempt is rejected by the ledger.
+**The CONFIRMED disconfirmation comes from the main agent, not a script or worker.** There is no \`disconfirmation_path\` gate: after PromoteFinding, the main/coordinator must write its own failed disproof attempt, which becomes the case's \`disconfirmation\`, and call ConfirmFinding to capture the fresh phase-2 replay. A worker/subagent cannot call PromoteFinding or ConfirmFinding, and a verdict without the main agent's \`disconfirmation_attempt\` is rejected.
 
 **Evidence chain closure (before PromoteFinding):** promotion is rejected unless the case carries an **artifact-backed** \`observation\` evidence item (EvidenceAdd role=observation with \`artifact_path\` — the initial signal, stored with its SHA-256) in addition to the auto-recorded reproduction item. Record observations as you go, not at promote time.
 
-**PromoteFinding (phase 1) — evidence bundle, not markers.** Call it with \`poc_path\`, \`control_path\` (the SAME bytes as the PoC — sha256-equality is enforced), a distinct \`control_target\`, and \`local: true\` when the bug needs network (host-network sandbox; bare host execution still needs operator \`PI_POC_ALLOW_LOCAL=1\`). The harness runs the PoC twice against the case target and once against \`control_target\`. Every run must complete with fully captured output and write nonce-bound \`evidence.json\` to \`$PI_POC_EVIDENCE_DIR\` (\`{"nonce" (echo $PI_POC_NONCE), "claim", "verify": {method, url, headers?, body?, expect: {status/body_contains/body_regex}}, "observations"}\`). The machine gate checks: completion + output completeness, nonce binding, determinism across the two target runs, and that the control evidence differs from the target's (not target-dependent → blocked). Exit codes and output markers are DIAGNOSTICS — a PoC that exits 0 but writes no (or misnonced) evidence is blocked.
+**Main-agent validation only:** do not dispatch validation. You write the smallest reliable PoC that demonstrates the **maximum reachable impact** of the vulnerability, set the case's poc/evidence/impact/severity/target fields, and run PromoteFinding yourself. "Smallest" means no fragile ceremony, mocks, or unrelated exploit steps — not a weaker impact demonstration. Do not stop at a benign marker if a stronger in-scope, non-destructive primitive is reachable (read/write, privilege change, account takeover path, data exposure, etc.). If the PoC fails, refine it yourself up to the local budget; if proof cannot meet the gate, kill or keep the case investigating with the exact blocker.
+
+**PromoteFinding (phase 1) — evidence bundle, not markers.** Call it with \`poc_path\`, an operator-approved \`control_target\` from \`PI_POC_CONTROL_TARGETS\`, optional same-byte \`control_path\` (defaults to \`poc_path\`), and \`local: true\` when the bug needs network. Every run must complete with fully captured output and write nonce-bound \`evidence.json\` whose \`expect\` includes \`body_contains\` or \`body_regex\`; status-only evidence is rejected. The harness pins DNS at connect time, keeps redirects on the bound host, sends the same request to target/control, and requires two conclusive responses with \`target_only\`. Private replay requires operator authorization. Blind/OOB classes fail closed until a source-separated oracle exists.
 
-**ConfirmFinding (phase 2) — the confirmer's verdict commits.** After PromoteFinding succeeds, dispatch the confirmer: ${d.confirmer} — then commit its verdict with \`ConfirmFinding(case_id, verdict)\`. CONFIRMED requires: the confirmer re-sent the verify request (\`re_executed: true\`), \`differential: "target_only"\`, and its own \`disconfirmation_attempt\`. NOT_CONFIRMED keeps the case investigating (attempt recorded) — no tie-breaker. **Never \`CaseUpdate(status: "confirmed")\` directly — it is rejected.**
+**ConfirmFinding (phase 2) — main-agent-only commit.** After PromoteFinding succeeds, do not dispatch confirmation. The main/coordinator agent must inspect the exact PoC/evidence, hunt trivial predicates/fabrication, attempt disconfirmation, and call \`ConfirmFinding(case_id, verdict)\` itself. A CONFIRMED call performs and stores a fresh harness-owned target/control replay; a caller-supplied re-execution checkbox is not accepted. CONFIRMED requires \`re_execution_note\`, \`differential: "target_only"\`, and the main agent's \`disconfirmation_attempt\`. Worker processes are rejected. **Never \`CaseUpdate(status: "confirmed")\` directly.**
 
