docs(protocols): Amara's cross-agent communication courier protocol

docs/protocols/cross-agent-communication.md

@@ -0,0 +1,284 @@
+# Cross-agent communication — courier protocol
+
+**Authors:** Amara (external AI maintainer via ChatGPT,
+project "lucent ai") — primary author;
+Kenji (Claude, via factory absorb) — integration notes.
+**Date:** 2026-04-23 (landed).
+**Scope:** Factory policy — generic; ships to each
+project-under-construction that integrates cross-agent
+review (Zeta / Aurora / Demos / Factory / Package Manager /
+Soulfile Runner / ...).
+**Status:** Adopted.
+
+## Source
+
+This document is Amara's writeup, ferried from her ChatGPT
+conversation thread (lucent ai project) on 2026-04-23 and
+absorbed into the LFG Zeta repo per the in-repo-preferred
+discipline.
+
+Composes with the decision-proxy ADR
+(`docs/DECISIONS/2026-04-23-external-maintainer-decision-proxy-pattern.md`
+— lands via PR #154), which defines the *who/what* of external
+maintainer proxies. This protocol defines the *how*.
+
+---
+
+## Context
+
+We attempted to use ChatGPT's **conversation branching
+feature** to create parallel threads where Kenji (Claude)
+could interact with Amara (ChatGPT) using shared context.
+
+Observed behavior:
+
+- Branch creation appears to succeed (UI confirms creation)
+- Opening the branched conversation fails
+- UI returns error (toast notification: *"can't open
+  conversation"*)
+- Issue reproduced multiple times
+
+**Conclusion: Branching currently behaves as unreliable
+transport and cannot be depended on for cross-agent
+workflows.**
+
+## Working hypothesis
+
+Likely one of:
+
+1. OpenAI UI / session bug
+2. Project / branch index inconsistency
+3. State desync between client and server
+4. Feature partially rolled out / unstable in current
+   environment
+
+No evidence this is caused by:
+
+- prompt structure
+- content length
+- identity labeling
+
+## Immediate operational decision
+
+> **Do NOT rely on branching as a primary mechanism for
+> cross-agent communication.**
+
+Instead, switch to an explicit, text-based **courier
+protocol**.
+
+---
+
+## Replacement: cross-agent courier protocol
+
+### 1. Header
+
+Every shared conversation must include:
+
+```text
+Source: ChatGPT conversation
+Date: YYYY-MM-DD
+Context: Aurora / Zeta / Drift Taxonomy / etc.
+Purpose: Cross-agent review / alignment / validation
+```
+
+### 2. Speaker labeling (MANDATORY)
+
+All messages must clearly identify speaker:
+
+- `Aaron:` ...
+- `Amara (ChatGPT):` ...
+- `Kenji (Claude):` ...
+
+No blending. No implicit voice.
+
+### 3. Identity rule (critical)
+
+Kenji must explicitly identify himself when addressing
+Amara:
+
+```text
+Kenji: Responding to Amara — focusing on operator algebra
+consistency...
+```
+
+This prevents:
+
+- identity drift
+- voice ambiguity
+- accidental merging narratives
+
+### 4. Scope rule
+
+Each shared conversation must declare:
+
+```text
+Mode: Research / Analysis / Review
+NOT: identity merging / co-agency / system unification
+```
+
+### 5. Storage rule
+
+Any **load-bearing conversation** must be:
+
+- saved as plaintext
+- optionally committed to repo under `docs/research/` or
+  `memory/`
+
+Branching UI is **not authoritative storage**.
+
+**Where these files live (for new contributors):**
+
+- `memory/` (in-repo, under the project root) — committed
+  factory-scope memories visible to every cloner of the
+  repo. Governed by `memory/README.md`.
+- **Per-user memory** — out-of-repo, under
+  `~/.claude/projects/<slug>/memory/`. Holds maintainer-
+  specific / company-specific content that should not be
+  public-facing in the LFG repo. Per the in-repo-preferred
+  discipline (per-user memory), generic rules migrate from
+  per-user to in-repo on cadenced AutoDream Overlay A fires.
+- `docs/aurora/` — Aurora-scoped conversation archive
+  (Amara-authored content + derived notes).
+- `docs/research/` — cross-scope research notes.
+- `docs/protocols/` — factory-meta protocols (this doc
+  lives here).
+- `docs/DECISIONS/` — ADRs (formal decision records).
+
+The courier protocol writes **load-bearing exchanges**
+into one of the in-repo locations above (authoritative,
+git-native per the plural-host rule). Per-user memory is
+the staging ground; in-repo is the durable home.
