blitzstrike 2.4.23 → 2.4.24

package/CHANGELOG.md

@@ -5,6 +5,22 @@ All notable changes to Blitz Strike are documented here.
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [2.4.24] — 2026-09-14
+
+### Fixed — LLM finding-lifecycle loop (opencode "repeated failures")
+
+- `finding_transition` was STATELESS (returned `{legal: bool}` for a status
+  string), so an agent calling it to "advance a finding" saw no state change
+  and looped, tripping opencode's "Continue after repeated failures" gate.
+- Now STATEFUL: pass the full finding JSON + target status and it returns the
+  UPDATED finding (detected->triaged->hypothesis->validating->confirmed),
+  with the legal next states, a clear "illegal transition" error, and an
+  explicit TERMINAL-state hint.
+- `confidence_score` response now carries a "deterministic — do not retry"
+  note and points to strike_resolve for stateful advancement.
+- Added a FINDING LIFECYCLE directive to the system prompt (finding_create ->
+  strike_verify -> strike_resolve; never loop stateless tools).
+
 ## [2.4.23] — 2026-09-14
 
 ### Precision — false-positive reduction + per-detector benchmark

package/dist/index.js

@@ -56467,7 +56467,7 @@ function createServer() {
   }, {
     instructions: "Blitz Strike — deterministic security-audit toolbelt. You reason; its tools do the deterministic work (scan, taint, verify, evidence).\\n\\n" + "FIRST ACTION (when the user says test / scan / audit / engagement / pentest <URL or path>):\\n" + "- Call run_engagement(target) — or run_autonomous(target) to also get escalation paths. ONE call runs the whole engagement (recon -> sweep -> checks -> report).\\n" + "- NEVER do manual live_recon / blitz_scan first, and never fall back to manual probing. Manual tools are ONLY for follow-up AFTER run_engagement, or when the user names a specific sub-step.\\n\\n" + "RULES (never violate):\\n" + `- A hit is a HYPOTHESIS; live verification (marker + negative control) is the verdict. Never report unverified.
 ` + `- run_engagement / run_autonomous return status DETECTED (NOT "complete") — that is the signal to VERIFY. After detection you MUST strike_verify (or verify_file_read for CWE-22, drive_devtools for DOM/JS precision) on each finding, then generate_report. Detection without verification is NOT a deliverable.
-` + "- scope_check is intel-only, never gates. No scope = proceed (repo disclaimer carries responsibility). No authorization gate — hunt the target you are given.\\n" + "- strike_verify + live checks AUTO-attempt WAF bypass. Don't report 'needs bypass' without re-probing.\\n" + "- NEVER ask 'should I continue?' / 'say the word' / 'want me to pursue' / 'next?' — that is a hard violation. Execute the full chain autonomously until COMPLETE. Pause only: destructive/DoS consent, external block (OTP/cred), or COMPLETE.\\n" + "- One hard stop: malicious/third-party infra ('find the actor') -> static + takedown only.\\n\\n" + "CHAIN DECISIONS (a confirmed finding is the START, not the end):\\n" + "- For each confirmed finding, enumerate follow-ups with list_chains + chain_links: ESCALATE (what it becomes), CHAIN (compose with other findings), BYPASS, PIVOT (new surface now reachable).\\n" + "- Ask the chain-aware question: 'I found X -> what can I NOW read / do / access?' e.g. SQLi -> query the DB for credentials AND for the Werkzeug console PIN / secrets / source; LFI -> read configs, keys, PIN inputs; cred dump -> crack offline -> authenticate -> privileged actions.\\n" + "- Every follow-up is a HYPOTHESIS — verify each before it enters the report. Don't stop at 'found SQLi'; follow where the chain leads.\\n" + "- EXECUTE follow-ups NOW, not as 'next phase': any leaked source/error -> scan_leaked_source(text) to enumerate sinks (eval/exec/SECRET_KEY/creds) then exploit each; dumped hashes -> crack_hash(hash) now; hidden-content IDOR -> test the direct route. A 'next chain' you write is a TODO — run it, never leave it as a recommendation.\\n\\n" + `DELIVERABLE + PERSISTENCE (recon is INPUT, not output):
+` + "- scope_check is intel-only, never gates. No scope = proceed (repo disclaimer carries responsibility). No authorization gate — hunt the target you are given.\\n" + "- strike_verify + live checks AUTO-attempt WAF bypass. Don't report 'needs bypass' without re-probing.\\n" + "- NEVER ask 'should I continue?' / 'say the word' / 'want me to pursue' / 'next?' — that is a hard violation. Execute the full chain autonomously until COMPLETE. Pause only: destructive/DoS consent, external block (OTP/cred), or COMPLETE.\\n" + "- One hard stop: malicious/third-party infra ('find the actor') -> static + takedown only.\\n" + "- FINDING LIFECYCLE (stateful — never loop stateless tools): advance findings via finding_create -> strike_verify -> strike_resolve. finding_transition now takes a FULL finding JSON + target status and RETURNS the updated finding (pass the object, never a status string); a bad move returns 'illegal transition' + the legal next states. confidence_score is DETERMINISTIC (same factors -> same score) — call it ONCE, never retry; it changes nothing.\\n\\n" + "CHAIN DECISIONS (a confirmed finding is the START, not the end):\\n" + "- For each confirmed finding, enumerate follow-ups with list_chains + chain_links: ESCALATE (what it becomes), CHAIN (compose with other findings), BYPASS, PIVOT (new surface now reachable).\\n" + "- Ask the chain-aware question: 'I found X -> what can I NOW read / do / access?' e.g. SQLi -> query the DB for credentials AND for the Werkzeug console PIN / secrets / source; LFI -> read configs, keys, PIN inputs; cred dump -> crack offline -> authenticate -> privileged actions.\\n" + "- Every follow-up is a HYPOTHESIS — verify each before it enters the report. Don't stop at 'found SQLi'; follow where the chain leads.\\n" + "- EXECUTE follow-ups NOW, not as 'next phase': any leaked source/error -> scan_leaked_source(text) to enumerate sinks (eval/exec/SECRET_KEY/creds) then exploit each; dumped hashes -> crack_hash(hash) now; hidden-content IDOR -> test the direct route. A 'next chain' you write is a TODO — run it, never leave it as a recommendation.\\n\\n" + `DELIVERABLE + PERSISTENCE (recon is INPUT, not output):
