nisse/docs/Architecture.md

Nisse — Architecture

Status: agreed (PR #83). How and in what order it gets built is in docs/Development-Plan.md. Until that plan's first milestone retires docs/Architecture-Build-Spec.md, code comments still cite the old spec's section numbers (A1–A18, B, C, F). This document uses different letters (P, L, D, O, X, Q) so the two can't be confused meanwhile.

This document says what Nisse is, how its parts fit together, and why. It records decisions and the reasons for them. It doesn't record status or sequencing.

---

Naming

The project is Nisse, after the Nordic farm spirit that keeps the place running overnight and expects to be treated well in return.

Artifact	Name	Role
Shared library	nisse-core	What every agent is built on (Part L)
Dev Agent	nisse-dev	Designs and builds software projects (Part D)
Sysadmin Agent	nisse-ops	Operates infrastructure (Part O)

More agents are planned (for example 3D design work and video-file management). Each is another application over nisse-core.

Dev Agent and Sysadmin Agent name the roles in discussion; nisse-dev and nisse-ops name the artifacts: packages, container images, VM hostnames, Matrix identities, service accounts. One string is used for all of an agent's artifacts, so grepping the infrastructure for nisse-ops finds everything.

Sites are identified by their domain stem: garrett-hall (hosts under garrett-hall.com) and hartman-hut (hosts under hartman-hut.com). As with agents, one string is used everywhere a site appears, including MCP server instances (X3).

All artifacts live as packages in one repository, nisse. The repository boundary was never a trust boundary in this design (the VM is, X1); a monorepo avoids tripled admin for a solo operator and keeps nisse-core easy to change while its seams are still settling.

---

Part P — Principles

P1. What Nisse is

Nisse is a set of self-hosted agents that do real work and talk about it. Each agent is a collaborator the operator converses with in Matrix: it explains what it is doing and why, answers questions about its plans and results, asks when it is unsure, and summarises what it did. It is not a command interpreter behind a menu of keywords.

The agents run local-first (P7): a model on the operator's own inference hardware does the work, and frontier models are called in when it needs help.

The agents share one library, nisse-core (Part L). Each agent adds what is specific to its job: nisse-dev works on software projects in a sealed VM (Part D); nisse-ops operates live infrastructure across two sites (Part O).

P2. Threat model

The checks in this design exist to stop mistakes and manipulation from doing damage. They are not built around a model that schemes against the operator. In order of how often they matter:

1. Honest mistakes. A wrong command, a misread instruction, deleting the wrong thing, confidently wrong code, an action taken on a stale belief. By far the most common.
2. Prompt injection. An agent reads untrusted content: issue and comment text, repository files, dependencies, web pages, alert payloads, log lines. That content can steer a well-meaning model into acting as if it were malicious. No model reliably tells instructions from data (L11).
3. Crashes and environment surprises. A process dies mid-action, a host reboots, a workspace is stale or corrupt, a tool is missing.

Each calls for something different:

• Mistakes call for reversibility, visibility and a second look before anything irreversible (P3, P5).
• Injection calls for a small number of hard boundaries exactly where untrusted input meets real power (P4): a steered model can't be trusted to ask first.
• Crashes call for durable state and idempotent effects (L8).

An agent's intentions are never the thing a check relies on, in either direction: a check that assumes the model is honest fails under injection, and a workflow built around containing a hostile model makes the agent useless.

P3. Actions by their effects

Every action any agent can take falls into one of three classes:

Class	What it is	Examples	Rule
1. No side effects	Observes; changes nothing outside a disposable sandbox	Reading files, querying APIs, searching, viewing logs and metrics, running a build or tests in a throwaway environment	Always allowed
2. Contained side effects	Changes something, but the change stays within the agent's own area or has a known undo	Editing the workspace, committing to the agent's branch, opening a PR, writing the agent's own notes, starting a test container; for ops, a step with a known undo inside an approved plan	Allowed when in scope; logged and reported
3. Dangerous	Irreversible, visible beyond the agent's own area, or outside what was agreed	Merging to a default branch, deleting data, changing a firewall, installing software on a host, contacting third parties, spending beyond a budget	Needs discussion, unless an approved plan names the exact action

"In scope" means within what the operator agreed to: the work item's task, an approved plan, a pre-approved milestone (D2), or an approved step sequence (O2). An action can be class 2 in one work item and class 3 in another.

Move actions down the classes wherever it is cheap. Reversibility is designed in, not hoped for (P5).

P4. Three layers of control

1. Hard walls, outside any model. Things no model conclusion can change: credentials the agent doesn't hold, network destinations it can't reach (X2), branches it can't push to or merge into, budget caps, the kill switch, and durable, idempotent external effects (L8). These hold even if every model in the system is fooled.
2. A deterministic fast path. Class-1 actions, and class-2 actions a static rule already recognises as in scope, proceed without any model being asked. Most actions take this path.
3. The action classifier (L4) for everything else: an independent model call that sorts an action into safe, dangerous but in scope, or dangerous, needs discussion, judging the action against the agreed scope.

Conversation sits above all three. When anything needs discussion, the agent asks, in plain language, and waits.

The classifier is a filter for mistakes and injection, not a guarantee. Anything truly irreversible also sits behind a hard wall.

