Skip to content

The Peer Workspace

The Peer does not inhabit a chat window. It inhabits the mass through a workspace: the durable office where the operating pair can see what is true, why it is true, what remains open, what may be changed, and how any change will be verified.

Flow keeps that office in one repository containing product code, operational records, evidence lineage, decisions, runbooks, deterministic operations, and instructions for the Peer. That is one implementation, not a requirement that every stone become a monorepo. A workspace may point into databases, object stores, business systems, or other repositories. What matters is that their meaning, authority, and safe access paths are durable and legible from one working whole.

A capable model can reconstruct a surprising amount from files. Reconstruction is still recurring labour, and reconstruction from the wrong files can become confident fiction. The workspace should answer these questions without relying on a prior conversation:

Question Durable answer
What product am I inside? Its purpose, scope, coherence boundary, vocabulary, and current status.
What do we believe now? Governed reads over current truth, with uncertainty shown rather than collapsed.
Why do we believe it? Evidence, source authority, provenance, and the reviewed changes beneath current values.
What is unfinished? Pending proposals, unanswered asks, queued arrivals, failed work, and known gaps.
What may I do? Declared operations, invariants, actor authority, and exact approval gates.
How do I know it worked? Deterministic validation, tests, inspection paths, and recovery runbooks.

This is the minimum inheritance that lets another capable Peer sit down in the office. Session history may help, but it is not part of the bond. If changing models or losing a conversation erases an operating fact, that fact never entered durable state.

An AGENTS.md can make a workspace immediately inhabitable. It should tell the Peer what to read first, what truths and invariants must not be bypassed, how to inspect the product, which operations are permitted, where authority changes, and how to verify work. Nested instructions can become more specific beside the records or code they govern.

It should not attempt to contain the whole office. A giant prompt becomes a second, stale implementation of the product. The front door points to durable domain documents, governed reads, decision records, operation catalogs, and runbooks; those sources remain authoritative in their own scopes.

Nor can a generic Facet template declare the product’s truth. The operator and the reality of the domain must name the coherence boundary, sources, authority, consequential transitions, publication policy, and people outside the bond. Copying doctrine into a repository does not make those decisions.

The first surface Flow built for its operator was a reconciliation web client. The operator did not need it. The Peer needed a governed read path through which it could inspect current truth, open obligations, evidence, changes, provenance, reconciliation candidates, and pending work without reverse engineering storage on every session.

That distinction matters. Raw access is not a read path. A stored token repeated to the operator is not a fact, and a query assembled ad hoc can silently choose the wrong coherence boundary. Human and machine forms should travel together: the Peer gets structured values for computation and the product’s declared meaning for explanation. Absence, uncertainty, suppression, and source status must remain visible in both.

A good read path changes nothing. It is deterministic, grounded in the same presentation rules the product trusts elsewhere, and broad enough for the Peer to see the whole relevant account before proposing a local answer.

The Peer should not need to remember which files are safe to edit or recreate a mutation protocol from prose. Consequential capability belongs behind typed operations that declare their inputs, expected state, invariants, effects, and authority. Proposal and dry-run forms let the pair inspect an exact delta before the declared gate. Application either advances the coherent truth and appends its decision together or advances nothing.

This does not require compiling every action in advance. The pair may inspect, analyze, draft, and explore from safe primitives. When repeated mechanical labour becomes clear, it can harden into a read or drafter. Promotion moves labour away from the Peer; it does not move judgment away from its authority.

The workspace should carry the distinctions the truth machine depends on:

  • original evidence remains distinct from extraction and interpretation;
  • source authority is explicit and bounded;
  • current truth is readable without erasing uncertainty or ancestry;
  • pending, failed, stale, and completed work are distinguishable;
  • generated, approved, sent, acknowledged, and incorporated remain separate;
  • decisions retain their reasons, and prohibitions retain their graves; and
  • verification and recovery paths are close enough to the work that the Peer can use them without handing commands back to the operator.

These meanings need not share a storage engine. They do need one legible route through the workspace. Links to external systems are enough when the access contract and authority are durable; copying their contents into a repository merely to make the office look self-contained creates another truth.

The operator works inside this office and needs no facet. People outside the bond may receive a deliberately cut worksheet, form, edition, or file, but the mass must continue when they use another channel. Email, calls, spreadsheets, and direct system observations can all enter as evidence if the workspace can preserve their provenance and admit them through the same rails.

This makes the cut genuinely disposable. The workspace does not break when a facet is ignored, shared with another outside person, revised after contact, or deleted. The outside experience still deserves excellent design; it simply is not allowed to become the product’s only route back to truth.

When the Peer is unavailable, the office must still contain the last valid truth, every completed decision, accepted contributions, queued work, and the instructions needed to recover. Deterministic reads and operations should keep working within their declared limits. Adaptive capacity may pause; the business must not become false.

Flow has paid for this workspace shape through real operator use. Its first outside contact separately paid for the claim that cuts are not workspace dependencies. It has not proved that the same arrangement transfers whole to another domain. The portable candidate is the office’s questions and boundaries, not Flow’s repository layout or command catalog.

To commission that test in an existing or new product, continue with Bootstrap a Workspace.