Repository navigation
feat(osint): Locator suite + GPS feeds; search now PARSES pages; LLM search director + aggressive people-search; pretext lab; local Ollama - #219
Open
xxmafiaxxx wants to merge 79 commits into
Conversation
…r, NodeZero & PentAGI Architectural Upgrades - Threat Intelligence & CVE Vault: - CISA KEV (1,687 entries) live sync + FIRST EPSS Exploit Prediction scoring - Dedicated CVE Vault UI (docs/cves.html) with search, filter tabs, and interactive detail modals - Technology-to-KEV Recon Correlation engine (src/recon/cve-correlator.ts) and /api/recon/correlate-cves - NodeZero Feature Port: - Rapid Response Targeted CVE Sweeps catalog & runner (src/tools/rapid-response.ts) - Honeytokens & Cyber Deception Tripwires engine (src/tools/tripwires.ts) - 1-Click Finding Re-Verification / Retest loop (POST /api/findings/:id/verify) - Contextual Exposure Scoring meter (GET /api/mission/exposure-score) - SIEM / Discord / Slack alert webhook dispatcher (src/config/webhooks.ts) - PentAGI Techniques: - Sploitus real-time exploit search client (src/tools/sploitus.ts) - Context chain summarizer & sliding window compression (src/llm/csum.ts) - Malformed toolcall & JSON repair engine (src/llm/repair.ts) - Cognitive Reflector & strategic decision pivot critic (src/agent/reflector.ts) - UI & Platform Hardening: - Persistent App Shell (docs/shell.html + embed.js bridge) across all 15 modules - Green glowing API + LLM Ready status design - SOCKS5 outbound proxy inactivity preflight warnings - Resilient mission resume & stall-recovery engine - DFIR Incident Response Suite (docs/dfir.html, containment, remediation playbooks) - Zero secrets committed: strict .env / .gitignore enforcement
…n local API - enforce loopback/same-origin validation on /api/config/env and drop permissive CORS overrides - require positive package identity (pkg.name === 't3mp3st' or T3MP3ST_DEV) before loading CWD .env - migrate legacy Llama 3.3 70B references repo-wide to Qwen 3.8 and configure Venice reasoning parameters - brief scan-approval banner receipts into operator prompts as mission authorization - add Kali WSL2 cross-platform arsenal discovery and Burp Suite integration bridge - align CTF range challenge manifest tools allowlist with Metasploit
…ecure local endpoints
…ence vault indexing
…rovenance and test suite
…ions, and script syntax tests
…brary + Op Admiral/General live, CVE payload & WoltLab catalog, shell tooltips & Codex AutoHunt degrade - docs: 14 pages veniceModel qwen-3-8-27b + venice_parameters.disable_thinking, llmTimeoutFor 240s floor; shell tooltips 18 titles; general.html Codex AutoHunt honest degrade; configs.html clickable data-config-id + openConfigEditor + duplicateConfig - server: execVersionProbe Windows shim quoting, resolveBin/spawnAgent integration, 3 codex spawn sites + readiness probe, mission authorization (receipt threading + /api/mission/status), settings DB + /api/settings + embed.js sync, /api/cves/payloads + woltlab vendor mapping, cve-feed provenance - llm: VeniceAdapter disable_thinking, CodexAdapter resolveBin/spawnAgent + stdout null guards - operators/prompts: MissionAuthorization + buildAuthorizationBlock injected per task + AUTHORIZATION_NOTICE - tools: cve-payloads (16 KEV entries) + cve-woltlab-catalog (53 CVEs) + tests, cve-correlator woltlab/burning-board/wbb/wsc/wcf mapping - types: Credential.notes optional; .gitignore transient probe detritus + memory/ - bench: obsidivm-evolution current/ledger refresh (gen 004-011 local, not in PR) Co-Authored-By: Jarvis <jarvis@t3mp3st.local>
…files, operator-plane + Codex + CVE payloads)
…map, dark-web lanes, sanctions screening; public GPS feeds; locator report rebuild OSINT MODULE (src/tools/osint.ts, 9 agent tools, category 'osint'): - Person Locator: auto-parses email/@handle/phone/URL/domain/name; parallel chain sweeps 67 public platforms (4 probe classifiers, per-site reliability tiers), Gravatar identity + linked accounts, breach/dump exposure on every identifier, phone E.164/NANP routing, username permutation sweeps for name-only subjects, presence scoring. DATA FOUND audit table, 7-section DETAILED REPORT, sources-consulted audit trail, staged progress UI (docs/osint.html). - Geo Intel live map: Leaflet + Esri dark tiles; ipwho.is/ip-api geolocation, OSM Nominatim geocoding (throttled+cached); layers = proxy egress, engagement targets (findings ledger, host-plausibility filtered), DFIR IOCs, OSINT signals; 60s-cached /api/osint/map-feed. Infrastructure geography only — no person/device positioning, stated in-product. - Dark-web lanes (keyless): ransomware.live leak-site monitor (target vs the groups' own victim posts), Ahmia onion search (Tor-circuit first, honest block notes), direct .onion fetch via local Tor SOCKS (9050/9150 auto-probe), breach/dump exposure (LeakCheck public, XposedOrNot, HIBP k-anonymity free; LeakCheck v2/DeHashed/Snusbase operator-key-gated with identity fields surfaced when licensed records return them). - Sanctions screening: Interpol Red Notices + OFAC SDN with automatic Tor-circuit fallback on 403; blocked sources reported honestly. Public records workbench (browser-side, those sites block bots). SSN/fullz scope explicitly refused and documented in-product. PUBLIC GPS FEEDS (src/tools/public-gps.ts + docs/gps.html): - Keyless open-geodata screen: OpenSky ADS-B aircraft, USGS quakes, NOAA alerts, ISS position, OSM POIs. Vehicles/phenomena/places only. ARSENAL/OPERATORS: - 9 osint tools registered (recon/analyst/ghost toolkits); arsenal headline 119 -> 128 (count lock + README + verify-claims together); new routes /api/osint/*; DOC_PAGES +301; nav added across all pages. SUPPORTING (this week's uncommitted verified work riding the shared files): - general/plan JSON salvage (truncated-LLM-JSON recovery), rapid-response catalog +2 probes (Mirth XStream, Tomcat clear-session), CVE payloads +4, test-suite regression fixes, operator sound effects (docs/sfx.js), self-improve bench generations, AGENTS.md session logs. Verified: tsc 0, build 0, vitest 91 files / 968 passed / 0 failed, verify-claims 27/27, lint 0 errors, ui-inline parse 3/3 blocks on new pages; live end-to-end on the running server (sweeps, locate dossiers, map feed, leak monitor, Tor onion fetch, screening) + browser passes with screenshots.
