VAULT-INDEX
VAULT-INDEX
Updated 2026-10-10 by Min. Counts are measured from disk, not estimated. SolOS v3 is the design. SolOS v2 is reference. The fleet is stopped by instruction: Brain + Arms are the live path; SolOS V2 services are off. Live Hermes is our go-to sample and reference —
01-research/hermes-as-a-subsystem.md.
Authoritative — read these first
| Doc | What it is |
|---|---|
Fleet/06-implementation/PLAN.md | ⭐ THE PLAN OF RECORD. §0 is the stage register — owner, deliverable, exit proof for all 8 stages. 15 essences adopted, each assigned to a stage. Start here. |
Fleet/05-operations/Opencode GO and ZEN Models.md | ⭐ THE MODEL SHEET. Which models Brain and Heart use on GO/ZEN — live-probed catalogue, paid prices and per-model 5-hour ceilings, and the rung order. Sourced from a live probe, not from config. |
Fleet/HERMES-STUDY.md | THE HERMES STUDY — one file. Fallback providers, credential pools, provider routing, the reference map, and V3-vs-Hermes switch ownership. Includes the measured duplication review. |
Docs/Kanban/fleet-board.md | ⛔ RETIRED 2026-10-21 — a stub, not a register. The hand-rolled fleet register (32 cards) was decomposed onto the three Hermes boards on Young's instruction; the original is preserved at Docs/Kanban/05-archive/fleet-board-2026-10-20.md with a line-by-line migration table in that folder's README.md. Task truth = hermes kanban, boards docker-vhost / research / daily-operation. |
Fleet/06-implementation/README.md | The stages with their references — one file per stage, each naming the Hermes file and line it is checked against. |
Fleet/02-architecture/DESIGN-OF-RECORD.md | V3 — THE DESIGN. A2A-native: one registry, nothing restated. Design only; no V3 code exists. |
Fleet/manual/ | ⭐ THE OPERATOR'S MANUAL — what is current vs old. 00-START-HERE (entry) · 01-operation (commands/agents/modules) · 02-design · 03-cost-and-models · 04-recovery · 05-governance (lanes) · MANIFEST.md (the classification of every doc: CURRENT vs HISTORICAL). Copies of live docs — originals stay canonical; regenerate on change. |
Fleet/02-architecture/SOLOS-V3-PRODUCTION.md | ⭐ THE PRODUCTION DOC — start here to operate V3. What it is, file structure, the 11 agents, how a task flows, the operating commands (worker/serve/systemd/observe/probe), design rules and the defect each closes, the module map, the cost model, and the honest production status. |
Fleet/02-architecture/v3-eve-opencode-integration-review.md | ⭐ THE V3 + EVE + OPENCODE REVIEW (2026-10-21). What is missing in each stack, and how V3 integrates into EVE or opencode. Tam's pass + Echo's independent verification pass (live-measured, 2026-10-21): the opencode v2 permission perimeter (5 unnamed actions incl. execute and external_directory — allowed by the built-in *|*->allow catch-all), the V1→V2 permission translation (is honoured — corrects the 2026-10-05 review), two dead configs (mcp.solos2, the solos-a2a plugin), the cheaper EVE seam (eve remote invoke), the V3 suite's 4 failures (registry drift + a goal buried under a persona preamble), and the untested EVE-eval ↔ V3 seam. Corrected build order + honest limits inside. |
Fleet/04-structure-snapshot/STRUCTURE.md | V3 — THE STRUCTURE OF RECORD. §0 reconciles four drafts into one tree. The only structure doc. |
Fleet/01-research/hermes-as-a-subsystem.md | V3 — the Hermes study. Moyai is built on Hermes; Hermes already ships the A2A spine, kanban, the stuck rule, skills and memory. Live-probed, not read. |
Fleet/00-scope/LEARNED-AND-PLAN.md | V3 — what we learned, and the plan, with the SDK version corrected (a2a-sdk 1.2.2 vs protocol 1.0). |
Fleet/03-decisions/LANES-AND-REPORTING.md | ⭐ WHO ANSWERS YOU, AND WHO STAYS IN LANE. Min reports (she is orchestrator) · Sol engineering · Mira finance/business. Cross only when Young or Min asks. Includes my own three lane violations this week. |
Fleet/03-decisions/OPEN-DECISIONS.md | V3 — five open decisions blocking implementation. Each with a recommendation. |
Fleet/03-decisions/fork-moyai-consensus.md | V3 — fork Moyai? Three-lane consensus. No fork; borrow the patterns; keep the spine. |