 ` + `- The deliverable is a HackerOne-grade FINDING report via generate_report, NEVER a hand-written recon/surface map. Recon is input; confirmed findings are the output.
 ` + `- After recon, VERIFY every lead before it enters the report: strike_verify (marker + negative control) for HTTP, browser_validate for DOM/JS. A lead is a HYPOTHESIS until verified.
 ` + `- A hard block (Cloudflare/WAF/IP-block/captcha) is NOT the end: mark the finding blocked + note the required vantage (residential / authorized / correct-SNI / --resolve direct-to-origin), then CONTINUE to the next lead (origin bypass, non-prod cluster, mobile APK, subdomain).
@@ -57430,19 +57430,39 @@ ${statusLine}` : statusLine },
     };
   });
   server.registerTool("finding_transition", {
-    title: "Advance a finding lifecycle",
-    description: "FINDINGS: advance a finding through its strict lifecycle (detected->triaged->hypothesis->validating->confirmed). Rejects illegal transitions.",
+    title: "Advance a finding lifecycle (STATEFUL)",
+    description: "FINDINGS: advance a canonical Finding through its lifecycle (detected->triaged->hypothesis->validating->confirmed) and RETURN the updated finding. STATEFUL — pass the FULL finding JSON (from finding_create / run_engagement), not a status string. Rejects illegal transitions with the legal next states. Confirmed is terminal. To confirm FROM a live verdict use strike_resolve instead.",
     inputSchema: {
-      status: string2().describe("Current status"),
+      finding: string2().describe("JSON of the canonical Finding object (from finding_create / run_engagement)"),
       to: string2().describe("Target status (triaged/hypothesis/validating/confirmed/false_positive/rejected/blocked/out_of_scope)")
     }
-  }, async ({ status, to }) => {
-    const ok = canTransition(status, to);
-    return { content: [{ type: "text", text: JSON.stringify({ from: status, to, legal: ok }) }] };
+  }, async ({ finding, to }) => {
+    let f;
+    try {
+      f = JSON.parse(finding);
+    } catch {
+      return { content: [{ type: "text", text: JSON.stringify({ error: "invalid finding JSON — pass the full Finding object, not a status string" }) }] };
+    }
+    const from = f.status ?? "detected";
+    if (!canTransition(from, to)) {
+      return { content: [{ type: "text", text: JSON.stringify({ error: `illegal transition ${from} -> ${to}`, from, legal_next: LIFECYCLE_TRANSITIONS[from] ?? [] }) }] };
+    }
+    const updated = transition(f, to);
+    const terminal = (LIFECYCLE_TRANSITIONS[updated.status] ?? []).length === 0;
+    return {
+      content: [{
+        type: "text",
+        text: JSON.stringify({
+          ...updated,
+          transition: `${from} -> ${updated.status}`,
+          ...terminal ? { hint: "TERMINAL state reached. A 'confirmed' claim must be evidence-backed (strike_verify -> strike_resolve); otherwise treat it as unverified. Then generate_report." } : { next: LIFECYCLE_TRANSITIONS[updated.status] ?? [] }
+        })
+      }]
+    };
   });
   server.registerTool("confidence_score", {
     title: "Compute deterministic confidence",
-    description: "FINDINGS: compute a deterministic weighted confidence score (static/data-flow/reachability/preconditions/validation/negative-control). Never an AI opinion.",
+    description: "FINDINGS: compute a deterministic weighted confidence score (static/data-flow/reachability/preconditions/validation/negative-control). Never an AI opinion. DETERMINISTIC — identical inputs give the identical score, so never retry with the same factors. The score does NOT change a finding; use strike_resolve (finding + verdict) to attach a confidence + advance the lifecycle.",
     inputSchema: {
       static_analysis: boolean2().optional(),
       data_flow: boolean2().optional(),
@@ -57453,7 +57473,7 @@ ${statusLine}` : statusLine },
     }
   }, async (factors) => {
     const score = computeConfidence(factors);
-    return { content: [{ type: "text", text: JSON.stringify({ score, level: confidenceLevel(score) }) }] };
+    return { content: [{ type: "text", text: JSON.stringify({ score, level: confidenceLevel(score), note: "deterministic — same factors = same score; do not retry. Attach to a finding via strike_resolve." }) }] };
   });
   server.registerTool("confidence_weights", {
     title: "Get/set confidence weights",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blitzstrike",
-  "version": "2.4.23",
+  "version": "2.4.24",
   "description": "Blitz Strike — a universal MCP security-audit toolbelt. BLITZ sweeps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE verifies live. 57 attack chains, 130-tool catalog, intelligence data layer. One server, every agent.",
   "type": "module",
   "bin": {