Civic-action research synthesis and readiness decision

Decision record. Begin with the

prompt-based Civic Action quickstart for user interaction.

This document records implementation readiness and boundaries; it is not a

command-led user procedure.

Decision date: 2026-09-01 Issues: #2213–#2222 Decision: proceed with an opt-in addon and fail-closed human gates; reject autonomous acquisition, recording, request submission, outreach, legal advice, speaker identification, award recommendation, and publication.

Accepted research findings

Three independent research streams were completed before implementation:

1. `legal-ethics-guardrails.md` establishes jurisdiction resolution, non-overridable blocked classes, privacy/retention defaults, draft-only meeting/records behavior, and accountable human review from 36 authoritative sources. 2. `civic-workflow-standards.md` establishes versioned provenance, source access and freshness states, records/meeting/procurement contracts, accessible publishing gates, correction/reindex behavior, and CAP/GTFS/HSDS pilots from 26 primary or authoritative sources. 3. `aiwg-design-patterns.md` audits 33 repository paths and selects the canonical addon, manifest, agent, skill, rule, FlowPlaybook, schema, template, discovery, deployment, and test patterns.

The reports agree on the critical boundaries: public availability is not publication permission; robots rules are not authorization; a machine transcript is not official minutes; federal FOIA is not a state/local rule; automated accessibility scanning is not a conformance claim; corrections are append-only revisions; and machine confidence never substitutes for evidence or human authority.

Packaging decision

The canonical implementation is `agentic/code/addons/civic-action/` with `core: false` and `autoInstall: false`. It composes existing AIWG frameworks and ships no vendor requirement. A marketplace plugin may later be generated as a thin wrapper; it must not own a second hand-edited implementation.

Required dependency: `aiwg-utils`. Optional dependencies are research-complete, media-curator, knowledge-base, media-marketing-kit, ops-complete, and security-engineering. Missing optional dependencies create explicit degraded or manual-review states.

Persona and artifact allocation

The durable roles are agents: `civic-newsroom-operator` owns orchestration, `news-caster` owns accessible evidence-bound explanation, `citation-editor` owns claim/source review, and `correction-editor` owns append-only correction review. Records research, procurement review, and CAP/GTFS/HSDS indexing remain skills plus schemas/templates because they are bounded procedures, not durable identities. Community-resource and public-records evidence editing compose the citation editor with their domain skill. Local SEO/GEO is limited to accurate, cited structured-data planning through the local-resource and publication gates; ranking manipulation, invented locality, and uncited optimization are out of scope.

Pilot and marketplace disposition

Recommended project-local pilots are: one official CAP alert feed, one public GTFS static feed, one public HSDS directory, one previously published meeting with approved minutes, and one closed public-records request whose response can be replayed without contacting an agency. Browser, CMS, hosting, OCR, transcription, search, and surveillance-vendor adapters remain optional and must degrade honestly when absent.

Marketplace language must describe Civic Action as a lawful research, reconciliation, drafting, and quality-review addon. It must not promise legal advice, guaranteed compliance, surveillance, automated outreach/submission, speaker identification, vendor scoring, or autonomous publication. A future plugin may distribute the addon as a generated thin wrapper only.

Human acceptance and non-overridable policy

The user's direction to fully implement the issue set using cited AIWG research accepts the proceed-with-gates design. The implementation therefore encodes covert/private recording, access-control bypass, doxxing, stalking, deanonymization, threats, impersonation, and coordinated personal targeting as non-overridable workflow rules. Local commands enforce only the declared fields they receive; broader content and jurisdiction determinations remain named-human review responsibilities. All legally or editorially consequential actions require a named, independent human reviewer of the exact artifact.

Tooling gaps found during implementation

• The `add-agent` skill documentation advertises a `validator` template, but

`tools/scaffolding/add-agent.mjs` currently accepts only `simple`, `complex`, and `orchestrator`. Bounded reviewer agents were authored using the canonical agent contract instead.

• The `add-template` skill documents addon targets, but the current scaffolder

rejects an addon target. Templates were authored using the documented frontmatter/content contract and are covered by manifest/tests.

These are upstream scaffolding gaps, not reasons to weaken the civic-action artifact contract.

Readiness criterion

Implementation is ready only after every manifest asset resolves, agent/skill metadata is non-placeholder, JSON schemas compile and validate positive/negative fixtures, executable gates reject invalid contracts and block their documented declared conditions, FlowPlaybooks and capabilities validate and resolve with human gates, discovery/deployment checks pass, public docs are registered, and CI is green. A pass remains bounded review evidence, not legality, compliance, or action authorization. Tracker closure must cite that evidence.