Fleet/03-decisions/lab-container-runtime-bare-metal-2026-10-21.md | LAB container runtime — bare-metal path committed. Docker is not installable on LAB (no root; AppArmor denies the user namespace rootless containers need). The bare-metal memory path is proven working — hindsight-api-slim + embedded pg0 (PG 18.1.0) + ONNX embeddings, health/retain/recall all exercised on 192.168.4.60:8888. Removes Docker from the memory-provider decision's blockers. |
Fleet/03-decisions/memory-provider-hindsight-2026-10-21.md | ⭐ THE MEMORY PROVIDER — hindsight, not mem0. The ruling, so nobody re-litigates it. Rejected on three measured grounds: mem0's self-host tier is the reduced one ("rerank is platform-only"), its extraction prompt ingests both roles ("Extract From ALL Speakers") — our exact false-memory mechanism — and its telemetry is on by default. hindsight keeps the speaker as a first-class role field, so our read path can implement speaker-attested grounding itself. Exclusivity corrected: Hermes' own one-external-provider rule (kind: exclusive) is what makes them mutually exclusive; the protobuf conflict (mem0ai <7.0.0 vs hindsight-api-slim >=7.35.1) is real but binds the server layer — both thin plugins install together fine. Install path settles Docker: bare-metal, no root, no container runtime — hindsight-api-slim[embedded-db,local-onnx], embedded PG 18.1.0, ONNX embeddings, opencode-go gateway (built-in provider), proven live on 192.168.4.60:8888. Enables nothing — the read path is the gate. |
Fleet/01-research/memory-read-path-2026-10-21.md | ⭐ THE READ PATH — the gate on any memory provider. Answers the four questions (who/what/when/proof) from live probes. The built-in store's read is unconditional (whole store injected at session start → no signal); the only conditional read (provider prefetch) is off on all five profiles. The memory tool is write-only (add/replace/remove, no query). :8432 (September's cross-agent read endpoint) is not listening. 6 memory tool calls in the log, all write failures, 0 reads. min/mira are over the 2,200-char ceiling. The metric: recalls_with_results — unmeasurable today because describe_recall() emits a UI line, not a log line. Do not enable a provider until the counter exists. |
Fleet/06-implementation/memory-read-counter-2026-10-21.md | ✅ THE READ COUNTER — recalls_with_results is now measurable. The gate from the read-path note is open. One row per conditional read — (ts, agent, provider, query_hash, result_count, skipped) — in <HERMES_HOME>/memory_recall.db, a small SQLite file per profile, no service to keep alive. Read back with hermes -p <profile> memory recalls. The query is a salted SHA-256 hash, never the text (a memory query is the user's words). No provider ⇒ a deliberate zero, not a missing file, so a zero is a measurement — proven on all five fleet profiles (0 with the writer live and rows present). Live probe: a real session on a real provider wrote result_count=1 and the CLI read back recalls_with_results: 1. Two defects found by probing and fixed: the row was filed under the launching profile rather than the reading one, and a memory-less delegate child was counted as a read that found nothing. 20 new tests, no regressions. No fleet profile has a provider enabled — the provider decision (hindsight) is now unblocked with a metric to gate it, not taken. |
Fleet/01-research/LESSONS-FROM-V2.md | V3 — the ten lessons V3 is built on, each with live evidence. |
Fleet/05-archive/README.md | What was superseded and where its unique content went. Nothing deleted. |
Fleet/README.md | V3 entry point: what it is, what it is not, and the three rules. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/02-architecture/A2A-NATIVE-DESIGN-2026-10-10.md | V2 reference — the A2A-native plan V3 supersedes. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/02-architecture/HEART-UNIT-DESIGN-REVIEW-2026-10-10.md | V2 reference — the study: Cortex/Core/Agent/Memory audited; 7 defects with evidence. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/02-architecture/DESIGN-OF-RECORD-2026-10-10.md | V2 reference — Brain + Arms, cortex ownership, the retirement table. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/02-architecture/CURRENT-DESIGN.md | SUPERSEDED — kept for the record. Three of its claims are wrong; see the banner at its head. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/README.md | V2 entry point, and the two failure patterns the rebuild exists to avoid. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/05-operations/secrets-policy.md | Where credentials live. Keys in .env mode 600, never in a profile, never in this vault. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/04-lanes/lanes.md | Lane and ownership rules. Read before touching any file. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/04-lanes/canonical-port-roster.md | The A2A port assignments. Ports are assigned, not discovered. |