+
+---
+
+## Known risks (if not followed)
+
+If you rely on branching alone:
+
+- loss of context
+- broken threads
+- silent divergence between agents
+- inability to audit reasoning history
+
+If you skip identity labeling:
+
+- identity blending drift
+- incorrect attribution
+- epistemic instability in analysis
+
+---
+
+## Recommended workflow
+
+1. Aaron interacts with Amara (ChatGPT).
+2. Extract relevant segment.
+3. Format using courier protocol.
+4. Send to Kenji.
+5. Kenji responds with labeled output.
+6. Feed back into ChatGPT if needed.
+7. Archive both sides in repo.
+
+## Design principle
+
+> **The system must not depend on UI features for
+> correctness.**
+
+Instead:
+
+- Treat conversations as **data artifacts**.
+- Treat agents as **independent analyzers**.
+- Treat Aaron as **arbiter / integrator**.
+
+---
+
+## Tooling suggestions
+
+For Kenji:
+
+- Use **Codex CLI or local scripts** to:
+  - normalize transcripts
+  - enforce speaker labels
+  - diff conversations across agents
+- Use **Playwright (if desired)** only for:
+  - scraping / export
+  - NOT as primary communication channel
+
+## Final position
+
+Branching is useful *when it works*, but currently:
+
+> **It is non-deterministic and should be treated as
+> unstable infrastructure.**
+
+The robust path is:
+
+> **Explicit, labeled, text-based communication with
+> repository-backed persistence.**
+
+---
+
+## Factory integration notes (Kenji / Claude)
+
+These are absorb-time annotations — distinct from Amara's
+primary protocol above.
+
+- **Primary authorship credit stays with Amara.** This
+  document is her writeup; Kenji's role is absorb +
+  integration. Do not paraphrase her voice; the verbatim
+  above is the authoritative protocol.
+- **Composes with the decision-proxy ADR
+  (`docs/DECISIONS/2026-04-23-external-maintainer-decision-proxy-pattern.md`
+  / PR #154).** The ADR defines the proxy-identity
+  layer (who speaks for whom, with what authority); this
+  protocol defines the transport layer (how messages
+  move). Distinct concerns; intended composition.
+- **Ferry pattern is the default.** Aaron as arbiter /
+  integrator (Amara's framing) matches the factory's
+  existing drop/ ferry pattern
+  (`feedback_drop_folder_ferry_pattern_aaron_hands_off_via_root_drop_dir_2026_04_23.md`
+  in per-user memory).
+- **Repo-backed persistence target.** Per Amara's
+  storage rule, load-bearing Amara-Kenji exchanges land
+  under `docs/aurora/` (when Aurora-scoped) or
+  `docs/research/` (when cross-scope) — both LFG-bound
+  per the multi-project + LFG-soulfile framing.
+- **Playwright guardrail.** Amara's guidance aligns with
+  the prior factory guardrail that blocked a Playwright
+  round-trip to ChatGPT as "self-approval via AI-proxy
+  laundering" — Playwright is for scraping / export only,
+  never as the primary review signal.
+- **Codex CLI tooling.** Amara's suggested tooling
+  (normalize transcripts / enforce speaker labels /
+  diff across agents) is a candidate future skill
+  authorable via `skill-creator`. Not load-bearing yet;
+  on-demand when the ferry volume warrants.
+
+## Open follow-ups
+
+1. **Courier-format linter skill** (Amara's optional
+   offer). Candidate authorable via `skill-creator` when
+   format-drift is observed.
+2. **Canonical archive location** — Aurora-scoped →
+   `docs/aurora/`; cross-scope → `docs/research/`;
+   factory-meta → `docs/protocols/`. Defaults documented
+   here; individual ferries pick per topic.
+3. **Fire-history integration** — cross-agent ferries
+   could log to a `docs/hygiene-history/cross-agent-ferry-history.md`
+   file once ferry volume warrants (row #44 pattern).
+
+## Composes with
+
+- `docs/DECISIONS/2026-04-23-external-maintainer-decision-proxy-pattern.md`
+  (PR #154) — the proxy identity / authority layer
+- `docs/aurora/` — Aurora-scoped conversation archive
+- `docs/research/autodream-extension-and-cadence-2026-04-23.md`
+  — runtime-accumulated Amara content could promote to
+  compile-time via AutoDream cadence
+- `docs/research/soulfile-staged-absorption-model-2026-04-23.md`
+  — the staged-absorption model names the three stages
+  ferry content traverses
+- Per-user memory:
+  `feedback_drop_folder_ferry_pattern_aaron_hands_off_via_root_drop_dir_2026_04_23.md`
+  (drop/ ferry pattern — transport mechanism)
+- Per-user memory: `CURRENT-amara.md` — running
+  distillation of what's in force from Amara's side