stack/docs/prd/mosaic-stack.md

# PRD: Mosaic Stack

- Status: approved, version 1.2. Jason approved draft 0.4 unchanged as
  1.0 on 2026-10-05 (lead decision 61). He approved version 1.1, the
  Organization requirements, in decision 03130d08-ddce-480d-a601-e59f64744fc3
  on 2026-10-14. He approved version 1.2, the chief of staff, the
  assistant, threads and agent limits, through Mos on 2026-10-14 (lead
  decisions 141 and 142). It is never edited in place. Changes are a new
  version.
- Version 1.2 source, Jason's words relayed by Mos:
  - Q147 (full text in lead decision 141): "COS should focus on COS
    duties. This will include hire/terminate authority for agents, org
    structure, etc. ... This necessitates the creation of a new role for
    user interaction. ... The user should be allowed to have 1 - N
    sessions with this agent to allow for topics of focus in a given
    thread. ... The interaction agent in each session is the user's main
    agent."
  - Q148, Mos's read-back of Q147: "Q148: Yes".
  - Q149, on the 1.2 draft: "approved as long as the agent limits are
    tunable via "Settings" in the WebUI and via the CLI tooling."
  - Q150: "No hard cap on agent limits. All tunable." The five kinds of
    limit are quoted under Launching.
  - Q151, Mos's read-back of Q150, with the default that launches over a
    limit are refused and logged and a run that hits its token limit is
    stopped: "Q151: yes".
- Owner: Jason. Sage writes it from the PRDY interview.
- Template: PRDY "software" (`v1/packages/prdy/src/templates.ts`), filled
  by hand until PRDY is ported.
- Interview record: Sage's thread 1ef1e4f8. Round 1 is lead decision 45.
  Round 2 is lead decision 48. Round 3 is lead decision 53. Design
  inputs are lead decisions 43, 44, 46, 47 and 49 to 52.

## Introduction

### Objective

The working north star, ratified as lead decision 44: "Jason declares
businesses, projects and roles. Agents in those roles carry the work end
to end under declared policy, and only gated decisions reach him."

### Context

See `docs/plans/2026-10-04_foundation-direction.md`. Agents already do
real work here. But a person has to launch them, relay between them, hand
them credentials and correct them. Jason has no CLI and no complete WebUI,
and outstanding decisions don't collect anywhere.

## Users

For now the only user is Jason. The stack must still be installable by
outside users later, which means:
- an installer;
- no paths or names specific to Jason;
- configuration through the variable layers.

(Round 1, answer D.)

## Problem statement

In priority order (round 1):
1. **Admin falls on Jason.** He launches agents, relays between them,
   finds credentials and corrects them.
2. **Jason can't see what the agents are doing or what's waiting on
   him.**
3. **Agents drift from what the business needs.** Nothing ties a piece of
   work to a stated goal.

## Scope and non-goals

### Hard limits

The stack doesn't do any of these. Each one is either a gated decision or
off the table (round 1, "all five"):
- A. Spend money without Jason.
- B. Speak externally as Jason without his approval.
- C. Give agents Jason's personal credentials. Each role gets its own.
- D. Become a multi-tenant SaaS in v1. There is no hosted SaaS in v1.
- E. Replace Claude Code, Codex or Pi. It wraps them.

### Where it runs

Mosaic Stack stays self-hosted open source. The licence is open, and Jason
decides it later.

Self-hosted, on Jason's machines and homelab (round 1, "A first"). Cloud
workers aren't ruled out later.

### Out of scope for v1

Round 2, "all":
- businesses other than Mosaic Stack (SetSpark joins after v1);
- minting credentials (v1 hands out tokens Jason creates);
- Codex (v1 covers Pi and Claude Code, and Codex comes next);
- Pocket ID and SSO;
- access from beyond localhost.

## Requirements

Every task the stack creates cites one of these ids. A task with no
parent requirement is refused.

### Roles

- **REQ-ROLE-1.** Each role is a reviewed file, `roles/<name>.json` at
  version 2. The file holds:
  - a contract;
  - an authority map over a closed vocabulary of actions;
  - the credentials it needs, by service and scope.

