Add fail-closed operational decision gates

[Deathcharge]

Summary

• add a zero-dependency operational decision-gate layer over existing consensus results
• support explicit pass choices, minority vetoes, closed vocabularies, required participants, and minimum successful weight
• return deterministic passed, blocked, or indeterminate verdicts with stable reason codes, the applied policy snapshot, and complete consensus evidence
• add optional versioned policy IDs, schema-v1 serialization, and deterministic SHA-256 policy digests
• add end-to-end release-gate and policy-panel examples plus API, integration, security, roadmap, and productization documentation
• close still-valid review feedback from merged PR Productize agent-consensus as a standalone library #4: shared setting validation, response hashing, real metadata serialization coverage, alias normalization, CI action pinning, least-privilege release guidance, and stale evidence wording

Why

The core tally engine was credible but stopped at “what did the group agree on?” Real release, output-safety, authorization, and routing workflows also need a deterministic answer to “may the application act on this evidence?” Keeping policy evaluation separate preserves the small provider-neutral core while giving the package concrete operational use cases.

Blocking evidence takes precedence: a configured veto or agreed non-pass choice is blocked. Missing reviewers, insufficient successful weight, unknown choices, quorum failure, or no consensus is indeterminate. Only a fully satisfied policy is passed.

Read-only inspection of optional Samsarix producer contracts informed the policy-panel recipe: action policy emits allow/deny with policy identity, ethics emits allow/deny/review with decision identity, orchestration needs an approval primitive, and LaunchGuard has release/readiness evidence. The package imports none of them.

Developer impact

New public API:

• DecisionPolicy
• DecisionStatus
• DecisionReason
• DecisionVerdict
• evaluate_decision()

DecisionPolicy.digest covers normalized schema-v1 policy content using canonical UTF-8 JSON and SHA-256. A fixed contract vector protects the encoding from drift. The digest is not a signature.

There are still no runtime dependencies, network calls, persistence, provider SDKs, sibling-repository imports, or protected-action side effects. Hosts remain responsible for authenticated identity, authorization, redaction, policy distribution, and enforcement.

Exact-head validation

Head: 310372a

• python -m ruff format --check . — 18 files formatted
• python -m ruff check . — passed
• python -m mypy agent_consensus — passed, 6 source files
• python -m pytest — 79 passed, 99.62% branch-aware coverage
• python -m build — sdist and wheel built; wheel built from sdist
• python -m twine check <artifacts> — both passed
• clean environment pip install --no-deps <wheel> — passed
• installed schema-v1 policy/digest smoke outside checkout — passed
• installed-wheel pip check — no broken requirements
• all seven offline examples — passed, including release and policy-panel gates
• hosted CI — passed twice across quality, Linux Python 3.10–3.14, Windows Python 3.14, package build/check, isolated wheel install, and all examples

Local artifact snapshot:

• wheel SHA-256: 0EC23FA1C984E7044B03CD23E3ABDBA7A0C2C260F7E4D0AFCF2637AFC6236402
• sdist SHA-256: 0245C34A52F6C51FE847C21332EEDA489BB3AD86A975F2D8C2B6FD4127F9C251

Review and release gates

• The full Codex Security workspace scan is pending its setup confirmation; targeted source/sink, credential-pattern, and placeholder scans found no matches.
• A separate real consumer with a pinned producer version remains the next adoption proof.
• PyPI name ownership, release version/tag, Trusted Publisher setup, and the final MIT/Apache-2.0/MPL-2.0 choice remain owner-controlled.
• No package was published and no production infrastructure was changed.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 320d28eb-0d84-46bc-add6-d26b117f7465

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}