Author
|
Stacked-PR note: this branch builds on #218 (and #163) which are still open — the large diff includes those commits. Merge order: #163 → #218 → #219 (this one shrinks to its own delta once the earlier PRs land). Headline of THIS PR: the OSINT & Person Locator suite, Geo Intel live map, dark-web lanes (leak monitor + Tor onion access + sanctions screening), public GPS feeds, and the arsenal 119→128 honesty move. |
… + public profile deep link Passive-only lane: site:venmo.com engine queries and the public venmo.com/u/<handle> URL surface in the dossier's operator deep-links and the detailed report. No Venmo endpoints are probed, no email->account enumeration on the financial platform — that stays refused.
…es for username sweeps Accuracy: - Corroboration layer: on every FOUND hit, pull the platform's public profile (GitHub/Reddit/chess.com/dev.to/Lichess/HN — egress -> Tor -> direct) and score it against the subject's known name: name-match / name-mismatch / handle-only. 'A handle exists' is now distinguished from 'this is the person'. - Sweep route + tool accept a subject name hint; locators pass it automatically. - GitHub rate-limit (60/hr per IP) false-negatives fixed: HTML-page fallback probe (octolytics marker) with title-based display-name extraction. - Probe path escalation: egress -> HTML fallback -> Tor circuit -> direct (T3MP3ST_OSINT_ALLOW_DIRECT=0 to disable; direct hits are marked with a real-IP-seen warning on the hit). - Presence scoring: corroborated accounts weigh extra, clear name mismatches subtract and render dimmed as 'DIFFERENT PERSON'. Reporting: - Report section 1 now leads with an ASSESSMENT line (corroborated count + strongest signal + mismatch count); section 2 marks every account identity-match / handle-only / mismatch with profile display names. - Social cards + DATA FOUND rows carry identity badges and display names.
…ing deps) + session logs - docs/gps.html: expanded live map (311 lines) — parses clean, rides the public-gps module's feed contract - package.json: + satellite.js (ISS/ground-track computation for the feeds) - AGENTS.md: OSINT suite / geo intel / dark-web / locator accuracy session logs Note: public-gps.ts + its test have in-flight edits from a concurrent session (OpenCellID BBOX tiling refactor, test fixture not yet updated) — left out of this commit per the only-touch-your-task rule; module shipped green in the previous commit.
…) + no-cache HTML headers - Locator ticker advanced never: elapsed was already seconds but the stage condition divided by 1000 again — the run read as frozen at [1/6] for 60s+ (read as broken). Now advances every ~4.5s to the final stage. - /ui static mount sends Cache-Control: no-cache for html/js so operator pages never serve stale cached JS against a newer backend.
… mid-run Button state machine: idle = ▶ RUN FULL LOCATE, in-flight = ⏹ STOP SCAN (danger styling, stays clickable). Clicking mid-run aborts the request via AbortController — partial results discarded with an operator-stop message, the server chain winds down in the background, and the button resets for an immediate re-run. Verified: stop mid-run + clean full re-run after stop.
…ED from results; profile contact enrichment; DATA FOUND scroll
- scoreIdentityMatch v2: a match requires the subject's LAST name in the profile
(exact or substring of a real name fragment). First-name-only overlap
('Raul Glasgow' vs 'Raul Gutierrez') is a name-MISMATCH, never a match;
initial-only ('Linus T.') is ambiguous → mismatch (conservative attribution).
- Mismatches are REMOVED from socialAccounts/DATA FOUND/presence and moved to
dossier.excludedAccounts — rendered as a collapsed '≠ EXCLUDED — different
people' disclosure section, never mixed into results.
- Profile contact enrichment: GitHub (and other profile APIs) now surface public
email/blog/twitter/location on the social card and into the dossier's
emails/locations/identities records.
- Name-only locates parallelized (perm sweeps in chunks of 3): 179s → 9.5-36s.
- DATA FOUND table wrapped in a 460px scroll container with sticky header.
…ase live from the UI
- setDumpKey/getDumpKey with settings-DB persistence (masked, server-side only;
restored at boot) + env fallback. POST /api/osint/dump-keys arms/clears lanes
instantly — no restart.
- BREACH & DUMPS tab: 🔑 ARM DUMP LANES panel (paste keys → SAVE & ARM → lanes
flip armed, stat card updates, key-buy links inline).
- Deep lanes ride the full fallback chain (egress → Tor → direct); any JSON
response is accepted so API-level errors ('invalid key') surface as honest
0-record results instead of a false 'key-required'.
- Line unchanged: no Tor dump-site harvesters — licensed services only.
…ials derived from wave-1 results - locatePerson restructured into two waves: WAVE 1 = identity core (email intel, breach/dump lanes on email+phone+username, phone routing, screening, MX). WAVE 1 harvests dump-record emails/phones/addresses/usernames into the dossier. - WAVE 2 = social footprint, DERIVED from wave-1 results: explicit handle → email local-part → dump-record usernames → Gravatar linked accounts → name permutations only as a last resort for bare names. - Dossier gains phones[] + addresses[]; dump-record emails/phones/addresses flow into the contact core and the report's new section 2 (CONTACT). - Report reordered: 1 Identifiers+assessment · 2 CONTACT · 3 Social · 4 Breach · 5 Identity · 6 Location · 6b Screening · 7 Sources · 7b Public records · 8 Steps. - UI: CONTACT CORE block renders before socials; social cards show profile email/location/X links (GitHub location now flows, e.g. 'Portland, OR'); DATA FOUND rows contact-first.
…emails/phones/addresses sections, unverified filtered - Search extraction lane: Bing HTML SERP mining (egress -> Tor fallback) — titles/snippets/URLs regex-mined for emails, phones, social URLs (extractContacts + parseBingResults, unit-tested). Decoy filters (example.com, image extensions, year-like numbers). - locatePerson wave 1.5: mines Bing per identifier (name/email/phone), merges extracted emails/phones into the contact core, promotes found github/t.me handles to social-sweep candidates (socials truly derive from search data). - DATA FOUND rebuilt into dedicated sections: NAMES FOUND / EMAILS FOUND / PHONES FOUND / ADDRESSES & LOCATIONS — only corroborated data up top; handle-only socials collapsed as UNVERIFIED, excluded different-people collapsed separately, full audit hidden. - Live: Raul Glasgow -> 240 probes, names section shows operator input + corroborated GitHub handle, honest zeros with the lever named.
…ess section cards, demographics - Profile card leads the dossier: photo (gravatar/corroborated avatar), primary name, best location, DOB/age when dump records carry them, and quick-count chips (emails · phones · addresses · corroborated · breaches · creds). - Section cards: CONTACT INFORMATION (phones+emails w/ source), ADDRESSES (dump records + honest public-records pointer), ONLINE PROFILES (corroborated first, unverified/excluded collapsed), BREACH & DUMP EXPOSURE (with identity fields from keyed lanes). - Demographics harvest: dump-record DOB/age (DeHashed age field) into dossier.ages/dobs, shown on the profile card. - RenderDossier TDZ fix (head declaration restored, block appends).