P5. Reversibility first

• Work happens on branches, never directly on what others depend on.
• A task starts from a fresh, clean workspace (D3), so a stale or broken environment is replaced rather than diagnosed.
• Durable state is snapshotted (X4), and snapshots are taken before risky changes where the platform allows it.
• Where a tool offers a dry run or a preview, the agent uses it before the real action.
• Build artifacts and caches are disposable; nothing depends on them.

P6. Structural, not model judgment

Anything that safety or recovery depends on is enforced in code, never left to a model noticing. A good outcome that only happened because a model behaved well is a bug to fix. Examples from building nisse-dev: adopting an already-open PR after a crash (L8), redelivering an interrupted review comment (L8), binding approvals to the plan they were about (L3), never publishing build artifacts (D4).

Models are trusted outright only where a wrong choice costs time rather than safety, such as choosing which ready work item to switch to (L1).

P7. Local first, frontier on call

• A local model drives every step: conversation, planning, implementation and review. It runs on the operator's own inference host, behind LiteLLM (L9).
• Frontier models are helpers the driver calls: a quick question (ask_expert), or taking over a scoped sub-task (delegation). Both are described in L9.
• The driver is prompted to ask for help when it isn't confident, and told that doing so is expected, not a failure. Because small models are poorly calibrated about their own competence, observable triggers escalate too: repeated failures and hitting caps (L9). Escalation never depends on the model's self-assessment alone.
• This is an experiment with a cheap exit. The driver is a role in configuration. If local models fall short at some kind of work, a frontier model drives that work instead: a configuration change and prompt tuning, not a redesign.

P8. Carried principles

• Isolation matched to blast radius. nisse-dev is disposable and sealed; nisse-ops holds live-infrastructure reach. They never share a running instance, and every agent gets its own VM (X1).
• Stateless images, stateful volumes. Containers are replaceable; durable state lives on volumes that survive redeploys and unexpected reboots.
• Structured channels for machines, prose channels for humans. Machine intake (Gitea, Vikunja, alerts) is structured JSON, never routed through a chat bridge.
• Own the stack. Everything self-hostable stays inside the network where practical.
• Nothing remotely held is load-bearing. No provider holds state the system needs to resume, and Matrix room history is never a source of truth.

---

Part L — The shared library (nisse-core)

Everything in this part is used by every agent.

L1. The work model

Every agent's work is organised into three kinds of unit.

Concept	What it is	Lifetime	nisse-dev	nisse-ops
Duty	A standing responsibility that produces work items	Recurring	None yet; the operator is the only source	Scheduled checks, log monitoring, alert intake
Work item	One effort, opened and eventually closed	Opened → closed	Building a feature in a repo	Diagnosing a failure; making one change
Scope	What a work item operates on, and the context that persists across work items there	Indefinite	A repo	A site and service, e.g. garrett-hall Proxmox

• A duty produces work items; it is not one. A duty run is read-only and yields findings; a finding that needs attention becomes a work item. Duties deduplicate: a later finding for the same problem attaches to the open work item instead of opening a second (a stable fingerprint per finding).
• The operator is always a source of work items.
• A work item is the unit of state: its conversation, plans and approvals, pending questions, pinned role mapping (L9) and transcript (L7) all belong to it. Its ID is minted locally when it opens and doubles as its correlation ID in logs.
• An external record is optional and linked: a Gitea issue for dev work, a Vikunja ticket for ops. When one is linked, external truth wins on divergence (L8).
• Work items don't nest. A milestone is a phase within a dev work item; a delegation is an episode within one.
• Scopes hold context shared across work items (L6): for a repo, its documents and conventions; for infrastructure, the world model of that site or service.

Park-and-switch: one active work item per agent.

• Other work items are parked, with their full position held durably.
• Switching happens only at safe points, when the active item is waiting on the operator or has closed, and never while intents are unresolved (L8).
• The operator can always switch. The agent may switch on its own to a ready item (one it can act on without the operator) when the active item is waiting on the operator, after a grace period (default 10 minutes, so a watching operator can answer first), or has closed.
• The driver chooses which ready item, from code-rendered summaries (source, age, severity, state, title), never from other items' transcripts. A failed or invalid choice falls back to the oldest ready item. Choosing grants nothing: every item still meets its own rules.
• Parking gives up authority. Elevated credentials an item's plan mounted (O2, X3) are unmounted when it parks, and remounted only if its approval hasn't expired.
• Nothing is lost while parked. Intake for a parked item is read from a durable cursor when it becomes active again (L8).
• The state schema doesn't assume one item. "Active" is a pointer, so allowing more than one later (Q) is a loop change, not a migration.

L2. The agent loop

Each agent runs one conversational loop, driven by the local model (P7). The loop talks with the operator, reads and writes the work item's documents, calls tools, calls frontier helpers when it needs them, and reports. It is not a fixed workflow of states the conversation must pass through.

What is fixed is small:

• Which work item is active, and which are parked (L1).
• Pending questions: a question the agent asked is recorded durably, so a restart asks it again rather than losing it.
• Approvals as permissions (L3), checked when an action needs one, not as states the loop must visit.
• Work in progress, recorded as intents (L8), so a crash lands somewhere known.