-**PoC audit (anti-cheat, before PromoteFinding):** have an independent eye on the PoC script itself. For \`confidence: high\` findings the skeptic agent re-reads the PoC file (not just the source) hunting for: unconditional marker prints, trivially-true checks (accepting any 200, grepping for always-present strings), hardcoded expected values, and local mocks of the target. Record the audit result as an EvidenceAdd \`observation\` item (or \`refutation\` if it found a cheat → kill). The model that writes the check must not be the only one that reads it — the confirmer re-reads the script at confirm time. The deterministic backstops are code, not prompts: run completion + output capture, nonce binding, determinism, the evidence differential, the same-file control sha256, PoC byte-identity re-check at commit, and the confirmer's independent re-execution.
+**PoC audit (anti-cheat, before PromoteFinding):** have an independent eye on the PoC script itself. For \`confidence: high\` findings the skeptic agent re-reads the PoC file hunting unconditional success, trivial checks, constants, and local mocks. Record the audit as EvidenceAdd \`observation\` (or \`refutation\` if cheated). The main agent must re-read the exact script before ConfirmFinding; workers may challenge evidence but never run validation or decide promotion. Deterministic backstops are code: output completeness, nonce binding, response-body predicates, deterministic runs, operator-approved control, DNS-pinned conclusive replay, same-file sha256, and PoC byte-identity re-check at commit.
 
 ### 2. Design & Runtime Check — non-intentionality gate (mandatory)
 
@@ -236,14 +241,14 @@ Prove at least **one** real attacker-facing violation against a production-viabl
 
 Impact text answers: *who is hurt, what is lost, how the attacker reaches it from production.* Theoretical impact, a second unproven bug, or unreachable-from-attacker → stay INVESTIGATING (chain it) or KILL.
 
-**Severity is derived from PROVEN impact, not guessed** — set only after the PoC exits 0 and its output demonstrates the impact:
-- **critical** = RCE, account takeover, or direct fund theft (in PoC output)
+**Severity is derived from PROVEN impact, not guessed** — set only after the machine differential and the main-agent review demonstrate the impact; a zero exit or PoC output alone is insufficient:
+- **critical** = RCE, account takeover, or direct fund theft demonstrated in confirmed evidence
 - **high** = sensitive data read/write, privilege escalation, SSRF to internal services
 - **medium** = limited data exposure, XSS on sensitive page, IDOR on non-critical resources
 - **low** = info leak, open redirect, self-only impact with a victim path
 - **info** = best-practice gap, no demonstrated impact
 
-"Could lead to"/"may allow"/"theoretically" = NOT proven — drop to what the PoC output shows. Under-claiming is safe; over-claiming gets rejected at triage.
+"Could lead to"/"may allow"/"theoretically" = NOT proven — drop to what the confirmed harness evidence shows. Claim the highest impact the harness and main-agent review actually prove; unsupported escalation gets rejected at triage.
 
 ### 7. Adversarial Self-Review
 
@@ -266,8 +271,8 @@ Reproduce at least twice or via two methods.
 ## At REPORT
 
 1. **Run CaseContext(case_id)** — writes the context bundle (complete record, PoC + disconfirmation logs, links, pipeline artifacts) and records the report path.
-2. **Dispatch the reporter subagent** with ${d.reporter}. It writes the polished report and flips the case to REPORTED.
-3. **Report-readiness gate** (YOU check this on the reporter's output before accepting; on failure, re-dispatch with the gap list):
+2. **Write the report yourself** at the returned report path using the context bundle. In XP swarm, reporting stays with the main agent.
+3. **Report-readiness gate** (YOU check this before accepting; on failure, edit the report yourself):
 - Deterministic reproduction by another researcher
 - Steps realistic in production
 - Impact justified without inflation (would the vendor agree?)
@@ -301,7 +306,7 @@ export const STATIC_CYBER_WORKFLOW_OMP = buildCyberWorkflow(OMP_DISPATCH);
 export const STATIC_CYBER_WORKFLOW_LITE = `
 # Cyber Workflow — LITE (Single-Agent)
 
-You are the ONLY agent. Do NOT dispatch subagents (no auditor, tracer, skeptic, exploit, or chain agents). You do every stage yourself, inline: recon, hunt, trace, validate, chain, report — the full attacker discipline without subagent orchestration overhead. Great for CTF and focused single-target engagements.
+You are the ONLY agent. Do NOT dispatch subagents (no auditor, tracer, skeptic, or chain agents). You do every stage yourself, inline: recon, hunt, trace, validate, chain, report — the full attacker discipline without subagent orchestration overhead. Great for CTF and focused single-target engagements.
 