  Any action the file doesn't list is gated.
- **REQ-ROLE-2.** Each role instance has one holder at a time. The
  assistant is the exception: it has one holder per thread at a time. A
  claim takes a lock and logs it, and releasing the role or ending the run
  frees it. The chief of staff may end another role's run, and that act is
  logged with its reason. Anyone else who revokes a stuck holder needs a
  resolved decision, and the revoke cites it.
- **REQ-ROLE-3.** An organization file declares the role instances,
  the arbiters and the credential references. The user and the chief of
  staff write it. The chief of staff's writes go through a broker verb
  that logs each change. No other role writes it. A missing or invalid
  file fails closed. (Q133, confirmed as Q140: "We need the human and the
  Chief of Staff to be able to edit the Business file(s)".)
- **REQ-ROLE-4.** The first organization is Mosaic Stack, with six role
  instances:
  - chief of staff;
  - assistant;
  - PM, held by Sage;
  - CTO, held by Darkwing;
  - coder;
  - reviewer.

  The schema can also declare CEO, CFO and the other C-suite roles. The
  assistant's name is set in the organization file. Jason's may be named
  Mos.

Added in version 1.2:

- **REQ-COS-1.** The chief of staff runs the organization. It hires and
  terminates role instances, owns the organization's structure and is
  the only role that launches sessions (REQ-LAUNCH-1). It takes work
  from the assistant and assigns it through the roles that do it. It
  never talks to the user. Gated decisions still reach the human
  through the inbox (REQ-DEC-2, REQ-DEC-4).
- **REQ-ASSIST-1.** The assistant is the user's main agent. Each user has
  one assistant role per organization. It talks with the user, and it
  passes work for the organization to the chief of staff. It doesn't
  direct other agents, create or assign tasks, or launch sessions.
- **REQ-ASSIST-2.** The broker limits the assistant's messages, not its
  contract alone:
  - the assistant may send only to the user, in the thread, and to the
    chief of staff;
  - only the user and the chief of staff may send to the assistant;
  - every refused send is logged with its reason.

### Variables

- **REQ-VAR-1.** Variables come in four layers: system, business, project
  and agent. Every key is declared in code with a type, the layers allowed
  to set it, and a merge rule.
  - The most specific layer wins for plain values.
  - Limits intersect, so a lower layer can only narrow them.
  - An unknown key is refused, and so is a key set at a layer it isn't
    allowed in.
- **REQ-VAR-2.** `~/.config/mosaic-dev/config.json` stays the only system
  config, and nothing writes it automatically. A secret appears only as a
  reference, either a file path or an environment variable name. Only
  the broker reads a secret's contents. It holds them in memory and
  never logs or writes them. Everything else checks a referenced file
  with `stat`.

### Credentials

- **REQ-CRED-1.** Every role instance has its own token for each service
  (Gitea, Vikunja). The broker holds them, and they never enter an agent's
  environment or files. In v1 Jason creates them by hand from a runbook
  Sage writes (round 3, 1A):
  - in Gitea, a new bot user per role, named `<business>-<role>-bot`,
    for example `mosaic-stack-pm-bot` (round 3, 2A; lead decision 58);
  - in Vikunja, a `bot-<business>-<role>` user per role, plus one
    read-only `bot-<business>-sync` that does all polling (lead
    decisions 52 and 58).
- **REQ-CRED-2.** A session that finds only founder credentials stops.

### Decisions

- **REQ-DEC-1.** Every decision is a record in an append-only SQLite
  table. The record holds:
  - the role and run that raised it;
  - its class and action;
  - the options and a recommendation;
  - who resolved it;
  - its timestamps.

  A correction is a new row.
- **REQ-DEC-2.** Routing follows the decision's class:
  - routine and within-role decisions are logged;
  - cross-role decisions go to the business's arbiter for that domain;
  - gated decisions go to the human.
- **REQ-DEC-3.** A human resolution, by an authorized human, comes only from a `mosaic` CLI session
  started outside any agent run. Agents reach the decision store only
  through a broker. The broker stamps role and run from the launch record.
- **REQ-DEC-4.** A gated decision that blocks work reaches an authorized human
  right away: in the CLI inbox, plus a Discord DM sent through the existing
  connector (#1509). Everything else goes into one daily digest at 08:00
  Central. (Round 2, answers 2A and 3A; round 3, 4A and 5A.)

### Messages

- **REQ-MSG-1.** Messages are addressed to a role. They are append-only
  and go to whoever currently holds the role. A message to or from the
  assistant also names its thread and goes to that thread's holder. When
  the assistant sends work to the chief of staff, the reply comes back to
  the same thread. T3 and Discord carry messages; the address is still
  the role.