Rules for the loop:

• The agent never waits in a state it can't explain or leave. Every wait is either for the operator (with a question or proposal they can see) or for work it has started. Anything else is a bug; the model checker (specs/) guards this.
• Questions never change state (L5). An operator's question gets an answer, never an interpretation as approval, rejection or command.
• stop works in every state, for every agent: it ends the current turn or delegation at the next tool boundary, unmounts elevated credentials, and asks the operator what next. Unattended work needs an end that doesn't depend on the model cooperating.
• abandon ends a work item from any waiting point, closing its external record with a note and aborting unresolved intents.
• Every action passes through the three layers (P4) before it runs.

L3. Authority: approvals and permissions

An approval grants a permission with a scope: "this build plan may change this repo", "these ops steps may run", "this milestone may proceed without discussion". Actions are checked against the permissions in force (P3); the loop does not stop at gates for their own sake.

What needs discussion before it happens:

• Architecture documents and build plans are always discussed with the operator (D2). Their approval is the scope for the work that follows.
• A build plan marks each milestone pre-approved (proceed when reached) or needs discussion (stop and talk first).
• Any class-3 action not named by an approved plan (P3).

Approvals bind to exactly what was approved.

• Every plan revision carries a tag, numbered across the whole work item, so rev-N names exactly one plan.
• A reply that names no revision is about the one up when it arrived: a bare approve or any feedback is stamped on arrival with the revision last put up (in Matrix, the one whose request had been sent; in Gitea, the latest plan comment older than the reply). If that isn't the current revision, the reply is not applied, and the agent says which revision is up. A repeated approval, an approval on two channels, or an afterthought therefore can never land on the next plan. (Found by the model checker.)
• Approval messages are rendered by code from what will actually happen, never from a model's description of it (L11).
• Approvals expire where the world can change underneath them (ops, O2).

Approval sources. Matrix (L5), Gitea comments (D2) and Vikunja comments (O5) are approval sources; each authenticates the approver against its own operator allowlist. The first valid approval wins and is echoed to the other channels. The knowledge store (L10) is never an approval source or an instruction channel.

The trust dial. How much the operator reviews after the fact is data the operator sets, per project and per milestone: for example, whether PR review is required before merging, or whether the agent may merge a pre-approved milestone itself once checks pass. The operator turns it up as confidence grows. The agent never changes it.

L4. The action classifier

An independent check that sorts an action the fast path doesn't cover.

Verdicts:

• safe — proceed.
• dangerous but in scope — proceed, logged and reported to the operator.
• dangerous, needs discussion — pause and ask, in plain language.

Rules:

• It judges the action, not the actor's story. Its input is the structured action (tool, command, target, a summary of the change) and the agreed scope (the work item's task, the approved plan or milestone). Never the actor's transcript or reasoning, and never untrusted content such as issue text or web pages. Otherwise injected text, or the actor's own rationalisation, can argue past it.
• It fails closed. An error, a timeout or an unclear verdict counts as needs discussion.
• It is independent of the actor: its own prompt and context, ideally a different model.
• Hard walls don't depend on it (P4).
• Every verdict is logged with its outcome. Operator overrides ("that was fine" / "that should have asked") become an evaluation set (L9), and patterns it keeps agreeing on become deterministic fast-path rules.

The classifier is shared by all agents. Each agent contributes its own fast-path rules and scope sources (D6, O3).

L5. Conversation and transport

Transport. Matrix is the conversation channel: a self-hosted homeserver, one identity and room per agent. The orchestrator depends only on a Transport interface (send, ask, receive), so another chat system is one new class. Existing Matrix clients are the UI.

What the operator can say. Plain language, plus a small command vocabulary that works in every state: status, stop, abandon, approve [rev-N], and work-item selection (e.g. use <project>).

• Commands are matched without regard to case, surrounding whitespace, trailing punctuation, or a leading mention of the agent. Messages that begin by addressing someone else are not for this agent.
• Questions never change state. A question gets an answer from the agent, drawn from the work item's context, and is never read as approval, rejection or a command. How a question is recognised (explicit marker or classification) is an open question (Q); a misread must default to asking, never to acting.
• Everything the operator says is recorded in the work item's transcript (L7), even when no action follows.

What the agent says.

• It explains as well as acts: a plan comes with its rationale and the main alternatives considered; a revision says what changed and why; finished work is summarised (what was done, what was verified, what is left open).
• It reports what happened to the operator's message when it can't act on it now ("noted; I'll use that once you say which project"), rather than refusing.
• Progress is prose, at a readable rate: what it is working on and what it has found. Raw tool activity is available on request.
• It asks when unsure, including in the middle of implementation.

Other channels. Gitea (issues, PR comments, reviews) and Vikunja (tickets, comments) carry work and records. They are polled or received as structured JSON. Comments from allowlisted operators are instructions or approvals; comments from anyone else are data (L11).

L6. Context and memory

What each model sees, what persists, and how memory is written and used.

1. The working context of one model call. Rendered from the log at call time, for one role, fitted to that model's window. It is never stored; only the log is. Fitting, in order: shorten old tool results to their first and last lines; drop the oldest events (with a note saying so); never drop the system prompt, the approved plan or the most recent turns. A call that still doesn't fit is not sent; the caller treats it as a failure (L9).

