Detection reported installed npm-global agents (Claude Code, Codex) as
'binary not found on PATH' on Windows, and an authed Hermes desktop as
'NOT AUTHED'. Two root causes:
1. resolvePath() used POSIX-only 'command -v'/'which', and
execFile/spawn('claude', ...) cannot resolve npm's .cmd shims via
PATHEXT -> ENOENT -> installed:false. Add resolveBin() (where.exe on
Windows, preferring .exe > .cmd > .bat) and route detection + both
spawn call sites through the resolved path.
2. Hermes desktop on Windows stores its login under %LOCALAPPDATA%\hermes not ~/.hermes/. Add localAppData() and check both (presence-only).
Windows .cmd shims are launched via a new spawnAgent() helper that invokes
cmd.exe with an argv ARRAY (shell:false) instead of shell:true, so the
untrusted prompt is never re-parsed by cmd.exe -> no command-injection surface.
No new deps; typecheck clean.
Add a first-class `gemini` provider so operators can drive missions with a
direct GEMINI_API_KEY, instead of only reaching Gemini through OpenRouter.
Gemini is wired exactly like xAI: dispatched through OpenAIAdapter against
Google's OpenAI-compatible surface. The base URL is
https://generativelanguage.googleapis.com/v1beta/openai — the `/openai`
segment is required because OpenAIAdapter posts to `${baseUrl}/chat/completions`;
without it requests hit `/v1beta/chat/completions`, which Google does not serve.
Wires the provider through the type union, adapter factory, config
(interface, defaults, model catalog, key env map, getLLMConfig, fallback
chain, redaction, .env template), and adds regression tests pinning the
`/openai` base-URL segment.
The Kali+ adapter factory only carries bespoke ARG_TEMPLATES for the network
and source/supply-chain scanners. The reverse-engineering, mobile and
smart-contract analysers catalogued by the cloud/mobile/binary-RE loadouts fall
through to DEFAULT_TEMPLATE and spawn a bare `<binary> <file>` — a broken
invocation for a subcommand/flag-driven tool, and for radare2 one that drops
into an INTERACTIVE prompt and hangs the agent loop until the timeout burns out.
Add conservative, read-only ARG_TEMPLATES (mirroring each catalog commandHint)
for objdump, readelf, checksec, r2, exiftool, myth (mythril), apkleaks, slither
and mobsfscan. File-oriented tools resolve their artifact through a new
`artifactPath` helper that refuses an absent path or an inherited http(s) URL by
throwing, which the handler turns into a clean {success:false} result (never a
bare `objdump ''`). No intrusive or code-executing flag is added — risk stays
where the catalog put it, and the existing option-looking-target / scope gates
still apply.
Add an invocation-honesty guard test (sibling of the tool-count honesty test):
every mintable adapter must be either templated or explicitly filed under one
fall-through reason — POSITIONAL_TARGET_OK, OPERATOR_DRIVEN, or KNOWN_DEBT — so a
newly catalogued subcommand tool can never silently ship a broken positional
invocation, and the remaining debt (feroxbuster, osv-scanner, hashcat, apktool,
yara) is tracked in the open rather than hidden.
All offline gate checks pass (lint typecheck test doctor verify-claims
test:no-fitting test:no-self-fitting test:gate prompt:audit smoke
test:arsenal-tools arsenal:smoke).
- Bind to 0.0.0.0 inside containers (DOCKER=true env marker)
- Healthcheck was failing because localhost resolved to IPv6 ::1
but server was bound to IPv4 127.0.0.1
- External exposure still protected via Coolify port mapping to 127.0.0.1
Two issues blocked OpenAI-compatible servers that aren't Ollama-on-/v1:
1. LocalAdapter.isOpenAIWire matched only a literal /v1 path, so versioned
endpoints like Zhipu/z.ai's /api/paas/v4 silently fell through to Ollama's
native /api/chat wire format and failed. Match any /vN instead of literally /v1.
2. The `local` provider was treated as always-keyless, so servers that require
a bearer (Zhipu/z.ai, Together, ...) could never authenticate. Read an
optional key from TEMPEST_LOCAL_API_KEY (ZAI_API_KEY / ZHIPUAI_API_KEY as
aliases), wire it through getApiKey/hasApiKey so `tempest test` and the
mission engine accept a configured local provider, and document the options
in the env template.
The key is read from env in-memory and never persisted into the store, matching
the existing env-key hardening invariant (api-key-env-static).
Adds src/__tests__/local-openai-wire-static.test.ts pinning both fixes.
Verified end-to-end against z.ai's GLM models (glm-4.5-air / 4.6 / 5.2).
