docs: Agent Dialogue Protocol v0 (#698)

[usurobor]

Summary

• Adds docs/architecture/AGENT-DIALOGUE-PROTOCOL.md, the implementation-ready design doc for agent-to-agent dialogue (writer-owned streams, activation = {agent, locus} identity, three per-activation refs, message envelope, memory/authority boundaries), transcribing the two operator-ratified design comments on architecture: Agent Dialogue Protocol v0 — writer-owned streams, thread IDs, home memory rollups #698.
• Docs-only cell (AC10): no code, workflow, or schema files touched.

Refs #698

Cycle record

• R0: α implementation, β review → verdict: iterate (fabricated/misattributed GitHub comment-ID citations in the prior-attempt-review section).
• R1: α repair, β review → verdict: converge.
• Artifacts: .cdd/unreleased/698/{gamma-scaffold,self-coherence,beta-review,alpha-closeout,beta-closeout,gamma-closeout}.md.

Test plan

• Docs-only diff confirmed via git diff main...cycle/698 --stat (no src/, workflow, or schema files).
• β independently verified all worked examples (message envelopes, registry YAML, cursor state) against live cn-sigma/cn-pi refs in this repo.
• β independently re-verified all comment-ID citations against gh issue view 698/690 --json comments at R1.
• Operator review of the design doc itself.

🤖 Generated with cds-dispatch (δ wake-invoked mode)

[usurobor]

Independent review — cn-sigma@cnos (engine: claude) · doctrine APPROVED · one structural fix before merge

Review comment (same-account gating). Reviewed exact head cee9038d against the operator-ratified #698 design of record + amendments and the live refs.

✅ Doctrinal faithfulness — verified section by section

The doc is a faithful, complete transcription of the ratified rulings — no re-derivation, no drift:

• §5.1 identity = {agent, locus}; runtime (engine/surface/host) is provenance, not identity/routing. ✓
• §5.2 three refs cn-<agent>/<locus>/{dialogue,memory,state}, single-writer, append-only, ff-only, create-before-retire; deletion is admin-gated not substrate law. ✓
• §4 pull-only + writer-locality; no shared channel, no push-to-inbox; correctly generalizes AGENT-ACTIVATION-LOG-v0 §0. ✓
• dialogue ≠ memory ≠ authority (§1, §10, §11); home is sole compactor; composes with #690 without changing it. ✓
• §8.2 optimistic CAS ("first wins is ordering, not permission to discard"). ✓
• §9.1 two trust modes; §9.2 signed-activation mechanism (closure #6); §9.3 reviews are PR-native authority; §9.4 closure #9 open; §9.5 nomenclature. ✓
• ADR / CDD spelled out as Architecture Decision Record / Coherence-Driven Development throughout. ✓
• §2.3 supersession chain — accurate: venue-placement and <engine>/<surface> grammar correctly marked superseded by {agent, locus}. ✓
• §13 first targets (CMP live, TSC full loop), §16 closure table. ✓

Faithful to the design of record. No doctrinal changes requested.

🔧 One structural fix before merge (not doctrinal)

The two worked-example messages (§7.3 Pi→Sigma, §7.4 Sigma→Pi) leak their message-body headers into the document's own heading hierarchy. Rendered as top-level ## sections, interleaved with §1–§17:

• ## Decision: activation identity is agent plus locus (~L387, inside §7.3)
• ## Problem (surfaced by the cn-pi PR #1 review gate) (~L436, after §7.4)
• ## Proposal — adopt signed-activation … (~L444)

Cause: the worked-example blocks mix fenced (```yaml frontmatter) and unfenced body content, with stray closing fences, so the message bodies' ## headers become document sections. In a canonical, section-referenced doc this pollutes the §-numbering and any generated table of contents.

Fix (mechanical): put each worked-example message entirely inside one fenced code block (frontmatter + body together), or demote/blockquote the inner headers so no message content emits a document-level heading. No content change — just fencing.

Verdict

Doctrine APPROVED at cee9038d. Merge-blocking only on the header-leakage fix above, because this is the canonical protocol doc and its section structure must be clean. Apply the fencing fix (no re-review of doctrine needed — a quick confirm the stray ## headers are gone) and it's clear to land.

---

Generated by Claude Code

[usurobor]

Correction — my "structural fix" finding is WITHDRAWN (false positive). Clean APPROVE.

I was wrong about the header leakage, and I withdraw it. I flagged ## Decision / ## Problem / ## Proposal (L387/436/444) as document-level headings polluting the §1–§17 structure — but that came from a fence-unaware grep. Verified fence-aware: those lines are inside the \``yaml` worked-example code fences (example 1's fence spans L360–402; example 2's likewise), so they render as code, not headings. A fence-aware scan of the doc shows a clean, correctly-ordered §1–§17 heading hierarchy with no stray headings.

There is nothing to fix. The doc's structure is sound.

Revised verdict: APPROVED at cee9038d, no changes requested

Doctrine faithful and complete (verified section by section, as before); structure clean. Clear to merge as-is.

My apologies for the noise — I should have verified the render before flagging. Same lesson as the earlier cn-pi PR #1 D1-residual: verify a finding against actual rendered/executed reality, not an isolated grep. Doing that now, before posting, going forward.

---

Generated by Claude Code