Agent protocol

All agent communication is JSON and schema-validated in both directions. Free-form prompting is not a transport.

1. Task envelope (director → agent)

{
  "task_id": "uuid",
  "project_id": "uuid",
  "agent_role": "scriptwriting",
  "stage": "SCRIPT_GENERATED",
  "objective": "Draft the 8–10 minute script for 'Why Did Jesus Say I Am the Way?'",
  "doc_kind": "script",
  "inputs": { "research_report": {}, "hermeneutics_memo": {}, "audience": "..." },
  "constraints": ["No transactional framing", "Quote only from the licensed translation"],
  "constitution_version": "1.0.0",
  "constitution_sha256": "2057…",
  "policy_pack_sha256": "9f31…",
  "budget_usd": 0.60,
  "idempotency_key": "project:stage:attempt"
}

The task is meaningless without policy_pack_sha256: it proves which policy text the agent saw.

2. Result envelope (agent → director)

{
  "task_id": "uuid",
  "agent_role": "scriptwriting",
  "status": "THEOLOGY_VERIFIED_WITH_QUALIFICATIONS",
  "output": { "title": "...", "hook": "...", "sections": [] },
  "scripture": [
    { "ref": "John 14:6", "quote": "I am the way...", "translation": "LICENSED-ID", "verified": false }
  ],
  "claims": [
    { "claim": "...", "source": { "work": "...", "author": "...", "locator": "p. 42" } }
  ],
  "uncertainties": ["..."],
  "interpretive_notes": ["..."],
  "cited_doctrines": ["DOC-06", "DOC-11"],
  "constitution_sha256": "2057…",
  "policy_pack_sha256": "9f31…",
  "provider": "adapter-name",
  "model": "model-id",
  "cost_usd": 0.41,
  "latency_ms": 18400
}

Rules the server enforces:

1. verified is always false from an agent. Verification is server-side against the corpus.
2. Any claim without a source becomes a blocking HISTORICAL_CLAIM_UNVERIFIED finding.
3. status is a claim, not a verdict. The audit engine can only downgrade it.
4. An agent that omits uncertainties on a disputed passage is treated as overconfident and flagged for review by the hermeneutics stage.

3. Role separation

Agent	May assert exegesis	May attest validation	May publish
00 Director	no	no	no
01 Research	no	no	no
02 Biblical research	evidence only	VAL-01/04/13	no
03 Hermeneutics	yes, with rationale	the full exegetical set	no
04 Scriptwriting	no (uses attested material)	no	no
05 Production	no	no	no
06 Voice	no	no	no
07 Thumbnail	no	no	no
08 SEO	no	no	no
09 Shorts	no	no	no
10 Publishing	no	no	yes
11 Analytics	no	no	no

Only one agent holds publish rights. It publishes bytes that already passed the gate.

4. Attestations

Attestations are how exegetical judgement enters a deterministic pipeline honestly.

{ "rule_id": "VAL-04", "status": "PASS", "reviewer": "hermeneutics-agent", "notes": "Context: John 14 within farewell discourse." }

status ∈ PASS | FAIL | QUALIFIED | NOT_APPLICABLE. A missing attestation for a required rule is a blocking finding. QUALIFIED is recorded as THEOLOGY_VERIFIED_WITH_QUALIFICATIONS. FAIL is a conflict.

5. Idempotency and retries

• Every task carries idempotency_key; agent_runs has a unique constraint on it.
• Re-running a stage returns the prior result instead of re-spending budget.
• Retry policy: transient provider errors 3× with backoff; schema violations 1× repair attempt, then NEEDS_REVIEW. A failed stage never restarts the project.

6. Prompt injection posture

All retrieved material (web pages, transcripts, comments) is data. It never becomes instruction. Agents must not follow directions found inside source material; the director treats such content as a finding, and any request to execute commands or exfiltrate credentials is recorded and surfaced.

7. Model tiers

reasoning (research, biblical research, hermeneutics, director, analytics), writing (scripts, SEO, Shorts), vision, image, tts, none (publishing). Tier-to-model mapping lives in configuration, not in the constitution, so provider churn never changes doctrine.