Canonical Flow Model
This model is the provider-neutral vocabulary for flow insights. It covers only concepts that change action for the MVP and leaves provider-specific shapes to adapters. Guardrail: model what changes action; reference the rest.
Concepts
Work item
- Definition: A unit of intent to change the system, visible to stakeholders (for example, a feature or defect slice).
- Why it changes action: It anchors ownership, review expectations, and validation scope.
- Minimum attributes: intent/title, current stage (implementation, review, validation), related change identifiers, accountable owner, service/team scope.
- Not: A full backlog ticket history or requirements document; details beyond the action-driving fields stay out of the model.
Change
- Definition: A concrete, reviewable modification to code, config, or content proposed by a work item.
- Why it changes action: Review, validation, and integration decisions attach to the change, not to the abstract work item.
- Minimum attributes: change identifier, target branch/environment, linked work item, submitter actor, created timestamp.
- Not: A deployment record or long-term audit log; it is the proposal under review.
Actor
- Definition: A person or automation that initiates, reviews, validates, or owns a change.
- Why it changes action: Signals must direct recommendations to the right accountable party.
- Minimum attributes: identity, role in the flow (author, reviewer, validator, owner), team or service affiliation.
- Not: A scheduler or workflow runner; those are tools, not accountable actors.
Review state
- Definition: The state of human review for a change.
- Why it changes action: Determines whether reviewers or authors need to act.
- States:
awaiting_review,under_review,changes_requested,approved,not_required. - Not: An automation gate or test result; that belongs to validation state.
Validation state
- Definition: The state of automated or manual checks that verify the change (tests, integration checks, policy gates).
- Why it changes action: Drives whether the change can advance or needs remediation.
- States:
pending,running,passed,failed,flaky,skipped. - Minimum attributes: scope (branch vs. trunk), last run timestamp, evidence link.
- Not: A deployment stage; deployment belongs to implementation stacks, not this model.
Evidence state
- Definition: The status of required evidence (security attestations, approvals, traceability artifacts) needed for compliance or risk acceptance.
- Why it changes action: Work may be blocked waiting for evidence even when implementation is complete.
- States:
not_required,required,pending,recorded,stale. - Minimum attributes: required evidence types, owner responsible for supplying them, freshness timestamp.
- Not: A full document store or audit archive; only pointers and freshness matter here.
Ownership state
- Definition: The clarity of accountability for a work item or change.
- Why it changes action: Signals need a clear escalation path when progress stalls.
- States:
owned,delegated,unclear,missing. - Minimum attributes: accountable owner, fallback team/service, last confirmation timestamp.
- Not: An org chart or CMDB graph; only the action-driving ownership signals are modeled.
Risk signal
- Definition: A structured indicator that a flow stage, service, or team carries heightened risk based on observed states.
- Why it changes action: Guides prioritization and escalation across competing work.
- Minimum attributes: scope (service/team/stage), contributing factors, confidence, recommended action.
- Not: A general health score or SLA metric; it is specific to the flow insight signals.
Illustrative provider mappings (non-contractual)
Adapters normalize provider-specific fields into the model above. Examples:
- GitHub: pull request → change; reviewers → actors; required checks → validation state; code owners → ownership state.
- GitLab: merge request → change; approvers → actors; pipelines → validation state; group or code owner rules → ownership state.
These mappings are illustrative only; the contract remains the canonical model definitions.