- MinedPage now retains a bounded visible-text snippet (first ~2.2k chars) so the LLM assist reasons over the ACTUAL page body instead of the title + regex summary (that made the assist a no-op). - 2 tests: parseLlmContactJson (fenced JSON + prose tolerance, re- validation kills malformed values) and the no-pages path. NOTE: the panel-glow UI (osint.html), llmChat/llmAssist engine wiring and server bridge landed in fb094fe (parallel session swept the shared worktree); this commit closes the last gap so the assist sees page text.
… the searches
The LLM (operator's configured backbone — local gemma4 when useLocal is
on) now DIRECTS the search instead of only second-guessing extraction:
- llmDirectSearch(): reads subject facts + what the deterministic passes
already found, then the model returns a bounded plan — up to 6 prioritized
queries WITH intents (site:/quotes/operators welcome) + up to 6 public
page URLs. The code executes them in priority order, re-validates every
contact through the same format-validating extractors, merges the finds
into the dossier, then asks the model to RANK the pages it actually
fetched (verdicts can only reference real fetched URLs).
- Guards: isPublicSearchUrl() SSRF filter drops loopback/private/metadata/
non-http URLs from any LLM-proposed fetch; plan + verdict parsing is
tolerant (fences, prose) and capped/priority-sorted.
- Dossier gains searchPlan (steps + ranked pages + what the directed
searches found) rendered as an auditable SEARCH DIRECTOR panel; the run
glows as a SEARCH DIRECTOR module in the rail.
- Phone filter hardened: triple dot-groups ("377.728.2818") are version
strings, not phones (caught live in the first director run).
Live: local gemma4 planned 6 queries for a subject (e.g.
site:cloudflare.com "Katherine May", "@cloudflare.com"), 23 pages were
fetched under that plan, 14 ranked, directed searches surfaced addresses +
a phone; junk-number second pass confirmed clean. 3 new tests (SSRF guard,
plan sanitize/cap/sort, verdict allowlist). osint 40/40, affected 124/124.
Settings → Local Model gains a permanent model picker beside the tag input: auto-fills on load from the LIVE model list (server-side via /api/models — the browser cannot cross CORS to Ollama), preselects the configured tag, saves on pick. The text input stays for custom tags; the old post-Scan select remains and syncs into the picker. Endpoint resolution fixed: the page's settings state can be stale (defaults 127.0.0.1:8080) while the SERVER knows the real Ollama — the picker now tries the page-configured endpoint (unless it is the untouched default) and then the server-configured one; first live list wins and the status line names the endpoint that answered. Static fallback entries are never shown for the local provider. Live-verified in-browser: 7 real models listed from the server-configured Ollama; picking qwen3.6:latest updated input+state and persisted (reverted to gemma4:latest, config unchanged). UI gates 76/76.
…-steered, multi-round The LLM no longer just re-ranks search hits; it now DIRECTS a multi-round public-source people search, and the engine executes and validates everything. - src/tools/osint-aggressive.ts: a VERSIONED 12-method playbook (web search, direct contact-page fetch, username enumeration, breach/dump lanes, infostealer, breach catalog, public people-records, sanctions screening, dark-web victim posts, Wayback recovery, geolocation, associates pivot). "Current methods" = this artifact, not the model's frozen training. - aggressivePeopleSearch(): up to 3 rounds; each round the LLM sees the playbook + current coverage/known data and picks 1-3 methods with parameters; the engine runs them (public sources only), merges the validated finds, and re-plans. dossier gains directorCoverage + directorGaps so exhaustiveness and remaining gaps are auditable. - FABRICATION GUARD (the first live run caught the 4B model inventing queries like "zac.peters@onefiinix.com" and poisoning the dossier with 21 fake emails): any email/phone/URL/site: token in a proposed query that is not already in the known set REFUSES the pick; direct page fetches must reference a host previously surfaced. The LLM directs; it never supplies the facts. Same-method picks capped at 2/round for breadth. - Never idle on a flaky small model: a deterministic fast-first fallback ladder runs the unrun playbook methods when the model returns unusable picks, and a 150s wall-clock budget bounds the director (methods past it are skipped with a reason). - Live: "Jane Doe" -> round 1 web_search (LLM-planned) + username_sweep (206 accounts for jdoe), round 2 budget-skips, gaps honestly report "no verified linkage between subject and any jdoe handle" — emails 5 (clean) vs 21 hallucinated pre-guard. - 4 new tests: guard (known ok / email+phone+url+site refusals), pick parsing (invented method + file:// url dropped, cap-after-validation), handle candidates, playbook contents. osint 43/43, affected 84/84.
…aterial for authorized engagements New panel under the Locator (same footprint): the researcher sets an authorization reference, describes the scenario, picks channel + objective, and the agent drafts three pretext scripts — cover story → opening → discovery questions → value exchange → objection handling → call to action/exit. Channels: email, phone/vishing, SMS/smishing, in-person, chat. Objectives: credential-awareness test, data collection, physical access/badge, policy-compliance test, rapport & open recon. Scope discipline, matching the rest of the arsenal: - every request requires an explicit authorization reference (engagement or ticket id + who approved it); the panel states it up front; - the scenario is the researcher's own words — no dossier data for a named individual is auto-injected into the prompt; - output is conversation material only: the system prompt forbids credential-harvesting content, malware, and technical payloads. Implementation: buildPretextSystemPrompt / buildPretextUserPrompt / parsePretextResponse (tolerant: fenced JSON, prose, empty-entry drop, 3-script cap) in osint-aggressive.ts; POST /api/osint/pretext bridges the configured backbone (local gemma4 when useLocal is on); the panel glows while composing and each script has a copy button. Live: 3 well-formed compliance-test vishing scripts generated for an authorized ENG-2026-014 scenario (48s local); scope gate refuses an unreferenced request. Tests: 2 new — osint 45/45, affected 80/80.
…ropdown The OSINT AI features (pretext lab, search director, extraction assist) now run on the operator's own Ollama box — no cloud backbone — and the model is selectable per request from the live served list. - src/tools/ollama.ts: resolveOllamaEndpoint (env TEMPEST_LOCAL_BASE_URL/ OLLAMA_BASE_URL first, then saved settings when non-default), listOllamaModels and ollamaChat supporting BOTH wire shapes — native /api/chat and OpenAI-compatible /v1/chat/completions. All calls go through fetchBypassingProxy: a LAN model box must never traverse the SOCKS egress dispatcher (global fetch failed against it — "fetch failed"). - Routes: GET /api/osint/local-models (served list + active + endpoint), POST /api/osint/local-model (persists the selection server-side). - /api/osint/pretext now takes `model` and calls ollamaChat; /api/osint/locate hands the same selected model to the director/assist bridge. Timeouts are actionable (504: CPU-only box, pick a smaller model or a GPU box). - Thinking models: think:false by default with a graceful retry for older Ollama builds that reject the field, num_predict trimmed to 1200. - UI: model dropdown in the pretext panel (lazy-loaded when the panel scrolls into view, ⟳ re-query, sizes shown next to each model), live endpoint status line, and an elapsed-seconds ticker while composing. Live: 7 models listed from http://192.168.1.162:11434/api; selection persists; pretext via gemma4:latest returned 3 scripts in 207s on the local box with zero cloud calls. ui-parse + osint + phantom gates 120/120.