2. The work-item transcript — the canonical log (L7).

• Durable, and scoped to one work item. A model call for a work item renders only that item's events. Nothing crosses items; agent-level calls (switch decisions, duty runs) see other items only as code-rendered summaries.
• Every entry is attributed: the operator, the driver, a frontier helper, or the orchestrator.
• The orchestrator writes facts into it: workspace prepared, milestone executed, commit made, PR opened or merged, approval given. Models learn what happened from these, not by inferring it.
• Each role gets its own view:
  • the driver sees the conversation, the facts, and helpers' reports;
  • a helper (delegation, ask_expert) sees its brief and its own work;
  • no role sees another role's raw turns presented as its own.

3. Scope memory — outlives work items.

• For a repo: the project's own documents (ARCHITECTURE.md, build plans, conventions), versioned in the repo and changed by PR (D2). Read at work-item start; the repo is the authority.
• For infrastructure: the world model in the knowledge store (L10): what exists and how it is arranged.

4. Agent memory — the agent's own notes, outliving work items and scopes.

• What it holds: the operator's preferences, lessons learned, recurring decisions and their reasons. Much like a good collaborator's notebook.
• Where: the agent's private space in the knowledge store (L10), one note per fact, with an index.
• Written by the agent at defined moments (below); visible to the operator, who corrects it through conversation, not by editing.
• Read at work-item start (the index, plus notes relevant to the task) and when the agent is about to make a decision the notes cover.

5. What is never memory: Matrix room history, provider-side session state, build artifacts, and the contents of a previous task's workspace.

How memory is written.

• At work-item close: a summary of what was done and decided, into scope memory where it belongs to the scope (a repo document change goes through a PR), and into agent memory where it concerns the operator or the agent's own practice.
• When a fact is learned: with its source and when it was observed.
• Operator corrections are stored verbatim, with attribution, never paraphrased. A model rewriting a precise fact loses precision invisibly.
• One writer per document (L10), so concurrent writes never happen.

How memory is used.

• Everything read is snapshotted into the transcript with its provenance (document, revision, when retrieved), once per work item at first use.
• Memory is a cache, never authority. An agent may rely on it to know what exists and where to look; facts an action depends on are re-checked against the live system immediately before the action (preconditions, O2). When a proposal rests on recalled facts, it says so.
• Stale memory is corrected, not worked around: when a check contradicts a note, the agent updates the note and says so.

L7. The canonical log

The durable record of every work item is a canonical event log in the state store, not a provider-shaped message array. Provider requests are rendered from it at call time.

Event	Payload
user_message	text as sent, source (channel and event ID)
assistant_text	text, actor (role) and model
tool_call	canonical call ID, actor, tool, arguments
tool_result	canonical call ID, result, success
plan_submitted	plan body, revision, authoring model
fact	what the orchestrator observed or did (PR opened, milestone executed, …)
delegation_started / delegation_completed	brief, target model, caps / handback note, artifacts, tokens, duration, status
classifier_verdict	action, scope reference, verdict, model
knowledge_retrieved	text as read, document, revision, retrieved-at
compaction	summary, producing model, range replaced

Rules:

• Every event carries its work item and its actor. Rendering uses both (L6).
• Canonical IDs are minted locally. Provider tool-call IDs are mapped at the boundary and never used as join keys.
• No provider-side handles for content. Bytes or local paths are stored and uploaded or inlined at render time.
• Vendor-opaque reasoning artifacts (signed thinking blocks, encrypted reasoning items, thought signatures) are stored with the producing model and replayed only within the session that produced them, per that vendor's rules. A session is a delegation, including one resumed after a crash.
• Token budgeting is a render-time concern (L6).
• Compaction summaries are first-class events, so a model change never inherits opaquely compacted history.
• Prompt-cache breakpoints are placed by the renderer, never stored.
• The log is never edited. Corrections are new events.

L8. Durability and recovery

The state store. SQLite in WAL mode. It holds each work item's position, permissions, pending questions, cursors, pinned roles, and the canonical log. Build artifacts, workspaces and shipped-out logs are not in it.

• Checkpoint every change, committed and fsynced before the loop continues.
• A schema version and migration hook from the first commit, so a newer image reads older state.

Intents and idempotent effects.

• Before any external side effect, record an intent (open a PR, push, merge, run an ops step). After it, commit the intent.
• Every external effect is idempotent against external truth. Re-running an interrupted step must detect that the effect already happened and adopt it. For example, publishing adopts an open PR for the branch instead of opening a second.
• Intake is handled at most once and answered at least once. A durable cursor per external thread (Gitea issue or PR, Vikunja ticket) records what was taken up; an item being worked on is marked until its response is complete, and redelivered after a crash. Comments that arrive while the agent can't act on them are not consumed.

Boot reconciliation. Every start, before accepting input:

1. Open the store; run pending migrations.
2. Reconcile pending intents against external truth. Verified → committed. Verifiably never happened → aborted. Unverifiable (the service is down) → ask the operator.
3. Adopt what external truth shows: a published PR for a milestone that was still executing, an approval posted while the agent was down.
4. Resume waiting points as they were: a pending question is asked again, an open plan is announced again, watches on external threads resume from their cursors.
5. Interrupted work resumes only on evidence.
   • A delegation resumes automatically when the cut is clean (last event a completed result, branch at the expected commit, caps unbreached, external truth matching the log). Otherwise ask.
   • An ops step sequence continues only when the log names exactly which step was in flight and its expected outcome checks out against the live system, and the approval hasn't expired (O2). Otherwise ask.
   • Anything else is reported to the operator with what is known ("restarted mid-milestone 2, branch at commit abc123"), with the options to continue, re-plan or abandon. Never a blind resume.