### Tasks

- **REQ-TASK-1.** Tasks live in Vikunja, reached through `/api/v2` with a
  bot token for each role. Each task cites a requirement id. Each task
  field has one writing role, listed in the slice 1 brief's field table,
  and the assigned role writes the task's state. Vikunja's token scopes
  cover whole route groups, so the broker's verbs enforce this, not the
  tokens. A person's edits come in as events.
- **REQ-TASK-2.** Polling is the source of truth. Every 30 seconds the
  sync bot reads each project's open-task board and the tasks updated
  since the last check, so column moves and deletions of open tasks show
  up within one poll. An hourly full reconcile catches the rest. Slice 1
  uses no webhooks. (Lead decisions 51 and 52.)
- **REQ-TASK-3.** The installer offers two choices: point at an existing
  Vikunja, or deploy the bundled one. The bundled one is the unmodified
  upstream image. No Vikunja code enters the repository.
- **REQ-TASK-4.** The queue keeps the agents' lock, review rounds and
  audit log. No field syncs in both directions.

### Interface

- **REQ-CLI-1.** `mosaic` is the front door. From it Jason can:
  - talk to the assistant in a project thread they choose or create;
  - see the decision inbox and resolve decisions;
  - see tasks and agents;
  - follow the trail of any piece of work.
- **REQ-CLI-2.** The assistant session that `mosaic` talks to is one the stack
  launches and owns, running Pi or Claude Code without a window. The
  product doesn't depend on T3. (Round 2, answer 5A.)
- **REQ-WEB-1.** The WebUI shows the same data as the CLI: conversations,
  the inbox, tasks, agents and trails. It has a chat area that lists the
  user's projects and the named threads in each, with one conversation
  view per thread where the user talks to the assistant. The user creates
  and names projects and threads there without a terminal. Its Settings
  view shows every agent limit, the usage counted against each, and the
  model each role instance runs, and lets the user change them
  (REQ-LIMIT-1 to REQ-LIMIT-3).

### Launching

- **REQ-LAUNCH-1.** The chief of staff launches role sessions within the
  agent limits written in the organization file (REQ-LIMIT-1 to
  REQ-LIMIT-3):
  - role instances only;
  - a launch that would exceed any agent limit is refused, and the
    refusal is logged with the limit it hit;
  - every launch is logged with its harness, provider, model, and the
    account or key it runs on;
  - revocable with one word.

  Assistant thread sessions are the exception to who launches. The stack
  starts one when the user opens or writes to a thread, and ends it after
  the thread has been idle for the time the organization file sets. A
  running thread session counts against the limits like any other. An
  idle thread holds no session and counts against none. The stack starts
  the chief of staff itself. (Round 2, answer 6A; Q147; Q150.)

Added in version 1.2. Jason, Q150, verbatim: "No hard cap on agent
limits. All tunable. Agent limit: A parallel execution limit arbitrarily
set by the user or COS. This limit can apply to: overall executing agent
limit across all models; provider parallel agent limit; provider parallel
agent limit by model; daily quota limit based on OAuth account limit or
CLIProxyAPI config in Mosaic Stack; absolute token limit per agent,
model, provider, or API key".

- **REQ-LIMIT-1.** Agent limits are settings with no ceiling in code. The
  user and the chief of staff view and change them from the WebUI's
  Settings view and from a `mosaic` command, without editing a file. The
  chief of staff changes them through its organization-file broker verb
  (REQ-ROLE-3). The same places view and change the model each role
  instance runs. The running-session limits are:
  - the total number of sessions running at once, across all models;
  - the number running at once per provider;
  - the number running at once per provider and model. The default is 4
    Opus and 4 Sonnet, as a default that can be changed, not a cap.

  Any kind may be left unset, which means no limit of that kind. A value
  is a whole number of 0 or more, and 0 stops new launches of that kind.
  A change is checked by the same rules as the organization file, then
  written to it and logged with who made it and the old and new values.
  An invalid value is refused, and the old value stays. A change applies
  to the next launch without a restart, and lowering a limit doesn't end
  sessions already running.
- **REQ-LIMIT-2.** A daily quota per OAuth account, or per client key of
  the CLIProxyAPI that Mosaic Stack configures. The stack counts the
  tokens its sessions use against the account or key each ran on, and the
  day resets at a time and time zone set with the quota. When an account
  or key reaches its quota, new launches on it are refused, and running
  sessions on it stop at the end of their current turn. Both are logged.
  The count covers sessions the stack ran. Use of the same account
  outside the stack isn't counted, and Settings says so. Measuring the
  quota in tokens is a draft position: Q150 names no unit, and OAuth
  plans meter by usage windows, not daily tokens.
- **REQ-LIMIT-3.** A token limit per role instance, model, provider or API
  key. Each limit counts the tokens its subject has used since the limit
  was set or last reset, and the user and the chief of staff can reset
  it. A count is the sum of every kind the harness reports (input,
  output, cache write and cache read), and Settings shows each kind. A
  launch whose subject is at its limit is refused, and a running session
  that reaches a limit is stopped at the end of its current turn. Both
  are logged with the limit and the count.