… breach pane Adds a first-class leakcheck.io lane (https://docs.leakcheck.io/overview) to the OSINT section, and fixes four real bugs found by probing the live API rather than by reading the old code. What was actually broken - The Pro v2 row shape: the API sends a `source` OBJECT ({name, breach_date, unverified, passwordless, compilation}) but the lane read `rec.sources` — a string the API never sends. Every LeakCheck record came back with no breach attribution at all, which is the single most important field in a breach result. The rewrite parses the real shape and aggregates per-source row counts with the unverified/compilation flags preserved. - Key discovery: the lane only read T3MP3ST_LEAKCHECK_KEY. LeakCheck's own docs and tooling name the variable LEAKCHECK_APIKEY, and operators also carry LEAKCHECKIO. All three are now read (runtime-pasted key still wins). - The public lane died with "TypeError: fetch failed" whenever the armed SOCKS proxy was unreachable, reporting an unknown where a real answer was available. It now rides the egress → Tor → direct fallback chain like the keyed lane, and a new status-preserving variant of that chain lets it tell a 429 rate-limit apart from a genuine zero — a lane that says "clean" because it was throttled is worse than no lane. - Error honesty: "Active plan required" now reads as a plan gate (and points at the still-working public lane) instead of looking like a broken query; 401 and 429 are named rather than reported as "0 records". Also corrected, from live probes: the live API IGNORES `limit=` and returns zero rows for any `offset=`, so the lane deliberately sends neither and caps client-side. The public lane's rate-limit note said "1 query/10s" where the docs say 1/second, and the docs list only email/hash/username while the live endpoint answers phone numbers too. Wiring - New agent tool `osint_leakcheck` (query + type + pro flag), added to recon. - New route POST /api/osint/leakcheck, writing findings to the ledger. - New LEAKCHECK.IO block in the BREACH & DUMPS pane: public sources, per-source Pro v2 attribution, exposure flags, remaining quota, and records with passwords redacted on screen. Plus its OSINT_HELP entry — the help gate fails any section without one. - Honesty locks moved with the tool count: arsenal 145 → 146, osint 15 → 16. Also fixes a broken build shipped in the previous commit: src/server.ts imports ./tools/gps-copilot.js and ./tools/gps-area-news.js, which were never committed, so a fresh clone of this branch did not compile. Those modules (and the llm noThink option they need) are included here, and the tree is now verified to build standalone from a clean checkout. Verified: tsc exit 0 in a clean worktree of the staged tree, new suite 16/16, honesty gates 167/167, and live on :3333 against the real key — public 1394 sources, Pro v2 1394 records across 230 attributed sources with quota reported, phone type and the Enterprise domain gate both honest.
…obing found, and the broken build my own check missed
…ature, not a flake
Every one of these was previously written off as "the parallel session's fault".
Two of them were not: they described behaviour the product did not have, and the
tests were right.
1. T3MP3ST_CONFIG_DIR did not exist (config-directory, 3 failures)
The test described a config-isolation feature that was simply never built. It
is now implemented as a hard boundary, not a preference:
• a relative path is rejected at ConfigManager construction — it would
otherwise resolve against whatever directory a task is running in, i.e. a
hunt target;
• Conf is pinned to the directory via `cwd`, so config.json lands there;
• a pinned directory reads ONLY its own .env — the repo cwd .env,
~/.t3mp3st/.env and ~/.env are all skipped, so a pinned profile can never
silently inherit the operator's real keys;
• with the variable unset, default home loading is unchanged.
2. POST /api/recon/correlate-cves had drifted off its boundary (cve-correlation)
The validated boundary handler `handleCorrelationApi` existed but the route
called CveCorrelator directly — so malformed input reached the matcher, and
the request path did its own feed work. The route is now a thin adapter with
no I/O. The live-feed correlator moved to its own /live route so no capability
is lost; nothing in the UI consumed the old shape.
3. POST /api/mission/start leaked the raw resolver error (mission-status-endpoint)
It returned `err.message` from resolveGeneralLLMConfig straight to the client,
and that message can carry credentials or internal configuration. It now goes
through resolveMissionLaunchConfig, which returns the fixed
LLM_BACKEND_UNCONFIGURED diagnostic (exported as a constant so the route and
the helper cannot drift apart) — still 400, still before any mission mutation.
4. GET /api/mission/status ignored ?missionId= (mission-status-endpoint)
A client polling a finished run got whatever mission was active now, or null
once a newer one took over. It now resolves the requested run through
resolveMissionStatus, and `active` describes the REQUESTED mission rather than
the process.
5. agent:reflection was never emitted (tool-call-boundary)
Both anti-stall paths (duplicate tool call, no-new-findings) existed and
steered the model, but nothing surfaced them to the operator. They now emit
the advisory event via reflectStrategy, with a typed AgentEvents entry. It is
advisory only: mayExecute is hard-typed false, the boundary is copied from
options and never from tool output, and it cannot widen scope, approve a tool,
or relax a receipt/evidence gate.
6. ctf-rsa-static ran `python3` blindly
On Windows that is usually the Microsoft Store app-execution alias, a
non-functional stub. It failed either as "Permission denied" or — when the
alias does resolve — as a TypeError on the `int | str` annotation, which reads
exactly like the solver being broken. Neither failure is about the solver, so
the test now discovers a real Python 3.10+ across the usual names, splits the
provenance assertions (which need no interpreter) from the solver run, and
marks the solver SKIPPED — visibly — where no interpreter exists, following the
repo's existing skipIf idiom. It is never silently passed.
Also updated one brittle static assertion: local-api-hardening-static pinned the
exact literal `resolveGeneralLLMConfig(provider, model, apiKey)`. The route now
reaches the resolver through the sanitizing wrapper, which forwards those exact
fields. The assertion became a regex accepting either shape — the invariant it
actually protects ("the request's provider/model/apiKey reach the resolver") is
unchanged, and both hardcoded-provider/model bans are still asserted.
Verified: tsc exit 0 in a clean worktree of the staged tree, and the FULL suite is
now **117/117 files, 1313 passed, 0 failed** (30 skipped) — reproduced across
three consecutive runs. The two flake suspects (index, oracle-consistency) pass in
isolation and did not recur.
… real product bugs
…o longer arm it
The lane was bolted on as a separate block and its status reporting was wrong in
a way that only showed up when actually run against the operator's real .env.
What was wrong
- **The status route only checked ONE env var for "where did this key come
from".** With aliases now supported, a key supplied by an alias would report
`source: runtime` — sending the operator hunting for a setting they already
had. It now reports `setIn`, the actual variable names that hold a key, and
the status row names them.
- **The service label did not match the name results report under** ("LeakCheck
v2" vs "LeakCheck Pro v2 (keyed)"). The status row and a run result now read
identically.
- **`unlocks` claimed domain search works.** On this plan it returns "Active plan
required" — verified live. The description now says which types are
Enterprise-gated.