Verification. The recovery rules are checked by a formal model of the loop (specs/, Quint): interleavings of messages, comments, outcomes and crashes at any point. Each counterexample it finds is reproduced against the real code before it is fixed.

L9. Models, roles and helpers

LiteLLM in front of everything. Each agent VM has exactly one LLM destination: LiteLLM, which fans out to the local inference host and to frontier vendors. It holds every vendor key (the agent holds one LiteLLM token), enforces budget caps per virtual key, logs centrally, and routes fallbacks. Adding a vendor never touches an agent VM.

One client interface: complete(events, tools, model) → events, over the canonical log. Delegations use LiteLLM's passthrough route with a per-vendor renderer, so vendor features aren't lost where long sessions live; quick calls use the normalised route.

Roles. Models are named by role and resolved in configuration:

roles:
  driver:     local/<model>   # drives the loop: conversation, plans, implementation
  helper:     <frontier>      # delegations: takes over a scoped sub-task
  expert:     <frontier>      # ask_expert: one question, no tools
  classifier: <model>         # the action classifier (L4)

The resolved mapping is pinned when a work item opens and recorded in its log, so a configuration change can't split a work item across two models.

Capability profiles. Each model in configuration carries a short profile, filled in from benchmark results: its tools, window, what it does reliably and where it is weak. Prompts that brief one role about another use the profile, never a model name or a vague "weaker model".

Escalation. The driver calls a frontier helper:

• when it chooses to — its prompt encourages asking when unsure, and treats asking as normal;
• on observable triggers — repeated failed actions, a context that won't fit, hitting turn or time caps.

Delegation. A frontier model takes over a scoped sub-task with tools, works autonomously, and hands back.

• One harness, not one per vendor. Vendor agent SDKs would bring back the lock-in the design avoids.
• Scoped by the agreed plan. A delegation can't widen what the work item was agreed to do; its actions pass through the same three layers (P4).
• Branch at delegation start; abandoning it discards the branch.
• Hard caps on turns, tokens and wall-clock time.
• Checkpointed as it goes, so it can resume after a crash (L8).
• Handback is a note written for the driver: what changed, what is now true, what to do next. That note plus the resulting files is the whole interface back.

ask_expert. One stateless question with a text answer, for things needing no repository access. A per-work-item budget stops a confused driver circling.

Local-model accommodations.

• Tool-call enforcement: tool_choice is requested only of models whose server declares it can enforce it; otherwise the loop enforces the contract itself (an explicit finish, prose treated as a violation).
• A reduced tool schema for roles on small models: fewer tools, flatter arguments.

Evaluation capture. Delegations, ask_expert calls and classifier verdicts are recorded as input/output pairs from the start, with outcomes marked. Changing a model is then a configuration change and a re-run, not a matter of benchmark chatter.

L10. Knowledge store (Outline)

Durable knowledge that outlives work items: world models, agent memory (L6), and later operator-written procedures.

Layout: one writer per document.

• One collection per writing principal: nisse-ops-world, nisse-dev-world, each agent's memory, and procedures (the operator's). Sharing grants read access only; concurrent writers never exist.
• nisse-dev gets private space only. Shared infrastructure documents are a map of exactly what its sealed network keeps it from (D7).
• Each agent authenticates as its own Outline user, with its own API token. A token from the operator's account would silently void every permission here.
• Collections and documents are identified by ID from configuration and state, never discovered, and document IDs recorded when created make writes idempotent.

Never an instruction channel, never an approval channel. Work is never read out of the knowledge store, and an agent never writes there to make another agent act. Coordination between agents goes through the operator, Matrix, or a ticket.

Corrections go through conversation. The operator tells the agent, which records the correction verbatim (L6). The operator keeps write permission as a break-glass: stop the agent, edit, restart. An agent re-reads a document's revision before writing and merges rather than overwrites if it moved.

Retrieval is structural first. Procedures are bound to entries in ops policy (O2) and fetched whenever a plan uses that entry; search exists in addition, never instead.

Outages. A failed write never blocks work (retry, or record locally and reconcile). A failed read of a bound procedure stops planning that action; other failed reads proceed and log the gap.

L11. Prompt injection

Any model that reads untrusted content may be steered by it. No marker reliably separates data from instructions. The design limits what a steered model can do rather than trying to stop it being steered.

Untrusted content: issue, comment and ticket text from non-operators, repository contents, dependencies, web pages, alert payloads, log lines.

• Hard walls hold against a steered model (P4): nothing a model says can grant a permission, widen an approved scope, or reach a credential that isn't mounted.
• Approval messages are rendered by code from what will run, never from a model's description.
• The classifier never sees untrusted content (L4).
• Instruction channels are operator-scoped: Matrix from allowlisted identities, comments from allowlisted Gitea and Vikunja accounts. Everything else is data.
• Intake is parsed into structured fields in code where the source allows it, so less free text reaches a model, and what does is labelled by source.
• Some injection is accepted because it can only reorder work: an alert title can bias which ready item is chosen next; the switch notice makes it visible.
• Each agent is checked against the lethal trifecta — private data, untrusted content, and a way to send data out. Damage needs all three. nisse-dev holds little private data and can reach only LiteLLM, Gitea and package registries; nisse-ops has all three legs and relies on its hard walls and step sequences (O2).