### Meta-harness

- **REQ-HARN-1.** The meta-harness generates each session's launch bundle
  from the role and the resolved variables:
  - the prompt;
  - the policy;
  - skills;
  - typed tools for the vocabulary actions;
  - a manifest.

  Pi comes first, then Claude Code. Codex comes after v1.
- **REQ-HARN-2.** Four layers actually enforce the rules:
  - credential scope;
  - verbs that refuse without a resolved decision;
  - the harness's tool limit;
  - a container.

  Hooks only refuse early and log, and no document may call them
  enforcement. A probe run against the live harnesses must pass before any
  build relies on a block.

### Events

- **REQ-EVT-1.** Every step in slice 1 emits an event to the append-only
  store. Measures are computed from events, never from tagging chat
  messages by hand.

### Organizations

Added in version 1.1 (decision 03130d08).

- **REQ-ORG-1.** "Business" is renamed "organization". One install holds
  one or more organizations, and each organization has its own bus as the
  tenant boundary.
- **REQ-ORG-2.** The data model is install, then organization, then
  projects, then threads, so v1 doesn't lock in one human per
  organization.
- **REQ-USER-1.** Users and role-based access control exist at install
  and organization level. An agent belongs to one organization and
  reaches only its allowed projects and threads.
- **REQ-USER-2.** Only the user who owns a thread reads it or writes to
  it, plus that user's assistant and the chief of staff's replies on work
  the thread sent. This fills in REQ-USER-1's "threads" for the chat
  area. Added in version 1.2.
- **REQ-THREAD-1.** Threads live under projects.
- **REQ-THREAD-2.** The user creates projects and names threads inside
  them, from the CLI or the WebUI. Each thread is a separate assistant
  session. A thread's messages are kept. When the user comes back to a
  thread, after an idle stop or a restart, the assistant picks up where
  it left off with that history. One thread's conversation never appears
  in another's. Added in version 1.2.
- **REQ-TPL-1.** Templates exist for agent roles and for organization
  types.
- **REQ-FED-1.** Installs can federate with each other. This is post-v1.

## Acceptance criteria and success measures

v1 is done when Jason gives one-sentence requests to the assistant, in a
thread, from the CLI or the WebUI, and Mosaic Stack builds pieces of
itself from them, five pieces in a row. For each piece:
- only gated decisions reach Jason;
- every step leaves a trail in the CLI and the WebUI.

(Round 1, answer A; the assistant and threads from version 1.2.)