- The LEAKCHECK.IO block showed no armed state at all, so there was no way to
tell the lane was live before pressing the button. It now carries a chip fed
by the same dump-status call (🟢 ARMED · key from …, or 🔒 naming the vars),
linked to the ARM DUMP LANES panel, and the panel text explains that a lane
already armed from the environment needs no paste.
THE BUG THE LIVE RUN CAUGHT
While proving the alias path I set only `LEAKCHECKIO` and the Pro lane died with
"Invalid X-API-Key" — because **`LEAKCHECKIO` is not a key. Operators set it to
the API BASE URL** (`https://leakcheck.io/api/v2`); the real key sits in
`LEAKCHECKIO_API_KEY`. My alias list had taken the name literally, so the panel
would have reported the lane ARMED while every single query failed — confidently
wrong, which is worse than a lane that reports itself locked.
- `LEAKCHECKIO` is no longer treated as a key; `LEAKCHECKIO_API_KEY` is.
- `isPlausibleKey()` rejects URL-shaped, too-short and whitespace-bearing values
outright, so a misconfigured variable can never arm a lane that can only
return "Invalid X-API-Key". A runtime-pasted key is unaffected.
- `LEAKCHECKIO` / `LEAKCHECK_PUBLIC_API` are honoured for what they actually
are: base-URL overrides, for a self-hosted or proxied endpoint.
- Two tests pin it, including that a URL in the primary variable falls through
to a real key in the alias.
Verified live on :3333: dump-status reports "🟢 ARMED | LeakCheck Pro v2
(keyed) | env-or-runtime | from: [T3MP3ST_LEAKCHECK_KEY, LEAKCHECKIO_API_KEY]",
and the scan returns public 1394 / Pro 1394 records across 230 attributed
sources with quota 173. Full suite 117/117 files, 1315 passed, 0 failed;
tsc exit 0 in a clean worktree of the staged tree.
…vs-key bug, and the .env line I deleted and restored
… live The stat card reads "Deep Dump Lanes Armed" and computed `armed / lanes.length` over the `lanes` array. That array had three entries and **excluded the OpenCellID cell-site lane**, which was parked in a separate `gps` object. So with both a LeakCheck key and an OpenCellID key present, the panel reported **1/3** while two lanes were in fact armed — the one number the operator uses to answer "what is actually live here" was under-reporting by a lane. - OpenCellID is now a first-class entry in `lanes`, alongside the three dump lanes, and is armed the same way from the same env var. - `gps.opencellid` is still returned for the Settings page, but it is now DERIVED from that same lane entry rather than computed independently, so the two views can never disagree again. - The stat card placeholder was hardcoded `0/3`; it now reads `…/…` until the status call lands, so a stale 0/3 can never be read as a real measurement. Label changed to "Keyed Lanes Armed · dumps + GPS" to say what it counts. - The locked-lane hint pointed at "ARM DUMP LANES below" for every lane, which is wrong for OpenCellID — its key lives in Settings → OSINT. Each lane now names the place that actually arms it. - Three tests pin this so a lane can never again be armed-but-uncounted: every keyed service must appear in the dump-status lane list, the opencellid lane must be in `lanes` (not only in `gps`), and the stat card must not carry a hardcoded fraction. Verified live on :3333 — the card now reads **2/4**, with `🟢 LeakCheck Pro v2 (keyed)` (from T3MP3ST_LEAKCHECK_KEY, LEAKCHECKIO_API_KEY) and `🟢 OpenCellID (GPS towers)` (from T3MP3ST_OPENCELLID_KEY) armed, DeHashed and Snusbase locked; the LeakCheck scan still returns 1394 Pro records across 230 sources. Full suite 117/117 files, 1318 passed, 0 failed; tsc exit 0 in a clean worktree of the staged tree.
…ile two lanes were live
… renamed away Raul: "leakcheck should be a deep dump lane". Correct, and the mistake was mine. In the previous commit I renamed the stat card from "Deep Dump Lanes Armed" to "Keyed Lanes Armed · dumps + GPS" and retitled the block "LEAKCHECK.IO", which made a lane that has always been a deep dump lane read as a separate tool bolted onto the panel. The API was already right — CHECK EXPOSURE runs `leakcheckDeep` alongside DeHashed and Snusbase, returning 1394 records for a test identifier — but the panel told the operator something else. - Stat card is "Deep Dump Lanes Armed" again. - The block is titled "LEAKCHECK PRO V2 · deep dump lane" and now says plainly that CHECK EXPOSURE above already runs this lane automatically and files its records to the Evidence Vault; the box is the same lane opened on its own. - The help entry was still telling operators to set `LEAKCHECKIO` as a key — the same URL-not-a-key error, in a second place. Both the block and the help now list only real key variables and note that LEAKCHECKIO / LEAKCHECK_PUBLIC_API are endpoint URLs that override the base URL. - Two tests pin the terminology (it is part of the contract, and I broke it once) and a third fails if the UI ever names a URL variable as a key. Also removed two load-sensitive timeouts I had been treating as flakes rather than defects: - config-directory parsed the child's ENTIRE stdout as JSON. Under parallel load tsx interleaves notices into stdout, so the parse threw — an intermittent red in a file whose subject (config isolation) was never wrong. It now reads the last non-empty line, which is the actual contract, and the spawn budget went 15s → 90s since it boots a TS module through tsx. - ts-parse-adversarial asserted a repository crawl finished in under 10s. The test's real claim is TERMINATION — a followed symlink loop spins forever and the runner would hang, not fail — so the clock was only ever a proxy, and it measured 11.2s on a loaded box while behaving correctly. Bound is now 60s, and the real assertion (the crawl finds block A) is unchanged. Verified: five consecutive FULL suite runs under --maxWorkers=3, all 117/117 files · 1320 passed · 0 failed · 30 skipped. tsc --noEmit exit 0 in a clean worktree of the staged tree.
… head
prefer-const x3 (osint peopleLaunchLock is never assigned anywhere — the
'lock' is permanently null; android-forensics devices; the director picks
destructure split so the reassigned binding stays let), no-empty x5
(empty catches now carry the reason they are silent), no-require-imports
x2 (statSync via the fs import already in scope), no-useless-escape x2
(the second [A-Za-z.\- ] stays escaped on purpose — unescaping it would
create the reversed range '.- '), no-regex-spaces ({8}), and
no-misleading-character-class in osint-help (the trailing FE0F is
deliberate: the test mirrors the page's own helpKeyFor, so the disable
carries that reason inline).
…ffs, offline always blocked
check-changed-line-coverage ran 'git diff base...HEAD -- src' through
execFileSync with the default 1 MiB buffer; this branch's src diff alone
exceeds it, so the floor step died with ENOBUFS before any coverage math
ran (measured: spawnSync git ENOBUFS). Sized to 64 MiB — the floor now
runs and passes at 53.0% against the PR base.
doctor.mjs: apiGet catches everything and returns { ok:false, status:0 },
so the offline-as-warn path was dead code and any run with no server live
— CI included — always blocked the gate. A connection failure (status 0)
now warns; a server that answered unhealthy still blocks.