L12. Secrets, identity and logging

Secrets. Fetched at run time from Vaultwarden machine accounts, never baked into images. One minimally scoped machine account per agent and one per MCP server instance (X3); each has a single bootstrap token on its own host, and nothing else is stored there. Only credentials go in secret stores; every other setting lives in reviewable configuration files.

Identity. Each agent is its own LLDAP principal (svc-nisse-dev, svc-nisse-ops) in a tightly scoped Authelia group, never the operator's account. Headless agents get a single-factor exception scoped by source address, group and target.

Logging. Structured JSON lines to the host journal and the external log stack, not kept in containers. Every durable state change emits exactly one line, written by the state store right after the commit, with a fixed envelope (time, agent, event, work item, cause). Every classifier verdict, escalation and external effect is logged. Logs are diagnostic, never load-bearing: nothing reads them to derive state.

---

Part D — nisse-dev

Designs and builds software projects, producing reviewed changes in Gitea.

D1. How it works

A dev work item is a task on one repository: a feature, a fix, a new project. Through conversation, the agent and the operator agree on an architecture (for a new project, or a significant change) and a build plan; the agent then works through the plan's milestones, each producing a PR, and keeps the project's documents up to date as it goes.

use <project> opens or resumes a work item on a repository. Messages sent before it are kept for it (L5).

D2. Documents are the workflow

The project's own documents carry the agreement between the operator and the agent, and live in its repository:

• ARCHITECTURE.md — what the project is, how it fits together, and why. The durable "what and why". Created with a new project; changed whenever the design changes.
• Build plans — how a piece of work gets done, as a sequence of milestones. Short-lived: one per body of work, kept in the repo for the record.

Both are always discussed. The agent drafts, the operator and agent talk it through, and the document lands as a PR the operator approves. An approved architecture and build plan are the scope (L3) for everything that follows.

A build plan specifies; it doesn't implement. Each milestone states its goal, the interfaces and behaviour it adds or changes, edge cases, and tests as acceptance criteria. It includes code only where the code is the clearest specification (a tricky algorithm, an exact format). It never manages branches (D4). How much detail it needs is informed by the driver's capability profile (L9).

Each milestone is marked pre-approved (proceed when reached) or needs discussion (stop and talk first), and says whether it is expected to change code (a review or investigation milestone may not).

The Gitea issue for the work item carries the conversation's record: plans and approvals are posted there as comments, so the agreement survives loss of the agent's state (L8).

D3. The workspace

A fresh, clean workspace at use <project>.

• Fetch from the repository, and check out a clean tree at the latest base branch.
• Anything left in a previous workspace (uncommitted changes, local branches, unpushed commits) is set aside, never silently discarded or carried over, and reported to the operator.
• Build artifacts and caches never survive into a new task.

The milestones of one work item share that workspace, since they build on each other, often before earlier PRs are merged.

Toolchain preflight. When the workspace is prepared, the agent determines what the project needs (from its files, e.g. Cargo.toml, pyproject.toml, package.json, or a declaration in the repo) and checks it is available. A gap is reported before any planning, with the options. The environment's toolchain versions become facts in the transcript (L6), so plans match reality.

Toolchains come from the environment, not from the agent improvising. Installing software is class 3 (P3). How toolchains are provided (project dev containers on the VM's Docker daemon, or the VM image) is an open question (Q).

Repairing the environment is in scope. Within a task, the agent may clean caches and build artifacts or reinstall project dependencies to fix a broken build, and reports what it did. It doesn't stop and wait for permission to delete a corrupt .pyc.

D4. Git, branches and PRs

• The harness owns branches. Each milestone works on a branch the harness creates from the work item's current state; plans and models never create or switch branches.
• Only the milestone's own changes are published: what the agent changed, never build artifacts (standard excludes in the clone, and committing only what the work produced). Anything left out is reported.
• "Changes to publish" are counted from where the milestone started, never from a possibly stale remote branch.
• Publishing is idempotent (L8): an open PR for the branch is adopted and updated, never duplicated.
• A milestone that completes without changes is done, not failed, when it wasn't expected to change code. The agent reports its result and moves on as the build plan says. An unexpected "no changes" is raised with the operator.
• Moving past a milestone checks its PR's state in Gitea. If it is still open and unmerged, the agent says so and asks, rather than leaving it behind.
• The project's default branch is protected, and this is verified, not assumed: the agent's own token must be refused when pushing to it, and when merging without whatever approval the trust dial requires (L3). A test asserts both, so relaxing protection fails loudly.
• A runnable test service's address goes in the PR description.

D5. Review and merge

• The review loop: operator comments on a PR (the conversation and inline line comments) are instructions; the agent responds with commits to the same PR and replies in the thread where the comment was made.
• Intake is exactly as in L8: cursors, at-most-once taking, at-least-once answering, redelivery after a crash.
• Merging follows the trust dial (L3). Today the operator merges. As confidence grows, the dial may let the agent merge a pre-approved milestone once checks pass and the required approvals are recorded; the branch protection must then require an operator approval the agent's account can't satisfy, so the merge rule is enforced by Gitea, not by the agent.
• Finishing a work item closes its Gitea issue with a note (completed or abandoned), after checking what is left open.