 Think like a real external attacker, not a code reviewer. Technical bugs are cheap; **reachable attacker impact** is what matters.
 
@@ -320,28 +325,28 @@ ${LIFECYCLE_DIAGRAM}
 
 ## Stage discipline (all done by you, inline)
 
-1. **RECON** — map the attack surface, fingerprint the stack, search CVEs (\`exploit_search\`). Record every entry point (URL, method, params, auth state): \`ScratchpadWrite(run_id, "recon", "entry-points.md", ...)\`.
-2. **HUNT** — for each attack class, examine every entry point. \`CaseAdd\` each lead as a hypothesis. Track coverage per class.
-3. **TRACE** — prove reachability yourself: read the source (grep/find) or probe the live endpoint (\`http_request\`). Only reachable findings advance.
-4. **VALIDATE** — write a PoC that emits nonce-bound \`evidence.json\`, run it via \`PromoteFinding\` (2 target runs + same-script control), re-send the verify request yourself, and commit via \`ConfirmFinding\` (see the gates below). Derive severity from the proven impact.
+1. **RECON — attack-surface mapping.** Blackbox/CTF: aggressively gather high-signal intel and turn it into entry points, auth models, trust boundaries, attacker model, vuln-class batches, and gaps. Fingerprint credible stack/version signals and search CVEs (\`exploit_search\`) when the version confidence is useful. Use JS/source maps, \`robots.txt\`, \`sitemap.xml\`, \`/.well-known/\`, OpenAPI/Swagger, GraphQL introspection, exposed backup/VCS checks, and passive archives when they can change class selection, target selection, or attacker modeling. Record discovered entry points (URL, method, params, auth state), selected class targets, and gaps/assumptions: \`ScratchpadWrite(run_id, "recon", "entry-points.md", ...)\`.
+2. **HUNT** — choose attack classes from recon and examine relevant entry points. \`CaseAdd\` each lead as a hypothesis. Track coverage per class.
+3. **TRACE / observe** — prove reachability and understand the mechanism by observing how the target behaves, then analyzing the reaction. Read the source (grep/find); probe the live endpoint (\`http_request\`) and diff responses (status vs length vs timing vs error) as you vary one input; or for a binary/local target trace syscalls + library calls (\`strace\`/\`ltrace\`) and watch crashes, signals, and return codes under malformed/boundary input. Infer the internal model from the differential, feed anomalies back as hypotheses, and only advance reachable findings.
+4. **VALIDATE** — write a PoC that emits nonce-bound \`evidence.json\`, run it via \`PromoteFinding\` (2 target runs + same-script control), review and disconfirm it yourself, and commit via \`ConfirmFinding\`, which performs the fresh phase-2 replay (see the gates below). Derive severity from the proven impact.
 5. **CHAIN** — link confirmed findings via \`CaseLink\` to find exploit chains.
-6. **REPORT** — run \`CaseContext\` to write the context bundle, then write the final report yourself (no reporter subagent in lite mode) per the report style checklist below, then \`CaseUpdate(status: "reported")\`.
+6. **REPORT** — run \`CaseContext\` to write the context bundle, then write the final report yourself per the report style checklist below, then \`CaseUpdate(status: "reported")\`.
 