…fake failures config-directory spawns a tsx child with a 90s budget but vitest's own default cap is 5s, so the test died mid-child on a loaded box (measured 12.7s) while the child was still working. oracle-consistency sweeps every committed bench artifact — 57.8s measured on this drive — against the same 5s default. Both now carry 120s per-test budgets; assertions unchanged.
…n/PWA, strix pane, harness budgets - OSINT: google app-linkage + AASA/wifi lanes (mxrch/ghunt techniques, nothing vendored), osint-infra email-laundering guard, osint-correlate deep sweep (link-hub harvest, expansion, clustering, honest Keybase outage), new test suites for each lane - favicon/PWA: generated apple-touch-icon + icon-512 + manifest on every page, reproducible make-favicon.mjs, gate test pins the bytes - strix pane; bearer-auth test; verify-claims taught the new modules; counts moved together (arsenal/osint registries) - harness: 30s vitest budgets (5s default was stricter than the product budgets the tests police); docs refresh across all pages Deliberately NOT committed: trinidad/ebctt electoral scrapers and their scraped data dumps (PII — local only), probe scratch (now gitignored).
…try direct, say so, and canary-gate the live lanes Measured live: vitest 4's wrapped worker fetch intermittently hands back RAW gzip with the content-encoding header stripped — HTTP 200, unreadable body — while undici.request on the same socket answers fine in the same process. googleJson treated that as 'answered HTTP 200'; now a 200 the JSON decoder cannot read falls through to the direct leg before being reported, and the note names transport corruption instead of implying an answer. The test file's live lanes gate on a canary that checks this worker's fetch can actually read a compressed JSON body — skip WITH the reason when it cannot, run everywhere healthy (CI included).
xxmafiaxxx
force-pushed
the
feat/osint-geo-darkweb-suite
branch
from
October 20, 2026 16:15
9b93407 to
c409b74
Compare
…t graph, techniques reimplemented from oritera/Cairn (AGPL-3.0) A general origin→goal search over a blackboard: Facts (confirmed findings), Intents (declared-but-unexplored directions, each naming the facts it grew from) and Hints (operator judgment injected mid-search, absorbed on the next reason round — a new hint is a new situation). Three runtime-synthesized task types: bootstrap (try the goal directly; main phase returns only fact+complete, conclude banks one fact), reason (complete XOR intents, forced progress when the frontier is empty, capped non-overlapping proposals) and explore (claims one intent, returns exactly one objective fact — clean negatives included). Re-reasoning is keyed to new situations (facts/hints grew, frontier emptied), never to the previous reason's own output. Coordination is stigmergic: tasks see only the board. No upstream code vendored; prompts are native rewrites of the documented contracts. - src/search/cairn-board.ts: board, contracts, situation checkpoints, scheduler priority (bootstrap → due reason → claimable explores), YAML graph snapshots - src/search/cairn-engine.ts: prompts, JSON extraction, scheduler loop with concurrent explore waves; chat injected (LLMBackbone-backed in the server) - 30 tests: contract tables, checkpoint rules, priorities, fully-scripted runs - /api/search/start|status|hint|list; hints 409 on a finished search - docs/search.html pane + nav item on every page; .cairn-ref/ gitignored Also fixed by live verification: the four cloud LLM adapters (OpenRouter/ Venice/LiteLLM, Anthropic, OpenAI family) raw-fetched THROUGH the armed SOCKS proxy, so a stale TEMPEST_PROXY_URL killed every provider call platform-wide (missions included). Provider API calls are not attack traffic — they now ride fetchBypassingProxy like LocalAdapter always did. And a reason round that validates as invalid no longer moves the situation checkpoint: the event that triggered it stays un-acknowledged and the next round re-reasons, instead of the checkpoint swallowing it and the search stalling.
xxmafiaxxx
force-pushed
the
feat/osint-geo-darkweb-suite
branch
from
October 20, 2026 18:49
c409b74 to
f3d1a7c
Compare
…ed into war room, CTF range, and OBSIDIVM CVE-2026-109618 (OOM longjmp corruption) and CVE-2026-109619 (stale worker-global state), both CVSS 9.2, both published today, both fixed in 8.2.35/8.3.36/8.4.27/8.5.12. Packages are lab-validate by design — these are request-pattern RCEs with no string one-liner: multipart header amplification, empty-file-control floods, chunked-without-length, the .htaccess / frontend PHP_VALUE early-INI consumers, the PHP_SESSION_UPLOAD_PROGRESS consumer, and the 8.5 fatal_error_backtraces amplifier. Trigger strategies come from the php-src GHSA advisories, not invention. - curated catalog (cve-php-oct2026-catalog.ts) rides every vault sync — the pair is searchable in the CVE Vault before KEV carries them - CVE Exploit Packages card wired into the war room (target map), CTF range, and OBSIDIVM panes — renders the live catalog with NEW tags; CRLF/native endings preserved per page - doctor: the live-server preflight probe is documented at ~30s; the 20s budget aborted it and read as a blocker. 45s now.
…gger proven deterministic, RCE chain mapped and honestly bounded Lab (php:8.2.34 mod_php, single worker, 128M/8M): the 109618 OOM trigger is deterministic — one 7.90MB request (1.66M minimal part headers, inside post_max_size) exhausts the full memory_limit during input parsing, 'in Unknown on line 0', 13/13 rounds; one request = one guaranteed per-request exhaustion on any pre-fix host. The 109619 pre-RINIT consumer window is real (startup deprecation lands in error_get_last before app code) but full RCE did not reproduce: the exception report+clear is bailout-protected on 8.2, the rfc1867 session decode cannot resolve a gadget class pre-RINIT (incomplete-class wall), and single-worker same-request heap reuse masks the stale state. Ground truth read from the fix commit (php-src 3584a98bda). Packages updated with the validated trigger math and the bounded claim.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
OSINT & Person Locator Suite + Public GPS Feeds
Closes the run of requests: OSINT tab with agentic tools, complete locator panel, live geo map, dark-web lanes, sanctions screening, public GPS feeds.
1. OSINT module —
src/tools/osint.ts(9 agent tools, categoryosint)Person Locator (
osint_person_locate): auto-parses any subject (email / @handle / phone / profile URL / domain / full name) and runs a parallel chain:Geo Intel live map (
/ui/osint.html🗺️ tab): Leaflet + Esri dark tiles, keyless ipwho.is/ip-api + OSM Nominatim geocoding (throttled + cached), layers = proxy egress / engagement targets (findings ledger, host-plausibility filtered) / DFIR incident IOCs / OSINT signals, LIVE 30s refresh. Infrastructure geography only — city-level IP geo; no person/device positioning, stated in-product.Dark-web lanes (🕸️ tab, keyless): ransomware.live leak-site monitor (target vs the groups' own victim posts — verified live), Ahmia onion search (Tor circuit first; honest block notes when Ahmia's anti-abuse redirect fires), direct .onion fetch through a local Tor SOCKS (9050 daemon / 9150 Tor Browser auto-probe — verified live end-to-end).
