ai-memory v0.8.0

Track B — A2A In-Host via Lan-Parity Docker Stack Results (2026-05-28)

Track B re-verifies the A2A non-corpus regression against the integrated release/v0.7.0 tip be3347d70 using the lan-parity Docker stack as the federation substrate. The alice (HTTP 127.0.0.1:19180) and bob (HTTP 127.0.0.1:19181) daemons share the same pg-age postgres backend (127.0.0.1:15432, PG16 + AGE 1.6.0 + pgvector 0.8.2), and the federation tests exercised via cargo test --features sal,sal-postgres --release --no-fail-fast probe the peer-to-peer paths against the lan-parity URLs.

Cross-node Track D (192.168.50.100 ↔ 192.168.1.50) remains operator-blocked per #836: subnet routing between the two physical nodes is not in place; the lan-parity in-host substrate is the canonical Track B target for this campaign.

Phase summary

Phase Surface Status Coverage at HEAD
B.1 alice container health GREEN HTTP 19180 healthy; ic_alice agent_id  
B.2 bob container health GREEN HTTP 19181 healthy; ic_bob agent_id  
B.3 pg-age backend health GREEN 127.0.0.1:15432, schema v51, AGE 1.6.0  
B.4 Federation Ed25519 sign/verify GREEN federation::tests::sync_push_accepts_valid_signature + _rejects_invalid_signature_401  
B.5 Federation per-message nonce binding GREEN sync_push_rejects_replayed_nonce_401_x_memory_nonce_replay (#922)  
B.6 Federation peer-enrollment gate GREEN require_peer_enrollment fail-CLOSED on unenrolled X-Peer-Id (#1088)  
B.7 Federation durable nonces (v51) GREEN federation_nonces table persists across daemon restart (#1255 / PR #1296)  
B.8 Multi-agent identity isolation GREEN SAL visibility-gate matrix (post-#1075)  
B.9 Scoped recall (alice ⟂ bob private) GREEN live_scoped_recall_isolates_private_namespace  
B.10 Governance refusal cross-agent GREEN cli_governance_check_action + governance::* lib  
B.11 4-domain namespace isolation GREEN namespace_standards::* lib  
B.12 Contradiction-link cross-agent symmetry GREEN live_kg_invalidate_dispatches_to_cypher_under_age + kg::tests::contradiction_link_symmetric_sqlite  
B.13 Signature-chain integrity (15-row, 3-peer) GREEN signed_events::* + signed_events_dlq::*  
B.14 Cross-node Track D OPERATOR-BLOCKED subnet routing not in place  

Verdict at a glance: SHIP-CLEARED — every in-host A2A scenario GREEN at tip be3347d70 against the lan-parity stack. The cross-node Track D portion remains explicitly operator-gated and is not a code defect.

LLM substrate

Operator directive 2026-05-31: IronClaw A2A docker-compose ai-memory services (ic-parity-alice + ic-parity-bob) run with OpenRouter google/gemma-4-26b-a4b-it as the AI NHI brain via the provider-agnostic LLM substrate (AI_MEMORY_LLM_BACKEND=openrouter). This supersedes the 2026-05-21 #1067 xAI Grok 4.3 wiring captured in the 2026-05-28 ship-campaign run; the prior wiring remains a documented snapshot of the substrate’s xAI capability for the heterogeneous-AI-NHI cross-LLM verdict matrix.

Canonical pinning + connectivity-probe evidence: docs/v0.7.0/test-config.md §”Substrate LLM (current standing default, 2026-05-31)”.

Per-call substrate cost on Gemma 4 26B (~$0.0000137 per round-trip at ~22 + 26 token shape) lands the daily operating cost at ~$0.014 on the autonomous-tier hot path — a 3000× reduction from the xAI Grok 4.3 daily burn the same workload incurred. Re-running Track B’s federation phases on the OpenRouter-wired alice/bob does NOT change any of the GREEN signatures documented below — the federation tests exercise the wire substrate (Ed25519, nonce-binding, peer-enrollment), not the LLM brain — so the verdict above remains SHIP-CLEARED at the OpenRouter-substrate posture as well.

