Skip to content
theofor teamsworking practice / not a workspace

Share the workflow.
Name the responsibilities.

Build your model orchestration with skills for the job and built-in MCIR memory for relevant history. Give sources, routing, tools, and memory boundaries named owners, so continuity never becomes unrestricted access.

A team rule worth keeping

No invisible change to an important path.

The practices below belong in your repository, ticketing process, and application. They are not a promise of built-in approvals or live co-editing.

01who owns what

Different questions need different reviewers.

Workflow maintainer
Owns the route and configuration. Checks that a reachable model, the intended per-model skills, and required policy bindings remain on the chosen path.
Knowledge owner
Approves source documents and revisions. Resolves conflicting guidance, corrects stale remembered context, and decides when a source should no longer be used.
Application engineer
Owns permissions, tool argument validation, error handling, and the human-review experience. Separates draft generation from external side effects.
Operational reviewer
Knows whether the result is fit for its actual purpose. Can reject the draft, request missing evidence, or escalate beyond the workflow.

02an authored review specimen

Make the change smaller than the review.

Example change record · revision B

Route equipment-return questions to a source-backed draft.

PreviousA general operations answer with no required source reference.
ProposedUse the approved equipment excerpt and require the draft to preserve its source label.
UnchangedNo write tools. Process owner reviews the final answer. A policy is enforced only if explicitly bound.
Review fixturesA supported question; a missing source; an unrelated request; an attempted write; an unbound-policy configuration.
Restore pointKeep the previous configuration and source revision in your own versioned record.

Review the route, not just the nicest output.

A short successful draft can hide a broader permission, an unrelated source, or a policy that was never bound. Review what changed in the configuration and what the negative fixtures now do.

Record the intended model capability, source revision, tool scope, policy identity, and owner alongside the change. A version label is useful only if it points to the configuration actually used.

The record shown here is a suggested working document. It is not an exported theo schema, a built-in audit receipt, or an importable recipe.

03a repeatable practice

Propose. Inspect. Approve. Observe.

An application-owned change-review loop
  1. 01Propose

    Describe one behavior change and the sources, keys, or tools it affects.

  2. 02Inspect

    Replay synthetic positive and negative cases. Compare the path, not only the answer.

  3. 03Approve

    The appropriate owners review the evidence in your existing change process.

  4. 04Observe

    Check deployed behavior, handle errors, and keep a known restore point.

This practice is not a collaboration feature claim. Production review, rollout, monitoring, and restoration are responsibilities of the deploying team.

04configuration is not consent

Shared does not mean universally enforced.

Does shared work mean everyone shares all memory?

No. MCIR is scoped to the authenticated user, account, or organization owner, not a global pool. Keep identity and resource permissions explicit when deploying agents or connecting another application. Cross-session continuity does not merge tenants or grant access.

Does this page describe built-in team approvals?

No. The roles, review record, and revision discipline are an operating practice for your team to implement with its existing tools. They do not imply built-in co-editing, sign-off queues, deployment branches, or source-control integration.

Can we replace human review with a guardrail?

Not for the responsibilities described here. A configured rule evaluates particular patterns or constraints; it does not determine whether a business, clinical, or legal decision is appropriate.

What should we test after changing a policy binding?

Check the real API key, policy status, phases and verdicts, plus the bound and unbound cases. Test tool denial, missing context, and the output behavior your client displays. A green browser simulation is not production enforcement evidence.

your next experiment

Review a workflow together without sharing real data.

Use a synthetic recipe as a common reference. Record what should happen when context is missing, a tool is denied, or a policy is unbound.