Autonomous coding safety rail

Reviewer packets before agent confidence theater.

A deterministic pipeline that turns a git diff into scoped specialist reviews, evidence bundles, and a final hard-gate judge. No LLM calls inside the factory. No repo mutation. No commit until approval evidence exists.

Open sample review plan View JSON contract Machine status

spec_factory_tests=pass cases=6
context_factory_tests=pass cases=6
review_factory_tests=pass cases=16
review_factory_judge_tests=pass cases=14
code_factory_tests=pass cases=3

fullsample risk tier

8review packets

14judge cases

Pipeline

Spec Factory

Turns the task into requirements, acceptance criteria, readiness, and approval gates.

Context Factory

Collects bounded repo context and project docs with hashes and source paths.

Review Factory

Routes changed files to specialist reviewer packets based on deterministic risk rules.

Reviewer Lane

Hermes delegates packets to specialists: tests, architecture, security, docs, ops, cost/context.

Final Judge

Aggregates findings. BLOCK beats REQUEST_CHANGES. Approval needs real local evidence.

Risk routing demo

The sample diff intentionally touches auth code, a dependency manifest, and a secret-like config path.

.env
requirements.txt
src/auth.py

Generated reviewer panel

spec_compliance
tests_regression
architecture
security
dependency_supply_chain
cost_context
docs_dx
final_judge

Safety fixes from review

Refuses review output inside the reviewed repo unless explicitly overridden.
Rejects git --base-ref option injection before running git diff.
Adds bounded untracked-file numstat evidence.
Routes .env/private-key/cert paths to security/full-tier review.
Final judge now lets closed approval gates BLOCK even when other evidence is also missing.
Redaction no longer mangles generated command flags when a task merely mentions secret-like paths.

Risk tiers

Tier	Triggered by	Reviewers
`lightweight`	Docs-only changes with no security/dependency/infra signal.	docs_dx → final_judge
`standard`	Normal code changes.	spec_compliance, tests_regression, architecture, docs_dx, final_judge
`full`	Auth/secrets/dependencies/infra/large diffs.	Full specialist panel including security, supply-chain, ops, cost/context.

Run it

REVIEW_OUT="/Users/slothuus/.hermes/rapid42/reviews/$(basename "$REPO")/$(date +%Y%m%d-%H%M%S)"
CONTEXT_OUT="/Users/slothuus/.hermes/rapid42/contexts/$(basename "$REPO")/$(date +%Y%m%d-%H%M%S)"

python3 /Users/slothuus/.hermes/rapid42/scripts/rapid42_review_factory.py   --repo "$REPO"   --task "$TASK"   --spec-contract /path/to/spec_contract.json   --output-dir "$REVIEW_OUT"   --context-output-dir "$CONTEXT_OUT"

python3 /Users/slothuus/.hermes/rapid42/scripts/rapid42_review_judge.py   --findings-dir /path/to/reviewer-findings   --output-dir /path/to/fresh-review-decision   --verification-evidence "<real local test/check output>"   --acceptance-evidence-file /path/to/acceptance_evidence.json   --approval-evidence "<required for closed gates>"   --spec-contract /path/to/spec_contract.json   --review-plan "$REVIEW_OUT/review_plan.json"

Sample artifacts

Loading…