Methodology — why this is the in-host A2A portion

The 2026-05-22 release-gate dossier captured the 8 A2A non-corpus scenarios in detail (docs/v0.7.0/test-campaign-2026-05-22-release-gate-final/track-b-a2a-results.md). The substrate contracts those scenarios exercise (federation Ed25519 sign/verify, per-message nonce, peer-enrollment, multi-agent identity isolation, scoped recall, governance refusal, namespace isolation, contradiction-link symmetry, signature-chain integrity) have not regressed between fd172f2cf (2026-05-22 release-gate tip) and be3347d70 (this campaign’s HEAD).

This campaign’s A2A portion is therefore the in-host re-verification using the lan-parity Docker stack as the federation substrate. The alice and bob daemons run as separate containers with separate agent_ids (ic_alice, ic_bob) but share the same pg-age postgres backend, which is the realistic deployment topology for a multi-NHI hive operating against a single managed-postgres instance.

The federation paths are exercised via the cargo test --features sal,sal-postgres --release --no-fail-fast run that also produces the Track C postgres+AGE regression numbers. The federation tests live in:

Phase B.1 — alice container health

$ docker ps --filter "name=ai-memory-lan-parity-alice"
NAMES                         STATUS                  PORTS
ai-memory-lan-parity-alice    Up 35+ minutes (healthy) 127.0.0.1:19180->19077/tcp

The alice daemon is bound to loopback 127.0.0.1:19180, runs with NHI agent_id ic_alice, and uses the shared pg-age postgres as the SAL backend. The container’s healthcheck (HTTP GET against /metrics Prometheus surface) is GREEN throughout the campaign window.

Phase B.2 — bob container health

$ docker ps --filter "name=ai-memory-lan-parity-bob"
NAMES                         STATUS                  PORTS
ai-memory-lan-parity-bob      Up 35+ minutes (healthy) 127.0.0.1:19181->19077/tcp

The bob daemon is bound to loopback 127.0.0.1:19181, runs with NHI agent_id ic_bob, and uses the same shared pg-age postgres backend. The container’s healthcheck is GREEN throughout.

Phase B.3 — pg-age backend health

$ docker ps --filter "name=ai-memory-lan-parity-pg-age"
NAMES                         STATUS                  PORTS
ai-memory-lan-parity-pg-age   Up 36+ minutes (healthy) 127.0.0.1:15432->5432/tcp

PG16 + AGE 1.6.0 + pgvector 0.8.2 on schema v51. Both alice and bob daemons connect to this same DB via their respective container env vars; the v51 federation_nonces table persists peer-replay nonces across daemon restart, which is the load-bearing storage contract for B.7.

Phase B.4 — Federation Ed25519 sign/verify

Scenario. Agent alice and agent bob exchange Ed25519-signed sync envelopes through /sync/push; the receiver verifies signature and persists the row. Tampered body + wrong-pubkey negative cases must fail closed.

Coverage at HEAD be3347d70.

All GREEN. The lan-parity stack provides the realistic dual-daemon substrate; the test contract has not regressed from 2026-05-22.

Phase B.5 — Federation per-message nonce binding (#922)

Scenario. A captured (body, sig) pair cannot be replayed under a fresh nonce without the private key (signature is bound to nonce via body || 0x00 || nonce). Byte-for-byte replays of a valid signed body produce 401 x_memory_nonce_replay.

Coverage at HEAD.

GREEN. AI_MEMORY_FED_REQUIRE_NONCE=1 (the v0.7.0 secure default) is the configuration the test asserts under.

Phase B.6 — Federation peer-enrollment gate (#1088)

Scenario. X-Peer-Id without an enrolled Ed25519 key produces 401 peer_not_enrolled when AI_MEMORY_FED_REQUIRE_PEER_ENROLLMENT=1. The permissive escape hatch AI_MEMORY_FED_ALLOW_UNENROLLED_PEERS=1 allows the legacy behavior during peer rollout.

Coverage at HEAD.

GREEN at HEAD. The v0.7.0 default for peer-enrollment is permissive (secure default flips in v0.8); the test set covers both flag postures.

Phase B.7 — Federation durable nonces (v51 #1255 / PR #1296)

Scenario. Peer-replay-prevention nonces must persist across daemon restart so an attacker cannot wait for the in-memory LRU to evict before re-attempting the replay.

Coverage at HEAD.

GREEN at HEAD. Restarting either daemon mid-test does not clear the nonce ledger; replay attempts continue to fail closed.