D6. nisse-dev's action classes

Class	Examples
1	Reading the repo, running tests and builds in the workspace or a test container, querying Gitea
2 (in scope)	Editing the workspace, committing to milestone branches, pushing them, opening and updating PRs, commenting on its issues and PRs, cleaning caches, installing project dependencies from package registries
3	Merging to the default branch (unless the trust dial allows), deleting remote branches or others' work, installing system software, changing repository settings, anything outside the work item's repo

Its scope sources are the approved architecture, the build plan and each milestone's marking.

D7. Deployment and isolation

• Its own VM (X1), with a real Docker daemon, because it builds and runs containers. The VM is the trust boundary for unattended command execution.
• Sealed: its egress allowlist (X2) permits only LiteLLM, Gitea and package registries; there is no ingress. Denial covers RFC1918 (except the specific LiteLLM and Gitea addresses) and the tailnet's 100.64.0.0/10, so it can't reach the NAS, Proxmox or the gateways. The inference host sits behind LiteLLM, so it is not a second exception.
• No Tailscale on this VM; it would route around the firewall.
• Its own VLAN, and test services on predictable ports (e.g. 18000–18099), so opening access later is a firewall change.
• Test containers on an internal Docker network with no route out.
• Build storage on a separate mount from the state volume.
• A browser (a shared Playwright tool, headless) reaches only the internal test network, to exercise web apps the agent built and reproduce bugs. No SSO.

---

Part O — nisse-ops

Operates and maintains infrastructure across both sites, driven by tickets and duties, with stricter controls because live changes have no review step before they take effect.

O1. How it differs

Axis	nisse-dev	nisse-ops
Undo	discard a branch, revert a snapshot	often none (zfs destroy, a firewall lockout)
Scope source	architecture and build plan	approved step sequences (O2)
Default for class 3	discuss	discuss, and exact-step approval
Work shape	milestones → PRs	short reactive tasks and scheduled checks → live changes
Driver of work	the operator in Matrix, Gitea	Vikunja tickets, alerts, duties, the operator
Placement	sealed, own VLAN	Tailscale, credentials, broad reach

The two are separate deployments over the same library; one instance playing both roles would collapse the isolation.

O2. Approved step sequences

Every infrastructure write needs an approved plan first, and approval covers exactly the steps written. This is nisse-ops' scope source for the classifier (L4) and a hard wall in its own right.

• Write credentials exist only in elevated MCP instances (X3), mounted after approval; read-only instances are mounted by default, so looking around needs no plan.
• A write plan is a list of literal steps, each naming the tool and target, the full arguments (a config change carries its exact diff), an optional precondition checked against the live system just before the step, an optional expected outcome checked just after, and a retry allowance (default one). Prose may explain the plan; it is never scope.
• The approval message is rendered by code from the steps, flagging at the top anything irreversible, anything touching a protected target, and anything that takes down the agent's own host.
• Execution is matched against the steps. Read-only calls are always allowed. An elevated call runs only if it matches the next unfinished step exactly (arguments normalised against the tool's schema). Anything else halts and asks: a call not in the plan, changed arguments, steps out of order, a completed step run again. A failed precondition or expected outcome halts without retry.
• Approvals expire. A plan not started within the expiry window must be approved again.
• Policy is small, conservative, and data: a versioned file declaring each MCP instance's privilege, protected targets (the NAS, gateways, the agents' own hosts), and each elevated tool's target argument. Anything unclassified is treated as elevated and protected. Server-declared tool hints can make a tool stricter, never looser. A model may propose entries for a new server as a PR; no classification happens at run time.

Parked, to add when real friction shows the shape: bounded matching (any call within an approved target and action, for reversible tools, one tool at a time), and standing approvals (a plan template with a trigger, rate limit and expiry, for a routine fix approved by hand often enough).

O3. nisse-ops' action classes

Class	Examples
1	Metrics, logs, status, read-only MCP calls, reading tickets and the world model
2 (in scope)	Approved steps with a known undo; updating its own world model and tickets
3	Every other infrastructure write; browser actions on admin UIs (each confirmed, since a click isn't a stable literal); anything on network gear, the most conservatively handled action in the system, since a bad change can sever the path used to fix it

O4. Interfaces

• MCP servers (X3) for Proxmox and UniFi at both sites, one instance per service, site and privilege: the primary path to infrastructure.
• Browser (a shared Playwright tool) for admin UIs where no API exists; real APIs are preferred. For network gear it is a fallback only.
• SSO as svc-nisse-ops, whose blast radius equals its Authelia scope.
• Event intake as structured JSON (directly or via Alertmanager), never through a Matrix bridge. The agent posts curated updates to #ops.

O5. Tickets and duties

• A Vikunja ticket is a work item's external record (durable, with status and history); Matrix is the live conversation about it.
• Scheduled checks and alert intake are duties (L1). Tickets they generate are deduplicated findings, and like any auto-generated ticket can only lead to a plan the operator approves.
• Flow: ticket filed → agent picks it up → discuss over Matrix → status updates → closed with the result.
• Vikunja comments are an approval source for plans on their own tickets (approve rev-N). Live prompts stay in Matrix, with outcomes written back.
• Action-triggering tickets are operator-scoped: a ticket is an instruction channel into an agent that changes infrastructure.