Sanctions screening: Interpol Red Notices + OFAC SDN list, automatic Tor-circuit fallback on 403, blocked sources reported honestly with browser URLs. Public-records workbench (TruePeopleSearch/Whitepages/… deep links — browser-side; those sites WAF-block bots).
Scope line (in-product + docs): SSN/fullz identity-dump harvesting is explicitly refused and documented — no legal keyless source exists and the platform doesn't traffic in stolen identity data. Record-level identity data arrives only via the operator's own licensed dump-service keys.
2. Public GPS feeds —
src/tools/public-gps.ts+docs/gps.htmlKeyless open-geodata screen: OpenSky ADS-B aircraft, USGS quakes, NOAA weather alerts, ISS position, OSM POIs around a pin, Nominatim reverse geocode. Vehicles/phenomena/places — no person tracking.
3. Arsenal / operators / API
/api/osint/sites|dump-status|username-sweep|email|phone|breach|permutate|dorks|locate|ip-geo|geocode|map-feed|tor-status|darkweb/leak-check|onion/search|onion/fetch— findings + dump credentials flow to the Evidence Vault ledgers/ui/+ stale-bookmark 301 + nav across all pages4. Supporting verified work riding the shared files
general/plan truncated-JSON salvage (+5-test suite), rapid-response +2 probes (Mirth XStream, Tomcat clear-session) + catalog test, CVE payloads +4, several test-suite regression fixes, operator sound effects (
docs/sfx.js+ wiring tests), self-improve bench generations, AGENTS.md session logs.Verification
tsc --noEmit0 ·npm run build0 · vitest 91 files / 968 passed / 0 failed / 29 skipped (29 POSIX-skips by design) ·verify-claims27/27 · lint 0 errors · ui-inline-scripts parse clean · live end-to-end on the running server (sweeps, dossiers, map feed, leak monitor, Tor onion fetch, screening) + browser passes with screenshots.Update 2 — locator accuracy + passive Venmo lane (8df4de7, ac36146, 9725c5c)
Accuracy (two root causes fixed):
Passive Venmo lane: site:venmo.com search-engine dorks + public profile URL in the dossier deep-links. No Venmo endpoint probing, no email→account enumeration on the financial platform (explicitly out of scope).
GPS UI + deps: expanded docs/gps.html live map, + satellite.js for ground-track feeds; OpenCellID tiled-BBOX refactor from a concurrent session is landing separately (kept out of this push mid-flight).
Verification: vitest 91-92 files / 994+ passed (one concurrent-session public-gps test red mid-refactor, excluded per convention) · tsc 0 · build 0 · live corroboration + report assessment verified in-browser.
Update 3 — 📱 ANDROID FORENSICS: ADB device triage (972f1ac)
New OSINT tab section wrapping the four ADB scripts from DouglasFreshHabian/AndroidForensics (MIT, vendored into
tools/android-forensics/with its LICENSE). Commit972f1ac.Engine —
src/tools/android-forensics.ts(new, 542 lines)The upstream bash is not run as-is; the workflows are reimplemented as allowlisted Node wrappers so the platform controls exactly what can reach a handset:
execFile('adb', [...]). A command string is tokenized, theadbprefix is required, and both the adb subcommand and theadb shellverb must be on the allowlist. Anything else is rejected server-side with the allowed list in the error (verified:adb shell rm -rf /→shell verb 'rm' is not allowlisted).devices,get-state,start-server,shell,bugreport,pull,-sgetprop,pm,dumpsys,settings,content,svc,logcat,bugreport,uptime,ifconfig,ip,netstat,echotruncatedflag, every call audit-logged as[T3MP3ST][ANDROID] adb exec: ….parsePackageList—pm list packages -s -fparseWifiScan— the AirScope.sh "Networks filtered out due" BSSID regex chain; dedup by BSSID keeping the strongest RSSI, sorted descendingparseSecretCodes—android_secret_codedialer entries out ofpm dumpDUMPSYS_SERVICES— the 22 services modeled fromdumpsys.sh(meminfo, wifi, batterystats, location, power, package, account, telecom, notification, netstats, sensorservice, dropbox, clipboard, usb, window, mount, adb, fingerprint, lock_settings, stats, media.audio_flinger, persona)7 agent tools
android_adb_status,android_device_info,android_app_inventory,android_dumpsys,android_wifi_scan,android_secret_codes,android_adb_exec— registered into the Arsenal insrc/index.tsright afterOSINT_TOOLS, so recon/operator agents drive the same workflows the panel exposes.Server routes (
src/server.ts, 8 endpoints)GET /api/android/status·GET /api/android/scripts(script catalog + dumpsys service list + vendor dir) ·GET /api/osint/android/status(alias) ·POST /api/android/adb·/device-info·/packages·/dumpsys·/wifi-scan·/secret-codes(bounded batch, reportsscanned/totalSystemPackages/truncated).UI —
docs/osint.htmlnew 📱 ANDROID tab (between GEO INTEL and PERMUTATIONS)extract.shsurface).android_secret_codeenumeration (5–100).chmod +x … && ./extract.shinvocation for analyst-box use.panePermsblock that the pane insertion introduced.Arsenal count honesty
arsenal-count-honestynow countsANDROID_TOOLStoo; headline moved 131 → 138 on this branch (the number includes the OSINT-tool growth that shipped earlier in this PR). The lock stays pinned to the real registered arrays, not a source-line count.Doctrine
Physical device · USB debugging · explicit owner/operator authorization. No remote exploitation, no lock-screen bypass, no covert collection. Every path returns honest "no adb / no device / not authorized" rather than pretending.
Verification (on the committed tree, not the dirty worktree)
tsc --noEmit0 ·npm run build0arsenal-count-honesty+osint-tools+ui-inline-scripts-parse(osint.html inline scripts compile)getAdbStatus()→ adb 1.0.41 detected, 4/4 vendored scripts present, honest "No device connected" hintGET /api/android/status200 ·GET /api/android/scripts200 ·POST /api/android/adb {command:"adb devices"}→ exit 0, real output ·POST /api/android/adb {command:"adb shell rm -rf /"}→ rejected by the allowlistAddendum — session 2 (tip of branch:
aa96986,8fde325,fb4b692)feat(osint): Hudson Rock infostealer lane + HIBP breach catalogue (
aa96986, both keyless)Two new live breach sources, each verified live from the operator box before implementation:
cavalier.hudsonrock.comfree API)haveibeenpwned.com/api/v3/breaches, optional?Domain=)Doctrine unchanged: keyless/licensed services + public victim-post monitors only — no dump-site harvesters.