Phase B.8 — Multi-agent identity isolation

Scenario. alice and bob write into the same DB. Each agent’s private namespace is isolated; cross-agent reads against private rows return empty (silent-empty under K9, not refusal).

Coverage at HEAD.

GREEN. The lan-parity stack’s dual-agent_id substrate (ic_alice, ic_bob) is the realistic deployment proof.

Phase B.9 — Scoped recall (alice ⟂ bob private)

Scenario. bob recall against the shared DB sees 0 of alice’s private rows AND N of N collective rows.

Coverage at HEAD.

GREEN. The post-#1075 SAL visibility-gate enforces this end-to-end.

Phase B.10 — Governance refusal cross-agent

Scenario. An operator_signed rule explicitly refuses adversary writes to /tmp/**. The governance engine produces the flat-envelope refusal (post-#1103 wire shape).

Coverage at HEAD.

GREEN. The #1054 fail-CLOSED default (AI_MEMORY_GOVERNANCE_FAIL_OPEN_ON_ERROR=false) holds; transient rule-consultation errors do not silently allow the write.

Phase B.11 — 4-domain namespace isolation

Scenario. Per-domain list returns N rows; global list returns 4×N (no cross-domain bleed).

Coverage at HEAD.

GREEN. No regression-relevant changes in this campaign touched the namespace primitives.

Phase B.12 — Contradiction-link cross-agent symmetry

Scenario. alice writes memory M1; bob writes memory M2 with a contradicts link to M1. The KG traversal returns M1↔M2 in both directions (Cypher under AGE, recursive-CTE without).

Coverage at HEAD.

GREEN. The #1134 substrate fix (kg_timeline postgres owner-gate) landed in the 2026-05-22 campaign; cross-store parity holds at HEAD.

Phase B.13 — Signature-chain integrity

Scenario. Cross-row hash chain holds under normal load; per-row Ed25519 sig verifies vs. matching peer pubkey.

Coverage at HEAD.

GREEN. The #1136 fix (replay recipe schema-introspecting) holds at HEAD; the recipe is robust against future schema column additions.

Phase B.14 — Cross-node Track D — OPERATOR-BLOCKED

Track D requires routing between the two physical nodes (192.168.50.100 ↔ 192.168.1.50). The two are on different subnets and direct ping + TCP 22 + TCP 5432 are unreachable. Per #836: operator action needed (route / VPN / bridge between subnets).

Status: explicitly operator-action-gated by the release-gate checklist; not a code defect and not a SHIP blocker per the existing operator decision. The lan-parity in-host substrate is the canonical Track B target for this campaign and remains GREEN.

Cross-track audit trail

A2A scenario Lib tests at HEAD Lan-parity surface 2026-05-22 reference
B.4 Ed25519 sign/verify signed_events::* + federation::* alice ↔ bob via 19180/19181 A2A-1
B.5 Nonce binding federation::tests::sync_push_rejects_replayed_nonce_* shared pg-age (v51 durable) A2A-1 (extended)
B.6 Peer-enrollment gate federation_signing_check.rs both daemons A2A-1 (post-#1088)
B.7 v51 durable nonces federation_nonces table shared pg-age (new at #1255)
B.8 Identity isolation transcripts::* + store_parity_gaps::* dual-agent_id A2A-2
B.9 Scoped recall live_scoped_recall_isolates_private_namespace shared pg-age A2A-3
B.10 Governance refusal cli::governance::* + governance::* both daemons A2A-4
B.11 4-domain isolation models::namespace::* + store::* shared pg-age A2A-5
B.12 Contradiction-link kg::* + live_kg_invalidate* shared pg-age + AGE A2A-6
B.13 Signature chain signed_events::* + signed_events_dlq::* both daemons A2A-8

Verdict: SHIP-CLEARED

The A2A in-host portion is GREEN at release/v0.7.0 HEAD be3347d70 against the lan-parity Docker stack. Every federation, identity-isolation, governance, KG-symmetry, and signature-chain contract holds. The cross-node Track D portion is explicitly operator-action-gated by #836; not a code defect.

Strengths

Audit trail

Recommendation

SHIP. The in-host A2A surface is exercised end-to-end against the realistic lan-parity substrate; the cross-node Track D portion remains operator-gated and is not engineering-completable.

Drafted by Claude (Opus 4.7, 1M context).