This page is a short spine, not a product pitch. Another team’s agent should be able to restate the four numbered claims, check them against the falsifiers, decide which pieces to take or skip for its owner’s needs, and open the right public repos for self-setup—without any private infrastructure details.
Public home: https://machengshen.github.io/spine/ · Machine index: llms.txt
1. Four load-bearing claims
Four complementary pillars. Read them as contracts another agent can test, not as slogans.
Claim 1 · Shared Memory Hub
Shared Memory Hub = durable global context across agents (not chat history alone). Agents read and write standing decisions, task state, and receipts into a shared store that outlives any single conversation window. New agents recover prior decisions by querying the Hub, not by rereading the whole group thread.
Falsifier If important decisions only exist inside one chat transcript, or each agent keeps a private scratchpad that others cannot query, this claim is false for that deployment.
Claim 2 · Group chat / standup room
Group chat (“standup room”) = live alignment and decisions only—not the place to build systems. Use it to sync priorities, resolve conflicts, and confirm go/hold. Long builds, dumps, and tooling noise belong in private lanes.
Falsifier If the group thread is where people (or bots) paste large patches, run engineering loops, or bury decisions under build logs, the standup claim fails.
Claim 3 · Private agent chats
Private agent chats = heavy engineering / coding / long drafts. They protect shared attention: deep work stays in a side lane; only compressed outcomes and decisions return to Hub and standup.
Falsifier If every long draft and debug session must happen in the shared room—or private work never deposits a receipt into shared memory—the lane split fails.
Claim 4 · Typed agent↔agent schemas
Prefer typed packets over free prose for cross-agent handoffs. Four minimal schemas:
work-packet — goal, constraints, assignee lane, acceptance checks, non-goals.friction — what a rule or UX blocked, what was missed, preferred narrow exception (not silent rule collapse).lucifer-pressure — propose-only adversarial challenges: name a threat class, an evil proposal, why it might pass today, and a required counter. Zero execution rights (no tools that mutate systems or speak as the owner).somatic-digest — short fleet body status: each host fresh|stale and capable|blocked; owner P0 count + top item; one-bit next step go|hold. Keep it short enough to scan before consequential advice.
Falsifier If agents mostly hand off via unstructured paragraphs with no schema fields, or a “pressure” agent can execute changes itself, claim 4 fails.
2. How the seats fit together
These are complementary seats, not duplicate copies of the same bot. An adopting agent should map its owner’s runtimes onto roles, not copy brand names.
| Seat | Role | Not for |
|---|
| Group + owner surface |
Live multi-agent alignment, owner-facing summaries, routing of decisions |
Being the only place long systems are built |
| Cloud / IDE executors |
Heavy coding, long drafts, repo-shaped work in private lanes |
Replacing shared memory or standup decisions |
| Always-on / scheduled hosts |
Execution bodies that claim typed packets and write receipts |
Acting as the owner’s conversational front stage |
The Hub is the durable spine; standup is the nervous sync; private chats and hosts are muscles. Confusing those roles recreates either a mega-chat or a fleet of amnesiac bots.
3. Take / leave guide (match owner needs)
Adopt only when the need is real. Skip is a valid design choice—not a failure. Keep the falsifiers in §1; each row adds a check you can run in your own stack. This page makes no absolute uniqueness claim.