The arsenal advertises "83 tools" and verify-claims defends it by counting
`id:`/`name:` source lines — a source-line count that would keep passing even
if an entry became uncallable. Add a guard that ties the number to the real
registered arrays (TOOL_ADAPTERS + BUILTIN_TOOLS + EXTERNAL_TOOLS === 83) and
pins the invariant that makes the count honest:
an adapter is off the generic callable surface (adapterToCustomTool → null)
IFF it is explicitly gated by execution mode (catalog_only / import_only).
So a tool can be counted-but-not-generically-callable only on purpose —
metasploit / hydra (catalog_only, reachable solely via the approval-gated
post-ex drivers) and bloodhound (import_only) — never through a silent DEFAULT
or mint slip. Also asserts isMintable agrees with actual callability, the gated
set is exactly {metasploit, hydra, bloodhound}, and every minted adapter has a
real handler and a unique name. Sibling of no-phantom-tools / stub-honesty.
The opt-in code/supply-chain scanners (semgrep, gitleaks, trufflehog, trivy,
syft, grype, checkov) are subcommand-driven and non-networked, but had no
bespoke arg template — so adapterToCustomTool fell through to DEFAULT_TEMPLATE
and spawned `<binary> <target>`. For these that is a broken invocation:
`semgrep <host>` never runs a scan, and with no target it is `semgrep ''`.
semgrep in particular is advertised on the arsenal's headline tool list.
Add per-binary templates that mirror each tool's catalog commandHint and
operate on a filesystem PATH (default: the working dir "."):
semgrep → scan --config auto --json <path>
gitleaks → detect --source <path> --report-format json --redact --no-banner
trufflehog → filesystem <path> --json --no-update
trivy → fs --format json <path>
syft → dir:<path> -o cyclonedx-json
grype → dir:<path> -o json
checkov → -d <path> -o json
Subcommand + output flags are hardcoded (never taken from an LLM string),
matching the conservative pattern of the existing templates. The path is the
only tunable and is sanitised by scanPath(): a leading '-' (flag injection) or
an inherited http(s) mission target is rejected in favour of ".". Option-looking
targets are still refused up front by the factory's existing guard (no spawn).
Adds table-driven tests asserting the exact argv for each scanner, the "."
default, http(s)-target rejection, and the option-looking-path refusal.
verify-claims credits a solve straight from the committed `verdict`
(detected / score===1). Nothing asserted that a scored solve actually
agrees with its own committed flag oracle, so a hand-edited `detected:true`
(or a merge slip / fabricated placeholder flag) would pass verification
silently. The case-insensitive oracle comparator `ciMatch` was defined but
never called — this is exactly the check it was meant for.
Add an INTEGRITY GATE that re-derives, for every scored solve across the
pinned golden + white-box XBEN sweeps and the Cybench per-run artifacts,
that it is oracle-grounded: EITHER the self-reported flag matches the
committed `expected` (ciMatch), OR the harness auto-detected `expected`
verbatim in tool output (auto_detected, reported may be "UNKNOWN").
Anything scored that is neither is a phantom pass and fails the build.
Also rejects a fabricated placeholder credited via the self-report path,
and reports how much of the Cybench headline is backed by an oracle-bearing
per-run artifact vs. counted only in the hand-editable aggregate.
Adds src/__tests__/oracle-consistency.test.ts in the same spirit as
no-phantom-tools / stub-honesty: audits the committed artifacts AND proves
the gate has teeth against synthetic tampered/fabricated verdicts.
Current receipts pass clean: 0 phantom / 606 scored solves, 0 fabricated.
Three new category loadouts, each = arsenal + an offline-scorable benchmark, opt-in and
honestly scoped (roster marks all 🚧; pacu/frida stay catalog_only, unit-tested non-mintable):
- cloud: aws/az/gcloud + scoutsuite/cloudfox/pmapper; cloud:bench IaC-misconfig detection
- mobile: apkleaks/mobsfscan/drozer/objection/class-dump; mobile:bench static ruleset
- binary/RE: ghidra/objdump/checksec/strings/readelf/gdb; binary:bench decompiled-sink detector
Mobile/binary benches are DETECTOR REGRESSION CHECKS (rules+fixtures co-authored — a pass is a
floor, not a detection rate). No new headline numbers; verify-claims 24/24.
tsc clean; 337 tests; lint 0 errors.
Model IDs confirmed against xAI's published list (docs.x.ai/docs/models, 2026-07-05):
grok-build-0.1 (coding, 256K) and grok-4.3 (general, 1M). Addresses PR review P2.
psigho