## Technical considerations

- Vikunja v2.7.0 is AGPL-3.0. Its API tokens can't create users or mint
  more tokens. A Mosaic-owned account mints one bot user and one scoped,
  expiring token per role. The owner password for that account is the one
  high-value Vikunja secret. Instance admin features need a paid licence,
  and the plan doesn't use them. Source:
  `agents/researcher/work/2026-10-04_vikunja-pocketid.md`.
- Gitea tokens never expire, and creating one needs a password or the host
  shell. Rotation will be a scheduled step with a record.
- Pocket ID v2.17.0 is BSD-2. It offers passkey-only login for people and
  a client-credentials flow, but it can't issue tokens that Vikunja or
  Gitea accept. Agents use per-service role tokens. Pocket ID is wired in
  after v1.
- Decisions, messages, role claims and events share one SQLite file,
  `<dataRoot>/bus/bus.sqlite`. A guard on every table refuses UPDATE,
  DELETE and REPLACE. Run records stay write-once files.
- None of the three harnesses' hooks fails closed by default except Pi's
  `tool_call`. Source: `agents/filbert/work/meta-harness-survey-2026-10-04.md`.
- Version 1.2:
  - The organization file's `projects` key names git repositories to
    clone (`packages/seat/src/checkout.mjs:282-296`). A REQ-ORG-2 project
    is the user's container for threads. The build has to keep the two
    apart, or define one in terms of the other, before the chat area
    stores projects.
  - Threads change the bus schema: messages, claims and deliveries gain a
    thread. The bus refuses a database whose schema doesn't match
    (`store.mjs:113-115`), so this needs a migration and a re-pin with
    downtime.
  - Claude Code resumes a conversation only inside one launch today, because
    its session directory and HOME are per run (`launcher.mjs:204-206`,
    `sandbox.mjs:157`). REQ-THREAD-2 needs a per-thread session store that
    outlives the run, or a rebuild of context from the thread's messages.
  - Limits today: `launch.max` is keyed by model family, and a launch's
    family is the one key its model name contains
    (`packages/seat/src/session.mjs:52-57`). Each count must be 0 to
    `LAUNCH_CEILING = { opus: 4, sonnet: 4 }`
    (`packages/business/src/vocabulary.mjs:86`, `business.mjs:129-137`).
    The launcher reads the file once at start (`launcher.mjs:123`), so a
    change needs a bus restart. REQ-LIMIT-1 removes the ceiling, replaces
    family matching with a model table that names each model's provider,
    and re-reads limits at each launch. Jason's file needs a one-time
    migration from `{ opus: 4, sonnet: 4 }`.
  - The stack doesn't know a session's provider today. The only provider
    setting is the system-level `execution.provider` (`vars.mjs:84-87`),
    read by `sessionModel` (`packages/harness/src/bundle.mjs:49-55`).
    Claude Code sessions reach whatever `ANTHROPIC_BASE_URL` points at, or
    Anthropic when it's unset.
  - Tokens are on disk today, and nothing reads them:
    - Claude Code: the adapter uses `--output-format text`
      (`adapters/claude/adapter.sh`), so the runner sees only the answer.
      Claude Code's own transcript, in the run's HOME at
      `home/.claude/projects/<dir>/<session id>.jsonl`, has `model` and
      `usage` (input, output, cache write, cache read tokens) on each
      answer line. Checked on run re32fc7bd8143, model claude-sonnet-5-5.
    - Pi: the session file in the run's session directory has `provider`,
      `model` and `usage` (input, output, cacheRead, cacheWrite,
      totalTokens, cost) on each answer line. Checked on a Researcher
      session, provider zai.
    - The launch log has harness, model and family per launch, and no
      tokens (`launcher.mjs:269`, `:294`).
  - Today every Claude Code session runs on one credential, which the bus
    unit's environment file sets (`SESSION_CREDENTIALS`,
    `packages/seat/src/session.mjs:76`, decision 119). The stack has no
    name for that account. A quota per account needs the organization file
    to name each account and its credential reference (never the secret,
    invariant 3), and the launch to pick one and log its name. More than
    one account at a time needs the bus to hand a credential to each
    session, which is the open question in the #237 design row.
  - CLIProxyAPI, from Researcher's stand-in test (commit 587e64a27,
    `agents/researcher/work/i1-proxy-test/REPORT.md`): usage counters are
    kept per upstream credential, not per client key. Per-request events
    name the client key, but they last 60 seconds, a read removes them, and
    they carry the full plaintext key. All of it is behind the management
    API, which no agent calls. The stand-in didn't model token counts.
    So the stack can't read per-key usage from the proxy today. It counts
    its own sessions' tokens instead, and a per-key quota needs the stack
    to issue the keys (#237).
  - Token limits stop a session at the end of a turn, because the runner
    learns usage only when the turn ends. A long turn can go past its
    limit. Stopping inside a turn needs the runner to read streamed usage
    from each harness, which changes the adapter contract (stdout is the
    answer only, `adapters/README.md`).

## Risks and open questions

- In slice 1, agents run as Jason's OS user (round 3, 3A). They can
  write anything that user can write. Until seats run in a container,
  which comes next, the broker is a rule they follow, not a wall. What
  does hold in slice 1 is that agents never hold service tokens.
- Vikunja doesn't enforce `If-Match`. The broker compares before it
  writes, and a person editing in the same moment can still be
  overwritten in an owned field (lead decision 51).
- The Vikunja owner password and Gitea token creation are gated. Jason
  holds them.
- With no ceiling in code, the chief of staff can raise its own launch
  capacity and spend (REQ-LIMIT-1, Q150). The controls are the change
  log, with who made each change and the old and new values, and the
  user's power to change it back.
- A quota or token limit counts only what the stack ran. It can't see the
  provider's own remaining quota or use of the same account elsewhere.

## Milestones

Slice 1, in the order on the foundation direction page. The slice 1 brief
maps each step to requirement ids. CHAT-03's Gate E demonstration happens
during the slice 1 WebUI step, not separately (round 3, 7B).