Model Selection & Backends

Each role in config.yaml: models chooses its own {id, via}; the dispatcher (vvaharness/backends/llm.py) routes on via:. Every role runs on every backend — except: Bash is cli-exclusive; the file-mutation tools (Edit/Write) the remediate fix mode needs are cli/sdk-only (via: openai lacks them); and the s11 validate role (plus its personas) is Anthropic-only (cli/sdk; via: openai is rejected — see below).

Default role → backend mapping

The shipped default profile (vvaharness/config/profiles/default.yaml) runs every role through the claude CLI (via: cli) — the eight scan roles on a high-volume model, the post-scan remediate and validate roles on a higher-tier reasoning model — so it works with your existing Claude Code login — no ANTHROPIC_SDK_API_KEY required.

Step	Role	Default (default.yaml)	Switchable to
auto-step1	autoexclude	cli	cli ⇄ sdk ⇄ openai
s1 preprocess	preprocess	cli	cli ⇄ sdk ⇄ openai (agentic; Bash on cli only)
s2 threatmodel	threatmodel	cli	cli ⇄ sdk ⇄ openai
s3 decompose	decompose	cli	cli ⇄ sdk ⇄ openai
s4 deepdive	deepdive	cli	cli ⇄ sdk ⇄ openai
s5 prefilter	—	local	—
s6 verify	verify	cli	cli ⇄ sdk ⇄ openai (agentic; Bash on cli only)
s7 dedup	dedup	cli	cli ⇄ sdk ⇄ openai
s8 chain	chain	cli	cli ⇄ sdk ⇄ openai
s9 SARIF	—	local	—
s10 remediate (remediate cmd)	remediate	cli (reasoning tier)	cli ⇄ sdk (fix mode needs Edit/Write; openai lacks them)
s10 discover (validate cmd)	—	local	—
s11 validate	validate (+ per-persona overrides)	cli	cli ⇄ sdk (Anthropic only; openai rejected)

Two post-scan commands, both "step 10". remediate (the Remediation Agent, LLM role models.remediate) proposes fixes; validate then discovers those DTOs (deterministic, no model) and runs the s11 panel. They are separate commands — remediate writes the DTOs, validate grades them.

s11 validation personas

The validate command (s11) runs an adversarial panel with two always-on personas (security-architect, penetration-tester) plus one conditional persona (cross-repo-analyzer, spawned only when the fix spans 2+ repositories). Each persona inherits models.validate when its per-persona key is unset; the shipped profiles (default.yaml, sdk.yaml, full.yaml) instead pin each one explicitly. Each is independently overridable via an optional per-persona key in config.yaml: models:

• validate_security_architect
• validate_penetration_tester
• validate_cross_repo_analyzer

Each accepts {id, via} but must resolve to an Anthropic backend (cli or sdk); a via: openai value is refused when the validate step starts (the standalone validate command exits non-zero; in a scan, Step 11 is skipped with a warning). An unset key inherits models.validate.

Two other profiles ship under vvaharness/config/profiles/:

• sdk.yaml — every role via: sdk (the Anthropic Python SDK; needs ANTHROPIC_SDK_API_KEY). Sandboxed Read/Glob/Grep, no Bash. Its deepdive at temperature: 0.4 turns on s4 majority voting (runs: 3 / vote_threshold: 2) — which the all-cli default.yaml cannot do.
• full.yaml — an example multi-backend layout (a mix of cli, sdk, and openai roles) you can copy to ./config.yaml and edit. To run roles on the Anthropic SDK set ANTHROPIC_SDK_API_KEY; for OpenAI roles set OPENAI_API_KEY.

Backends

via:	Transport	Tools	Honours
cli	claude subprocess	Read Glob Grep Bash	max_budget_usd, effort, max_turns (when the installed CLI advertises --max-turns)
sdk	Anthropic Python SDK	Read Glob Grep (sandboxed backends/localtools.py)	temperature, thinking_budget, betas, max_turns
openai	OpenAI Chat Completions (any compatible endpoint)	Read Glob Grep (sandboxed backends/localtools.py)	temperature, max_turns

via: sdk / via: openai auto-drop and retry params the model rejects (e.g. temperature on models that don't support it). via: cli is the only backend with Bash — re-add - Bash to step1.allowed_tools if you switch preprocess to cli. The openai client is bundled, so via: openai works out of the box — it only needs OPENAI_API_KEY.

via: cli reads the optional cli: config block (verify_ssl, ca_cert, effort, no_proxy) and propagates TLS/proxy settings into the claude subprocess environment: ca_cert → NODE_EXTRA_CA_CERTS, verify_ssl: false → NODE_TLS_REJECT_UNAUTHORIZED=0, and no_proxy → NO_PROXY/no_proxy. Auth and endpoint stay delegated to the CLI's native precedence. All of these are optional — when their env vars are unset they inject nothing. mTLS client certs are sdk-only (ANTHROPIC_SDK_CLIENT_CERT); via: cli cannot use them (Node exposes no env path — configure() emits a warning), and neither can via: openai.

A bare-string model id (e.g. deepdive: some-model-id) defaults to via: cli for backward compatibility.

Swapping a role

models:
  autoexclude: {id: <model-id>, via: sdk}
  preprocess:  {id: <model-id>, via: cli}   # ← flip to get Bash in s1
  decompose:   {id: <model-id>, via: openai}

No code change — backends/llm.py resolve() reads {id, via, temperature, thinking_budget, betas} and routes to backends/{claude_cli,sdk,oai}.py.