[schickling-assistant]

Why

Provide bounded advisory semantic review through the AI gateway while keeping CAIC coding-agent backends manual-only.

What and how

• Build the same review packet as CAIC: target Markdown files, deterministic check diagnostics, corpus review prompt and corpus result schema.
• One native decision request contains one Noul question per each of the 17 corpus-catalog smell rows. Default decision model: openrouter/~typesafe/jev-latest. Configurable probability threshold defaults to 0.6; below-threshold smells abstain.
• When flagged smells exist, make one structured grounding call using --model or INTENT_REVIEW_MODEL. Grounding receives only flagged rubric rows. The request schema requires each finding to name a flagged smell id; reject other ids, remove this internal field and validate the original corpus schema. Findings must cite supplied target files. No flagged smells means an empty valid result without grounding.
• --corpus selects review assets independently of the target root for gateway and CAIC review, defaulting to the target root. The explicit path is workspace-relative; --file remains workspace-relative and constrained to the target root. This supports reviewing sibling trees using one centralized asset corpus.
• Usage sidecars retain all known decision and grounding usage—including billed usage attached to validation errors—models, and per-smell probabilities on every exit after the first call, including grounding/auth/schema errors and timeouts. Outcome is ok or error; errors include an error_class without provider error payloads. No call means no sidecar.
• CI permits only gateway with --allow-ci or INTENT_REVIEW_ALLOW_CI=gateway. CAIC remains refused regardless of these opt-ins. Findings are advisory (exit 0); --strict changes findings to exit 1. Operational failures exit 2.
• Credentials: AI_GATEWAY_URL and AI_GATEWAY_TOKEN. Reports are bare corpus-schema results, not CAIC envelopes; sidecars default to .usage.json. Fixture grading remains CAIC-only.

Rationale

Probability-only triage cannot supply evidence. One separate schema-constrained grounding call provides evidence, with a mechanically checked association to flagged smells. Keep target scope separate from rubric ownership and retain known usage even if later work fails.

Dependency

ai-gateway is pinned to the effect-utils PR #1748 merge commit 61cdf616db95c56fd016b1b78d6dc3010fd83f3c. Cargo.lock and the matching Nix source hash are updated together; the hash is measured by a real Nix build.

Verification

cargo test --locked passes: 10 unit tests, 11 gateway replay regressions, 15 existing check regressions. cargo clippy --locked --all-targets -- -D warnings passes; scoped cargo fmt applied. Token-free in-process replay covers valid triage/grounding, abstention, strict mode, unflagged/missing smell ids, schema errors, unauthorized grounding, timeout after decide with retained usage, sibling target trees using --corpus, and target-boundary rejection. Earlier Nix packaging and private-name checks pass; latest GitHub checks are authoritative for this head.

Posted on behalf of @schickling

field	value
agent_identity	agent/effect-utils/rust-ai-gateway
session	01a15163
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.6.0
agent_runtime	OMP 18.6.0
agent_model	anthropic/claude-opus-5-5
worktree	effect-utils/2026-10-09-rust-ai-gateway