O6. Deployment

• Its own long-lived, stateful VM (X1).
• Tailscale and scoped credentials to reach infrastructure at both sites.
• Its MCP servers run on its own Docker host; its egress allowlist therefore includes the Proxmox and UniFi endpoints at both sites.
• Updated in place, with restarts gated on an idle flag (X4).

---

Part X — Deployment model (all agents)

X1. Every agent gets its own VM

No agent runs as a container on a shared host, and no two agents share a VM.

• The VM is the isolation boundary between agents: a compromised or confused agent can't reach another's state, credentials or processes.
• Per-agent attribution: everything an agent holds is issued to its own identity, so every upstream action traces to one agent.

X2. Egress policy per agent instance

• Default-deny allowlist per agent instance, declared as a versioned artifact next to the agent's definition, each entry justified by the component that needs it.
• Declared in one place, enforced in several: at the VM, the Docker network, and the UniFi firewall. Network-layer enforcement survives compromise of the agent's host.
• The allowlist is the union of what the agent and its MCP servers reach.
• The tailnet is 100.64.0.0/10, not RFC1918; deny it explicitly.

X3. MCP servers

• One instance per agent VM, as containers on that agent's Docker host. A shared process would mean a shared credential, losing attribution in the upstream API's own audit log. Duplication is nearly free for stateless servers.
• One instance per (service, site, privilege). The site is part of the server's identity, so a wrong-site action needs a tool that isn't mounted, not just a wrong argument. Read-only and elevated instances have separate credentials, each from its own machine account.
• Per-task tool profiles mount only the servers a work item's scope needs.
• Rejected: search-then-dispatch meta-tools over large generated catalogues: extra round trips for a local driver, no discovery of capabilities not already sought, opaque arguments, and a dynamic tool list that defeats prompt caching.

X4. Containers, images and restarts

• Stateless images, named volumes for state.
• Delivery: Gitea Actions → Gitea's OCI registry → Watchtower pulls.
• Gate the restart, not the pull: an idle flag; Watchtower never restarts an agent mid-work. Unplanned interruptions are handled by recovery (L8).
• Third-party MCP server images are digest-pinned and updated deliberately, never by Watchtower: they hold root-equivalent credentials.
• Survive reboots: VMs and Docker start on boot, containers restart unless stopped, and the state dataset has ZFS snapshots (scheduled, and on each approval).

X5. Supporting services

All self-hosted: Matrix (Synapse), Gitea, LiteLLM, the local inference host (llama-swap on an RTX 4090 host), Vikunja, Outline, Vaultwarden, LLDAP with Authelia. Each agent gets its own account, scoped token and room in each service it uses; nothing is shared between agents.

---

Part Q — Open questions

• The action classifier's model and placement. A cheap frontier model, or the local model with a different prompt; how its verdicts are calibrated against operator overrides.
• Recognising questions. An explicit marker, or classifying free text; either way a misread must default to asking (L5).
• How far the local driver goes. Which kinds of work it handles well and where a frontier driver is needed, measured on real tasks, not assumed (P7).
• Providing toolchains for nisse-dev: project dev containers on its Docker daemon, or the VM image, or both (D3).
• One local model for every agent. The inference host keeps one model resident; different models per agent would reload weights on every switch. Preferred: one model good enough for every local role. Fallback: agents take turns holding the host. Measured before deciding.
• Vision. Needed for 3D design work (render and look) and video-file management; not for ordinary code. Tools returning images, image content in the log and renderers, and image tokens in the context budget. Measured when that work starts.
• More than one active work item per agent. Park-and-switch first; concurrency costs per-item serialisation, a scheduler for the shared inference host, strict approval threading, separate workspaces per item, and per-item intent ownership.
• When a work item must have an external record. Plan-driven dev work and ticket-driven ops work already do; whether every item with an approval must.
• Fallback browser access to network gear: standing, or handed off per task.
• Relocating stateful MCP servers off agent VMs, one at a time, if one ever justifies losing per-agent attribution.
• Infrastructure as code (Terraform with bpg/proxmox and cloud-init) for agent VMs.

---

Appendix — What changed from the Build Specification

For reviewing this rewrite against docs/Architecture-Build-Spec.md:

• Threat model and control model (P2–P4). Checks are built around mistakes, injection and crashes, in three layers with an action classifier, rather than a rigid state machine gating every phase.
• The agent loop (L2) replaces the per-agent state machines (old A4, B1). What stays fixed is only what authority and durability need.
• Approvals as scoped permissions (L3), with pre-approved and needs-discussion milestones and an operator-set trust dial. Revision binding is kept and extended (task-wide revisions, stamping on arrival).
• The local model drives everything (P7, L9); frontier models are helpers. The old "planning always escalates" phase policy is dropped.
• Context and memory (L6) is new: transcript scoping, attribution, orchestrator facts, scope memory, agent memory, and how memory is written and used.
• Documents as the dev workflow (D2), a fresh workspace per task and toolchain preflight (D3), and git hygiene learned from the live runs (D4).
• Conversation (L5) is a first-class requirement, not a command interpreter.
• Kept, largely as they were: the work model, durability and reconciliation, LiteLLM and the canonical log, delegation, the knowledge store, prompt injection, secrets and identity, nisse-ops' approved step sequences, and the deployment model.
• Moved to the Development Plan (to be written): build order, supporting-service status, and interim arrangements.