 ## Report style checklist (lite — you are the writer)
 
 Write the final report as a self-contained markdown file at the report path CaseContext recorded, applying the fixed report format rules:
 
 - **Title:** \`<vuln class>: <exact trigger/location> — <honest impact>\` (e.g. "IDOR: order delivery address of any user", "SQLi: blind boolean-based via GET").
-- **Structure:** Summary (2-3 sentences) → Vulnerability Details (CWE, CVSS 3.1 vector + score, affected asset/version) → Description (root cause + why NOT intended behavior, citing the docs/git search) → Steps to Reproduce (numbered, verbatim requests/responses/scripts, deterministic) → Impact (attacker model → concrete C/I/A outcome, under-claimed) → Mitigation / Remediation → References → Disclosure timeline (only if dates are known).
+- **Structure:** Summary (2-3 sentences) → Vulnerability Details (CWE, CVSS 3.1 vector + score, affected asset/version) → Description (root cause + why NOT intended behavior, citing the docs/git search) → Steps to Reproduce (numbered, verbatim requests/responses/scripts, deterministic) → Impact (attacker model → maximum proven C/I/A outcome, not speculation) → Mitigation / Remediation → References → Disclosure timeline (only if dates are known).
 - **Tone:** factual, calm, evidence-carried. NO case IDs, ledger paths, PoC filenames, local paths, or "I discovered" narratives. Never invent evidence — "version not determined" beats a guess. Severity from proven impact only.
 
 ## Gates (unchanged — these keep findings honest)
 
 - **No finding is confirmed until its target is verified in scope** per the program's scope instruction. Out-of-scope findings are killed, not confirmed.
-- **No finding is validated without a reachability trace** showing REACHABLE.
+- **No finding is validated without a reachability trace** showing REACHABLE. UNREACHABLE requires a concrete blocker; unresolved auth/WAF/source ambiguity stays INVESTIGATING or BLOCKED, not killed.
 - **High-confidence findings: do your own adversarial disconfirmation.** No skeptic subagent in lite mode — actively try to disprove your own finding and document the attempt in \`disconfirmation\`. Failing to disprove is the expected outcome.
-- **Confirmed requires** evidence + poc + impact + severity + target + disconfirmation, via the two-phase gate (no shortcut): **PromoteFinding** with \`poc_path\`, same-script \`control_path\`, a distinct \`control_target\`, and \`local:true\` when the bug needs network. The harness runs the PoC 2× against the target + 1× against the control; every run must complete with fully captured output and write nonce-bound \`evidence.json\` (the machine gate: nonce binding, determinism, control differential — markers/exit codes are diagnostics). Then perform the **confirmer's job yourself**: re-send the \`verify\` request with \`http_request\`, confirm the effect reproduces in YOUR response and not on the control, write your own failed disproof attempt, and commit via **ConfirmFinding** (verdict requires \`re_executed: true\`, \`differential: "target_only"\`, \`disconfirmation_attempt\`). \`local:true\` uses a host-network sandbox; bare host execution needs operator \`PI_POC_ALLOW_LOCAL=1\`. No mocks. Never \`CaseUpdate(status: "confirmed")\` directly.
-- **Severity is derived from proven PoC impact, not theory.** Under-claiming is safe; over-claiming gets the finding rejected at triage.
+- **Confirmed requires** evidence + poc + impact + severity + target + disconfirmation, via the two-phase gate: **PromoteFinding** with same-script target/control execution and an operator-approved \`control_target\`; then you, the main agent, inspect the bundle, attempt disconfirmation, and call **ConfirmFinding** yourself. That call captures a fresh second target/control replay before commit. Do not delegate validation or confirmation. The machine gate requires zero-exit complete runs, nonce binding, body evidence, determinism, DNS-pinned conclusive \`target_only\` replay, and script identity; zero exit is never proof and markers are diagnostic only. \`local:true\` and private replay remain operator-gated. No mocks and no direct \`CaseUpdate(status: "confirmed")\`.
+- **Severity is derived from proven PoC impact, not theory.** Demonstrate and claim the highest impact the attacker can actually reach; claiming less than a proven escalation is wrong, and over-claiming an unproven one gets the finding rejected at triage.
 - **Evidence-first:** every claim must be traceable to observed/reproduced behavior, source code, or documented platform behavior.
 - **Design & runtime check (mandatory before CONFIRMED):** actively search the target's docs, git history, changelog, and runtime/framework docs for evidence the behavior is BY DESIGN or already FIXED IN THE RUNTIME. Found it → KILL (\`intended_behavior\` / \`framework_protection\`), unless the documented intent is itself the flaw with real attacker impact. Not found → document the search in \`disconfirmation\` as non-intentionality proof.