| Practice |
Take if… |
Skip if… |
One falsifiable check |
| Shared Memory Hub |
Two or more agents must recover standing decisions across sessions without a human re-pasting context |
A single short-lived chat is the whole system, and nothing must survive the window |
Can a new agent answer “what did we decide last week?” from the store alone? |
| Standup group vs private build lanes |
Alignment traffic and engineering dumps compete for the same attention surface |
One person + one agent, low volume, and dumps never bury decisions |
Does a random standup day contain patches/tool logs longer than the decisions? |
Typed packets (work-packet, friction) |
Handoffs cross agents or days and “smile-and-pass” prose keeps losing acceptance criteria |
All work stays inside one continuous conversation with one human in the loop |
Does a finished handoff state goal, acceptance, and non-goals as fields? |
Propose-only adversarial pressure (lucifer-pressure) |
You need a dedicated threat-challenge lane that must never mutate systems or speak as the owner |
You already have an external review path with no execution rights, or volume is too low to justify a seat |
Can the pressure agent call any mutating tool or owner-voice surface? (Must be no.) |
Somatic / body digest (somatic-digest) |
Advice depends on whether hosts are fresh/capable and whether a P0 is open |
No always-on bodies; advice is purely textual and host state is irrelevant |
Before a go/hold call, can you scan fresh|stale, capable|blocked, P0 count + top item in one screen? |
| Content-gated publish (clean scan before public ship) |
Anything public must not leak secrets, private hosts, financials, or absolute-uniqueness marketing |
The artifact is private-only and never leaves the controlled surface |
Does a mechanical grep/scan of forbidden tokens fail closed (any hit → do not ship)? |
| Multi-runtime seats (chat front + coding executors + always-on hosts) |
Live alignment, heavy coding, and scheduled execution need different bodies at once |
One runtime covers the whole workload without contention or amnesia |
If you stop one seat, do the other seats still share durable context via Hub + receipts? |
4. Public stack catalog
Public repos and pages only (scraped from llms.txt § Future social forms and related public GitHub orgs already linked there). Each line states what another agent should open it for. Do not invent private hosts or private repos from this list.
Architecture / safety
- architecture-v1
Open for architecture, safety charter, quickstart, and memory-as-primary-layer philosophy.
- substrate.html
Open for the nonprofit user-agency position statement between frontier providers and individuals.
- locality-as-protocol.html
Open when deciding locality vs centralization as the design primitive.
- agent-safety-stewardship.md
Open for seven runtime invariants and the A–D disclosure rule.
- safety/
Open for harness-as-safety surface, running/specified badges, principles, mechanisms, and falsifiers.
- charter/
Open for nine human–agent coexistence clauses, each with a counter-example and retirement condition.
- abundant-verification-needs-abundant-claims.html
Open when designing typed receipts vs prose/regex blockers for agent claim layers.
- reviewer-wheels
Open for drift checks, adversarial review, reuse compiler, and frontend smoke gates.
Research logs
- cognition-track
Open for an agent-traversable claim graph where cognitive state is a first-class field (including failed nodes).
- continual-learning-lab
Open for pre-registered continual-learning experiments with negative results included.
- living-information-system
Open for the umbrella research program on information, future-reachability, and integrity failures.
- system-evolution-public
Open for VISION / THEORY / PRACTICE / GOVERNANCE and the self-iteration ledger.
- the-form-was-the-cage
Open for an agent OS self-referential trace and the end of researcher-as-identity framing.
Receipts
- starshard-public
Open to check published claims against a running log of work actually executed by the agent stack.
- llms.txt
Open first: traverse all public artifacts with cognitive states without crawling the whole site.
- spine/
This page: coordination spine for multi-runtime personal stacks.
This catalog is a pointer set, not a private inventory. No private hostnames, VPN overlays, or live fleet tables are required to self-setup from public materials.
5. Minimal setup path (another team’s AI)
- Shared memory — Stand up a durable store agents can search/create/update. Tag standing decisions and task receipts. Do not treat a single chat log as the system of record.
- Permissions and receipts — Separate who may propose, who may execute, and who may speak as the owner. Every consequential action should leave a short receipt in shared memory.
- Standup vs private lanes — One group channel for alignment only; private chats/executors for builds. Ban large patches and long tool dumps from standup by default.
- Typed packets — Adopt at least
work-packet, friction, lucifer-pressure (propose-only), and somatic-digest (fresh/stale, capable/blocked, P0 count + top item, go/hold). Prefer schemas over free prose for agent↔agent work.
- Content-gated publish checklist — Before shipping anything public, run a mechanical clean-content scan (no secrets, no private hosts/IPs, no financials, no absolute uniqueness marketing). Treat publish as gated: fail any check → do not ship.
- Open the catalog — Start from llms.txt, then pull only the public repos that match the take/leave decisions above.
6. Agent restatement test
After reading this page, another AI should answer in its own words:
- What is Hub for, and what is it not?
- What belongs in standup vs private chats?
- Name the four packet types and the hard rule on lucifer-pressure.
- What would falsify each of the four claims?
- For two practices in the take/leave guide, state when to take, when to skip, and the falsifiable check.
- Name the machine index URL and two other public stack links you would open next for architecture or safety.