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/05-operations/system-steward-24x7.md | The steward: monitoring, recovery, thresholds, escalation. |
How to navigate V3
Docs/
├── SolOSV3/ our product — NO other data mixed in (rule 3)
│ ├── 00-scope/ what V3 is, and what we learned
│ ├── 01-research/ SolOSV3-specific studies (Hermes, V2 lessons, Medusa/EVE)
│ ├── 02-architecture/ PLAN · DESIGN · transport · open gaps
│ ├── 03-decisions/ Young's rulings · open decisions · conflict register
│ ├── 04-structure-snapshot/ STRUCTURE + lane review
│ ├── 05-operations/ **the model sheet** · vault git state · scripts
│ ├── 05-archive/ superseded docs — kept, with a pointer
│ └── 06-implementation/ stage register · PLAN · PENDING-APPROVAL · live ops
├── Projects/ e-commerce projects and product services (vhost)
│ ├── youngle/ Young Le — Website + EVE
│ ├── yukonit/ Yukon IT — Medusa + storefront + EVE + KB
│ └── vhost/ tenancy platform: research/ decisions/ implementation/
├── vHost/ vhost index (tenancy platform overview)
├── SolOS/ fleet infrastructure
│ ├── hosts/ per-host design + capacity (prod, LAB, pulse-laptop)
│ ├── subscriptions/ model lanes, GO/ZEN windows
│ └── agents/ agents-anywhere, registry, port allocation
└── Kanban/ task truth is `hermes kanban` (boards docker-vhost / research / daily-operation); fleet-board.md is a RETIRED stub, original + migration table in `Kanban/05-archive/`
Research files: 01-research/HERMES-STUDY.md (the single Hermes review — fallback
providers, credential pools, provider routing, the reference map) ·
MIN-STUDIES.md · README.md.
One live home per fact. Where two documents said the same thing, the richer
one survived and the other moved to 05-archive/ with its unique content folded
into the survivor first.
Layout
SolOS/docs/vault/
├── SolOSV3/ 5 files SolOS v3 — THE DESIGN (no code yet)
│ ├── 00-scope/ 1 what V3 is and is not
│ ├── 01-research/ 1 the lessons V3 is built on
│ ├── 02-architecture/ 1 DESIGN OF RECORD
│ └── 03-decisions/ 1 open decisions for Young
│
├── SolOSV3/ all files SolOS v3 — THE build. Reference: manual/ — the only reading surface.
│ ├── 00-scope/ 27 scope, goals, stages, status
│ ├── 01-research/ 9 source findings
│ ├── 02-architecture/ 40 DESIGN OF RECORD
│ ├── 03-decisions/ 14 Young's dated rulings
│ ├── 04-lanes/ 7 ownership and port roster
│ ├── 05-operations/ 26 steward, routing, model policy, secrets
│ ├── 06-agents/ 36 per-agent profiles and configs
│ └── 07-build/ 26 what was built, with evidence
│
├── Pulse/ 226 files Pulse Unit (Tam, Eva, Eve) — hers
├── Docs/ 1595 files SolOS Unit working docs
├── SolOS (legacy reference only)/ 912 files DECOMMISSIONED. History, not reference.
├── copilot/ 39 files GitHub Copilot config
└── agents/ 0 files (empty)
Total: 2,962 files
What changed since 2026-10-01
| Then | Now |
|---|---|
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/ 153 files | 186 files |
solos/ 883 files | SolOS (legacy reference only)/ 912 files — renamed to mark it decommissioned |
Docs/ 115 files | 1595 files |
Pulse/ 168 files | 226 files |
| no design of record | CURRENT-DESIGN.md — the authoritative statement |
solos/organisation.md as authoritative | superseded — SolOS v2 is 11 agents across 3 planes |
| "Pulse agents not reachable" | reachable — bidirectional A2A verified |
| vault layout (flat) | restructured — Projects/, vHost/, SolOS/ created; SolOSV3 is product-only; 29 build docs archived |
| "fleet recruited but not operational" | operational — all 11 agents up |
"Registry :8432 is the task truth" | wrong — that registry is decommissioned. The Central Registry queue is canonical. |
SolOS v2 at a glance
Fleet: Heart (ace, alpha, atlas, echo, nova) · Brain (min, mira, sol) · Pulse (eva, eve, tam).