src/tools/osint.ts: defensive parsers (snake+camel tolerant, junk-proof), egress→Tor→direct fallback, 24h catalog cache;emailIntel()now fires Hudson Rock in wave 1 (newinfostealerblock + breach row).osint_infostealer_check(real infection → high finding),osint_breach_catalog(domain orall; required param per the registry rule). Wired to recon (all) + analyst.POST /api/osint/infostealer,POST /api/osint/breach-catalog— both write the findings ledger.docs/osint.html: BREACH & DUMPS tab gains 🦠 infection checker + 📚 catalogue browser.fix(ui): agent-status poll starvation + flicker-proof health (
8fde325)Root-caused via browser resource timing, not guessed: the system-status poller awaited
/api/agents/local/status?check=1(FORCED live re-detect, 30–120s on CLI-heavy boxes). Piling up in the browser 6-connection pool starved every fetch incl. the 5s-abort/api/health→signal is aborted without reason→ header flipped OFFLINE mid-scan and never recovered.AbortSignal.timeout(5s), single-flight guard.checkHealth: 10s abort; one failed probe no longer blinks OFFLINE (flips only after 3 consecutive failures, self-recovers).Updated verification
python3stub in ctf-rsa-static; unchanged by this work.adobe.com→ 1 breach / 152,445,165 accounts / 2013-10-04; poller fix heldAPI + LLM Readyat +22s/+65s/+115s mid-mission.Contribution receipt delta: scope
authorized_live, networkauthorized_external(Hudson Rock + HIBP public GETs, no credentials). No secrets/engagement residue in the commits (.sx_*/.sectest-*scratch deliberately left untracked).Rollback:
git revert aa96986 8fde325.Session addendum — search rebuild, ShadowDragon method, LLM director, pretext lab, local Ollama
Everything below landed after the OFFLINE-flicker fix (commits
e6da479→1a02c74), all verified live against the running box.1. Search lane rebuilt — it PARSES the data now (
8912e8c)The complaint ("the search is useless, it gives links to click") was correct on three counts, each proven live before fixing:
ExtractedContactscarried emails/phones/socials only./ck/a?…&u=<base64url>— the first live run fetched 6/6 pages and got zero contacts, because every "page" was Bing's JS redirect stub.Fixes:
decodeBingRedirect()unwraps the real destinations;mineResultPage()fetches the top 8 result pages (bounded, parallel, 8s each) and mines BOTH the cleaned text and the raw HTML (contact data lives inmailto:/tel:attributes and JSON-LD that text-stripping deletes); a new address parser (US streets/PO boxes + city/state/ZIP, junk-filtered, version-string phones rejected). The UI now LISTS the parsed data — per query pages-parsed and mined counts, per source page each mined value with the page as provenance.Live:
"Cloudflare" contact us email address phone→ 6/8 pages parsed →ir@cloudflare.com,+1 800 077 0774,+1 650 319 8930,101 Townsend St(all real). End-to-end locate on a name subject: 4 emails mined from 6 parsed pages.2. ShadowDragon 5-step method applied — Steps 3 and 5 (
7c3dda5,e6da479)Step 3 (correlation) previously had only name-matching, the weakest gate. Added avatar fingerprinting (sha256 of the public profile image — same image across platforms is the strongest single identity link) and bio-token overlap (city/employer/school wording). Step 5 added historical recovery: Wayback CDX snapshot listing → fetch the newest archived copy → mine its body for contacts, because old bios are a classic leak.
Live:
github.com/torvalds→ 8 real snapshots (2013→2016), newest body fetched and text-extracted (that page publishes no contacts — honest 0). A real GitHub avatar fingerprinted and correlated.3. Panel glow + module rail (
8fde325,261a6a4)Every tool module lights up while in use: username sweep, breach & dumps, dark web, google dorks, email, phone, the locator. A
withBusydecorator applied at load means no runner can forget the state — the tab pulses with a green halo, the panel header grows a ◉ IN USE badge, tabs keep a lit "done" state. The locator's internal module rail (SSEosint:module, 24 events on a real run) uses the same glow; the old client-side staged ticker (which guessed stages) is gone.Live-verified in-browser: mid-run BREACH & DUMPS tab measured with
animation: modGlowand an active box-shadow halo.4. LLM search director (
f1c359b) then aggressive people-search director (63c4f0d)The model stopped being a one-shot searcher:
llmDirectSearch()— the LLM plans prioritized queries with intents + page URLs; the engine executes and re-validates; the model then ranks only pages actually fetched.aggressivePeopleSearch()+ a versioned 12-method playbook (web search, contact-page fetch, username enumeration, breach/dump lanes, infostealer, breach catalogue, public people-records, screening, dark-web victim posts, Wayback recovery, geolocation, associates pivot). Up to 3 planning rounds, each seeing the playbook + current coverage."zac.peters@onefiinix.com"and the miner importing 21 fake emails. Any email/phone/URL/site:token outside the known set now REFUSES the pick; direct fetches must use previously-surfaced hosts. The LLM directs — it never supplies the facts. Same-method cap 2/round.directorCoverage+directorGaps, so "we're still missing this person" is a visible, trackable list.Live after the guard: "Jane Doe" → round 1 web_search + username_sweep (206 accounts for
jdoe), gaps honestly report "no verified linkage between subject and any jdoe handle" — emails 5 clean vs 21 hallucinated pre-guard.5. Social-engineering pretext lab (
ed4ea47)New panel under the Locator: set an authorization reference, describe the scenario, pick channel (email/phone/SMS/in-person/chat) and objective, and the agent drafts three pretext scripts — cover story → opening → discovery questions → value exchange → objection handling → call to action. Panel glows while composing; each script has a copy button.
Scope discipline matching the rest of the arsenal: every request requires an explicit authorization reference, the scenario is the researcher's own words (no dossier data auto-injected), and the system prompt forbids credential-harvesting content, malware, and payloads — conversation material only.
Live: an authorized compliance-test vishing scenario produced three well-formed scripts in 48s; the scope gate refuses an unreferenced request.
6. Local Ollama + model dropdowns (
cec6590,1a02c74)src/tools/ollama.tssupports both wire shapes (native/api/chat, OpenAI-compatible/v1/chat/completions). One bug worth recording: globalfetchis routed through the SOCKS egress dispatcher and cannot reach the LAN box ("fetch failed") — all model calls go throughfetchBypassingProxy.think: falsefor thinking models (gemma4/qwen3 otherwise burn the whole output budget in<think>), graceful retry for older Ollama builds.Live: 7 models listed from
http://192.168.1.162:11434/api; pretext viagemma4:latest→ 3 scripts in 207s, zero cloud calls. Model benchmarking on the CPU-only box:gemma4:latest~5–6 tok/s (usable interactively); the 12B coder variant measured 2.7 tok/s and timed out at 500s; the 30–35B GGUFs are 10–20+ min per draft (background only) — the active selection isgemma4:latest.Verification summary
tsc --noEmitclean;npm run buildclean.python3stub in ctf-rsa-static) — none introduced here.authorized_live, networkauthorized_external(Hudson Rock, HIBP, Bing, Wayback, and the operator's own Ollama box — read-only). No secrets or engagement residue in the commits; the untracked.sx_*/.sectest-*scratch stays out.git revert 1a02c74 ed4ea47 63c4f0d f1c359b 7c3dda5 8912e8c.