Skip to main content
Sigilix runs five specialists per review: four domain specialists in parallel — Metis, Argus, Iris, Eunomia — unified by Harmonia, the synthesizer that runs after them. Each domain specialist has a focused prompt, a model tuned to its role, and a cross-provider fallback that makes a same-family provider outage unlikely to silence a specialist. A specialist can still be skipped if both primary and fallback fail — see the retry + fallback policy below. Internally each specialist has a role name (logic, security, performance, tests) that appears in sigilix.json under rules.<role>. The public-facing brand names (Metis, Argus, Iris, Eunomia) are the same specialists with the same prompts.
BrandRoleDomain
MetislogicArchitecture, layering, dead code
ArgussecurityAuth, input validation, secret leakage
IrisperformanceN+1, hot paths, big-O regressions
EunomiatestsNaming, semantics, test coverage
Harmonia is covered on its own page — see Synthesizer. This page is the four domain specialists.

​
Metis · Architecture

What it catches:
  • Circular dependencies between modules
  • Layer-boundary violations (e.g., domain code importing from infrastructure)
  • Interface drift between code and its tests/types
  • Module cohesion issues
Model selection: Metis runs a reasoning-heavy model — its proof-style chain-of-thought reasoning is well-suited for architectural arguments: tracing imports across files, evaluating layered architecture rules, and judging cohesion. When its primary provider is unavailable, Sigilix falls back to a model on a different provider with an uncorrelated outage profile, so a single provider incident can’t silence the logic role. Sample finding: 
[Metis] Boundary violation detected
domain/invoice.ts imports infra/stripe.ts
Rule: Domain may not depend on Infrastructure.
Suggested fix: Introduce `PaymentGateway` port in domain/.

​
Argus · Security

What it catches:
  • Unsanitized inputs (SQL injection, XSS, SSRF, formula injection)
  • Secret leakage in responses, logs, or version control
  • Authentication and authorization bypasses
  • Insecure regex patterns prone to ReDoS
  • OWASP Top-10 patterns
Model selection: Argus runs a faster, higher-volume model than Metis because security findings are higher-volume per PR and need turnaround. The cross-provider fallback ensures that if one provider is down, Argus still produces a verdict — security checks should never silently skip. Sample finding: 
[Argus] Critical: Potential SSRF
utils/fetcher.ts:71 — user-supplied URL passed to fetch()
without an allowlist. Validate against approvedHosts[].

​
Iris · Performance

What it catches:
  • N+1 query patterns in ORM-heavy code
  • Hidden quadratic loops (especially newly inlined ones)
  • Memory leaks (event listeners not removed, growing caches)
  • Unbounded recursion or iteration
  • Big-O regressions vs. the previous implementation
Model selection: Iris runs a model tuned for cost-effective, high-throughput pattern-spotting, with a cross-provider fallback. The exact model choice is tuned over time from shadow telemetry — Sigilix keeps the primary that produces the highest-quality performance findings at the lowest latency, without changing this page each time. Sample finding: 
[Iris] O(n²) render loop
components/Table.tsx:112 — sorting inside render()
Memoize with useMemo keyed by sortKey.

​
Eunomia · Semantics

What it catches:
  • Dead code reachable only by impossible conditions
  • Naming that doesn’t describe behavior (e.g., handleStuff when the function sends an email)
  • Logic errors that pass type-checking (off-by-one, inverted conditions, swapped arguments)
  • Missing test coverage on non-trivial branches added in the diff
  • Inconsistent error handling between two functions added in the same PR
Model selection: Eunomia’s surface is broad and shallow — it needs speed and breadth more than depth, so it runs a fast, high-coverage model with a cross-provider fallback. It often shares a model family with Argus, which lets Sigilix amortize warm-cache costs across two specialists. Sample finding: 
[Eunomia] Unreachable branch
checkout.ts:45 — early return bypasses tax calc
when total < 0. Remove or handle as error path.

​
Retry + fallback policy

Each specialist runs against its primary model with one retry on transient failure (503, 429, timeout). If retries are exhausted, the cross-provider fallback fires once. If the fallback also fails, that specialist’s contribution is skipped — Harmonia synthesizes from the remaining specialists and the verdict is posted with a _3 of 4 specialists succeeded_ footnote. The cross-provider design — primary and fallback on different infrastructure — is deliberate. Same-family fallbacks fail together when the upstream provider has an outage, defeating the fallback’s purpose.

​
Why these four domain specialists (and not more or fewer)?

The split was chosen empirically:
  • Architecture, security, performance, semantics are the four most-cited categories in code review failure-mode taxonomies.
  • Adding a fifth domain specialist (e.g., a “documentation” specialist) didn’t materially improve catches — Eunomia already covers naming and stale comments.
  • Reducing to three domain specialists (e.g., merging architecture and semantics) left blind spots in cross-module reasoning.
The architecture supports adding domain specialists in the future. If a customer-driven category emerges (e.g., a domain-specific compliance specialist for financial code), it can plug into Harmonia’s synthesis without changing the existing four.

​
Tuning per-repo

You can’t disable individual specialists — keeping the ensemble whole is what makes the synthesis work. What you can do per-repo:
  • Influence what each catches via rules.<role> in sigilix.json (Rules & Guidance).
  • Scope the review via pathFilters so a specialist sees less (Path Filters & Profile).
  • Shift flag-worthiness via profile: "chill" | "assertive".
Harmonia (the synthesizer) is always on. It’s how Sigilix produces a single coherent comment.

Synthesizer

How Harmonia deduplicates, calibrates, and posts the final review.

Confidence Scoring

Proof-tier receipts and the grounding gate that govern what posts.