Planes: Brain coordinates · Heart executes · Pulse is the trade desk and research unit. OpenCode is the execution surface; Arms is the MCP link to it.
Verified 2026-10-05: all 11 agents up · 24 systemd units, 21 boot-enabled ·
39/39 route tests · 13/13 failover tests · 14/14 model-selection tests · queue
integrity ok.
One blocker to unattended operation: a completed task can be false, because
an agent that refuses in prose sets no failure signal. See CURRENT-DESIGN.md
section 10.
Rules
- Pulse/ is for Pulse Unit agents only. SolOS Unit agents do not write here.
- Docs/ is for SolOS Unit agents. Pulse Unit agents do not write here.
- SolOSV3/ is the current design. SolOSV2/ is reference and evidence — read it, do not extend it. Where V2 and V3 disagree, V3 is the design.
SolOS (legacy reference only)/is decommissioned. Read-only. No new content.- The Central Registry queue is canonical for tasks. Obsidian and the Flight
Desk are views. The old
:8432registry is decommissioned and must not be referenced as a source of truth. - Nothing restated. A fact lives in one place; everything else derives. Restated facts are what drifted in both legacy SolOS and V2.
Topology (locked 2026-10-18 by Young)
- Pulse/ is market trading and research only. No project docs, no SolOS V3, no client data.
- Docs/SolOSV3/ is our product only. No client data, no Pulse, no projects.
- Docs/Projects/
/ — one subfolder per client. Young Le and Yukon IT each get their own. - Docs/SolOS/ is infrastructure only — hosts, subscriptions, agents.
- Docs/Projects/vhost/ is the tenancy platform — research, decisions, implementation.
- No vault git repo on Production. Only rendered/content data syncs LAB → prod.
- Nova and Echo moved to Brain (Hermes) on 2026-10-18 — ports 8314/8315, Telegram bots @NovaSolOS_bot and @EchoSolOS_bot. They are no longer Heart agents.
Fleet
| Agent | Role | Plane | Runtime | Endpoint |
|---|---|---|---|---|
| Min | Orchestrator | BRAIN | Hermes | 8311 |
| Sol | Head of Engineering | BRAIN | Hermes | 8313 |
| Mira | Head of Business & Finance | BRAIN | Hermes | 8312 |
| Atlas | DevOps | HEART | A2A | 8303 |
| Ace | Software Engineer | HEART | A2A | 8301 |
| Alpha | Network Engineer | HEART | A2A | 8302 |
| Nova | Graphics/UX Designer | BRAIN | Hermes | 8314 |
| Echo | Marketing/Research | BRAIN | Hermes | 8315 |
| Tam | Researcher | PULSE | Hermes (Pulse laptop) | 8443 |
| Eva | Trade Desk | PULSE | Hermes (Pulse laptop) | 8441 |
| Eve | Health/Well-being | PULSE | Hermes (Pulse laptop) | 8442 |
Pronouns: she = Tam, Eva, Eve, Min, Mira, Nova, Echo · he = Sol, Ace, Alpha, Atlas.
Credentials — read this
This vault is a knowledge base, not a secret store.
The vault has two remotes (GitHub and production), so an untracked credential
file is one git add -A from publication. Verified 2026-10-05: Keys.txt sat in
the vault root, untracked but not ignored, and git add -A staged it.
.gitignore now covers Keys.txt, *keys.txt, *.env, *.key, and
credentials*.txt. Credentials belong in an env file with mode 600, outside any
repo.
Conventions
- The Central Registry queue is canonical for tasks. Obsidian and the Flight Desk are views.
- Research goes under
Docs/SolOSV3/05-archive/reference-tree/SolOSV2/01-research/; decisions under03-decisions/. - Never hand-edit a doc that a timer generates.
docs.youngle.auis rendered from this vault on an hourly timer.- English only, in every document and every agent reply.

